Pulling together the semester’s work into my final project.
While my overarching research agenda is concerned with the ways that ethical standards are communicated to, and disseminated through, academic and professional communities, the specific research question for this project is:
What can recent articles on experimental research in political sceince journals teach us about how ethical standards are discussed in publications?"
In order to explore this question, the following steps are taken.
The article collection process is described here.
Using code I’ve developed for this project, I will pull all the articles into R. In addition, I’ll pull in the csv of all the article citation data from EndNotes.
####Load in libraries:
####Load in csv of citations:
###Load in all articles as a single corpus:
####Apply the function to the directory and create corpus:
Question: my texts and corpus have 121 elements, which is the correct number. Why does textsSummary only have 100?
In addition, one article did not download properly as a pdf of the whole article, so I have fixed that.
####Check for metadata:
data frame with 0 columns and 121 rowsNo metadata available.
I am wondering if I do, in fact, need to pull and and create as corpus all the articles by journal so that I can create journal title metadata? and then join the corpora? I will do that for now as I don’t know how to extract the title from the texts or textCorpus.
First, pull in each journal individually (I am suppressing this code as well as all warnings for the sake of brevity, but am happy to provide it upon request).
####Add Journal identifier and unique ID to each summary and apply it to the Corpus:
Summary of each journal’s metadata:
[1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" [1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" [1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" [1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" [1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" [1] "Text" "Types" "Tokens" "Sentences" "Journal"
[6] "ID" ####Join individual corpora into one combined corpus:
Text
1 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Able and Mostly Willing_ An Empirical Anatomy of Information's Effect on Voter‐Driven Accountability in Senegal.pdf
2 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/All in the Family_ Partisan Disagreement and Electoral Mobilization in Intimate Networks—A Spillover Experiment.pdf
3 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Are Voters More Likely to Contribute to Other Public Goods_ Evidence from a Large‐Scale Randomized Policy Experiment.pdf
4 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf
5 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Brokers, Social Networks, Reciprocity, and Clientelism.pdf
6 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Building Cooperation among Groups in Conflict_ An Experiment on Intersectarian Cooperation in Lebanon.pdf
7 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Campaign Contributions Facilitate Access to Congressional Officials A Randomized.pdf
8 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Campaign Finance Transparency Affects Legislators’ Election Outcomes and Behavior.pdf
9 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Can Incarcerated Felons Be (Re)integrated into the Political System_ Results from a Field Experiment.pdf
10 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Causes of Noncompliance with International Law_ A Field Experiment on Anonymous Incorporation.pdf
11 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Channels for Influence or Maps of Behavior_ A Field Experiment on Social Networks and Cooperation.pdf
12 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Comparing and Combining List and Endorsement Experiments_ Evidence from Afghanistan.pdf
13 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Compulsory Voting and Parties’ Vote‐Seeking Strategies.pdf
14 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Distorted Communication, Unequal Representation_ Constituents Communicate Less to Representatives Not of Their Race.pdf
15 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Do Perceptions of Ballot Secrecy Influence Turnout_ Results from a Field Experiment.pdf
16 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Do Politicians Discriminate Against Internal Migrants_ Evidence from Nationwide Field Experiments in India.pdf
17 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Does Race Affect Access to Government Services_ An Experiment Exploring Street‐Level Bureaucrats and Access to Public Housing.pdf
18 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Elections and Deceptions_ An Experimental Study on the Behavioral Effects of Democracy.pdf
19 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Electoral Institutions and Electoral Cycles in Investment Incentives_ A Field Experiment on Over 3,000 U.S. Municipalities_.pdf
20 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Encouraging Political Voices of Underrepresented Citizens through Coproduction_ Evidence from a Randomized Field Trial.pdf
21 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Ethnic Networks.pdf
22 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Explaining Explanations_ How Legislators Explain their Policy Positions and How Citizens React.pdf
23 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Greater Expectations_ A Field Experiment to Improve Accountability in Mali.pdf
24 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/How Markets Shape Values and Political Preferences A Field Experiment.pdf
25 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/How to Elect More Women_ Gender and Candidate Success in a Field Experiment.pdf
26 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Is Voting Habit Forming_ New Evidence from Experiments and Regression Discontinuities.pdf
27 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Non‐Governmental Monitoring of Local Governments Increases Compliance with Central Mandates_ A National‐Scale Field Experiment in China.pdf
28 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Norms versus Action_ Why Voters Fail to Sanction Malfeasance in Brazil.pdf
29 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Policy Deliberation and Voter Persuasion_ Experimental Evidence from an Election in the Philippines.pdf
30 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Political Determinants of Economic Exchange_ Evidence from a Business Experiment in Senegal.pdf
31 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Positive Spillovers from Negative Campaigning.pdf
32 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Preaching to the Choir_ Americans Prefer Communicating to Copartisan Elected Officials.pdf
33 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Social Esteem and Participation in Contentious Politics A Field Experiment.pdf
34 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Sources of Authoritarian Responsiveness- A Field Experiment in China.pdf
35 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Economic Consequences of Partisanship in a Polarized Era.pdf
36 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Effect of Fact‐Checking on Elites_ A Field Experiment on U.S. State Legislators.pdf
37 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Hijab Penalty_ Feminist Backlash to Muslim Immigrants.pdf
38 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Impact of Elections on Cooperation_ Evidence from a Lab‐in‐the‐Field Experiment in Uganda.pdf
39 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Moderating Effect of Debates on Political Attitudes.pdf
40 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/The Unequal Distribution of Opportunity_ A National Audit Study of Bureaucratic Discrimination in Primary School Access.pdf
41 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Universal Love or One True Religion_ Experimental Evidence of the Ambivalent Effect of Religious Ideas on Altruism and Discrimination.pdf
42 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Urbanization Patterns, Information Diffusion, and Female Voting in Rural Paraguay.pdf
43 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Which Narrative Strategies Durably Reduce Prejudice_ Evidence from Field and Survey Experiments Supporting the Efficacy of Perspective‐Getting.pdf
44 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Why Vote with the Chief_ Political Connections and Public Goods Provision in Zambia.pdf
45 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/An Experimental Test for “Backlash”.pdf
46 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Ballot Secrecy Concerns and Voter Mobilization.pdf
47 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Do Negatively Framed Messages.pdf
48 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/How Content Preferences Limit the Reach of Voting Aids.pdf
49 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Jury Service as Civic Engagement.pdf
50 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Mentoring the Next Generation of Women Candidates.pdf
51 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Mobilizing Latino Voters.pdf
52 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Partisan Social Pressure and Voter Mobilization.pdf
53 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Political Ambition and Constituent Service.pdf
54 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/The Comparative Effectiveness on Turnout .pdf
55 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/The Generalizability of Social Pressure Effects.pdf
56 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/Vote Self-Prediction Hardly Predicts Who Will Vote, and Is (Misleadingly) Unbiased.pdf
57 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Politics Research/When Does Increasing Mobilization Effort.pdf
58 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Can_Employment_Reduce_Lawlessness_and_Rebellion_A_Field_Experiment_with_High-Risk_Men_in_a_Fragile_S.pdf
59 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Can_Social_Contact_Reduce_Prejudice_and_Discrimination_Evidence_from_a_Field_Experiment_in_Nigeria.pdf
60 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Childhood_Skill_Development_and_Adult_Political_Participation.pdf
61 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Does_Electoral_Competition_Exacerbate_Interethnic_or_Interpartisan_Economic_Discrimination_Evidence_.pdf
62 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Educative_Interventions_to_Combat_Misinformation_Evidence_from_a_Field_Experiment_in_India.pdf
63 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Empowering_Women_through_Development_Aid_Evidence_from_a_Field_Experiment_in_Afghanistan.pdf
64 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Establishing_the_Rule_of_Law_in_Weak_and_War-torn_States_Evidence_from_a_Field_Experiment_with_the_L.pdf
65 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Group_Size_and_Protest_Mobilization_across_Movements_and_Countermovements.pdf
66 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/How_Does_Development_Assistance_Affect_Collective_Action_Capacity_Results_from_a_Field_Experiment_in.pdf
67 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Information_Dissemination_Competitive_Pressure_and_Politician_Performance_between_Elections_A_Field_.pdf
68 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Information_Provision_Voter_Coordination_and_Electoral_Accountability_Evidence_from_Mexican_Social_N.pdf
69 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/International_Interventions_to_Build_Social_Capital_Evidence_from_a_Field_Experiment_in_Sudan.pdf
70 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Is_Deliberation_an_Antidote_to_Extreme_Partisan_Polarization_Reflections_on_America_in_One_Room.pdf
71 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Is_Position-Taking_Contagious_Evidence_of_Cue-Taking_from_Two_Field_Experiments_in_a_State_Legislatu.pdf
72 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/On_the_Limits_of_Officials_Ability_to_Change_Citizens_Priorities_A_Field_Experiment_in_Local_Politic.pdf
73 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Outside_Lobbying_over_the_Airwaves_A_Randomized_Field_Experiment_on_Televised_Issue_Ads.pdf
74 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Participation_Government_Legitimacy_and_Regulatory_Compliance_in_Emerging_Economies_A_Firm-Level_Fie.pdf
75 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/Reducing_Exclusionary_Attitudes_through_Interpersonal_Conversation_Evidence_from_Three_Field_Experim.pdf
76 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/The_Adverse_Effects_of_Sunshine_A_Field_Experiment_on_Legislative_Transparency_in_an_Authoritarian_A.pdf
77 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/The_Curse_of_Good_Intentions_Why_Anticorruption_Messaging_Can_Encourage_Bribery.pdf
78 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/The_Minimal_Persuasive_Effects_of_Campaign_Contact_in_General_Elections_Evidence_from_49_Field_Exper.pdf
79 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/The_Organizational_Roots_of_Political_Activism_Field_Experiments_on_Creating_a_Relational_Context.pdf
80 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/The_Psychology_of_State_Repression_Fear_and_Dissent_Decisions_in_Zimbabwe.pdf
81 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/What_Do_I_Need_to_Vote_Bureaucratic_Discretion_and_Discrimination_by_Local_Election_Officials.pdf
82 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Political Science Review/When_Are_Legislators_Responsive_to_Ethnic_Minorities_Testing_the_Role_of_Electoral_Incentives_and_Ca.pdf
83 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Building_Confidence_in_Elections_The_Case_of_Electoral_Monitors_in_Kosova.pdf
84 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Can_Conversing_with_a_Computer_Increase_Turnout_Mobilization_Using_Chatbot_Communication.pdf
85 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Choi-2021-Linguistic-assimilation-does-not-re.pdf
86 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Coethnicity_and_Corruption_Field_Experimental_Evidence_from_Public_Officials_in_Malawi.pdf
87 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Commitment_to_Political_Ideology_is_a_Luxury_Only_Students_Can_Afford_A_Distributive_Justice_Experim.pdf
88 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Does_Deliberative_Education_Increase_Civic_Competence_Results_from_a_Field_Experiment.pdf
89 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Does_Exposure_to_Gender_Role_Models_Increase_Womens_Political_Ambition_A_Field_Experiment_with_Polit.pdf
90 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Does_the_Message_Matter_A_Field_Experiment_on_Political_Party_Recruitment.pdf
91 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Drowned_Out_by_the_Noise_The_Downstream_Mobilisation_Effects_of_Party_Campaigning_between_Local_and_.pdf
92 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Email_Mobilization_Messages_Suppress_Turnout_Among_Black_and_Latino_Voters_Experimental_Evidence_Fro.pdf
93 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Encouraging_Small_Donor_Contributions_A_Field_Experiment_Testing_the_Effects_of_Nonpartisan_Messages.pdf
94 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Friends-and-Neighbors_Mobilization_A_Field_Experimental_Replication_and_Extension.pdf
95 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/How_Much_GOTV_Mail_is_Too_Much_Results_from_a_Large-Scale_Field_Experiment.pdf
96 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/If_We_Build_It_Only_Some_Will_Come_An_Experimental_Study_of_Mobilization_for_Seattles_Democracy_Vouc.pdf
97 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Mobilizing_Candidates_Political_Actors_Strategically_Shape_the_Candidate_Pool_with_Personal_Appeals.pdf
98 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Response_Bias_in_Survey_Measures_of_Voter_Behavior_Implications_for_Measurement_and_Inference.pdf
99 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Service_Representation_in_a_Federal_System_A_Field_Experiment.pdf
100 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/When_Do_Politicians_Pursue_More_Policy_Information.pdf
101 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Journal of Experimental Political Science/Women_Want_an_Answer_Field_Experiments_on_Elected_Officials_and_Gender_Bias.pdf
102 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Does_Social_Media_Promote_Civic_Activism_A_Field_Experiment_with_a_Civic_Campaign.pdf
103 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Gaffe_Appeal_A_Field_Experiment_on_Partisan_Selective_Exposure_to_Election_Messages_.pdf
104 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/How_Does_Media_Influence_Social_Norms_Experimental_Evidence_on_the_Role_of_Common_Knowledge.pdf
105 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Is_it_worth_door-knocking_Evidence_from_a_United_Kingdom-based_Get_Out_The_Vote_GOTV_field_experimen.pdf
106 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Nongovernmental_Campaign_Communication_Providing_Ballot_Secrecy_Assurances_Increases_Turnout_Results.pdf
107 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Parties_are_No_Civic_Charities_Voter_Contact_and_the_Changing_Partisan_Composition_of_the_Electorate.pdf
108 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Pivotality_and_Turnout_Evidence_from_a_Field_Experiment_in_the_Aftermath_of_a_Tied_Election.pdf
109 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Temperature_and_outgroup_discrimination.pdf
110 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/The_motivational_basis_of_constituency_work_how_intrinsic_and_extrinsic_motivations_interact.pdf
111 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Political Science Research and Methods/Voting_Made_Safe_and_Easy_The_Impact_of_e-voting_on_Citizen_Perceptions.pdf
112 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Bottom-up accountability and public service provision.pdf
113 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Combining bottom-up monitoring and top-down accountability.pdf
114 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Does digital advertising affect vote choice.pdf
115 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Does direct democracy increase communicative responsiveness.pdf
116 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Intended and unintended consequences of democracy promotion assistance to Georgia after the Rose Revolution.pdf
117 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Mobilizing opposition voters under electoral authoritarianism.pdf
118 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Photo identification laws and perceptions of electoral fraud.pdf
119 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Studying policy diffusion at the individual level.pdf
120 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Traditional institutions and social cooperation.pdf
121 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/Research and Politics/Voting and civic engagement.pdf
Types Tokens Sentences Journal
1 3276 18676 460 Journal of Political Science
2 2722 13047 536 Journal of Political Science
3 2676 12122 543 Journal of Political Science
4 2919 13332 505 Journal of Political Science
5 2811 14311 528 Journal of Political Science
6 2683 13067 443 Journal of Political Science
7 2622 11819 515 Journal of Political Science
8 2818 14511 652 Journal of Political Science
9 2515 13466 469 Journal of Political Science
10 3222 14080 551 Journal of Political Science
11 2847 14570 544 Journal of Political Science
12 2886 16080 468 Journal of Political Science
13 2958 13633 680 Journal of Political Science
14 2687 13032 533 Journal of Political Science
15 2572 13149 462 Journal of Political Science
16 3415 13300 583 Journal of Political Science
17 3328 15037 666 Journal of Political Science
18 2844 12942 675 Journal of Political Science
19 3171 13966 531 Journal of Political Science
20 2515 11727 522 Journal of Political Science
21 2789 12336 422 Journal of Political Science
22 3300 17474 659 Journal of Political Science
23 2737 11950 430 Journal of Political Science
24 3147 15711 543 Journal of Political Science
25 2782 13223 513 Journal of Political Science
26 3094 15132 532 Journal of Political Science
27 3082 14813 577 Journal of Political Science
28 3118 13427 557 Journal of Political Science
29 2582 11630 336 Journal of Political Science
30 2957 13263 440 Journal of Political Science
31 2631 14997 495 Journal of Political Science
32 3076 13497 663 Journal of Political Science
33 2509 10648 422 Journal of Political Science
34 3356 15209 557 Journal of Political Science
35 2646 12495 416 Journal of Political Science
36 2599 11339 533 Journal of Political Science
37 3032 12848 538 Journal of Political Science
38 3735 19631 704 Journal of Political Science
39 2848 14514 517 Journal of Political Science
40 3036 13281 590 Journal of Political Science
41 3038 15261 638 Journal of Political Science
42 2968 15922 545 Journal of Political Science
43 2435 12818 450 Journal of Political Science
44 2704 13616 462 Journal of Political Science
45 2397 13281 638 American Politics Research
46 2228 14766 597 American Politics Research
47 1800 8759 456 American Politics Research
48 2393 13398 740 American Politics Research
49 2189 12247 560 American Politics Research
50 2334 12995 600 American Politics Research
51 2156 11057 563 American Politics Research
52 2134 12667 705 American Politics Research
53 1746 8784 352 American Politics Research
54 1338 7356 316 American Politics Research
55 2008 12119 422 American Politics Research
56 2101 13424 588 American Politics Research
57 2262 15013 385 American Politics Research
58 3845 17906 760 American Political Science Review
59 4379 21635 866 American Political Science Review
60 2974 13656 610 American Political Science Review
61 4045 20150 901 American Political Science Review
62 3487 15868 705 American Political Science Review
63 3250 17358 596 American Political Science Review
64 3438 17567 660 American Political Science Review
65 3188 16799 785 American Political Science Review
66 3440 19689 580 American Political Science Review
67 3826 17922 656 American Political Science Review
68 3589 21021 664 American Political Science Review
69 3808 24080 862 American Political Science Review
70 3509 17146 800 American Political Science Review
71 2594 12572 604 American Political Science Review
72 2534 13100 547 American Political Science Review
73 1532 6027 232 American Political Science Review
74 4189 21979 882 American Political Science Review
75 3121 16944 679 American Political Science Review
76 4767 25468 865 American Political Science Review
77 3122 15727 614 American Political Science Review
78 3516 19490 831 American Political Science Review
79 2672 12971 688 American Political Science Review
80 3336 16643 700 American Political Science Review
81 3129 15299 612 American Political Science Review
82 3159 16875 591 American Political Science Review
83 1056 5012 211 Journal of Experimental Political Research
84 1646 6644 388 Journal of Experimental Political Research
85 1646 6635 319 Journal of Experimental Political Research
86 1061 3309 185 Journal of Experimental Political Research
87 1338 5996 293 Journal of Experimental Political Research
88 1470 6319 306 Journal of Experimental Political Research
89 1495 6248 324 Journal of Experimental Political Research
90 1213 5200 258 Journal of Experimental Political Research
91 1293 6516 290 Journal of Experimental Political Research
92 1096 3614 205 Journal of Experimental Political Research
93 1209 3987 213 Journal of Experimental Political Research
94 1695 7684 358 Journal of Experimental Political Research
95 1508 6039 338 Journal of Experimental Political Research
96 2107 9834 500 Journal of Experimental Political Research
97 1853 8668 638 Journal of Experimental Political Research
98 950 3411 179 Journal of Experimental Political Research
99 1657 7938 368 Journal of Experimental Political Research
100 1323 5524 260 Journal of Experimental Political Research
101 1611 7631 485 Journal of Experimental Political Research
102 2437 11901 501 Political Science Research and Methods
103 3356 15290 839 Political Science Research and Methods
104 2259 11765 487 Political Science Research and Methods
105 1907 10609 425 Political Science Research and Methods
106 1574 8590 244 Political Science Research and Methods
107 2033 9901 435 Political Science Research and Methods
108 1678 7318 355 Political Science Research and Methods
109 1627 5806 234 Political Science Research and Methods
110 2222 11181 470 Political Science Research and Methods
111 2573 13463 552 Political Science Research and Methods
112 1543 5508 207 Research and Politics
113 1943 7012 208 Research and Politics
114 1740 5737 207 Research and Politics
115 1635 6412 235 Research and Politics
116 2535 10303 327 Research and Politics
117 1588 5618 188 Research and Politics
118 1565 6342 190 Research and Politics
119 1444 6199 243 Research and Politics
120 1682 6269 185 Research and Politics
121 775 2248 90 Research and Politics
ID
1 101
2 102
3 103
4 104
5 105
6 106
7 107
8 108
9 109
10 110
11 111
12 112
13 113
14 114
15 115
16 116
17 117
18 118
19 119
20 120
21 121
22 122
23 123
24 124
25 125
26 126
27 127
28 128
29 129
30 130
31 131
32 132
33 133
34 134
35 135
36 136
37 137
38 138
39 139
40 140
41 141
42 142
43 143
44 144
45 301
46 302
47 303
48 304
49 305
50 306
51 307
52 308
53 309
54 310
55 311
56 312
57 313
58 201
59 202
60 203
61 204
62 205
63 206
64 207
65 208
66 209
67 210
68 211
69 212
70 213
71 214
72 215
73 216
74 217
75 218
76 219
77 220
78 221
79 222
80 223
81 224
82 225
83 401
84 402
85 403
86 404
87 405
88 406
89 407
90 408
91 409
92 410
93 411
94 412
95 413
96 414
97 415
98 416
99 417
100 418
101 419
102 501
103 502
104 503
105 504
106 505
107 506
108 507
109 508
110 509
111 510
112 601
113 602
114 603
115 604
116 605
117 606
118 607
119 608
120 609
121 610So now I have a corpus with metadata including journal title.
How many documents in my combinedCorpus?
[1] 121Now I’m going to tokenize my combinedCorpus, as well as remove numbers and punctuation:
Tokens consisting of 121 documents and 6 docvars.
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Able and Mostly Willing_ An Empirical Anatomy of Information's Effect on Voter‐Driven Accountability in Senegal.pdf :
[1] "Able" "and" "Mostly" "Willing"
[5] "An" "Empirical" "Anatomy" "of"
[9] "Information's" "Effect" "on" "Voter-Driven"
[ ... and 12,579 more ]
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/All in the Family_ Partisan Disagreement and Electoral Mobilization in Intimate Networks—A Spillover Experiment.pdf :
[1] "All" "in" "the" "Family"
[5] "Partisan" "Disagreement" "and" "Electoral"
[9] "Mobilization" "in" "Intimate" "Networks"
[ ... and 10,019 more ]
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Are Voters More Likely to Contribute to Other Public Goods_ Evidence from a Large‐Scale Randomized Policy Experiment.pdf :
[1] "Are" "Voters" "More" "Likely" "to"
[6] "Contribute" "to" "Other" "Public" "Goods"
[11] "Evidence" "from"
[ ... and 8,824 more ]
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf :
[1] "Black" "Politicians" "Are" "More"
[5] "Intrinsically" "Motivated" "to" "Advance"
[9] "Blacks" "Interests" "A" "Field"
[ ... and 10,228 more ]
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Brokers, Social Networks, Reciprocity, and Clientelism.pdf :
[1] "Brokers" "Social" "Networks" "Reciprocity"
[5] "and" "Clientelism" "Nico" "Ravanilla"
[9] "University" "of" "California" "San"
[ ... and 10,936 more ]
/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Building Cooperation among Groups in Conflict_ An Experiment on Intersectarian Cooperation in Lebanon.pdf :
[1] "Building" "Cooperation" "among"
[4] "Groups" "in" "Conflict"
[7] "An" "Experiment" "on"
[10] "Intersectarian" "Cooperation" "in"
[ ... and 10,052 more ]
[ reached max_ndoc ... 115 more documents ]First, I’m going to do a simple term search on the word “ethics” (eventually I will want to use lemmatization, I believe but this is just a start)
Keyword-in-context with 6 matches.
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4963]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 9594]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Building Cooperation among Groups in Conflict_ An Experiment on Intersectarian Cooperation in Lebanon.pdf, 4969]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Campaign Finance Transparency Affects Legislators’ Election Outcomes and Behavior.pdf, 7527]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Channels for Influence or Maps of Behavior_ A Field Experiment on Social Networks and Cooperation.pdf, 4976]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Distorted Communication, Unequal Representation_ Constituents Communicate Less to Representatives Not of Their Race.pdf, 2648]
of the experidiscussion of the | ethics |
Staff Polity McClendon Gwyneth H | Ethics |
GROUPS IN CONFLICT Compensation and | Ethics |
observed used data on congressional | ethics |
network behave more similarly Hypothesis | Ethics |
any experiment on human subjects | ethics |
of experimenting on public offiment's
of Using Public Officials Haider-Markel
in Table The dependent variable
investigations from correlations between scandals
Second we have three experimental
were naturally raises questions about [1] 38Can I add other terms?
Keyword-in-context with 6 matches.
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 3872]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4366]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4400]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4712]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4963]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 9594]
C  T consistent with | ethical |
as follows First in words | Ethical |
their extrinsic motivations comexperiment several | ethical |
of their districts However the | ethical |
of the experidiscussion of the | ethics |
Staff Polity McClendon Gwyneth H | Ethics |
considerations see below If this
Considerations note that if nonblack
concerns were still important pose
status of example as a
of experimenting on public offiment's
of Using Public Officials Haider-Markel[1] 117I think I’ll just use the stem of “ethic” and see what that gets me:
Keyword-in-context with 6 matches.
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 3872]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4366]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4400]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4712]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 4963]
[/Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf, 9594]
C  T consistent with | ethical |
as follows First in words | Ethical |
their extrinsic motivations comexperiment several | ethical |
of their districts However the | ethical |
of the experidiscussion of the | ethics |
Staff Polity McClendon Gwyneth H | Ethics |
considerations see below If this
Considerations note that if nonblack
concerns were still important pose
status of example as a
of experimenting on public offiment's
of Using Public Officials Haider-Markel[1] 129So this gives us 129 matches of words starting with ethic.
Now I’m interested in how many documents it appears in:ethic
3 ethica
1 ethics
26 ethical
37 ethically
3 What I don’t know though, is if there is overlap amongst these words in one document, so I think this is where lemmatize is will be helpful.
I’m having trouble making this work so I’m going back to Week 4 NLP tools and stemming:
Text
1 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Able and Mostly Willing_ An Empirical Anatomy of Information's Effect on Voter‐Driven Accountability in Senegal.pdf
2 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/All in the Family_ Partisan Disagreement and Electoral Mobilization in Intimate Networks—A Spillover Experiment.pdf
3 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Are Voters More Likely to Contribute to Other Public Goods_ Evidence from a Large‐Scale Randomized Policy Experiment.pdf
4 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf
5 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Brokers, Social Networks, Reciprocity, and Clientelism.pdf
6 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Building Cooperation among Groups in Conflict_ An Experiment on Intersectarian Cooperation in Lebanon.pdf
Types Tokens Sentences Journal ID
1 3276 18676 460 Journal of Political Science 101
2 2722 13047 536 Journal of Political Science 102
3 2676 12122 543 Journal of Political Science 103
4 2919 13332 505 Journal of Political Science 104
5 2811 14311 528 Journal of Political Science 105
6 2683 13067 443 Journal of Political Science 106 Text
1 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Able and Mostly Willing_ An Empirical Anatomy of Information's Effect on Voter‐Driven Accountability in Senegal.pdf
2 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/All in the Family_ Partisan Disagreement and Electoral Mobilization in Intimate Networks—A Spillover Experiment.pdf
3 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Are Voters More Likely to Contribute to Other Public Goods_ Evidence from a Large‐Scale Randomized Policy Experiment.pdf
4 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests_ A Field Experiment Manipulating Political Incentives.pdf
5 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Brokers, Social Networks, Reciprocity, and Clientelism.pdf
6 /Users/lissie/DACSS/697D Text as Data/Final Project Materials/Articles pdfs/American Journal of Political Science/Building Cooperation among Groups in Conflict_ An Experiment on Intersectarian Cooperation in Lebanon.pdf
Types Tokens Sentences Journal ID
1 3276 18676 460 Journal of Political Science 101
2 2722 13047 536 Journal of Political Science 102
3 2676 12122 543 Journal of Political Science 103
4 2919 13332 505 Journal of Political Science 104
5 2811 14311 528 Journal of Political Science 105
6 2683 13067 443 Journal of Political Science 106
text
1 Able and Mostly Willing: An Empirical Anatomy of Information’s Effect on Voter-Driven Accountability in Senegal Abhit Bhandari Institute for Advanced Study in Toulouse Horacio Larreguy Harvard University, Institute for Advanced Study in Toulouse, Toulouse School of Economics John Marshall Columbia University Abstract: Political accountability may be constrained by the reach and relevance of information campaigns in developing democracies and—upon receiving information—voters’ ability and will to hold politicians accountable. To illuminate voter-level constraints and information relevance absent dissemination constraints, we conducted a field experiment around Senegal’s 2017 parliamentary elections to examine the core theoretical steps linking receiving different types of incumbent performance information to electoral and nonelectoral accountability. Voters immediately processed information as Bayesians, found temporally benchmarked local performance outcomes particularly informative, and updated their beliefs for at least a month. Learning that incumbents generally performed better than expected, voters durably requested greater politician contact after elections while incumbent vote choice increased among likely voters and voters prioritizing local projects when appraising incumbents. In contrast, information about incumbent duties did not systematically influence beliefs or accountability. These findings suggest voters were able and mostly willing to use relevant information to hold politicians to account. Verification Materials: The materials required to verify the computational reproducibility of the results, procedures, and analyses in this article are available on the American Journal of Political Science Dataverse within the Harvard Dataverse Network, at: https://doi.org/10.7910/DVN/QAJQXP. I nforming voters about their incumbent’s perforever, recent studies identifying the effects of informamance in office is thought to help citizens retional campaigns on electoral accountability (Banertain high-quality politicians (Fearon 1999) and hold jee et al. 2011; Chong et al. 2015; Cruz, Keefer, and politicians to account beyond the ballot box (Aker, ColLabonne 2019; Dunning et al. 2019; Ferraz and Finan lier, and Vicente 2017; Gottlieb 2016). In practice, how2008; Humphreys and Weinstein 2012) and nonelectoral Abhit Bhandari is Postdoctoral Research Fellow, Institute for Advanced Study in Toulouse, University of Toulouse Capitole, 1 Esplanade de l’Université, 31080 Toulouse, France (abhit.bhandari@iast.fr). Horacio Larreguy is Associate Professor, Government Department, Harvard University, 1737 Cambridge Street, Cambridge, MA 02138 (hlarreguy@fas.harvard.edu) and Visiting Researcher, Institute for Advanced Study in Toulouse and Toulouse School of Economics, University of Toulouse Capitole, 1 Esplanade de l’té, 31080 Toulouse, France (hlarreguy@tse-fr.eu). John Marshall is Assistant Professor, Department of Political Science, Columbia University, International Affairs Building, 420 West 118th Street, New York, NY 10027 (jm4401@columbia.edu). We thank Fodé Sarr and his team of enumerators for invaluable research assistance, and Elimane Kane, Thierno Niang, and LEGS-Africa for partnering with us to implement this project. We thank Antonella Bandiera, Nilesh Fernando, Matthew Gichohi, Lakshmi Iyer, Kate Orkin, Julia Payson, Amanda Robinson, Arturas Rozenas, Cyrus Samii, Moses Shayo, Jay Shon, Alberto Simpser, and participants at talks at APSA, MPSA, Notre Dame, NYU, NYU CESS Experimental Political Science Conference, and WGAPE-NYU Abu Dhabi for excellent comments. This project received financial support from the Spencer Foundation, and was approved by the Columbia Institutional Review Board (IRB-AAAR3724) and the Harvard Committee on the Use of Human Subjects (IRB17-0880). Bhandari and Larreguy acknowledge funding from the French National Research Agency (ANR) under the Investments for the Future program (Investissements d’Avenir, grant ANR-17-EURE-0010). Our preanalysis plan was registered with the Social Science Registry, and is available at https://www.socialscienceregistry.org/trials/2324. American Journal of Political Science, Vol. 00, No. 00, xxxx 2021, Pp. 1–27 ©2021, Midwest Political Science Association DOI: 10.1111/ajps.12591 1 2 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL political engagement (see Casey 2018; Lieberman, Postary elections to examine these voter-level mechanisms ner, and Tsai 2014) yield mixed findings. underpinning political accountability among deputies Given the complex chain of conditions linking seeking reelection for a second term. Across 450 rural vilthe provision of information to better governance lages from 5 of Senegal’s 45 districts, we trained enumer(Dunning et al. 2019; Lieberman, Posner, and Tsai ators to personally distribute and explain informational 2014), it is hard to know where voter–politician acleaflets to voters aged 20–38 in treated villages in the countability breaks down. For example, the limited month preceding the election. Our factorial design vareffects of providing incumbent performance informaied whether respondents were informed about (1) parliation on vote choice in the six-country Metaketa study mentary deputies’ duties and (2) their current deputy’s (Dunning et al. 2019) could reflect difficulties of disparticipation in legislative affairs and the projects and seminating information, failures to provide sufficiently transfers received by their district, either with or without relevant information, voters’ inability to internalize a comparison with their district’s previous deputy. Our information, or voters’ low willingness to electorally panel survey tracked voters’ beliefs, vote intentions and reward (punish) better (worse) performing incumbents. ultimate choices, and post-election contact requests of Furthermore, although community empowerment incumbents immediately before and after treatment and interventions have received significant attention (see again after the election. Casey 2018), little is known about whether incumbent Our findings first demonstrate that rural Senegalese performance information can influence nonelectoral citizens processed incumbent performance information accountability. A particularly important nonelectoral in sophisticated ways. Immediately after receiving the means of communicating information, preferences, or information, voters favorably updated their beliefs in line requests is citizen contact with incumbent politicians with their relatively pessimistic prior beliefs and the fact once in office (Bussell 2019; Grossman, Humphreys, and that current incumbents mostly outperformed previous Sacramone-Lutz 2014).1 deputies. These changes in beliefs indicate that voters This article dissects voters’ ability and will to use difcare principally about local outcomes (projects and ferent types of incumbent performance information to transfers), rather than legislative efforts within parliahold legislative deputies to account. By personally disment. Moreover, although information about deputy dutributing and explaining such information, we abstract ties did not affect beliefs, temporally benchmarked inforfrom dissemination challenges to focus on three links mation further improved voter appraisals of the incumbetween receiving incumbent performance information bent and increased the precision of such beliefs. We find and voter engagement in electoral and nonelectoral acsimilar—albeit somewhat smaller—effects of local percountability. First, we illuminate voter internalization of formance indicators and temporal benchmarks on beliefs information and its decay over time by examining the exaround a month after treatments were administered. tent to which voters update their beliefs in a Bayesian Immediately after receiving incumbent performance manner, both immediately after receiving information information, voters also sought electoral and nonelecand a month later. Second, we vary the content of the toral means of holding politicians to account. The avinformation disseminated to understand what informaerage treated voter—who updated more favorably about tion voters regard as relevant. Specifically, we combine current incumbents than challengers—became 3 perindicators of the current incumbent’s national and locentage points more likely to intend to vote for the incal performance with (i) information about deputy ducumbent. Heterogeneity in such electoral rewards reties and (ii) a temporal benchmark against previous inflected the degree of voter belief updating and whether cumbents’ performance that helps voters to abstract from performance information was the most important facdistrict-specific factors affecting every incumbent’s pertor determining vote choices. Treated voters also became formance. Third, we study whether persistent changes in significantly more likely to request a visit from, or an opbeliefs translate into greater electoral support for, and portunity to express their views or demands to, winning greater post-election effort to request contact with, betincumbent deputies after the election. ter performing incumbents, and how this behavior varies Although voters persistently updated their beliefs with the information’s relevance to voters. and demonstrated an initial willingness to hold politiTogether with a local civil association, we designed a cians to account, electoral accountability was ultimately field experiment in Senegal around the 2017 parliamenonly exercised among likely voters and the voters that most valued performance on local outcomes. While our 1 Such efforts could involve articulating programmatic demands or treatments did not affect self-reported vote choices on seeking pork. average, the treated respondents who cared most about ABLE AND MOSTLY WILLING 3 incumbents lobbying for local development projects or against previous incumbents more relevant than inforhad turned out at the last election did reward the inmation solely about current incumbents. This finding cumbents overseeing more local projects and transfers. contrasts with studies that report no additional effect Consistent with substantial within-village diffusion of of combining cross-sectional comparisons alongside inour information—by voters and political parties—to the cumbent performance information in other developing more experienced voters most likely to respond to it, we contexts (Arias et al. 2018a; Campello and Zucco 2016). further find greater incumbent vote shares at polling staThese contrasting findings suggest that future research, tions that received information revealing higher rates of possibly comparing spatial and temporal benchmarks local projects and transfers. within the same experiment, is required to identify Nonelectoral requests for incumbent contact afwhen different types of benchmark are most relevant. ter the election increased more uniformly, even a Furthermore, we find that Senegalese voters prioritize month after receiving treatment. The average respondent politicians bringing projects and higher value transcontinued to make more requests of incumbents, who fers to their district. Conversely, greater involvement won in each race, especially in districts that received in parliamentary activities is—if anything—punished more projects and transfers. This increase in requests reby voters (see also Adida et al. 2020; Humphreys and flects not only relatively costless requests for the winWeinstein 2012). However, we find little evidence that ning incumbents to call respondents or visit their village, information about incumbent responsibilities influences but also citizens incurring the cost of sending SMS or voter appraisals on its own, or that it systematically subvoicemail messages to winning incumbents. This effect stitutes or complements the provision of performance was also most pronounced among voters who received information. This suggests that any accountabilitybenchmarked information. enhancing effects of civic education programs (e.g., Our core finding that receiving relevant forms of inGottlieb 2016) may operate through components of cumbent performance information can induce persistent the program beyond information about incumbent voter belief updating and facilitate electoral and nonelecresponsibilities. toral accountability makes several main contributions. Finally, we show that incumbent performance inFirst, by unpacking the key links in the accountability formation also influences a costly nonelectoral means chain once voters receive relevant information, we show through which voters can seek political accountability, that accountability failures are unlikely to reflect coglikely by altering expectations that politicians will be nitive constraints—voters’ inability to process informaresponsive to their constituents. This finding completion in a Bayesian manner (Gomez and Wilson 2006) ments evidence that civic education and communication or retain updated beliefs (Zaller 1992)—or voter untechnologies can stimulate nonelectoral political enwillingness to hold politicians to account. In this regagement (Aker, Collier, and Vicente 2017; Gottlieb gard, voters’ sophisticated responses—even in a hyper2016; Grossman, Humphreys, and Sacramone-Lutz presidential context where almost half our respondents 2014). Future research is required to establish whether lacked any formal schooling—chime with Arias et al. voters’ greater efforts to contact the best performing (2018b), Humphreys and Weinstein (2012), and Kendall, legislators reflect particularistic desires for “pork” or— Nannicini, and Trebbi (2014). We advance this literaperhaps less likely—a desire for more programmatic ture by showing that voters’ initial belief updating perpolicies. sisted for the most relevant pieces of information, but only translated into electoral accountability among experienced voters and voters who care about the topics about which information was provided. This suggests that less effective information dissemination campaigns Incumbent Performance may instead reflect limited campaign reach (e.g., DunInformation and Bottom-Up ning et al. 2019), a lack of relevant or credible inforPolitical Accountability mation (e.g., Boas, Hidalgo, and Melo 2019), or competing community or political responses to information The canonical selection model of electoral accountability campaigns (e.g., Banerjee et al. 2011; Cruz, Keefer, and reflects the agency relationship between voters and politiLabonne 2019). cians. In its simplest formulation, voters use performance Second, we illuminate the types of information that indicators to identify high-quality incumbents, and then can facilitate “bottom-up” political accountability. We vote to retain them (Fearon 1999). This framework predemonstrate that voters find temporal benchmarking dicts that information that favorably updates voter beliefs 4 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL about incumbent quality, relative to challenger quality, Kendall, Nannicini, and Trebbi 2014).2 Following prior increases support for the incumbent, especially among research, we thus hypothesize the following: voters for whom politician quality outweighs other facHypothesis 1: Incumbent performance informators entering their voting calculus. After elections, voters tion will increase (decrease) incummay similarly become more willing to engage in costly bent support and requests to the exefforts to contact incumbents who they expect will be retent that such information causes sponsive and effective. Appendix section A (p. SI2) forvoters to favorably (unfavorably) upmally summarizes these logics. date about incumbent quality. By providing information about incumbent performance—that incumbents could not have anticipated would be publicized when deciding how to act (cf. Grossman and Michelitch 2018)—just before Relevance of Novel Information elections in which incumbents sought reelection, our Even if novel incumbent performance information is design sidesteps strategic policy and candidacy choices. credible and internalized, voters must also perceive it Furthermore, by directly providing information to as relevant—that is, pertaining to incumbent quality— voters, we also abstract from potential failures in the to influence political accountability. Beyond considering process through which information is disseminated different indicators of incumbent performance, we study and consumed. two further aspects of relevance that could complement Whether receiving credible incumbent performance the provision of incumbent performance indicators: ininformation causes voters to hold incumbents seeking formation about incumbent duties and temporal perforreelection to account electorally and nonelectorally thus mance benchmarks. rests upon (i) voters’ cognitive capacity to process and inInformation about incumbent duties could help votternalize novel information, (ii) information’s relevance ers to infer incumbent quality from performance sigto voters, and (iii) voters’ willingness to act on their nals in at least two ways. First, voters may only recognize updated beliefs about the incumbent’s quality. The folperformance information as relevant upon learning that lowing subsections theorize key conditions under which politicians possess the capacity to feasibly influence such each element of this anatomy of political accountability performance indicators (Gottlieb 2016). Second, specific may hold. information about an incumbent’s duties may help voters to assign responsibility across multiple layers of government (Powell and Whitten 1993). Although incumbent duties are often implicit when performance information Internalization of Novel Information is provided, or outlined alongside performance information (Gottlieb 2016), we explicitly separate between proPolitical accountability relies on voters comprehending viding information about duties and performance to test incumbent performance information and somewhat whether: durably updating their beliefs about incumbent quality. Because the information that voters read, hear, and Hypothesis 2: Receiving information about incumobserve is often complex, they may only superficially bent duties, either alongside or understand it and therefore not meaningfully update without corresponding incumbent their beliefs (Gomez and Wilson 2006). Moreover, beperformance information, increases cause voters may reject novel information challenging (decreases) incumbent support and their preexisting beliefs or fail to internalize it over requests among better (worse) pertime (Zaller 1992), belief updating about incumbent forming incumbents. quality may be too transient to influence voting behavior Benchmarked incumbent performance information (Humphreys and Weinstein 2012). could increase the accuracy of voters’ posterior beCredible and comprehensible information is most liefs through two main channels. First, receiving multilikely to alter behavior when it differs from voters’ prior ple performance signals helps voters to filter out combeliefs about their incumbent’s quality. Bayesian votmon shocks influencing the performance of all agents ers update the position and precision of their posterior in a given period or location (Aytaç 2018; Meyer and beliefs most when the information provided is precise, the information deviates from their prior beliefs, and 2 Voters may also update about challengers from incumbent pertheir prior beliefs are imprecise (e.g., Arias et al. 2018b; formance, if their types are correlated. ABLE AND MOSTLY WILLING 5 Vickers 1997). Second, benchmarks enable voters to upbeliefs about the incumbent to their available actions date about the absolute quality level of other politicians (Gomez and Wilson 2006). Second, in the case of voting, who resemble challengers, especially where benchmarks voters must attach significant weight to beliefs about inare from a different political party from the incumbent. cumbent quality in their voting calculus. Third, the proBoth channels facilitate more accurate and precise beliefs cess of providing information could set in motion other about absolute and relative incumbent, and possibly chalforces that override the influence of voter beliefs on vote lenger, candidate quality—the key drivers of political acchoice, including voter coordination around particular countability in our conceptual framework. candidates (Arias et al. 2019) or candidate campaign reThe relative utility of cross-sectional and intertemsponses to information revelations (Banerjee et al. 2011; poral benchmarks in a particular context depends on Cruz, Keefer, and Labonne 2019). If such equilibrium rethe accuracy and uncertainty of voters’ prior beliefs sponses affect vote choices, changes in voter beliefs may about timeand unit-specific shocks and their magnot ultimately translate into changes in behavior. In sum, nitude. Thus far, extant studies focusing on spatial we anticipate the following: benchmarks—that help filter out period-specific shocks that equally affect all incumbents holding office in Hypoothesis 4: The magnitude of information’s efdifferent districts contemporaneously (e.g., changing nafects on incumbent support and eftional budgets)—generally find limited evidence that forts to contact incumbents after such benchmarks influence beliefs beyond providing the election is greatest among votinformation about only the incumbent’s performance ers who are civically educated, value (Arias et al. 2018a; Aytaç 2018; Campello and Zucco incumbent performance indicators, 2016).3 We focus on temporal benchmarks, which have and are less susceptible to forces yet to be tested experimentally. Temporal benchmarks counteracting the information. help filter out the effects of time-invariant features of a district that affect all incumbents serving that district, such as geographical constraints or demographic political importance. We hypothesize the following: Parliamentary Accountability in Hypothesis 3: Relative to only providing incumbent Senegal performance information, temporal benchmarks increase (decrease) inSenegal is one of Africa’s oldest and strongest democcumbent support and requests when racies. It has generally experienced robust multiparty (i) incumbent performance is above political competition—including peaceful transitions in (below) voters’ prior belief and/or 2000 and 2012, following fair democratic elections— (ii) the previous incumbent’s persince 1981, and is known for its vibrant civil society and formance was below (above) voters’ freedom of press and expression. However, voters are ofprior belief. ten poorly informed about legislative politics, and political accountability remains low. Appendix section B (pp. SI2–SI7) demonstrates formally that case (i) reflects benchmarked information facilitating more precise inferences, whereas case (ii) reThe Assemblée Nationale’s Role flects benchmarked information updating beliefs about district-specific characteristics influencing performance. The Assemblée Nationale (Parliament) plays a limited role in democratic representation in Senegal’s hyperpresidential context (Beck 2012; Thomas and Sissokho 2005). Acting on Internalized Beliefs Deputies are elected for 5-year terms by a mixed system, where competing coalitions form a national list and Even if information meaningfully updates voters’ beliefs, submit lists for each of Senegal’s 45 departments (which bottom-up political accountability requires that voters serve as parliamentary districts). In each department, the ultimately act on such beliefs. This likely requires that coalition winning most votes receives all seats allotted several conditions hold. First, voters must connect their to the department. In 2017, 105 deputies were elected 3 from 12 single and 33 multimember departments and 60 The clearest electoral evidence comes from cross-national macroeconomic comparisons in advanced democracies (Aytaç seats were allocated in proportion to a coalition’s national 2018). vote share. In the 2012 legislative elections, President 6 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL FIGURE 1 The Most Important Factor Driving Individuals’ Vote Choices Village or community of deputy 14 Ethnicity or religion of deputy 3 Education or profession of deputy 7 Party of deputy 4 Political experience of deputy 7 Ability of deputy to amend 8 national laws and budgets Ability of deputy to lobby for 46 projects/transfers to department Campaign promises of deputy 7 Electoral gifts distributed by deputy 2 0 10 20 30 40 50 Percentage of voters citing factor as the most important in determining vote choice Note: Data are drawn from the baseline of the authors’ panel survey of rural Senegaleese voters aged 20–38. Macky Sall’s coalition—Benno Bokk Yakaar (BBY)— difference between good and bad deputies as their “cawon 87 of 90 majoritarian departmental seats and appacity to lobby successfully.” proximately half the proportionally allocated seats. Our study examines deputies elected from departmental majoritarian lists because of their stronger electoral ties to constituents. Voter Engagement with Parliamentary The primary constitutional role of elected deputies Elections and Deputies is amending and voting on laws drafted by government Voter turnout in Senegal reached 54% in the 2017 ministries. However, few laws are rejected by the Assemparliamentary elections, slightly below the sub-Saharan blée Nationale, and its role in checking executive power African mean. Nevertheless, Senegal’s 2016 Afrobaromis often questioned by civil society. Deputies can also inieter round indicates that 87% of respondents viewed tiate laws themselves, although this is rare in practice Senegal as a democracy, and 64% reported being satisfied (Thomas and Sissokho 2005). with the functioning of Senegalese democracy. Although Nonetheless, deputies can—and do—affect legisladirect citizen interaction with deputies is rare (only 9% of tive decisions through their parliamentary duties. First, respondents in our sample had contacted a deputy within deputies can serve on the Assemblée’s 11 parliamentary the last year) and voters are pessimistic about whether committees, through which they can make recommendadeputies listen to voters and respond to requests, intertions and amendments to ministerial bills before plenary actions with party officials and brokers who report to debates. Second, deputies can submit questions to the deputies are relatively common. government to defend and publicize their constituents’ Although election outcomes often reflect nationwide interests, which relevant ministers answer in open sesswings in coalition support, our baseline survey data insions. Third, although deputies do not receive specific dicate that many voters also seek to elect deputies who funds for local development projects, they are widely bring local development projects to their department. believed—and themselves claim—to influence the alloFigure 1 indicates that 46% of voters claim that a deputy’s cation of local projects and government transfers by lobpotential to lobby for projects and transfers benefiting bying ministers. Indeed, one deputy described the biggest their department is the most important factor driving ABLE AND MOSTLY WILLING 7 their vote choice. Fewer voters regard national-level polis less educated and developed than the national average. icy engagement as important. Moreover, when asked to Within each village, we aimed to survey nine registered choose between hypothetical deputies seeking to improve voters aged 20–38 who had lived in the village prior to voters’ welfare, 71% of respondents favored locally orithe age of primary school enrollment. Appendix section ented politicians to nationally oriented ones. D (pp. SI7–SI9) provides further sampling information. However, actually holding deputies to account has proved challenging for several reasons. First, voters often lack the information needed to identify the bestInformation Treatments performing deputies. Only 35% of voters in our sample could name at least one of their parliamentary repreOur treatments entailed distributing and explaining sentatives, and only 61% could correctly identify the scorecards detailing combinations of legislator duties, incumbent party in their department. Moreover, voters’ current incumbent performance, and previous incumprior beliefs are uncorrelated with the incumbent perforbent legislator performance in the month preceding the mance metrics that our treatment provides (see Figure 5 election. Regarding legislator duties, we highlighted that below). The paucity of reliable information partly relegislators can (1) serve on the 11 parliamentary comflects the limited penetration of mass media and election mittees, (2) participate in parliamentary debates, and campaigns in rural communities. Second, attempts to (3) lobby government ministers to allocate projects and hold deputies to account often compete against clientransfers to their departments. telistic incentives and coordinated group voting pushing Regarding incumbent legislator performance, we vote choices in different directions. Political parties can provided five nationally and locally oriented measures of heavily influence rural vote choice via village chiefs and performance in office over the 5-year electoral cycle that other local brokers as intermediaries (Koter 2013). relate to deputies’ primary duties: (1) committee memberships, (2) positions of leadership within parliament, (3) the number of parliamentary debates participated in, (4) the number of local projects budgeted for their Research Design department in parliamentary documents, and (5) the number and (inflation-adjusted) per capita per year We designed an information dissemination campaign in value of ministry transfers received by the department, partnership with a local civil association—LEGS-Africa, decomposed by transfer category.4 All deputyand a transparency-oriented organization in Dakar—to perdepartment-specific data were obtained from the Assonally deliver and explain incumbent performance insemblée Nationale or ministries, and their accuracy, formation to voters prior to Senegal’s 2017 parliamentary relevance, and impartiality were validated by the head elections. We randomized core components of the inforof legislative services at the Assemblée Nationale, the mation’s content across villages, and used a panel study librarians and archivists at the Assemblée Nationale, and to track voter beliefs and actions before the information several active and former deputies. treatments were delivered, immediately after their delivBased on the performance metrics just described, we ery, and a month later (after the election). This design used a 2 × 3 factorial design to randomly assign villages thus traces the key links between voters’ receipt of into one of the six experimental conditions in Table 1. formation and whether and how voters hold incumbent Treatment conditions vary along two dimensions of condeputies to account. tent, and include a pure control group. First, the “duties” dimension informed voters of the three main functions (enumerated above) that deputies can perform. Second, Sample Selection the “performance” dimension varied whether voters We conducted our study in the five departments shown received “incumbent” information relating to the/an in Figure 2: Fatick, Foundiougne, Kanel, Oussouye, and incumbent representative’s performance on the five Ranérou Ferlo. In each department, the current deputies measures described above or “benchmark” information were from the BBY coalition—the president’s ruling additionally providing the same information pertaincoalition—and the previous deputies were from the Sopi ing to the performance of the/a department’s previous coalition—the previous president’s ruling coalition and 4 BBY’s rival. Within these departments, we selected 450 Annual transfer data were available from 2010 to 2016, and normalized by 2013 population size. Transfers affecting multiple derural villages containing 200–4,000 people for our sampartments were distributed in proportion to each department’s ple. Appendix Table D1 (p. SI10) shows that this sample 2013 population. 8 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL FIGURE 2 Sampled Senegalese Departments Note: The five departments included in this study are shaded in gray. incumbent representative. In the multimember departincumbent–previous incumbent pairing. Points above ments where two incumbent deputies sought reelection the 45o line represent cases where the current incum(Kanel) or more than one deputy held office during bent outperformed the previous incumbent. The curthe previous legislative session (Fatick, Foundiougne, rent incumbents often outperformed preceding incumand Kanel), we maximized treatment homogeneity by bents, especially with respect to debates, projects, and randomly selecting one current deputy seeking reelectransfers. We accordingly anticipated that our perfortion and one benchmark deputy per randomization mance information would increase voters’ favorability block (defined below) to be reported on in every leaflet toward current incumbents, on average, across departdelivered within the block. ments. However, we also examine heterogeneity by perFigure 3 reports the distribution of the performance formance level. metrics provided, where each point represents a current TABLE 1 Treatment Conditions Type of performance information provided: Whether duties information is provided: None Incumbent Benchmark None 75 villages [pure control] 75 villages 75 villages Duties 75 villages 75 villages 75 villages ABLE AND MOSTLY WILLING 9 FIGURE 3 Distribution of Treatment Information across Departments (45o Line in Gray) Committee memberships Leaderships positions Debates participated in Kanel Kanel Oussouye Kanel Oussouye 10 3 1 Fatick Fatick .8 Current incumbent Current incumbent Current incumbent 8 Oussouye 2 Fatick Fatick Foundiougne .6 6 .4 4 1 Ranérou Ferlo Kanel Kanel .2 2 Ranérou Ferlo Kanel Foundiougne Kanel Kanel 0 0 0 Ranérou Ferlo Fatick Foundiougne Foundiougne 0 1 2 3 0 .2 .4 .6 .8 1 0 2 4 6 8 10 Previous incumbent Previous incumbent Previous incumbent Local projects Number of transfers Transfer value per capita-year 10000 15000 15 6 Fatick Fatick Fatick Current incumbent Current incumbent Current incumbent 10 4 5000 5 2 Kanel Oussouye Foundiougne Kanel Kanel Oussouye Ranérou Ferlo Oussouye Ranérou Ferlo Ranérou Ferlo 0 0 0 Foundiougne Foundiougne 0 5 10 15 0 1 2 3 0 2000 4000 6000 Previous incumbent Previous incumbent Previous incumbent Note: Cases within departments where previous incumbents performed identically are not duplicated. Each information treatment was distributed to votenumerators—mostly university graduates—themselves ers through leaflets like the one in Figure 4. The leaflets understood and could clearly explain the leaflets’ content were professionally designed in partnership with LEGSin both languages. On average, treatment delivery took Africa. Each leaflet variant showed the LEGS-Africa logo around 5 minutes. Suggesting that the treatment was realongside a statement that the organization is nonpargarded as credible, 82% of treated respondents reported tisan at the top, whereas data sources and (redacted) that the leaflet came from an nongovernmental organicontact information were provided at the bottom. The zation. example in Figure 4 depicts the duties and benchmark Our intervention is thus heavier handed than most treatment variant—the maximum amount of informaprior information dissemination campaigns. Previous tion that was provided.5 The three paragraphs below campaigns have posted fliers, sent SMS messages, crethe LEGS-Africa logo were provided to all participants ated newspaper articles, or arranged dissemination meetreceiving a “duties” variant. The current incumbent ings or video viewings (e.g., Banerjee et al. 2011; Chong performance information on the left of the leaflet was et al. 2015; Dunning et al. 2019). Unlike studies proprovided to participants receiving the “incumbent” viding access to information, we ensure that voters revariant, whereas the performance information on the ceived and understood the information to focus on beleft and right was provided to participants receiving the lief updating and voter behavior absent dissemination “benchmark” variant. The leaflet was piloted to ensure constraints. Several prior studies conduct similarly intencomprehensibility. sive interventions (e.g., Boas, Hidalgo, and Melo 2019; The leaflet was delivered and explained in person, on Humphreys and Weinstein 2012). behalf of LEGS-Africa, to respondents during our baseline survey. Enumerators gave each voter several minutes to read the leaflet in French and then spent several minInformation Provision Randomization utes explaining the meaning of each component in the respondent’s local language. Our training ensured that Leaflet treatment conditions were block-randomized at the village level to mitigate contamination arising from 5 Appendix Figures E1– E4 (pp. SI11–SI13) show our other leaflets. within-village spillovers. Specifically, we constructed 75 10 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL FIGURE 4 Example of “Duties + Benchmark” Treatment in Oussouye blocks, and assigned each experimental condition to one tween August 4 and August 26. We also mapped each vilof six similar villages from within the same departlage to its associated polling station. ment.6 In multimember departments, we used complete randomization to assign an incumbent–previous incumMeasurement of Primary Outcomes. Our primary bent pair to each block.7 classes of outcomes focus on voter beliefs and whether voters ultimately engaged in electoral accountability and requests to contact winning incumbents after the elecData Collection tion. First, we measured voters’ beliefs about how well inWe designed a two-wave panel survey and collected cumbents have done overall since they were elected in polling station–level electoral returns. The baseline sur2012, how they compare with the previous incumbent, vey was conducted in person between July 4 and July and how the current incumbent seeking reelection would 29, and our treatments were administered after enudo if reelected on 5-point scales from “very bad” (1) to merators collected respondents’ characteristics, baseline “very good” (5).8 For each variable, we also elicited the beliefs, previous behaviors, and intentions. The shorter strength of voters’ assessment on a 10-point scale ranging post-election survey was conducted by telephone befrom “not at all certain” (1) to “completely certain” (10).9 These beliefs, and their associated certainty, were elicited 6 After stratifying by department, village similarity was determined at baseline before and after information treatments were by Mahalanobis distance across 24 pretreatment covariates. 8 7 “Don’t know” responses are coded at the midlevel of the scale. Appendix Table E1 (p. SI11) reports the distribution of configu9 rations by deputy. “Don’t know” responses are coded as the lowest level of certainty. ABLE AND MOSTLY WILLING 11 delivered for both treated and control respondents. The cators of requests to contact incumbents.11 By standardfirst two questions were repeated at endline. izing all indexes with respect to the control group, effect Second, we elicited voting behavior: vote intention magnitudes represent standard deviation changes in conbefore and after treatment in the baseline survey, and trol group outcomes. self-reported turnout and vote choice at endline. We use indicators for respondents stating that they would or Compliance. We encountered two minor forms of data did vote for the incumbent. We address self-reporting missingness. First, we could not access 7 villages. Howconcerns at endline with a robustness check that only ever, because villages were surveyed identically by enucounts votes as valid where the respondent correctly remerators and not informed of treatment status in adcalled features of the ballot and its party-specific color. vance, the opportunity to conduct surveys was unafWe also elicited certainty about intended vote choice fected by treatment assignment. Second, 4% of responon a 10-point scale. Furthermore, we use the electoral dents attrited between baseline and endline surveys, but returns to calculate incumbent party vote share at the not differentially so across treatment conditions. Appolling station corresponding to each village in our sampendix section F (pp. SI10–SI13) provides additional deple. Although fewer than 2% of voters were treated tails, and reports balance tests supporting the randomwithin polling stations, information could spread within ization’s integrity. our tight-knit set of rural villages to affect this behavioral outcome. Third, we measure costly efforts to contact winning Estimation incumbents after the election. At baseline, we offered respondents the opportunity to request a visit from, or Following our preanalysis plan, the following fully satsign up to be contacted to express their views to, any urated ordinary least squares (OLS) regression specificaparty or candidate if they were subsequently elected.10 tion estimates the average treatment effect of different inThese behavioral measures capture nonelectoral means formational components of the leaflet:12 of seeking accountability, akin to Aker, Collier, and ViYiv = γb + δe + αYivbaseline + β1 duties v + β2 incumbent v cente (2017), Bussell (2019), and Grossman, Humphreys, + β3 benchmark v + β4 (incumbent v × duties v ) and Sacramone-Lutz (2014). The post-election endline survey again offered respondents the opportunity to re+ β5 (benchmark v × duties v ) + εiv , (1) quest a visit from and sign up to be contacted by the where Yiv is an outcome for respondent i in village v, γb winning candidate—the incumbent in each department. are randomization block fixed effects, and δe are enuAt endline, we also created a hotline where respondents merator fixed effects. Wherever possible, the outcome’s could send text messages (costing around US$0.04, or pretreatment baseline counterpart Yivbaseline is included to 5% of rural per capita daily expenses) or leave voiceincrease estimation efficiency. For polling station–level mails (US$0.18, or 21% of rural per capita daily exoutcomes, we replace the iv subscript with a p subscript. penses) requesting to be contacted by the winning canTo recover the village-level average treatment effect, all didate. We measure this by linking telephone numbers survey-based regressions are weighted by the inverse of to the respondent. These types of opportunity to engage the number of respondents in the corresponding basedirectly with elected politicians are rare in rural Seneline or endline survey. Standards errors are clustered by gal, as in other African contexts; see Appendix section C village. One-sided t tests are applied to prespecified di(p. SI7). rectional hypotheses. Two-sided t tests are applied to hyGiven the large number of outcomes—which enpotheses that were not prespecified or were prespecified gender concerns about multiple comparisons and noise without a hypothesized direction, and—denoted by disin specific variables—we combine related individualtinct symbols—to estimates in the opposite direction to level outcomes using indexes. Separately within baseline our prespecified hypothesis. and endline panel waves, we created inverse-covariance We test additional hypotheses underpinning the acweighted (ICW) indexes to summarize two groups of countability process by further estimating heterogeneity items: (i) incumbent evaluation, which aggregates all atin treatment effects by the content of the information titudinal and voting outcomes, and (ii) behavioral indi11 The ICW approach accounts for correlation among items; see Appendix section G (pp. SI13–SS14). 10 12 The voter’s name and village were shared with the party by LEGSAppendix section H (pp. SI14–SI15) explains minor deviations Africa. from the preanalysis plan. 12 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL TABLE 2 Leaflet Comprehension (Baseline Survey) Respondent correctly states… …number of …deputies …number of …number of parliamentary lack incumbent’s previous committees department local incumbent’s fund projects debates (1) (2) (3) (4) Duties 0.663∗∗ 0.459∗∗ (0.022) (0.022) Incumbent 0.729∗∗ (0.024) Benchmark 0.717∗∗ 0.461∗∗ (0.025) (0.027) Two-sided test: Incumbent = Benchmark (p value) 0.61 Observations 3,999 3,999 3,999 3,999 Outcome range {0,1} {0,1} {0,1} {0,1} Control outcome mean 0.05 0.14 0.08 0.07 Control outcome std. dev. 0.22 0.35 0.26 0.26 Note: Each regression includes randomization block and enumerator fixed effects. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from prespecified one-sided t tests. provided, voter’s prior beliefs, or the importance of conColumns (1) and (2) indicate that receiving any duties intent for a voter’s decision making. Although such predeformation increased the proportion of respondents cortermined moderators are not randomly assigned, these rectly identifying the number of parliamentary committests further evaluate consistency with the political actees from 5% to 71% and that deputies lack individual countability logic. funds for department projects from 14% to 60%. Column (3) demonstrates that incumbent performance information increased the proportion of respondents correctly identifying the number of local projects received Immediate Effects of Information under the current incumbent from 8% to around 80%. Provision Finally, column (4) shows that the benchmark leaflet increased correct answers regarding the number of debates We start by examining immediate responses to receiving that the previous incumbent participated in from 7% to information at the end of the baseline survey. This en53%. Voters’ inability to generally comprehend the inforables us to assess voter-level links in the accountability mation provided therefore does not seem to represent a logic upon receiving information and before any further bottleneck in the political accountability process. interactions with other voters or political actors occur. Voters Update Their Beliefs in a Bayesian Voters Comprehend the Leaflet’s Manner Information Given that information may not be perceived as crediWe first verify that voters comprehended the treatment ble or relevant, voter comprehension does not necessarinformation. All respondents were asked four factual ily imply that our leaflets’ information would alter voter multiple-choice questions pertaining to different compobeliefs. Furthermore, regurgitating information may not nents of the leaflet’s information. imply sophisticated internalization. Restricting attention to the corresponding informaColumns (1)–(3) in Table 3 show that, on avertional treatments, the results in Table 2 demonstrate age, voters favorably updated their posterior beliefs imthat most respondents comprehended the information. mediately after receiving the information—in line with TABLE 3 Average Effects of Information Treatments (Baseline Survey) Incumbent evaluation outcomes Incumbent contact request outcomes Incumbent Relative Prospective Incumbent Incumbent Request Request Incumbent overall performance incumbent vote evaluation incumbent incumbent contact request performance (vs. previous) performance intention index (ICW) visit conversation index (ICW) ABLE AND MOSTLY WILLING (1) (2) (3) (4) (5) (6) (7) (8) Panel A: All information treatment conditions Duties 0.062 −0.043 0.066 0.003 0.003 −0.023 −0.027 −0.056 (0.065) (0.056) (0.053) (0.013) (0.042) (0.020) (0.018) (0.041) Incumbent 0.362∗∗ 0.221∗∗ 0.239∗∗ 0.030∗∗ 0.227∗∗ 0.008 0.005 0.016 (0.063) (0.051) (0.054) (0.013) (0.040) (0.019) (0.018) (0.038) Incumbent × Duties −0.014 0.127∗ 0.044 0.002 0.052 0.061∗∗ 0.046∗ 0.120∗∗ (0.088) (0.077) (0.076) (0.020) (0.060) (0.027) (0.027) (0.058) Benchmark 0.457∗∗ 0.353∗∗ 0.376∗∗ 0.037∗∗ 0.331∗∗ 0.002 0.001 0.004 (0.066) (0.057) (0.057) (0.015) (0.045) (0.020) (0.019) (0.042) Benchmark × Duties −0.051 0.041 −0.098 −0.004 −0.036 0.028 0.038† 0.074 (0.091) (0.084) (0.077) (0.020) (0.064) (0.028) (0.027) (0.059) Panel B: Pooling duties treatment conditions Incumbent 0.356∗∗ 0.285∗∗ 0.262∗∗ 0.031∗∗ 0.253∗∗ 0.039∗∗ 0.029∗∗ 0.076∗∗ (0.044) (0.038) (0.037) (0.010) (0.029) (0.014) (0.013) (0.029) Benchmark 0.432∗∗ 0.375∗∗ 0.328∗∗ 0.035∗∗ 0.313∗∗ 0.017 0.021† 0.042† (0.045) (0.042) (0.039) (0.010) (0.032) (0.014) (0.013) (0.030) Benchmark − Incumbent 0.076∗ 0.089∗ 0.066∗ 0.005 0.060∗ −0.022 −0.008 −0.034 (0.042) (0.039) (0.039) (0.011) (0.032) (0.014) (0.014) (0.029) Observations 3,942 3,932 3,928 3,999 3,891 3,999 3,998 3,998 Outcome range {1,…,5} {1,…,5} {1,…,5} {0,1} [−2.3,1.9] {0,1} {0,1} [−1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.00 0.70 0.70 −0.00 Control outcome std. dev. 1.07 0.90 1.09 0.49 1.00 0.46 0.46 1.00 Note: Each regression includes randomization block and enumerator fixed effects and a lagged dependent variable (or pretreatment incumbent vote intention as a proxy). Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from prespecified one-sided t tests. 13 14 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL FIGURE 5 Correlation between Pretreatment Incumbent Overall Performance Assessments and Incumbent Deputy Performance Village mean evaluation of incumbent overall performance Very good Good Neither Bad Very bad -3 -2 -1 0 1 Overall performance (ICW) Note: Villages are jittered around the overall performance (ICW) levels for each of the eight deputies. Hypothesis 1. Panel A indicates that treated voters who ing most about deputies bringing projects back to their received incumbent-only, and especially benchmarked, department and already believed that incumbents were performance information experienced around a third of capable of influencing this process. Henceforth, we focus a standard deviation increase in favorability toward their primarily on the comparison between benchmarked and incumbent deputy across each assessment of their suitnonbenchmarked incumbent performance information ability for office.13 Appendix Table I1 (p. SI16) further by pooling the control and duties conditions. shows that treatment increased voters’ certainty in their The pooled specification in panel B of Table 3 beliefs about the incumbent’s current and future perfordemonstrates significant differential effects of withinmance by nearly 0.2 standard deviations. Such updatdepartment temporal benchmarks. Consistent with ing is consistent with voters being poorly informed— voters using the previous incumbents’ generally worse Figure 5 reports a somewhat negative correlation beperformance to filter out department-specific effects tween reported performance and pretreatment beliefs— (Hypothesis 3), the tests at the foot of columns (1) and and receiving credible performance indicators that gen(3) show that benchmarked information increased voterally exceeded prior expectations. ers’ overall and prospective posterior assessments more Conversely, information about deputies’ duties did than receiving incumbent-only information. Appendix not systematically affect voter evaluations—whether on Table I1 (p. SI16) further shows that the benchmark its own or in conjunction with performance indicators. also differentially increased belief precision. Suggesting This lack of support for Hypothesis 2 indicates that that this greater precision increased the weight attached voters do not need additional information about into the signal of incumbent performance in voters’ cumbent duties to internalize incumbent performance posterior beliefs, column (2) also reports that the benchinformation. The lack of systematic effects of informamarked information had a larger effect on the relative tion about duties is a consistent pattern throughout this comparison between current and previous incumbents study. This could reflect a plurality of respondents carthan incumbent-only information.14 13 14 Around 25% of treated voters favorably updated their pretreatAppendix Table I2 (p. SI17) shows that benchmarks did not alment beliefs; very few updated unfavorably. ter average prospective challenger performance evaluations. PreviABLE AND MOSTLY WILLING 15 Finally, the heterogeneous effect results in Table 4 (p. SI18) similarly shows that voters with the least faindicate that voter processing of information is consisvorable and least precise prior beliefs were significantly tent with Bayesian updating (Hypothesis 1). Panel A first more likely to intend to vote for the incumbent. Second, shows that voters updated significantly more favorably and supporting Hypothesis 4, the effects of performance about their incumbent when the leaflet indicated higher information were greatest among the voters for whom performance on a standardized ICW scale combining the information was most relevant. Panel C demonstrates our six reported performance indicators. Panel B further that the 54% of respondents who ranked the incumbent’s shows that voters’ increased favorability almost entirely ability to amend laws and budgets or lobby for projects reflects the “local” projects and transfers components in the department as the most important determinant of of the performance index, suggesting that—consistent their vote choice (before treatment dissemination) were with their own stated preference—voters are mostly consignificantly more likely to intend to vote for the incumcerned with the resources that deputies bring to their debent after treatment. Similar results hold in column (4) of partments. As in Adida et al. (2020), engaging in “napanel D among the 71% of voters who expressed a prefertional” legislative efforts, if anything, appears to conence for a locally oriented, as opposed to nationally oristitute a negative signal. Moreover, Appendix Table I3 ented, politician in our pretreatment vignette. (p. SI18) shows that the voters with the least favorable and least precise prior beliefs updated most favorably about the incumbent. Together, these findings imPerformance Information Increases ply sophisticated voter learning about incumbent qualIncumbent Contact Requests ity, suggesting that neither cognitive capacity nor resistance to information impeded political accountability in Our behavioral outcomes capturing requests to contact rural Senegal. winning incumbents after the election demonstrate that revelations of better-than-expected performance also encouraged nonelectoral accountability. Columns (6) and Performance Information Alters Vote (7) in panel B of Table 3 show that, on average, treated Intentions respondents became significantly more willing to request a visit from, or a conversation with, incumbents if reWe next examine whether these changes in beliefs transelected (Hypothesis 1). The index outcome estimates in late into vote intentions. Column (4) of Table 3 indicates column (8) imply around a 0.05 standard deviation inthat receiving incumbent performance information increase relative to the control group. For such behaviors, creased intentions to vote for the incumbent deputy by benchmarked information did not differentially increase 3 percentage points (Hypothesis 1).15 This effect is again requests, although this may have been limited by high larger for benchmarked than incumbent-only informarates of take-up among incumbent supporters. Broadly tion (Hypothesis 3), although not significantly so. We in line with vote intentions, the heterogeneous treatalso again find no evidence to suggest that incumbent dument effects in Table 4 and Appendix Table I3 report ties information directly affected vote intention or modlarger treatment effects on incumbent requests where inerated the effects of performance information (Hypothcumbent performance levels were greater and especially esis 2). among the voters who cared most about performance Effects on vote intentions also vary in line with (Hypotheses 1 and 4). Although not always statistically the electoral accountability logic. First, changes in insignificant, these estimates suggest that voters became cumbent vote intentions are consistent with changes more likely to seek contact with incumbent politicians afin voters’ beliefs (Hypothesis 1). Column (4) of panter learning that the incumbent may be more responsive els A and B of Table 4 shows that treatments revealthan expected. ing better performance—in terms of more local projects Post-election incumbent contact requests represent and transfers—were substantially more likely to increase the one area where information about deputies’ duties intent to vote for the incumbent. Appendix Table I3 might complement performance indicators. Columns (6)–(8) of panel A in Table 3 show that learning that the ous incumbent performance falling below prior expectations thus does not appear to drive the differential effects of benchmarked incumbent is generally performing better than expected information. primarily translated into nonelectoral accountability ef15 Voters’ already-high vote choice certainty did not signififorts when voters were aware of what incumbents can do cantly increase. (Hypothesis 2). However, any suggestion that voters must TABLE 4A Effects of Information Treatments (Baseline Survey), by Incumbent Performance Level Incumbent evaluation outcomes Incumbent contact request outcomes 16 Incumbent Relative Prospective Incumbent Incumbent Request Request Incumbent overall performance incumbent vote evaluation incumbent incumbent contact request performance (vs. previous) performance intention index (ICW) visit conversation index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) Panel A: Heterogeneity by (standardized) reported performance level Incumbent 0.355∗∗ 0.285∗∗ 0.261∗∗ 0.031∗∗ 0.253∗∗ 0.039∗∗ 0.029∗∗ 0.076∗∗ (0.042) (0.036) (0.036) (0.010) (0.028) (0.014) (0.013) (0.029) ∗∗ ∗∗ ∗∗ Incumbent × Overall performance (ICW) 0.233∗∗ 0.179∗∗ 0.177 0.025 0.115 0.012 0.008 0.023 (0.050) (0.044) (0.044) (0.009) (0.029) (0.017) (0.017) (0.037) Benchmark 0.430∗∗ 0.373∗∗ 0.327∗∗ 0.035∗∗ 0.313∗∗ 0.016 0.021† 0.041† (0.044) (0.041) (0.038) (0.010) (0.031) (0.014) (0.013) (0.030) Benchmark × Overall performance (ICW) 0.230∗∗ 0.140∗∗ 0.156∗∗ 0.037∗∗ 0.137∗∗ 0.014 0.003 0.019 (0.051) (0.047) (0.050) (0.010) (0.033) (0.019) (0.018) (0.041) Overall performance (ICW) range [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] Panel B: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.359∗∗ 0.287∗∗ 0.264∗∗ 0.030∗∗ 0.255∗∗ 0.039∗∗ 0.028∗∗ 0.076∗∗ (0.039) (0.034) (0.035) (0.010) (0.026) (0.014) (0.013) (0.029) Incumbent × National performance (ICW) 0.031 0.068∗∗ 0.012 0.001 0.013 0.000 0.016 0.017 (0.043) (0.034) (0.033) (0.009) (0.025) (0.015) (0.014) (0.031) Incumbent × Local performance (ICW) 0.317∗∗ 0.239∗∗ 0.250∗∗ 0.025∗∗ 0.201∗∗ 0.009 −0.004 0.006 (0.045) (0.038) (0.038) (0.011) (0.030) (0.016) (0.015) (0.033) Benchmark 0.431∗∗ 0.373∗∗ 0.327∗∗ 0.035∗∗ 0.313∗∗ 0.016 0.021† 0.041† (0.042) (0.038) (0.037) (0.010) (0.029) (0.014) (0.013) (0.030) ‡ § Benchmark × National performance (ICW) −0.032 −0.127§§ −0.071 0.006 −0.063 −0.005 0.003 −0.002 (0.048) (0.044) (0.038) (0.009) (0.030) (0.015) (0.014) (0.031) Benchmark × Local performance (ICW) 0.236∗∗ 0.271∗∗ 0.221∗∗ 0.035∗∗ 0.220∗∗ 0.005 0.013 0.020 (0.047) (0.042) (0.039) (0.011) (0.032) (0.016) (0.015) (0.033) National performance (ICW) range [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] Local performance (ICW) range [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] Observations 3,942 3,932 3,928 3,999 3,891 3,999 3,998 3,998 Outcome range {1,...,5} {1,...,5} {1,...,5} {0,1} [−2.3,1.9] {0,1} {0,1} [−1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.00 0.70 0.70 −0.00 Control outcome std. dev. 1.07 0.90 1.09 0.49 1.00 0.46 0.46 1.00 Notes: Each regression includes randomization block and enumerator fixed effects a lagged dependent variable (or pre-treatment incumbent vote intention as a proxy). Lower-order ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL (standardized) interaction terms are included but not shown. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from pre-specified one-sided t tests; ‡ p < 0.1, § p < 0.05, §§ p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. TABLE 4B Effects of Information Treatments (Baseline Survey), by Voter Criteria for Preferred Politicians Incumbent evaluation outcomes Incumbent contact request outcomes Incumbent Relative Prospective Incumbent Incumbent Request Request Incumbent overall performance incumbent vote evaluation incumbent incumbent contact request performance (vs. previous) performance intention index (ICW) visit conversation index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) ABLE AND MOSTLY WILLING Panel A: Heterogeneity by importance of performance in determining vote choice Incumbent 0.355∗∗ 0.284∗∗ 0.261∗∗ 0.031∗∗ 0.253∗∗ 0.039∗∗ 0.029∗ 0.077∗∗ (0.044) (0.037) (0.037) (0.010) (0.029) (0.014) (0.013) (0.029) Incumbent × Performance most important −0.007 −0.008 0.009 0.020∗ 0.021 0.021 0.019 0.045† (0.034) (0.029) (0.030) (0.009) (0.022) (0.013) (0.013) (0.027) Benchmark 0.431∗∗ 0.373∗∗ 0.328∗∗ 0.036∗∗ 0.313∗∗ 0.017 0.022 0.043 (0.045) (0.041) (0.039) (0.010) (0.032) (0.014) (0.013) (0.030) Benchmark × Performance most important −0.001 −0.023 −0.010 0.016† 0.007 0.024† 0.019 0.048† (0.031) (0.028) (0.030) (0.010) (0.022) (0.013) (0.013) (0.027) Performance most important range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} Panel B: Heterogeneity by preference for locally-oriented deputies Incumbent 0.370∗∗ 0.192∗∗ 0.198∗∗ −0.018 0.172∗∗ 0.041† 0.035 0.085† (0.074) (0.070) (0.063) (0.023) (0.054) (0.029) (0.029) (0.063) Incumbent × Prefer locally-oriented −0.019 0.130∗ 0.089 0.068∗∗ 0.114∗ −0.002 −0.009 −0.012 deputies (0.078) (0.077) (0.071) (0.027) (0.062) (0.033) (0.033) (0.071) Benchmark 0.502∗∗ 0.286∗∗ 0.302∗∗ 0.014 0.279∗∗ 0.028 0.028 0.062 (0.074) (0.074) (0.066) (0.022) (0.057) (0.028) (0.027) (0.060) Benchmark × Prefer locally-oriented −0.098 0.123† 0.036 0.029 0.046 −0.016 −0.010 −0.028 deputies (0.078) (0.076) (0.070) (0.024) (0.060) (0.032) (0.032) (0.068) Prefer locally-oriented deputies range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} Observations 3,942 3,932 3,928 3,999 3,891 3,999 3,998 3,998 Outcome range {1,...,5} {1,...,5} {1,...,5} {0,1} [−2.3,1.9] {0,1} {0,1} [−1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.00 0.70 0.70 −0.00 Control outcome std. dev. 1.07 0.90 1.09 0.49 1.00 0.46 0.46 1.00 Notes: Each regression includes randomization block and enumerator fixed effects a lagged dependent variable (or pre-treatment incumbent vote intention as a proxy). Lower-order (standardized) interaction terms are included but not shown. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from pre-specified one-sided t tests; ‡ p < 0.1, § p < 0.05, §§ p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. 17 18 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL TABLE 5 Leaflet Recall (Endline Survey) Received Received Received Received leaflet duties incumbent previous information information incumbent information (1) (2) (3) (4) ∗∗ Any treatment 0.921 (0.010) Duties 0.881∗∗ (0.009) Incumbent 0.920∗∗ (0.009) Benchmark 0.937∗∗ 0.924∗∗ (0.009) (0.007) Two-sided null: Incumbent = Benchmark (p value) 0.03 Observations 3,875 3,875 3,875 3,875 Outcome range {0,1} {0,1} {0,1} {0,1} Control outcome mean 0.07 0.01 0.01 0.00 Control outcome std. dev. 0.25 0.09 0.09 0.06 Note: Each regression includes randomization block and (baseline and endline) enumerator fixed effects. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from prespecified one-sided t tests. believe that politicians possess the capacity to respond efentiating features of the leaflet’s content. These high refectively before engaging in costly requests is tentative becall rates substantially exceed those documented in other cause it did not persist at endline. field experiments generally using lower intensity information dissemination mechanisms.16 This suggests that limited recall is unlikely to represent a significant barrier to political accountability, at least within a month of inLonger Term Effects of Information formation dissemination. Provision To understand whether the immediate increases in elecBeliefs About Incumbent Performance toral and nonelectoral efforts to hold incumbents to acPersist After the Election count translate into persisting beliefs, voting behavior, Updated beliefs about the incumbent also largely perand costly attempts to contact incumbents, we turn to sisted a month after treatment. Columns (1) and (2) of our endline survey and polling station electoral returns. Table 6 show that voters who received performance information continued to register significantly higher ratings of the incumbent, and believe that the incumbent perVoters Correctly Recall Leaflet Content Type formed better than previous incumbents (Hypothesis 1). After the Election Appendix Table I1 (p. SI16) further reports that treated voters continued to express greater certainty about their The endline results in Table 5 show that most responbeliefs. Treatment effects roughly halved relative to the dents continued to recall the treatment information immediate effect of providing incumbent performance around a month after its dissemination. Column (1) information. Given the lack of evidence that information shows that virtually all treated respondents correctly recalled receiving the LEGS-Africa leaflet, whereas only 7% 16 of control respondents incorrectly recalled receiving the For example, the average treated respondent across the Metaketa studies was only 7 percentage points more likely to recall the inleaflet. Columns (2)–(4) further demonstrate that almost formation’s substance than generally uninformed control responas many respondents correctly remembered the differdents (Dunning et al. 2019). TABLE 6 Average Effects of Information Treatments (Endline Survey) Incumbent evaluation outcomes Incumbent contact request outcomes Incumbent Relative Incumbent Incumbent Incumbent Request Request Request Called Incumbent overall performance vote vote evaluation incumbent incumbent hotline hotline contact request performance (vs. previous) (validated) index (ICW) visit conversation number index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) ABLE AND MOSTLY WILLING Panel A: All information treatment conditions Duties 0.012 −0.018 −0.049 −0.017 −0.074 0.006 −0.004 0.014† 0.016 0.085† (0.053) (0.052) (0.032) (0.032) (0.062) (0.006) (0.038) (0.010) (0.022) (0.053) Incumbent 0.149∗∗ 0.113∗∗ −0.042 −0.033 0.059 0.006 −0.015 0.006 0.013 0.065 (0.050) (0.047) (0.031) (0.030) (0.055) (0.007) (0.039) (0.010) (0.020) (0.052) Incumbent × Duties 0.024 0.060 0.036 0.030 0.102 −0.013 −0.005 −0.001 −0.004 −0.070 (0.070) (0.069) (0.043) (0.043) (0.081) (0.009) (0.051) (0.014) (0.029) (0.073) Benchmark 0.235∗∗ 0.256∗∗ −0.017 −0.004 0.202∗∗ 0.017∗∗ −0.017 0.004 0.056∗∗ 0.170∗∗ (0.049) (0.045) (0.032) (0.031) (0.056) (0.005) (0.047) (0.011) (0.022) (0.051) Benchmark × Duties 0.020 −0.021 0.038 0.020 0.028 −0.015‡ 0.157 0.004 −0.070§ −0.164§ (0.072) (0.069) (0.044) (0.044) (0.082) (0.008) (0.151) (0.015) (0.030) (0.074) Panel B: Pooling duties treatment conditions Incumbent 0.161∗∗ 0.144∗∗ −0.024 −0.018 0.110∗∗ 0.000 −0.017 0.006 0.011 0.029 (0.035) (0.034) (0.021) (0.021) (0.039) (0.005) (0.029) (0.007) (0.015) (0.036) Benchmark 0.246∗∗ 0.246∗∗ 0.002 0.007 0.217∗∗ 0.009∗∗ 0.063 0.006 0.021† 0.087∗∗ (0.036) (0.034) (0.022) (0.022) (0.041) (0.004) (0.054) (0.007) (0.015) (0.037) Benchmark 0.085∗∗ 0.102∗∗ 0.026† 0.024 0.107∗∗ 0.009∗ 0.080 0.000 0.010 0.058∗ Incumbent (0.033) (0.031) (0.020) (0.020) (0.036) (0.004) (0.065) (0.007) (0.014) (0.035) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,…,5} {1,…,5} {0,1} {0,1} [−2.8,1.9] {0,1} {0,1} {0,1} {0,1} [−7.3,1.5] Control outcome mean 3.08 3.46 0.64 0.53 0.00 0.98 0.98 0.95 0.11 −0.00 Control outcome std. 0.93 0.95 0.48 0.50 1.00 0.14 0.14 0.21 0.32 1.00 dev. Note: Each regression includes randomization block and (baseline and endline) enumerator fixed effects and a lagged dependent variable (or pretreatment incumbent vote intention as a proxy). Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from pre-specified one-sided t tests; ‡ p < 0.1, § p < 0.05, §§ p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. 19 20 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL spilled across villages (see Table I8), these estimates sugidence. Although the leaflet’s content did not signifigest that the smaller treatment effects at endline could recantly influence the average respondent’s self-reported flect decay in information’s effects, the influence of other vote, panels C and particularly D suggest that the inforfactors during the election campaign, and/or posttreatmation treatments did relatively increase incumbent supment interactions within villages relating to the informaport among respondents who—at baseline—regarded tion’s provision (see below). performance information as the most important factor The persistent changes in beliefs remain generally in determining their vote choice or preferred locally oriconsistent with Bayesian updating. First, the hypothesis ented deputies (Hypothesis 4).18 These findings suggest tests at the foot of panel B in Table 6 indicate that benchthat only a small share of the younger rural voters in our marked information continued to induce more favorable sample ultimately acted on their updated beliefs. updating than receiving incumbent-only performance However, our sample of voters may respond difindicators (Hypothesis 3). Second, the heterogeneous ferently to incumbent performance information than treatment effects with respect to reported incumbent the broader, more politically experienced, electorate. In performance in Table 7 suggest that voters concentrated particular, younger voters are far less likely to turn out: their attention on certain types of information over time: within our sample, a 20-year-old was more than 20 pera comparison of panel A with panel B indicates that votcentage points less likely to turn out than a 33-year-old. ers increasingly prized higher levels of local performance, Moreover, voters who have not previously voted were and also became more likely to view national-oriented significantly less likely to value performance indicators legislative activity negatively (Hypotheses 1 and 4). The and locally oriented politicians. Consequently, if the lack of heterogeneous effects in panels C and D again relatively politically inexperienced voters in our sample suggests that voters valuing deputy performance and often do not turn out or vote on the basis of other priorities differently nevertheless updated similarly from factors, even when they persistently update about the the information provided. The results thus suggest that, incumbent’s performance, then our theoretical framealthough the effects of information persisted, voters work suggests that electoral accountability may rely on increasingly emphasized local performance indicators more seasoned likely voters receiving the information and benchmarked information over time. and updating similarly. To better approximate electorate-level voting behavior, Table 8 first restricts our survey sample to the Performance Information Influences the 38% of voters who reported turning out in the 2012 Vote Choices of Likely Voters parliamentary election. Such voters were 14 percentage points more likely to report voting in 2017. The point esWe next examine whether the beliefs that persisted timates in columns (1)–(4) indicate that previous voters through endline carried through to voting behavior. We immediately updated their vote intentions and persistest this crucial link in the chain of electoral accountabiltently updated their posterior beliefs similarly to our full ity by examining self-reported vote choices, before anasample of young registered voters, suggesting that any lyzing polling station–level electoral returns. differences in behavior are unlikely to reflect differential The self-reported survey data provide mixed evpriors or differential updating from the information idence that incumbent performance information ultireceived. However, validated vote choice depicts a stark mately enhances electoral accountability. First, panel A contrast. Unlike the full sample of voters, columns (3) of Table 6 offers little systematic evidence of an inand (5) show that previous voters remained 2–3 percentcrease in incumbent voting on average, even after colage points more likely to vote for the incumbent across umn (4) applies our vote validation criteria (Hypothesis baseline and endline surveys (Hypothesis 4), although 1).17 The pooled estimates in panel B are also indistinthis is imprecisely estimated in this subsample. Furtherguishable from zero. Nevertheless, consistent with voters’ more, in contrast with the full sample, the heterogeneous more favorable updating from benchmarked informaeffects in column (6) show that the persisting belief tion, panel B indicates that the benchmark increased BBY that incumbents with higher performance scores on the voting by around 2.5 percentage points more than the local performance dimension—the primary driver of incumbent-only performance information. Second, the differences in election-time beliefs—are better overall heterogeneous effects in Table 7 also yield mixed ev18 Appendix Table I7 (p. SI21) shows that treatment information 17 Appendix section I.10 (pp. SI25–SI28) shows that treatment did did not increase the importance of performance in determining not significantly affect turnout. vote choice. TABLE 7A Effects of Information Treatments (Endline Survey), by Incumbent Performance Level Incumbent evaluation outcomes Incumbent contact request outcomes Incumbent Relative Incumbent Incumbent Incumbent Request Request Request Called Incumbent overall performance vote vote evaluation incumbent incumbent hotline hotline contact request performance (vs. previous) (validated) index (ICW) visit conversation number index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Heterogeneity by (standardized) reported performance level Incumbent 0.161∗∗ 0.143∗∗ −0.024 −0.018 0.109∗∗ 0.000 −0.017 0.006 0.010 0.029 ABLE AND MOSTLY WILLING (0.035) (0.034) (0.021) (0.021) (0.039) (0.004) (0.029) (0.007) (0.015) (0.036) Incumbent × Overall −0.060 −0.048 0.001 −0.001 −0.056 −0.006 −0.007 0.015† 0.020 0.035 performance (ICW) (0.044) (0.044) (0.027) (0.027) (0.053) (0.005) (0.012) (0.010) (0.021) (0.053) Benchmark 0.245∗∗ 0.245∗∗ 0.002 0.006 0.215∗∗ 0.009∗∗ 0.063 0.006 0.021† 0.087∗∗ (0.035) (0.034) (0.022) (0.022) (0.040) (0.004) (0.054) (0.007) (0.015) (0.037) Benchmark × Overall −0.049 −0.040 0.021 0.013 −0.032 −0.009‡ −0.004 −0.000 0.012 −0.029 performance (ICW) (0.045) (0.046) (0.027) (0.028) (0.055) (0.005) (0.012) (0.011) (0.023) (0.056) Overall performance (ICW) [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] [−2.37,1.12] range Panel B: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.162∗∗ 0.145∗∗ −0.024 −0.018 0.111∗∗ 0.000 −0.016 0.006 0.010 0.028 (0.034) (0.034) (0.021) (0.021) (0.039) (0.004) (0.028) (0.007) (0.015) (0.036) Incumbent × National −0.028 −0.021 −0.000 −0.013 −0.022 −0.008 −0.007 0.003 0.024∗ 0.003 performance (ICW) (0.034) (0.036) (0.024) (0.025) (0.046) (0.007) (0.029) (0.007) (0.013) (0.041) Incumbent × Local 0.068∗∗ 0.097∗∗ 0.004 0.015 0.065† 0.009∗ 0.016 0.014∗ 0.012 0.097∗∗ performance (ICW) (0.034) (0.036) (0.023) (0.024) (0.044) (0.005) (0.018) (0.007) (0.014) (0.037) Benchmark 0.245∗∗ 0.245∗∗ 0.002 0.006 0.215∗∗ 0.009∗∗ 0.065 0.006 0.020† 0.086∗∗ (0.035) (0.034) (0.022) (0.022) (0.040) (0.004) (0.054) (0.007) (0.015) (0.036) Benchmark × National −0.039 −0.080§ 0.012 0.009 −0.060 −0.017§ −0.097 −0.020§ 0.012 −0.113§ performance (ICW) (0.034) (0.037) (0.026) (0.026) (0.044) (0.007) (0.074) (0.009) (0.013) (0.051) Benchmark × Local 0.023 0.083∗∗ 0.012 0.018 0.059† 0.011∗ 0.003 0.022∗∗ 0.009 0.117∗∗ performance (ICW) (0.036) (0.036) (0.024) (0.024) (0.044) (0.006) (0.016) (0.008) (0.014) (0.044) National performance (ICW) [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] [−1.42,2.21] range Local performance (ICW) [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] [−1.39,1.35] range Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5} {1,...,5} {0,1} {0,1} [−2.8,1.9] {0,1} {0,1} {0,1} {0,1} [−7.3,1.5] Control outcome mean 3.08 3.46 0.64 0.53 0.00 0.98 0.98 0.95 0.11 −0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 1.00 0.14 0.14 0.21 0.32 1.00 Notes: Each regression includes randomization block and (baseline and endline) enumerator fixed effects and a lagged dependent variable (or pre-treatment incumbent vote intention as 21 a proxy). Lower-order (standardized) interaction terms are included but not shown. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from pre-specified one-sided t tests; ‡ p < 0.1, § p < 0.05, §§ p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. 22 TABLE 7B Effects of Information Treatments (Endline Survey), by Voter Criteria for Preferred Politicians Incumbent evaluation outcomes Incumbent contact request outcomes Incumbent Relative Incumbent Incumbent Incumbent Request Request Request Called Incumbent overall performance vote vote evaluation incumbent incumbent hotline hotline contact request performance (vs. previous) (validated) index (ICW) visit conversation number index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Heterogeneity by importance of performance in determining vote choice Incumbent 0.161∗∗ 0.144∗∗ −0.024 −0.018 0.110∗∗ −0.000 −0.016 0.006 0.011 0.029 (0.035) (0.034) (0.021) (0.021) (0.039) (0.005) (0.028) (0.007) (0.015) (0.037) Incumbent × Performance 0.021 −0.007 0.001 0.011 0.015 −0.006 0.007 −0.001 0.003 −0.026 most important (0.034) (0.033) (0.019) (0.019) (0.037) (0.004) (0.013) (0.007) (0.012) (0.035) Benchmark 0.246∗∗ 0.246∗∗ 0.002 0.006 0.217∗∗ 0.009∗ 0.063 0.006 0.021 0.087∗ (0.036) (0.034) (0.022) (0.021) (0.041) (0.004) (0.053) (0.007) (0.015) (0.037) Benchmark × Performance −0.018 0.002 0.025 0.031 0.032 −0.005 0.060 −0.003 0.028∗ 0.010 most important (0.035) (0.033) (0.019) (0.019) (0.037) (0.004) (0.052) (0.007) (0.014) (0.033) Performance most important {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} range Panel B: Heterogeneity by preference for locally-oriented deputies Incumbent 0.104† 0.103∗ −0.095§§ −0.073§ −0.008 −0.018§ −0.004 0.006 0.018 −0.049 (0.064) (0.060) (0.035) (0.035) (0.067) (0.007) (0.038) (0.014) (0.025) (0.065) Incumbent × Prefer 0.079 0.057 0.099∗∗ 0.077∗ 0.164∗∗ 0.025∗∗ −0.015 −0.001 −0.009 0.108† locally-oriented deputies (0.079) (0.070) (0.042) (0.043) (0.080) (0.010) (0.045) (0.017) (0.028) (0.082) Benchmark 0.191∗∗ 0.227∗∗ −0.048 −0.045 0.121† −0.003 0.283 0.002 0.050∗ 0.062 (0.062) (0.062) (0.038) (0.038) (0.074) (0.006) (0.281) (0.013) (0.029) (0.060) Benchmark × Prefer 0.077 0.025 0.070† 0.072† 0.134† 0.016∗∗ −0.306 0.005 −0.041 0.035 locally-oriented deputies (0.075) (0.073) (0.043) (0.045) (0.086) (0.008) (0.323) (0.015) (0.032) (0.073) Prefer locally-oriented deputies {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} range Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5} {1,...,5} {0,1} {0,1} [−2.8,1.9] {0,1} {0,1} {0,1} {0,1} [−7.3,1.5] Control outcome mean 3.08 3.46 0.64 0.53 0.00 0.98 0.98 0.95 0.11 −0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 1.00 0.14 0.14 0.21 0.32 1.00 Notes: Each regression includes randomization block and (baseline and endline) enumerator fixed effects and a lagged dependent variable (or pre-treatment incumbent vote intention as a proxy). Lower-order (standardized) interaction terms are included but not shown. Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from pre-specified one-sided t tests; ‡ p < 0.1, § p < 0.05, §§ p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL ABLE AND MOSTLY WILLING 23 did translate into a significantly higher probability of comprise at least 50% of registered voters at their polling treatment increasing self-reported incumbent votes. station.19 These findings suggest that incumbent performance The results in Table 9 largely mirror the self-reported information may have induced electoral accountabilbehavior of the survey respondents who reported votity among experienced voters, who rewarded highly ing in 2012. We do not observe a notable average treatperforming locally oriented incumbents. ment effect of incumbent performance information on Due to high levels of within-village information incumbent vote share in columns (1) and (3). Howdiffusion, such responses could translate into polling ever, columns (2) and (4) show that treatment effects instation–level voting outcomes. Indeed, Appendix section creased with performance on the local dimension that I.6 (pp. SI20–SI22) shows that almost 40% of our nine voters value most (Hypothesis 1). The estimates imply treated voters per village, and at least one within every that a standard deviation increase in an incumbent’s lovillage, discussed the leaflet with others in their village. cal performance increased the incumbent’s vote share by Appendix section I.8 (pp. SI22–SI25) further reports that around 3 percentage points. In sum, these findings sugincumbents, and to a lesser extent challengers, also disgest that likely voters in treated villages who did not reseminated treatment information through their operaceive leaflets directly nevertheless learned about incumtives and community meetings. In the small and densely bent performance and engaged in electoral accountabilconnected rural villages that comprise our sample, treatity. In contrast, younger and first-time voters who were ment information thus likely reached many citizens beless likely to vote did not. fore the election through interpersonal interactions. To examine official electoral returns that are not susceptible 19 to self-reporting biases, we restrict our analysis to polling Appendix Table I11 (p. SI27) reports similar results weighting stations containing the 284 villages in our sample that all polling stations by the share of voters at a given polling station from the experimental village. TABLE 8 Effects of Information Treatments, Among Respondents Those Who Turned Out in 2012 (Baseline and Endline Surveys) Incumbent overall Incumbent vote Incumbent vote performance (endline) intention (validated) (1) (2) (3) (4) (5) (6) Incumbent 0.134∗ 0.136∗ 0.015 0.018 0.034 0.034 (0.059) (0.058) (0.015) (0.015) (0.033) (0.032) Benchmark 0.235∗∗ 0.231∗∗ 0.022 0.025 0.034 0.041 (0.057) (0.056) (0.016) (0.016) (0.031) (0.030) Incumbent × National performance (ICW) −0.039 −0.022 −0.060† (0.053) (0.014) (0.032) Incumbent × Local performance (ICW) 0.072 0.032† 0.058 (0.060) (0.016) (0.037) Benchmark × National performance (ICW) −0.077 −0.013 −0.019 (0.060) (0.015) (0.035) Benchmark × Local performance (ICW) 0.027 0.036∗ 0.083∗ (0.060) (0.017) (0.035) Observations 1,469 1,469 1,528 1,528 1,435 1,435 Outcome range {1,…,5} {1,…,5} {0,1} {0,1} {0,1} {0,1} Control outcome mean 3.10 3.10 0.63 0.63 0.59 0.59 Control outcome std. dev. 0.93 0.93 0.48 0.48 0.49 0.49 Note: Each regression includes randomization block and (baseline and, where relevant, endline) enumerator fixed effects and a lagged dependent variable (or pretreatment incumbent vote intention as a proxy). Observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. Given that these hypotheses were not prespecified, † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from two-sided t tests. 24 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL TABLE 9 Effects of Information Treatments (Polling Station Data) Incumbent vote Incumbent vote share share (proportion (proportion of of turnout) registered voters) (1) (2) (3) (4) Incumbent 0.001 −0.013 0.006 −0.006 (0.023) (0.025) (0.018) (0.020) Benchmark −0.003 −0.016 −0.003 −0.014 (0.024) (0.026) (0.018) (0.020) Incumbent × National performance (ICW) −0.000 0.012 (0.032) (0.021) Incumbent × Local performance (ICW) 0.043∗ 0.030∗ (0.025) (0.017) Benchmark × National performance (ICW) 0.021 0.011 (0.033) (0.021) Benchmark × Local performance (ICW) 0.020 0.023† (0.026) (0.017) Observations 284 284 284 284 Outcome range [0.06,0.99] [0.06,0.99] [0.02,0.73] [0.02,0.73] Control outcome mean 0.71 0.71 0.41 0.41 Control outcome std. dev. 0.17 0.17 0.13 0.13 Note: Each regression includes randomization block fixed effects and a lagged dependent variable. Observations are unweighted, and polling stations for which the village in our sample comprises less than 50% of registered voters are excluded. Robust standard errors are in parentheses. † p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01 from prespecified one-sided t tests. Performance Information Increases through which to contact incumbents were almost uniIncumbent Contact Requests versally sought, the 20% increase in the costly act of actually texting or calling the hotline in column (9) provides Although intentions to engage in electoral accountabilthe most compelling evidence. Aggregated as an index, ity somewhat weakened between the time of information column (10) of panel B reports that benchmarked perdelivery and the election itself, nonelectoral requests to formance information induced a 0.09 standard deviation contact incumbents represent a different kind of effort to increase across such behaviors. hold incumbents to account. The contact requests that As with electoral accountability, the heterogeneous we study are less likely than vote choice to depend on effects in Table 7 further demonstrate that increased the relative importance of performance metrics to voters, hotline usage was greatest in departments where the and are less susceptible to campaign-based interactions incumbent-only and benchmarked treatments reported between survey waves. greatest incumbent performance (Hypothesis 1). In line Consistent with these intuitions, voters receiving with self-reported behaviors among likely voters, this performance information continued to make greater efprincipally reflected local performance. Consistent with fort to contact the reelected incumbents a month afthe baseline results, requests were no greater among reter information was delivered. As at baseline, columns spondents who value performance information more in (6)–(10) of Table 6 indicate that the benchmarked inmaking vote choices. These results indicate that treatformation significantly increased requests to contact inment caused voters to durably engage in costly efforts to cumbents (Hypotheses 1 and 3). Given that low-cost reseek accountability from better performing incumbents, quests for visits, conversations, and a hotline number consistent with voters expecting greater responsiveness. ABLE AND MOSTLY WILLING 25 Conclusion ties to maximize the effectiveness of information dissemination campaigns. Finally, by focusing on information Given the mixed evidence that information campaigns provision within election campaigns, we abstracted from can support political accountability, this article examthe incentives and external constraints driving incumined the extent to which accountability failures reflect bent performance in the first place and the processes of voter-level constraints or the types of information that candidate (de)selection that dictate the candidates that voters receive. By abstracting from issues of informavoters ultimately choose between. Each factor could contion dissemination and take-up, we dissect the process strain political accountability at an earlier stage and merlinking the personal delivery and explanation of incumits further examination (see e.g. Grossman and Michebent performance information to electoral and nonlitch 2018). electoral efforts by voters to hold incumbent legislaAlthough rural areas where deputies from the prestors to account. Our findings show that rural Senegalese ident’s party generally win may be exposed to fewer voters engaged in sophisticated information processing, competing political influences, there are good reasons largely retained their updated beliefs, and regarded local to believe that our anatomy of political accountability projects/transfers and temporal benchmarks as particuextends beyond our sample and intervention. First, solarly informative about incumbent quality. In contrast, phisticated responses among voters with low educations information about national performance and incumbent levels suggest that voters across the world could draw duties had little systematic effect in a context where delivsimilar inferences. Second, parliamentary elections in ering “pork” is widely regarded as a legislator’s primary Senegal share many features with elections in other function. Persisting beliefs and initial intentions to hold developing democracies, including the dominance of politicians to account did translate into electoral rewards the president’s party and nontrivial levels of clientelism. for better performing incumbents among more experiBecause information is likely to have weaker effects in enced voters, but not among unlikely voters or voters contexts where electoral competition is more limited prioritizing other issues. With respect to nonelectoral ac(Grossman and Michelitch 2018), information’s capacity countability, voters persistently engaged in greater costly to enhance bottom-up accountability may be larger efforts to contact the best performing winning incumwhere party switching is more common and national bents after the election. waves influence voting behavior less than in Senegal. Taken together, these findings illustrate that voters— Third, because our leaflets are similar in design to upon receiving information they deem credible and previous studies (Chong et al. 2015; Dunning et al. relevant—are able and mostly willing to hold politicians 2019; Gottlieb 2016; Humphreys and Weinstein 2012), to account. This highlights the importance of underour findings may help to direct researchers in other standing how factors other than voter-level constraints contexts toward the types of impediments to electoral may sustain low-accountability equilibria in developing accountability that we highlight here. contexts. First, future research might establish both the Nevertheless, several features of our study’s context most effective and scalable means through which informerit further research to explore the generality and mation can be communicated to voters en masse and broader implications of our findings. First, because the factors driving demand for and supply of such incurrent incumbents’ performance generally exceeded formation absent external campaigns. Second, our parboth voters’ prior beliefs and previous incumbents’ pertial equilibrium focuses only briefly addressed election formance in our sample, it is natural to wonder whether campaign responses to information dissemination. Alvoters would respond differently to unexpectedly or though such responses are often documented (Banerjee comparably poor performance. Although future studies et al. 2011; Cruz, Keefer, and Labonne 2019), little is yet should explore this more extensively, Appendix section known about whether or when politicians influence poI.4 provides evidence that voters updated negatively litical accountability by complementing or refuting inand sought to sanction the current incumbent in (pp. formation dissemination campaigns. Third, our finding SI17–SI20) Oussouye—a department where the previous that voters are keen to take advantage of a rare nonelecincumbent’s local performance exceeded the current intoral opportunity to engage with incumbents suggests cumbent’s. Second, it is similarly important to establish that a lack of contact opportunities may also limit politwhether voters regard across-party temporal benchical accountability. Civil society organizations and policy marks, like those that we provided, as more relevant makers seeking to increase political accountability might than within-party benchmarks that filter out party-level therefore consider facilitating more of these opportunifactors. This may be pertinent in regimes where elections 26 ABHIT BHANDARI, HORACIO LARREGUY, AND JOHN MARSHALL effectively select candidates from within the dominant Casey, Katherine. 2018. “Radical Decentralization: Does Comparty. Finally, further research should explore exactly munity Driven Development Work?” Annual Review of Economics 10:139–65. what citizens demand when they perceive their representatives to be more responsive or effective, and if they ulChong, Alberto, Ana De La O, Dean Karlan, and Leonard Wantchekon. 2015. “Does Corruption Information Inspire timately receive it. Whether this constitutes standard disthe Fight or Quash the Hope? A Field Experiment in Mexico tributive politics is important to understand for governon Voter Turnout, Choice and Party Identification.” Journal ments and donors, who often seek to use accountability of Politics 77(1):55–71. mechanisms to promote programmatic policy making. Cruz, Cesi, Philip Keefer, and Julien Labonne. 2019. “Buying Informed Voters: New Effects of Information on Voters and Candidates.” Working paper, https://julienlabonne.files. wordpress.com/2020/04/ppcrv_web_042020.pdf (accessed References June 26, 2020). Dunning, Thad, Guy Grossman, Macartan Humphreys, Susan Adida, Claire, Jessica Gottlieb, Gwyneth McClendon, and Eric Hyde, Craig McIntosh, and Gareth Nellis. 2019. InformaKramon. forthcoming. “Breaking the Clientelistic Equilibtion, Accountability, and Cumulative Learning: Lessons from rium: The Joint Importance of Salience and Coordination.” Metaketa I. Cambridge: Cambridge University Press. Comparative Political Studies 53(6):851–891. Fearon, James D. 1999. “Electoral Accountability and the ConAker, Jenny C., Paul Collier, and Pedro C Vicente. 2017. “Is Introl of Politicians: Selecting Good Types versus Sanctionformation Power? Using Mobile Phones and Free Newspaing Poor Performance.” In Democracy, Accountability, and pers During An Election in Mozambique.” Review of EcoRepresentation, eds. Adam Przeworski, Susan Stokes, and nomics and Statistics 99(2):185–200. Bernard Manin. Cambridge: Cambridge University Press, 55–97. Arias, Eric, Horacio Larreguy, John Marshall, and Pablo Querubín. 2018a. “Does the Content and Mode of DelivFerraz, Claudio, and Frederico Finan. 2008. “Exposing Corery of Information Matter for Electoral Accountability? Evirupt Politicians: The Effects of Brazil’s Publicly Released Audence from a Field Experiment in Mexico.” Working paper, dits on Electoral Outcomes.” Quarterly Journal of Economics https://tinyurl.com/y6tkfb7p (accessed June 26, 2020). 123(2):703–45. Arias, Eric, Horacio Larreguy, John Marshall, and Pablo Gomez, Brad T., and J. Matthew Wilson. 2006. “Cognitive HetQuerubín. 2018b. “Priors Rule: When Do Malfeasance erogeneity and Economic Voting: A Comparative Analysis Revelations Help and Hurt Incumbent Parties?” NBER of Four Democratic Electorates.” American Journal of PolitWorking Paper 24888, https://www.nber.org/papers/ ical Science 50(1):127–45. w24888 (accessed June 26, 2020). Gottlieb, Jessica. 2016. “Greater Expectations: A Field ExperiArias, Eric, Pablo Balán, Horacio Larreguy, John Marshall, and ment to Improve Accountability in Mali.” American Journal Pablo Querubín. 2019. “Information Provision, Voter Coorof Political Science 60(1):143–57. dination, and Electoral Accountability: Evidence from MexGrossman, Guy, and Kristin Michelitch. 2018. “Informaican Social Networks.” American Political Science Review tion Dissemination, Competitive Pressure, and Politi113(2):475–98. cian Performance Between Elections: A Field Experiment Aytaç, Selim Erdem. 2018. “Relative Economic Performance in Uganda.” American Political Science Review 112(2): and the Incumbent Vote: A Reference Point Theory.” Jour280–301. nal of Politics 80(1):16–29. Grossman, Guy, Macartan Humphreys, and Gabriella Banerjee, Abhijit V., Selvan Kumar, Rohini Pande, and Felix Sacramone-Lutz. 2014. “I Wld Like U WMP to Extend Su. 2011. “Do Informed Voters Make Better Choices? Electricity 2 Our VillageÑ: On Information Technology and Experimental Evidence from Urban India.” Working paper, Interest Articulation.” American Political Science Review https://epod.cid.harvard.edu/sites/default/files/2018-02/ 108(3):688–705. do_informed_voters_make_better_choices_experimental_ Humphreys, Macartan, and Jeremy Weinstein. 2012. “Policevidence_from_urban_india.pdf (accessed June 26, 2020). ing Politicians: Citizen Empowerment and Political Beck, Linda J. 2012. “Senegal.” In Countries at the Crossroads Accountability in Uganda Preliminary Analysis.” Working 2011: An Analysis of Democratic Governance, ed. Jake Dizard, paper, https://tinyurl.com/y4h6uasz (accessed June 26, Christopher Walker, and Vanessa Tucker. Lanham, MD, 2020). USA: Rowman & Littlefield, 581–600. Kendall, Chad, Tommaso Nannicini, and Francesco Trebbi. Boas, Taylor C., F. Daniel Hidalgo, and Marcus André Melo. 2014. “How Do Voters Respond to Information? Evidence 2019. “Norms Versus Action: Why Voters Fail to Sanction from a Randomized Campaign.” American Economic Review Malfeasance in Brazil.” American Journal of Political Science 105(1):322–53. 63(2):385–400. Koter, Dominika. 2013. “King Makers: Local Leaders Bussell, Jennifer. 2019. Clients and Constituents: Political Reand Ethnic Politics in Africa.” World Politics 65(2): sponsiveness in Patronage Democracies. Oxford: Oxford Uni187–232. versity Press. Lieberman, Evan S., Daniel N. Posner, and Lily L. Tsai. 2014. Campello, Daniela, and Cesar Zucco, Jr. 2016. “Presiden“Does Information Lead to More Active Citizenship? Evtial Success and the World Economy.” Journal of Politics idence from an Education Intervention in Rural Kenya. ” 78(2):589–602. World Development 60:69–83. ABLE AND MOSTLY WILLING 27 Meyer, Margaret A., and John Vickers. 1997. “Performance Appendix A: Overview of our theoretical framework Comparisons and Dynamic Incentives.” Journal of Political Appendix B: The effect of providing incumbent-only and Economy 105(3):547–81. benchmarked information Powell, Jr., G. Bingham, and Guy D. Whitten. 1993. “A CrossAppendix C: Senegalese parliamentary electoral and National Analysis of Economic Voting: Taking Account of the Political Context.” American Journal of Political Science non-electoral political engagement in comparative con37(2):391–414. text Thomas, Melissa A., and Oumar Sissokho. 2005. “Liaison LegAppendix D: Additional details about sample selection islature: The Role of the National Assembly in Senegal.” Appendix E: Additional information about information Journal of Modern African Studies 43(1):97–117. treatments Zaller, John R. 1992. The Nature and Origins of Mass Opinion. Appendix F: Compliance and experimental validation Cambridge: Cambridge University Press. checks Appendix G: ICW index construction Supporting Information Appendix H: Deviations from pre-analysis plan Appendix I: Additional results Additional supporting information may be found online in the Supporting Information section at the end of the article.
2 All in the Family: Partisan Disagreement and Electoral Mobilization in Intimate Networks— A Spillover Experiment Florian Foos University of Zurich Eline A. de Rooij Simon Fraser University Abstract: We advance the debate about the impact of political disagreement in social networks on electoral participation by addressing issues of causal inference common in network studies, focusing on voters’ most important context of interpersonal influence: the household. We leverage a randomly assigned spillover experiment conducted in the United Kingdom, combined with a detailed database of pretreatment party preferences and public turnout records, to identify social influence within heterogeneous and homogeneous partisan households. Our results show that intrahousehold mobilization effects are larger as a result of campaign contact in heterogeneous than in homogeneous partisan households, and larger still when the partisan intensity of the message is exogenously increased, suggesting discussion rather than behavioral contagion as a mechanism. Our results qualify findings from influential observational studies and suggest that within intimate social networks, negative correlations between political heterogeneity and electoral participation are unlikely to result from political disagreement. Replication Materials: The data, code, and any additional materials required to replicate all analyses in this article are available on the American Journal of Political Science Dataverse within the Harvard Dataverse Network, at: http://dx.doi.org/10.7910/DVN/ZFLG25. I n deciding whether to vote in an election, and Sinclair 2012; Zuckerman 2005; Zuckerman, Dasović, and which party to support, informal political discusFitzgerald 2007). sions with family, friends, and acquaintances play an Consequently, political scientists have long discussed important role. These discussions reinforce social norms, the potential of campaigns to mobilize voters indirectly provide us with political information, allow us to underby encouraging political discussion within personal netstand the views and reasoning of others, and enable us to works, such as that of the family, in the run-up to an express—and maybe even persuade others of—our own election (McClurg 2003; Rolfe 2012). Glaser, for instance, views (Conover, Searing, and Crewe 2002; Huckfeldt and argued that “the most influential canvassers are the rest of Sprague 1987, 1991; Mutz 2002b, 2006; Sinclair 2012; the people in the voter’s family” and that “if party workers Zuckerman, Dasović, and Fitzgerald 2007). Ultimately, or civic organizations want the greatest return on the use they are said to result in increased political engagement of their scarce services, they should make contact with and electoral mobilization (Cutts and Fieldhouse 2009; the most politicized member of a household [ . . . ] and Huckfeldt and Sprague 1995; Klofstad 2007, 2011, 2015; motivate that member to bring all the other members to Lake and Huckfeldt 1998; McClurg 2003; Rolfe 2012; the polls” (1959, 570). Florian Foos is Postdoctoral Researcher, Department of Political Science, University of Zurich, Affolternstrasse 56, CH-8050 Zurich, Switzerland (foos@ipz.uzh.ch). Eline A. de Rooij is Assistant Professor, Department of Political Science, Simon Fraser University, 8888 University Drive, Burnaby, B.C., V5Y 1S6 Canada (eline_de_rooij@sfu.ca). We are grateful to the Labour Party organizers and volunteers in Birmingham, UK, for implementing the telephone canvassing campaign. Without their dedication, hard work, and support, this study would not have been possible. We are also thankful to Marcus A. Roberts for his support in facilitating contact with the campaign. Finally, we are indebted to Peter Aronow, Kevin Collins, Alexander Coppock, Geoffrey Evans, Stephen D. Fisher, Edward Fieldhouse, Fabrizio Gilardi, Donald P. Green, Christian Müller, Mark Pickup, Vincent Pons, Marco Steenbergen, Catherine de Vries, the conference participants of the 2014 Annual Meetings of the American and Midwest Political Science Associations, the participants of the publication seminar at the Department of Political Science at the University of Zurich, and three anonymous reviewers for their advice and helpful comments. Florian Foos acknowledges studentships from the Nuffield College Sociology and Politics groups for the period when conducting fieldwork for this study. American Journal of Political Science, Vol. 61, No. 2, April 2017, Pp. 289–304 \002 C 2016, Midwest Political Science Association DOI: 10.1111/ajps.12270 289 290 FLORIAN FOOS AND ELINE A. DE ROOIJ But, while studies show a strong relationship becongruence between household members’ party and tween political discussion within personal networks and policy preferences (Stoker and Jennings 2005), still no political participation (Bond et al. 2012; Klofstad 2011, more than an estimated 40–60% of two-voter households 2015; McClurg 2003; Rolfe 2012), there is an unresolved can be described as politically homogeneous in partisan debate about the extent to which indirect mobilization terms (Bélanger and Eagles 2007; Johnston et al. 2005; is dependent on the degree of political agreement within Zuckerman, Fitzgerald, and Dasović 2005). We ask to the network (Bélanger and Eagles 2007; Fitzgerald and what extent intrahousehold mobilization during an Curtis 2012; Huckfeldt, Johnson, and Sprague 2002, election campaign is conditioned by both the degree of 2004; Klofstad 2011; Klofstad, Sokhey, and McClurg heterogeneity of party preferences within the household 2013; McClurg 2003, 2006; Mutz 2002a, 2006; Nir 2011; and the partisan intensity of a campaign message. We Pattie and Johnston 2009). In an influential series of focus on two-voter households, the majority of which observational studies, Mutz (2002a, 2002b, 2006), while are likely to consist of family members, and in particular, acknowledging the benefits of mixed political company couples. Following Huckfeldt, Johnson, and Sprague for political tolerance, has expressed concerns about the (2004), we conceptualize disagreement as partisan correlation between network heterogeneity and electoral disagreement (for a discussion, see Klofstad, Sokhey, and abstention. Her findings suggest that there might be a flip McClurg 2013). Instead of relying on only one network side to increasing tolerance, a demobilizing, “dark side of member’s report of the extent of partisan disagreement in mixed political company” (Mutz 2006, 89). On the other his or her network, though, we use the self-reported party hand, studies such as those conducted by Huckfeldt, preference of each household member, thus reducing the Johnson, and Sprague (2002, 2004), Nir (2005), and systematic biases in respondents’ perceptions (Klofstad, McClurg (2006) dispute the demobilizing impact of Sokhey, and McClurg 2013; Osborn and Mendez disagreement within a network. 2011). The literature on network heterogeneity and politOur research design combines data from a previical participation has faced difficulty in making strong ously conducted randomized campaign experiment in causal inferences (Fowler et al. 2011; Klofstad 2007, 2011; the United Kingdom (UK), which includes information Mutz 2006). Findings from studies relying on randomon the party preferences of household members, with ized campaign experiments, which allow for stronger validated turnout data from the official voter register on causal inferences, support the notion of indirect mobiboth assigned subjects and those household members exlization through small, personal networks by showing cluded from the original experiment. We conduct an imthat conversations with campaign volunteers mobilize not plicit mediation analysis (Gerber and Green 2012) by utionly the contacted individual, but also his or her houselizing campaign messages of different partisan intensity hold members—a phenomenon that is called “spillover” that manipulate the potential for partisan disagreement (Nickerson 2008; Sinclair 2012; Sinclair, McConnell, and within heterogeneous and homogeneous households, and Green 2012). Yet spillover experiments have not addressed we then assess their impact on turnout. Our results show the potential consequences of political (dis)agreement that intrahousehold mobilization effects are larger as a for intrahousehold mobilization. In addition, these exresult of campaign contact in households that consist of periments have exclusively relied on nonpartisan mobiindividuals who hold competing party preferences, and lization messages, notwithstanding the partisan nature larger still when the partisan intensity of the message is of most campaigns. All the while, the partisan intenincreased. Thus, we find no evidence in support of a desity of mobilization messages likely impacts mobilization mobilizing impact of disagreement. In contrast, partisan among members of a household differently depending disagreement within the confined space of the household on whether they are politically like-minded or non-likemight be beneficial for encouraging political debate and minded. ultimately participation. In this article, we integrate the literature on political homogeneity and heterogeneity in social networks and the experimental literature on indirect campaign mobilization. We focus on what is arguably the most Social Influence in Partisan important context of interpersonal influence for voters: Households the household (Cutts and Fieldhouse 2009; Fieldhouse and Cutts 2012; Nickerson 2008; Sinclair, McConnell, Early voting studies established the importance of poand Green 2012; Stoker and Jennings 1995; Zuckerman, litical discussion within personal networks for voting beDasović, and Fitzgerald 2007). While there is a high havior (Berelson, Lazarsfeld, and McPhee 1954; Campbell PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 291 et al. 1960; Downs 1957; Lazarsfeld, Berelson, and Gaudet members if we do not perceive them as politically agree1948). Ever since these early studies, there has been a deable, and therefore useful, sources of information. Among bate about the importance of the nature of the network the few observational studies that have looked explicitly at (Lazarsfeld, Berelson, and Gaudet 1948). Currently, an households, there is some empirical support for the noimportant aspect of this debate concerns the question of tion that individuals in homogeneous—all-Democratic whether discussion is more likely in networks that consist or all-Republican—households are indeed more likely of politically like-minded or non-like-minded individto vote than those in households in which members uals, and how this so-called network homogeneity and differ in their party registration (Bélanger and Eagles heterogeneity (Mutz 2006) ultimately impact on political 2007). participation. On the other hand, exactly because the household It is important to note that much of this current tends to constitute our most intimate personal network, debate does not explicitly address households (for noit can arguably endure a substantial amount of disagreetable exceptions, see Bélanger and Eagles 2007; Bello and ment when it comes to issues that do not directly affect Rolfe 2014). Yet family or household members are regpersonal relationships between family members (Morey, ularly named as the most important or frequent discusEveland, and Hutchens 2012). In fact, political discussions sion partners (Mutz 2006, 126; Sinclair 2012, 27). Couples might be difficult to avoid with people with whom one have been shown to influence each other, to become more interacts so frequently, often occurring as by-products of similar in their party preferences (Stoker and Jennings sharing the same household, and certainly not as easy to 2005; Verba, Schlozman, and Burns 2005; Zuckerman, shun as unwelcome TV channels or online news sites. As Dasović, and Fitzgerald 2007), and to exhibit similar levels with sports, we might enjoy discussing and watching a of turnout (Glaser 1959; Straits 1990; Verba, Schlozman, game with household members even if we support differand Burns 2005). ent teams. Mutz (2002a, 2006) shows that individuals in heteroThe expectation that individuals who live in politigeneous networks are less likely to participate in politics, cally heterogeneous households are able to disagree withand she argues that this is due to the experienced crossout being disagreeable is also based on the observation pressures resulting from the diverging political views that individuals in heterogeneous networks are more tolwithin the network. Specifically, Mutz (2006) suggests erant of other political viewpoints than individuals in that there might be two, interconnected psychological homogeneous discussion networks (Mutz 2002b, 2006). processes at work. First, membership in a politically hetIndeed, Mutz’s (2002b) finding that political discussion erogeneous network may lead to internal ambivalence and political tolerance are higher in heterogeneous netregarding one’s own political position. This ambivalence, works appears to stand in contrast to a theory that would in turn, might lead to negative participation decisions. expect household members to disengage from politics in Second, and most importantly, because individuals try order to avoid conflict and confrontation. Bello and Rolfe to maintain social harmony within their personal net(2014) show that Mutz’s (2002b) finding also generalworks by pleasing all members, they avoid discussion of izes to the UK, and specifically to the household. Moredivisive political issues, which in turn discourages parover, Klofstad (2011) shows that social intimacy (trust) ticipation. An additional potential mechanism, discussed strengthens the relationship between discussion and civic by Huckfeldt, Johnson, and Sprague (2004, 7), builds on participation. Downs’s (1957) argument that individuals seek out disNot only might political discussion be as frequent in cussion partners within their personal networks who are heterogeneous as in homogeneous households, but there well informed and have similar political viewpoints as are also reasons to suspect it might be more frequent. Bello a means to obtain low-cost information. Consequently, and Rolfe (2014) show that not only are “spouses and famthis mechanism also suggests that political disagreement ily members [ . . . ] more likely to be retained as political within personal discussion networks is avoided, albeit indiscussants than other friends, even close friends,” but advertently. “respondents are more likely to continue discussing poliWhen it comes to households, it is certainly plausitics with spouses who hold different political views when ble that these processes work similarly. Maintaining social compared to spouses who hold the same view” (Bello and harmony might be particularly important with those with Rolfe 2014, 141). This might be because, given the already whom we live under one roof, and unlikely to be jeophigh level of homophily in terms of socioeconomic charardized over political issues that are of limited interest acteristics and the living environment, discussing with to the majority of individuals (Mutz 2006, 106). In addia household member who thinks differently about polition, we might avoid political discussion with household tics provides a more effective means of gaining additional 292 FLORIAN FOOS AND ELINE A. DE ROOIJ information about an upcoming election than talking it comes to campaign-induced discussion, homogeneous to a copartisan. Alternatively, within the context of an partisan households might have a higher probability of election campaign, political disagreement might create being contacted by election campaigns than heterogea situation of friendly competition between household neous partisan households because the former provide members that leads to higher levels of political engagethe opportunity to mobilize several potential party ment, and ultimately electoral participation. Politically supporters with one contact. interested individuals—often partisans—are more likely A handful of scholars have conducted studies in to discuss politics, even though this increases the probawhich households are randomly allocated to be contacted bility of political disagreement, because they simply enjoy by nonpartisan campaigns, ensuring that households political discussion (Huckfeldt and Mendez 2008). Comexposed to such an external shock are in expectation the petition between household members might therefore eisame as households that are not (Nickerson 2008; Sinclair ther increase electoral participation indirectly through 2012; Sinclair, McConnell, and Green 2012). Differences fostering political discussion or directly increase turnout in the outcome of interest should consequently be due to if household members decide to “counter-mobilize,” votsocial influence within the household, and not to preexing for the opposing candidate in order to “cancel out” isting compositional and contextual differences between the vote of their household member (Straits 1990, 64–65; households. These randomized field experiments have Wolff 2002, 33). shown that turnout levels among household members are significantly higher because their family members were contacted by election campaigns, confirming that The Challenge of Making Causal contagion of campaign messages indeed occurs (Sinclair, Inferences McConnell, and Green 2012).1 One of the major challenges of research on the impact of network diversity on political participation is separating Contribution and Limitations self-selection into politically like-minded personal discussion networks, and contextual influences, from While randomized spillover experiments have consocial influence within networks (Fowler et al. 2011; tributed much to our understanding of the household Klofstad 2007, 2011; Manski 1993). Self-selection can as the most important location for intrapersonal mobioccur when characteristics of individuals impact both lization, the household itself has remained a “black box” the level of heterogeneity in their discussion network and for experimentalists (Fowler et al. 2011; Nickerson 2008; their level of political participation (Bélanger and Eagles Sinclair, McConnell, and Green 2012). Field experiments 2007, 850; Mutz 2006, 115). For instance, individuals who have neither been able to distinguish between households are more politically interested or have stronger partisan according to the composition of their members’ preexpreferences might be more likely to participate as well isting partisan preferences, nor have they been able to as to surround themselves with like-minded discussion ascertain the impact of (discussions about) explicitly parpartners (Mutz 2006, 115). Contextual influences occur tisan campaigns on electoral mobilization, as all spillover when the external environment has a similar, exogenous experiments have been based on nonpartisan campaign impact on all members of a network—for example, interventions. a political campaign could change network members’ The contribution of this study lies in addressing these political views and participation rates in the same omissions. By utilizing a spillover experimental design, direction, without individuals influencing each other we are able to separate selection effects and preexisting (Fowler et al. 2011, 450; Manski 1993). contextual differences from social influence in assessA few studies have exploited the benefits of panel ing to what extent intrahousehold mobilization occurs data (Bello and Rolfe 2014) or statistical techniques such as a result of an election campaign. Subsequently, we test as propensity score matching (Klofstad, Sokhey, and whether such mobilization is conditioned by the degree of McClurg 2013) for causal inference. While certainly an heterogeneity of party preferences within the household. improvement on earlier correlational work, such studies We perform an implicit mediation analysis (Gerber and do not fully resolve the issue of separating self-selection 1 and contextual influences from social influence, as they Based on a natural experiment that exploits random assignment do not overcome the issue of unobserved individual-level to college dorm rooms, Klofstad (2007, 2011, 2015) also provides evidence of social influence by showing that being randomly exor contextual-level confounders (Arceneaux, Gerber, posed to political discussion leads to increased civic and political and Green 2006; Wooldridge 2010). For example, when participation in and after college. PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 293 FIGURE 1 Variables and Expected Nickerson 2008). As Bullock, Green, and Ha (2010) have Relationships demonstrated, it is almost impossible to identify a causal mediator, even if a direct measure of discussion were Heterogeneous available in our data set. Hence, we must acknowledge partisan household the possibility of alternative explanations for any indirect (pretreatment) mobilization effects identified in our experiment that (-) conflict avoidance; (+) friendly competition; may occur in the (near) absence of discussion. Houseseeking like-minded seeking new information from non-like-minded hold members might simply have listened in on the discussion partner discussion partner experimental subject’s telephone conversation with the Partisan intensity party volunteer and adjusted their behavior accordingly. Household of campaign Discussion Alternatively, they might have felt social pressure after + + member voting intervention (unobserved) (outcome) observing the experimental subject going out to vote, or (randomized) they might have considered the lower costs of voting by going to the polls together (Nickerson 2008, 55). Most of these alternative mechanisms, however, differ in their Green 2012, 333–36) by randomly assigning one of two observable implications. In the latter two scenarios in campaign messages of varying partisan intensity, hence particular, there would be little reason to expect different further manipulating the potential for discussion between levels of indirect mobilization depending on whether the household members. Our argument is that overtly parhousehold is homogeneous or heterogeneous, or dependtisan messages exacerbate partisan sentiment compared ing on the partisan intensity of the campaign message. to more neutral messages, making preexisting partisan A third limitation concerns the question of generdisagreement within the household more salient. In part, alizability (Fowler et al. 2011, 466). The mobilization this might be because overtly partisan messages function messages used in our study were communicated by a local as a social identity cue, rallying support for the team Labour Party. Work by Jost and colleagues suggests that and pitting party supporters against each other (Green, individuals who hold conservative ideologies are more Palmquist, and Schickler 2002; Rogers, Fox, and Gerber likely to avoid situations that arouse cognitive dissonance 2014). An overview of our study variables and hypotheand less likely to be open to new experiences and argusized relationships is shown in Figure 1. ments than liberals (Carney et al. 2008; Garrett 2009; Jost While our study thus allows for a unique contribuet al. 2003; Nam, Jost, and Van Bavel 2013). Consistent tion to both the literatures on social network composition with this idea, Mutz (2006, 33) shows that conservatives and on voter mobilization, it is important to point out are less likely to be embedded in cross-cutting political its limitations. First, although our study design allows discussion networks than liberals. Consequently, we us to identify the causal impact of campaign contact on cannot exclude the possibility that conservatives might intrahousehold mobilization within both heterogeneous be less responsive to a Labour campaign message than and homogeneous partisan households, we prime, but Labour supporters to a Conservative campaign message. do not randomly assign, partisanship within households. Moreover, within the household, conservatives might Consequently, we cannot exclude the possibility that combe less likely to share opinion-challenging messages positional partisan differences might be confounded by with Labour supporters than vice versa. Hence, we a number of (un)observable background attributes. Inwould expect indirect mobilization effects to be more dividuals in heterogeneous partisan households might, pronounced if contact was made by a Conservative for instance, be—ex ante—less likely to participate in campaign and mediated by a subject who has liberal politics than individuals who live in homogeneous parleanings than the other way around. Our focus on tisan households. While we therefore cannot claim that campaign contact also means that our findings might not the conditional effects that we uncover are causal in nageneralize to other settings in which political discussion ture, we utilize information about turnout behavior in is likely to occur within the household, such as when multiple previous elections to estimate the effects net of watching news programs together. ex ante turnout differences between heterogeneous and Finally, although households are important locations homogeneous households. of social interaction, they are also unique because personal Second, we share the limitation present in the relationships between household members are usually experimental and observational literature that we cannot stronger and may therefore withstand political disagreeconclusively show that contagion within households is ment more easily than ties that exist between colleagues or a result of actual discussion (Fowler et al. 2011, 463; acquaintances. This study therefore seeks to qualify, but 294 FLORIAN FOOS AND ELINE A. DE ROOIJ not directly contradict, the empirical findings of previous constituency.3 The party aims to contact every voter living work such as Mutz’s (2006), whose data on cross-cutting in the parliamentary constituency to find out which party exposure include weak ties such as those formed at the he or she supports. It is common practice for constituents workplace or in places of worship. to volunteer their voting intentions to party canvassers Notwithstanding these limitations, the household is in the UK. The database covered around 38% of regisarguably the most important personal discussion network tered voters. Although the accuracy of the pretreatment and hugely influential in our turnout decisions. By exparty support measure obtained through local Labour amining the conditions under which campaign-induced Party canvassing is difficult to assess, a comparison to partisan disagreement between household members city council election results suggests that neither Labour translates into turnout, we aim to shed some light on insupporters nor supporters of rival parties are overrepretrahousehold mobilization and to more closely integrate, sented in the data (supporting information, Table A7).4 In and make valuable contributions to, the literatures on addition to party preference, the database also provided social network composition and campaign mobilization. information on gender, year of birth, whether individuals were registered as postal voters, the electoral ward in which they reside, and individuals’ validated turnout hisResearch Design tories (as available from public records). This extensive set of covariates allows us to check whether our findings We rely on a previously conducted randomized campaign are robust to covariate adjustment. experiment to identify mobilization effects between household members in two-voter households and to gauge the extent to which they are conditioned by Experimental Assignment both the degree of heterogeneity of party preferences within the household and the partisan intensity of To fulfill the noninterference assumption (Stable Unit the campaign messages. We conducted this campaign Treatment Value Assumption, short SUTVA), we ranexperiment in several electoral wards in the British city of domly selected one individual per household included Birmingham in the context of the West Midlands Police in the electoral database to be part of the original exand Crime Commissioner (PCC) election, which was periment. The household members of these experimenheld on November 15, 2012.2 This campaign experiment tal subjects were not assigned to receive any treatment. combined two design innovations that make it especially For random assignment, a total of 13,065 households well suited for our current study. with landline numbers (of which 5,190 were two-voter First, the experiment was conducted in cooperation households) were first stratified into three blocks based with a local Labour Party. Existing field experimental on the latest recorded party preference of the experimental studies of interpersonal influence within households have subject: “Labour” supporter, “rival party” supporter (i.e., so far exclusively relied on identifying spillover effects Conservative, Liberal Democrat, Green, British National of nonpartisan get-out-the-vote (GOTV) interventions. Party, Respect, UK Independence Party, or those who othWhen trying to assess the impact of an election camerwise indicated explicitly opposing Labour), and those paign on indirect mobilization within households, workwho were “unattached” to a party or whose latest recorded ing from within a party’s campaign mimics reality more party preference was unknown (i.e., “won’t say,” “don’t closely given that most mobilization efforts tend to be know,” “non-voter,” or “missing”). The choice of these partisan in nature. As we discuss below, phone bank vol3 unteers were instructed to use two different messages: one We were provided with access to a fully anonymized version of the database that included registered voters with available landline or of low and one of high partisan intensity. This difference mobile phone numbers, gained informed consent from the Conallows us to identify how varying the potential for partisan stituency Labour Party to conduct this experiment, and received disagreement affects intrahousehold mobilization. approval from the internal review boards of Oxford and Simon Second, we had access to detailed information on Fraser Universities. 4 the pretreatment party preferences of both household For the purpose of a separate field experiment conducted elsemembers. This information was included in the Labour where Foos (2015, chap. 3) validated the party support measure used in this study with a comparable measure obtained from indeParty’s extensive targeting database and is based on prior pendently conducted phone interviews on a sample of rival party canvassing efforts conducted by party volunteers in the supporters. Interviews with subjects classified as rival party supporters showed that 72% of respondents in the untreated control group were correctly classified. Only 9% classified as rival party 2 For a detailed description of the data, see the Replication Data supporters by canvassers told interviewers that they supported the Codebook, available at http://dx.doi.org/10.7910/DVN/ZFLG25. Labour Party or a Labour candidate instead. PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 295 FIGURE 2 Experimental Assignment = = = = = = = = = = = = = = three groups was motivated by the design of the original randomly assigned to either one of two treatment groups experiment, which aimed to distinguish Labour from rior to the control group. The control group did not receive val party supporters. Homogeneous partisan households any form of contact from the campaign. An overview of are accordingly defined as households that include either the random assignment within partisan blocks for twotwo Labour voters or two rival party supporters, heterovoter households is shown in Figure 2. geneous partisan households as consisting of one Labour We followed the standard randomization inference and one rival party supporter, and unattached households procedure (Gerber and Green 2012) to check whether any as consisting of at least one unattached voter.5 Next, for covariate imbalances in our sample of two-voter housetwo-voter households, experimental subjects (and their holds were larger than expected given random sampling household members) were, within each partisan group, variability. The resulting p-value of .50 indicates that we cannot reject the sharp null hypothesis that the covariates 5 Among rival party supporters in our sample, 49% indicated suptaken together are not systematically related to treatment port for the Conservative Party, less than 5% explicitly mentioned the name of another party or of an independent candidate, and 47% assignment of the household member. Detailed balance stated to be “against” Labour. The latter are most likely Conservatables and figures can be found in the supporting infortive supporters or, alternatively, supporters of another right-wing mation (Table A1 and Figure A1). Furthermore, as we party. This means that we cannot exclude the possibility that some would expect given that turnout data are collected from homogeneous partisan households consisting of two rival party supporters are actually heterogeneous. In the analysis section, we the public register, missing outcome data for any housedemonstrate that our findings remain robust when distinguishing hold member are neither systematically related to treathomogeneous Labour households from households consisting of ment assignment, nor are they a function of treatment two rival party supporters. assignment, pretreatment covariates, and the interaction 296 FLORIAN FOOS AND ELINE A. DE ROOIJ between treatment assignment and covariates (see Figure any reasons for why contact had failed (i.e., answering A2 in the supporting information). machine, no answer, hang-up), and the number of In order to estimate the within-household spillover callbacks made. After the election, the local Labour Party effects resulting from the two treatments, we compare provided us with an updated version of the (anonymized) turnout among unassigned household members living database, which included validated turnout data for the with subjects who were assigned to one of the two treatPCC election from the marked electoral register for both ment conditions to turnout among unassigned housethe experimental subjects and their household members. hold members living with subjects assigned to the control Before the analysis, all two-voter households with condition. In doing so, our spillover model relies on the missing outcome data were removed from the sample. common assumption that effects spill over within houseWe took great care in the parallel administration holds, but not between neighboring households. This asof both treatments. The total contact rate, defined as sumption has been tested and confirmed in previous field the percentage of individuals in the respective treatment experiments (Sinclair, McConnell, and Green 2012), and groups who answered the phone when called by volunwe have no reason to believe that it was violated in the teers, among two-voter households, is 45% for both treatcontext of our experiment. ment groups (see Table 1). Thus, we failed to administer the treatment to 55% of those assigned to treatment. This Treatments rate of noncompliance is similar to the rate recorded in other GOTV studies in the UK (John and Brannan 2008). Experimental subjects allocated to the two treatment groups were called by telephone by Labour Party volunteers in the week leading up to the election (November Analysis and Results 10–15) and were encouraged to vote in the PCC election on November 15, 2012. Phone bank volunteers were enWe start by examining to what extent low and high partirely blind to subjects’ pretreatment party preferences, tisan intensity phone messages mobilized experimental and they were instructed to ask to speak to the person in subjects and their household members. Table 1 shows the experimental group before delivering the messages. turnout rates for experimental subjects and their houseIn case the person was unavailable, volunteers were inhold members assigned to the control group and to the structed to politely finish the conversation without retwo treatment groups. These rates are shown by experivealing the intent of the call. mental block, meaning conditional on the pretreatment In formulating the message scripts, we worked closely party preference of the assigned subject. with the local Labour Party in order to ensure the mesThe final column shows that turnout rates among sages paralleled normal campaign efforts. Both messages subjects assigned to the control group and their unasprovided practical information to subjects on the elecsigned household members are almost identical at around tion date and their local polling station, and encouraged 19–20%. We would expect this, given the stratified ranthem to vote for the same candidate, providing infordom sampling procedure used to allocate one subject per mation about the candidate’s background and his most household to the experimental sample. Among subjects important policy stance (opposition to cuts in police assigned to receive either a low or high partisan intensity numbers). However, the content of the scripts varied drasphone call, turnout reaches 22–23%. Turnout among the tically across the two treatment groups. The “high parhousehold members of these subjects is slightly higher at tisan intensity” treatment was formulated in a strongly 24%. Using randomization inference (Aronow and Samii partisan tone, explicitly mentioning the Labour Party 2012; Gerber and Green 2012) to estimate p-values and and policies multiple times, while taking an antagonisconfidence intervals, we find that, for all partisan groups tic stance toward the main rival party, the Conservative combined, both the direct and the indirect intent-to-treat Party, by pointing out the Labour candidate’s opposition (ITT) effects are statistically different from the respective to the “Tory cuts” in police numbers. In contrast, the “low control groups, with p < .05. The results suggest that partisan intensity” treatment message avoided all stateboth the low and high partisan intensity campaign ments about party competition and mentioned neither messages resulted in intrahousehold mobilization. As the candidate’s party affiliation nor the rival party. Both we would expect, Table A2 in the supporting informamessage scripts are displayed in Figures A3 and A4 in the tion shows that the estimates are robust to covariate supporting information. adjustment. Campaign volunteers were asked to fill in a form on We next test whether the rates of intrahousehold mowhether contact with the targeted individual was made, bilization differ significantly between homogeneous and PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 297 TABLE 1 Turnout and Contact Percentages in Treatment and Control Groups Partisanship of Assigned Subject Labour Rival Party Unattached Combined Assigned Subjects Turnout Control 25.0 24.1 11.6 19.2 Turnout Low Partisan Intensity 27.2 25.4 17.3 22.6 Turnout High Partisan Intensity 29.6 26.2 14.7 22.4 ITT Low Partisan vs. Control 2.2 1.4 5.8∗∗ 3.4∗ [−3.7, 8.3] [−4.0, 6.7] [2.1, 9.4] [0.6, 6.2] ITT High Partisan vs. Control 4.6 2.1 3.1 3.2∗ [−1.5, 10.7] [−3.1, 7.3] [−0.4, 6.9] [0.5, 5.9] Unassigned Household Members Turnout Control 24.7 24.7 12.7 19.8 Turnout Low Partisan Intensity 28.3 27.7 19.3 24.4 Turnout High Partisan Intensity 28.2 28.7 16.6 23.6 ITT Low Partisan vs. Control 3.5 3.0 6.6∗∗∗ 4.6∗∗ [−2.4, 9.6] [−2.5, 8.3] [2.7, 10.5] [1.8, 7.5] ITT High Partisan vs. Control 3.4 3.9 3.9∗ 3.8∗∗ [−2.5, 9.4] [−1.3, 9.2] [0.2, 7.9] [1.0, 6.6] Contact Rates Low Partisan Intensity 53.6 46.0 39.7 45.4 Contact Rates High Partisan Intensity 47.3 50.4 38.8 44.8 N 1,273 1,635 2,022 4,930 Note: Two-tailed tests based on randomization inference. Table excludes subjects in households with missing turnout data. Combined column includes inverse probability weights accounting for different probabilities of assignment to experimental conditions between blocks. ∗∗∗ p < .001, ∗∗ p < .01, ∗ p < .05. heterogeneous partisan households. Figure 3a displays interaction, this relationship remains unchanged regardthe changes in the predicted probabilities of turnout for less of whether we adjust for pretreatment individual unassigned subjects as a function of the treatment asand compositional differences in age, gender, place of signment of their household member, conditional on the residence, registration type (postal voter or not), and partisan composition of the household. The predicted turnout history (see Models I–III in Table A3 in the probabilities are based on a logistic regression of turnout supporting information). Moreover, Figure 3b shows on treatment assignment of the household member, the that our results are robust to further breaking down partisan composition of the household, the interaction the homogeneous partisan household category into between treatment assignment and household partisan homogeneous Labour households and homogeneous composition, pretreatment covariates, and the interacrival party households, and the unattached category into tion between pretreatment covariates and treatment ashouseholds that include only one unattached subject and signment (see Equation 14 and Table A3, Model III, in households in which both members are unattached.6 the supporting information). What do these results tell us? We find no evidence to The key result is that intrahousehold mobilization support the hypothesis that individuals are more likely to effects are stronger, not weaker, in households where mobilize a copartisan than a household member who individuals support different parties. While indirect supports another party as a result of being contacted mobilization effects are around 3.5 percentage points by the Labour campaign. Although we find that camin homogeneous partisan households and around 7 paign messages spill over in homogeneous households percentage points in households with at least one unattached voter, spillover effects in heterogeneous par6 tisan households peak at around 15 percentage points. The marginal changes in predicted probabilities in Figure 3b are based on estimates from Table A4, Model III, in the supporting Although we are dealing with a treatment-by-covariate information. 298 FLORIAN FOOS AND ELINE A. DE ROOIJ FIGURE 3 ITTs on Turnout of Unassigned Household Members, Conditional on Household Party Preferences Note: Bars indicate 95% confidence intervals. consisting of two Labour Party supporters or that inFigure 4a shows that rival party supporters mobiclude an unattached voter, mobilization effects in these lize Labour supporters if they are contacted by a Labour households are no greater than in heterogeneous partisan campaign. Figure 4b also shows that Labour Party suphouseholds. In contrast, we find greater spillover effects porters mobilize rival party supporters as a consequence when partisans support different parties. of Labour campaign contact. While the effects for heterogeneous households are noisy due to small sample sizes, they are substantially large and seem even more Interaction Dynamics between Household pronounced than indirect mobilization effects between Members two Labour supporters (although the difference between the two effects is not statistically significant). These reThe finding that campaign spillover effects appear more sults thus suggest that party supporters discuss campaign pronounced in heterogeneous than in homogeneous parmessages with their household members even if a party tisan households is difficult to reconcile with a theoretical they oppose initiates contact. conception of household members as conflict avoiders. In what follows, we try to further disentangle the partisan mobilization dynamics within the household by examinImplicit Mediation Analysis: Does ing whether spillover is more or less likely depending on Increasing Partisan Intensity Affect the partisan preferences of both household members. Spillover? Figure 4 displays the marginal changes in predicted probabilities resulting from a logistic regression The previous results demonstrate that within-household of turnout on assignment to either of the two treatments, mobilization plays an important role in partisan electhe partisanship of the experimental subject, the partition campaigns and that, in contrast to what a conflictsanship of his or her household member, and the twoavoidance mechanism would suggest, heterogeneous and three-way interactions between the treatment and the partisan households are conducive to indirect campaign party preferences of both household members. The model mobilization. We implicitly test the effects of raising the also includes pretreatment covariates and interactions belevel of partisan discussion within the household by extween covariates and the treatment (for estimates, see ploiting the design feature that the partisan intensity Table A5, Models I–III, in the supporting information). of the campaign message was randomly assigned. MesFigure 4a displays the indirect mobilization effects if sages with stronger partisan language should exacerbate the assigned household member supports a rival party, partisan sentiment compared to more neutral messages, 4b if the assigned household member supports Labour, making preexisting partisan disagreement within the and 4c if the assigned subject is unattached to any party. household more salient, and resulting in a higher probaThe rival, Labour, and unattached labels listed on the bility of partisan discussions in the household. x-axis refer to the party preference of the unassigned Table 2 and Figure 5 show the logistic regression household member. estimates and the corresponding predicted conditional PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 299 FIGURE 4 Interactions between Treatment Assignment and Party Preferences of Assigned and Unassigned Subject Note: Bars indicate 95% confidence intervals. TABLE 2 Logistic Regression Results: CACE of High vs. Low Partisan Intensity Call on Turnout of Unassigned Subjects, Conditional on Household Party Preferences I II III IV V Phone Call −.109 −.154 1.337 1.407 1.824 (.142) (.165) (.687) (.785) (1.364) Heterogeneous Partisan Reference Group Homogeneous Partisan −.202 −.687 .688 .256 .406 (.325) (.383) (.581) (.648) (.656) Unattached −.571 −.648 .399 .363 .534 (.364) (.431) (.604) (.680) (.690) Heterogeneous × Call Reference Group Homogeneous × Call −1.437∗ −1.578 −1.823∗ (.716) (.818) (.835) Unattached × Call −1.606∗ −1.713∗ −1.976∗ (.719) (.824) (.841) Covariates No Yes No Yes Yes Covariates × Call No No No No Yes N 965 Note: Results based on two-tailed tests, compliers only, with errors in parentheses. Covariates are turnout in seven previous elections, postal voter, gender, age, and electoral ward. Dummies for experimental blocks are included. ∗∗∗ p < .001, ∗∗ p < .01, ∗ p < .05. complier average causal effects (CACE) comparing outwhether they were about to receive the high or the low comes directly for subjects whose household members partisan intensity call, and compliance is defined as living answered the high-intensity phone call to subjects whose in a household where the assigned subject would answer household members answered the low partisan intensity the phone when called, the share of compliers in the phone call. Since subjects included in the original high and the low partisan intensity call groups should, experiment did not know before answering the phone in expectation, be identical (Gerber et al. 2010, 302–5). 300 FLORIAN FOOS AND ELINE A. DE ROOIJ FIGURE 5 CACE of High vs. Low Partisan Message for Unassigned Household Members, Conditional on Household Party Preferences Note: Bars indicate 95% confidence intervals. Following Gerber et al. (2010), this “perfect blindness” which both members support rival parties (based on assumption can be assessed empirically. If, as in our case, Model V in Table A6 in the supporting information). compliance at 45.4% and 44.8% does not vary as a funcIf partisan discussion is a plausible mediator linking tion of treatment assignment, we can compare turnout campaign contact to intrahousehold mobilization, rates for household members of compliers directly to raising the partisan intensity of the message should each other (for the formal assumption and the correincrease, not decrease, indirect mobilization effects in sponding CACE spillover estimator, see Equations 23 and heterogeneous households. 24 in the supporting information). In line with this expectation, we find that indirect Across partisan blocks, the results in columns 1 and mobilization effects in heterogeneous households are 2 of Table 2 confirm that receiving a high-intensity par35 percentage points larger among individuals living with tisan message did not result in significantly lower levels someone who received a call priming his or her partisan of intrahousehold mobilization than receiving a message identity than among subjects whose household members without partisan content. The relative effectiveness of the received a call of low partisan intensity. This substantial two messages again seems to differ according to the partibut noisy CACE estimate might be a function of the low san makeup of the household. Columns 3, 4, and 5 display baseline turnout rate, and the relatively small subsample the results of models that include interactions between the size when comparing household members of compliers partisan composition of the household and the partisan directly to each other. It is hence likely that we are dealing content of the message. While column 4 shows the estiwith an upper bound on the true effect size. Although mates from a model unadjusted for compositional differwe therefore caution against overinterpreting the size of ences between homogeneous and heterogeneous housethis effect, we clearly do not find any evidence that would holds, column 5 shows the estimates for a model that support the notion that priming household members’ includes interactions between pretreatment covariates party preferences would decrease indirect mobilization and the treatment. The statistically significant interaction between subjects who are predisposed to disagree on terms indicate that increasing the potential for partisan which party to support. disagreement had a significantly more positive effect on turnout for household members living in heterogeneous partisan households than for household members living Discussion and Conclusion in homogeneous and unattached partisan households. Figure 5 displays the marginal changes in predicted An increasing number of political scientists view turnout probabilities of turnout for unassigned subjects based on as a social activity, and few doubt that family, friends, and Model V in Table 2. Figure 5b shows a robustness check neighbors influence decisions about whether or not to that compares heterogeneous households separately to participate politically and whom to support (Huckfeldt homogeneous Labour households and households in and Sprague 1995; Klofstad 2011; Mutz 2006; Rolfe 2012; PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 301 Sinclair 2012; Zuckerman 2005; Zuckerman, Dasović, and increase indirect mobilization in heterogeneous partisan Fitzgerald 2007). The debate within the literature now households to a greater extent than campaign messages centers on whether political disagreement between netof a lower partisan intensity, our findings are suggeswork members is harmful or beneficial for political partive of household members talking about the message, ticipation. In this article, we contribute to this debate rather than solely acting as a consequence of observing by focusing on a specific, but very common, instance their household member’s act of voting. Our assertion of political disagreement that occurs among household that members in two-voter households do not refrain members in the context of a partisan election campaign from partisan discussions in order to avoid conflict is also and that affects the decision to vote. Our study is unique in line with research that shows how couples who live because it combines the exogenous assignment of partitogether become increasingly similar over time in their san telephone calls with an unusually rich data set, made policy and party preferences (Stoker and Jennings 2005; available by a local branch of the UK Labour Party, which Zuckerman, Fitzgerald, and Dasović 2005). In order to includes the party preferences and validated turnout of adopt each other’s preferences, surely couples must be all household members. Our results demonstrate that talking. within-household mobilization effects are considerably In addition to the literature on disagreement in larger as a result of campaign contact in households that social networks, this article also makes an important consist of members who disagree in their party preference contribution to the literature on partisan election than in households in which members share a similar campaigns. While the existence of an indirect mobilizaparty preference. tion effect within households has been demonstrated These results provide an important qualification in nonpartisan GOTV experiments (Nickerson 2008; regarding the role of disagreement in discouraging Sinclair, McConnell, and Green 2012), we show that electoral participation in the context of the household. this effect is also present in partisan campaigns. As Rather than supporting a demobilizing impact of partisan Nickerson, Friedrichs, and King (2006) caution, partisan disagreement, our results offer new experimental eviand nonpartisan campaigns differ in multiple ways, and dence for the participatory benefits of disagreement that intrahousehold partisan dynamics are widely expected to advance on observational work carried out by Huckfeldt, influence the effectiveness of indirect campaign mobilizaJohnson, and Sprague (2002, 2004), Nir (2005), and tion (Huckfeldt, Johnson, and Sprague 2004). It remains McClurg (2006). Importantly, our study is the first to to be seen to what extent indirect mobilization effects causally identify the indirect effects of partisan messages between partisans are replicated, or even amplified, when on turnout in heterogeneous and homogeneous personal conservative parties initiate contact with supporters of networks. We recognize that our findings may be limited more liberal rival parties, as liberals are said to be more to the household or other equally intimate settings, and open to opinion-challenging information than consermay not apply to social networks in which ties are weaker vatives (Jost et al. 2003; Nam, Jost, and Van Bavel 2013). and political discussion is more easily avoidable. We have The relatively large spillover effects reported in this suggested that because households tend to be high-trust article (cf. Nickerson 2008; Sinclair, McConnell, and environments, the risk of potential fallout from political Green 2012) might partly be due to the low-information, discussion is likely lower than in environments such as low-saliency nature of the election. Alternatively, they the workplace. If social costs are lower, individuals can might be a function of the research design. In a recent pamore easily enjoy the entertainment value and political per, Mann and Sinclair (2014) argue that spillover effects information that discussions about partisan politics might differ depending on whether the lowor the highprovide. As such, our findings qualify, but do not directly propensity voter in a household is targeted. As a funccontradict, Mutz’s (2006) earlier work on turnout in tion of our research design, the lower-propensity voter heterogeneous networks that consist of both strong and had an equal probability of being assigned to contact as weak ties. the higher-propensity voter. In contrast, in Nickerson’s How confident are we that this intrahousehold mo(2008) experimental design, the higher-propensity voter bilization is a result of discussion rather than some other might have been more likely to self-select into opening the form of social influence? Social influence can occur in door to the canvasser. Even if our results are on the high a number of imaginable ways. For instance, observing end of the scale, they nonetheless strongly suggest that a household member voting might function as a simple partisan election campaigns can have substantial indirect reminder to vote, or it may provide social pressure to mobilization effects that should be taken into account comply with a perceived norm of voting. By showing eviwhen assessing the cost-effectiveness of different voter dence that campaign messages of a highly partisan nature mobilization methods. 302 FLORIAN FOOS AND ELINE A. DE ROOIJ Our results thus raise intriguing strategic questions Aronow, Peter M., and Cyrus Samii. 2012. “The R Packfor partisan campaigns. Partisans appear to be more reage for Performing Randomization-Based Inference for Experiments. R Package. Version 0.9.” http://CRAN.Rsponsive to campaign messages from rival parties if these project.org/package=ri: Available on CRAN. messages are mediated by household members than if Bélanger, Paul, and Munroe Eagles. 2007. “Partisan Crossthey are transmitted directly by party canvassers. Our Pressure and Voter Turnout: The Influence of Micro and results suggest that campaigns can indirectly affect the Macro Environments.” Social Science Quarterly 88(3): 850– turnout behavior of a rival party’s supporters via sympa67. thetic household members. Future research should hence Bello, Jason, and Meredith Rolfe. 2014. “Is Influence Mightinvestigate whether a campaign strategy that targets hetier Than Selection? Forging Agreement in Political Discussion Networks during a Campaign.” Social Networks 36(1): erogeneous partisan households is prone to being strate134–46. gically ineffective by mobilizing votes for opposing parBerelson, Bernard R., Paul F. Lazarsfeld, and Willian N. McPhee. ties, or if such a strategy might be beneficial to parties if 1954. Voting: A Study of Opinion Formation in a Presidential contacted supporters persuade their household members Campaign. Chicago: University of Chicago Press. to switch their vote. Bond, Robert M., Christopher J. Fariss, Jason J. Jones, Are these results specific to the UK context? We sinAdam D. I. Kramer, Cameron Marlow, Jaime Settle, and cerely doubt so. Compared to the United States, the partiJames H. Fowler. 2012. “A 61-Million-Person Experiment in Social Influence and Political Mobilization.” Nature san context in the UK is less polarized, and fewer individ489(7415): 295–98. uals hold strong party preferences (Ford 2014; Pattie and Bullock, John G., Donald P. Green, and Shang E. Ha. 2010. Johnston 2010). As such, one might have expected no dif“Yes, But What’s the Mechanism? (Don’t Expect an Easy ference in intrahousehold mobilization between houseAnswer).” Journal of Personality and Social Psychology holds in which members support the same party and those 98(4): 550–58. in which they do not. Our finding that spillover effects Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes. 1960. The American Voter. New York: are larger in heterogeneous than in homogeneous partisan Wiley. households as a result of campaign contact, and larger still Carney, Dana R., John T. Jost, Samuel D. Gosling, and Jeff when the partisan intensity of the message is increased, Potter. 2008. “The Secret Lives of Liberals and Conserwould lead us to expect an even more pronounced impact vatives: Personality Profiles, Interaction Styles, and the of partisan disagreement within households on electoral Things They Leave Behind.” Political Psychology 29(6): 807– participation in the U.S. context. 40. Political discussion within personal networks is Conover, Pamela, Don Searing, and Ivor Crewe. 2002. “The Deliberative Potential of Political Discussion.” British Journal widely believed to go hand in hand with increasing levels of Political Science 32(1): 21–62. of political knowledge and tolerance of competing politCutts, David, and Edward Fieldhouse. 2009. “What Small Spaical viewpoints, ultimately benefiting the functioning of tial Scales Are Relevant as Electoral Contexts for Individual democracies (Conover, Searing, and Crewe 2002; Mutz Voters? The Importance of the Household on Turnout at the 2002b, 2006). However, scholars have doubted that these 2001 General Election.” American Journal of Political Science 53(3): 726–39. civic characteristics of politically heterogeneous networks Downs, Antony. 1957. An Economic Theory of Democracy. New also facilitate political participation (McClurg 2003; Mutz York: Harper and Row. 2002a, 2006). Our article shows that political disengageFieldhouse, Edward, and David Cutts. 2012. “The Companion ment is not necessarily the flip side of the political diverEffect: Household and Local Context and the Turnout of sity and tolerance fostered by sharing the same household Young People.” Journal of Politics 74(3): 856–69. with individuals who hold different political views. A betFitzgerald, Jennifer, and Amber K. Curtis. 2012. “Partisan Dister understanding of how partisan political disagreement cord in the Family and Political Engagement: A Comparative within the family can motivate voter participation proBehavioral Analysis.” Journal of Politics 74(1): 129–41. vides valuable insights into the general conditions under Foos, Florian. 2015. Bringing the Party Back In: Mobilization and Persuasion in Constituency Election Campaigns. D.Phil which political discussion in personal networks occurs dissertation, University of Oxford. and under which it generates civic benefits. Ford, Robert. 2014. “In America Polarization Is a Problem. In Britain It Could Be a Solution.” http://www.washingtonpost. com/blogs/monkey-cage/wp/2014/02/20/in-americaReferences polarization-is-a-problem-in-britain-it-could-be-asolution. Arceneaux, Kevin, Alan S. Gerber, and Donald P. Green. 2006. Fowler, James, Michael Heaney, David Nickerson, John Padgett, “Comparing Experimental and Matching Methods Using a and Betsy Sinclair. 2011. “Causality in Political Networks.” Large-Scale Voter Mobilization Experiment.” Political AnalAmerican Politics Research 39(2): 437–80. ysis 14(1): 37–62. PARTISAN DISAGREEMENT AND ELECTORAL MOBILIZATION IN INTIMATE NETWORKS 303 Garrett, R. Kelly. 2009. “Politically Motivated Reinforcement Lake, Ronald La Due, and Robert Huckfeldt. 1998. “Social CapSeeking: Reframing the Selective Exposure Debate.” Journal ital, Social Networks, and Political Participation.” Political of Communication 59(4): 676–99. Psychology 19(3): 567–83. Gerber, Alan S., and Donald P. Green. 2012. Field Experiments: Lazarsfeld, Paul F., Bernard Berelson and Hazel Gaudet. 1948. Design, Analysis and Interpretation. New York: Norton. The People’s Choice: How the Voter Makes Up His Mind in Gerber, Alan S., Donald P. Green, Edward H. Kaplan, and Holger a Presidential Campaign. New York: Columbia University L. Kern. 2010. “Baseline, Placebo, and Treatment: Efficient Press. Estimation for Three-Group Experiments.” Political Analysis Mann, Christopher B., and Betsy Sinclair. 2014. “Two Votes for 18(3): 297–315. the Price of One: Targeted Messages and Spillover Effects Glaser, William A. 1959. “The Family and Voting Turnout.” in Early Voting.” Paper presented at the Annual Meeting Public Opinion Quarterly 23(4): 563–70. of the Southern Political Science Association, New Orleans, Louisiana. http://citation.allacademic.com. Green, Donald P., Bradley Palmquist, and Eric Schickler. 2002. Partisan Hearts and Minds: Political Parties and the Social Manski, Charles. 1993. “Identification of Exogenous Social EfIdentities of Voters. New Haven, CT: Yale University Press. fects: The Reflection Problem.” Review of Economic Studies 60(3): 531–42. Huckfeldt, Robert R., Paul E. Johnson, and John D. Sprague. 2002. “Political Environments, Political Dynamics, and the McClurg, Scott D. 2003. “Social Networks and Political ParSurvival of Disagreement.” Journal of Politics 64(1): 1–21. ticipation: The Role of Social Interaction in Explaining Political Participation.” Political Research Quarterly 56(4): Huckfeldt, Robert R., Paul E. Johnson, and John D. Sprague. 448–64. 2004. Political Disagreement: The Survival of Diverse Opinions within Communication Networks. Cambridge: Cambridge McClurg, Scott D. 2006. “The Electoral Relevance of Political University Press. Talk: Examining Disagreement and Expertise Effects in Social Networks on Political Participation.” American Journal Huckfeldt, Robert, and Jeanette Morehouse Mendez. 2008. of Political Science 50(3): 737–54. “Moths, Flames, and Political Engagement: Managing Disagreement within Communication Networks.” Journal of Morey, Alyssa C., William P. Eveland, and Myiah J. Hutchens. Politics 70(1): 83–96. 2012. “The ‘Who’ Matters: Types of Interpersonal Relationships and Avoidance of Political Disagreement.” Political Huckfeldt, Robert R., and John D. Sprague. 1987. “Networks in Communication 29(1): 86–103. Context: The Social Flow of Political Information.” American Political Science Review 81(4): 1197–216. Mutz, Diana C. 2002a. “The Consequences of Cross-Cutting Networks for Political Participation.” American Journal of Huckfeldt, Robert R., and John D. Sprague. 1991. “Discussant Political Science 46(3): 838–55. Effects on Vote Choice: Intimacy, Structure, and Interdependence.” Journal of Politics 53(1): 122–58. Mutz, Diana C. 2002b. “Cross-Cutting Social Networks: Testing Democratic Theory in Practice.” American Political Science Huckfeldt, Robert R., and John D. Sprague. 1995. Citizens, PolReview 96(1): 111–26. itics and Social Communication: Information and Influence in an Election Campaign. Cambridge: Cambridge University Mutz, Diana C. 2006. Hearing the Other Side: Deliberative versus Press. Participatory Democracy. Cambridge: Cambridge University Press. John, Peter, and Tessa Brannan. 2008. “How Different Are Telephoning and Canvassing? Results from a ‘Get Out the Vote’ Nam, H. Hannah, John T. Jost, and Jay J. Van Bavel. 2013. “Not Field Experiment in the British 2005 General Election.” for All the Tea in China! Political Ideology and the AvoidBritish Journal of Political Science 38(3): 565–74. ance of Dissonance-Arousing Situations.” PloS ONE 8(4): e59837. Johnston, Ron, Kelvyn Jones, Carol Propper, Rebecca Sarker, Simon Burgess, and Anne Bolster. 2005. “A Missing Level in Nickerson, David W. 2008. “Is Voting Contagious? Evidence the Analyses of British Voting Behaviour: The Household as from Two Field Experiments.” American Political Science ReContext as Shown by Analyses of a 1992–1997 Longitudinal view 102(1): 49–57. Survey.” Electoral Studies 24(2): 201–25. Nickerson, David W., Ryan D. Friedrichs, and David C. King. Jost, John T., Jack Glaser, Arie W. Kruglanski, and Frank J. 2006. “Partisan Mobilization Campaigns in the Field: Results Sulloway. 2003. “Political Conservatism as Motivated Social from a Statewide Turnout Experiment in Michigan.” Political Cognition.” Psychological Bulletin 129(3): 339–75. Research Quarterly 59(1): 85–97. Klofstad, Casey A. 2007. “Talk Leads to Recruitment: How DisNir, Lilach. 2005. “Ambivalent Social Networks and Their Concussions about Politics and Current Events Increase Civic sequences for Participation.” International Journal of Public Participation.” Political Research Quarterly 60(2): 180–91. Opinion Research 17(4): 422–42. Klofstad, Casey A. 2011. Civic Talk: Peers, Politics, and the Future Nir, Lilach. 2011. “Disagreement and Opposition in Social Netof Democracy. Philadelphia: Temple University Press. works: Does Disagreement Discourage Turnout?” Political Studies 59(3): 674–92. Klofstad, Casey A. 2015. “Exposure to Political Discussion in College Is Associated with Higher Rates of Political ParticiOsborn, Tracy, and Jeanette Morehouse Mendez. 2011. “Two pation over Time.” Political Communication 32(2): 292–309. Become One? Spouses and Agreement in Political Opinions.” American Politics Research 39(5): 783–803. Klofstad, Casey A., Anand E. Sokhey, and Scott D. McClurg. 2013. “Disagreeing about Disagreement: How Conflict in Pattie, Charles J., and Ron J. Johnston. 2009. “Conversation, Social Networks Affects Political Behavior.” American JourDisagreement and Political Participation.” Political Behavior nal of Political Science 57(1): 120–34. 31(2): 261–85. 304 FLORIAN FOOS AND ELINE A. DE ROOIJ Pattie, Charles J., and Ron J. Johnston. 2010. “Voting and Identity.” In The Oxford Handbook of British Politics, ed. Matthew Supporting Information Flinders, Andrew Gamble, Colin Hay, and Michael Kenny. Oxford: Oxford University Press, 461–83. Additional Supporting Information may be found in the Rogers, Todd, Craig R. Fox, and Alan S. Gerber. 2014. online version of this article at the publisher’s website: “Rethinking Why People Vote: Voting as Dynamic Social Expression.” In The Behavioral Foundations of Policy, ed. Table A1: Covariate Balance among Unassigned Subjects Eldar Shafir. Princeton, NJ: Princeton University Press, Table A2: Direct and Indirect Mobilization Effects, 91–107. Covariate-adjusted Rolfe, Meredith. 2012. Voter Turnout: A Social Theory of Political Table A3: Logistic and IV Regression of Turnout of UnasParticipation. Cambridge: Cambridge University Press. signed Subject on Treatment Assignment of Assigned Sinclair, Betsy. 2012. The Social Citizen: Peer Networks Subject, Conditional on Household Party Preferences and Political Behavior. Chicago: University of Chicago Press. Table A4: Robustness Check: Logistic Regression of Sinclair, Betsy, Margaret McConnell, and Donald P. Green. Turnout of Unassigned Subject on Treatment Assignment 2012. “Detecting Spillover Effects: Design and Analysis of of Assigned Subject, Conditional on Household Party Multi-Level Experiments.” American Journal of Political SciPreferences, 5-Categories Operationalization ence 56(4): 1055–69. Table A5: Logistic Regression of Turnout on Treatment Stoker, Laura, and M. Kent Jennings. 1995. “Life-Cycle TranAssignment and Party Preference of Both Subjects, and sitions and Political Participation: The Case of Marriage.” Interactions between Assignment and Party Preferences American Political Science Review 98(2): 421–33. Table A6: Robustness Check: Logistic Regression Results: Stoker, Laura, and M. Kent Jennings. 2005. “Political Similarity and Influence between Husbands and Wives.” In The Social CACE of High vs. Low Partisan Intensity Call on Turnout Logic of Politics: Personal Networks as Contexts for Political of Unassigned Subjects, Conditional on Household Party Behavior, ed. Alan S. Zuckerman. Philadelphia: Temple UniPreferences, 5-Categories Operationalization versity Press, 51–74. Table A7: Comparison between Official City Council Straits, Bruce. 1990. “The Social Context of Voter Turnout.” Elections Results for May 2011 and May 2012, and Public Opinion Quarterly 54(1): 64–73. Complete November 2012 Labour Party Databasea (PreVerba, Sidney, Kay Lehman Schlozman, and Nancy Burns. 2005. treatment) across All Electoral Wards Included in the “Family Ties: Understanding the Intergenerational Transmission of Political Participation.” In The Social Logic of Experiment Politics: Personal Networks as Contexts for Political Behavior. Figure A1: Balance Check: Estimated Log Likelihood RePhiladelphia: Temple University Press, 95–114. sulting from Multinomial Logistic Regression of TreatWolff, Jonathan. 2002. “Are We Good Enough for Democracy?” ment Assignment of Household Member on Pretreatment Think 1: 29–34. Covariates Compared to Sampling Distributions of SimWooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Seculated Log Likelihood Statistics under Sharp Null. a) Left: tion and Panel Data. 2nd ed. Cambridge, MA: MIT Press. Assigned Subjects (p=.20) b) Right: Household Members Zuckerman, Alan S. 2005. The Social Logic of Politics: Personal Networks as Contexts for Political Behavior. Philadelphia: (p=.50) Temple University Press. Figure A2: Attrition Check: Estimated Log Likelihood Zuckerman, Alan S., Josip Dasović, and Jennifer Fitzgerald. Resulting from Multinomial Logistic Regression of Treat2007. Partisan Families: The Social Logic of Bounded Parment Assignment on Missing Outcome Data Compared tisanship in Germany and Britain. New York: Cambridge to Sampling Distributions of Simulated Log Likelihood University Press. Statistics under Sharp Null. a) Left: Excluding Household Zuckerman, Alan S., Jennifer Fitzgerald, and Josip Dasović. Composition Interactions (p=.26) b) Right: Including 2005. “Do Couples Support the Same Political Parties? Sometimes: Evidence from British and German Household Household Composition Interactions (p=.32) Panel Studies.” In The Social Logic of Politics: Personal NetFigure A3: High Partisan Intensity Script works as Contexts for Political Behavior, ed. Alan S. ZuckerFigure A4: Low Partisan Intensity Script man. Philadelphia: Temple University Press, 75–94. Figure A5: Questionnaire
3 Are Voters More Likely to Contribute to Other Public Goods? Evidence from a Large-Scale Randomized Policy Experiment Toby Bolsen Georgia State University Paul J. Ferraro Georgia State University Juan Jose Miranda The World Bank Voting has been described as a contribution to a public good. Are people who vote frequently therefore more likely to contribute to other public goods? Does partisanship affect how likely a person is to engage in these cooperative behaviors? Although surveys suggest that the answer to these questions is “Yes,” few empirical studies examine these questions using observed behaviors. We examine them in the context of a large-scale, randomized controlled trial to induce voluntary action in a common-pool resource dilemma. During a drought in the southeastern United States, pro-social messages that encouraged water conservation were randomly assigned to 35,000 out of 106,000 households. Frequent voters in primary and general elections (1990–2008) were substantially more responsive to the messages, but there was no detectable difference in the responses of Republican and Democrat households. Our results suggest that internalized pro-social preferences promote action for the public good across behavioral contexts. F or decades, scholars have argued that voting is a contribute to other public goods. Recent studies show that contribution to a public good and that many votsome individuals possess pro-social preferences, which ers view it as a civic duty (Campbell et al. 1960; affect behavior in laboratory experiments and stated inGerber and Green 2004; Knack 1992; Mueller 1989). In tentions to engage in collective actions like voting (Dawes, fact, voters often receive an “I voted” sticker after castLoewen, and Fowler 2011; de Oliveira, Croson, and Eckel ing their ballots in order to publicize their participa2008; Edlin, Gelman, and Kaplan 2007; Fowler 2006; tion in this socially valued activity (Riker and Ordeshook Fowler and Kam 2007; Jankowski 2002, 2007; Sandroni 1968). Recent experiments have strengthened this view and Fedderson 2006). Results from these studies suggest of voting as a pro-social action by demonstrating the that some individuals behave as if they have internalized power of nonpecuniary incentives to affect voter turnout pro-social preferences that guide whether or not they con(Arceneaux and Nickerson 2009; Davenport et al. 2010; tribute to public goods (e.g., cooperators versus egoists). Gerber, Green, and Larimer 2008, 2010; Gerber and In democratic systems, the degree to which, and cirRogers 2009). These pioneering experiments have made cumstances under which, individuals act pro-socially afimportant contributions to our understanding of the fects how governments solve social dilemmas. Individparadox of voting—i.e., why individuals vote despite inuals’ decisions determine collective outcomes and thus centives to free ride (Downs 1957; Fowler 2006; Knack affect the types of policies needed to promote the collecand Kropf 1998; Olson 1965). tive good. Given the provision of public goods is a priWe move this literature in a different direction by mary purpose of government, understanding pro-social examining the extent to which voters are more likely to behavior is thus critical for political science. In particular, Toby Bolsen is Professor of Political Science, Georgia State University, Atlanta, GA 30303 (tbolsen@gsu.edu). Paul J. Ferraro is Professor of Economics, Andrew Young School of Policy Studies, Georgia State University, PO Box 3992, Atlanta, GA 30302–3992 (pferraro@gsu.edu). Juan Jose Miranda is Economist of the Sustainable Development Department in the Latin America & the Caribbean Region at The World Bank, Washington, DC 20433 (unicojm@gmail.com). This material is based upon work supported by the Cooperative State Research, Education, and Extension Service, U.S. Department of Agriculture, under Agreement No. 2003–38869–02007. We thank Rachel Croson, Sean Richey, and seminar participants at Georgia State University and the Southern Economic Association (2011 meeting) for helpful comments. We thank Aries Arugay and Lindsey Herbel for research assistance. American Journal of Political Science, Vol. 58, No. 1, January 2014, Pp. 17–30 \002 C 2013, Midwest Political Science Association DOI: 10.1111/ajps.12052 17 18 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA elucidating the stability of pro-social preferences across and Gachter 2000; Fishbacher, Gachter, and Fehr 2001; decision domains is an important task for political scienHamman, Weber, and Woon 2011; Ostrom 2006). Few tists: is the propensity to take pro-social action a stable such experiments, however, examine the stability of propredisposition across domains, or is it domain-specific?1 social behavior across domains, and the few that do rely To answer this question, we examine the behavior almost exclusively on comparisons between laboratory of voters and nonvoters and ask, “Are people who vote studies and field experiments—with conflicting results frequently also more likely to contribute to other public (de Oliveira, Croson, and Eckel 2011; Grossman 2011; goods?” Additionally, we explore the role of partisanship: Levitt and List 2007; Voors et al. 2012). In contrast, our does partisanship affect how likely a voter is to engage study uses two naturally occurring decision domains (i.e., in pro-social behavior? Although surveys of self-reported a “field-field” comparison). behaviors suggest that the answers to these questions may During a drought in the southeastern United States, be “Yes,” no empirical study has examined these questions a water utility randomly assigned pro-social, norm-based using revealed (observable) behaviors. We examine these messages that encouraged households to conserve water questions with data from a randomized controlled trial (Ferraro and Price 2013).2 These messages caused housethat aimed to induce voluntary action in a common-pool holds to use less water, on average. We merge data from resource dilemma (water conservation during a drought). this field experiment with individual-level data on voting Although field experiments in political science are histories in primary, general, and special elections (1990– increasingly common, the vast majority focus on voter 2008) to explore whether a household’s voting frequency turnout (Arceneaux and Nickerson 2009; Gerber and is correlated with its responsiveness to a randomized proGreen 2000a, 2000b; Green 2011; Green and Nickersocial message promoting water conservation. son 2003; Michelson, Bedolla, and McConnell 2009) and We find that frequent voters are more responsive to none explore voter willingness to contribute to the provithe randomized pro-social messages, a result that suggests sion of other public goods. There is clear evidence from internalized pro-social preferences are stable across conthe political science literature that pro-social stimuli can texts.3 Our experimental design reduces the likelihood affect turnout decisions (Davenport et al. 2010; Gerber of spurious inferences because past voting decisions are and Green 2010; Gerber, Green, and Larimer 2008, 2010; uncorrelated with whether or not a household receives a Gerber and Rogers 2009; Green 2011). But political scienconservation message. Correlation is absent because the tists know less about the degree to which pro-social action sample is large (>100,000 households), and treatments in these contexts is driven by internalized pro-social prefwere randomized within small neighborhood units (see erences to contribute to the public good as opposed to Data and Methods). external threats of making noncompliance public. In the Our analysis also sheds light on the poorly unlatter case, pro-social action results from the desire to derstood relationship between partisanship and willingavoid public sanction or shame rather than an internalness to contribute to environmental public goods. We ized predisposition to cooperate. The design we describe below allows us to examine whether individuals display 2 Although conservation responses to the messages can also yield stable pro-social preferences across domains in the abprivate benefits, subsequent analyses of the experimental data supsence of a threat to make one’s behavior public. port the interpretation that the treatment effects were driven by Our field experiment builds on a vast literature in pro-social preferences rather than rival, private-preference-based mechanisms (e.g., signals of privately optimal behavior to agents the social sciences that uses laboratory experiments to with incomplete information) (Ferraro and Miranda 2013). elucidate the voluntary provision of public goods, in3 Identifying the specific form or forms of these stable internalized cluding the exploitation of common-pool resources (e.g., preferences is beyond the scope of this study. The economic and Andreoni and Croson 2008; Dawes and Thaler 1988; Fehr political science literatures postulate a range of plausible forms of pro-social preferences in collective action settings, including adherence to a civic duty or social norm (Bolsen 2013; Chong 2000; 1 Data and replication files for this study are available at AJPS Data Cialdini 2003; Davenport et al. 2010; Gerber, Green, and Larimer Archive on Dataverse: hdl:1902.1/21394. A preference comprises 2008, 2010; Gerber and Rogers 2009; Knack and Kropf 1998; Kropf “comparative evaluations of (i.e., a ranking over) a set of objects” 2009; Schultz 1999; Thogersen 2008), conditional cooperation (Druckman and Lupia 2000, 2). The stability of private (selfish) (Allcott 2011; Alpı́zar, Carlsson, and Johansson-Stenman 2008; preferences is a long-standing subject of debate among scholars Frey and Meier 2004; Shang and Croson 2009), altruistic pref(Converse 1964; Tversky and Kahneman 1981; Zaller 1992). Preferences (Fowler 2006; Fowler and Kam 2007), and civic norms erence stability is central to normative debates regarding the expromoting cooperation (Beck et al. 2002; Huckfeldt and Sprague tent to which citizens are capable of playing a meaningful role 1992; Ikeda and Richey 2005; Nickerson 2008; Putnam 1966). See in democratic governance (Bartels 2003; Druckman 2004). MoreGächter, Nosenzo, and Sefton (2012) for one recent study that tries over, economic models of decision making typically assume stable to differentiate rival pro-social preference forms (norms vs. altrupreferences. ism) in a laboratory gift-exchange experiment. STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 19 investigate whether responses to the experimental treatlatent variable—that is a sense of cooperation—affects ments are conditional on partisanship, as captured by incidences of these behaviors” (2009, 544; see also Brehm voting patterns in primary elections. Survey results sugand Rahn 1997; Fisher and Ackerman 1998; Knack 1992; gest that Republicans are less willing than Democrats to Knack and Kropf 1998). contribute to environmental public goods, but we know In addition to informing our understanding of the of only one study that measures observed behavior, rather stability of pro-social preferences, our study also conthan self-reported attitudes. We find that Republican tributes to related literature that explores how drawand Democratic primary voters respond similarly to proing attention to social norms can increase voter turnout social conservation requests. Overall, the results from our (Davenport et al. 2010; Gerber, Green, and Larimer 2008, study have implications for understanding when and why 2010; Gerber and Rogers 2009). These studies provide individuals take actions that promote the public good, as clear evidence that turnout is shaped by nonpecuniary well as for crafting policy interventions targeting behavior incentives, but the experimental pro-social messages ofchange. ten include external pressure to comply with prevailing norms—e.g., fear that one’s neighbors will be informed that he or she did not vote. Thus, it can be difficult to disentangle whether pro-social action results from (1) the Private Actions for the Public Good manifestation of an internalized pro-social disposition or (2) compliance with prevailing norms so as to avoid isoPrevious research that explores the stability of prolation or social sanction. In our study, the threat of public social preferences across decision contexts makes labknowledge of one’s actions is absent. to-field comparisons of behavior (Benz and Meier 2006; Inspired by these disparate literatures, our first reCarpenter and Seki 2011; de Oliveira, Croson, and Eckel search question asks: Are individuals who contribute to 2008; Karlan 2005; Voors et al. 2012) or uses self-reported the public good in one policy-relevant context also more measures of field behavior (Knack 1992; Knack and Kropf responsive when asked to contribute in a different context? 1998; Kropf 2009). The external validity of laboratory exOur study provides the first answer to this question usperiments that measure pro-social behavior, however, is ing revealed behavior measures and tests the following controversial (Barmettler, Fehr, and Zehnder 2012; Levitt hypotheses: and List 2007; also see Jerit, Barabas, and Clifford 2013), Null Hypothesis: There is no relationship between the freand self-reported measures of intentions are subject to quency with which people vote and their response to a social desirability biases (and the studies that use them pro-social appeal to conserve water during a drought. fail to control for confounding variables). Alternative Hypothesis: The more frequently one votes, In contrast, our study uses observed behaviors in two the more responsive he or she will be to a pro-social naturally occurring decision domains and an experimenappeal to conserve water during a drought. tal design to aid with inferences. Our design allows us to observe whether a tendency to express pro-social behavior A second distinct research question we explore in in one context (voting) is correlated with responsiveness this article relates to whether or not partisanship affects to a randomized request to contribute to the public good responsiveness to a randomized intervention requesting in an unrelated common-pool resource dilemma. Voting pro-social action in a common-pool resource dilemma. is a contribution to a pure public good because the benRecently, scholars have begun to explore the impact of efits are nonrival and nonexcludable. The likelihood that partisan identification on behavior in laboratory setone’s vote will be pivotal in any election is close to zero, tings that require trust and cooperation (Carlin and Love resulting in incentives to free ride and abstain from con2013; Fowler 2006; Fowler and Kam 2007). For example, tributing to the public good. Conversely, an individual’s Carlin and Love (2013) find that copartisans in the United water use is not a pro-social behavior per se; it is a priStates exhibit greater trust compared to rival partisans vate consumption decision. However, an explicit request in a dictator game. Rand et al. (2009) demonstrate that to reduce water consumption during a local drought is an in-group biases shaped the behavior of partisans in a dicappeal to cooperate by refraining from using a rival but tator game during the 2008 presidential election season. nonexcludable public good. As in voting, water conservaFor instance, in experiments conducted during the camtion is costly and free riding on the nonexclusive benefits paign, supporters of Clinton and Obama gave systematis possible. As Kropf noted, a positive correlation between ically more money to supporters of the same candidate pro-social behavior in one domain and pro-social behavthan to supporters of the other candidate. In another reior in a different domain suggests “the same underlying cent lab experiment, Gromet, Kunreuther, and Larrick 20 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA (forthcoming) find that conservative students were less Null Hypothesis: Democratic and Republican primary likely to purchase a more expensive energy-efficient light voters, on average, respond equally to a pro-social bulb when it was labeled with a proenvironmental mesappeal to conserve water during a drought. sage. These studies suggest that looking at the relationship Alternative Hypothesis: Relative to Republican primary between partisanship and willingness to contribute to a voters, Democratic primary voters are more responpublic good would be a fruitful path of inquiry. sive, on average, to a pro-social appeal to conserve It is no secret that concern for the environment water during a drought. has become a liberal identifier. In 2008, a survey revealed a 34% gap between Democrats and Republicans in their agreement that global warming exists (Dunlap and McCright 2008; also see McCright and Dunlap 2011). In Data and Methods 2009, a survey revealed a 23% gap between the DemoStudy Site crat and Republican agreement “that people ought to pay higher prices to protect the environment” (Pew Research Ferraro and Price (2013) report that in 2006, Cobb Center 2009). These gaps are even greater among elected County, Georgia, contained an estimated 679,325 officials. The 2010 National Environmental Scorecard of people—71% of whom identified as white/Caucasian and the League of Conservation Voters gives the House Demo23% as black. The 2000 census reported 227,487 housecratic leadership a score of 100 (out of 100) and the Reholds, approximately two-thirds of which were owner publican leadership a score of 0. The Senate Democratic occupied. In addition, 89% of adults over the age of 25 leadership also received a score of 100, while Republihad a high school degree, and 40% had a bachelor’s decans received a 5. Although environmental concern has gree. In 2004, median household income was estimated not always correlated strongly with partisan identification at $52,936. Cobb County, the home district of the forin the United States, the last decade has seen increasing mer Republican House Speaker Newt Gingrich, is not polarization along party lines. Guber explains, “Today, known for environmentalism. Its current congressmen political ideology and partisan identification are imporhave some of the lowest League of Conservation Voter tant determinants of general environmental concern, and scores (LCV 2008) recorded in 2007 and 2008 (Price 9% are not exclusive to global warming” (2013, 94). Thus, and Gingrey 0%). partisans may differ in the way in which they respond to requests for pro-social action in a common-pool resource dilemma. Experimental Data We know of only one study to examine the relationThe field experiment tested whether conservation mesship between citizen partisanship and environmental acsages affected monthly household water consumption betion (rather than stated preferences or intentions). In the tween June and September 2007 (Ferraro and Price 2013). context of household electricity consumption, Costa and Ferraro partnered with the Cobb County Water System Kahn (2013) found that a subgroup of liberals in liberal (CCWS)—an agency of the Cobb County Government in neighborhoods responds more to the receipt of a Home the metropolitan area of Atlanta, Georgia—to administer Energy Report and reduces its energy consumption more three versions of a water conservation message to 35,093 than a subgroup of conservatives in conservative neighhouseholds, with another 71,779 households assigned borhoods.4 to a control group (106,872 in total, all single-family These studies motivate our second research question: dwellings).5 In 2000, Cobb County was Georgia’s second Does partisanship affect responsiveness to pro-social appeals to contribute to environmental public goods? Our study tests 5 the following hypotheses: CCWS distributes treated surface water to about 170,000 Cobb County customers. About 150,000 of these customers reside in single-family dwellings. Monthly preand postexperiment water 4 Specifically, Costa and Kahn stated, “We find that among political data were provided by the CCWS billing department. Customers liberals who purchase electricity from renewable resources, who whose billing address had changed between May 2006 and April donate to environmental causes, and who live in a census block 2007 were excluded from the experimental sampling because one group where the share of liberals is in the top 75th percentile, reof the treatments required water-use history. Customers with conceiving a HER led to reductions in electricity usage of 3.6%. In sumption lower than 4,000 gallons per month were excluded by contrast, among political conservatives who do not pay for renewthe utility. Water meters are read and bills are sent daily based on a able electricity, who do not donate to environmental groups, and household’s assignment to one of 390 “meter routes.” Ferraro and who live in a census block group where the share of liberals is in Price explain, “To ensure that we have no systematic differences the bottom 25th percentile, receiving a HER led to reductions in across treatments in the day of the month an outcome is meaelectricity usage of 1.1%” (2013, 682). sured, we randomized treatment assignment within meter routes STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 21 largest user of the public water supply, using almost 8% of Voting Data Georgia’s public water supply (Ferraro and Price 2013). The experiment was designed to provide feedback to We merged data from the field experiment with inCCWS on the effectiveness of mail-based conservation dividual records of voter turnout (purchased from programs. www.aristotle.com) and county tax-assessor data on the The experiment comprised a control group and three home.6 We linked households from the experiment with treatment groups: the voter database by matching addresses of individuals within households. Using addresses and address(1) Treatment 1: An “information only” message matching software (reclink command in Stata; Blasnik that comprised a two-sided tip sheet about ways 2010), we were able to match over 85% of the households the household could conserve water. in the water-conservation experiment with the registered(2) Treatment 2: The tip sheet (Treatment 1) plus voter database. We assume that members of unmatched a pro-social appeal encouraging customers to households are not registered in the electoral system and “do their part” and “work together to use water include a dummy variable for these households in the stawisely.” tistical analyses (see last subsection below). The results we (3) Treatment 3: The tip sheet (Treatment 1), the report below are robust to excluding these households: the pro-social appeal (Treatment 2), and a soestimated treatment effects differ by less than 8% when cial comparison. The social comparison conthese households are excluded (results are available upon trasted each household’s water use from June to request). October 2006 to the median county household use for the same period and indicated the percentile in which the household fell during this Measures period (see supporting information, Appendix The dependent variable in our analyses is a measure of A1, for a copy of each message). each household’s water use for June through September The first treatment, a technical advice letter, works 2007 (monthly, in thousands of gallons). Given the treatthrough a single channel—scrutiny—and has the smallment assignment is at the household level, we must agest effect on household water consumption. The second gregate voting data from the individual to the household treatment augmented the technical advice letter by inlevel. We measure voting frequency and partisanship as cluding an appeal to pro-social preferences highlighting follows: the importance of conserving water. This appeal high(1) Vote frequency = The number of times every lights a social norm—conservation and concern for the registered voter in the house voted in a primary, environment—and led to additional reductions in avergeneral, or special election (1990–2008) divided age household water use. Treatment 3 makes the social by the number of times every registered voter norm more salient by including a social comparison and in the household could have voted, which demay heighten the extent to which the household believes pends on the birth year of each registered voter. its actions are scrutinized. Ferraro and Price (2013) preThis measure treats differently aged individuals dict and find evidence that this higher level of scrutiny who vote in every election equally. In contrast, generated the largest reductions in overall water cona simple count of vote frequency would also sumption. The messages were all mailed during the week be picking up the effects of age (older people, of May 21, 2007, via first-class mail in official CCWS by definition, have more opportunities to vote envelopes. Four weeks later, all treatment households rein our panel). As robustness checks (see supceived a second copy of the tip sheet (and no other comporting information, Tables A4–A6), we also ponent of the treatments). Households were unaware that run our analysis using a simple count of votthe messages were part of an experimental design (a soing frequency as well as three other methods called “natural field experiment”). Our sample includes all households in Ferraro and Price’s field experiment. 6 We verified the accuracy of the vote history data from Aristotle with data on voting history purchased from the Cobb County Board of Elections. Although the experiment took place in the summer of 2007, we include voting data from the general election in 2008 which correspond to neighborhood sections. . . . [This] increases due to the high turnout in that election. In separate analyses, we the precision of the estimates of treatment effects provided that unconfirmed that receiving a treatment administered in the summer observables affecting treatment response are more similar within of 2007 promoting water conservation did not affect the likelihood rather than between meter routes” (2013, 69). of voting in the 2008 general election. 22 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA of aggregation: (a) considering only the oldest in any observational study, one must be cautious when registered voter in the household, (b) considerestimating heterogeneous treatment effects. Randomizaing only the most frequent voter in the housetion does not guarantee covariate balance within subhold, and (c) selecting at random one member groups, and thus one might mistake spurious correlations of the household. All results are similar and for heterogeneous causal effects (Imai and Strauss 2011). do not change the inferences drawn from the However, our sample size is large, and our randomizadata. tion was done within small neighborhood groups (almost (2) Democrat = “1” if the number of times every 400 meter-route groups). Thus, one would expect balregistered voter in the house voted in a Demoance among observable and unobservable characteristics cratic primary election is greater than the numacross the experimental treatment arms within groups deber of times every registered voter in the house fined by our voting frequency and partisanship measures. voted in a Republican primary election (1990– To provide evidence of this balance, we examine pretreat2008); “0” otherwise. ment water use across the treatment and control groups (3) Republican = “1” if the number of times every within each voting frequency decile (supporting informaregistered voter in the house voted in a Retion, Table A2) and within our households labeled Demopublican primary election is greater than the crat, Republican, and “Neither” (supporting information, number of times every registered voter in the Table A3). For example, we test (F-test) whether prehouse voted in a Democratic primary election treatment mean water uses across treatment and control (1990–2008); “0” otherwise. groups are statistically indistinguishable from each other within the first decile of voter frequency, within the second Among the households we label as Democrat, 76% decile, etc. With 10 sequential tests and Type I error rate only vote in Democratic primaries, and 65% of houseset to 0.05, we would expect approximately one of them to holds we label as Republican only vote in Republican reject the null hypothesis of no difference through chance primaries. As robustness checks, we also run our analalone. In no test is the null hypothesis rejected. Alternaysis removing the households that have members who tively, one could test whether the mean voting-frequency vote in both primaries, as well as using the three aggregameasure is equal across treatment arms. In all treatment tion methods described for the Vote Frequency measure. arms, the mean is identical at 0.10, and we fail to reject the We use voting frequency in primary elections to measure null of equality. Looking at the three sequential tests for partisanship rather than party registration for two reapartisanship, we also fail to reject the null hypothesis of no sons. First, voters in Georgia do not register with a party differences across treatment arms in all cases. Thus, our affiliation and are allowed to vote in any primary they data appear balanced with respect to the pretreatment wish (one primary per year). Second, even if they were to outcome variables within subgroups across treatment register with a party, we believe that the act of voting in arms. a primary election is an equally valid measure of partiBased on recommendations from Bruhn and McKensanship compared to a party affiliation someone declares zie (2009) and to increase the statistical precision of when first registering to vote. We assume that individuour estimates, we include dummy variables for the als who incur the costs associated with voting in primary water-meter routes in which randomization was conelections are more partisan, on average, than individuals ducted. These are excluded from the tables for prewho do not. Thus, our measure picks up the two tails of sentational simplicity. To further increase statistical partisanship much more clearly at the cost of lumping toprecision, we also include in our regression models pregether independents and voters who would register with treatment household water use, fair market value of the a party but never vote in the primaries.7 house (dollars), age of the house (years), and a dummy variable indicating if the home is owner occupied. We also create another dummy variable for households with no Covariate Balance and Statistical Controls registered voters—i.e., the 15% of households we could In experimental studies in which treatments were not not match to the registered voter list. Our results are randomized within subgroups (e.g., partisan groups) or robust to excluding these covariates. Indeed, the estimated effects barely change (at second decimal place), further strengthening our assertion that the control sub7 The omitted registered voter group thus comprises households groups are valid counterfactuals for the treated subgroups with registered voters who never vote in a primary (25,703) and households with registered voters whose primary election counts even though randomization was not conducted within cancel each other out (2,568). subgroups. STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 23 TABLE 1 Linear Regression: Water Use from tion. Each regression model includes meter-route dummy June through September 2007 variables and household characteristics described in the previous section.8 In Model A, we collapse the treatments Model A Model B to a single dummy variable to increase statistical preci∗∗∗ Vote Frequency (%) 4.303 4.316∗∗∗ sion and because we are interested in whether frequent (0.651) (0.651) voters are more responsive to any appeal for voluntary Any Treatment (= 1) −0.691∗∗∗ — conservation action. The key variable of interest for our (0.162) first research question is the interaction between receivAny Treatment ∗ Vote −2.507∗∗ — ing any treatment and our household measure of vote Frequency (1.032) frequency (Any treatment ∗ Vote frequency). Treatment 1 — 0.005 Registered voters who have no voting history and are (0.279) exposed to treatment reduce their water consumption by Treatment 2 — −0.695∗∗∗ 691 gallons on average (second row, Table 1). The inter(0.240) action term of voting frequency and treatment (−2.507) Treatment 3 — −1.380∗∗∗ indicates that households with a voting history of 100% (0.227) reduce water consumption by an estimated additional Treatment 1 ∗ Vote Frequency — −1.625 2,507 gallons, on average, as a result of receiving a conser(1.691) vation request by mail (over 3,000 gallons in total). The Treatment 2 ∗ Vote Frequency — −2.685∗ size of the additional estimated average treatment effect (1.579) for the most frequent voting household represents a 6.2% Treatment 3 ∗ Vote Frequency — −3.237∗∗ reduction in water consumption in summer 2007 com(1.475) pared to the estimated mean counterfactual use (dividing Unregistered (= 1) −0.091 −0.087 2,507 by the average consumption of the control group (0.173) (0.173) in 2007). To better illustrate the magnitude of this differWater Use from June to 0.333∗∗∗ 0.333∗∗∗ ence, consider that a five-minute shower uses anywhere November 2006 (0.013) (0.013) from 10 to 25 gallons of water, and the average top-load Water Use in April and May 0.812∗∗∗ 0.811∗∗∗ washing machine between 40 and 45 gallons of water per 2007 (0.046) (0.046) load. Fair Market Value 1.80e-05∗∗∗ 1.80e-05∗∗∗ Drawing from the literature in behavioral psychology (2.56e-06) (2.56e-06) and economics, Ferraro and Price present a theory that Age of Home 0.028∗∗ 0.028∗∗ posits that as one moves from treatment 1 to treatment (0.011) (0.011) 3, the strength of pro-social preferences increases (see the Ownership Status (= 1 if 0.491∗∗∗ 0.485∗∗∗ fourth section). Model B in the right-hand column of owner) (0.182) (0.182) Table 1 reports the results after each treatment is indiAge of Homeowner (= 1 if 0.012 0.003 vidually interacted with our measure of vote frequency. >65 years old) (0.192) (0.192) If one accepts Ferraro and Price’s theory and if pro-social Constant −2.420 −2.425 preferences were stable across our two contexts, the esti(1.731) (1.732) mated mean effects of each treatment interacted with our Observations 103,340 103,340 vote frequency measure would also be ordered similarly: R2 0.64 0.64 treatment 3 + treatment 3 ∗ vote frequency > treatment 2 + treatment 2 ∗ vote frequency > treatment 1 + treatment Note: Cell entries are parameter estimates (robust standard deviations in parentheses) for a linear regression estimation of water 1 ∗ vote frequency. We indeed see this predicted ordering. use (in thousands of gallons) in summer 2007 on the covariates. Among nonvoters, treatment 1 has no detectable effect Not listed are 390 dummy variables for water meter routes that on water use, whereas treatment 2 reduces water use by control for neighborhood effects. We use one-tailed significance 695 gallons, and treatment 3 reduces water use by 1,380 tests because our alternative hypothesis is directional. ∗ p < 0.05, ∗∗ p < 0.025, ∗∗∗ p < 0.005. gallons, on average. For households with the highest voting frequency, treatment 1 reduces water use by an addiResults tional 1,625 gallons, treatment 2 reduces water use by an Table 1 reports the results from ordinary least squares (OLS) linear regression models of household water use 8 We lose about 3,000 observations in Tables 1 and 2 because of in the four months following the randomized intervenmissing covariate values in the tax-assessor data. 24 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA additional 2,685 gallons, and treatment 3 reduces water texts. Framing this hypothesis in another way, the data use by an additional 3,237 gallons, on average. are consistent with the existence of an internalized sense The tests in Table 1 lead us to reject our null hypothof civic-mindedness that drives some individuals to conesis that there is no relationship between the frequency tribute to public goods across contexts. with which people vote and their response to a pro-social For completeness, we also estimate a second model appeal to conserve water during a drought: frequent votin Table 2 (Model B) in which we interact each treaters are more responsive to pro-social requests to reduce ment individually with partisanship variables. Because water consumption. The results in Table 1 thus provide we break up our sample into many subgroups, we do not evidence of stable social preferences across behavioral have enough statistical power to discriminate among the domains. different treatment messages. A couple of coefficients are significantly different from zero in a potentially intriguing pattern, but the only definitive conclusion that can be Partisanship and Cooperative Preferences drawn from the model is the same conclusion we draw from Model A: Democratic and Republican primary votOur second research question asks whether there is heters respond similarly when receiving a request to take erogeneity in responses to receiving an experimental pro-social collective action for the environment. treatment based on partisanship. To explore this question, The analysis in Table 2 attempts to estimate the effect we estimate four models to infer the effect of a household’s of partisanship on responsiveness to a pro-social appeal partisanship, as defined by voting history in primary electo contribute to an environmental public good. In contions between 1990 and 2008, on responsiveness to protrast to the analysis in Table 1, a potential concern with social appeals to conserve water (Table 2 and Table 3). the analysis in Table 2 is that there may be time-invariant Table 2 reports the results from two OLS linear regresunobservable factors associated with being a Republican sions that test whether Republicans or Democrats are or Democrat that influence one’s willingness to reduce more responsive to the experimental treatments. Model water use in response to receiving a randomly assigned A collapses the three different versions of the treatment treatment message. To address this concern, we take adinto a single binary measure (Any treatment). vantage of monthly preand postexperimental measures The estimates in the third row imply that unregistered of each household’s water use to estimate two fixed-effects and registered, nonprimary voting households receiving panel-data regression models. These models give up staa treatment message reduced summer 2007 water contistical power in exchange for greater control over timesumption by an estimated 656 gallons. The fourth and invariant unobservable characteristics at the household fifth row estimates imply that Democrat and Republican level. Table 3 reports these results. households respond more strongly to a treatment mesReceiving any treatment significantly reduces water sage than the nonprimary voting households. However, use an estimated 295 gallons (monthly; p < 0.001). In the responses of Democrat and Republican households contrast to the coefficient estimates in Table 2 that imply are similar in magnitude and not statistically different Democratic and Republican households reduce water use from each other (F-statistic = 0.29). Thus, in contrast to more than nonpartisans after receiving a treatment mesour first hypothesis test, we are unable to reject the second sage, the coefficients in Table 3 imply no difference among null hypothesis that, among Democratic and Republican these households: the coefficients for Any Treatment ∗ primary voters, there is no difference in the average reDemocrat ∗ Postexperiment period and for Any Treatment ∗ sponse to a pro-social appeal to conserve water during a Republican ∗ Postexperiment period are not statistically drought. different from zero. The estimated coefficients, however, Because both Democratic and Republican primary are consistent with the main result in Table 2: Demovoters significantly reduce water consumption in response cratic and Republican households respond similarly after to receiving an experimental treatment, the results from receiving a request to take pro-social collective action for Model A in Table 2 also bolster the results in Table 1, which the environment.9 we argued are consistent with the hypothesis that frequent voters display stable cooperative preferences across con9 Our results seem to conflict with Costa and Kahn (2013), but texts. In other words, our results in Table 1 are not simply a closer look suggests they are more similar than different. The driven by Democratic voters responding to the expericoefficient on their “Liberal” coefficient (which combines Greens, mental treatments, but rather they are consistent with Peace and Freedom, and Democrat party members) is tiny and not statistically significantly different except in the one regression (out the hypothesis that a common factor—e.g., adherence to of five) that has the fewest controls. Only when they create subnorms of cooperation—drives behavior across both congroups by combining the weak effect of the political party variable STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 25 TABLE 2 Linear Regression: Partisanship and Water Use from June through September 2007 Model A Model B ∗ Democrat (= 1) 0.318 0.319∗ (0.186) (0.186) Republican (= 1) 1.051∗∗∗ 1.052∗∗∗ (0.168) (0.168) Any Treatment (= 1) −0.656∗∗∗ — (0.186) Any Treatment ∗ Democrat −0.619∗∗ — (0.294) Any Treatment ∗ Republican −0.459∗ — (0.267) Treatment 1 — 0.079 (0.334) Treatment 2 — −0.669∗∗ (0.273) Treatment 3 — −1.354∗∗∗ (0.253) Treatment 1 ∗ Democrat — −1.008∗∗ (0.474) Treatment 2 ∗ Democrat — −0.522 (0.438) Treatment 3 ∗ Democrat — −0.364 (0.431) Treatment 1 ∗ Republican — −0.100 (0.446) Treatment 2 ∗ Republican — −0.559 (0.397) Treatment 3 ∗ Republican — −0.762∗∗ (0.387) Unregistered (= 1) −0.115 −0.115 (0.182) (0.182) Water Use from June to November 2006 0.333∗∗∗ 0.332∗∗∗ (0.013) (0.013) Water Use in April and May 2007 0.811∗∗∗ 0.811∗∗∗ (0.046) (0.046) Fair Market Value 1.78e-05∗∗∗ 1.79e-05∗∗∗ (2.56e-06) (2.56e-06) Age of Home 0.028∗∗∗ 0.028∗∗∗ (0.011) (0.011) Ownership Status (= 1 if owner) 0.520∗∗∗ 0.515∗∗∗ (0.180) (0.180) Age of Homeowner (= 1 if >65 years old) 0.146 0.140 (0.192) (0.192) Constant −2.538 −2.549 (1.728) (1.728) Observations 103,448 103,448 R2 0.64 0.64 Note: Cell entries are parameter estimates (robust standard deviations in parentheses) for a linear regression estimation of water use (in thousands of gallons) in summer 2007 on the covariates. Not listed are 390 dummy variables for water meter routes that control for neighborhood effects. We use one-tailed tests because our alternative hypothesis is directional. ∗ p < 0.05, ∗∗ p < 0.025, ∗∗∗ p < 0.005. 26 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA TABLE 3 Fixed-Effects Linear Regression: Water Use Partisanship (Panel Data, Monthly) Model A Model B ∗∗∗ Postexperiment 0.457 0.456∗∗∗ (0.068) (0.068) Democrat ∗ Postexperiment −0.126 −0.125 (0.089) (0.089) Republican ∗ Postexperiment 0.510∗∗∗ 0.511∗∗∗ (0.071) (0.071) Unregistered ∗ Postexperiment −0.295∗∗∗ −0.294∗∗∗ (0.107) (0.107) Any Treatment ∗ Postexperiment −0.297∗∗∗ — (0.114) Any Treatment ∗ Democrat ∗ Postexperiment 0.039 — (0.126) Any Treatment ∗ Republican ∗ Postexperiment 0.156 — (0.117) Treatment 1 ∗ Postexperiment — −0.027 (0.091) Treatment 2 ∗ Postexperiment — −0.491 (0.318) Treatment 3 ∗ Postexperiment — −0.369∗∗∗ (0.073) Treatment 1 ∗ Democrat ∗ Postexperiment — −0.146 (0.128) Treatment 2 ∗ Democrat ∗ Postexperiment — 0.202 (0.325) Treatment 3 ∗ Democrat ∗ Postexperiment — 0.0560 (0.117) Treatment 1 ∗ Republican ∗ Postexperiment — 0.114 (0.116) Treatment 2 ∗ Republican ∗ Postexperiment — 0.319 (0.321) Treatment 3 ∗ Republican ∗ Postexperiment — 0.027 (0.099) Constant 8.552∗∗∗ 8.552∗∗∗ (0.006) (0.006) Observations (month-household) 1,813,590 1,813,590 Number of households 106,682 106,682 Note: Cell entries are parameter estimates (standard deviations in parentheses) for a fixed-effects panel data estimator of monthly water use (in thousands of gallons). Fixed effects are modeled at the household level. We use one-tailed significance tests because our alternative hypothesis is directional. ∗ p < 0.05, ∗∗ p < 0.025, ∗∗∗ p < 0.005. Conclusions ence. The degree to which individuals are willing to make voluntary contributions to the public good determines Understanding why some citizens, but not others, take acthe policies that need to be in place to reach collectively tion for the public good strikes at the heart of political scidesirable outcomes. To contribute to a broader understanding of pro-social behavior, we study the expression with the effects of other more influential variables (e.g., donating to of pro-social preferences across domains and shed light environmental causes, living in top-quartile liberal neighborhood) on the mechanisms by which pro-social messages impact do they find significant differences. politically relevant actions. STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 27 Our first research question explores the stability of Our study also has implications for policy makers pro-social preferences: is the propensity to take proand officials who craft messages to influence private acsocial action a stable predisposition across domains, or tions with public consequences. Messages promoting acis it domain-specific? In contrast to previous studies that tion for the public good may be more cost-effective if make lab-to-field comparisons of behavior or use selfthey target individuals who are more predisposed to coreported measures of behavior, we use a field-to-field operate (e.g., frequent voters).10 More research, however, comparison of revealed preferences with data from a ranwould be needed to confirm these conjectures as well as domized policy experiment and the voting histories of our findings on the stability of pro-social preferences and households. Consistent with the hypothesis of stable prothe relationship between partisanship and responsiveness social preferences across domains, we find that frequent to pro-social appeals to contribute to an environmental voters were more responsive to a randomized policy incommon-pool resource. Future research on partisanship tervention that used pro-social appeals to encourage parin other environmental contexts is needed to elucidate in ticipation in a collective action: frequent voters reduced what ways partisanship affects responses to environmenwater use significantly more than less frequent voters and tal policies and programs. The results in this case suggest nonvoters. that, counter to common intuition, Democrats and ReOur second research question contributes to the inpublican voters respond similarly to a pro-social request choate literature on the impact of partisan identification for conservation; however, these results may be restricted on cooperative behavior. Counter to expectations that to water use and not apply to global warming–related beDemocrat households would be more responsive to prohaviors or other environmentally relevant actions. Future social appeals to contribute to an environmental public studies also should, like our study, take advantage of the good, we find no evidence that Republican and Demogrowing number of randomized controlled social expercrat households respond differently in our sample. This iments using large sample sizes. By testing our hypotheresult also implies that the relationship we find between ses in other geographic and behavioral contexts, we can voting and responsiveness to pro-social appeals to congreatly improve our understanding of the way in which serve water is not being driven by reductions in water pro-social preferences and partisanship shape collective use among Democrats only (or Republicans only). Thus, action. a common factor—e.g., adherence to civic norms, altruistic preferences, or a sense of civic-mindedness—may be driving the increased responsiveness to pro-social reReferences quests among partisans of both stripes. In addition to informing our understanding of proAllcott, Hunt. 2011. “Social Norms and Energy Conservation.” social preferences and partisanship, our study also conJournal of Public Economics 95(9–10): 1082–95. tributes to the growing literature in political science that Alpı́zar, Francisco, Fredrick Carlsson, and Olof Johanssonexamines how norm-based interventions affect politiStenman. 2008. “Anonymity, Reciprocity, and Conformity: cally relevant actions (Bolsen 2013; Davenport et al. 2010; Evidence from Voluntary Contributions to a National Park in Costa Rica.” Journal of Public Economics 92(5–6): Gerber, Green, and Larimer 2008, 2010). This literature 1047–62. highlights how social pressure shapes willingness to enAndreoni, James, and Rachel Croson. 2008. “Partners versus gage in collective actions. Less clear, however, is the extent Strangers: Random Rematching in Public Goods Experto which pro-social actions are driven by internalized beiments.” Handbook of Experimental Economics Results 1: havioral predispositions as opposed to compliance out of 776–83. fear of social sanction. Our study demonstrates that, in the Arceneaux, Kevin, and David W. Nickerson. 2009. “Who Is absence of external pressure or fear that one’s behavior is goMobilized to Vote? A Re-Analysis of 11 Field Experiments.” American Journal of Political Science 53(1): 1–16. ing to be revealed publicly, taking action for the public good Blasnik, Michael. 2010. “Reclink: Stata Module to Probin one domain is correlated with responsiveness to a proabilistically Match Records.” http://ideas.repec.org/c/boc/ social message encouraging collective action in a different bocode/s456876.html domain. Of course, political scientists have long known Barmettler, Franziska, Ernst Fehr, and Christian Zehnder. that voters possess pro-social preferences that drive them 2012. “Big Experimenter Is Watching You! Anonymity and to vote (Fowler 2006). What is novel about our study is Prosocial Behavior in the Laboratory.” Games and Economic Behavior 75(1): 17–34. that it is the first to examine revealed preferences with respect to whether voters are more likely to participate in 10 Of course, the total aggregate effect from targeting, say, frequent an unrelated collective action, in a different domain, when voters would depend on how many frequent voters are in the poppresented with an explicit request for cooperation. ulation. 28 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA Bartels, Larry M. 2003. “Democracy with Attitudes.” In ElecFraming Effects.” American Political Science Review 98(4): toral Democracy, ed. Michael Bruce MacKuen and George 671–86. Rabinowitz. Ann Arbor: University of Michigan Press, 48– Druckman, James N., and Arthur Lupia. 2000. “Preference For82. mation.” Annual Review of Political Science 3(1): 1–24. Beck, Paul Allen, Russell J. Dalton, Steven Greene, and Robert Dunlap, Riley E., and Aaron M. McCright. 2008. “A WidenHuckfeldt. 2002. “The Social Calculus of Voting: Interpering Gap: Republican and Democratic Views on Climate sonal, Media, and Organizational Influences on Presidential Change.” Environment 50(5): 26–35. Choices.” American Political Science Review 96(1): 57–73. Edlin, Aaron, Andrew Gelman, and Noah Kaplan. 2007. “Voting Benz, Matthias, and Stephan Meier. 2008. “Do People Behave as a Rational Choice: Why and How People Vote to Improve in Experiments as in the Field? Evidence from Donations.” the Well-Being of Others.” Rationality and Society 19(3): Experimental Economics 11(3): 268–81. 294–314. Bolsen, Toby. 2013. “A Light Bulb Goes On: Invoking Norms Fehr, Ernst, and Simon Gächter. 2000. “Cooperation and Punin Appeals for Collective Action.” Political Behavior 35(1): ishment in Public Goods Experiments.” American Economic 1–20. Review 90(4): 980–94. Brehm, John, and Wendy Rahn. 1997. “Individual-Level EviFerraro, Paul J., and Juan Jose Miranda. 2013. “Heterogeneous dence for the Causes and Consequences of Social Capital.” Treatment Effects and Mechanisms in Information-Based American Journal of Political Science 41(3): 999–1023. Environmental Policies: Evidence from a Large-Scale Field Bruhn, Miriam, and David McKenzie. 2009. “In Pursuit of BalExperiment.” Resource and Energy Economics 35: 356–79. ance: Randomization in Practice in Development Field ExFerraro, Paul J., and Michael K. Price. 2013. “Using Non– periments.” American Economic Journal 1(4): 200–232. Pecuniary Strategies to Influence Behavior: Evidence from Campbell, Angus, Phillip E. Converse, Warren E. Miller, and a Large–Scale Field Experiment.” Review of Economics and Donald E. Stokes. 1960. The American Voter. Chicago: UniStatistics 95(1): 64–73. versity of Chicago Press. Fischbacher, Urs, Simon Gächter, and Ernst Fehr. 2001. “Are Carlin, Ryan E., and Gregory J. Love. 2013. “The Politics of People Conditionally Cooperative? Evidence from a Public Interpersonal Trust and Reciprocity: An Experimental ApGoods Experiment.” Economic Letters 71(3): 397–404. proach.” Political Behavior 35(1): 43–63. Fisher, Robert J., and David Ackerman. 1998. “The Effects Carpenter, Jeffrey, and Erika Seki. 2011. “Do Social Preferof Recognition and Group Need on Volunteerism: A Soences Increase Productivity? Field Experimental Evidence cial Norm Perspective.” Journal of Consumer Research 25(3): from Fishermen in Toyama Bay.” Economic Inquiry 49(2): 262–75. 612–30. Fowler, James H. 2006. “Altruism and Turnout.” Journal of Chong, Dennis. 2000. Rational Lives: Norms and Values in PoliPolitics 68(3): 674–83. tics and Society. Chicago: University of Chicago Press. Fowler, James H., and Cindy D. Kam. 2007. “Beyond the Self: Cialdini, Robert B. 2003. “Crafting Normative Messages to ProSocial Identity, Altruism, and Political Participation.” Jourtect the Environment.” Current Directions in Psychological nal of Politics 69(3): 813–27. Science 12(4): 105–9. Frey, Bruno, and Stephen Meier. 2004. “Social Comparisons Converse, Philip. 1964. “The Nature of Belief Systems in Mass and Pro-Social Behavior: Testing ‘Conditional CooperaPublics.” Critical Review 18(1–3): 1–74. tion’ in a Field Experiment.” American Economic Review 94: Costa, Dora, and Matthew Kahn. 2013. “Energy Conservation 472–82. ‘Nudges’ and Environmentalist Ideology: Evidence from a Gächter, Simon, Daniele Nosenzo, and Martin Sefton. 2012. Randomized Residential Electricity Field Experiment.” Jour“Peer Effects in Pro-Social Behavior: Social Norms or Social nal of the European Economic Association 11(3): 680–702. Preferences?” IZA Discussion Paper No. 6345. Davenport, Tiffany C., Alan S. Gerber, Donald P. Green, Gerber, Alan S., and Donald P. Green. 2000a. “The Effects Christopher W. Larimer, Christopher B. Mann, and Costas of Canvassing, Telephone Calls, and Direct Mail on Voter Panagopoulous. 2010. “The Enduring Effects of Social PresTurnout: A Field Experiment.” American Political Science sure: Tracking Campaign Experiments over a Series of ElecReview 94(3): 653–63. tions.” Political Behavior 32(3): 423–30. Gerber, Alan S., and Donald P. Green. 2000b. “The Effect of Dawes, Christopher T., Peter J. Loewen, and James H. Fowler. a Nonpartisan Get-Out-the-Vote Drive: An Experimental 2011. “Social Preferences and Political Participation.” JourStudy of Leafleting.” Journal of Politics 62(3): 846–57. nal of Politics 73(3): 845–56. Gerber, Alan S., and Donald P. Green. 2004. Get Out the Vote! Dawes, Robyn M., and Richard H. Thaler. 1988. “Anomalies: How to Increase Voter Turnout. Washington, DC: Brookings Cooperation.” Journal of Economic Perspectives 2(3): 187–97. Institution Press. de Oliveira, Angela C. M., Rachel T. A. Croson, and Catherine Gerber, Alan S., and Donald P. Green. 2010. “Introduction to Eckel. 2011. “The Giving Type: Identifying Donors.” Journal Social Pressure and Voting: New Experimental Evidence.” of Public Economics 95(5–6): 428–35. Political Behavior 32(3): 331–36. Downs, Anthony. 1957. An Economic Theory of Democracy. New Gerber, Alan S., Donald P. Green, and Christopher W. Larimer. York: Harper. 2008. “Social Pressure and Voter Turnout: Evidence from Druckman, James N. 2004. “Political Preference Formaa Large-Scale Field Experiment.” American Political Science tion: Competition, Deliberation, and the (Ir)Relevance of Review 102(1): 33–48. STABILITY OF PRO-SOCIAL PREFERENCES ACROSS DOMAINS 29 Gerber, Alan S., Donald P. Green, and Christopher W. Larimer. Levitt, Steven, and John List. 2007. “What Do Laboratory 2010. “An Experiment Testing the Relative Effectiveness of Experiments Measuring Social Preferences Reveal about Encouraging Voter Participation by Inducing Feelings of the Real World?” Journal of Economic Perspectives 21(2): Pride or Shame.” Political Behavior 32(3): 409–22. 153–74. Gerber, Alan, and Todd Rogers. 2009. “Descriptive Social McCright, Aaron, and Riley Dunlap. 2011. “The PoliticalizaNorms and Motivation to Vote: Everybody’s Voting and So tion of Climate Change and Polarization in the American Should You.” Journal of Politics 71(1): 178–91. Public’s Views of Global Warming, 2001–2010.” Sociological Green, Donald. 2011. “Harnessing Elemental Psychological Quarterly 52(2): 155–94. Forces to Promote Voter Turnout.” Newsletter of the APSA Michelson, Melissa R., Lisa Garcia Bedolla, and Margaret A. Experimental Methods Section 2(2): 2. McConnell. 2009. “Heeding the Call: The Effect of Targeted Green, Donald P., and David W. Nickerson. 2003. “Getting Out Two-Round Phonebanks on Voter Turnout.” Journal of Polthe Vote in Local Elections: Results from Six Door-to-Door itics 71(4): 1549–63. Canvassing Experiments.” Journal of Politics 65(4): 1083–96. Mueller, Dennis C. 1989. Public Choice II. Cambridge: CamGromet, Dena M., Howard Kunreuther, and Richard P. Larrick. bridge University Press. Forthcoming. “Political Ideology Affects Energy-Efficient Nickerson, David. 2008. “Is Voting Contagious? Evidence from Attitudes and Choices.” Proceedings of the National Academy Two Field Experiments.” American Political Science Review of Sciences. 102(1): 49–57. Grossman, Guy. 2011. “Lab-in-the-Field Experiments.” NewsOlson, Mancur. 1965. The Logic of Collective Action: Public Goods letter of the APSA Experimental Section 2(2): 13. and the Theory of Groups. Cambridge, MA: Harvard UniverGuber, Deborah Lynn. 2013. “A Cooling for Climate Change? sity Press. Party Polarization and the Politics of Global Warming.” Ostrom, Elinor. 2006. “The Value-Added of Laboratory ExAmerican Behavior Scientist 57(1): 93–115. periments for the Study of Institutions and Common-Pool Hamman, John R., Roberto A. Weber, and Jonathan Woon. Resources.” Journal of Economic Behavior & Organization 2011. “An Experimental Investigation of Electoral Delega61(2): 149–63. tion and the Provision of Public Goods.” American Journal Pew Research Center. 2009. “Independents Take Center Stage of Political Science 55(4): 738–52. in Obama Era: Trends in Political Values and Core AttiHuckfeldt, Robert, and John Sprague. 1992. “Political Parties tudes: 1987–2009.” May 21. http://www.people-press.org/ and Electoral Mobilization: Political Structure, Social Struc2009/05/21/section-9-the-environment-and-the-economy/ ture, and the Party Canvass.” American Political Science RePutnam, Robert D. 1966. “Political Attitudes and the Loview 86(1): 70–86. cal Community.” American Political Science Review 60(3): Ikeda, Ken’ichi, and Sean E. Richey. 2005. “Japanese Network 640–54. Capital: The Impact of Social Networks on Japanese Political Rand, David G., Thomas Pfeiffer, Anna Dreber, Rachel W. Participation.” Political Behavior 27(3): 239–60. Sheketoff, Nils C. Wernerfelt, and Yochai Benkler. 2009. Imai, Kosuke, and Aaron Strauss. 2011. “Estimation of Hetero“Dynamic Remodeling of In-Group Bias during the 2008 geneous Treatment Effects from Randomized Experiments, Presidential Election.” Proceedings of the National Academy with Application to the Optimal Planning of the Get-Outof Sciences 106(15): 6187–91. the-Vote Campaign.” Political Analysis 19(1): 1–19. Riker, William, and Peter Ordeshook. 1968. “A Theory of the Jankowski, Richard. 2002. “Buying a Lottery Ticket to Help the Calculus of Voting.” American Political Science Review 62(1): Poor: Altruism, Civic Duty, and Self-Interest in the Decision 25–41. to Vote.” Rationality and Society 14(1): 55–77. Sandroni, Alvaro, and Timothy Fedderson. 2006. “A Theory Jankowski, Richard. 2007. “Altruism and the Decision to Vote: of Participation in Elections.” American Economic Review Explaining and Testing High Voter Turnout.” Rationality 96(4): 1271–82. and Society 19(10): 5–34. Schultz, Wesley P. 1999. “Changing Behavior with NormaJerit, Jennifer, Jason Barabas, and Scott Clifford. 2013. “Comtive Feedback Interventions: A Field Experiment on Curbparing Contemporaneous Laboratory Field Experiments on side Recycling.” Basic and Applied Social Psychology 21(1): Media Effects.” Public Opinion Quarterly 77(1): 256–82. 25–36. Karlan, Dean S. 2005. “Using Experimental Economics to Shang, Jen, and Rachel Croson. 2009. “Field Experiments in Measure Social Capital and Predict Financial Decisions.” Charitable Contribution: The Impact of Social Influence on American Economic Review 95(5): 1688–99. the Voluntary Provision of Public Goods.” Economic Journal 119: 429–34. Knack, Stephen. 1992. “Civic Norms, Social Sanctions, and Voter Turnout.” Rationality and Society 4(1): 133–56. Thogersen, John. 2008. “Social Norms and Cooperation in RealLife Social Dilemmas.” Journal of Economic Psychology 29(4): Knack, Stephen, and Martha Kropf. 1998. “For Shame! 458–72. The Effect of Community Cooperative Context on the Probability of Voting.” Political Psychology 19(3): 585– Tversky, Amos, and Daniel Kahneman. 1981. “The Framing of 99. Decisions and the Psychology of Choice.” Science 211(4481): 453–58. Kropf, Martha. 2009. “Won’t You Be My Neighbor? Norms of Cooperation, Public Broadcasting, and the Collective Voors, M. J., T. Turley, A. Kontoleon, E. H. Bulte, and J. List. Action Problem.” Social Science Quarterly 90(3): 538– 2012. “Exploring Whether Behavior in Context-Free Exper52. iments Is Predictive of Behavior in the Field: Evidence from 30 TOBY BOLSEN, PAUL J. FERRARO, AND JUAN JOSE MIRANDA Lab and Field Experiments in Rural Sierra Leone.” Economic balance across treatment arms within voting subgroups Letters 114(3): 308–11. on outcome variable] Zaller, John. 1992. The Nature and Origins of Mass Opinion. Appendix 4. Mean Water Consumption on Summer New York: Cambridge University Press. 2006 by Partisanship and Treatment [to demonstrate pretreatment balance across treatment arms within partisan subgroups on outcome variable] Supporting Information Appendix 5. Figure SI.1 Experimental Treatment Response by Voting Frequency Quartile using Unconditional Data (i.e., not regression-adjusted) Additional Supporting Information may be found in the Appendix 6. Linear Regression – Water Use June through online version of this article at the publisher’s website: September 2007 (Cross-Sectional) [using alternative ways Appendix 1. Rival Explanations and Robustness Checks of measuring voting frequency] Appendix 2. Treatment Messages Appendix 7. Linear Regression – Partisanship and WaTreatment 1: tip-sheet only ter Use June through September 2007 (Cross Sectional) Treatment 2: tip-sheet and Treatment 2 letter [using alternative ways of measuring voting frequency] Treatment 3: tip-sheet and Treatment 3 letter Appendix 8. Linear Regression – Water Use Partisanship Appendix 3. Mean Water Consumption on Summer 2006 (Panel Data, Monthly) [using alternative ways of measurby Deciles and Treatment [to demonstrate pre-treatment ing voting frequency]
4 Black Politicians Are More Intrinsically Motivated to Advance Blacks’ Interests: A Field Experiment Manipulating Political Incentives David E. Broockman University of California, Berkeley Why are politicians more likely to advance the interests of those of their race? I present a field experiment demonstrating that black politicians are more intrinsically motivated to advance blacks’ interests than are their counterparts. Guided by elite interviews, I emailed 6,928 U.S. state legislators from a putatively black alias asking for help signing up for state unemployment benefits. Crucially, I varied the legislators’ political incentive to respond by randomizing whether the sender purported to live within or far from each legislator’s district. While nonblack legislators were markedly less likely to respond when their political incentives to do so were diminished, black legislators typically continued to respond even when doing so promised little political reward. Black legislators thus appear substantially more intrinsically motivated to advance blacks’ interests. As political decision making is often difficult for voters to observe, intrinsically motivated descriptive representatives play a crucial role in advancing minorities’ political interests. “What’s the proof that I care about the black However, there have long been two compelling excommunity? Every time I get a letter from a planations for this robust relationship between descripblack person outside my district, I respond.” tive and substantive political representation. On the —Anonymous Black State Legislator one hand, this link is often attributed to purely intrinsic motivations politicians are said to have to aid P oliticians are reliably more likely to advance the those like them due to feelings of group identification interests of those who share their personal charor shared personal preferences (e.g., Mansbridge 1999, acteristics, including their gender, race, profes2003; Whitby 1997). On the other, this link is often prision, class, and sexual orientation (e.g., Burden 2007; see marily attributed to politicians’ differing electoral innext section for review); whether measured by roll-call centives (e.g., Canon 1999; Grose 2011, 30–37)—for votes, responsiveness to constituent requests, or policy example, black politicians who expect difficulty winoutcomes, scholars reliably find that descriptive reprening support from white voters might advance blacks’ sentatives provide greater substantive representation to interests to a greater extent merely out of a strategic constituents like them. calculus.1 David E. Broockman is a Graduate Student in Political Science, University of California, Berkeley, 210 Barrows Hall, Berkeley, CA 94720 (broockman@berkeley.edu). I thank John Bullock, Dan Butler, Ryan Enos, Sean Gailmard, Don Green, Zoltan Hajnal, Greg Huber, Gabe Lenz, Todd Rogers, Eric Schickler, Chris Skovron, Laura Stoker, Jessica Suits, Rob van Houweling, and the anonymous reviewers for helpful feedback. Eleanor Powell deserves special thanks for her guidance. The interviews discussed were conducted in June–August 2010 during participant observation research with 27 state legislators (14 white, 13 black) who at the time represented majority-minority state legislative districts. Interest group employees and legislative black caucus leaders also provided invaluable advice, information, and assistance during this period. I owe great thanks to these legislators for inviting me into their districts and their homes. I also owe special thanks to Martha Grant, April Lawson, the Bills family, Will and Nicola Wilson, Sara Meacham, and members of Couchsurfing.org for making this participant observation work possible. The Yale College Dean’s Office Fellowship in the Social Sciences, the Yale Political Science Department’s Frank M. Patterson Grant, and the Jonathan Edwards College Richter Travel Fellowship were all generous in their financial support for these activities. I also acknowledge financial support from the National Science Foundation Graduate Research Fellowship Program. The Yale Institution for Social and Policy Studies and Dan Butler also supported the collection and purchase of the datasets I employ. Replication data are available at the AJPS Dataverse. 1 Much work, including Whitby (1997) and Grose (2011), acknowledges both potential mechanisms. American Journal of Political Science, Vol. 57, No. 3, July 2013, Pp. 521–536 \002 C 2013, Midwest Political Science Association DOI: 10.1111/ajps.12018 521 522 DAVID E. BROOCKMAN Both these theories yield predictions fully consistent of political representatives’ personal backgrounds more with scholars’ findings; indeed, they typically yield the generally. same predictions. Yet this observational equivalence in existing data leaves open significant questions for the design of representative institutions: to what extent are Are Politicians More Intrinsically strong relationships of electoral accountability necessary Motivated to Represent Their Group? if politicians are to be expected to advance their group’s interests? Or, alternatively, to what extent are political Over the last three decades, an impressive body of evirepresentatives intrinsically motivated to advance their dence has demonstrated that politicians provide greater group members’ interests even when they do not have substantive political representation to those who share strong political incentives to do so? their personal characteristics, including their gender, race, I present a field experiment that sheds light on these profession, class, and sexual orientation. Much of this litquestions by granting a rare look at how politicians behave erature has shown that legislators are more responsive when their political incentives to advance their group’s inin their roll-call voting to the interests and preferences terests are significantly reduced. In the experiment, which of those who share their personal characteristics (e.g., draws on interviews conducted with legislators who repBurden 2007; Canon 1999; Carnes 2012; Grose 2005; resent black-minority districts, I presented state legislaHutchings 1998; Tate 2003; Whitby 1997).2 Other work tors with an ostensible opportunity to improve the welfare shows that this pattern also holds for committee behavior of a black person but varied the degree of political incen(Gamble 2007; Minta 2009, 2011) and the provision of tive they had to do so. Specifically, I emailed all 6,928 U.S. constituency service (Butler and Broockman 2011; Grose state legislators asking for help filing for unemployment 2011). And crucially, still further research demonstrates benefits. All the emails came from an ostensibly black that these differences in policymaking behavior also have alias, Tyrone Washington. I then randomized the legisreal effects on substantive outcomes for members of these lators into two conditions: in one condition the sender groups (e.g., Bratton and Ray 2002; Chattopadhyay and claimed to live in cities in the legislators’ districts, while Duflo 2004; Haider-Markel 2007; Haider-Markel, Joslyn, in the other condition the sender ostensibly resided in and Kniss 2000; Meier and England 1984; Pande 2003; cities far outside their districts. Preuhs 2006).3 This treatment was intended to vary the degree of Though differences between black and nonblack political incentive legislators had to respond to the putapoliticians are the most studied, much work has shown tively black individual’s request, and the results indicate that this link holds for other groups as well (e.g., Carnes that it was successful in doing so: legislators were half as 2012). Burden (2007), for example, even finds that legislikely to respond to the sender who claimed to live in a lators who smoke are more likely to vote consistently with city outside their district. However, black and nonblack the tobacco industry’s preferences. legislators did not react equally to this lessening of their incentives: black legislators were far more likely to conElites’ Extrinsic and Intrinsic Motivations tinue responding to the request from the out-of-district as Mechanisms individual than were their nonblack peers. This large difference between how black and nonMembers of many groups have thus been consistently black legislators reacted to a lessening of their electoral shown to fare better when their fellow group members incentives provides unique and strong support for the serve in government. However, why this link holds is both hypothesis that black legislators are significantly more unclear from existing evidence and crucially important intrinsically motivated to advance blacks’ interests. Imto designing representative institutions. portantly, black legislators thus appear to be significantly On the one hand, scholars often assert that the link more likely than their counterparts to work to advance between descriptive and substantive representation is blacks’ interests in contexts where political incentives are driven by politicians’ own personal intrinsic motivation weak and, as has long been theorized, to therefore en2 gage in significant “virtual” or “surrogate” representation See also Kingdon (1981), Carroll (1994), Hood and Morris (1998), (Mansbridge 2003; Pitkin 1967) of fellow group members Cobb and Jenkins (2001), Griffin and Newman (2007), Grose, Mangum, and Martin (2007), and Juenke and Preuhs (2012). beyond the bounds of their official constituencies. These 3 findings have significant implications for the study of See also Nye, Rainer, and Stratmann (2011), Shayo and Zussman (2011), Iyer et al. (2011), and Franck and Rainer (2012). This large descriptive representation, the design of institutions for literature is also marked by remarkably few dissenting studies (e.g., facilitating minorities’ voice in government, and the role Swain 1993). BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 523 to promote the welfare of their group or enact the prefDefining these claims formally can help elucidate erences held by their group (e.g., Burden 2007). I follow these theories’ observational equivalence in existing data Ariely, Bracha, and Meier in defining such “intrinsic modespite their differing substantive implications. For ease tivation” as “private preferences for others’ wellbeing” of reading and consistency with the empirical sections, (2009, 544). In this sense, a politician motivated to adI use blacks and nonblacks as an example, though the vance the interests of his or her group for intrinsic reaframework could generalize to any setting where legissons alone does so “purely [for the] psychic . . . benefits” lators have the opportunity to improve the welfare of a involved (Citrin and Green 1990, 6). For example, black person or group. politicians are often theorized to work to advance other Suppose when deciding whether to advance blacks’ blacks’ interests due to a sense of group solidarity with interests in some form (e.g., by answering a letter from other blacks (Whitby 1997).4 More generally, theories of a black person, or by casting a supportive roll-call vote representation have long supposed that legislators’ peron a civil rights issue), a legislator expects to gain both sonal preferences and backgrounds importantly influence benefits that aid her in winning reelection, which I call their behavior (e.g., Burden 2007; Miller and Stokes 1963). extrinsic benefits and denote E, and psychic benefits that On the other hand, it has also been widely suggested are completely separate from expected tangible rewards, that legislators advance their groups’ interests to a greater intrinsic benefits denoted I. A politician’s expected utilextent because they face stronger political incentives to ity for performing an act that advances blacks’ interests do so (e.g., Whitby 1997, 85). The logic for this claim is thus given in total by the accounting identity EU = is straightforward. For example, consider Chattopadhyay E + I;5 it is determined by motivations that are extrinsic and Duflo’s (2004) analysis of female legislators in India, and intrinsic, depending on to what extent they either who they demonstrate are markedly more likely to supfacilitate the receipt of some outside reward (E) or confer port policies that benefit women. These female legislators’ psychic benefits (I). behavior might result from their greater personal concern The thesis that descriptive representation leads to for fellow women’s welfare (intrinsic motivation), but it greater substantive representation for blacks holds that, may also be that female legislators’ best reelection strategy all else equal, black legislators place higher overall value is to focus on securing support with female voters because on advancing blacks’ interests than nonblack legislators they expect winning male support will be more diffido. To denote this counterfactual formally, let EB and EW cult in a sexist society (e.g., Fowler and McClure 1989). respectively refer to the extrinsic political rewards black Likewise, black politicians may be expected to focus on and nonblack legislators expect to receive from performwinning blacks’ votes because they anticipate comparaing an identical act to advance blacks’ interests and IB and tively greater difficulty winning whites’ (e.g., Grose 2011, IW to their respective expected intrinsic rewards for the 30–37). Politicians’ strategic pursuit of reelection could same. thus perhaps fully explain why they tend to advance their In these terms, existing evidence regarding the link group members’ interests to a greater extent. between descriptive and substantive representation shows us that black politicians are more motivated to advance blacks’ interests overall, EB + IB > EW + IW , though this Why Intrinsic Motivation Has Been could be for intrinsic or extrinsic reasons. In turn, the inDifficult to Identify and Why It Matters trinsic motivation hypothesis explains this statement with the more specific claim that black legislators have greater Why have scholars had difficulty distinguishing between intrinsic motivation to advance blacks’ interests than do these explanations, and why does it matter that we do? nonblack legislators, IB > IW , while the extrinsic motivation hypothesis claims that black legislators perceive 4 greater extrinsic incentives to advance blacks’ interests, Some of this research uses the term “linked fate,” a concept typically associated with a sense of shared destiny among blacks in the EB > EW . mass public (e.g., Dawson 1994; Gay 2004). I choose not to emCrucially, note that the observation that black politiploy the term because it is sometimes used to refer to a somewhat cians advance blacks’ interests to a greater extent (EB + self-interested motivation group members have for advancing their group’s interests because they believe their group’s fate is linked to IB > EW + IW ) does not directly imply that black polititheir own. This mechanism is sometimes hypothesized to underlie cians are more intrinsically motivated to do so, IB > IW , elites’ behavior, though the explanation I seek to investigate in this article is a strictly altruistic one. (Interested readers are directed 5 to McClain et al. 2009 for a thorough review of the distinctions Assuming that nonpolitical costs (like the time it takes to answer between various concepts related to group membership, identity, a letter) are identical across legislators for identical acts and so can consciousness, and linked fate.) be ignored for the sake of parsimony. 524 DAVID E. BROOCKMAN unless one makes assumptions about black and nonblack However, we can still attempt to extrapolate how politicians’ political payoffs, EW and EB : black legislapoliticians would behave if they faced no political intors clearly find greater reason to vote liberally on civil centives by observing what happens when their political rights issues, but this observation alone does not tell us if incentives change in magnitude. Though no scholar has they do so out of a personal commitment to these issues formally defended this practice, the literature on race and or a perceived strategic imperative to win blacks’ votes. politics contains notable examples of this logic. HutchThis observational equivalence problem likely extends to ings (1998), for example, considers congresspeople’s votes all the activities scholars have measured; indeed, roll-call on the Civil Rights Act of 1990. He finds that southvotes, constituency service provision, constituency comern Democrats were highly responsive to constituency munication, and committee behavior—the key variables pressures on the well-publicized final passage of the bill, in most studies of descriptive representation—have all though far less responsive on an important amendment long been recognized to be central to legislators’ efforts with lower public salience. Hutchings interprets this patto pursue reelection (e.g., Mayhew 1974). Political incentern as indicating that legislators act more in line with tives alone could thus explain patterns of behavior such their extrinsic incentives when they believe they are unas that observed by Butler and Broockman (2011) that der great scrutiny but that less-scrutinized behavior rewhite and black Democrats are both more responsive to veals something about their true preferences. Similarly, inquiries from copartisan constituents of their race. Minta (2009, 2011) analyzes the degree to which legisYet intrinsic motivation matters precisely because lators participate in oversight hearings, which he argues politicians are often not strongly incentivized to perform they expect few political rewards for doing and therefore a variety of acts that still matter a great deal: they make implies that they are intrinsically motivated. many decisions concerning groups that pay little attenA black legislator I interviewed best laid out the tion to their actions (Delli Carpini and Keeter 1996), in intuition for this reasoning (in so doing inspiring the contexts that are difficult (or impossible) for voters to experiment to be described) when I pressed him for evobserve (Arnold 2009; Hall 1996), and concerning the idence that he was more intrinsically motivated to adinterests of those beyond the boundaries of their convance black’s interests than his nonblack colleagues. He stituencies (Mansbridge 2003). The decisions politicians responded: make when their political incentives are weak clearly have “What’s the proof that I care about the black comconsequences, yet we cannot reliably infer what they do in munity? Every time I get a letter from a black person such settings from what they do in salient public behavior. outside my district, I respond. The white [legislators] don’t do that.” Intuition for Identification Strategy To express these scholars’ and this legislator’s intuition formally, suppose  is a term bounded by 0 ≤  As the above discussion suggests, researchers would thus that captures the degree to which a legislator perceives ideally identify politicians’ intrinsic motivation by oba political incentive to perform a particular act—that is, serving how they act when they have no political incenthe degree to which a legislator’s expected utility might tives whatsoever.6 We might refer to this intuition as the be affected by voters or other actors capable of exercising fact that a politician’s intrinsic motivation would be laid political accountability rewarding (or punishing) him or bare when he “thinks nobody is watching.”7 Of course, her for a particular act. Higher values of  would correexisting literature has had difficulty pinning down intrinspond to situations where accountability is greater, such sic motivation precisely because there are few observable as, following from Hutching (1998), a roll-call vote on environments where politicians are likely to believe they a very salient bill. In turn, low values of  might be asface no political consequences for their actions. sociated with acts like asking questions during oversight 6 hearings (Minta 2009, 2011) or engaging in closed-door As we cannot apply the standard approach of inducing or exploiting variation in intrinsic motivation in order to directly observe its negotiations with other legislators. effects. (Potentially see Washington 2008 for such a design, though Taking  into account, legislators’ expected utility note that having a daughter might affect legislators’ own personal would be described by EU = E + I, such that legbenefit from various policy outcomes or degree of knowledge about islators expect fewer extrinsic benefits from advancing their likely effects, not just their degree of intrinsic motivation to advance women’s interests.) their groups’ interests in situations where  is lower. Even 7 though we can never really observe what politicians do One might likewise recall the proverb (attributed to Malcolm Forbes) that one can “judge the character of a man by how he treats when they “believe no one is watching” (i.e., when  = 0), those who can do nothing for him.” we thus still might extrapolate what might occur when BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 525  = 0 (and thus make inferences about I alone) BOX 1 Text of Email Sent to State by comparing how politicians’ behavior changes when Legislators  changes—for example, following Hutchings’s (1998) logic, when the salience of a roll-call vote declines.8 The From: Tyrone Washington black state legislator quoted above appealed to this very Subject: unemployment benefits in [TREATMENT logic: he claimed to be less sensitive than his white peers CITY NAME] to whether he will receive political rewards for advancing Text: Dear [Mr./Ms.] [STATE REPRESENTATIVE’S blacks’ interests, which he argued was evidence that his LAST NAME], motivations must be more intrinsic than theirs.9 My name’s Tyrone Washington and I live in [TREATWatching how politicians change their behavior when MENT CITY NAME]. Can you tell me how to get untheir political incentives are reduced can thus help us employment benefits? I lost my job but nobody will tell infer their degree of intrinsic motivation, an important me where to get them and I don’t know what to do. task precisely because politicians often act in situations Thank you where their political incentives are weak. The next section Tyrone Washington applies this intuition to the experimental design. Experimental Design within each legislator’s district (e.g., for a legislator representing Dallas, Texas, “Dallas”) and a well-known city Consistent with the intuition developed above, I delocated far outside their district but within their state signed and implemented an experiment attempting to (e.g., for a Dallas-based legislator, “Houston”). I then grant a rare look at how politicians change their behavior randomly assigned each legislator to the inand out-ofwhen their electoral incentives are significantly weakened. district treatment groups with block randomization on Specifically, drawing on my interview with the black state state, party, and race. The online supporting information legislator quoted above, I sent every state legislator in the details a randomization check that indicates this was sucUnited States serving in mid-November 2010 (N = 6,928) cessful, shows a map of the out-of-district cities I used for an email asking for help enrolling in state unemployment each state, and fully elaborates the technical details of the benefits. The emails all came from the alias Tyrone Washcity assignment process. ington, which strongly signals being black.10 The text of The names of the cities appeared in the subject and the email appears in Box 1. the first line of the email’s text, as shown in the “treatment Crucially, Tyrone purported to live in, randomly, eicity name” field in Box 1. Each legislator only received one ther a city in each legislator’s district or a city far from email. each legislator’s district. This manipulation was designed I chose to ask about signing up for unemployment to vary legislators’ incentives to respond,  in the framebenefits for three reasons. First, the objective interests work developed previously, consistent with the intuition of the letter’s sender were clear: receiving a response developed by the black state legislator quoted. would improve his welfare. Rather than measuring To implement this manipulation, I first assigned each how responsive legislators are to constituency service legislator the names of two cities using ArcGIS: a city requests in a general sense, therefore, the experiment 8 See Lee, Moretti, and Butler (2004), Gailmard and Jenkins (2009), putatively presents politicians with an opportunity Snyder and Stromberg (2010), and van Houweling (2012) for exto tangibly improve the welfare of a black person in amples of similar logic being applied elsewhere. However, note that their state though at a cost of time and effort, one of actors besides voters (e.g., interest groups) may succeed in altering politicians’ incentives for action under the conditions these works the very same opportunities they have when deciding analyze. whether to advance blacks’ interests in the context of 9 Likewise, another black legislator interviewed accused a white making policy (Hall 1996). Choosing unemployment colleague who represents a black district of not being “sincere” in benefits (along with the letters’ errors in grammar his representation of his black constituents, offering as evidence his and diction) also minimized the political benefits weaker efforts on his constituents’ behalf behind closed doors. legislators might have perceived from answering the 10 Essentially no whites are named Tyrone (Fryer and Levitt 2004), out-of-district email. Last, because requests for help and 89.9% of those with the last name Washington are black (Word filing for unemployment benefits are so commonplace, I et al. n.d.). Though it is not certain that every legislator thought Tyrone Washington was black, any patterns to the contrary would ensured that the costs associated with answering the letter bias the results toward zero. would be both relatively equal between legislators and, 526 DAVID E. BROOCKMAN TABLE 1 Formal Summary of Identification Strategy In-District Letters Out-of-District Letters Difference Black Legislators B + C EB + IB B + T EB + IB (C T )EB Nonblack Legislators C EW + IW T EW + IW (C T )EW Difference-in-Differences: (EB EW )(C T ) consistent with ethical considerations (see below), If this proves to be the case, we can then draw inferrelatively low. ences about politicians’ intrinsic motivation, I, so long as one final assumption is satisfied. As this article seeks to understand a mechanism for the link between descriptive Identification Strategy and substantive representation, I assume that the extensive literature on descriptive representation is correct that Legislators thus all received a letter from a putatively black black legislators are at least as interested in advancing individual who purported to live either within or far from blacks’ interests (for a sum of potentially extrinsic and intheir district. Table 1 summarizes how this manipulation trinsic reasons) as are nonblack legislators, all else equal. shed light on legislators’ degree of extrinsic and intrinsic With it established that black legislators are not as motivation for responding. The entries in Table 1 refer to motivated by the receipt of extrinsic rewards for respondthe determinants of legislators’ response rates consistent ing to a black person as are their nonblack colleagues, yet with the expected utility function described previously. also that black legislators are at least as motivated as their C refers to the degree that responding is electorally inwhite colleagues to do so overall, it then must be that they centivized in the condition where they received a letter do so because they have greater intrinsic motivation to do from a putative constituent and T when they received so. Formally, if EB + IB ≥ EW + IW (black legislators are a letter from a putative nonconstituent. The difference at least as interested in advancing blacks’ interests as are C – T thus captures the degree to which the treatnonblack legislators) and EW > EB (nonblack legislators ment lessened legislators’ expected political rewards for are more animated by political incentives than black legresponding between conditions. islators when representing blacks, which the experiment Though I wish to infer whether black politicians can establish), it must be that IB > IW (intrinsic motivaare more intrinsically motivated to advance blacks’ intion to represent blacks is greater among black legislators terests than are nonblacks (IB > IW ), as discussed, this than nonblack legislators). is not straightforward: intrinsic motivation to improve In sum, consistent with the intuition described by the the welfare of blacks, denoted I, is equally present in all black legislator quoted above, if black legislators engage conditions, as is some degree of extrinsic motivation, E. in greater “surrogate representation” (Mansbridge 2003) Moreover, there is also selection bias, denoted B, that by responding to letters from blacks that their nonblack prevents inferences based on direct comparisons of how counterparts would not have when their expected political likely black and nonblack legislators are to respond to the rewards are reduced, this implies that black legislators’ out-of-district letter. intrinsic motivation to advance blacks’ interests is greater. However, as the interviewed state legislator suggested, the experiment allows for an indirect comparison of legislators’ intrinsic motivation, I, as follows. First, in words, Ethical Considerations note that if nonblack legislators are more likely to cease responding because their extrinsic incentives decrease, Though the human subjects committee approved this this would suggest that their extrinsic motivations comexperiment, several ethical concerns were still important pose a larger share of their reasons for responding when to consider in its design that any future work employing they do. Formally, as shown in Table 1, the experimenthis approach should carefully consider as well. tal difference-in-differences estimates the quantity (EW – First, the experiment was designed so as to place as EB )(C – T ). A positive value of this coefficient would minimal a burden on legislators’ time as possible: the resuggest that nonblack legislators advance blacks’ interests quest made was selected to be commonplace and hence for reasons that are to a greater extent extrinsic than are easy for legislators to answer. From the replies, it appears black legislators’ (assuming C – T , the relative incentive that this was successful: the median reply was 298 characfor responding to a constituent, is greater than zero). ters, or about 50 words, less than a third of the length of BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 527 this article’s abstract. As Hall (1996) notes, investigating Some of these observations were excluded from the how legislators choose to spend their time is an important dataset before the analysis. First, 297 observations were way to learn about their priorities. However, researchers dropped as the emails immediately bounced as undelivalso have a clear obligation to minimize the burden they erable because the addresses were reported incorrectly on impose on others, and future work should continue to the legislative websites or were entered with typographitake this concern seriously. cal errors. Second, when the experiment ran, a very small Next, deception was largely unavoidable in this exnumber of legislators carbon copied their replies to the periment in order to determine how legislators would relegislators who actually represented the cities with names spond differently to people they believed lived inside and I used to construct the out-of-district city treatment. For outside of their districts.11 However, the ethical status of example, as a courtesy, a legislator who represented Fort deception in the many audit experiments academics have Worth, Texas, replied to an email that claimed to come conducted is the subject of continuing debate (e.g., Riach from “Tyrone in Houston” with a carbon copy to a legand Rich 2004), and researchers should employ deception islator who represents Houston. However, the Houston with great care and avoid its use when it is unnecessary. legislators’ office was also part of the experiment and In addition, to minimize any harm that might come thus had received their own copy of the letter. I therefore to subjects, the replication data will not include inforexcluded all 713 legislators like this Houston legislator mation that could identify the legislators. Especially with who might have received letters via their colleagues by politicians, it is important to keep in mind that others virtue of representing one of the 100 places with names might use data on whether they responded to harm them. I used to construct the treatments. Finally, all remaining Even though as researchers we appreciate that we cannot 334 legislators in Georgia and Indiana were excluded bemake inferences about why any one particular legislator cause legislators in these states share staff, rendering it did or did not respond, others may not appreciate this unclear to which legislator the replies should be credited. limitation. These criteria removed 19.3% of the sample, resulting Last, as McClendon (2012) points out in a thoughtful in 5,593 usable observations. However, all of the experidiscussion of the ethics of experimenting on public offiment’s results remain the same when the excluded obsercials, such research also has a potential cost for other revations remain, and the supporting information shows searchers because these officials control research budgets. that these criteria were uncorrelated with treatment asElected officials have so far not shown signs of reacting signment. There were 4,965 white legislators, 364 black negatively to this work in debates over support for politlegislators, and 264 legislators from other racial groups in ical science and themselves also regularly authorize and the final dataset. fund audit trials of their own agencies and the private market (see, e.g., Fix and Turner 1998, chap. 6; Yinger 1998). However, researchers should remain carefully atExperimental Results tuned and sensitive to this concern. I received 2,365 replies to these 5,593 emails in total, an overall response rate of 42.3%. Following Butler and Data on Legislators and Excluded Broockman (2011), the analysis employs this objective Observations dependent variable of whether I received a reply from the To implement the experiment, I first gathered data on leglegislator at all. However, I also collected data on whether islators’ races and email addresses in mid-2010 from pubor not the replies were helpful (which 87.6% were) and lic state legislative websites for all 50 states. This yielded found that all the results held when using this alternative a sample of 6,928 legislators in total. The data also independent variable. These results and the criteria used for cluded covariates about the legislators’ districts, including coding “helpful” responses are available in the supporting their total population, the percentage of the districts that information. were black, the rural and urban makeup of the districts, median household incomes for blacks and whites in the Legislators Less Likely to Respond to district, and the Squire (2007) index of state legislative Out-of-District Emails professionalism. The experiment relies on the assumption that legislators found less political reason to respond to requests osten11 sibly from a person living hundreds of miles away from However, see recent attempts to reduce deception in audit studies by cooperating with real constituents (Butler, Karpowitz, and Pope their district than requests which purport to be from a 2012; Carnes and Holbein 2012). constituent living in their district (i.e., that C > T ). 528 DAVID E. BROOCKMAN TABLE 2 Experimental Treatment Effects (4, (5, Logistic OLS, CEM Specification (1, OLS) (2, OLS) (3, OLS) Regression) Matched/Weighted) Experimental Treatment Effects Out-of-District Email −0.266∗∗ −0.275∗∗ −0.276∗∗ −1.185∗∗ −0.311∗∗ (0.013) (0.013) (0.013) (0.059) (0.012) Out-of-District Email x Black Legislator 0.128∗ 0.128∗ 0.552∗ 0.161∗∗ (0.052) (0.051) (0.227) (0.046) Covariates Black Legislator −0.097∗∗ −0.112∗ −0.462∗ −0.085∗∗ (0.036) (0.045) (0.197) (0.033) Other Nonblack Minority Legislator −0.035 −0.162 0.018 (0.031) (0.140) (0.041) Democratic Legislator −0.051∗∗ −0.232∗∗ 0.088∗∗ (0.014) (0.061) (0.022) State Senator 0.089∗∗ 0.398∗∗ 0.191∗∗ (0.016) (0.072) (0.016) South −0.004 −0.018 −0.041∗ (0.017) (0.076) (0.019) Black Population Percent 0.084 0.377 −0.039∗∗ (0.067) (0.298) (0.036) Black Median HH Income ($10,000s) −0.000 −0.001 0.069∗∗ (0.007) (0.032) (0.019) White Median HH Income ($10,000s) 0.021∗ 0.093∗ −0.113∗∗ (0.010) (0.045) (0.015) Squire Index 0.489∗∗ 2.183∗∗ −0.036∗∗ (0.071) (0.322) (0.105) District Total Population (10,000s) −0.004∗∗ −0.016∗∗ 0.006 (0.001) (0.004) (0.001) Urban Percent 0.014 0.065 0.232∗∗ (0.023) (0.106) (0.023) Constant 0.555∗∗ 0.561∗∗ 0.442∗∗ −0.281∗∗ 0.584∗∗ (0.009) (0.009) (0.025) (0.113) (0.043) R2 .073 .074 .091 .240 N 5593 5593 5593 5593 5125 Note: Dependent variable in all regressions is whether the legislator responded to the email. ∗ p < .05, ∗∗ p < .01 (two-tailed tests). The data strongly validate this assumption: overall, legisBlack Legislators Are Less Sensitive Than lators were 26.6 percentage points less likely to respond Nonblacks to Their Political Incentives for to emails in the out-of-district treatment (p < .0001; Responding all p-values two-tailed). Legislators assigned to the indistrict group responded to 55.5% of emails, whereas Though all legislators were less responsive when their about half that number responded to the out-of-district political incentives were decreased, how did the effect emails, or only 28.9%. This number was surprisingly high of decreasing legislators’ political incentives vary beand suggests that the baseline level of intrinsic motivatween legislators? The data strongly reject the null hytion may not be trivial, yet it also shows that legislators pothesis that black and nonblack legislators were equally are highly responsive to their electoral incentives, as exconcerned with the political rewards they might receive pected. These results are reported in the first column of from responding to the letter: all else equal, nonblacks Table 2. were much less likely to respond to the out-of-district BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 529 FIGURE 1 Rates of Reply by Treatment Group and Legislators’ Race (Matched Dataset) letters than were blacks. Column 2 shows that nonblack are simply “incomparable” to blacks (i.e., King and Zeng legislators responded to the out-of-district emails 27.5 2006), because legislators from districts with few blacks percentage points less frequently (p < .001). However, might have been suspicious of a letter from a putatively black legislators responded to the out-of-district emails black individual or less knowledgeable about unemployonly 14.7 percentage points less frequently, a treatment ment benefits, and because the most policy-relevant difeffect 12.8 percentage points lower (p < .01) than that ferences occur where black legislators could plausibly be of their nonblack counterparts. The results thus imply elected. that about half of the nonblack legislators who did not Therefore, I also present the results after using coarsrespond to the email because it came from outside their ened exact matching (Iacus, King, and Porro 2012) to district would have responded if they were black (after improve balance between the districts with and without accounting for selection bias). black legislators by matching on the districts’ population This result remains robust to a variety of controls percentage black, the median household income in the in column 3, which includes covariates for the legisladistrict, and the legislators’ party. The procedure comtors’ race, party, whether the legislator is a state senator, pletely removed 467 observations from the dataset and whether the legislator is from the south, the black popgreatly reduced the statistical weight placed on an adulation of the district, the district black and white popditional 4,409 observations; 717 observations—378 deulations’ median household incomes, the Squire (2007) scribing nonblack legislators and 339 describing black index of state legislative professionalism, the district’s tolegislators—were identified as good matches. tal population, and the percent of the district which is Despite this thorough narrowing of the dataset, the urban. The fourth column of Table 2 also shows that the difference between black and nonblack legislators’ beresults hold just as strongly with logistic regression. havior when their incentives were weakened remained large and significant; in fact, it was even larger among Matched Observations. A weakness with the above analthese observations. The top four bars in Figure 1 ysis is that many nonblack legislators represent districts depict the rates of reply among black and nonblack legwhere blacks would almost never be elected or are Repubislators in each treatment group in the matched data; licans, though essentially no black legislators are. Likecolumn 5 of Table 2 reports these results with controls. wise, though legislators’ treatment condition is randomly As Figure 1 makes clear, black and nonblack legislators assigned, their race is not. responded similarly to the in-district letters, yet nonblack These differences may be problematic because some legislators were much less responsive once their political legislators’ circumstances may be so different that they incentives are diminished. This difference-in-differences 530 DAVID E. BROOCKMAN of 18.5 percentage points between these groups’ treatment concern to some extent (see the third section), though effects is highly significant (p < .001); the treatment eftwo additional analyses can do so further. First, column fect among nonblack legislators in the weighted sample 1 of Table 3 shows that blacks’ responsiveness to the outis more than double the size of the effect among black of-district letter was not negatively moderated by the dislegislators. tance between the black legislators’ district and the city I also check the robustness of the results with a simTyrone claimed to be from. In fact, though statistically inilar analysis based on linear regression in the supporting significant, this relationship is positive, inconsistent with information by analyzing the main results in the presence this concern. Column 2 verifies this implication further of additional heterogeneous treatment effect estimates for by showing that the main result of the experiment continthe black population of the district, the median houseued to hold even when considering observations where hold incomes of blacks and whites, whether the state is in the sender purported to live in a city more than 200 miles the American south, and the Squire (2007) index of state away from the legislators’ districts.12 legislative professionalism. Consistent with the matching Second, it might be of concern that black legislators analysis, in all cases the original results remain the same, could perceive greater electoral incentives to respond to and these rival explanations are statistically insignificant out-of-district letters than whites due to progressive am(see Table SI3). bition for statewide office; they may expect to need to In sum, the results of the field experiment strongly win black votes across their state in the future. Three facts support the view that extrinsic concerns constitute a make it seem unlikely this was the case. First, as Johngreater share of the reasons why nonblack legislators adson, Oppenheimer, and Selin (2012) recently document, vance blacks’ interests when they do. By contrast, black black politicians have very rarely run for or met success in legislators appear relatively insensitive to whether they statewide elections in recent decades for a variety of conwill receive outside reward for promoting blacks’ intertextual reasons, despite growing ranks of black officeholdests; that is, unlike nonblacks, most black legislators coners showing ambition for higher offices like U.S. House tinued to advance the black individuals’ interests even seats. Further allaying this concern, the results also remain when their electoral incentives for doing so dramatically significant at the 0.05 level when limiting the scope of the decreased. Moreover, a matching analysis showed that this analysis to states where Barack Obama received less than pattern held especially strongly in the very districts where 45% of the vote in 2008 (and thus where black legislators, institutions to promote descriptive representation matter 98% of whom are Democrats, should be relatively unlikely most. to expect to win statewide office any time soon).13 This In the next section, I evaluate a number of alternative result is shown in column 3 of Table 3. One final placebo explanations for these findings. test adds further confidence to this assumption still. If the response rates were driven by legislators’ perceived need to win black votes in future statewide elections, one would Evaluating Alternative Explanations expect the experimental manipulation to affect nonblack Potential Threats to Internal Validity Republicans more than nonblack Democrats (since Republicans almost never win black votes and Democrats, Possibility That Black Legislators Have Greater Incenespecially in states with numerous blacks, almost always tives to Respond to Blacks Living Far Away Than Do rely on them). However, as column 4 of Table 3 shows, Whites. The identification strategy assumes that blacks nonblack Democrats and Republicans did not differ in and whites perceive a similar decline in their relative extheir treatment of the out-of-district letters. There seems trinsic incentives for responding to the letter from the to be little evidence that black legislators would have exout-of-district writer instead of the in-district writer (that pected meaningfully greater extrinsic rewards from the is, that C – T is identical for whites and blacks). There out-of-district letter than would have whites. may be two reasons, however, to doubt this assumption. First, black legislators may expect black political networks spanning their states (e.g., Cho 2003) to monitor 12 This holds for other distances as well, including cases where the their interactions with black individuals across their state out-of-district cities are more than 50 miles away, 100 miles away, more closely. That is, black legislators might fear that not and 300 miles away. By 400 miles, the sample size grows too small (N = 5 black legislators). Distances were calculated between the responding to the out-of-district letter would endanger geographic ‘centroid’ of each district and the official geographic their reputation in black political networks in ways that coordinates of the out-of-district cities. would have repercussions back in their districts. Unem13 The result also holds at the 0.05 level when excluding states from ployment benefits were used in the letter to help allay this this group that have few black legislators (KS, KY, OK, WV). BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 531 TABLE 3 Robustness Checks and Placebo Tests for Alternative Explanations Specification (1, OLS) (2, OLS) (3, OLS) (4, OLS) (5, OLS) (6, OLS) Out-ofStates Where Black District City Obama Received Observations Legislators >200 Miles Away <45% Nonblack Nonblack All Used Only from District of 2008 Vote Legislators Legislators Legislators Experimental Treatment Effects Out-of-District Email −0.230ˆ −0.266∗∗ −0.269∗∗ −0.264∗∗ (0.120) 0.291∗∗ 0.268∗∗ (0.019) (0.015) (0.017) (0.020) (0.023) Out-of-District Email x Distance 0.045 Between (0.059) Legislator’s District and Out-of-District City (in 100s of Miles) Out-of-District Email x Black 0.180∗ 0.174∗ 0.217ˆ Legislator (0.078) (0.083) (0.119) Out-of-District Email x −0.018 Democratic Legislator (0.026) Out-of-District Email x Female −0.005 Legislator (0.031) Out-of-District Email x Percent of −0.002 Legislators Who Are Black (0.002) Out-of-District Email x −0.005 Black Legislator x Percent of (0.008) Legislators Who Are Black Covariates Distance Between Legislator’s −0.022 District and Out-of-District City (0.041) (in 100s of Miles) Black Legislator −0.119∗ 0.091 (0.054) 0.211∗∗ (0.083) (0.059) Democratic Legislator −0.032ˆ (0.019) Female Legislator −0.002 (0.022) Percent of Legislators Who Are 0.003∗ Black (0.001) Percent of Legislators Who Are −0.015∗∗ Black x Black Legislator (0.005) Constant 0.507∗∗ 0.564∗∗ 0.499∗∗ 0.578∗∗ 0.562∗∗∗ 0.545∗∗∗ (0.084) (0.014) (0.016) (0.013) (0.011) (0.012) R2 .024 .082 .078 .079 .077 .078 N 364 2322 1720 5229 5229 5593 Note: Dependent variable in all regressions is whether the legislator responded to the email. ˆp < 0.10, ∗ p < 0.05, ∗∗ p < 0.01 , ∗∗∗ p < 0.001 (two-tailed tests). 532 DAVID E. BROOCKMAN Possibility That Black Legislators Respond Differently the unemployed largely explained the results, one would to the Treatment Due to Being of Higher Quality. One expect Democrats, who are significantly more supportchallenge in designing the experiment was that there were ive of unemployment insurance nationwide, to be more so few black legislators in the United States (as of Novemlikely to respond to the letter. However, as shown in colber 2010) that a set of white letter placebo conditions umn 4 of Table 3, Democrats were no more responsive could not be implemented due to a lack of adequate samthan Republicans to the out-of-district letter. ple size. With fewer than 90 black legislators in each cell In addition, as Dawson’s theory of linked fate argues, under such a design (and further exacerbated by the bithe “economic component of group interests” (1994, 85) nary nature of the dependent variable), the experiment is at the heart of black group consciousness in the United would have been significantly underpowered and only States. Even if some of black legislators’ intrinsic moable to reliably detect implausibly large differences in retivation to respond was driven by the economic differsponse rates. However, this lack of parallel white letter ences underlying the racial divide in America, this would conditions introduces several assumptions. still have largely similar and important substantive conOne such assumption is that black legislators themsequences for how black and nonblack legislators differ in selves do not tend to respond more to out-of-district their treatment of blacks in the United States. However, emails because they are of generally higher quality. As future work should attempt to further distinguish the efAnzia and Berry (1990) have shown, because the barriers fects of race and class, an important topic many scholars to office are higher for women, women who are elected have long and no doubt will continue to pursue. tend to be more talented in a variety of ways. Might a similar legislator quality effect explain black legislators’ Legislators Expect Tyrone’s Black Legislator Will Redecreased sensitivity to their electoral incentives to respond. Finally, what of the possibility that black legspond? A placebo test based on Anzia and Berry’s (2011) islators do or do not respond to the out-of-district emails own theory can help address this concern: women are because they expect another black state legislator already also underrepresented at the state house level just as they represents the sender? Column 6 of Table 3 examines are at the federal level (in fact, even more so than blacks). the three-way interaction between the legislator’s race, Column 5 of Table 3 thus evaluates whether female legisthe out-of-district email treatment, and the percentage lators were more likely than men to respond to the letter of legislators in the state who are black. The results show overall or the out-of-district letter in particular and finds that legislators in general (the treatment x percentage of no support for either hypothesis. It consequently appears legislators who are black coefficient) and black legislators that discrepancies in legislative quality (in the Anzia and in particular (the three-way interaction) do not appear Berry sense) across races would not lead legislators to significantly less responsive to the out-of-district emails respond to this experiment differently.14 in states where many other black legislators serve.15 In summary, while there are many tempting alternaBlack Legislators May Be More Motivated or Able to tive explanations for the experiment’s findings, a number Provide Assistance to All Unemployed. An additional of robustness checks and placebo tests suggest they are concern with my interpretation of the results is that, deunlikely to account for the results observed. spite the advantages to using a request for unemployment benefits, the responses to such a request might actually reflect black legislators being more intrinsically motivated Possibility That Staff Results Limit than nonblack legislators to represent the unemployed in Generalizability general, regardless of race. In thinking about the experiment’s broader generalizTwo responses can help address this concern about ability, one final concern is that I treat state legislators’ how to interpret the results. First, while data on U.S. state email addresses, not the legislators themselves. In highly legislators’ personal views on unemployment benefits are professionalized legislatures where staff answers email, prohibitively difficult to gather, one variable does correthe treatment effect thus sometimes captures the effect late extremely well with such views: party. If concern for of treating a legislative office instead of the legislator 14 Another alternative account may be that black legislators are simper se. To evaluate the possibility that the effects might ply less likely to differentiate between emails based on their content in general (e.g., they simply read correspondence from constituents largely reflect staff behavior only, Table SI3 in the online less attentively). However, Butler and Broockman (2011) find that 15 blacks were actually far more likely than nonblacks to react to the The interaction between the treatment and the percentage of a partisanship and race signals in their experiment (see Butler and legislature that is black is also insignificant when examining black Broockman 2011, Table SI2, Parts C and D). legislators only. BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 533 supporting information shows that the results remain rostantive political representation. Underscoring this point, bust to the inclusion of a heterogeneous treatment effect a matching analysis showed that the effects strongly perfor state legislative professionalism (as measured by the sisted even in the very districts where institutions to proSquire index). The results also hold even when only conmote descriptive representation would have the greatest sidering states where legislators have no staff who help to effect. answer email, and the legislators themselves answered all Moreover, these results also strongly support the hythe emails. Legislative staff were unlikely to be responsible pothesis articulated by Mansbridge (2003) that legislators for the patterns observed. engage in “surrogate” representation of group members outside their official constituencies (see also Pitkin’s 1967 concept of “virtual” representation). Americans whose Discussion legal representatives are not of their group (a situation nearly all minority groups in the United States find themAn enduring and significant question about democratic selves in) thus do have an interest in ensuring that group representation is to what extent politicians act on intrinmembers do serve in the collective bodies that govern sic motivation to advance the interests of their group. them. Such insights are especially important to note as Such intrinsic motivation matters because politicians ofdebates over the future of the Voting Rights Act continue. ten have a great deal of leeway in their behavior (Bianco More broadly, the results also provide strong em1994): they make many decisions concerning groups that pirical evidence that intrinsic motivation can play an pay little attention to their actions (Delli Carpini and important role in shaping politicians’ behavior. Though Keeter 1996), in contexts that are difficult to observe it is often viewed as analytically productive to think of (Arnold 2009; Hall 1996), and regarding the interests of politicians as motivated by reelection alone (e.g., Maythose beyond the boundary of their formal constituencies hew 1974), the experiment showed that politicians ap(Mansbridge 2003). Nevertheless, though many scholpear willing to pay costs to achieve goals they personars argue that politicians place great personal value on ally value: political incentives alone cannot easily explain achieving certain outcomes for their group (e.g., Burden the behavior observed in the experiment, either on the 2007; Mansbridge 1999; Whitby 1997), the role of such part of the many legislators of all races who responded intrinsic motivation is challenging to empirically explore to the out-of-district letter or the black legislators who because politicians may have incentives to appear intrinwere especially responsive to it. A focus on legislators’ sically motivated even if they are not. Politicians’ intrinsic baseline and differing level of intrinsic motivation can motivation thus plays a potentially large role in political thus yield crucial insights into their behavior (see Mansrepresentation even as detecting it in action has remained bridge 2009). Though future work can and should cona considerable challenge for the very same reasons. sider whether these findings generalize to other groups, In this article, I attempted to shed light on these they strongly demonstrate the potential importance of long-standing questions about politicians’ intrinsic momany politicians’ personal preferences and desires. tivation with a field experiment that offered a rare look at The results also suggest several important avenues for how politicians behave when their political incentives are future research. First, as Democrats and Republicans apweakened. Specifically, I emailed state legislators from a pear to differ little in their treatment of black interests in putatively black alias asking for help signing up for state private settings, future research should further consider unemployment benefits and randomly varied whether the trade-off between the relative lack of concern nonthe letters purported to come from a person living within black Democrats appear to show for black interests in less or far from each legislator’s district. The results showed public behavior and the relatively robust substantive repthat while most black legislators continued to respond to resentation they generally provide blacks on many salient a putatively black person even when they expected little policy issues (Lublin 1997). political reward for doing so, nonblacks were much less reFrom an institutional point of view, the findings also sponsive once their political incentives were diminished. grant further urgency to efforts to devise institutions Intrinsic motivation to advance blacks’ interests thus apthat can better monitor elected officials’ behavior toward pears to be a substantial determinant of black legislators’ groups they are not a part of. Given how differently politibehavior. cians appear to behave when their incentives diminish, the To the extent black Americans face structural barriresults also underscore the clear importance of electoral ers to exercising political accountability, mechanisms for accountability to minorities’ equal representation, espeencouraging the election of blacks to office thus appear cially when they are not represented by a member of their especially crucial to ensuring that blacks receive equal subgroup. 534 DAVID E. BROOCKMAN Finally, the results underscore that Canon’s (1999) Butler, Daniel M., Christopher Karpowitz, and Jeremy C. Pope. and Dovi’s (2002) points about the “supply side” of 2012. “A Field Experiment on Legislators’ Home Style: Service versus Policy.” Journal of Politics 74(2): 474–86. candidates—which candidates run for office—should not be neglected. A rich literature explores variation in group Canon, David T. 1999. Race, Redistricting, and Representation: The Unintended Consequences of Black Majority Districts. identification and policy preferences in the mass public Chicago: University of Chicago Press. (e.g., Gay 2004; Hochschild and Weaver 2007), though Carnes, Nicholas. 2012. “Does the Numerical Underrepresentamuch less has done so among elites. Future research tion of the Working Class in Congress Matter? Evidence from should do so—who governs does have consequences. Roll Call Voting in the House of Representatives.” Legislative Studies Quarterly 37(1): 5–34. Carnes, Nicholas, and John Holbein. 2012. “Don’t Take It PerSupporting Information sonally: Affluence, Influence, and Prejudice.” Working Paper, Duke University. Additional Supporting Information may be found in the Carroll, Susan J. 1994. Women as Candidates in American Politics. Bloomington: Indiana University Press. online version of this article at the publisher’s website: Chattopadhyay, Raghabendra, and Esther Duflo. 2004. • Randomization Check “Women as Policy Makers: Evidence from a Random◦ Table SI1: Randomization Check. DV = Assignment to ized Policy Experiment in India.” Econometrica 72(5): 1409–43. Out-of-District Treatment Group • Results with “Helpful Reply” as the Dependent VariCho, Wendy K. Tam. 2003. “Contagion Effects and Ethnic Contribution Networks.” American Journal of Political Science able 47(2): 368–87. ◦ Table SI2: Results With “Helpful Reply” as the DepenCitrin, Jack, and Donald P. Green. 1990. “The Self-Interest Modent Variable tive in American Public Opinion.” Research in Micropolitics ◦ Table SI3: Heterogeneous Treatment Effects 3: 1–28. • Inand Out-of-District City-Selection Method Cobb, Michael D., and Jeffery A. Jenkins. 2001. “Race and the ◦ Figure SI1: Example Inand Out-of-District City SelecRepresentation of Blacks’ Interests during Reconstruction.” Political Research Quarterly 54(1): 181–204. tion Results Dawson, Michael C. 1994. Behind the Mule. Princeton, NJ: ◦ Figure SI2: Out-of-District Assignment Areas Princeton University Press. ◦ Figure SI3: Guide to Reproducing the City Name AsDelli Carpini, Michael X., and Scott Keeter. 1996. What Amersignment Process (5 Panels) icans Know about Politics and Why It Matters. New Haven, CT: Yale University Press. Dovi, Suzanne. 2002. “Preferable Descriptive Representatives: References Will Just Any Woman, Black, or Latino Do?” American Political Science Review 96(4): 729–43. Anzia, Sarah F., and Christopher R. Berry. 2011. “The Jackie Fix, Michael, and Margery A. Turner, eds. 1998. A National Re(and Jill) Robinson Effect: Why Do Congresswomen Outport Card on Discrimination in America: The Role of Testing. perform Congressmen?” American Journal of Political Science Washington, DC: Urban Institute Press. 55(3): 47–93. Fowler, Linda L., and Robert D. McClure. 1989. Political AmbiAriely, Dan, Anat Bracha, and Stephan Meier. 2009. “Doing tion: Who Decides to Run for Congress. New Haven, CT: Yale Good or Doing Well? Image Motivation and Monetary InUniversity Press. centives in Behaving Prosocially.” American Economic ReFranck, Raphael, and Ilia Rainer. 2012. “Does the Leader’s Ethview 99(1): 544–55. nicity Matter? Ethnic Favoritism, Education, and Health in Arnold, R. Douglas. 1990. The Logical of Congressional Action. Sub-Saharan Africa.” American Political Science Review 106: New Haven, CT: Yale University Press. 294–325. Bianco, William T. 1994. Trust: Representatives and Constituents. Fryer, Roland G., Jr., and Steven D. Levitt. 2004. “The Causes Ann Arbor: University of Michigan Press. and Consequences of Distinctively Black Names.” Quarterly Bratton, Kathleen A., and Leonard P. Ray. 2002. “Descriptive Journal of Economics 119(3): 767–805. Representation, Policy Outcomes, and Municipal Day-Care Gailmard, Sean, and Jeffrey A. Jenkins. 2009. “Agency Problems, Coverage in Norway.” American Journal of Political Science the 17th Amendment, and Representation in the Senate.” 46(2): 428–37. American Journal of Political Science 53(2): 324–42. Burden, Barry C. 2007. Personal Roots of Representation. PrinceGamble, Katrina L. 2007. “Black Political Representation: An ton, NJ: Princeton University Press. Examination of Legislative Activity within US House ComButler, Daniel M., and David E. Broockman. 2011. “Do Politimittees.” Legislative Studies Quarterly 32(3): 421–47. cians Racially Discriminate Against Constituents? A Field Gay, Claudine. 2004. “Putting Race In Context: Identifying Experiment on State Legislators.” American Journal of Politthe Environmental Determinants of Black Racial Attitudes.” ical Science 55(3): 463–77. American Political Science Review 98: 547–62. BLACK POLITICIANS’ INTRINSIC MOTIVATION TO ADVANCE BLACKS’ INTERESTS 535 Griffin, John D., and Brian Newman. 2007. “The Unequal RepMansbridge, Jane. 2003. “Rethinking Representation.” Ameriresentation of Latinos and Whites.” Journal of Politics 69(4): can Political Science Review 97(4): 515–28. 1032–46. Mansbridge, Jane. 2009. “A ‘Selection Model’ of PolitiGrose, Christian R. 2005. “Disentangling Constituency and Legcal Representation.” Journal of Political Philosophy 17(4): islator Effects in Legislative Representation: Black Legislators 369–98. or Black Districts?” Social Science Quarterly 86(2): 427–43. Mayhew, David R. 1974. Congress: The Electoral Connection. Grose, Christian R. 2011. Congress in Black and White. CamNew Haven, CT: Yale University Press. bridge: Cambridge University Press. McClain, Paula D., Jessica D. Johnson Carew, Eugene Walton, Grose, Christian R., Maurice Mangum, and Christopher MarJr., and Candis S. Watts. 2009. “Group Membership, Group tin. 2007. “Race, Political Empowerment, and Constituency Identity and Group Consciousness: Evolving Racial Identity Service: Descriptive Representation and the Hiring of in American Politics.” Annual Review of Political Science 12: African-American Congressional Staff.” Polity 39(4): 449– 471–85. 78. McClendon, Gwyneth H. 2012. “Ethics of Using Public Officials Haider-Markel, Donald P. 2007. “Representation and Backlash: as Field Experimental Subjects.” The Experimental Political The Positive and Negative Influence of Descriptive RepreScientist 3(1): 13–20. sentation.” Legislative Studies Quarterly 32: 107–33. Meier, Kenneth J., and Robert E. England. 1984. “Black Haider-Markel, Donald P., Mark R. Joslyn, and Chad J. Kniss. Representation and Educational Policy: Are They Re2000. “Minority Group Interests and Political Representalated?” American Political Science Review 78(2): 392– tion: Gay Elected Officials in the Policy Process.” Journal of 403. Politics 62(2): 568–77. Miller, Warren E. and Donald W. Stokes. 1963. “Constituency Hall, Richard L. 1996. Participation in Congress. New Haven, Influence in Congress.” American Political Science Review 57: CT: Yale University Press. 45–57. Hochschild, Jennifer L., and Vesla Weaver. 2007. “The Skin Minta, Michael D. 2009. “Legislative Oversight and the Color Paradox and the American Racial Order.” Social Forces Substantive Representation of Black and Latino Inter86(2): 643–70. ests in Congress.” Legislative Studies Quarterly 34: 193– Hood, M. V., III, and Irwin L. Morris. 1998. “Boll Weevils and 218. Roll-Call Voting: A Study in Time and Space.” Legislative Minta, Michael D. 2011. Oversight: Representing the Interests Studies Quarterly 23(2): 245–69. of Blacks and Latinos in Congress. Princeton, NJ: Princeton Hutchings, Vincent L. 1998. “Issue Salience and Support for University Press. Civil Rights Legislation among Southern Democrats.” LegNye, John V. C., Ilia Rainer, and Thomas Stratmann. 2011. islative Studies Quarterly 23(4): 521–44. “Do Black Mayors Improve Black Employment Outcomes? Iacus, Stefano M., Gary King, and Giuseppe Porro. 2012. Evidence from Large US Cities.” Working Paper, George “Causal Inference without Balance Checking: Coarsened ExMason University. act Matching.” Political Analysis 20:(1) 1–24. Pande, Rohini. 2003. “Can Mandated Political RepresentaIyer, Lakshmi, Anandi Mani, Prachi Mishra, and Petia Topalova. tion Increase Policy Influence for Disadvantaged Minori2011. “The Power of Political Voice: Women’s Political Repties? Theory and Evidence from India.” American Economic resentation and Crime in India.” Working Paper, Harvard Review 93(4): 1132–51. Business School. Pitkin, Hanna F. 1967. The Concept of Representation. Berkeley: Johnson, Gbemende, Bruce I. Oppenheimer, and Jennifer L. University of California Press. Selin. 2012. “The House as a Stepping Stone to the SenPreuhs, Robert R. 2006. “The Conditional Effects of Minorate: Why Do So Few African American House Members ity Descriptive Representation: Black Legislators and Policy Run?” American Journal of Political Science 56(2): 387– Influence in the American States.” Journal of Politics 68(3): 99. 585–99. Juenke, Eric G., and Robert R. Preuhs. 2012. “Irreplaceable LegRiach, P., and J. Rich. 2004. “Deceptive Field Experiments islators? Rethinking Representatives in the New Century.” of Discrimination: Are They Ethical?” Kyklos 57(3): 457– American Journal of Political Science 56(3): 705–15. 70. King, Gary, and Langche Zeng. 2006. “The Dangers of Extreme Shayo, Moses, and Asaf Zussman. 2011. “Judicial Ingroup Bias Counterfactuals.” Political Analysis 14: 131–59. in the Shadow of Terrorism.” Quarterly Journal of Economics Kingdon, John W. 1981. Congressmen’s Voting Decisions. New 126(3): 1447–84. York: Harper and Row. Snyder, James M., and David Stromberg. 2010. “Press Coverage Lee, D. S., E. Moretti, and M. J. Butler. 2004. “Do Voters Affect and Political Accountability.” Journal of Political Economy or Elect Policies? Evidence from the U.S. House.” Quarterly 118(2): 355–408. Journal of Economics 119(3): 807–59. Squire, Peverill. 2007. “Measuring State Legislative ProfessionLublin, David. 1997. The Paradox of Representation. Princeton, alism: The Squire Index Revisited.” State Politics and Policy NJ: Princeton University Press. Quarterly 7(2): 211–27. Mansbridge, Jane. 1999. “Should Blacks Represent Blacks and Swain, Carol M. 1993. Black Faces, Black Interests: The RepreWomen Represent Women? A Contingent ‘Yes’.” Journal of sentation of African Americans in Congress. Cambridge, MA: Politics 61(3): 628–57. Harvard University Press. 536 DAVID E. BROOCKMAN Tate, Katherine M. 2003. Black Faces in the Mirror: African AmerWhitby, Kenny J. 1997. The Color of Representation: Congresicans and Their Representatives in the US Congress. Princeton, sional Behavior and Black Interests. Ann Arbor: University of NJ: Princeton University Press. Michigan Press. van Houweling, Robert P. 2012. “Parties as Enablers: Individual Word, David L., Charles D. Coleman, Robert Nunziata, and Incentives for Partisan Legislative Organization.” University Robert Kominski. n.d. “Demographic Aspects of Surnames of California, Berkeley. Unpublished manuscript. from Census 2000.” Technical Report for the U.S. Census Washington, Ebonya L. 2008. “Female Socialization: How Bureau. http://www.census.gov/genealogy/www/surnames Daughters Affect Their Legislator Fathers’ Voting on .pdf. Women’s Issues.” American Economic Review 98(1): 311– Yinger, John. 1998. “Evidence on Discrimination in Consumer 32. Markets.” Journal of Economic Perspectives 12(2): 23–40.
5 Brokers, Social Networks, Reciprocity, and Clientelism Nico Ravanilla University of California San Diego Dotan Haim Florida State University Allen Hicken University of Michigan Abstract: Although canonical models of clientelism argue that brokers use dense social networks to monitor and enforce vote buying, recent evidence suggests that brokers can instead target intrinsically reciprocal voters and reduce the need for active monitoring and enforcement. Combining a trove of survey data on brokers and voters in the Philippines with an experiment-based measure of reciprocity, and relying on local naming conventions to build social networks, we demonstrate that brokers employ both strategies conditional on the underlying social network structure. We show that brokers are chosen for their central position in networks and are knowledgeable about voters, including their reciprocity levels. We then show that, where village social networks are dense, brokers prefer to target voters who have many ties in the network because their votes are easiest to monitor. Where networks are sparse, brokers target intrinsically reciprocal voters whose behavior they need not monitor. Verification Materials: The data and materials required to verify the computational reproducibility of the results, procedures and analyses in this article are available on the American Journal of Political Science Dataverse within the Harvard Dataverse Network, at: https://doi.org/10.7910/DVN/UTRXT7. E lectoral clientelism—the targeting of money, clientelist campaign machines (Schmidt et al. 1977; Scott goods, or favors to influence voters—is a 1972; Stokes et al. 2013). widespread phenomenon, common in nearly all Despite the pervasiveness of brokers, the logic bedeveloping democracies and many established demochind the strategies they employ remains a key point of racies as well (Hicken 2011; Kitschelt and Wilkinson debate. Seminal models of clientelism argue that bro2007).1 Whether the goal is vote buying (Stokes 2005), kers are valuable to campaigns because their social emturnout buying (Nichter 2008), abstention buying (Cox beddedness allows them to monitor voter behavior and and Kousser 1981), or ticket buying (Aspinall and Sukjaenforce clientelistic exchanges (Brusco, Nazareno, and mati 2016), a common theme across all these varieties of Stokes 2004; Camp 2017; Kitschelt and Wilkinson 2007; clientelism is the crucial role of brokers—the intermediStokes et al. 2013). However, there is growing evidence aries between candidates and voters, and the lynchpins of that in some cases, brokers target electoral clientelism Nico Ravanilla is at School of Global Policy & Strategy, University of California, San Diego, 9500 Gilman Drive, #0519, La Jolla, San Diego, CA 92093 (nravanilla@ucsd.edu). Dotan Haim is at Department of Political Science, Florida State University, 113 Collegiate Loop, Tallahassee, FL 32306 (dhaim@fsu.edu). Allen Hicken is at Department of Political Science, University of Michigan, 505 S. State Street, Ann Arbor, MI 48109 (ahicken@umich.edu). We thank Grace Labalan (Innovations for Poverty Action) for unparalleled project management, Michael Davidson for outstanding research assistance, and seminar participants at SPSA 2017, AAS 2017, MPSA 2017, University of California San Diego, Northern Illinois University, Brigham Young University, Leiden University, University of Virginia, and Duke University. We thank Stephan Haggard and Craig McIntosh for valuable feedback. This project was part of a larger project on electoral dynamics in Malaysia, Indonesia, and the Philippines, and was conducted in collaboration with Edward Aspinall (ANU), Paul Hutchcroft (ANU), and Meredith Weiss (SUNYAlbany). Funding came from a Discovery Project grant from the Australian Research Council (DP140103114). Our work in the Philippines was supplemented with grants from the Australian Department of Foreign Affairs and Trade and De La Salle University. This study received ethical approval from the UM-IRB (HUM00111251). Preanalysis plan for this study is documented at e-gap.org (20160517AB). nravanilla@ucsd.edu 1 It is also present in many nondemocratic contexts (Corstange 2018). American Journal of Political Science, Vol. 00, No. 0, XXXX 2021, Pp. 1–18 ©2021, Midwest Political Science Association DOI: 10.1111/ajps.12604 1 2 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN to “intrinsically” reciprocal voters, attenuating the need Finally, we show that the targeting of voters contingent for active monitoring and enforcement (Schaffer 2007). on social network structures helps explain voter behavior For example, Finan and Schechter (2012) find that camat the polls. Being offered any money by a campaign is paigns successfully identify which individuals have high associated with a doubling of the likelihood of voters’ levels of reciprocity and then target those individuals reporting having voted for the campaign’s candidate, with private inducements. and this pattern is almost entirely driven by the behavior Here, we posit that brokers pursue both strategies of central voters in dense networks. but that the decision to use one strategy over another Our article contributes to ongoing debates about the is contingent on the social network structure in the nature of clientelism by painting a particularly comprearea where brokers operate. Specifically, we argue that in hensive picture of the strategies adopted by brokers. We dense social networks, brokers primarily target socially are among the first to show evidence in support of the key central voters who are more likely to believe that knowlassumptions behind the “information asymmetry” modedge of their vote choice will spread, making them more els of brokers: they have strong social connections with, susceptible to monitoring and social sanctions. In sparse and are knowledgeable about, voters. We find evidence networks, where targeting central voters is less effective, of widespread overall vote buying, targeted in a manner brokers instead identify and target intrinsically reciprocal that suggests that two prevailing targeting strategies idenvoters.2 tified by the literature are at play, but also find that each To test the observational implications of our theory, strategy is contingent on social network structures. Fiwe use an original survey paired with a lab-in-the-field nally, we confirm the fundamental mechanisms underexperiment to measure reciprocity of 199 brokers and lying each strategy, including the greater observability of 701 randomly sampled voters in Mahamot, Philippines, voter behavior in denser social networks. during the 2016 elections.3 For the broker survey, we The findings from this article also have important worked with the campaign manager of a nonincumbent implications for recent work suggesting that brokers are candidate contesting the mayoral seat to survey the full less effective than existing theories assume. Driven by the roster of brokers employed by the campaign. We then use observation that “unmonitored” clientelism is prevalent the methods pioneered by Cruz, Labonne, and Queruin the developing world and by the empirical difficulty bin (2017) to build the complete family-based social netin testing the key assumptions behind information works of brokers and voters in Mahamot. asymmetry models of brokers, scholars have called into We find that network centrality is a strong predicquestion whether brokers are truly a sine qua non of elector of who becomes a broker, crucial for the argument toral clientelism (Brierly and Nathan 2019; Chauchard that brokers rely on downward ties with voters for target2018; Kramon 2017; Muñoz 2014; Nichter 2018). For ing clientelism. Moreover, in contrast to recent findings example, recent studies have found that, absent reliable elsewhere (Brierly and Nathan 2019; Schneider 2019), party cues, local brokers have a difficult time identifying we find that, compared to the average person, Philippine how citizens in their areas voted (Schneider 2019) and brokers correctly identify voter characteristics (including that brokers overestimate their own ability to mobilize reciprocity levels) remarkably well. votes (Aspinall and Sukjamati 2016). By contrast, we find Consistent with our central hypothesis, we then that brokers in our context are impressively knowledgeshow that brokers distribute electoral handouts continable about citizens, including how they voted. Our results gent on the social network structure. Brokers prioritize suggest that, by not accounting for the structure of social targeting socially central voters when social networks networks, scholars might miss out on cases where brokers are dense overall. When the social network is sparsely successfully target vote buying in a way that incentivizes connected, brokers then target reciprocal voters. Using voters to follow through. In addition, our mechanism the detailed voter survey, we show evidence for one tests suggest that even in the absence of clear evidence important mechanism driving this result: that socially that brokers are actively monitoring voters, socially connected voters in dense networks are more likely to central voters in dense social networks are more likely to believe that brokers can find out about their voting perceive that brokers can discover their votes and have the behavior, and that they will be cut off from future power to cut them off from future benefits, resulting in clientelistic exchanges if they do not follow through. these voters being more likely to follow through on a vote bought. 2 This core argument is documented in our preanalysis plan (see Overall, we contend that our core findings about the Section 2, Theoretical Framework). importance of social network density for shaping broker 3 The municipality name has been anonymized. targeting strategies are generalizable across a wide variety BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 3 of political contexts. The details of the social networks Second, brokers can identify voters whose posiat play will certainly vary, and we should keep in mind tion in the social network makes them more likely the particularities about the Philippines political context to follow through on clientelistic exchanges because (e.g., village size and the weakness of political parties), they expect future benefits to be conditioned on their but in general, we expect the adaptability of brokers’ voting behavior. This type of reciprocity motivated by strategies to different network contexts to be a finding forward-looking self-interest is sometimes referred to as that travels. We will return to the scope conditions of our “instrumental” reciprocity (Sobel 2005). Actors respond argument in the conclusion. to those with whom they have ongoing relationships, anThe rest of the article proceeds as follows. In the ticipating that reciprocity will generate more benefits in next section, we outline our theory for how social netfuture interactions. In the case of vote buying, activating work structures play a role in determining brokers’ tarinstrumental reciprocity requires voters to believe that geting of handouts. The “Brokers and Vote Buying in they have an increased likelihood of facing consequences the Philippines” and “Measuring Social Networks, Recifor not following through, placing a greater emphasis procity, and Vote Buying” sections provide information on brokers’ perceived ability to monitor and enforce on the political context in the Philippines and present the exchanges. data we use to measure vote buying, intrinsic reciprocity, There are two possible mechanisms that might unand social network structures. The “Empirical Results” derlay the connection between an individual’s centrality section presents the main findings of the article, whereas in social networks and the increased tendency to display the “Conclusion” section concludes with a discussion on instrumental reciprocity. First, information that spreads the generalizability of our results and implications for futhrough social networks allows brokers to more conture scholarship. fidently ascertain voters’ vote choice and identify voters who should be sanctioned for not following through (Keefer 2007; Stokes 2005). When voters have many social ties in their village,4 we expect them to believe that Theory information about their vote choice is more likely to spread and that brokers will be more likely to access inA rich literature in comparative clientelism argues formation about their behavior. Social network centrality that clientelistic exchanges are sustained, in large part, also creates opportunities for social sanctioning against through dense social networks that allow brokers to defecting voters. Voters do not want to be viewed as demonitor and enforce vote buying. However, the role fecting because it may cut them off from future streams played by social networks in ensuring the reciprocal beof benefits, or undercut their ability to get additional sohavior of voters—following through on a handout by cial resources from people within their network. Thus, supporting the broker’s preferred candidate—is underthe ability to activate instrumental reciprocity increases specified. We argue that voter reciprocity can either be when targeting voters who have more extensive social ties instrumental or intrinsic, and that brokers can use social in their community, making them more vulnerable to networks to harness both these types of reciprocity to fasanctioning.5 cilitate the clientelist exchange. We expect that brokers will attempt to harness both First, social networks can help brokers harness the intrinsic and instrumental reciprocity when mobilizing power of voters’ intrinsic reciprocity by helping brokers support for their candidate, but that the viability of each identify which voters are most likely to follow through strategy depends on the social network density in the vilon clientelist exchanges without the need for monitorlage as a whole. Specifically, we predict that brokers will ing and enforcement. Intrinsic reciprocity is motivated prefer to target socially central voters when networks in by the pleasure a person receives in increasing the mathe village are more dense, creating the conditions necterial payoffs of people who helped them (Leider et al. essary to activate these voters’ increased propensity to 2009). This trait is strongly associated with altruism and display instrumental reciprocity. Dense social networks the desire to punish individuals who renege on promises made. Brokers can disproportionately target voters who 4 possess this trait in order to make it more likely that they In our theory, we use the term “village” as shorthand for the local political unit, whether it is urban or rural. will vote for the broker’s preferred candidate (Finan and 5 Schechter 2012). We expect that brokers’ social proximThis prediction is consistent with Cruz (2019) who finds that central voters are more likely to be targeted with vote buying ity to voters in their network will allow them to better and more concerned with the ramifications of reneging on voteidentify which individuals display intrinsic reciprocity. buying agreements. 4 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN are generally far more conducive to information diffuhypotheses. Our main hypothesis about how brokers will sion, especially from centrally located individuals (Yogatarget voters contingent on social network structure is narasimhan 2012). In our context, we posit that dense summarized in Figure 1. networks at the village level will result in central voters being even more likely to believe that brokers can monitor and sanction them. In addition, dense networks in the 1. Hypothesis 1: Brokers will target clientelism village as a whole make it more likely that social sanctions to voters depending on the social network for reneging will be widespread, especially for central architecture. individuals. (a) In dense social networks, brokers will target On the other hand, when village networks are central voters for whom they can leverage sparse, instrumental reciprocity is less easily harnessed instrumental reciprocity. to compel voters to follow through on the clientelist (b) In sparse social networks, brokers will target bargain—even the most socially central voters have a intrinsically reciprocal voters. low probability of their vote choice becoming widely 2. Hypothesis 2: Central voters in dense social netknown. As a result, in these cases brokers will resort works will feel that they are more likely to be efto the strategy that is still available to them: relying on fectively monitored and sanctioned by brokers. intrinsic reciprocity. In short, in villages with sparsely (a) In dense social networks, central voters will connected social networks, brokers will forgo targeting be more likely to believe that brokers can acsocially central individuals to instead target individuals cess information about their vote choice. with high levels of intrinsic reciprocity.6 (b) In dense social networks, central voters will Our theory rests on two key assumptions. First, we be more likely to believe that they will stop assume that across all villages, (Assumption 1) campaigns receiving handouts in the future if they fail will select socially central brokers. This attribute allows to follow through on a vote bought. brokers to use their social proximity to voters to better 3. Hypothesis 3: The efficacy of electoral handouts identify reciprocal voters and also makes it more likely will depend on the social network architecture. that brokers have preexisting ties to socially central vot• In dense social networks, central voters will be ers. In other words, we assume that a broker’s value flows, more likely to follow through on voting for a in part, from their position in relevant social networks. candidate from whom they accepted money. One challenge with this assumption from a causal standpoint is that by being invited to be a broker, an individual may become a key figure in a political network. Alternative Hypotheses Thus, rather than network position driving broker selection, it could be that broker selection drives network poThus far, we have argued that brokers’ targeting stratesition. Fortunately our empirical strategy relies on slowgies are contingent on the social network structure they to-change family networks, allowing us to largely sidestep face. However, before turning to our empirical results, it this endogeneity concern. is worth considering whether differential broker targetA second key assumption is that (Assumption 2) ing based on social network density may instead be capbrokers know voters considerably well. That is, they can turing evidence consistent with other prominent theories correctly identify voters in their constituency, and know of clientelism. First, it is possible that the types of voters basic information about those individuals. This would targeted by brokers (including their network characterisinclude demographic information such as age, income, tics) differ in “swing” or “core” districts. This could be and employment status. Importantly, we assume that not an important confounder if campaign strongholds tend only can brokers identify key demographic and politito overlap with dense social networks (c.f. Auyero 2000). cal characteristics of their voters, they can also correctly Alternatively, it may be that brokers’ targeting strategies identify voter levels of intrinsic reciprocity. Given these are contingent not on the density of social networks per assumptions, our theory generates the following testable se, but on the size of the polling stations in such networks. Previous work has shown that brokers are better 6 Note that ours is a ceteris paribus argument. Brokers may take able to monitor voters in places with a smaller number other factors into consideration when choosing who to target, as of voters per polling station (Rueda 2017). Failing to acwe discuss in the “Alternative Hypotheses” section. However, concount for this dynamic could bias our results if the dentrolling for those factors we expect voter centrality, intrinsic reciprocity, and these variables’ interaction with network density to sity of social networks tends to be correlated with the size shape brokers’ choice of strategy. of polling stations. BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 5 FIGURE 1 Hypothesized Targeting Strategy: (A) Dense Social Network and (B) Sparse Social Network Note: The targeting strategy in the hypothetical networks above corresponds to our core hypothesis (H1). If precinct size or the core versus swing nature of toral clientelism carries with it a negative moral and the district determined broker strategies, then the signifsocial stigma (Corstange 2009), the offer and acceptance icance of social networks should disappear after controlof cash or goods during election season is considered ling for these factors in our empirical analyses. In our standard practice in the Philippines and generally carries results section, we test whether our hypotheses hold afno such stigma. ter accounting for these alternatives. Consistent with our Politics in the Philippines has always been a decidpre-analysis plan, we do not have an explicit theory for edly local affair, with national party organizations playhow targeting dynamics based on network structures ining a negligible role in electioneering, and enduring party teract with precinct size or the swing/core nature of disattachments virtually nonexistent among voters. Instead, tricts, though we think this is an interesting avenue for a more informal chain of patron-broker relationships future inquiry.7 connects local politicians to their provincial and national counterparts (Hutchcroft and Rocamora 2003; Ravanilla 2017; Ravanilla, Sexton, and Haim Forthcoming). Brokers and Vote Buying in the In contrast, many of the major studies of clientelism have Philippines been conducted in contexts where party machines and voter attachments to parties are both strong (e.g., Calvo and Murillo 2004; Stokes et al. 2013). We discuss some of Our study was conducted in the Philippines. Electhe ways weak parties might affect the generalizability of tions for municipal, provincial, and national offices are our results in the conclusion, but the fact of the presence held every three years and are generally competitive, of weak parties is, in part, why we find the Philippines free, and fair, and—with some notable and important case so interesting. How does clientelism work in an enexceptions—relatively free of violence and intimidation vironment where parties are weak? How do brokers oper(Hicken, Aspinall, and Weiss 2019). Electoral clientelism ate where they are not permanently attached to a national is a prominent feature of provincial and municipal party? And we note that although the Philippines may elections in the Philippines, as numerous studies have look different from Argentina or Mexico in this regard, documented (Cruz, Keefer, and Labonne 2016; Hicken it finds kinship with many other developing democracies et al. 2015; Hicken et al. 2018; Hicken, Aspinall, and where parties remain underdeveloped and underinstituWeiss 2019). Although in some country contexts electionalized (Aspinall and Sukjamati 2016; Muñoz 2019). 7 Within the Philippines, our study was conducted in We include regressions in SI 17 that explicitly look at these interactions to facilitate future theory building and out-of-sample Mahamot, a rural municipality in Southern Luzon. With testing. a poverty level near the national average and an economy 6 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN centered around agriculture, Mahamot is fairly typical of However, their real work is done in the few days leadthe nearly 1,500 rural municipalities in the Philippines.8 ing up to election day. Brokers develop lists of voters who Politics in Mahamot have been fairly competitive, with will be targeted with handouts in their barangay (village), no one group or family able to completely control local which are then passed on to the central campaign and ofpolitics.9 ten checked against the list of registered voters for verifiThe organization and conduct of electioneering in cation. Money then flows down from candidates to broMahamot is representative of what we have directly obkers who are charged with distributing those funds to the served in other areas of the Philippines, and what other voters on their lists. Ethnographic evidence and the soresearchers report as well (Cruz 2019; Hicken et al. 2018; ciological literature on the Philippines suggest that broHicken, Aspinall, and Weiss 2019). The bases for camkers rely on norms of reciprocity, both intrinsic and inpaigning are local political machines built by and around strumental, as the following quotes suggest. On intrinsic political candidates or groups of candidates (e.g., famireciprocity: lies or clans). Although these local political machines are Our Filipino trait of utang na loob [debt of gratpromiscuous when it comes to national partisan affiliaitude] is evident… Once a person has granted tion, locally they often endure across multiple elections, us something, a favor, we would do everything and brokers (called liders in the Philippines) are generto pay that favor back to him or her, sometimes ally loyal to a local machine as long as it exists and is even at the expense of ourselves. We tend to competitive.10 view persons who did us some good things as The structure of local campaign organizations looks benefactors, and we view ourselves as beneficiasimilar to what we see in many other parts of the world ries who can please them by doing the same for where clientelism is a feature of elections—networks them. (Bava 1998, cited in Schaffer 2005) of brokers connect candidates to voters in each locality (Hicken 2011). Brokers include local officials, famOn instrumental reciprocity: ily members and allies, and those who received assistance from the candidate in the past (Schmidt et al. 1977; Sidel What kept them together and loyal to [the politi1999). Of the brokers we surveyed, more than half held cian] was the belief that they would be the prian elected position in their village at some point (premary beneficiaries of government programmes dominantly village councilors) and two-thirds are acin the form of scholarships, medical benefits, tively involved in religious or civil society organizations. and the like, after the election. (Rocina 2019, Many of these elements are similar to what Holland and 161) Palmer-Rubin (2015) term “organizational” brokers.11 Past research has not adequately specified under Although our theory is not specific to the type of browhat conditions each logic dominates. As we endeavor ker employed, the results from this study may be most to show in the next sections, the nature of this distribuapplicable to other context where brokers share simition of cash for votes differs systematically based on social lar features (see Mares and Young 2016, for a useful network structures. typology). The official campaign period lasts for about three months. Local brokers may play a role earlier in the campaign period by acting as the candidate’s guide in the vilMeasuring Social Networks, lage or by encouraging people to attend campaign events. Reciprocity, and Vote Buying Building Social Networks 8 As of 2015, the Philippines population was divided nearly evenly between a small handful of urban cities and rural municipalities To measure features of social networks structures and (51.2–48.8%) (PSA 2019). the positions of brokers in these networks, we rely on 9 This stands in contrast to some parts of the Philippines, where lothe methods pioneered by Cruz, Labonne, and Querucal political dynasties dominate (Mendoza, Venida, and Yap 2016). bin (2017) to establish familial links between individuals 10 Of the brokers we surveyed, 82% had worked for a campaign in appearing on the 2016 Certified Voter Lists.12 For each the past, of which 75% worked for the same candidate. barangay (village) in the survey area, we created a net11 There also exist parallels with other contexts where brokers are work in which a tie exists between all pairs of individuals local elites (e.g., Auerbach and Thachil 2018; Schneider 2019) and, who share at least one common surname. though partisan attachments are weak in the Philippines, the loyalty to a particular machine is reminiscent of partisan middle men 12 (e.g., Finan and Schechter 2012; Stokes 2005). See also Ravanilla, Davidson, and Hicken (2020). BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 7 FIGURE 2 Village Family Networks: (A) Dense Village Network and (B) Dense Village Network Note: The figures represent the family network structures of two villages in Mahamot with similar populations (Village A—365; Village B—314) but different network densities. Nodes represent individuals and are sized by degree centrality. Ties represent direct family relationships between those individuals. Dark squares represent brokers. Two features of the Filipino naming convention are work density by dividing the number of existing ties by crucial to the success of this method. First, the Spanish the number of possible ties. The following equation is naming convention adopted in the colonial period allows used to calculate this measure for each barangay (b), us to track both paternal and maternal family lines. Chilwhere t is the total number of connections between indren have one given name and two surnames: a “middle” dividuals in the barangay and n is the total population of name from their mother (her maiden surname) and a the barangay: “last” name from their father.13 Second, in the nineteenth t century, Provincial Colonial leadership assigned each Densit yb = . n × (n − 1)/2 barangay a set of new surnames from the Spanish “CatBecause the density measure is correlated with alogo Alfabetico de Apellidos” (the Alphabetical Catabarangay population—larger villages appear to be less logue of Surnames) to distribute to family heads in the dense because the number of possible ties grows quickly barangay. Because each barangay was assigned a different with an increase in population—we account for populaset of names, the legacy of this policy is that individuals tion in all empirical specifications. Figure 2 provides an who share surnames (especially in small geographic arexample of a dense and a sparse village network drawn eas) are highly likely to actually share a family link. from villages in Mahamot. In supporting information (SI) Figure E.1 (SI 9), we show that the distribution of Network Density. To calculate the relevant network village network density in Mahamot closely matches that measures, we begin by drawing the full network of all of rural barangays in the Philippines as a whole. As a reregistered voters in each barangay. 2016 voter registrasult, we think that our results are most likely to be genertion rates in Mahamot were above 87% on average, givalizable to the two-thirds (65%) of barangays nationwide ing us a highly representative view of the full family netthat fit into this category. work. In this network, individual voters are designated as nodes and family ties between these individuals (indiVoter and Broker Centrality. In addition to the villagecated by a shared middle or last name) are represented level measures, we calculate a number of statistics regardas unweighted, undirected edges. Barangays in our saming the position of each individual surveyed broker and ple have an average of approximately 600 registered votvoter in the network. When surveyed, each respondent ers and an average individual has 33 direct family members in their barangay. Using the full network for each 13 When women marry, they adopt the last name of their husband barangay, we calculate a barangay-level measure of netand keep their paternal surname as a middle name. 8 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN was asked to give their full legal name, which we then timates in the comparative literature (Hicken, Aspinall, matched with names on the voter list. All but 5 of the 199 and Weiss 2019; Finan and Schechter 2012; Stokes 2005; surveyed brokers and all 701 surveyed voters were present Vicente 2014). on the voter list (see details of sampling in SI Section D). For each individual, we then calculated their degree and betweenness centrality. Degree centrality is a count of the Measuring Voter Perceptions of Broker number of direct family ties that connect to an individMonitoring and Enforcement ual. Betweenness centrality is calculated as the number of shortest paths between all pairs of individuals in the netOur voter survey includes questions that allow us to test work on which the relevant individual is present, repredifferent mechanisms for why brokers are effective at ensenting whether an individual connects to different parts forcing clientelistic exchanges. In particular, we asked of the network. voter respondents who in their social networks knows about their vote (e.g., brokers, candidates, spouse, family members, etc.) We also asked voters what would happen if they did not vote for a candidate whose broker gave Measuring Social Preferences them money or goods (e.g., brokers cannot do anything, Following Finan and Schechter (2012), our measure of I will feel ashamed and will not be able to approach broreciprocity is constructed from play in a trust game. As kers; I will stop receiving handouts in the future; brokers all play was one shot and anonymous (see SI Section B will verbally or physically abuse me). for details), this measures intrinsic (rather than instrumental) reciprocity. Average reciprocity among brokers and voters are 0.075 and 0.073, respectively. These values Control Variables are almost twice as high as the average reciprocity levels among the respondents in Finan and Schechter (2012) In addition to our primary independent variables we col(0.043). lected data on a number of respondent characteristics, including, gender, age, educational attainment, religion, employment, marital status, household size, household income, trust, and altruism. Measuring Vote Buying In the broker survey (see SI Section D), each respondent was presented a randomly sampled set of voters from their respective barangays who were also respondents in Empirical Results our voter survey. We asked the broker if his/her team (we call team A) as well as the other team (we call Team B) Here we present evidence for how social networks shape offered money or in-kind goods to each of the voters on brokers’ targeting strategy. We start by testing the two the list and, if so, the value of such offers. We also asked key assumptions of our theory, which are that clientelist surveyed voters whether they had received money or incampaign machines employ brokers who have significant kind goods from a campaign, which campaign made the ties to voters, and that brokers know their brokers, inoffer, and the amount offered. cluding their level of intrinsic reciprocity. We then show Based on this broker survey, 57% of the voter samsupport for our core hypotheses, finding that brokers tarple were offered something in exchange for their votes get vote buying at central voters in dense social networks in the 2016 municipal elections (see summary statistics and highly intrinsically reciprocal voters in sparse netin SI Table F.1). The average value of the transfer offered works. As a test of the mechanism driving this pattern, was $17.50. A day laborer in agriculture earns between $4 we show that socially central voters are more likely to and $5 so this is a sizable amount. This amount is about think that brokers know who they voted for and that they 15% of the voters’ reported average monthly household would lose access to future benefits if they renege on a income ($117.00). vote bought. Next, we show evidence for the efficacy of Brokers claim that their campaign (Team A) offered broker targeting strategies. Highly central voters in vilsomething to 34% of voters in our sample. These eslages with dense social networks are far more likely to timates of the incidence and amount of vote buying follow through and vote for candidates who bought their are consistent with other reports in the Philippines, the vote. Finally, we examine the robustness of our findings numbers from our voter survey, as well as several esagainst alternative theories. BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 9 Who Becomes a Broker? TABLE 1 Who Becomes a Broker? A key assumption underpinning our theory—and inforRespondent mation asymmetry models of brokers in general – is that Is a Broker clientelist campaign machines select brokers who have Reciprocity 0.004 strong social ties to voters (Assumption 1). To test this (0.164) assumption, we estimate the following equation using OLS: Degree centrality (DC) 0.001 (0.000) Broker jm = β0 + β1 Reci procit y j + β2 DC j + β3 BC j Betweenness centrality (BC) 0.009∗ (0.004) +\002 \002 X j + ηm + \003 j , Female −0.228∗ (0.038) where Broker j takes on a value of 1 if respondent j is a Age 0.004∗ broker in barangay m, and 0, if a voter. Reci procit y j is the (0.001) measure of respondent j’s intrinsic reciprocity. DC j and Elementary and below −0.164∗ BC j are measures of respondent j’s degree and between(0.038) ness centrality, respectively. X j is a vector of demographics and other characteristics (other controls) summarized High school and below −0.018 in SI Table F.1. Last, ηm is barangay fixed effect. Robust (0.038) standard errors are clustered at the barangay level. Non-Catholic 0.094 Table 1 gives us an idea of the profile of a typical bro(0.080) ker. Compared to the average voter, a typical broker is Employed 0.133∗ more likely to be male and employed. They are also older (0.022) and better educated. Interestingly, brokers are no differNever married −0.072 ent from voters in terms of their income, religious affili(0.040) ation, family status, or their level of intrinsic reciprocity. Number of family members −0.002 They are also no different from voters in terms of many (0.006) of their attitudes: altruism, negative reciprocity, risk prefMonthly household income −0.000 erence, time preference, and trust. (0.000) The most important finding for our purposes is that brokers tend to be central in their family networks. InterIntegrity 0.017∗ estingly, brokers tend to exhibit higher than average be(0.007) tweenness centrality but are no more likely than average Altruism −0.002 citizens to exhibit high degree centrality. In plain terms, (0.002) brokers do not have more direct family members relative Negative reciprocity −0.002 to an average citizen but instead are more likely to occupy (0.029) positions that connect to different families within the Risk preference 0.004 network. This is consistent with our PAP, where we state (0.005) that “broker connectivity maybe important in both low Time preference −0.012 and high density networks, but raw connectivity (e.g. de(0.031) gree centrality) is less important than betweenness cenTrust −0.014 trality (between group brokerage power) because these (0.032) individuals can connect distant parts of the network to the candidate [p. 4].”14 Constant 0.122 (0.098) Adjusted R2 0.172 Number of observations 900 Note: Unit of observation is respondent (broker and voter pooled). 14 Robust standard errors clustered at the barangay level in parentheIn SI Table F.2 (SI 12), we interact the reciprocity and network centrality measures of brokers with the village family network denses. Barangay fixed effects included; ∗ p < .05. sity. We find that the density of the village social network does not seem to affect who is selected as a broker. 10 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN FIGURE 3 Relative to the Average Voter, How Well Do Brokers Know Voters? Note: Each graph above displays a violin plot of the predicted probability of correctly guessing voter attributes, based on the OLS model specified above. The plots include a marker for the median predicted value for brokers and voters, along with a distribution of predicted responses. Do Brokers Know the Voters? say being married, and 0 otherwise. On the right-hand side, we include an indicator for whether the guesser was A second key assumption required by our theory is that a broker, along with a measure of the social distance bebrokers know voters considerably well (Assumption 2). To tween each guesser and subject in the network (SDi j ). All test this assumption, we quizzed brokers on whether and the other variables are as defined in the “Who Becomes a how well they know a list of 10 randomly sampled votBroker?” section. Finally, νi are voter i fixed effects. Roers from all registered voters in their barangay, whose atbust standard errors are clustered at the barangay level. tributes we know based on voter survey responses. We Figure 3 and SI Table F.3 (SI 13) report the realso quizzed voter respondents on how well they know sults. We find that brokers know their voters remarkthe same 10 randomly sampled other voters in their ably well. Compared to the average citizen, brokers are barangay. We test whether brokers are significantly more far more likely to know a given voter in their barangay knowledgeable about voters than the average person in (∼20 percentage points).15 Additionally, brokers do their locality using the following OLS model: much better than the average voter in terms of correctly guessing the age (±5 years) (∼15 percentage points), Vot er_chari j = β0 +β1 Broker j +β2 SDi j +β3 Reci procit y j + β4 DC j +β5 BC j +\002 \002 X j +νi +\003i j . 15 This is probably a lower bound estimate. A separate study shows that respondents can identify more individuals when shown a Vot er_chari j is a binary outcome that takes on a value picture along with a nickname (Haim, Nanes, and Davidson of 1 if respondent j correctly guessed voter i’s attribute, Forthcoming). BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 11 marital status (∼14 percentage points), and educational TABLE 2 Social Networks, Reciprocity, and Vote attainment (∼13 percentage points) of voters in their Buying village.16 Importantly, we find that brokers also have informaHandout = 1 tion about the social attributes and preferences of vot(1) (2) (3) ers. For example, they are significantly better able to correctly predict how voters would play in the first round of Voter reciprocity (VR) 0.017 0.008 0.383∗ the trust game described above (lower middle graph; col(0.109) (0.095) (0.183) umn 5). They are also significantly better able to correctly Degree centrality (DC) 0.000 0.000 −0.003∗ guess whether voters report that they are likely going to (0.001) (0.001) (0.001) punish someone who put them in a compromising situBetweenness centrality (BC) 0.012∗ 0.010∗ 0.024∗ ation (lower right graph; column 6). (0.003) (0.003) (0.009) Consistent with the idea that brokers’ positions in VR × Village network density −6.794+ the network facilitate their ability to collect information (3.526) about voters, we also find that the more socially proxiDC × Village network density 0.050∗ mate a respondent (guesser) is to a voter, the more likely (0.020) they are to correctly guess their attributes (SI Table F.3). BC × Village network density −0.221 This effect is significantly more pronounced for brokers (0.148) than it is for other voters (SI Table F.4).17 Full set of controls No Yes Yes In summary, for brokers to rely either on intrinsic or 2 instrumental reciprocity, they need to know voters reaAdjusted R 0.207 0.218 0.222 sonably well along these dimensions. And indeed, we find Number of observations 4,375 4,375 4,375 that Philippine brokers are remarkably skilled at identifyMean of DV 0.340 ing voters and their characteristics, including which have Note: Unit of observation is broker–voter dyad (i.e., a broker j’s high levels of reciprocity. These results are consistent report of whether their campaign offered money to voter i). Rowith the findings by Finan and Schechter (2012) in their bust standard errors clustered at the barangay level in parentheses. ∗ p < .05; + p < .10. Broker fixed effects, demographic controls, and study of brokers in Paraguay, but contrast with the works “other controls” included. of Schneider (2019) and Brierly and Nathan (2019) on relatively uniformed brokers in India and Ghana, where Handouti j is an indicator variable that takes on a respectively. value of 1 if broker j reported that voter i was offered money by their campaign, and 0 otherwise. Reci procit yi , DCi , and BCi are measures of voter i’s intrinsic reciSocial Networks, Reciprocity, and procity, degree centrality, and betweenness centrality, reVote Buying spectively. V N Dm is the network density of village m. Last, σ j are broker fixed effects. Robust standard errors Given that in our context, brokers are well positioned in are clustered at the barangay level. the social network and possess good information about Table 2 reports our results.18 Contra Finan and voter characteristics, how do they leverage their priviSchechter (2012) we find that in the full sample, brokers leged position and information about voters? We test our are not more likely to target voters with high levels of incore hypothesis that brokers will target clientelism to voters trinsic reciprocity. Instead, brokers are more likely to tardepending on the social network architecture (Hypothesis get centrally located voters with offers of electoral hand1), by estimating the following equation using OLS: outs (column 1). In particular, brokers are more likely to Handouti j = β0 +β1 Reci procit yi +β2 DCi +β3 BCi target voters with high betweenness centrality. Recall that these are voters who are more likely to occupy positions + β4 Reci procit yi ∗ VNDm +β5 DCi ∗ VNDm that connect to different family communities within the + β6 BCi ∗ VNDm +β7 VNDm +\002 \002 Xi +σ j +\003i j , overall network. These results hold even after controlling for voter demographics (column 2). 16 Although knowing these types of characteristics is not sufficient To test our core hypothesis, we examine whether to explain a broker’s capacity to monitor voters, it is necessary. If a broker does not know their voters, it is unlikely they will be unable voter targeting is contingent on the nature of the social to monitor their behavior (Schneider 2019). network. Column 3 in Table 2 interacts voter intrinsic 17 See, for example, coefficient estimates for the interaction term RB × SD in columns 2–4. 18 All results are robust to using jackknife resampling method. 12 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN FIGURE 4 Marginal Effects of Key Independent Variables on the Likelihood of Being Targeted for Vote Buying Note: The figures display marginal effects from column 3 in Table 2. As overall village network density increases, brokers are more likely to target voters that are central in the social network. reciprocity, betweenness (BC) and degree (DC) with lage, and increases by 17 percentage points in a highvillage network density. The key results are shown in density village. Figure 4. Where overall network density is low, brokers do target intrinsically reciprocal voters, but as those ties grow more dense brokers switch strategies and begin tarDo Social Networks Activate Instrumental geting voters with high degree centrality over intrinsically Reciprocity? reciprocal individuals. That is, brokers start targeting more central voters with a greater number of family The results so far show evidence of brokers’ contingent members, even as they continue to rely on their pritargeting strategy, but we also want to investigate the mary strategy of targeting voters with high betweenness mechanisms driving these results. According to our story, centrality.19 social networks activate instrumental reciprocity. SpecifTo get a sense of the magnitude of these effects, we ically, although the ballot is secret, voters more central in calculated the probability that a voter will be targeted the network will perceive that information about their vote with handouts if her intrinsic reciprocity increased from choice is more likely to be accessed by brokers (Hypothesis the 10th to the 90th percentile of the sample distribu2a). In addition, they will feel they are more likely to be tion. In a low-density village (at the 10th percentile of the sanctioned if they defected (Hypothesis 2b). To test these distribution), this voter’s probability of being targeted mechanisms, we estimate the following equation using increases by 3 percentage points. In a high-density vilOLS: lage (at the 90th percentile), this voter’s probability of Vot er_perce pt ioni = β0 +β1 Reci procit yi +β2 DCi +β3 BCi being targeted decreases by 6 percentage points. Likewise, if a voter’s degree centrality increased from the 10th + β4 Reci procit yi ∗ VNDm +β5 DCi ∗ VNDm to the 90th percentile, her probability of being targeted + β6 BCi ∗ VNDm +β7 VNDm +\002 \002 Xi +\003i , decreases by 10 percentage points in a low-density vilwhere Vot er_perce pt ioni is voter i’s perception of monitoring and enforcement of the clientelistic exchange, and 19 all left-hand side variables are as defined in “Social netIn SI Table F.5, we dropped voters who received inducements from both campaigns as reported by Team A brokers. We find that works, reciprocity, and vote buying” section. our findings are stronger for the conditionality of targeting inTo learn who voters believe know about their vote, trinsically reciprocal voters (while the findings on the conditionalwe examine the correlates of voters’ responses to the ity of targeting central voters essentially remained the same). This question, “Other than yourself, who among the followsuggests that targeting intrinsically reciprocal voters becomes less important when voters are receiving inducements from multiple ing persons knows about your vote this last election— candidates. brokers, family members, and ‘others’ (i.e. friends, BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 13 TABLE 3 Do Voters Perceive Brokers to Be Effective Monitors? Brokers Family Others Know Knows Know (1) (2) (3) (4) (5) (6) (7) (8) (9) Voter reciprocity (VR) 0.069 0.042 0.102 0.291 0.006 −0.909 −0.083 −0.140 −0.346 (0.147) (0.157) (0.537) (0.208) (0.154) (0.449) (0.153) (0.157) (0.420) Degree centrality (DC) 0.001 0.001 −0.002 −0.001 0.000 −0.001 0.001 0.001 0.002 (0.001) (0.001) (0.002) (0.001) (0.001) (0.002) (0.001) (0.001) (0.003) Betweenness centrality (BC) 0.005 0.004 0.002 0.009 0.003 −0.015 −0.002 −0.001 −0.024 (0.004) (0.005) (0.012) (0.007) (0.007) (0.021) (0.006) (0.006) (0.023) VR × Village network density −1.303 15.231∗ 3.198 (9.942) (7.102) (6.887) DC × Village network density 0.051∗ 0.022 −0.017 (0.021) (0.040) (0.041) BC × Village network density 0.047 0.297 0.341 (0.199) (0.352) (0.315) Mean of DV among voter respondents 0.054 0.415 0.072 Full set of controls No Yes Yes No Yes Yes No Yes Yes Adjusted R 2 0.010 0.043 0.050 0.002 0.192 0.194 −0.002 0.018 0.024 Number of observations 668 668 668 668 668 668 668 668 668 Note: Unit of observation is the voter respondent. Robust standard errors clustered at the barangay level are in parentheses; ∗p < .05. Voter demographic controls, and “other controls” are included. coworkers, and religious leaders)?” Table 3 reports the when degree centrality increases from the 10th to the results. Not surprisingly, a significant share of the voter 90th percentile. The same is not true for reciprocal voters respondents (42%) report that family members know (Figure 5a, right figure). about their vote, while only about 7% of the responWe also asked voters what would happen if they did dents report that other individuals in their social network not vote for a candidate whose broker gave them money know about their vote. This highlights the importance or goods: Will they feel ashamed? Will brokers reprimand of family networks in particular for spreading the type them? Will they stop receiving handouts in the future? of information that would allow brokers to monitor the Or are brokers unable to do anything? The goal is to clientelist exchange.20 show that not only do socially central voters believe broFor our purpose, we are interested in voters’ percepkers can monitor their behavior but that they also believe tion of broker monitoring. Although only about 5% of there are consequences for defecting from the clientelist the respondents, on average, believe brokers know about exchange. Table 4 shows that this is indeed the case. Vottheir vote, this statistic masks the heterogeneity in voters with high-degree centrality are not only more likely ers’ perceptions driven by their degree centrality and to believe brokers will reprimand them (column 5), imvillage network density (Table 3, column 3; Figure 5a, portantly, they are also more likely to believe that they left figure). According to the model, in a high-density will stop receiving handouts in future elections should village the probability of thinking that a broker knows they defect (column 8). Moreover, as the village network about one’s vote choice increases by 18 percentage points grows denser, voters with high degree centrality are more likely believe that they will stop receiving handouts in the 20 Interestingly, reciprocal and central voters (measured by either future (Table 4, column 9; Figure 5b, left). In fact, the their degree or betweenness centrality) are no more likely than the average respondent to report that family members and other indiprobability that a voter in a high-density village believes viduals know about their vote (columns 4, 5, 7, and 8). However, that she will stop receiving handouts in the future if her as the village network becomes denser, reciprocal voters are much degree centrality increased from the 10th to the 90th permore likely to report that their family members know about their vote (column 6). centile increases by as much as 35 percentage points. 14 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN FIGURE 5 Marginal Effects of Key Independent Variables on the Mechanisms of Vote Buying: (A) Monitoring and (B) Enforcement Note: The top panel shows marginal effects from Table 3, column 3. The bottom panel shows marginal effects from Table 4, column 9. As overall village network density increases, central voters are more likely to perceive that they can be monitored (broker knows their vote) and that punishments can be enforced (will not receive handouts in the future). Efficacy of Vote Buying where Vot ed_ f or_TeamA_candidat ei j is an indicator that takes on a value of 1 if voter i supported Team A’s The last piece in our puzzle is to show that brokers’ tarmayoral candidate, and 0 otherwise. Note that this outgeting strategy contingent on network structure is effeccome variable is based on the voter survey (and not on tive. Given that the contingent exchange in our context is brokers’ guesses of voter behavior). money for votes, it should be the case that being offered TAi j is an indicator that equals 1 if Team broker j remoney by Team A translates into votes for Team A canported that voter i was offered money by Team A, and 0 didates. We test Hypothesis 3, that electoral handouts will otherwise. We include in the regression the variable T Bi j , be effective, and its efficacy will depend on the social netwhich is an indicator for whether Team B offered money work architecture, by estimating an OLS model using the to voter i, according to Team A’s brokers. All the other following equation: right-hand side variables are as previously defined. RoVot ed_ f or_TeamA_candidat ei j bust standard errors are clustered at the village level. Table 5 reports on the suggestive efficacy of vote buy= β0 +β1 TAi j +β2 Reci procit yi +β3 DCi +β4 BCi ing. We find that the targeting of voters does in fact + β5 TAi j ∗ Reci procit yi +β6 TAi j ∗ DCi +β7 TAi j ∗ BCi correlate with voters’ self-reported behavior at the polls. This is true both at the extensive (i.e., offered money or + β8 TAi j ∗ Reci procit yi ∗ V N Dm not) and intensive (i.e., amount of money offered) mar+ β9 TAi j ∗ DCi ∗ V N Dm +β10 TAi j ∗ BCi ∗ V N Dm gins. At the extensive margin, being offered any money + β V N D ++β T B +\002 X +σ +\003 , \002 by a campaign is associated with a 28 percentage point 11 m 12 ij i j i BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 15 TABLE 4 Do Social Networks Activate Instrumental Reciprocity? Will Feel Brokers Will Stop Receiving Brokers Can’t Ashamed Will Reprimand Handouts in the Future Do Anything (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Voter reciprocity (VR) 0.224 −0.001 −0.010 0.234 0.123 −0.720 0.262 0.128 −0.500 0.343 0.075 −0.542 (0.185) (0.174) (0.492) (0.191) (0.190) (0.653) (0.207) (0.209) (0.524) (0.199) (0.176) (0.443) Degree centrality (DC) 0.000 0.001 −0.001 0.001∗ 0.002∗ 0.001 0.002∗ 0.002∗ −0.003 −0.000 0.000 −0.000 (0.001) (0.001) (0.002) (0.001) (0.001) (0.002) (0.001) (0.001) (0.002) (0.001) (0.001) (0.002) Betweenness centrality (BC) 0.001 −0.005 0.023 −0.001 −0.003 0.029 −0.014∗ −0.015∗ 0.017 0.005 0.004 0.004 (0.006) (0.006) (0.019) (0.005) (0.005) (0.020) (0.005) (0.005) (0.013) (0.007) (0.006) (0.023) VR × Village network density 0.151 14.982 10.943 10.575 (6.891) (10.621) (8.178) (7.092) DC × Village network density 0.037 0.001 0.077∗ 0.006 (0.032) (0.024) (0.026) (0.024) BC × Village network density −0.469 −0.511 −0.468∗ 0.009 (0.241) (0.303) (0.209) (0.366) Mean of DV among voter respondents 0.306 0.172 0.182 0.428 Full set of controls No Yes Yes No Yes Yes No Yes Yes No Yes Yes Adjusted R2 −0.002 0.058 0.057 0.005 0.019 0.024 0.013 0.038 0.056 0.001 0.113 0.110 Number of observations 668 668 668 668 668 668 668 668 668 668 668 668 Note: Unit of observation is the voter respondent. Robust standard errors clustered at the barangay level are in parentheses; ∗p < .05. Voter demographic controls, and “other controls” are included. increase in the likelihood of voting for the campaign’s stronghold), and the log of precinct population. Overall, candidate. At the intensive margin, each additional we find that brokers are more likely to target vote buyPHP50 offered is associated with a 1% increase in the ing at areas where their candidate has strong preexisting likelihood of voting for the campaign’s candidate, on avsupport but are no more likely to target voters in small erage (see SI Table F.6). precincts.21 All of our main findings remain statistically Consistent with Hypothesis 3, we find that there is significant and are of similar substantive magnitude after substantial heterogeneity in how being offered money controlling for the additional variables, suggesting that translates to vote choice. In particular, money offered village density is not merely a proxy for the measures proseems to be more effective in mobilizing the support of posed by the alternative hypotheses. central voters in dense social networks (column 3). We In addition, we explore whether brokers’ targeting do not find, however, that intrinsic reciprocity matters strategy is also contingent on a precinct’s population or more in the efficacy of vote buying in sparse social netcandidate support by interacting these alternative meaworks, counter to what we expected, although not sursures with voter reciprocity and centrality (columns 3 prising in light of the fact the brokers do not primarily and 5). Interestingly, as the campaign’s baseline vote target intrinsically reciprocal voters. share increases (i.e., when the precinct is a stronghold), brokers are more likely to target high betweenness voters. Although we did not have a specific theory for what Alternative Hypotheses might be driving this effect, we think the way that network targeting strategies differ in core and swing districts As we discussed in the “Alternative hypotheses” section, is an interesting avenue for future research. For the purour findings may be picking up alternative broker targetposes of this study, even after controlling for these ading strategies documented in other contexts. We test our ditional interactions terms, our main results remain retheory against the two alternative hypothesis that brokers markably consistent (columns 4 and 6). The choice to will target clientelism depending on whether it is a “swing” target central voters in dense networks does not seem or “core” precinct, and that brokers will target clientelism to voters depending on the number of voters per polling station. Results are reported in SI Table F.7 (SI 17). Columns 21 (1) and (2) report the results analogous to columns 2 and We note, however, that all of Mahamot’s precincts are relatively small and of similar size (see SI Figure E.3). Rueda’s (2017) 3 of Table 2, controlling for the campaign’s precinct-level argument could very well hold where precincts are larger and vote share in 2013 (a proxy for whether the precinct is a more variable. 16 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN TABLE 5 Is Vote Buying Effective? and character of electoral clientelism. We hope this article contributes to this iterative process of empirical evalVoted for Team A Candidate = 1 uation and theoretical refinement. Specifically, we have (1) (2) (3) endeavoured to directly probe two sets of findings in Offered by Team A (TA) = 1 0.280∗ 0.251∗ 0.256∗ the literature—(1) the importance of both instrumental (0.027) (0.056) (0.053) Voter reciprocity (VR) 0.002 −0.065 −0.063 and intrinsic reciprocity in sustaining clientelism (Finan (0.188) (0.250) (0.250) and Schechter 2012; Stokes 2005), and (2) the capacity Degree centrality (DC) −0.000 −0.000 −0.000 (or lack thereof) of brokers to identify and utilize voter (0.001) (0.001) (0.001) reciprocity to facilitate clientelistic exchanges (Brierly Betweenness centrality (BC) −0.006 −0.007 −0.007 and Nathan 2019; Schneider 2019). We also considered (0.006) (0.008) (0.008) the strength of our argument in light of other explanaTA × VR 0.115 0.339 (0.231) (0.429) tions for targeting behavior—namely, swing versus core TA × DC 0.001 −0.002 arguments (Auyero 2000; Stokes et al. 2013), and polling (0.001) (0.001) station size (Rueda 2017)—and find that our results hold TA × BC −0.004 −0.008 when controlling for these alternative explanations. We (0.009) (0.015) believe that our core finding about the importance of soTA × VR × Village network density −4.110 (7.394) cial network density for shaping broker targeting strateTA × DC × Village network density 0.043∗ gies is generalizable across a wide variety of political con(0.017) texts, though the details of particular social networks will TA × BC × Village network density 0.103 certainly vary. (0.228) That said, there are features of the Philippines conAdjusted R2 0.175 0.175 0.176 text that may place important scope conditions on our Number of observations 3,839 3,839 3,839 Mean of DV 0.379 findings. First, although our study site is typical of most Note: Unit of observation is broker–voter dyad. Robust standard municipalities in the Philippines, it is smaller and more errors clustered at the barangay level; ∗p < .05. Regression conrural than other contexts where scholars have studied trols for being offered by Team B, broker fixed effects, and voter clientelism. Though we find that our results hold regarddemographic controls. less of the size of the polling precincts, it may be that the utility of the strategies we describe in this article dito be contingent on the candidates’ level of preexisting minish as districts become significantly larger or more support.22 urban (see Nathan 2019, for an analysis of urban clientelism). Second, the Philippine context is one is which political parties are extremely weak. For us, this is a feature, not a bug. Much of the foundational work on clienConclusion telism focuses on environments in which parties are relatively strong and voter attachments relatively stable. But We provide new empirical evidence consistent with the many developing democracies look more like the Philipcentral prediction of the classic theories of clientelism: pines than Argentina and thus our results may help shed namely, that brokers use dense social networks to monmore light on how clientelism operates in such contexts itor and enforce vote-buying bargains. However, our (see also Muñoz 2019). Third, although national parties findings also accommodate recent empirical evidence are weak, local political machines tend to be strong and suggesting that brokers target electoral clientelism to stable in the Philippines. It is possible that the dynamic intrinsically reciprocal voters. we describe is different where brokers are more like guns One question regarding studies like ours is about the for hire than a semi-permanent members of an ongoing generalizability of the results. To what extent might our political machine.23 Fourth, our context is characterized findings travel to other political or country contexts, and by multiple campaigns competing with each other, which what are the limits to generalizability? These are legitilimits our ability to test whether our theory would apply mate questions. As discussed, we view this article as part to contexts of a single vote-buying party. We suspect the of a productive dialogue in the field about the causes dynamics would be different in the latter case. 22 23 The triple interaction term of voter centrality, barangay denSee Aspinall and Hicken (2020) for an analysis of the implication sity, and 2013 candidate vote share is not a significant predictor of different local brokerage structures in the Philippines and Inof vote buying. donesia. BROKERS, SOCIAL NETWORKS, RECIPROCITY, AND CLIENTELISM 17 One important caveat is that our study occurred in Cruz, Cesi, Julian Labonne, and Pablo Querubin. 2017. “Politia locality in which family networks are the salient socian Family Networks and Electoral Outcomes: Evidence from the Philippines.” American Economic Review 107(10): cial network. As mentioned, other local networks will 3006–37. be more salient in other contexts. Although, again, we Cruz, Cesi, Philip Keefer, and Julien Labonne. 2016. “Inbelieve our findings might be useful in understanding cumbent Advantage, Voter Information, and Vote Buythe targeting strategies brokers employ using other neting.” IDB Publications (Working Papers) 7730. Washworks, this is something that needs further testing. An ington, DC: IDB Publications. https://publications. encouraging sign for the external validity of our findiadb.org/publications/english/document/IncumbentAdvantage-Voter-Information-and-Vote-Buying.pdf. ings is the recent study of Duarte et al. (2019), which Duarte, Raúl, Frederico Finan, Horacio Larreguy, and Laura finds a similar pattern of relationships in Paraguay. Taken Schechter. 2019. “Brokering Votes with Information Spread together, our findings may help explain why clientelism Via Social Networks.” NBER Working Paper No. 26241. can be so durable and adaptable. Campaigns are able Berkeley, CA: University of California. to draw on both instrumental reciprocity engendered by Finan, Frederico, and Laura Schechter. 2012. “Vote-Buying and social ties, and intrinsic reciprocity on the part of the Reciprocity.” Econometrica 80(2): 863–81. voters. Haim, Dotan, Matthew Nanes, and Michael Davidson. Forthcoming. “Family Matters: The Double-Edged Sword of Police-Community Connections.” Journal of Politics. Hicken, Allen. 2011. “Clientelism.” Annual Review of Political References Science 14: 289–310. Hicken, Allen, Edward Aspinall, and Meredith Weiss. 2019. Aspinall, Edward, and Allen Hicken. 2020. “Guns for Hire and Electoral Dynamics in the Philippines: Money Politics, PatronEnduring Machines: Clientelism Beyond Parties in Indoneage and Clientelism and the Grassroots. Singapore: National sia and the Philippines.” Democratization 27(1): 137–56. University of Singapore Press. Aspinall, Edward, and Mada Sukjamati. 2016. Electoral DynamHicken, Allen, Stephen Leider, Nico Ravanilla, and Dean Yang. ics in Indonesia: Money Politics, Patronage and Clientelism at 2015. “Measuring Vote-Selling: Field Evidence from the the Grassroots. National University of Singapore, Singapore: Philippines.” American Economic Review 105: 352–56. NUS Press. Hicken, Allen, Stephen Leider, Nico Ravanilla, and Dean Yang. Auerbach, Adam, and Tariq Thachil. 2018. “How Clients Select 2018. “Temptation in Vote-Selling: Evidence from a Field Brokers: Competition and Choice in India’s Slums.” AmeriExperiment in the Philippines.” Journal of Development Ecocan Political Science Review 112: 775–91. nomics 131: 1–14. Auyero, Javier. 2000. “The Logic of Clientelism in Argentina: Holland, Alisha C., and Brian Palmer-Rubin. 2015. “Beyond An Ethnographic Account.” Latin American Research Rethe Machine: Clientelist Brokers and Interest Organizaview 35: 55–81. tions in Latin America.” Comparative Political Studies 48(9): Brierly, Sarah, and Noah Nathan. 2019. “Connections of Party 1186–1223. Brokers.” Working Paper. https://tinyurl.com/sgys4xy. Hutchcroft, Paul, and Joel Rocamora. 2003. “Strong Demands Brusco, Valeria, Marcelo Nazareno, and Susan Stokes. 2004. and Weak Institutions: The Origins and Evolution of the “Vote Buying in Argentina.” Latin American Research ReDemocratic Deficit in the Philippines.” Journal of East Asian view 39: 66–88. Studies 3(2): 259–92. Calvo, Ernesto, and Maria Victoria Murillo. 2004. “Who DeKeefer, Philip. 2007. “Clientelism, Credibility, and the Policy livers? Partisan Clients in the Argentine Electoral Market.” Choices of Young Democracies.” American Journal of PolitiAmerican Journal of Political Science 48(4): 742–57. cal Science 51(4): 804–21. Camp, Edwin. 2017. “Cultivating Effective Brokers: A Party Kitschelt, Herbert, and Steven Wilkinson. 2007. “CitizenLeader’s Dilemma.” British Journal of Political Science 47: Politician Linkages: An Introduction.” In Patrons, Clients 521–43. and Policies: Patterns of Democratic Accountability and Political Competition, eds. Herbert Kitschelt and Steven WilkinChauchard, Simon. 2018. “Electoral Handouts in Mumbai son. New York: Cambridge University Press. Elections: The Cost of Political Competition.” Asian Survey 58(2): 341–64. Kramon, Eric. 2017. Money for Votes: The Causes and Consequences of Electoral Clientelism in Africa. New York: CamCorstange, Daniel. 2009. “Sensitive Questions, Truthful Anbridge University Press. swers? Modeling the List Experiment with LISTIT.” Political Analysis 17(1): 45–63. Leider, Stephen, Markus Mobius, Tanya Rosenblat, and QuocAnh Do. 2009. “Directed Altruism and Enforced ReciCorstange, Daniel. 2018. “Clientelism in Competitive and Unprocity in Social Networks.” Quarterly Journal of Economics competitive Elections.” Comparative Political Studies 51(1): 124(4): 1815–51. 76–104. Mares, Isabela, and Lauren Young. 2016. “Buying, ExpropriatCox, Gary, and J. Morgan Kousser. 1981. “Turnout and Rural ing, and Stealing Votes.” Annual Review of Political Science Corruption: New York as a Test Case.” American Journal of 19: 267–88. Political Science 25(4): 646–63. Mendoza, Ronald U., Edsel L. Beja Jr, Victor S. Venida, and Cruz, Cesi. 2019. “Social Networks and the Targeting of Vote David B. Yap. 2016. “Political Dynasties and Poverty: MeaBuying.” Comparative Political Studies 52: 382–411. 18 NICO RAVANILLA, DOTAN HAIM, AND ALLEN HICKEN surement and Evidence of Linkages in the Philippines.” OxSchmidt, Steffen, James Scott, Carl Lande, and Laura Guasti, ford Development Studies 44(2): 189–201. eds,. 1977. Friends, Followers, and Factions: A Reader in PolitMuñoz, Paula. 2014. “An Informational Theory of Campaign ical Clientelism. Berkeley, CA: University of California Press. Clientelism.” Comparative Politics 47: 79–98. Schneider, Mark. 2019. “Do Local Leaders Know Their Voters? Muñoz, Paula. 2019. Buying Audiences: Clientelism and ElecA Test of Guessability in India.” Electoral Studies 61. https:// toral Campaigns When Parties are Weak. Cambridge: Camdoi.org/10.1016/j.electstud.2019.102049. bridge University Press. Scott, James. 1972. “Patron-Client Politics and Political Change Nathan, Noah L. 2019. Electoral Politics and Africa’s Urban in Southeast Asia.” American Political Science Review 66(1): Transition: Class and Ethnicity in Ghana. New York: Cam91–113. bridge University Press. Sidel, John. 1999. Capital Coercion and Crime: Bossism in the Nichter, Simeon. 2008. “Vote Buying or Turnout Buying? MaPhilippines. Stanford, CA: Stanford University Press. chine Politics and the Secret Ballot.” American Political SciSobel, Joel. 2005. “Interdependent Preferences and Recience Review 102: 19–31. procity.” Journal of Economic Literature 43: 392– Nichter, Simeon. 2018. Votes for Survival: Relational Clientelism 436. in Latin America. New York:Cambridge University Press. Stokes, Susan. 2005. “Perverse Accountability: A Formal Model PSA. 2019. Urban Population in the Philippines. Philippines Staof Machine Politics with Evidence from Argentina.” Ameritistical Authority. http://www.psa.gov.ph/content/urbancan Political Science Review 99(3): 315–25. population-philippines-results-2015-census-population. Stokes, Susan, Thad Dunning, Marcelo Nazareno, and Valeria Ravanilla, Nico, Michael Davidson, and Allen Hicken. 2020. Brusco. 2013. Brokers, Voters, and Clientelism: The Puzzle of “Voting in Clientelistic Social Networks: Evidence from the Distributive Politics. Cambridge, UK: Cambridge University Philippines.” Working Paper. https://bit.ly/3mqR8Dr. Press. Ravanilla, Nico. 2017. “Motives in Pork Distribution: Partisan Vicente, Pedro. 2014. “Is Vote Buying Effective? Evidence Bias or Patronage.” Working Paper. San Diego, CA: Univerfrom a Field Experiment in West Africa.” Economic Journal sity of California San Diego. https://tinyurl.com/ql6msjk. 124(574): 356–87. Ravanilla, Nico, Renard Sexton and Dotan Haim. ForthYoganarasimhan, Hema. 2012. “Impact of Social Network coming. “Deadly Populism: How Local Political Outsiders Structure on Content Propagation – A Study Using YouTube Drives Duterte.s War on Drugs in the Philippines.” Journal Data.” Quantitative Marketing and Economics 10(1): 111– of Politics. 50. Rocina, Jose Aims R. 2019. “Muntinlupa City: The Use of Monikers as a Manifestation of Programmatic Politics.” In Supporting Information Electoral Dynamics in the Philippines: Money Politics, Patronage, and Clientelism at the Grassroots, eds. Allen Hicken, Additional supporting information may be found online Edward Aspinall, and Meredith Weiss. Singapore: National University of Singapore Press. in the Supporting Information section at the end of the Rueda, Miguel R. 2017. “Small Aggregates, Big Manipulaarticle. tion: Vote Buying Enforcement and Collective Monitoring.” Appendix A: Recruitment Script American Journal of Political Science 61(1): 163–77. Appendix B: Measuring Intrinsic Reciprocity Schaffer, Frederic. 2005. “Clean Elections and the Great Unwashed: Educating Voters in the Philippines.” Occasional Appendix C: Measuring Voter Behavior Paper 21. Princeton, NJ: School of Social Science, IAS. Appendix D: Surveying Brokers and Voters Schaffer, Frederic, ed,. 2007. Elections for Sale: The Causes and Appendix E: SI Figures Consequences of Vote Buying. Boulder, CO: Lynne Rienner. Appendix F: SI Tables
6 Building Cooperation among Groups in Conflict: An Experiment on Intersectarian Cooperation in Lebanon Han Il Chang New York University–Abu Dhabi Leonid Peisakhin New York University–Abu Dhabi Abstract: Societies divided along ethnic or religious lines suffer from persistent conflict and underprovision of public goods. Scholarly understanding of how to strengthen intergroup cooperation remains limited. In this study, we set out to test the effectiveness of two interventions on intergroup cooperation: cross-group expert appeal and participation in a cross-group discussion. The laboratory-in-the-field experiment is set in Lebanon’s capital, Beirut, and involves interactions between 180 Shia and 180 Sunni Muslim participants. We find that the expert appeal increases intersectarian cooperation in settings that do not entail reciprocal exchange. On average, cross-sectarian discussions do not improve cooperation, but those discussions in which participants delve deeply into the conflict’s causes and possible remedies are associated with greater cooperation. Neither intervention diminishes the effectiveness of sectarian clientelistic appeals. The policy implication of our study is that intergroup cooperation can be strengthened even in regions as bitterly divided as the Middle East. Replication Materials: The data, code, and any additional materials required to replicate all analyses in this article are available on the American Journal of Political Science Dataverse within the Harvard Dataverse Network, at: https://doi.org/10.7910/DVN/CQXHTH. S ocieties that are home to multiple ethnic or reother” (Allport 1954; Chen and Li 2009; Tajfel et al. 1971) ligious groups are known to have lower levels of and by making sanctioning more difficult because of limpublic goods provision (Alesina, Baqir, and Easterly ited information about the other group (Habyarimana 1999) and higher levels of conflict (Horowitz 1985). That et al. 2007). We advance the literature by exploring how is because cooperation is generally lower across group intergroup cooperation can be fostered. Much work in boundaries than within groups, and prevalence of inthe social sciences has explored the determinants of cogroup favoritism impedes outcomes that are beneficial to operation within groups; ours is among the few studies society as a whole (Akerlof and Kranton 2000). Difficulty in political science to examine how to build cooperawith cooperation across group boundaries is a common tion across group boundaries (e.g., Cederman et al. 2011; problem globally and has been particularly acute in the Fearon and Laitin 1996; Larson 2017; Miguel 2004). Middle East in recent decades. There, sectarian tensions Our starting point is the insight from the common have contributed to political instability and economic ingroup identity theory (Gaertner et al. 1993) that interunderperformance. In this project, we test the various group cooperation can be increased through activation means of encouraging cooperation across sectarian lines of a common overarching identity (for an empirical ilbetween Shia and Sunni Muslims. lustration, see Miguel 2004). There are different ways to There is a consensus in the literature on intergroup activate a common identity: through a top-down appeal cooperation that differences in group identity undermine or a bottom-up engagement. One way to build consencooperation by inhibiting altruistic behavior toward “the sus and construct overarching identity is by facilitating a Han Il Chang is Research Associate, New York University–Abu Dhabi, P.O. Box 129188, Abu Dhabi, United Arab Emirates (hc665@nyu.edu). Leonid Peisakhin is Assistant Professor of Political Science, New York University–Abu Dhabi, P.O. Box 129188, Abu Dhabi, United Arab Emirates (leonid.peisakhin@nyu.edu). We thank for helpful comments Thad Dunning, Susan Hyde, Peter van der Windt, Jeff Jensen, Rebecca Morton, Cole Tanigawa-Lau, three anonymous reviewers, and seminar participants at UC-Berkeley, NYUAD, and the 2016 WESSI Alumni Workshop. We are grateful to Jad Mahmoud Halabi and Vladimir Kurnosov, who provided excellent research assistance, and to Alicia Jammal at Information International, whose help with the logistics of the study was invaluable. The project was funded by an NYUAD Research Enhancement Grant. It was approved by NYUAD’s institutional review board under protocol #080-2016. American Journal of Political Science, Vol. 63, No. 1, January 2019, Pp. 146–162 \002 C 2018, Midwest Political Science Association DOI: 10.1111/ajps.12397 146 BUILDING COOPERATION AMONG GROUPS IN CONFLICT 147 discussion aimed at establishing cross-group commonThe content of the expert appeal catalyzes subsequent ality between members of competing groups. Studies group discussions in the same way that reading materion the causes of intergroup cooperation (e.g., Kerr and als provide a prompt for discussions in most studies on Kaufman-Gilliland 1994) and also the related literature deliberation and in the way that all real-life discussions on deliberative democracy (e.g., Fishkin 1997) have confollow on from some stimulus. We back out the effect of sidered the effectiveness of group discussions. A major group discussion by comparing that intervention against limitation of experimental work in this vein is that the the baseline of the expert appeal. theoretical expectations have been tested mostly in setThe effect of the two interventions on the levels of tings where group identities are arbitrarily assigned (for intergroup cooperation is tested in a series of games. We a review, see Balliet 2010). In this project, we explore the measure unconditional cooperation by observing particeffectiveness of a discussion intervention in a setting with ipants’ vote for sectarian candidates in simulated elechighly salient real-world religious identities. We also test tions and, separately, by recording how much of a certhe effectiveness of a completely novel intervention—an tain fixed financial resource participants allocate among appeal by experts—on intergroup cooperation. Experts anonymous members of their own and the opposite sect are commonly believed to influence public opinion eiin a standard other–other allocation game. Conditional ther by providing new information or legitimating certain cooperation is measured by observing the level of contripoints of view (Henrich and Gil-White 2001; Zaller 1992). butions in a standard public goods game. We distinguish between two different types of coThe study also examines the effects of a clientelistic operation: unconditional and conditional. Unconditional exchange by randomly exposing some participants to a cooperation entails friendly behavior toward a member of financial incentive to vote for their co-sectarian in the a different group with no expectation of reciprocity. Conelection game. We introduce this manipulation out of ditional cooperation takes place only if one expects the a concern that clientelism in the form of one-shot payopposite group to respond cooperatively. Unconditional ments in exchange for votes is common in Lebanon and cooperation is a form of altruistic behavior that signals across much of the Middle East (Corstange 2016; Lust a high regard toward the outgroup and a lower level of 2009). The presence of this intervention allows us to test ingroup favoritism. In the world of politics, voting for a whether expert appeal and group discussion reduce the candidate from the opposite ethnic or religious group out effectiveness of clientelism. Ours is the first study to conof support for their policy proposals can be an instance sider how interventions designed to increase intergroup of unconditional cooperation insofar as one is voting in cooperation interact with clientelism. support of a policy that might have no direct bearing on We find that watching experts discuss the benefits one’s self-interest. of intersectarian cooperation increases unconditional coIn a laboratory-in-the-field experiment involving 180 operation, but there is no effect on conditional cooperShia and 180 Sunni Muslims in Beirut, Lebanon, we test ation. Leveraging self-reported scores of cross-sectarian what effect watching a short, prerecorded expert appeal trust, we find that the expert appeal intervention fails to cooperate and, separately, participation in a crossto increase cross-sectarian trust, whereas increased trust sectarian discussion have on the level of intersectarian comight be a precondition for higher levels of conditional operation.1 The first intervention—the expert appeal— cooperation. We theorize that expert appeals might fail to is modeled after a televised political talk show. In an improve unconditional cooperation because such appeals unscripted conversation, four prominent journalists— more easily activate intrinsic motivations for cooperation two Shias and two Sunnis—discuss Lebanon’s problems but are not very effective at building intergroup trust. caused by sectarianism and conclude that the counParticipation in a cross-sectarian group discussion try would benefit substantially from greater cooperation about cooperation appears to have no effect either on across sectarian lines. In the second intervention, after unconditional or conditional cooperation. However, we watching the expert appeal, participants in mixed sixpresent suggestive observational evidence that the effecperson Sunni–Shia groups engage in face-to-face discustiveness of group discussion is highly heterogeneous and sions about issues relating to intersectarian cooperation. varies with substantive discussion depth. A deep and substantive exchange that touches on many topics relating 1 Lebanon is home to a large Christian population. We exclude to intergroup cooperation is correlated with higher levChristians from this study. Substantively, the problem of Shia– els of both types of cooperation. We also illustrate how Sunni relations is of broad relevance across the region, including clientelism impedes intergroup cooperation in divided in countries with Christian populations. Logistically, inclusion of a third group would have made the study more difficult to execute societies. Neither intervention succeeds in diminishing and more costly. the effectiveness of clientelism. 148 HAN IL CHANG AND LEONID PEISAKHIN Our research suggests that intergroup cooperation Lebanese electorate can be inferred to have sold their can be strengthened under certain conditions even in votes in the 2009 parliamentary election, according to places as bitterly divided as the Middle East. The primary Corstange (2012, 493). policy implication of this study is that dissemination of consensual cross-sectarian expert calls to cooperate—a top-down intervention—may in fact bring about greater Hypotheses cooperation across group lines in unconditional exchanges. Cooperation conditional on one’s expectation We hypothesize that both interventions—expert appeal about the behavior of the other group is likely to improve and participant discussion—should increase uncondionly once intergroup trust levels increase. Our findings tional and conditional intersectarian cooperation. Our also indicate that intersectarian cooperation in divided starting point is the insight from common ingroup idensocieties is unlikely to take root as long as clientelistic tity theory that models how activation of an overarching vote buying remains a common practice. group identity might decrease bias among members of the smaller constituent groups that make up the larger whole (Gaertner and Dovidio 2000; Gaertner et al. 1993). Context In this case, the smaller constituent groups are Shia and Sunni residents of Lebanon. The overarching group This project entailed interactions between residents identity is the Lebanese national identity.2 We expect of Beirut, the capital of Lebanon. Lebanon is a highly the common Lebanese national identity to come to the fractured society. Fissures run along the lines of religious fore for those exposed to the interventions. According to sectarian identities and are deeply historically rooted. common ingroup identity theory, increased salience of Sectarianism is institutionalized. The 1943 National shared membership in the Lebanese national community Pact heralding Lebanon’s independence from France should increase cooperation levels among Shia and and the 1989 National Reconciliation Accord (the Taif Sunni by increasing altruistic regard and trust across the Agreement) that concluded the intersectarian civil war sectarian divide. (1975–90) organize the system of government around The literature suggests several mechanisms by which sectarian differences. a shared overarching identity might be made more salient The very question of the relative size of the three by the treatments. In the expert appeal treatment, the largest sects is extremely sensitive, and the only national four experts—two Shias and two Sunnis—agree that inpopulation census in Lebanon’s fraught history took tersectarian cooperation is necessary to overcome the ills place in 1932. There is a general consensus that Shia and plaguing Lebanese society and that all religious commuSunni Muslims are now close to parity, each at about nities must band around the common national identity. 30% of the population, and that Christians make up We expect this treatment to increase the salience of the most of the remainder (Faour 2007). The three sects have Lebanese national identity insofar as a consensus among quotas in parliament: 64 seats are assigned to Christians, experts from different sects is persuasive. That experts can 27 each to Shia and Sunni Muslims, and the remaining be persuasive in affecting political attitudes and behavior 10 seats to other Muslim minorities. Basic services is well established empirically (Henrich and Gil-White are commonly provided along sectarian lines, and the 2001; Zaller 1992). population is highly segregated regionally within the As to the effectiveness of group discussion in increascountry and by neighborhood in Beirut. Marital practices ing the salience of the overarching group identity, studies are regulated independently by separate confessional on group decision making and deliberative discourse courts, and most high school students attend schools suggest that face-to-face communication—specifically, dominated by a single confession (Baytiyeh 2016). the need to stress shared identity characteristics in order External actors exert pressure on behalf of the competing to persuade listeners—helps participants to recognize the sects: Iran supports Lebanon’s Shias, whereas Saudi existence of commonalities that bind all group members Arabia supports the Sunnis. together, thus increasing cooperation within the broader This environment of institutionalized sectarianism group (for a review, see Delli Carpini, Cook, and Jacobs gives rise to clientelism. Most commonly, clientelism takes 2004). This leads us to hypothesize that group discussions the form of vote buying on the polling day (Corstange 2012, 2016), although longer-term clientelistic arrange2 Common Arab and Muslim identities are available as other overarments that entail selective service provision are also ching identities, but neither the experts nor the participants stressed present (Cammett and Isaar 2010). About 55% of the these in discussions. BUILDING COOPERATION AMONG GROUPS IN CONFLICT 149 that bring together participants from conflicting groups appeal about the benefits of intersectarian cooperation, should result in increased mutual regard among the (2) participation in a small-group discussion about interparticipants. sectarianism after seeing the expert appeal, and (3) being In the context of a laboratory-in-the-field experioffered money to vote for a co-sectarian in the subsement, we are able to test the observable implications of quent election game. All tasks, from watching the expert the theoretical arguments laid out above. Those implicadiscussion to playing the games, are undertaken within tions are that exposure to a discussion among experts that six-person groups in a series of face-to-face interactions. stresses the importance of a common Lebanese national Each group is made up of three Shia and three Sunni paridentity and participation in a group discussion about the ticipants; membership in these six-person groups remains benefits of intersectarian cooperation should both lead to fixed over the course of the study. increased cooperation across sectarian lines. This gives rise to the following hypotheses: Study Population H1: Participants randomized into the expert apParticipants are drawn from across Beirut, with a view to peal intervention will be more likely to coopassembling a representative sample of Shia and Sunni reserate with members of the opposite sect in idents of Lebanon’s capital. Participants’ characteristics exchanges entailing both unconditional and are described in Table 1. Two-thirds are from neighborconditional cooperation. hoods dominated by a single sect, and the remainder are H2: Participants randomized into the group disfrom mixed neighborhoods. This is a fair approximation cussion intervention will be more likely to coof how Muslims are distributed across Beirut neighboroperate with members of the opposite sect in hoods (Salamey and Tabar 2008). The average age of both types of exchanges. participants is 37.3 On average, participants have 12 years The study also contains a vote-buying intervention. of schooling. They come from households with monthly Clientelism is effective because financial incentives ofincomes of around $2,100. Men and women, and Shia fered by brokers might override individuals’ social preferand Sunni Muslims, are all equally represented. In Apences. Given that the expert appeal and group discussion pendix A in the supporting information (SI), we compare interventions strengthen the overarching group identity, the demographic characteristics of our participants to we expect these interventions to also strengthen intrinsic Arab Barometer data on Muslim residents of Beirut and motivations such as positive attitudes toward a member Lebanon. The comparison suggests that participants in of the opposite religious or ethnic group. We hypothesize this study are generally representative of the reference that appeal to intrinsic motivations will be more powerful populations with regard to age, education level, and than the pull of financial incentives, as those who accepted income. The Shia and Sunni communities are similar clientelistic payments might be made to feel ashamed of in size in Lebanon; this is also the case by design in our having taken money to vote for a co-sectarian. The test of study. this hypothesis entails a comparison of those who were randomized into the vote-buying treatment and either the expert appeal or group discussion intervention (an Setting interaction) against the baseline of those who were only On arrival, individuals were randomly assigned to sixsubject to vote buying. member groups. Each group was balanced with regard to sect (three Shia and three Sunni), gender (three men and H3: Participants assigned to expert appeal and three women), and age (three participants aged 18–40 group discussion interventions will be less and three aged 41–64). None of the group participants likely to vote for a co-sectarian candidate in had met beforehand, and we asked participants not to simulated elections after being offered a finanspeak with one another until instructed to do so. Particcial incentive to do so than those only subject ipants were asked to keep their names secret throughout to the vote-buying treatment. the study because names can be a marker of sectarian affiliation. Our assistants acted as group moderators: They Research Design informed participants that there were three Shia and This laboratory-in-the-field experiment is built around 3 Age was capped at 64 to maximize the chance that participants three interventions: (1) viewing of a prerecorded expert were literate. 150 HAN IL CHANG AND LEONID PEISAKHIN TABLE 1 Description of Participants: Summary Statistics Mean Standard Deviation Min Max Shia/Sunni, % 0.50 0.50 0 1 Female/Male, % 0.50 0.50 0 1 Age, Years 37 13 18 64 Education Level+ 3.56 1.50 0 6 Monthly Household Income Level++ 1.06 1.24 0 6 Predominantly Shia Neighborhood, % 0.30 0.46 0 1 Predominantly Sunni Neighborhood, % 0.35 0.48 0 1 Level of Understanding of Tasks+++ 2.64 0.60 0 3 Note: Survey question wording is in SI Appendix B. + Education levels: 3 = secondary; 4 = technical. ++ Income: 1 = 1,000,001 to 3,000,000LBP; 2 = 3,000,001 to 5,000,000LBP. +++ Level of understanding: 2 = understood instructions quite well; 3 = understood instructions fully. three Sunni Muslims in the group without revealing The expert appeal intervention is set up to resemwhich specific participants belonged to which sect or ble a political talk show. In the video, four experts—two introducing them, read out instructions for every task in Shia and two Sunni male journalists in their 50s—discuss Arabic, and moderated the group discussion. Assistants pressing problems in contemporary Lebanese politics, the followed written scripts and were trained not to deviate promise of intersectarian cooperation, and the coming tofrom agreed-upon instructions. They were randomly gether around a common national identity in solving the assigned to groups, but in the analyses that follow, we problems. The discussion is moderated by a female Chrisinclude a control for group moderator effects. tian journalist and is unscripted. The four experts and The experiment took place in November 2016 over the moderator are all prominent in Lebanese media, conthe course of 12 separate sessions that were completed in tributing opinion pieces to major newspapers and appear8 days. Five six-person groups were recruited for every ing on national television. In their public appearances, all session. Participants were randomly assigned to sessions the experts consistently advocate for intersectarian coopsubject to the constraints of sect, gender, and age requireeration and stress the importance of a shared Lebanese ments. All those attending a specific session were subject national identity. Forty-four percent of respondents reto the same experimental manipulation. There was no membered encountering at least one of the experts’ names communication across groups. either on television or in print.4 Respondents found the experts to be persuasive: 71% of those who watched the expert appeal considered all four experts to be either very Interventions or quite persuasive (for individual expert scores disaggregated by participants’ sect, see SI Appendix C). The experimental conditions are summarized in The expert appeal is 15 minutes in duration. The Table 2. There are three of them: baseline, expert appeal, discussion opens with all the experts agreeing that secand expert appeal plus participant discussion. This structarianism is the main problem in Lebanese politics. The ture is replicated twice—once in the absence of a cliennature of the problem is then illustrated: Sectarianism telistic intervention (conditions 1–3) and once in the presis linked to corruption and nepotism, lack of a unifyence of clientelism (conditions 4–6). The total number of ing Lebanese national identity, infrequent social mixing participants in the study is 360, the total of 60 six-member across sects, and institutional sclerosis and poor quality teams. Ten groups (i.e., 60 individuals) were randomly of basic public services. The discussion closes with parassigned to each intervention. This is a relatively small ticipants appealing for intersectarian cooperation, with number of participants per intervention, which raises a the aim of creating a common Lebanese national identity concern about the study’s statistical power. In designing so that problems of corruption, inefficiency in the provithe experiment, we faced a trade-off between statistical sion of public services, and fostering of preconditions for power (increasing the number of participants per intervention) and the study’s realism (including a clientelism 4 There is no difference in the rate of expert recognition among intervention). We opted for including a clientelism interparticipants assigned to the expert appeal and group discussion vention over increasing the number of subjects. conditions. BUILDING COOPERATION AMONG GROUPS IN CONFLICT 151 TABLE 2 Experimental Conditions Baseline Expert Appeal Expert Appeal + Participant Discussion No Clientelism Experimental condition 1: Experimental condition 2: Experimental condition 3: 60 participants 60 participants 60 participants Clientelism Experimental condition 4: Experimental condition 5: Experimental condition 6: 60 participants 60 participants 60 participants violent conflict can be overcome. The transcript of the our confederate one-on-one. We used two confederates: video (in English) and the video itself (in Arabic) are one Shia and one Sunni, both male and in their midavailable in SI Appendices D and E. 20s. Participants always met with the confederate who The participant discussion intervention is designed was their co-sectarian. Our confederates verified particito follow on from experts’ appeal in order to ensure pants’ contact information and then briefly described the that participants talk specifically about intersectarian first game, a simulated election with multiple rounds in cooperation. In real-world interactions, political appeals which participants would be asked to vote for a Shia or are often followed by conversations in a group setting. Sunni candidate proposing to divide a fixed amount of Our treatment takes on the form of precisely such a money in different ways across different sects, and that group discussion in which participants have a chance the Shia or Sunni candidates were actors. Then those not to react to the message that they had just been exposed randomized into the clientelism intervention were asked to. Experiments on deliberation commonly provide a to return to the group, and those in the clientelism treatstimulus—usually a written text—prior to discussion. ment were made the clientelistic offer. The confederate In this instance, a written prompt is replaced by a explained that unbeknownst to researchers, a good friend prerecorded message from experts. and co-sectarian of the confederate (and therefore also a The group discussions lasted 28 minutes on average; co-sectarian of the subject) was the actor delivering mesthe shortest was 18 minutes and the longest 39 minutes. sages in the simulated elections. The broker then offered Discussions were relatively freewheeling; group moderathe subject $10 “in secret” to vote for the candidate of tors were asked to pose a series of questions only in order their sect across all the elections irrespective of the canto direct the conversation (for moderator instructions, see didate’s policy platform.6 The election brokers’ script is SI Appendix F). Group discussion transcripts (translated available in SI Appendix H. Subjects were free to accept to English) are available in SI Appendix G.5 or reject the money. Seventy-six percent of the particiIn designing the study, we tried to approximate pants accepted the offer; those who rejected it mostly did Lebanon’s actual political environment. One obvious so stating explicitly that they wanted to hear the candichallenge to studying the effectiveness of expert appeals dates’ proposals about how the financial rewards were to or deliberative discussions is that political behavior in be divided. Lebanon, just as in many developing democracies, is not simply a product of individual preferences but is subject to pressure from vote-buying intermediaries. The quesOutcome Measures tion then is not just whether our interventions can sway The first game, designed to measure unconditional coopmembers of different sects to cooperate across the sectareration, was a simulated election and took place over the ian divide, but, more realistically, whether these intervencourse of four rounds. In each round, two candidates— tions can be effective in the presence of clientelism. To test one Shia, one Sunni; both besuited men in their late this proposition, we assigned 180 participants in their six50s of similar physical appearance—appeared in perperson groups to a clientelism intervention within which son before the six-person group to deliver competing the three experimental conditions were replicated. policy proposals with regard to how to divide a certain The clientelistic appeal was delivered by our confedamount of money between Shia and Sunni Muslims in the erates who acted as “election brokers.” This happened group. Prior to this, all participants were endowed with in the context of checking participants’ contact details. Participants were called to a private room to meet with 6 U.S. dollars are used commonly alongside the national currency (Lebanese pounds), and all the transactions in this study took place 5 One participant objected to being recorded, and therefore tranin U.S. dollars. Ten dollars is equivalent to about 2 hours of work scripts are available for 19 of 20 group discussions. on an average wage in Beirut. 152 HAN IL CHANG AND LEONID PEISAKHIN TABLE 3 Candidates’ Messages in the Election Game Shia Candidate Sunni Candidate Election 1 Sectarian distribution: Shia only (40/0) Egalitarian distribution: (20/20) Election 2 Egalitarian distribution (20/20) Sectarian distribution: Sunni only (0/40) Election 3 Sectarian distribution: Shia and Sunni (30/10) Egalitarian distribution: (20/20) Election 4 Egalitarian distribution (20/20) Sectarian distribution: Shia and Sunni (10/30) Note: In parentheses is the number of tokens offered to Shia and Sunni participants, respectively. 40 tokens each (a single token is equivalent to $0.50), and In the other–other allocation game, participants cannot 20 tokens were collected from every participant by way of benefit personally from their allocation decisions, unlike what was described to participants as a tax. The resultant in the simulated election where in the context of iterated 120 tokens ($60) were placed at the center of the table; votes, it might be in participants’ self-interest to supthis amount was to be divided according to the proposal port a candidate from the opposite sect. In short, the made by the winning candidate. other–other allocation game provides a clean measure of Over the course of four elections, the candidates dealtruistic regard toward the outgroup and, consequently, livered competing proposals for how to divide “the tax” a clearer test of the hypothesized psychological channel between the two sects. Votes were cast in writing by sebehind increased cooperation. cret ballot. Election winners were determined at the very The third and final game was the standard public end of the study.7 Instructions read out to participants goods game, which was used to measure the strength of by group moderators for this and all other games are conditional cooperation within groups.8 In this task, paravailable in SI Appendix I. Templates of decision sheets ticipants had to decide how many of 10 tokens to keep that participants had to complete in each game are in SI for themselves and how many to surrender into the comAppendix J. mon pool. At the end of every round, all tokens in the The proposals delivered by election candidates are common pool were multiplied by two, and the resulsummarized in Table 3. The order of elections was fixed tant sum was shared equally across all group members. to facilitate the implementation of the study, and the two Higher contributions to the common pool are reflective of candidates alternated in delivering speeches. In every eleca higher willingness to cooperate with others and speciftion, one of the candidates proposed equal distribution ically across sectarian lines. Expectations about others’ of tokens—20 per person—to both Shia and Sunni parwillingness to contribute are central to the public goods ticipants. The other candidate proposed some form of game. We know from existing work that cooperation in unequal distribution favoring his own sect by suggestheterogeneous groups is more difficult to achieve than in ing that either all or most of the tokens should go to homogeneous settings (Balliet, Wu, and De Dreu 2014). his co-sectarians. The script of statements delivered from Thus, the setting of a mixed group—mimicking a mixed memory by the two candidates is in SI Appendix K. We society—is an especially hard case in which to expect codefine sectarian voters as those who voted for a candidate operation. from their sect across all four elections irrespective of the The public goods game was repeated for five rounds. candidate’s policy proposal. Before playing the game, participants had a chance to The second game is the other–other allocation game, practice three hypothetical scenarios to see how the freewhich is commonly used as a measure of unconditional riding incentive operates in practice. During the game, cooperation and ingroup favoritism in an intergroup conevery group member recorded his or her contribution in text (Chen and Li 2009; Grosskopf and Pearce 2017). In writing secretly, and the group moderator wrote out the this game, participants were given 10 tokens that they individual contributions and resultant distributions on had to spend and were asked to distribute these between a whiteboard at the end of every round without linking an anonymous co-sectarian and an anonymous person individuals to specific contributions. from the opposite sect. Participants did not know which specific individuals in the group were affected. The al8 The sequence of games and of elections in the first game was location decision had to be recorded secretly in writing. fixed for simplicity of implementation. Note that games’ results were not revealed until the very end of the study. Fixed order of games is commonly used in experimental research in which tasks 7 In the event of a tie, the election winner was determined by a coin are complex and a fixed sequence facilitates implementation (e.g., toss. Chen and Li 2009; Gilligan, Pasquale, and Samii 2014). BUILDING COOPERATION AMONG GROUPS IN CONFLICT 153 Compensation and Ethics in Table 4.9 The dependent variable takes on the value of 1 if a participant votes for a co-sectarian across all four With the three games finished, participants were asked to elections irrespective of how the candidate proposed to complete a brief survey, and then earnings were tallied. divide the funds between Shia and Sunni participants.10 One election of four and a single round of the public The models are logistic regressions; the coefficients are goods game were picked at random for the purposes of nonexponentiated. For a substantive interpretation of calculating the winnings from games 1 and 3. Allocathe coefficients, we computed marginal effects of varytions from game 2 were added to these. In addition, every ing the treatment variable from 0 to 1 while holding all participant received a show-up fee of $18. Participants other variables at their means. We expect both intervenearned $48 on average, a little more than a day’s average tions to decrease the likelihood of voting for a co-sectarian wage in Beirut. Those who had been assigned to the cliencandidate. telism treatment and accepted the $10 reward kept that The results indicate that exposure to the expert appeal too. decreases sectarian voting, thus increasing unconditional We purposefully kept remuneration high in order to cooperation. Computing the marginal effects for the exmotivate participants to think carefully through their depert appeal intervention, we find that those randomly cisions and to try to imitate the high stakes of regular assigned to the expert appeal treatment are less likely to political interactions. This study involved deception for vote along sectarian lines by about 19–23% (Models 1 those assigned to the clientelism intervention. All those and 4, respectively). These effects are statistically signifiin the clientelism intervention were debriefed on the nacant. In contrast, participant discussion does not appear ture of deception; there were no adverse reactions in the to diminish sectarian voting. If anything, the likelihood of debriefing. voting for a co-sectarian after participating in a group discussion increases by 7–9% (Models 8 and 5, respectively), although this effect is not statistically significant. Given Results that there are relatively few participants in this study, there might be some concern that the results in Table 4 Balance are subject to Type I error and are not representative of By design, there is the same number of women and men true population dynamics. To address this possibility, we (50%) and Shia and Sunni participants (50%) across the disaggregated election results by sect and treatment; the six treatment conditions. The groups are also identical results are in SI Appendix O. The disaggregated pattern with regard to age, wealth, and participants’ ability to unof voting is consistent with expectations: After exposure derstand the group moderator’s instructions. Full results to expert appeal, fewer co-sectarians vote for a candidate of a randomization check are in SI Appendix L. from their own sect who proposes to give all tokens to In the analyses that follow, we include the following members of his sect. variables as controls: gender, sect, age, education, wealth, Treatment effects in the other–other allocation game level of understanding of tasks, and type of residential are explored in Table 5. Here, the dependent varineighborhood. The control variables are summarized in able is the average number of tokens (of 10) allocated Table 1. We also cluster standard errors at the level of to an anonymous co-sectarian. The model is ordinary six-person groups to control for the possible confoundleast squares (OLS). We expect the number of tokens ing effect of group dynamics and include controls for given to co-sectarians to decrease as a result of both group moderator effects and clientelism where appropriinterventions. ate. Groups are nested within sessions. In order to capConsistent with the findings in the election game, ture this dynamic, we replicate the main analyses using we find that exposure to the expert appeal diminishes multilevel modeling. All of the results from multilevel the number of tokens allocated to co-sectarians by apmodeling analyses are reported in SI Appendix M; these proximately one token, from seven to six. This effect is results are consistent with ones reported in the body of the statistically significant and suggests that the expert appeal article. 9 The number of observations in these and subsequent models is short of 240 because some participants made ambiguous markings Unconditional Cooperation on decision sheets. There is also missingness in analyses that include controls from survey questions. The effects of expert appeal (panel A) and participant dis10 Distribution of votes by treatment group across elections is recussion (panel B) in the simulated elections are explored ported in SI Appendix N. 154 HAN IL CHANG AND LEONID PEISAKHIN TABLE 4 Effect of Expert Appeal (Panel A) and Participant Discussion (Panel B) on Likelihood of Voting for a Co-sectarian A. Expert Appeal (Logistic) B. Participant Discussion (Logistic) Model 1 2 3 4 5 6 7 8 ∗ ∗ ∗∗ ∗∗ Expert Appeal –0.82 –0.86 –0.90 –1.04 (0.36) (0.33) (0.31) (0.34) Participant 0.41 0.43 0.45 0.38 Discussion (0.39) (0.37) (0.35) (0.37) Constant –0.17 –0.62∗ –0.10 3.77∗ –0.99∗∗ –1.61∗∗ –1.37∗∗ 2.84 (0.22) (0.29) (0.47) (1.73) (0.28) (0.33) (0.52) (1.58) Clientelism N Y Y Y N Y Y Y Group Moderator N N Y Y N N Y Y Indicator Control variables N N N Y N N N Y Observations 238 238 238 212 238 238 238 205 R2 0.03 0.06 0.10 0.14 0.01 0.06 0.10 0.13 Note: Models 1–4: treatment effects in conditions 2 and 5 versus baseline of 1 and 4. Models 5–8: treatment effects in conditions 3 and 6 versus baseline of 2 and 5. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗∗ p < .01. TABLE 5 Effect of Expert Appeal (Panel A) and Participant Discussion (Panel B) on Allocation of Tokens to an Anonymous Co-sectarian A. Expert Appeal (OLS) B. Participant Discussion (OLS) Model 1 2 3 4 5 6 7 8 Expert Appeal –1.03∗ –1.03∗ –1.03∗∗ –1.11∗∗ (0.42) (0.40) (0.36) (0.39) Participant 0.41 0.41 0.40 0.49 Discussion (0.35) (0.34) (0.30) (0.30) Constant 7.01∗∗ 6.60∗∗ 6.73∗∗ 10.88∗∗ 5.98∗∗ 5.77∗∗ 5.83∗∗ 7.98∗∗∗ (0.35) (0.36) (0.63) (1.68) (0.23) (0.24) (0.38) (1.46) Clientelism N Y Y Y N Y Y Y Group Moderator N N Y Y N N Y Y Indicator Control Variables N N N Y N N N Y Observations 238 238 238 212 239 239 239 206 R2 0.04 0.07 0.13 0.19 0.01 0.02 0.07 0.14 Note: Models 1–4: treatment effects in conditions 2 and 5 versus baseline of 1 and 4. Models 5–8: treatment effects in conditions 3 and 6 versus baseline of 2 and 5. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗ p < .01. increases altruistic regard toward the outgroup, as we had whereas participation in a group discussion has no effect hypothesized. As in the previous task, participation in a on this type of cooperation. group discussion appears to increase allocation of tokens to co-sectarians by about 0.4 of a token, although, once again, this effect is not statistically significant. All in all, we Conditional Cooperation find that exposure to the expert appeal seems to increase To explore whether exposure to an expert appeal and parintersectarian cooperation in unconditional exchanges, ticipation in a group discussion might affect conditional BUILDING COOPERATION AMONG GROUPS IN CONFLICT 155 TABLE 6 Effect of Expert Appeal (Panel A) and Participant Discussion (Panel B) on Number of Tokens Contributed to the Common Pool A. Expert Appeal (OLS) B. Participant Discussion (OLS) Model 1 2 3 4 5 6 7 8 Expert Appeal 0.32 0.32 0.36 0.39 (0.45) (0.45) (0.41) (0.42) Participant 0.07 0.07 0.10 0.19 Discussion (0.33) (0.32) (0.32) (0.35) Lagged Group 0.07∗∗ 0.07∗∗ 0.05∗∗ 0.05∗ 0.09∗∗ 0.09∗∗ 0.08∗∗ 0.08∗∗ Contribution (0.02) (0.02) (0.02) (0.02) (0.01) (0.01) (0.01) (0.01) Constant 3.35∗∗ 3.51∗∗ 4.42∗∗ –0.80 2.94∗∗ 3.05∗∗ 3.50∗∗ –1.09 (0.77) (0.87) (0.88) (2.12) (0.66) (0.72) (0.65) (1.51) Clientelism N Y Y Y N Y Y Y Group Moderator N N Y Y N N Y Y Indicator Control Variables N N N Y N N N Y Observations 960 960 960 856 960 960 960 828 R2 0.30 0.30 0.26 0.28 0.29 0.29 0.27 0.31 Note: Models 1–4: treatment effects in conditions 2 and 5 versus baseline of 1 and 4. Models 5–8: treatment effects in conditions 3 and 6 versus baseline of 2 and 5. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗ p < .01. cooperation in a setting involving calculations about the the financial incentive declined to accept it. Therefore, in behavior of the opposite group, we had the participants testing the effectiveness of vote buying, we use two difplay a public goods game within their mixed-sect groups. ferent measures of the treatment: whether the participant Average contributions to the common pool remained was randomly assigned to the treatment (Clientelism) and stable across rounds at about six tokens (of 10); the whether he or she accepted the financial incentive (Acpattern of average contributions is described in SI cepted Clientelistic Offer). It might be useful to think of Appendix P. We hypothesized that contributions to the this difference as that between the average treatment effect common pool and therefore the level of conditional and the complier average causal effect. cooperation will increase in both treatment conditions. The effectiveness of the clientelism manipulation in The impact of the two interventions on conditional encouraging sectarian voting and the impact of expert cooperation is explored in Table 6. The models are OLS; appeal and group discussion on clientelism are explored Lagged Group Contribution is a control for average group in Table 7. The dependent variable is a dummy indicating contribution in the preceding round of the game. Both sectarian voting. The models are logistic, and the coeffiinterventions increase the amount that participants concients are nonexponentiated. tribute, but these effects are not statistically significant Consistent with expectations, we find that clientelism across any specification. Contrary to expectations, our hyis highly effective at getting participants to vote along pothesis about the treatments strengthening conditional sectarian lines. Those subject to the clientelistic appeal cooperation is not supported by the data. are 21% more likely to vote for a co-sectarian (marginal effects from Models 1 and 5). Among those who accept the clientelistic offer, the predicted probability of Clientelism voting for a co-sectarian increases by 34% (Models 3 We hypothesized that the interventions would diminish and 7). the effectiveness of clientelism. We now test this hypothWhether the interventions mitigate the positive efesis in the context of the simulated elections. The clienfect of clientelism on sectarian voting is tested using intelism intervention incentivizes participants to vote for teraction terms between Clientelism/Accepted Clientelistic co-sectarians irrespective of the content of the candiOffer and each of the treatments in Models 2 and 4 for dates’ policy proposals. However, 24% of those offered the expert appeal and Models 6 and 8 for participant 156 HAN IL CHANG AND LEONID PEISAKHIN TABLE 7 Effects of Clientelism on Likelihood of Sectarian Voting A. Expert Appeal (Logistic) B. Participant Discussion (Logistic) Model 1 2 3 4 5 6 7 8 ∗ ∗∗ ∗ Clientelism 0.94 0.90 1.07 1.06 (0.38) (0.53) (0.36) (0.52) Accepted Clientelistic Offer 1.53∗∗ 1.77∗∗ 1.80∗∗ 1.27∗ (0.42) (0.62) (0.40) (0.56) Expert Appeal –1.04∗∗ –1.10∗ –1.15∗∗ –0.92 (0.34) (0.53) (0.33) (0.50) Clientelism × Expert Appeal 0.11 (0.74) Accepted Clientelistic Offer × –0.51 Expert Appeal (0.87) Participant Discussion 0.38 0.38 0.58 0.09 (0.37) (0.56) (0.36) (0.52) Clientelism × Participant 0.00 Discussion (0.74) Accepted Clientelistic Offer × 1.11 Participant Discussion (0.86) Constant 3.77∗ 3.81∗ 2.96 2.69 2.84 2.85 2.85 3.31 (1.73) (1.79) (1.88) (1.96) (1.58) (1.60) (1.74) (1.73) Group Moderator Indicator Y Y Y Y Y Y Y Y Control Variables Y Y Y Y Y Y Y Y Observations 212 212 212 212 205 205 205 205 R2 0.14 0.14 0.19 0.19 0.13 0.13 0.19 0.20 Note: Accepted Clientelistic Offer is coded as 0 for those not subject to clientelism and those who received the treatment but did not accept the offer. Models 1–4: data from conditions 1–2 and 4–5. Models 5–8: data from conditions 2–3 and 5–6. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗ p < .01. discussion. Neither of the interventions appears to diWhy No Effect of Interventions on minish the effectiveness of clientelism. The interaction Conditional Cooperation?11 coefficients are consistently not statistically significant. Contrary to what we hypothesized, the results suggest Conditional cooperation entails calculations about the that the treatments are ineffective in negating the effects of likelihood of a member of the opposite group reciprocatclientelism. ing in response to a benevolent action. We hypothesize that intergroup trust might be a prerequisite for a positive expectation that a member of the competing group will reciprocate. We have measures of cross-sectarian trust Discussion from the survey that was administered after the games had been completed. Using survey data, we constructed The results give rise to two important questions. First, measures of absolute and relative trust in the opposite why does the expert appeal intervention fail to increase sect. The absolute cross-sectarian trust measure reflects cooperation in exchanges that involve calculations about a participant’s trust in the opposite sect on a 4-point the likelihood of reciprocal response by the opposite scale, where 1 is no trust at all and 4 is complete trust group? Second, given the extensive literature on the im(for question wording, see SI Appendix B). The relative portance of discussion for consensus building, why did participant discussion fail to increase cooperation? In ad11 dressing these questions, we use observational data from We reran the analyses in Table 6 using the Tobit model to test for the presence of ceiling effects that might be responsible for absence a postexperimental survey and participant discussion of treatment effects. The results are reported in SI Appendix Q and transcripts. rule out ceiling effects. BUILDING COOPERATION AMONG GROUPS IN CONFLICT 157 TABLE 8 Effect of Expert Appeal and Participant Discussion on Cross-Sectarian Trust (Models 1–4), and Effect of Trust on Contributions in a Public Goods Game (Models 5–6) A. Absolute Cross-Sectarian B. Relative Cross-Sectarian C. Contribution to Public Trust (Ordinal Logistic) Trust (Ordinal Logistic) Goods (OLS) 1 2 3 4 5 6 Expert Appeal 0.19 0.45 (0.31) (0.28) Participant Discussion 0.42 0.15 (0.24) (0.24) Absolute Cross-Sectarian Trust 0.69∗∗ (0.23) Relative Cross-Sectarian Trust 0.46∗∗ (0.18) Lagged Group Contribution 0.06∗∗ 0.06∗∗ (0.02) (0.02) Clientelism Y Y Y Y Y Y Group Moderator Indicator Y Y Y Y Y Y Control Variables Y Y Y Y Y Y Constant –2.52 –1.12 (1.48) (1.58) /cut1 –1.23 0.24 –1.50 –2.08 (1.60) (1.36) (1.50) (1.45) /cut2 1.18 2.82 –0.18 –0.65 (1.57) (1.31) (1.50) (1.41) /cut3 2.55 4.33 0.92 0.44 (1.58) (1.31) (1.52) (1.40) /cut4 4.12 4.04 (1.58) (1.48) /cut5 5.26 5.05 (1.76) (1.63) /cut6 5.67 5.46 (1.83) (1.69) Observations 176 179 173 178 1060 1044 R2 0.05 0.04 0.03 0.02 0.32 0.31 Note: Models 1 and 3: data from conditions 1–2 and 4–5. Models 2 and 4: data from conditions 2–3 and 5–6. Models 5 and 6: data from all experimental conditions. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗ p < .01. cross-sectarian trust measure captures how much a parcoefficients are nonexponentiated. Exposure to the exticipant trusts the other group relative to her group. The pert appeal and participation in a group discussion both relative trust measure is on a 7-point scale and is calcuappear to somewhat increase cross-sectarian trust. Howlated by subtracting a participant’s trust in her own sect ever, this effect is not statistically significant. These results from her trust in the opposite sect. suggest that the interventions failed to increase crossFirst, in panels A and B of Table 8, we establish sectarian trust. whether the two interventions increased cross-sectarian We test for the existence of a relationship between trust. We analyze the effect of treatments on trust using cross-sectarian trust and conditional cooperation in panel the ordinal logistic regression. For ease of comparison, C of Table 8, where we regress cross-sectarian trust (abthe regression in panel B is also ordinal logistic.12 The solute and relative in Models 5 and 6, respectively) on the number of tokens that participants contributed to the 12 Results from an OLS regression are reported in SI Appendix R. common pool in the public goods game. The results are 158 HAN IL CHANG AND LEONID PEISAKHIN from an OLS regression. Consistent with expectations, discussion, as existing studies suggest that women and we find that a one-unit increase in absolute trust (on a 4young people might participate at lower rates (Karpowitz point scale) is associated with an additional contribution and Mendelberg 2014).13 of 0.7 of a token (of 10). The magnitude of the correlaIn Models 1–2, 4–5, and 7–8, we examine the tion for relative cross-sectarian trust is similar. Both estieffect of the overall depth of group discussion in each mates are statistically significant. In short, higher levels of of the tasks. Higher quality of discussion appears to cross-sectarian trust do indeed tend to correlate with inbe strongly associated with more cooperation across creased conditional cooperation across sectarian lines. sectarian lines. A move from the shallowest to the deepest Because the treatments seem to fail to increase crossdiscussion is associated with a decrease in the probability sectarian trust, they also fail to increase conditional of sectarian voting of 76% in the election game (Model cooperation. 2), allocation of four fewer tokens to a co-sectarian in the other–other allocation game (Model 5), and an increase in the contribution in the public goods game of Heterogeneous Effects of Discussion Quality 2.9 tokens (Model 8). This result suggests that the overall ineffectiveness of the group discussion intervention Results reported thus far suggest that participation in a masks a great deal of heterogeneity. group discussion has no effect on unconditional or condiOne possible concern about these associations is tional cooperation. This is contrary to initial expectations. that our score of discussion depth stands in for someOne possible reason why might be that group discussions thing else, for instance, discussion tone or duration, or varied substantially as to their quality. length of interventions by group moderators. To rule out To test for possible heterogeneity in group discussion these concerns, we replicated the analyses in Table 9 and effects, we constructed an index that measures their subincluded measures of discussion tone14 (SI Appendix T), stantive depth. Substantive depth might matter insofar as well as discussion duration and the length of moderas discussions that lack it might fail to alter participants’ ator interventions (SI Appendix U). Inclusion of these preferences. A discussion depth index for the group is additional variables does not change the magnitude or an aggregate of six individual discussion depth scores. significance of the reported coefficients. A participant’s position on each of 25 issues commonly Another potential concern is that, given that disbrought up over the course of discussions forms the bacussion depth was not assigned randomly, both active sis of her score. The 25 common issues fall into three contribution to discussion and higher propensity to cobroad categories: negative effects of sectarianism, causes operate across sectarian lines might be a product of some of sectarianism, and ways to encourage cooperation. A unmeasured set of factors. For instance, in SI Appendix V, participant receives 1 point if she said that the issue is we demonstrate that older participants had higher indirelevant to cooperation, –1 if she stated that it was irrelevidual discussion scores. The data allow us to disaggrevant, or 0 if she did not express an opinion on the issue. gate the effect of the depth of a participant’s own conThere is no double-counting of repeated opinions on the tribution to group discussion from the effect of being same issue. Coding rules for discussion depth scores are in an especially engaging discussion. We have discusin SI Appendix S. Individual scores range from the low of sion scores for individuals and also aggregate discussion 0 to the high of 11. Group scores are compiled by adding scores for the five remaining group members. If the qualthe individual scores of the six participants. The discusity of one’s own contributions to group discussion is the sion depth scores for groups range from the lowest of 15 only thing that matters in the association between disto the highest of 41. cussion depth and intersectarian cooperation, then the The results of the analyses examining the effect of score for others’ contributions to discussion should not be discussion depth on the outcomes of the three games significant. are presented in Table 9. The outcomes are a dummy This is what we test for in Models 3, 6, and 9 of Table for voting for a co-sectarian across four elections (panel 9. The results suggest that it is the quality of others’ and A), the number of tokens allocated to a co-sectarian not one’s own comments that shape one’s willingness to in the other–other allocation game (panel B), and the 13 size of the contribution in the public goods game (panel These measures are constructed based on moderators’ evaluations of the dynamics of each group discussion. C). Panel A reports nonexponentiated coefficients from 14 a logistic regression; OLS is used in panels B and C. Discussion tone is calculated by using automated sentiment analysis; negative sentiment is subtracted from positive, and the reWe add two additional control variables that measure sultant coefficient divided by highest discussion score among all how active women and those under 40 were in the groups to standardize scores. BUILDING COOPERATION AMONG GROUPS IN CONFLICT 159 TABLE 9 Effect of Discussion Depth on Sectarian Voting (Task 1), Allocation in the Other–Other Allocation Game (Task 2), and Contribution in the Public Goods Game (Task 3) B. Allocation to an Anonymous Co-sectarian in A. Sectarian Voting the Other–Other Allocation C. Contribution in the (Logistic) Game (OLS) Public Goods Game (OLS) Model 1 2 3 4 5 6 7 8 9 Group Discussion –0.11∗∗ –0.17∗∗ –0.14∗∗ –0.16∗∗ 0.11∗∗ 0.11∗ Score (15–41) (0.04) (0.05) (0.03) (0.03) (0.03) (0.05) Individual Discussion –0.32 –0.25 0.07 Score (0–11) (0.19) (0.13) (0.14) Others’ Aggregate –0.14∗∗ –0.13∗∗ 0.12∗ Discussion Score (0.05) (0.05) (0.06) (11–36) Female Participation –0.36 –0.00 0.06 –0.34 –0.47 –0.52 0.80 1.24∗ 1.22∗ (0.84) (0.86) (0.90) (0.22) (0.31) (0.33) (0.43) (0.60) (0.61) Youth Participation 0.04 0.13 0.17 –0.49∗∗ –0.70∗∗ –0.70∗∗ 1.08∗∗ 0.88∗ 0.89∗ (0.35) (0.39) (0.38) (0.15) (0.21) (0.22) (0.24) (0.38) (0.39) Lagged 0.07∗∗ 0.07∗∗ 0.07∗∗ Group Contribution (0.01) (0.02) (0.02) Constant 1.56 7.99∗ 7.19 10.85∗∗ 11.55∗∗ 11.07∗∗ –0.33 –4.80 –5.03 (1.52) (3.57) (3.68) (0.87) (2.08) (2.27) (0.99) (2.76) (3.11) Clientelism Y Y Y Y Y Y Y Y Y Group Moderator Y Y Y Y Y Y Y Y Y Indicator Control Variables N Y Y N Y Y N Y Y Observations 114 95 95 114 95 95 456 380 380 R2 0.20 0.33 0.34 0.16 0.24 0.25 0.28 0.32 0.32 Note: All models use data from conditions 3 and 6. Standard errors, clustered at group level, are in parentheses. ∗ p < .05, ∗∗ p < .01. cooperate across sectarian lines. More substantive comwould be needed to confirm the validity of these suggesments by other group members seem to be associated tive associations. with more cooperative behavior by individual members of the group, and this association is statistically significant. The coefficient for an individual’s discussion score is consistently in the right direction, but it never reaches Conclusion statistical significance.15 We have to be cautious not to overinterpret these asIn a laboratory-in-the-field experiment among the resisociations. Discussion depth was not varied randomly in dents of Beirut, Lebanon, we set out to examine how to this study, and our results are based on the analyses of increase cooperation between Shia and Sunni Muslims. only 19 group discussions. Therefore, much more work Our results indicate that exposure to a cross-sectarian consensual message calling for greater cooperation and delivered by experts can effectively increase uncondi15 Another way to tackle this concern is to compare the level of intional intersectarian cooperation (i.e., exchanges in setdividual contributions when a participant finds herself in a group tings where no calculations about reciprocal action are where discussion is flagging to one where there is engaging discusnecessary). Pro-cooperation appeal by experts has no efsion. In SI Appendix W, we report the findings from these pairwise fect on conditional cooperation where expectation of recidifferences of means comparisons. We find that inactive participants consistently cooperate more across sectarian lines when they procity is a precondition for cooperative behavior because find themselves in a group where others participate actively. the appeal seems to fail to increase cross-sectarian trust. 160 HAN IL CHANG AND LEONID PEISAKHIN Participation in a group discussion with members of the by including clientelistic appeals instead of having fewer opposite sect does not appear to increase either type of interventions with more subjects per treatment. cooperation in an experimental setting. We did find sugOn a different note, there are important issues of exgestive evidence that the effect of discussion might be ternal validity to consider. In Lebanon, Shia and Sunni highly heterogeneous. A substantive discussion touchMuslims are close to being numerically balanced. In coning on many aspects of intersectarian relations appears trast, in many Middle Eastern countries, one group domto be associated with higher levels of unconditional and inates the other numerically (e.g., Sunnis in Pakistan or conditional cooperation. Finally, we find that neither inAfghanistan) or politically (e.g., Sunnis in Bahrain). In futervention succeeds in diminishing the effectiveness of ture research, it would be important to explore whether clientelism. the treatments have the same effect in societies in which These findings give rise to several important questhe two groups are highly unequal either in terms of reltions for follow-up research. First, the tentative finding ative numbers or political influence. on the heterogeneous effect of group discussion suggests The primary policy implication of this project is that that substantive discussions might be important, and addissemination of a cross-sectarian appeal by experts to ditional research is needed to establish how discussion cooperate across group boundaries may increase intersecpartners might be incentivized to engage in a more subtarian cooperation in exchanges that do not entail calcustantive exchange (e.g., Humphreys, Masters, and Sambu lations about the likelihood of reciprocal response. These 2006). Second, in the aggregate, only the expert appeal has results from the laboratory suggest the need for a field any substantive effect on increasing cooperation (only of intervention to test this relationship in a less controlled the unconditional variety) among the Shia and Sunni and more realistic setting. Our findings also suggest that Muslims in this study. This hints at a possibility that topa reduction in the incidence of clientelism is likely to down attempts to engender intergroup cooperation— result in greater intersectarian cooperation. Given what ones that are propagated by authoritative figures perwe know about the pervasiveness of clientelism in many haps in televised debates, on the radio/podcasts, or in the young democracies (Hicken 2011), the practice of vote press—might be more effective than bottom-up efforts buying is likely to continue to pose a major impediment to achieve the same goal (e.g., like group discussions) to voting in support of candidates from ethnic or religious given that several conditions might have to be met for groups other than one’s own. the bottom-up efforts to be effective.16 Finally, with regard to conditional cooperation, our findings indicate that cross-sectarian trust might be an important prerequisite for conditional cooperation across sectarian lines. References More research is needed to establish how to augment Akerlof, George, and Rachel Kranton. 2000. “Economics and cross-sectarian trust. Identity.” Quarterly Journal of Economics 115(3): 715–53. As with any experiment, a few words of caution are Alesina, Alberto, Reza Baqir, and William Easterly. 1999. “Public in order. First, with only 60 participants per intervention, Goods and Ethnic Divisions.” Quarterly Journal of Economics this study has relatively low statistical power, and some of 114(November): 1243–84. our conclusions must therefore be tentative. Given that Allport, Gordon. 1954. The Nature of Prejudice. Reading, MA: for most of the nonstatistically significant treatment coefAddisonWesley. ficients, associated standard errors are large, we acknowlBalliet, Daniel. 2010. “Communication and Cooperation in Soedge the problem of low power but are not extremely concial Dilemmas: A Meta-Analytic Review.” Journal of Conflict Resolution 54(1): 39–57. cerned about it because the point estimates are close to Balliet, Daniel, Junhui Wu, and Carsten K. W. De Dreu. 2014. zero. Furthermore, in subsidiary analyses, we also demon“Ingroup Favoritism in Cooperation: A Meta-Analysis.” Psystrated that effects reported in regressions are also present chological Bulletin 140(6): 1556–81. when the data are disaggregated by sect and treatment. Baytiyeh, Hoda. 2016. “Has the Educational System in Lebanon In settling on a study with relatively few participants per Contributed to the Growing Sectarian Divisions?” Education intervention, we prioritized realism over statistical power and Urban Society April: 1–14. Cammett, Melanie, and Sukriti Isaar. 2010. “Bricks and Mortar Clientelism: Sectarianism and the Logics of Welfare Alloca16 tion in Lebanon.” World Politics 62(3): 381–421. In this study, the group discussion intervention combines a topdown appeal with bottom-up engagement, as group discussion Cederman, Lars-Erik, Nils Weidmann, and Kristian Gleditsch. follows on from watching the expert appeal. Therefore, we are not 2011. “Horizontal Inequalities and Ethnonationalist Civil able to cleanly disentangle the effect of top-down and bottom-up War: A Global Comparison.” American Political Science Reinterventions. view 105(3): 478–95. BUILDING COOPERATION AMONG GROUPS IN CONFLICT 161 Chen, Yan, and Sherry Xin Li. 2009. “Group Identity and Social Lust, Ellen. 2009. “Democratization by Elections? Competitive Preferences.” American Economic Review 99(1): 431–57. Clientelism in the Middle East.” Journal of Democracy 20(3): Corstange, Daniel. 2012. “Vote Trafficking in Lebanon.” Inter122–35. national Journal of Middle East Studies 44(3): 483–505. Miguel, Edward. 2004. “Tribe or Nation? Nation Building and Corstange, Daniel. 2016. The Price of a Vote in the Middle East: Public Goods in Kenya versus Tanzania.” World Politics Clientelism and Communal Politics in Lebanon and Yemen. 56(April): 327–62. New York: Cambridge University Press. Salamey, Imad, and Paul Tabar. 2008. “Consociational DemocDelli Carpini, Michael, Fay Lomax Cook, and Lawrence racy and Urban Sustainability: Transforming the ConfesJacobs. 2004. “Public Deliberation, Discursive Participation, sional Divides in Beirut.” Ethnopolitics 7(2–3): 239–63. and Citizen Engagement: A Review of Empirical Literature.” Tajfel, Henri, M. G. Billig, R. P. Bundy, and Claude Flament. Annual Review of Political Science 7: 315–44. 1971. “Social Categorization and Intergroup Behaviour.” Faour, Muhammad. 2007. “Religion, Demography, and Politics European Journal of Social Psychology 1(2): 149–78. in Lebanon.” Middle Eastern Studies 43(6): 909–21. Zaller, John. 1992. The Nature and Origins of Mass Opinion. Fearon, James, and David Laitin. 1996. “Explaining InterethNew York: Cambridge University Press. nic Cooperation.” American Political Science Review 90(4): 715–35. Fishkin, James. 1997. The Voice of the People: Public Opinion and Democracy. New Haven, CT: Yale University Press. Supporting Information Gaertner, Samuel, and John Dovidio. 2000. Reducing Intergroup Bias: The Common Intergroup Identity Model. New York: Taylor and Francis. Additional supporting information may be found online Gaertner, Samuel, John Dovidio, Phyllis Anastasio, Betty in the Supporting Information section at the end of the Bachman, and Mary Rust. 1993. “The Common Group article. Identity Model: Recategorization and the Reduction of Intergroup Bias.” European Review of Social Psychology 4(1): Appendix A: Comparison of Demographic Character1–26. istics for Study Participants and Reference Populations Gilligan, Michael, Benjamin Pasquale, and Cyrus Samii. 2014. (Muslim Residents of Beirut and Muslim Citizens of “Civil War and Social Cohesion: Lab-in-the-Field Evidence Lebanon) from Nepal.” American Journal of Political Science 58(3): Appendix B: Survey Questions 604–19. Appendix C: Experts’ Persuasiveness Scores DisaggreGrosskopf, Brit, and Graeme Pearce. 2017. “Discrimination in a Deprived Neighborhood: An Artefactual Field Expergated by Individual Expert and Participants’ Sect iment.” Journal of Economic Behavior and Organization 141: Appendix D: Expert Video Transcript (English) 29–42. Appendix E: Expert Video File (Arabic) Habyarimana, James, Macartan Humphreys, Daniel Posner, and Appendix F: Guidance Questions for Moderators for Jeremy Weinstein. 2007. “Why Does Ethnic Diversity UnderGroup Discussions (English) mine Public Goods Provision?” American Political Science Appendix G: Transcripts of Group Discussions (English) Review 101(4): 709–25. Appendix H: Election Brokers Scripts (English) Henrich, Joseph, and Francisco Gil-White. 2001. “The Evolution of Prestige: Freely Conferred Deference as a Mechanism Appendix I: Instructions for All the Games (English) for Enhancing the Benefits of Cultural Transmission.” EvoAppendix J: Decision Sheets (English) lution and Human Behavior 22(3): 165–96. Appendix K: Script Delivered by Election Candidates Hicken, Allen. 2011. “Clientelism.” Annual Review of Political (English) Science 14: 289–310. Appendix L: Randomization Check Horowitz, Donald. 1985. Ethnic Groups in Conflict. Berkeley, Appendix M: Results from Multivel Modeling CA: University of California Press. Appendix N: Vote Choice Disaggregated by Election and Humphreys, Macartan, William Masters, and Martin Sandbu. Sect 2006. “The Role of Leaders in Democratic Deliberations: Results from a Field Experiment in São Tomé and Prı́ncipe.” Appendix O: Vote Choice Disaggregated by Sect, Election, World Politics 58(4): 583–622. and Treatment Karpowitz, Christopher, and Tali Mendelberg. 2014. The Silent Appendix P: Average Contributions in the Public Goods Sex: Deliberation and Institutions. Princeton, NJ: Princeton Game University Press. Appendix Q: Testing for Presence of Ceiling Effects in Kerr, Norbert, and Cynthia Kaufman-Gilliland. 1994. “ComContributions to the Common Pool in the Public Goods munication, Commitment, and Cooperation in Social Dilemmas.” Journal of Personality and Social Psychology Game Using the Random-Effects Tobit Model 66(3): 513–29. Appendix R: Relative Cross-Sectarian Trust Regression Larson, Jennifer. 2017. “Networks and Interethnic Cooperafrom Table 8 (OLS) tion.” Journal of Politics 79(2): 546–59. Appendix S: Discussion Coding Rules 162 HAN IL CHANG AND LEONID PEISAKHIN Appendix T: Table 9 Replicated with Addition of AuAppendix V: Determinants of the Extent of Individual tomated Discussion Tone Score Alongside Discussion Participation in the Discussion Depth Score Appendix W: Distribution of Discussion Scores and Appendix U: Effect of Discussion Duration and Length Pairwise Comparisons of Inactive and Active Parof Moderator Intervention on Cooperation Levels across ticipants’ Discussion Scores as a Product of Group the Three Tasks Characteristics[NOTE: I AM GOING TO TAKE THESE OUT AS ACTUAL CODE CHUNKS SO I CAN GET THIS TO KNIT FOR THE PURPOSES OF POSTING]
Now we’re going to annotate:
{r, warning=FALSE} cnlp_init_udpipe() annotated <- cnlp_annotate(myData)
[Wow that took a long time!]
Let's look at the tokens!
{r}
head(annotated$token)
head(annotated$document)Join the token and document objects.
{r} annoData <- left_join(annotated\(document, annotated\)token, by = “doc_id”) head(annoData)
now I'm going to try and filter down just to the lemma ethic:
{r}
ethics_docs <- annoData %>%
filter(lemma == "ethic") %>%
group_by(Journal)This gives us nowhere near enough results, and a visual inspection of the data shows us the the lemmatization isn’t great - we have six different words:
ethic ethica ethical ethicality ethically ethics
Let’s see if we can get all of those into one search?
{r} ethics_docs <- annoData %>% filter(lemma == “ethic” | lemma ==“ethica” | lemma == “ethical” | lemma == “ethicality” | lemma == “ethically” | lemma == “ethics”) %>% group_by(Journal)
This looks like we now have all of our documents that use one of our search terms in one dataframe
{r}
yes_ethics <- ethics_docs %>% group_by(doc_id) %>% summarize(count=n())
yes_ethics
So, 51 of our 121 documents have one of our ethics terms in it.
Now let’s make a dataframe of our no ethics as well.
{r} #no_ethics_docs <- annoData %>% # filter(lemma != “ethic” | lemma !=“ethica” | lemma != “ethical” | lemma != “ethicality” | # lemma != “ethically” | lemma != “ethics”) %>% # group_by(Journal)
#no_ethics <- no_ethics_docs %>% group_by(doc_id) %>% summarize(count=n()) #no_ethics
#annoDataAll <- annoData %>% group_by(doc_id) %>% summarize(count = n()) #no_ethics <- annoDataAll %>% filter(doc_id == yes_ethics$doc_id)
This really isn't working and I'm not sure how to do this!
Going back to my tokens:
{r}
docvars(combinedTokens)
I think at this point, I’m going to submit this as a blog post.
{r} annoDocs <- annotated$document
{r}
setwd("~/DACSS/697D Text as Data/Final Project Materials")
write_as_csv(annoDocs, "annoDocs.csv")
[Note: I ended up writing this to csv and bringing it into excel to do some quick and dirty bar charts.]