Autonomous Weapons Survey Results

library(readxl)
library(dplyr)
library(ggplot2)
library(sandwich)
library(lmtest)
library(knitr)

Autonomous Weapons Survey

Background

This report summarizes a survey experiment on public perceptions of autonomous weapons use in a U.S.-China military crisis scenario. The study asks whether respondents see missile strikes as more or less intentional depending on the country scenario and whether the weapons system was human-operated or fully autonomous AI-enabled.

The primary outcome is perceived intentionality. Secondary outcomes examine escalation expectations, the perceived justifiability of escalation, and responsibility attributions to political leaders, military officials, human operators, and AI systems.

The preregistered analyses focus on the 2 x 2 factorial design: country scenario and weapons-system autonomy. Exploratory analyses examine manipulation-check passers, secondary outcomes, moderators, and downstream associations.

Method

Data

The Qualtrics export and demographic append file are in this folder.

qualtrics_file <- "Technology and National Security_August 5, 2026_09.34.xlsx"
demographics_file <- "6a68eb13df3330a45698e7cc_appends.rds"

Read the two files and merge respondent demographics onto the Qualtrics data by vsid.

qualtrics <- read_excel(qualtrics_file)
demographics <- readRDS(demographics_file)

survey <- qualtrics %>%
  filter(vsid != "vsid") %>%
  mutate(vsid = as.character(vsid)) %>%
  left_join(
    demographics %>%
      mutate(vsid = as.character(vsid)),
    by = "vsid"
  )

names(survey) <- gsub("\u00A0", " ", names(survey))

survey <- survey %>%
  rename(
    comfort_doctor = comfort_1,
    comfort_news = comfort_2,
    comfort_banks = comfort_3,
    ai_us = `ai_effectiveness _1`,
    ai_china = `ai_effectiveness _7`,
    ai_sk = `ai_effectiveness _8`,
    intentionality = intentionality_1,
    escalation = escalation_1,
    justified = justified_1,
    resp_leader = `responsibility _1`,
    resp_military = `responsibility _2`,
    resp_pilot = `responsibility _4`,
    resp_ai = `responsibility _5`
  )

Quick merge check.

nrow(qualtrics)

## [1] 3081

nrow(demographics)

## [1] 2915

nrow(survey)

## [1] 3080

sum(!survey$vsid %in% demographics$vsid)

## [1] 122

Recode Party ID

survey <- survey %>%
  mutate(
    pid = case_when(
      pid_base == "Democrat" ~ "Democrat",
      pid_base == "Republican" ~ "Republican",
      pid_base %in% c("Independent", "Other or none") &
        pid_lean == "The Democratic Party" ~ "Democrat",
      pid_base %in% c("Independent", "Other or none") &
        pid_lean == "The Republican Party" ~ "Republican",
      pid_base %in% c("Independent", "Other or none") &
        (pid_lean == "Neither" | is.na(pid_lean)) ~ "Independent",
      TRUE ~ NA_character_
    )
  )

Data Dictionary

The table below describes the variables used in the analyses. Variables listed as constructed are created later in this file from the raw Qualtrics and demographic variables.

Variable	Source	Description
vsid	Raw / merge key	Respondent identifier used to merge the Qualtrics data with the demographic append file.
surveyid	Raw	Survey identifier from Qualtrics or the survey panel.
consent	Raw	Consent response. Analyses keep respondents who indicated that they were 18 or older and agreed to participate.
condition_agressor	Raw treatment assignment	Country whose weapons system launched the missiles. Values are China or US. The variable name is misspelled in the export and is left unchanged to match the raw data.
condition_autonomy	Raw treatment assignment	Weapons-system condition. Values are human or autonomous.
country_scenario	Constructed treatment indicator	Equals 1 for Chinese missiles striking a U.S. ship and 0 for U.S. missiles striking a Chinese ship.
autonomous_ai	Constructed treatment indicator	Equals 1 for the autonomous-AI condition and 0 for the human-operated condition.
country_condition	Constructed label	Readable label for the country scenario used in tables and plots.
weapons_condition	Constructed label	Readable label for the weapons-system condition used in tables and plots.
weight	Raw demographic / panel variable	Respondent weight appended to the survey data.
survey_weight	Constructed robustness variable	Numeric copy of weight used in the weighted primary-model robustness check.
comfort_doctor	Raw pre-treatment item	Comfort with AI helping doctors identify possible medical conditions.
comfort_news	Raw pre-treatment item	Comfort with AI helping news organizations check factual claims before publication.
comfort_banks	Raw pre-treatment item	Comfort with AI helping banks detect unusual account activity.
ai_comfort	Constructed covariate	Average of the three comfort items, recoded from -2 to 2, where higher values indicate more comfort with AI.
ai_comfort_c	Constructed covariate	Mean-centered version of ai_comfort.
ai_us	Raw pre-treatment item	Respondent rating of U.S. AI model effectiveness on a 0 to 100 scale.
ai_china	Raw pre-treatment item	Respondent rating of Chinese AI model effectiveness on a 0 to 100 scale.
ai_sk	Raw pre-treatment item	Respondent rating of South Korean AI model effectiveness on a 0 to 100 scale.
relative_ai	Constructed exploratory covariate	U.S. AI effectiveness minus Chinese AI effectiveness, used in earlier exploratory plots and models.
relative_china_us_ai	Constructed preregistered covariate	Chinese AI effectiveness minus U.S. AI effectiveness, used in preregistered models.
relative_china_us_ai_c	Constructed covariate	Mean-centered version of relative_china_us_ai.
us_ai_effectiveness_c	Constructed moderator	Mean-centered U.S. AI effectiveness rating.
china_ai_effectiveness_c	Constructed moderator	Mean-centered Chinese AI effectiveness rating.
southkorea_ai_effectiveness_c	Constructed moderator	Mean-centered South Korean AI effectiveness rating.
intentionality	Raw outcome	Respondent rating of how intentional the missile strike was, measured on a 0 to 100 scale.
perceived_intentionality	Constructed outcome	Numeric copy of intentionality used in regression models.
escalation	Raw outcome	Respondent rating of the likelihood of escalation, measured on a 0 to 100 scale.
escalation_likelihood	Constructed outcome	Numeric copy of escalation used in preregistered secondary models.
justified	Raw outcome	Respondent rating of the justifiability of escalation, measured on a 0 to 100 scale.
escalation_justifiability	Constructed outcome	Numeric copy of justified used in preregistered secondary models.
resp_leader	Raw outcome	Responsibility attributed to senior political leaders and their top advisers.
resp_military	Raw outcome	Responsibility attributed to military officials overseeing the operation.
resp_pilot	Raw outcome	Responsibility attributed to the human operator of the weapons system. Asked only in human-operated conditions.
resp_ai	Raw outcome	Responsibility attributed to the AI weapons system itself. Asked only in autonomous-AI conditions.
pid_base	Raw demographic variable	Initial party identification response from the demographic append file.
pid_lean	Raw demographic variable	Party-leaning follow-up for respondents who identify as independent or other/none.
pid	Constructed demographic covariate	Party ID recode. Leaning independents are assigned to Democrat or Republican; non-leaning independents and missing leaners are kept as Independent.
pid_republican	Constructed demographic covariate	Binary party indicator used in regression models. 1 = Republican; 0 = Democrat or Independent.
age_group4	Raw demographic variable	Four-category age group: 18-29, 30-49, 50-64, or 65+.
age_ordinal	Constructed demographic covariate	Ordinal age variable coded 1 = 18-29, 2 = 30-49, 3 = 50-64, 4 = 65+.
gender	Raw demographic covariate	Gender from the demographic append file.
gender_female	Constructed demographic covariate	Binary gender indicator used in regression models. 1 = Female; 0 = Male or Other.
education	Raw demographic variable	Education category from the demographic append file.
education_ordinal	Constructed demographic covariate	Ordinal education variable coded from lowest to highest educational attainment.
income	Raw demographic variable	Household income category from the demographic append file.
income_ordinal	Constructed demographic covariate	Ordinal income variable coded from lowest to highest income category.
race	Raw demographic variable	Race category from the demographic append file.
race_white	Constructed demographic covariate	Binary race indicator. Reference category is Non-white; comparison category is White.
hispanic	Raw demographic covariate	Hispanic identity. Reference category in models is No.
metro	Raw demographic covariate	Metropolitan status. Reference category in models is Non-metro.
compCheck_1	Raw manipulation check	Respondent answer identifying which country’s weapons system launched the strike.
compCheck_2	Raw manipulation check	Respondent answer identifying whether the weapons system was human-operated or autonomous.
comp_country_correct	Constructed manipulation check	Whether the respondent correctly identified the country whose weapons system launched the strike.
comp_autonomy_correct	Constructed manipulation check	Whether the respondent correctly identified the weapons-system condition.
comp_both_correct	Constructed manipulation check	Whether the respondent passed both manipulation checks.

Respondents per Condition

survey %>%
  count(condition_agressor, condition_autonomy)

## # A tibble: 5 × 3
##   condition_agressor condition_autonomy     n
##   <chr>              <chr>              <int>
## 1 China              autonomous           743
## 2 China              human                742
## 3 US                 autonomous           743
## 4 US                 human                742
## 5 <NA>               <NA>                 110

Results

Preregistered Analyses

The following sections implement the analysis plan in OSF_PreReg.docx. The primary analyses are intent-to-treat: manipulation-check failures are not excluded from the confirmatory models. The experimental indicators are coded so that the reference category is the human-operated U.S.-strike / Chinese-victim condition.

Preregistered Analysis Data

comfort_levels <- c(
  "Very uncomfortable",
  "Somewhat uncomfortable",
  "Neither comfortable or uncomfortable",
  "Somewhat comfortable",
  "Very comfortable"
)

education_levels <- c(
  "Some high school or less",
  "High school graduate or GED",
  "Some college, no degree",
  "2-year or associate degree",
  "4-year or bachelor degree",
  "Post-graduate degree"
)

income_levels <- c(
  "Less than $15,000",
  "$15,000 to under $50,000",
  "$50,000 to under $75,000",
  "$75,000 to under $100,000",
  "$100,000 to under $150,000",
  "$150,000 to under $200,000",
  "More than $200,000"
)

age_levels <- c("18-29", "30-49", "50-64", "65+")

prereg_analysis_data <- survey %>%
  filter(!is.na(consent),
         grepl("I am 18 or older", consent),
         !is.na(condition_agressor),
         !is.na(condition_autonomy),
         !is.na(intentionality)) %>%
  mutate(
    perceived_intentionality = as.numeric(intentionality),
    escalation_likelihood = as.numeric(escalation),
    escalation_justifiability = as.numeric(justified),
    country_scenario = ifelse(condition_agressor == "China", 1, 0),
    autonomous_ai = ifelse(condition_autonomy == "autonomous", 1, 0),
    country_condition = ifelse(country_scenario == 1, "Chinese strike / U.S. victim", "U.S. strike / Chinese victim"),
    weapons_condition = ifelse(autonomous_ai == 1, "Autonomous AI", "Human-operated"),
    survey_weight = suppressWarnings(as.numeric(weight)),
    us_ai_effectiveness = as.numeric(ai_us),
    china_ai_effectiveness = as.numeric(ai_china),
    southkorea_ai_effectiveness = as.numeric(ai_sk),
    relative_china_us_ai = china_ai_effectiveness - us_ai_effectiveness,
    age_group4 = factor(age_group4, levels = age_levels, ordered = TRUE),
    age_ordinal = as.numeric(age_group4),
    gender = factor(gender, levels = c("Male", "Female", "Other")),
    education = factor(education, levels = education_levels, ordered = TRUE),
    education_ordinal = as.numeric(education),
    income = factor(income, levels = income_levels, ordered = TRUE),
    income_ordinal = as.numeric(income),
    race = factor(
      race,
      levels = c(
        "White",
        "Black or African-American",
        "Asian or Asian-American",
        "Native American or American Indian",
        "Pacific Islander",
        "Mixed race",
        "Some other race"
      )
    ),
    race_white = factor(ifelse(race == "White", "White", "Non-white"),
                        levels = c("Non-white", "White")),
    hispanic = factor(hispanic, levels = c("No", "Yes")),
    metro = factor(metro, levels = c("Non-metro", "Metropolitan")),
    ai_comfort = (
      (match(comfort_doctor, comfort_levels) - 3) +
      (match(comfort_news, comfort_levels) - 3) +
      (match(comfort_banks, comfort_levels) - 3)
    ) / 3,
    pid = case_when(
      pid_base == "Democrat" ~ "Democrat",
      pid_base == "Republican" ~ "Republican",
      pid_base %in% c("Independent", "Other or none") &
        pid_lean == "The Democratic Party" ~ "Democrat",
      pid_base %in% c("Independent", "Other or none") &
        pid_lean == "The Republican Party" ~ "Republican",
      pid_base %in% c("Independent", "Other or none") &
        (pid_lean == "Neither" | is.na(pid_lean)) ~ "Independent",
      TRUE ~ NA_character_
    ),
    pid = factor(pid, levels = c("Independent", "Democrat", "Republican")),
    pid_republican = case_when(
      pid == "Republican" ~ 1,
      pid %in% c("Democrat", "Independent") ~ 0,
      TRUE ~ NA_real_
    ),
    gender_female = case_when(
      gender == "Female" ~ 1,
      gender %in% c("Male", "Other") ~ 0,
      TRUE ~ NA_real_
    ),
    comp_country_correct = case_when(
      condition_agressor == "China" ~ compCheck_1 == "A Chinese weapons system",
      condition_agressor == "US" ~ compCheck_1 == "A U.S. weapons system",
      TRUE ~ NA
    ),
    comp_autonomy_correct = case_when(
      condition_autonomy == "autonomous" ~ grepl("fully autonomous", compCheck_2),
      condition_autonomy == "human" ~ grepl("human-operated", compCheck_2),
      TRUE ~ NA
    ),
    comp_both_correct = comp_country_correct & comp_autonomy_correct
  )

prereg_primary_data <- prereg_analysis_data %>%
  filter(!is.na(perceived_intentionality),
         !is.na(ai_comfort),
         !is.na(relative_china_us_ai),
         !is.na(pid_republican),
         !is.na(age_ordinal),
         !is.na(gender_female),
         !is.na(education_ordinal),
         !is.na(income_ordinal),
         !is.na(race_white),
         !is.na(hispanic),
         !is.na(metro)) %>%
  mutate(
    ai_comfort_c = ai_comfort - mean(ai_comfort, na.rm = TRUE),
    us_ai_effectiveness_c = us_ai_effectiveness - mean(us_ai_effectiveness, na.rm = TRUE),
    china_ai_effectiveness_c = china_ai_effectiveness - mean(china_ai_effectiveness, na.rm = TRUE),
    southkorea_ai_effectiveness_c = southkorea_ai_effectiveness - mean(southkorea_ai_effectiveness, na.rm = TRUE),
    relative_china_us_ai_c = relative_china_us_ai - mean(relative_china_us_ai, na.rm = TRUE)
  )

prereg_sample_summary <- data.frame(
  sample = c("Assigned condition and primary outcome", "Covariate-complete primary model with demographics"),
  n = c(nrow(prereg_analysis_data), nrow(prereg_primary_data))
)

kable(prereg_sample_summary, caption = "Preregistered analytic sample sizes.")

Preregistered analytic sample sizes.

sample	n
Assigned condition and primary outcome	2935
Covariate-complete primary model with demographics	2833

Covariate Balance Checks

Coding note: these tables are diagnostics for the randomized design. They are not used to decide whether to keep or drop covariates. The goal is simply to show whether pretreatment characteristics look similar across the four randomly assigned experimental cells.

balance_data <- prereg_analysis_data %>%
  mutate(
    condition_cell = paste(country_condition, weapons_condition, sep = " | "),
    republican = pid_republican,
    female = gender_female,
    white = ifelse(race_white == "White", 1, 0),
    hispanic_yes = ifelse(hispanic == "Yes", 1, 0),
    metro_yes = ifelse(metro == "Metropolitan", 1, 0)
  )

numeric_balance <- bind_rows(
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "AI comfort",
      n = sum(!is.na(ai_comfort)),
      mean = mean(ai_comfort, na.rm = TRUE),
      sd = sd(ai_comfort, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "China minus U.S. AI effectiveness",
      n = sum(!is.na(relative_china_us_ai)),
      mean = mean(relative_china_us_ai, na.rm = TRUE),
      sd = sd(relative_china_us_ai, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Age group (ordinal)",
      n = sum(!is.na(age_ordinal)),
      mean = mean(age_ordinal, na.rm = TRUE),
      sd = sd(age_ordinal, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Education (ordinal)",
      n = sum(!is.na(education_ordinal)),
      mean = mean(education_ordinal, na.rm = TRUE),
      sd = sd(education_ordinal, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Income (ordinal)",
      n = sum(!is.na(income_ordinal)),
      mean = mean(income_ordinal, na.rm = TRUE),
      sd = sd(income_ordinal, na.rm = TRUE),
      .groups = "drop"
    )
) %>%
  arrange(covariate, condition_cell)

binary_balance <- bind_rows(
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Republican",
      n = sum(!is.na(republican)),
      proportion = mean(republican, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Female",
      n = sum(!is.na(female)),
      proportion = mean(female, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "White",
      n = sum(!is.na(white)),
      proportion = mean(white, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Hispanic",
      n = sum(!is.na(hispanic_yes)),
      proportion = mean(hispanic_yes, na.rm = TRUE),
      .groups = "drop"
    ),
  balance_data %>%
    group_by(condition_cell) %>%
    summarize(
      covariate = "Metropolitan",
      n = sum(!is.na(metro_yes)),
      proportion = mean(metro_yes, na.rm = TRUE),
      .groups = "drop"
    )
) %>%
  arrange(covariate, condition_cell)

kable(
  numeric_balance,
  digits = 2,
  caption = "Covariate balance check for continuous and ordinal pretreatment variables by experimental cell."
)

Covariate balance check for continuous and ordinal pretreatment variables by experimental cell.

condition_cell	covariate	n	mean	sd
Chinese strike / U.S. victim | Autonomous AI	AI comfort	733	0.52	1.09
Chinese strike / U.S. victim | Human-operated	AI comfort	735	0.37	1.08
U.S. strike / Chinese victim | Autonomous AI	AI comfort	724	0.47	1.09
U.S. strike / Chinese victim | Human-operated	AI comfort	730	0.43	1.10
Chinese strike / U.S. victim | Autonomous AI	Age group (ordinal)	726	2.72	0.95
Chinese strike / U.S. victim | Human-operated	Age group (ordinal)	728	2.72	0.94
U.S. strike / Chinese victim | Autonomous AI	Age group (ordinal)	725	2.73	0.94
U.S. strike / Chinese victim | Human-operated	Age group (ordinal)	725	2.70	0.95
Chinese strike / U.S. victim | Autonomous AI	China minus U.S. AI effectiveness	720	-7.38	21.80
Chinese strike / U.S. victim | Human-operated	China minus U.S. AI effectiveness	722	-8.52	25.11
U.S. strike / Chinese victim | Autonomous AI	China minus U.S. AI effectiveness	716	-8.65	26.67
U.S. strike / Chinese victim | Human-operated	China minus U.S. AI effectiveness	718	-6.80	24.39
Chinese strike / U.S. victim | Autonomous AI	Education (ordinal)	727	3.91	1.51
Chinese strike / U.S. victim | Human-operated	Education (ordinal)	729	3.90	1.54
U.S. strike / Chinese victim | Autonomous AI	Education (ordinal)	725	3.89	1.52
U.S. strike / Chinese victim | Human-operated	Education (ordinal)	725	3.98	1.48
Chinese strike / U.S. victim | Autonomous AI	Income (ordinal)	727	3.68	1.74
Chinese strike / U.S. victim | Human-operated	Income (ordinal)	729	3.50	1.70
U.S. strike / Chinese victim | Autonomous AI	Income (ordinal)	725	3.52	1.64
U.S. strike / Chinese victim | Human-operated	Income (ordinal)	724	3.60	1.70

kable(
  binary_balance,
  digits = 3,
  caption = "Covariate balance check for binary demographic variables by experimental cell."
)

Covariate balance check for binary demographic variables by experimental cell.

condition_cell	covariate	n	proportion
Chinese strike / U.S. victim | Autonomous AI	Female	727	0.549
Chinese strike / U.S. victim | Human-operated	Female	729	0.539
U.S. strike / Chinese victim | Autonomous AI	Female	725	0.570
U.S. strike / Chinese victim | Human-operated	Female	725	0.539
Chinese strike / U.S. victim | Autonomous AI	Hispanic	727	0.122
Chinese strike / U.S. victim | Human-operated	Hispanic	729	0.132
U.S. strike / Chinese victim | Autonomous AI	Hispanic	725	0.124
U.S. strike / Chinese victim | Human-operated	Hispanic	725	0.112
Chinese strike / U.S. victim | Autonomous AI	Metropolitan	727	0.864
Chinese strike / U.S. victim | Human-operated	Metropolitan	728	0.889
U.S. strike / Chinese victim | Autonomous AI	Metropolitan	725	0.884
U.S. strike / Chinese victim | Human-operated	Metropolitan	725	0.887
Chinese strike / U.S. victim | Autonomous AI	Republican	727	0.374
Chinese strike / U.S. victim | Human-operated	Republican	729	0.362
U.S. strike / Chinese victim | Autonomous AI	Republican	725	0.374
U.S. strike / Chinese victim | Human-operated	Republican	725	0.370
Chinese strike / U.S. victim | Autonomous AI	White	727	0.757
Chinese strike / U.S. victim | Human-operated	White	729	0.734
U.S. strike / Chinese victim | Autonomous AI	White	725	0.746
U.S. strike / Chinese victim | Human-operated	White	725	0.757

Complete-Case Diagnostics

The primary adjusted model uses complete cases for the outcome, preregistered covariates, and demographic controls. This section reports how many respondents are excluded by that requirement and whether the retained and excluded respondents differ on observed variables. These diagnostics help contextualize the covariate-adjusted model; they do not change the preregistered estimand.

complete_case_diagnostics <- prereg_analysis_data %>%
  mutate(
    primary_model_sample = ifelse(ResponseId %in% prereg_primary_data$ResponseId, "Included", "Dropped"),
    republican = pid_republican,
    female = gender_female,
    white = ifelse(race_white == "White", 1, 0),
    hispanic_yes = ifelse(hispanic == "Yes", 1, 0),
    metro_yes = ifelse(metro == "Metropolitan", 1, 0)
  )

complete_case_counts <- complete_case_diagnostics %>%
  count(primary_model_sample) %>%
  mutate(percent = n / sum(n))

complete_case_numeric <- bind_rows(
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Perceived intentionality", n = sum(!is.na(perceived_intentionality)), mean = mean(perceived_intentionality, na.rm = TRUE), sd = sd(perceived_intentionality, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "AI comfort", n = sum(!is.na(ai_comfort)), mean = mean(ai_comfort, na.rm = TRUE), sd = sd(ai_comfort, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "China minus U.S. AI effectiveness", n = sum(!is.na(relative_china_us_ai)), mean = mean(relative_china_us_ai, na.rm = TRUE), sd = sd(relative_china_us_ai, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Age group (ordinal)", n = sum(!is.na(age_ordinal)), mean = mean(age_ordinal, na.rm = TRUE), sd = sd(age_ordinal, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Education (ordinal)", n = sum(!is.na(education_ordinal)), mean = mean(education_ordinal, na.rm = TRUE), sd = sd(education_ordinal, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Income (ordinal)", n = sum(!is.na(income_ordinal)), mean = mean(income_ordinal, na.rm = TRUE), sd = sd(income_ordinal, na.rm = TRUE), .groups = "drop")
) %>%
  arrange(variable, primary_model_sample)

complete_case_binary <- bind_rows(
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Republican", n = sum(!is.na(republican)), proportion = mean(republican, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Female", n = sum(!is.na(female)), proportion = mean(female, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "White", n = sum(!is.na(white)), proportion = mean(white, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Hispanic", n = sum(!is.na(hispanic_yes)), proportion = mean(hispanic_yes, na.rm = TRUE), .groups = "drop"),
  complete_case_diagnostics %>%
    group_by(primary_model_sample) %>%
    summarize(variable = "Metropolitan", n = sum(!is.na(metro_yes)), proportion = mean(metro_yes, na.rm = TRUE), .groups = "drop")
) %>%
  arrange(variable, primary_model_sample)

kable(complete_case_counts, digits = 3, caption = "Respondents included in and dropped from the covariate-complete primary model.")

Respondents included in and dropped from the covariate-complete primary model.

primary_model_sample	n	percent
Dropped	102	0.035
Included	2833	0.965

kable(complete_case_numeric, digits = 2, caption = "Observed differences between included and dropped respondents for numeric variables.")

Observed differences between included and dropped respondents for numeric variables.

primary_model_sample	variable	n	mean	sd
Dropped	AI comfort	89	0.17	1.19
Included	AI comfort	2833	0.46	1.09
Dropped	Age group (ordinal)	71	2.75	1.02
Included	Age group (ordinal)	2833	2.71	0.94
Dropped	China minus U.S. AI effectiveness	43	-9.28	24.01
Included	China minus U.S. AI effectiveness	2833	-7.82	24.57
Dropped	Education (ordinal)	73	4.04	1.42
Included	Education (ordinal)	2833	3.92	1.51
Dropped	Income (ordinal)	72	3.58	1.75
Included	Income (ordinal)	2833	3.57	1.69
Dropped	Perceived intentionality	102	60.57	28.12
Included	Perceived intentionality	2833	62.92	24.84

kable(complete_case_binary, digits = 3, caption = "Observed differences between included and dropped respondents for binary variables.")

Observed differences between included and dropped respondents for binary variables.

primary_model_sample	variable	n	proportion
Dropped	Female	73	0.507
Included	Female	2833	0.550
Dropped	Hispanic	73	0.096
Included	Hispanic	2833	0.123
Dropped	Metropolitan	72	0.903
Included	Metropolitan	2833	0.880
Dropped	Republican	73	0.301
Included	Republican	2833	0.372
Dropped	White	73	0.726
Included	White	2833	0.749

Cell Means for Primary Outcome

prereg_cell_means <- prereg_analysis_data %>%
  group_by(country_condition, weapons_condition) %>%
  summarize(
    n = n(),
    mean = mean(perceived_intentionality),
    se = sd(perceived_intentionality) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

kable(
  prereg_cell_means,
  digits = 2,
  caption = "Mean perceived intentionality and 95% confidence intervals by experimental cell."
)

Mean perceived intentionality and 95% confidence intervals by experimental cell.

country_condition	weapons_condition	n	mean	se	ci_low	ci_high
Chinese strike / U.S. victim	Autonomous AI	737	62.09	0.87	60.38	63.80
Chinese strike / U.S. victim	Human-operated	737	72.11	0.85	70.45	73.77
U.S. strike / Chinese victim	Autonomous AI	728	54.21	0.94	52.37	56.06
U.S. strike / Chinese victim	Human-operated	733	62.83	0.91	61.06	64.61

prereg_cell_means_plot_data <- prereg_cell_means %>%
  mutate(
    country_condition = factor(
      country_condition,
      levels = c("Chinese strike / U.S. victim", "U.S. strike / Chinese victim"),
      labels = c("Chinese strike", "U.S. strike")
    ),
    weapons_condition = factor(
      weapons_condition,
      levels = c("Human-operated", "Autonomous AI")
    )
  )

ggplot(
  prereg_cell_means_plot_data,
  aes(x = weapons_condition, y = mean, color = country_condition, group = country_condition)
) +
  geom_line(linewidth = 1, alpha = 0.45) +
  geom_point(size = 3.3) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.8) +
  scale_color_manual(
    values = c("Chinese strike" = "#A63A33", "U.S. strike" = "#235A8C"),
    name = "Country scenario"
  ) +
  scale_y_continuous(
    limits = c(50, 80),
    breaks = seq(50, 80, 5)
  ) +
  labs(
    title = "Preregistered Cell Means for Perceived Intentionality",
    x = "Weapons-system condition",
    y = "Perceived intentionality",
    caption = "Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 14) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank()
  )

Primary Preregistered Intentionality Model

This model regresses perceived intentionality on the country-scenario indicator, the autonomous-AI indicator, their interaction, preregistered covariates, and demographic controls. Continuous covariates are demeaned. HC2 heteroskedasticity-consistent standard errors are used for two-sided tests with alpha = .05.

prereg_primary_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

prereg_primary_vcov <- vcovHC(prereg_primary_model, type = "HC2")
prereg_primary_coeftest <- coeftest(prereg_primary_model, vcov. = prereg_primary_vcov)
prereg_primary_ci <- coefci(prereg_primary_model, vcov. = prereg_primary_vcov)

prereg_primary_results <- data.frame(
  term = rownames(prereg_primary_coeftest),
  estimate = prereg_primary_coeftest[, 1],
  hc2_se = prereg_primary_coeftest[, 2],
  statistic = prereg_primary_coeftest[, 3],
  p_value = prereg_primary_coeftest[, 4],
  ci_low = prereg_primary_ci[, 1],
  ci_high = prereg_primary_ci[, 2],
  row.names = NULL
)

kable(
  prereg_primary_results,
  digits = 3,
  caption = "Primary preregistered OLS model with HC2 standard errors."
)

Primary preregistered OLS model with HC2 standard errors.

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
(Intercept)	64.156	2.451	26.180	0.000	59.351	68.962
country_scenario	9.039	1.253	7.214	0.000	6.582	11.496
autonomous_ai	-9.030	1.316	-6.864	0.000	-11.609	-6.450
ai_comfort_c	1.284	0.455	2.825	0.005	0.393	2.176
relative_china_us_ai_c	0.047	0.021	2.251	0.024	0.006	0.088
pid_republican	2.034	0.975	2.087	0.037	0.123	3.945
age_ordinal	-0.578	0.500	-1.155	0.248	-1.558	0.403
gender_female	0.142	0.925	0.154	0.878	-1.672	1.956
education_ordinal	-0.178	0.345	-0.516	0.606	-0.854	0.498
income_ordinal	-0.047	0.325	-0.144	0.886	-0.683	0.590
race_whiteWhite	-1.212	1.091	-1.110	0.267	-3.351	0.928
hispanicYes	2.772	1.378	2.012	0.044	0.071	5.473
metroMetropolitan	1.390	1.391	0.999	0.318	-1.338	4.117
country_scenario:autonomous_ai	-1.231	1.806	-0.682	0.496	-4.773	2.311

Unadjusted and Adjusted Treatment Estimates

Coding note: the unadjusted model estimates the treatment-cell differences using only the randomized country-scenario indicator, the autonomous-AI indicator, and their interaction. The adjusted model is the preregistered primary model above. Showing them side by side makes it easier to see whether covariate adjustment changes the substantive treatment estimates.

prereg_unadjusted_full_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai,
  data = prereg_analysis_data
)

prereg_unadjusted_complete_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai,
  data = prereg_primary_data
)

prereg_unadjusted_full_vcov <- vcovHC(prereg_unadjusted_full_model, type = "HC2")
prereg_unadjusted_full_coeftest <- coeftest(prereg_unadjusted_full_model, vcov. = prereg_unadjusted_full_vcov)
prereg_unadjusted_full_ci <- coefci(prereg_unadjusted_full_model, vcov. = prereg_unadjusted_full_vcov)

prereg_unadjusted_complete_vcov <- vcovHC(prereg_unadjusted_complete_model, type = "HC2")
prereg_unadjusted_complete_coeftest <- coeftest(prereg_unadjusted_complete_model, vcov. = prereg_unadjusted_complete_vcov)
prereg_unadjusted_complete_ci <- coefci(prereg_unadjusted_complete_model, vcov. = prereg_unadjusted_complete_vcov)

prereg_unadjusted_full_results <- data.frame(
  model = "Unadjusted, full outcome sample",
  term = rownames(prereg_unadjusted_full_coeftest),
  estimate = prereg_unadjusted_full_coeftest[, 1],
  hc2_se = prereg_unadjusted_full_coeftest[, 2],
  p_value = prereg_unadjusted_full_coeftest[, 4],
  ci_low = prereg_unadjusted_full_ci[, 1],
  ci_high = prereg_unadjusted_full_ci[, 2],
  row.names = NULL
)

prereg_unadjusted_complete_results <- data.frame(
  model = "Unadjusted, covariate-complete sample",
  term = rownames(prereg_unadjusted_complete_coeftest),
  estimate = prereg_unadjusted_complete_coeftest[, 1],
  hc2_se = prereg_unadjusted_complete_coeftest[, 2],
  p_value = prereg_unadjusted_complete_coeftest[, 4],
  ci_low = prereg_unadjusted_complete_ci[, 1],
  ci_high = prereg_unadjusted_complete_ci[, 2],
  row.names = NULL
)

prereg_adjusted_treatment_results <- prereg_primary_results %>%
  mutate(model = "Adjusted, preregistered primary model", .before = term)

unadjusted_adjusted_treatment_terms <- c(
  "(Intercept)",
  "country_scenario",
  "autonomous_ai",
  "country_scenario:autonomous_ai"
)

prereg_unadjusted_adjusted_comparison <- bind_rows(
  prereg_unadjusted_full_results,
  prereg_unadjusted_complete_results,
  prereg_adjusted_treatment_results
) %>%
  filter(term %in% unadjusted_adjusted_treatment_terms)

kable(
  prereg_unadjusted_adjusted_comparison,
  digits = 3,
  caption = "Unadjusted and adjusted treatment estimates for perceived intentionality. All standard errors are HC2."
)

Unadjusted and adjusted treatment estimates for perceived intentionality. All standard errors are HC2.

model	term	estimate	hc2_se	p_value	ci_low	ci_high	statistic
Unadjusted, full outcome sample	(Intercept)	62.834	0.906	0.000	61.058	64.609	NA
Unadjusted, full outcome sample	country_scenario	9.276	1.240	0.000	6.845	11.708	NA
Unadjusted, full outcome sample	autonomous_ai	-8.622	1.306	0.000	-11.182	-6.062	NA
Unadjusted, full outcome sample	country_scenario:autonomous_ai	-1.397	1.784	0.434	-4.894	2.100	NA
Unadjusted, covariate-complete sample	(Intercept)	63.218	0.918	0.000	61.417	65.019	NA
Unadjusted, covariate-complete sample	country_scenario	8.958	1.253	0.000	6.501	11.415	NA
Unadjusted, covariate-complete sample	autonomous_ai	-9.023	1.318	0.000	-11.608	-6.439	NA
Unadjusted, covariate-complete sample	country_scenario:autonomous_ai	-1.090	1.805	0.546	-4.629	2.449	NA
Adjusted, preregistered primary model	(Intercept)	64.156	2.451	0.000	59.351	68.962	26.180
Adjusted, preregistered primary model	country_scenario	9.039	1.253	0.000	6.582	11.496	7.214
Adjusted, preregistered primary model	autonomous_ai	-9.030	1.316	0.000	-11.609	-6.450	-6.864
Adjusted, preregistered primary model	country_scenario:autonomous_ai	-1.231	1.806	0.496	-4.773	2.311	-0.682

Planned Contrasts for H1-H4

prereg_contrast_weights <- matrix(
  0,
  nrow = 4,
  ncol = length(coef(prereg_primary_model))
)
colnames(prereg_contrast_weights) <- names(coef(prereg_primary_model))
rownames(prereg_contrast_weights) <- c(
  "H1: Chinese strike vs. U.S. strike among human-operated systems",
  "Autonomous AI effect in U.S.-strike / Chinese-victim condition",
  "Autonomous AI effect in Chinese-strike / U.S.-victim condition",
  "Country-scenario by autonomous-AI interaction"
)

prereg_contrast_weights["H1: Chinese strike vs. U.S. strike among human-operated systems", "country_scenario"] <- 1
prereg_contrast_weights["Autonomous AI effect in U.S.-strike / Chinese-victim condition", "autonomous_ai"] <- 1
prereg_contrast_weights["Autonomous AI effect in Chinese-strike / U.S.-victim condition", "autonomous_ai"] <- 1
prereg_contrast_weights["Autonomous AI effect in Chinese-strike / U.S.-victim condition", "country_scenario:autonomous_ai"] <- 1
prereg_contrast_weights["Country-scenario by autonomous-AI interaction", "country_scenario:autonomous_ai"] <- 1

prereg_contrast_estimate <- as.numeric(prereg_contrast_weights %*% coef(prereg_primary_model))
prereg_contrast_se <- sqrt(diag(prereg_contrast_weights %*% prereg_primary_vcov %*% t(prereg_contrast_weights)))
prereg_contrast_statistic <- prereg_contrast_estimate / prereg_contrast_se
prereg_contrast_p <- 2 * pt(abs(prereg_contrast_statistic), df = df.residual(prereg_primary_model), lower.tail = FALSE)
prereg_contrast_critical <- qt(0.975, df = df.residual(prereg_primary_model))

prereg_contrasts <- data.frame(
  contrast = rownames(prereg_contrast_weights),
  estimate = prereg_contrast_estimate,
  hc2_se = prereg_contrast_se,
  statistic = prereg_contrast_statistic,
  p_value = prereg_contrast_p,
  ci_low = prereg_contrast_estimate - prereg_contrast_critical * prereg_contrast_se,
  ci_high = prereg_contrast_estimate + prereg_contrast_critical * prereg_contrast_se,
  row.names = NULL
)

kable(
  prereg_contrasts,
  digits = 3,
  caption = "Planned treatment contrasts using HC2 standard errors."
)

Planned treatment contrasts using HC2 standard errors.

contrast	estimate	hc2_se	statistic	p_value	ci_low	ci_high
H1: Chinese strike vs. U.S. strike among human-operated systems	9.039	1.253	7.214	0.000	6.582	11.496
Autonomous AI effect in U.S.-strike / Chinese-victim condition	-9.030	1.316	-6.864	0.000	-11.609	-6.450
Autonomous AI effect in Chinese-strike / U.S.-victim condition	-10.261	1.237	-8.295	0.000	-12.686	-7.835
Country-scenario by autonomous-AI interaction	-1.231	1.806	-0.682	0.496	-4.773	2.311

prereg_contrast_plot_data <- prereg_contrasts %>%
  mutate(
    contrast_label = case_when(
      grepl("^H1", contrast) ~ "H1: Chinese vs. U.S. strike\namong human-operated systems",
      grepl("U.S.-strike", contrast) ~ "Autonomous effect\nin U.S.-strike condition",
      grepl("Chinese-strike", contrast) ~ "Autonomous effect\nin Chinese-strike condition",
      TRUE ~ "Country scenario x\nautonomous AI interaction"
    ),
    contrast_label = factor(
      contrast_label,
      levels = c(
        "H1: Chinese vs. U.S. strike\namong human-operated systems",
        "Autonomous effect\nin U.S.-strike condition",
        "Autonomous effect\nin Chinese-strike condition",
        "Country scenario x\nautonomous AI interaction"
      )
    )
  )

ggplot(prereg_contrast_plot_data, aes(x = estimate, y = contrast_label)) +
  geom_vline(xintercept = 0, linewidth = 0.5, color = "gray55") +
  geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.14, linewidth = 0.8) +
  geom_point(size = 3) +
  scale_y_discrete(limits = rev(levels(prereg_contrast_plot_data$contrast_label))) +
  labs(
    title = "Preregistered Treatment Contrasts for Intentionality",
    x = "Estimated effect on perceived intentionality",
    y = NULL,
    caption = "Points are OLS estimates; error bars are 95% CIs using HC2 standard errors."
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.major.y = element_blank()
  )

## Warning: `geom_errobarh()` was deprecated in ggplot2 4.0.0.
## ℹ Please use the `orientation` argument of `geom_errorbar()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

## `height` was translated to `width`.

Weighted Primary Model

Coding note: this is a robustness check using the appended respondent weight. It keeps the same outcome, treatment indicators, interaction, preregistered covariates, and demographic controls as the primary model. The unweighted covariate-adjusted model remains the primary preregistered analysis unless the preregistration explicitly says weights are primary.

prereg_weighted_data <- prereg_primary_data %>%
  filter(!is.na(survey_weight),
         survey_weight > 0)

prereg_weighted_primary_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_weighted_data,
  weights = survey_weight
)

prereg_weighted_primary_vcov <- vcovHC(prereg_weighted_primary_model, type = "HC2")
prereg_weighted_primary_coeftest <- coeftest(prereg_weighted_primary_model, vcov. = prereg_weighted_primary_vcov)
prereg_weighted_primary_ci <- coefci(prereg_weighted_primary_model, vcov. = prereg_weighted_primary_vcov)

prereg_weighted_primary_results <- data.frame(
  term = rownames(prereg_weighted_primary_coeftest),
  estimate = prereg_weighted_primary_coeftest[, 1],
  hc2_se = prereg_weighted_primary_coeftest[, 2],
  statistic = prereg_weighted_primary_coeftest[, 3],
  p_value = prereg_weighted_primary_coeftest[, 4],
  ci_low = prereg_weighted_primary_ci[, 1],
  ci_high = prereg_weighted_primary_ci[, 2],
  row.names = NULL
)

kable(
  prereg_weighted_primary_results,
  digits = 3,
  caption = paste0("Weighted version of the primary preregistered model using HC2 standard errors (N = ", nrow(prereg_weighted_data), ").")
)

Weighted version of the primary preregistered model using HC2 standard errors (N = 2833).

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
(Intercept)	60.343	2.962	20.375	0.000	54.536	66.150
country_scenario	8.639	1.527	5.658	0.000	5.645	11.633
autonomous_ai	-9.394	1.638	-5.735	0.000	-12.606	-6.182
ai_comfort_c	1.278	0.570	2.244	0.025	0.161	2.395
relative_china_us_ai_c	0.029	0.024	1.187	0.235	-0.019	0.077
pid_republican	1.869	1.186	1.575	0.115	-0.457	4.195
age_ordinal	0.131	0.590	0.223	0.824	-1.025	1.288
gender_female	-0.282	1.143	-0.247	0.805	-2.524	1.960
education_ordinal	-0.141	0.433	-0.326	0.744	-0.991	0.708
income_ordinal	0.064	0.428	0.149	0.882	-0.775	0.903
race_whiteWhite	-2.445	1.386	-1.764	0.078	-5.162	0.273
hispanicYes	3.827	1.551	2.467	0.014	0.786	6.869
metroMetropolitan	3.462	1.715	2.019	0.044	0.100	6.824
country_scenario:autonomous_ai	0.460	2.238	0.206	0.837	-3.928	4.848

Alternative Demographic Coding Robustness Check

The primary model treats age, education, and income as ordinal controls. This robustness check treats those variables as categorical factors instead. The purpose is to check whether the treatment estimates depend on the linearity assumption implied by ordinal coding.

prereg_categorical_demographics_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_group4 + gender_female + education + income +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

prereg_categorical_demographics_vcov <- vcovHC(prereg_categorical_demographics_model, type = "HC2")
prereg_categorical_demographics_coeftest <- coeftest(prereg_categorical_demographics_model, vcov. = prereg_categorical_demographics_vcov)
prereg_categorical_demographics_ci <- coefci(prereg_categorical_demographics_model, vcov. = prereg_categorical_demographics_vcov)

prereg_categorical_demographics_results <- data.frame(
  term = rownames(prereg_categorical_demographics_coeftest),
  estimate = prereg_categorical_demographics_coeftest[, 1],
  hc2_se = prereg_categorical_demographics_coeftest[, 2],
  statistic = prereg_categorical_demographics_coeftest[, 3],
  p_value = prereg_categorical_demographics_coeftest[, 4],
  ci_low = prereg_categorical_demographics_ci[, 1],
  ci_high = prereg_categorical_demographics_ci[, 2],
  row.names = NULL
)

prereg_categorical_demographics_treatment_results <- prereg_categorical_demographics_results %>%
  filter(term %in% c("country_scenario", "autonomous_ai", "country_scenario:autonomous_ai"))

kable(
  prereg_categorical_demographics_treatment_results,
  digits = 3,
  caption = "Treatment coefficients from the robustness model with age, education, and income entered as categorical controls."
)

Treatment coefficients from the robustness model with age, education, and income entered as categorical controls.

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
country_scenario	9.007	1.254	7.183	0.000	6.548	11.466
autonomous_ai	-9.066	1.314	-6.897	0.000	-11.644	-6.489
country_scenario:autonomous_ai	-1.221	1.806	-0.676	0.499	-4.762	2.319

Treatment-Estimate Robustness Summary

This table gathers the focal treatment coefficients across the main specification checks above. It is intended as a compact audit of whether the main treatment pattern changes across unadjusted, adjusted, weighted, and categorical-demographic specifications.

treatment_robustness_results <- bind_rows(
  prereg_unadjusted_full_results %>%
    mutate(specification = "Unadjusted, full outcome sample", .before = term),
  prereg_unadjusted_complete_results %>%
    mutate(specification = "Unadjusted, covariate-complete sample", .before = term),
  prereg_primary_results %>%
    mutate(specification = "Adjusted primary model", .before = term),
  prereg_weighted_primary_results %>%
    mutate(specification = "Weighted adjusted primary model", .before = term),
  prereg_categorical_demographics_results %>%
    mutate(specification = "Adjusted model with categorical demographics", .before = term)
) %>%
  filter(term %in% c("country_scenario", "autonomous_ai", "country_scenario:autonomous_ai")) %>%
  mutate(
    term_label = case_when(
      term == "country_scenario" ~ "Chinese vs. U.S. strike among human-operated systems",
      term == "autonomous_ai" ~ "Autonomous effect in U.S.-strike condition",
      TRUE ~ "Country scenario x autonomous AI interaction"
    )
  ) %>%
  select(specification, term_label, estimate, hc2_se, p_value, ci_low, ci_high)

kable(
  treatment_robustness_results,
  digits = 3,
  caption = "Focal treatment coefficients across primary and robustness specifications."
)

Focal treatment coefficients across primary and robustness specifications.

specification	term_label	estimate	hc2_se	p_value	ci_low	ci_high
Unadjusted, full outcome sample	Chinese vs. U.S. strike among human-operated systems	9.276	1.240	0.000	6.845	11.708
Unadjusted, full outcome sample	Autonomous effect in U.S.-strike condition	-8.622	1.306	0.000	-11.182	-6.062
Unadjusted, full outcome sample	Country scenario x autonomous AI interaction	-1.397	1.784	0.434	-4.894	2.100
Unadjusted, covariate-complete sample	Chinese vs. U.S. strike among human-operated systems	8.958	1.253	0.000	6.501	11.415
Unadjusted, covariate-complete sample	Autonomous effect in U.S.-strike condition	-9.023	1.318	0.000	-11.608	-6.439
Unadjusted, covariate-complete sample	Country scenario x autonomous AI interaction	-1.090	1.805	0.546	-4.629	2.449
Adjusted primary model	Chinese vs. U.S. strike among human-operated systems	9.039	1.253	0.000	6.582	11.496
Adjusted primary model	Autonomous effect in U.S.-strike condition	-9.030	1.316	0.000	-11.609	-6.450
Adjusted primary model	Country scenario x autonomous AI interaction	-1.231	1.806	0.496	-4.773	2.311
Weighted adjusted primary model	Chinese vs. U.S. strike among human-operated systems	8.639	1.527	0.000	5.645	11.633
Weighted adjusted primary model	Autonomous effect in U.S.-strike condition	-9.394	1.638	0.000	-12.606	-6.182
Weighted adjusted primary model	Country scenario x autonomous AI interaction	0.460	2.238	0.837	-3.928	4.848
Adjusted model with categorical demographics	Chinese vs. U.S. strike among human-operated systems	9.007	1.254	0.000	6.548	11.466
Adjusted model with categorical demographics	Autonomous effect in U.S.-strike condition	-9.066	1.314	0.000	-11.644	-6.489
Adjusted model with categorical demographics	Country scenario x autonomous AI interaction	-1.221	1.806	0.499	-4.762	2.319

Equivalence-Style Tests for Negligible Effects

For the preregistered equivalence-style tests, a meaningful effect is defined as Cohen’s D = .20. The table converts that threshold into raw intentionality points using the standard deviation of the primary model sample. These tests are useful when the autonomous-AI coefficients are small: instead of only asking whether an effect is statistically distinguishable from zero, they ask whether the estimate is statistically contained inside the preregistered negligible-effect range.

sesoi_intentionality <- 0.20 * sd(prereg_primary_data$perceived_intentionality, na.rm = TRUE)

prereg_equivalence_weights <- prereg_contrast_weights[
  c(
    "Autonomous AI effect in U.S.-strike / Chinese-victim condition",
    "Autonomous AI effect in Chinese-strike / U.S.-victim condition"
  ),
  ,
  drop = FALSE
]

prereg_equivalence_estimate <- as.numeric(prereg_equivalence_weights %*% coef(prereg_primary_model))
prereg_equivalence_se <- sqrt(diag(prereg_equivalence_weights %*% prereg_primary_vcov %*% t(prereg_equivalence_weights)))
prereg_equivalence_p_lower <- pt(
  (prereg_equivalence_estimate - (-sesoi_intentionality)) / prereg_equivalence_se,
  df = df.residual(prereg_primary_model),
  lower.tail = FALSE
)
prereg_equivalence_p_upper <- pt(
  (prereg_equivalence_estimate - sesoi_intentionality) / prereg_equivalence_se,
  df = df.residual(prereg_primary_model),
  lower.tail = TRUE
)

prereg_equivalence_tests <- data.frame(
  contrast = rownames(prereg_equivalence_weights),
  estimate = prereg_equivalence_estimate,
  lower_bound = -sesoi_intentionality,
  upper_bound = sesoi_intentionality,
  hc2_se = prereg_equivalence_se,
  p_lower = prereg_equivalence_p_lower,
  p_upper = prereg_equivalence_p_upper,
  p_tost = pmax(prereg_equivalence_p_lower, prereg_equivalence_p_upper),
  equivalent_at_alpha_05 = pmax(prereg_equivalence_p_lower, prereg_equivalence_p_upper) < 0.05,
  row.names = NULL
)

kable(
  prereg_equivalence_tests,
  digits = 3,
  caption = paste0("Equivalence-style tests using +/- ", round(sesoi_intentionality, 2), " intentionality points as the meaningful-effect threshold.")
)

Equivalence-style tests using +/- 4.97 intentionality points as the meaningful-effect threshold.

contrast	estimate	lower_bound	upper_bound	hc2_se	p_lower	p_upper	p_tost	equivalent_at_alpha_05
Autonomous AI effect in U.S.-strike / Chinese-victim condition	-9.030	-4.967	4.967	1.316	0.999	0	0.999	FALSE
Autonomous AI effect in Chinese-strike / U.S.-victim condition	-10.261	-4.967	4.967	1.237	1.000	0	1.000	FALSE

equivalence_critical <- qt(0.95, df = df.residual(prereg_primary_model))

prereg_equivalence_plot_data <- prereg_equivalence_tests %>%
  mutate(
    ci90_low = estimate - equivalence_critical * hc2_se,
    ci90_high = estimate + equivalence_critical * hc2_se,
    contrast_label = case_when(
      grepl("U.S.-strike", contrast) ~ "Autonomous effect\nin U.S.-strike condition",
      TRUE ~ "Autonomous effect\nin Chinese-strike condition"
    ),
    contrast_label = factor(
      contrast_label,
      levels = c(
        "Autonomous effect\nin U.S.-strike condition",
        "Autonomous effect\nin Chinese-strike condition"
      )
    )
  )

ggplot(prereg_equivalence_plot_data, aes(x = estimate, y = contrast_label)) +
  annotate(
    "rect",
    xmin = -sesoi_intentionality,
    xmax = sesoi_intentionality,
    ymin = -Inf,
    ymax = Inf,
    alpha = 0.12,
    fill = "gray60"
  ) +
  geom_vline(xintercept = 0, linewidth = 0.5, color = "gray50") +
  geom_vline(xintercept = c(-sesoi_intentionality, sesoi_intentionality),
             linewidth = 0.6,
             linetype = "dashed",
             color = "gray40") +
  geom_errorbarh(aes(xmin = ci90_low, xmax = ci90_high), height = 0.14, linewidth = 0.8) +
  geom_point(size = 3) +
  scale_y_discrete(limits = rev(levels(prereg_equivalence_plot_data$contrast_label))) +
  labs(
    title = "Equivalence Tests for Negligible Autonomous-AI Effects",
    x = "Estimated autonomous-AI effect on perceived intentionality",
    y = NULL,
    caption = paste0("Shaded area is the preregistered negligible-effect region (+/- ", round(sesoi_intentionality, 2), " points). Error bars are 90% CIs for TOST.")
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.major.y = element_blank()
  )

## `height` was translated to `width`.

Manipulation Check Descriptives

The current Qualtrics export contains manipulation checks for the country whose weapons system launched the missiles and whether the system was autonomous or human-operated. It does not contain a separate victim-country manipulation check.

manipulation_check_overall <- prereg_analysis_data %>%
  summarize(
    n = n(),
    country_check_response_n = sum(!is.na(comp_country_correct)),
    country_check_correct = mean(comp_country_correct, na.rm = TRUE),
    autonomy_check_response_n = sum(!is.na(comp_autonomy_correct)),
    autonomy_check_correct = mean(comp_autonomy_correct, na.rm = TRUE),
    both_checks_correct = mean(comp_both_correct, na.rm = TRUE)
  )

manipulation_check_by_condition <- prereg_analysis_data %>%
  group_by(country_condition, weapons_condition) %>%
  summarize(
    n = n(),
    country_check_correct = mean(comp_country_correct, na.rm = TRUE),
    autonomy_check_correct = mean(comp_autonomy_correct, na.rm = TRUE),
    both_checks_correct = mean(comp_both_correct, na.rm = TRUE),
    .groups = "drop"
  )

kable(manipulation_check_overall, digits = 3, caption = "Overall manipulation-check pass rates.")

Overall manipulation-check pass rates.

n	country_check_response_n	country_check_correct	autonomy_check_response_n	autonomy_check_correct	both_checks_correct
2935	2905	0.738	2935	0.808	0.66

kable(manipulation_check_by_condition, digits = 3, caption = "Manipulation-check pass rates by condition.")

Manipulation-check pass rates by condition.

country_condition	weapons_condition	n	country_check_correct	autonomy_check_correct	both_checks_correct
Chinese strike / U.S. victim	Autonomous AI	737	0.800	0.821	0.722
Chinese strike / U.S. victim	Human-operated	737	0.749	0.775	0.644
U.S. strike / Chinese victim	Autonomous AI	728	0.693	0.834	0.638
U.S. strike / Chinese victim	Human-operated	733	0.709	0.802	0.634

Exploratory Model Among Manipulation-Check Passers

prereg_pass_data <- prereg_primary_data %>%
  filter(comp_both_correct)

prereg_pass_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_pass_data
)

prereg_pass_vcov <- vcovHC(prereg_pass_model, type = "HC2")
prereg_pass_coeftest <- coeftest(prereg_pass_model, vcov. = prereg_pass_vcov)
prereg_pass_ci <- coefci(prereg_pass_model, vcov. = prereg_pass_vcov)

prereg_pass_results <- data.frame(
  term = rownames(prereg_pass_coeftest),
  estimate = prereg_pass_coeftest[, 1],
  hc2_se = prereg_pass_coeftest[, 2],
  statistic = prereg_pass_coeftest[, 3],
  p_value = prereg_pass_coeftest[, 4],
  ci_low = prereg_pass_ci[, 1],
  ci_high = prereg_pass_ci[, 2],
  row.names = NULL
)

kable(
  prereg_pass_results,
  digits = 3,
  caption = paste0("Exploratory primary model among manipulation-check passers only (N = ", nrow(prereg_pass_data), ").")
)

Exploratory primary model among manipulation-check passers only (N = 1880).

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
(Intercept)	70.515	3.189	22.113	0.000	64.261	76.769
country_scenario	11.415	1.555	7.339	0.000	8.365	14.466
autonomous_ai	-12.483	1.725	-7.235	0.000	-15.867	-9.100
ai_comfort_c	1.322	0.572	2.311	0.021	0.200	2.444
relative_china_us_ai_c	0.043	0.029	1.483	0.138	-0.014	0.099
pid_republican	1.735	1.195	1.452	0.147	-0.609	4.080
age_ordinal	-1.031	0.617	-1.671	0.095	-2.242	0.179
gender_female	-0.357	1.134	-0.315	0.753	-2.581	1.868
education_ordinal	-0.780	0.436	-1.788	0.074	-1.635	0.075
income_ordinal	-0.237	0.394	-0.602	0.548	-1.009	0.535
race_whiteWhite	-0.797	1.416	-0.563	0.573	-3.574	1.980
hispanicYes	3.558	1.800	1.977	0.048	0.028	7.089
metroMetropolitan	0.848	1.776	0.478	0.633	-2.635	4.331
country_scenario:autonomous_ai	-0.860	2.263	-0.380	0.704	-5.299	3.579

Secondary Outcome Models

The preregistration lists escalation likelihood, escalation justifiability, and responsibility attributions as exploratory dependent variables. The following models use the same treatment specification and preregistered covariates as the primary model.

Likelihood of Escalation

prereg_escalation_likelihood_data <- prereg_primary_data %>%
  filter(!is.na(escalation_likelihood))

prereg_escalation_likelihood_model <- lm(
  escalation_likelihood ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_escalation_likelihood_data
)

prereg_escalation_likelihood_vcov <- vcovHC(prereg_escalation_likelihood_model, type = "HC2")
prereg_escalation_likelihood_coeftest <- coeftest(prereg_escalation_likelihood_model, vcov. = prereg_escalation_likelihood_vcov)
prereg_escalation_likelihood_ci <- coefci(prereg_escalation_likelihood_model, vcov. = prereg_escalation_likelihood_vcov)

prereg_escalation_likelihood_results <- data.frame(
  term = rownames(prereg_escalation_likelihood_coeftest),
  estimate = prereg_escalation_likelihood_coeftest[, 1],
  hc2_se = prereg_escalation_likelihood_coeftest[, 2],
  statistic = prereg_escalation_likelihood_coeftest[, 3],
  p_value = prereg_escalation_likelihood_coeftest[, 4],
  ci_low = prereg_escalation_likelihood_ci[, 1],
  ci_high = prereg_escalation_likelihood_ci[, 2],
  row.names = NULL
)

kable(
  prereg_escalation_likelihood_results,
  digits = 3,
  caption = "Likelihood of escalation model with HC2 standard errors."
)

Likelihood of escalation model with HC2 standard errors.

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
(Intercept)	65.757	2.232	29.464	0.000	61.381	70.133
country_scenario	1.081	1.137	0.950	0.342	-1.149	3.310
autonomous_ai	0.741	1.099	0.674	0.500	-1.415	2.897
ai_comfort_c	0.593	0.412	1.442	0.149	-0.214	1.400
relative_china_us_ai_c	0.006	0.020	0.327	0.744	-0.032	0.045
pid_republican	-0.556	0.878	-0.633	0.527	-2.278	1.167
age_ordinal	-0.673	0.447	-1.507	0.132	-1.549	0.203
gender_female	-0.909	0.847	-1.074	0.283	-2.570	0.751
education_ordinal	0.005	0.320	0.017	0.986	-0.622	0.633
income_ordinal	0.179	0.282	0.634	0.526	-0.374	0.731
race_whiteWhite	-1.035	1.026	-1.008	0.313	-3.047	0.977
hispanicYes	-2.217	1.385	-1.600	0.110	-4.933	0.500
metroMetropolitan	0.510	1.281	0.399	0.690	-2.001	3.022
country_scenario:autonomous_ai	-1.644	1.636	-1.005	0.315	-4.852	1.563

Justifiability of Escalation

prereg_escalation_justifiability_data <- prereg_primary_data %>%
  filter(!is.na(escalation_justifiability))

prereg_escalation_justifiability_model <- lm(
  escalation_justifiability ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_escalation_justifiability_data
)

prereg_escalation_justifiability_vcov <- vcovHC(prereg_escalation_justifiability_model, type = "HC2")
prereg_escalation_justifiability_coeftest <- coeftest(prereg_escalation_justifiability_model, vcov. = prereg_escalation_justifiability_vcov)
prereg_escalation_justifiability_ci <- coefci(prereg_escalation_justifiability_model, vcov. = prereg_escalation_justifiability_vcov)

prereg_escalation_justifiability_results <- data.frame(
  term = rownames(prereg_escalation_justifiability_coeftest),
  estimate = prereg_escalation_justifiability_coeftest[, 1],
  hc2_se = prereg_escalation_justifiability_coeftest[, 2],
  statistic = prereg_escalation_justifiability_coeftest[, 3],
  p_value = prereg_escalation_justifiability_coeftest[, 4],
  ci_low = prereg_escalation_justifiability_ci[, 1],
  ci_high = prereg_escalation_justifiability_ci[, 2],
  row.names = NULL
)

kable(
  prereg_escalation_justifiability_results,
  digits = 3,
  caption = "Justifiability of escalation model with HC2 standard errors."
)

Justifiability of escalation model with HC2 standard errors.

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
(Intercept)	60.008	2.545	23.581	0.000	55.018	64.998
country_scenario	2.338	1.324	1.765	0.078	-0.259	4.934
autonomous_ai	-2.403	1.316	-1.827	0.068	-4.983	0.177
ai_comfort_c	1.684	0.492	3.425	0.001	0.720	2.649
relative_china_us_ai_c	0.012	0.023	0.520	0.603	-0.033	0.057
pid_republican	4.471	1.029	4.344	0.000	2.453	6.489
age_ordinal	0.234	0.529	0.443	0.658	-0.803	1.271
gender_female	-2.942	0.981	-2.998	0.003	-4.866	-1.018
education_ordinal	0.172	0.369	0.466	0.641	-0.551	0.895
income_ordinal	-0.045	0.329	-0.136	0.892	-0.689	0.600
race_whiteWhite	-2.575	1.123	-2.293	0.022	-4.776	-0.373
hispanicYes	1.205	1.460	0.825	0.410	-1.659	4.068
metroMetropolitan	1.977	1.515	1.305	0.192	-0.993	4.947
country_scenario:autonomous_ai	1.375	1.900	0.724	0.469	-2.350	5.100

prereg_secondary_treatment_results <- bind_rows(
  prereg_escalation_likelihood_results %>%
    filter(term %in% c("country_scenario", "autonomous_ai", "country_scenario:autonomous_ai")) %>%
    mutate(outcome = "Likelihood of escalation", .before = term),
  prereg_escalation_justifiability_results %>%
    filter(term %in% c("country_scenario", "autonomous_ai", "country_scenario:autonomous_ai")) %>%
    mutate(outcome = "Justifiability of escalation", .before = term)
) %>%
  mutate(
    term_label = case_when(
      term == "country_scenario" ~ "Chinese vs. U.S. strike\namong human-operated systems",
      term == "autonomous_ai" ~ "Autonomous effect\nin U.S.-strike condition",
      TRUE ~ "Country scenario x\nautonomous AI interaction"
    )
  )

prereg_secondary_treatment_results$term_label <- factor(
  prereg_secondary_treatment_results$term_label,
  levels = c(
    "Chinese vs. U.S. strike\namong human-operated systems",
    "Autonomous effect\nin U.S.-strike condition",
    "Country scenario x\nautonomous AI interaction"
  )
)

ggplot(prereg_secondary_treatment_results, aes(x = estimate, y = term_label)) +
  geom_vline(xintercept = 0, linewidth = 0.5, color = "gray55") +
  geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.14, linewidth = 0.8) +
  geom_point(size = 2.8) +
  facet_wrap(~ outcome) +
  scale_y_discrete(limits = rev(levels(prereg_secondary_treatment_results$term_label))) +
  labs(
    title = "Treatment Coefficients for Secondary Outcomes",
    x = "Estimated effect",
    y = NULL,
    caption = "Points are OLS estimates; error bars are 95% CIs using HC2 standard errors."
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.major.y = element_blank(),
    strip.text = element_text(face = "bold")
  )

## `height` was translated to `width`.

prereg_responsibility_data <- bind_rows(
  prereg_primary_data %>%
    transmute(country_scenario, autonomous_ai, ai_comfort_c, relative_china_us_ai_c, pid_republican,
              age_ordinal, gender_female, education_ordinal, income_ordinal, race_white, hispanic, metro,
              actor = "Senior political leaders", responsibility = as.numeric(resp_leader)),
  prereg_primary_data %>%
    transmute(country_scenario, autonomous_ai, ai_comfort_c, relative_china_us_ai_c, pid_republican,
              age_ordinal, gender_female, education_ordinal, income_ordinal, race_white, hispanic, metro,
              actor = "Military officials", responsibility = as.numeric(resp_military)),
  prereg_primary_data %>%
    transmute(country_scenario, autonomous_ai, ai_comfort_c, relative_china_us_ai_c, pid_republican,
              age_ordinal, gender_female, education_ordinal, income_ordinal, race_white, hispanic, metro,
              actor = "Human operator", responsibility = as.numeric(resp_pilot)),
  prereg_primary_data %>%
    transmute(country_scenario, autonomous_ai, ai_comfort_c, relative_china_us_ai_c, pid_republican,
              age_ordinal, gender_female, education_ordinal, income_ordinal, race_white, hispanic, metro,
              actor = "AI weapons system", responsibility = as.numeric(resp_ai))
) %>%
  filter(!is.na(responsibility))

responsibility_treatment_terms <- c(
  "country_scenario",
  "autonomous_ai",
  "country_scenario:autonomous_ai"
)

Responsibility: Senior Political Leaders

prereg_senior_political_leaders_data <- prereg_responsibility_data %>%
  filter(actor == "Senior political leaders")

prereg_senior_political_leaders_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_senior_political_leaders_data
)

prereg_senior_political_leaders_vcov <- vcovHC(prereg_senior_political_leaders_model, type = "HC2")
prereg_senior_political_leaders_coeftest <- coeftest(prereg_senior_political_leaders_model, vcov. = prereg_senior_political_leaders_vcov)
prereg_senior_political_leaders_ci <- coefci(prereg_senior_political_leaders_model, vcov. = prereg_senior_political_leaders_vcov)

prereg_senior_political_leaders_results <- data.frame(
  term = rownames(prereg_senior_political_leaders_coeftest),
  estimate = prereg_senior_political_leaders_coeftest[, 1],
  hc2_se = prereg_senior_political_leaders_coeftest[, 2],
  statistic = prereg_senior_political_leaders_coeftest[, 3],
  p_value = prereg_senior_political_leaders_coeftest[, 4],
  ci_low = prereg_senior_political_leaders_ci[, 1],
  ci_high = prereg_senior_political_leaders_ci[, 2],
  row.names = NULL
)

prereg_senior_political_leaders_treatment_results <- prereg_senior_political_leaders_results %>%
  filter(term %in% responsibility_treatment_terms) %>%
  mutate(actor = "Senior political leaders", .before = term)

kable(
  prereg_senior_political_leaders_treatment_results,
  digits = 3,
  caption = "Treatment coefficients for responsibility attributed to senior political leaders."
)

Treatment coefficients for responsibility attributed to senior political leaders.

actor	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
Senior political leaders	country_scenario	4.968	1.306	3.804	0.000	2.407	7.529
Senior political leaders	autonomous_ai	1.514	1.355	1.117	0.264	-1.144	4.171
Senior political leaders	country_scenario:autonomous_ai	-1.419	1.862	-0.762	0.446	-5.070	2.232

Responsibility: Military Officials

prereg_military_officials_data <- prereg_responsibility_data %>%
  filter(actor == "Military officials")

prereg_military_officials_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_military_officials_data
)

prereg_military_officials_vcov <- vcovHC(prereg_military_officials_model, type = "HC2")
prereg_military_officials_coeftest <- coeftest(prereg_military_officials_model, vcov. = prereg_military_officials_vcov)
prereg_military_officials_ci <- coefci(prereg_military_officials_model, vcov. = prereg_military_officials_vcov)

prereg_military_officials_results <- data.frame(
  term = rownames(prereg_military_officials_coeftest),
  estimate = prereg_military_officials_coeftest[, 1],
  hc2_se = prereg_military_officials_coeftest[, 2],
  statistic = prereg_military_officials_coeftest[, 3],
  p_value = prereg_military_officials_coeftest[, 4],
  ci_low = prereg_military_officials_ci[, 1],
  ci_high = prereg_military_officials_ci[, 2],
  row.names = NULL
)

prereg_military_officials_treatment_results <- prereg_military_officials_results %>%
  filter(term %in% responsibility_treatment_terms) %>%
  mutate(actor = "Military officials", .before = term)

kable(
  prereg_military_officials_treatment_results,
  digits = 3,
  caption = "Treatment coefficients for responsibility attributed to military officials."
)

Treatment coefficients for responsibility attributed to military officials.

actor	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
Military officials	country_scenario	0.801	1.209	0.663	0.507	-1.569	3.171
Military officials	autonomous_ai	0.462	1.221	0.378	0.705	-1.933	2.856
Military officials	country_scenario:autonomous_ai	0.550	1.725	0.319	0.750	-2.833	3.933

Responsibility: Human Operator

prereg_human_operator_data <- prereg_responsibility_data %>%
  filter(actor == "Human operator")

prereg_human_operator_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_human_operator_data
)

prereg_human_operator_vcov <- vcovHC(prereg_human_operator_model, type = "HC2")
prereg_human_operator_coeftest <- coeftest(prereg_human_operator_model, vcov. = prereg_human_operator_vcov)
prereg_human_operator_ci <- coefci(prereg_human_operator_model, vcov. = prereg_human_operator_vcov)

prereg_human_operator_results <- data.frame(
  term = rownames(prereg_human_operator_coeftest),
  estimate = prereg_human_operator_coeftest[, 1],
  hc2_se = prereg_human_operator_coeftest[, 2],
  statistic = prereg_human_operator_coeftest[, 3],
  p_value = prereg_human_operator_coeftest[, 4],
  ci_low = prereg_human_operator_ci[, 1],
  ci_high = prereg_human_operator_ci[, 2],
  row.names = NULL
)

prereg_human_operator_treatment_results <- prereg_human_operator_results %>%
  filter(term %in% responsibility_treatment_terms) %>%
  mutate(actor = "Human operator", .before = term)

kable(
  prereg_human_operator_treatment_results,
  digits = 3,
  caption = "Treatment coefficients for responsibility attributed to the human operator."
)

Treatment coefficients for responsibility attributed to the human operator.

actor	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
Human operator	country_scenario	1.68	1.416	1.186	0.236	-1.098	4.458

Responsibility: AI Weapons System

prereg_ai_weapons_system_data <- prereg_responsibility_data %>%
  filter(actor == "AI weapons system")

prereg_ai_weapons_system_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_ai_weapons_system_data
)

prereg_ai_weapons_system_vcov <- vcovHC(prereg_ai_weapons_system_model, type = "HC2")
prereg_ai_weapons_system_coeftest <- coeftest(prereg_ai_weapons_system_model, vcov. = prereg_ai_weapons_system_vcov)
prereg_ai_weapons_system_ci <- coefci(prereg_ai_weapons_system_model, vcov. = prereg_ai_weapons_system_vcov)

prereg_ai_weapons_system_results <- data.frame(
  term = rownames(prereg_ai_weapons_system_coeftest),
  estimate = prereg_ai_weapons_system_coeftest[, 1],
  hc2_se = prereg_ai_weapons_system_coeftest[, 2],
  statistic = prereg_ai_weapons_system_coeftest[, 3],
  p_value = prereg_ai_weapons_system_coeftest[, 4],
  ci_low = prereg_ai_weapons_system_ci[, 1],
  ci_high = prereg_ai_weapons_system_ci[, 2],
  row.names = NULL
)

prereg_ai_weapons_system_treatment_results <- prereg_ai_weapons_system_results %>%
  filter(term %in% responsibility_treatment_terms) %>%
  mutate(actor = "AI weapons system", .before = term)

kable(
  prereg_ai_weapons_system_treatment_results,
  digits = 3,
  caption = "Treatment coefficients for responsibility attributed to the AI weapons system."
)

Treatment coefficients for responsibility attributed to the AI weapons system.

actor	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
AI weapons system	country_scenario	-1.339	1.66	-0.807	0.42	-4.595	1.917

prereg_responsibility_treatment_results <- bind_rows(
  prereg_senior_political_leaders_treatment_results,
  prereg_military_officials_treatment_results,
  prereg_human_operator_treatment_results,
  prereg_ai_weapons_system_treatment_results
) %>%
  mutate(
    term_label = case_when(
      term == "country_scenario" ~ "Chinese vs. U.S. strike",
      term == "autonomous_ai" ~ "Autonomous vs. human",
      TRUE ~ "Interaction"
    )
  )

prereg_responsibility_treatment_results$term_label <- factor(
  prereg_responsibility_treatment_results$term_label,
  levels = c("Chinese vs. U.S. strike", "Autonomous vs. human", "Interaction")
)

ggplot(prereg_responsibility_treatment_results, aes(x = estimate, y = term_label)) +
  geom_vline(xintercept = 0, linewidth = 0.5, color = "gray55") +
  geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.14, linewidth = 0.8) +
  geom_point(size = 2.8) +
  facet_wrap(~ actor, ncol = 2) +
  scale_y_discrete(limits = rev(levels(prereg_responsibility_treatment_results$term_label))) +
  labs(
    title = "Treatment Coefficients for Responsibility Attributions",
    x = "Estimated effect on perceived responsibility",
    y = NULL,
    caption = "Error bars are 95% CIs using HC2 standard errors. Human-operator and AI-system responsibility were only asked in their corresponding weapons-system conditions."
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.major.y = element_blank(),
    strip.text = element_text(face = "bold")
  )

## `height` was translated to `width`.

Exploratory Responsibility Profiles by Actor

Coding note: senior political leaders and military officials are observed in all four treatment cells. Human-operator responsibility is only observed in the human-operated condition, and AI-system responsibility is only observed in the autonomous-AI condition. For that reason, this plot compares responsibility levels descriptively within each actor panel rather than treating all actor comparisons as if they were asked of the same respondents in every condition.

responsibility_profile_data <- bind_rows(
  prereg_analysis_data %>%
    transmute(country_condition, weapons_condition, actor = "Senior political leaders", responsibility = as.numeric(resp_leader)),
  prereg_analysis_data %>%
    transmute(country_condition, weapons_condition, actor = "Military officials", responsibility = as.numeric(resp_military)),
  prereg_analysis_data %>%
    transmute(country_condition, weapons_condition, actor = "Human operator", responsibility = as.numeric(resp_pilot)),
  prereg_analysis_data %>%
    transmute(country_condition, weapons_condition, actor = "AI weapons system", responsibility = as.numeric(resp_ai))
) %>%
  filter(!is.na(responsibility)) %>%
  mutate(
    country_condition = factor(
      country_condition,
      levels = c("Chinese strike / U.S. victim", "U.S. strike / Chinese victim"),
      labels = c("Chinese strike", "U.S. strike")
    ),
    weapons_condition = factor(weapons_condition, levels = c("Human-operated", "Autonomous AI")),
    actor = factor(
      actor,
      levels = c("Senior political leaders", "Military officials", "Human operator", "AI weapons system")
    )
  )

responsibility_profile_summary <- responsibility_profile_data %>%
  group_by(actor, country_condition, weapons_condition) %>%
  summarize(
    n = n(),
    mean = mean(responsibility),
    se = sd(responsibility) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

kable(
  responsibility_profile_summary,
  digits = 2,
  caption = "Descriptive responsibility means and 95% confidence intervals by actor and experimental condition."
)

Descriptive responsibility means and 95% confidence intervals by actor and experimental condition.

actor	country_condition	weapons_condition	n	mean	se	ci_low	ci_high
Senior political leaders	Chinese strike	Human-operated	728	74.64	0.87	72.94	76.34
Senior political leaders	Chinese strike	Autonomous AI	729	74.83	0.91	73.04	76.62
Senior political leaders	U.S. strike	Human-operated	720	69.63	0.98	67.72	71.55
Senior political leaders	U.S. strike	Autonomous AI	720	71.28	1.00	69.32	73.23
Military officials	Chinese strike	Human-operated	728	74.82	0.86	73.14	76.50
Military officials	Chinese strike	Autonomous AI	728	76.09	0.87	74.38	77.79
Military officials	U.S. strike	Human-operated	721	74.14	0.85	72.47	75.81
Military officials	U.S. strike	Autonomous AI	722	74.45	0.90	72.68	76.22
Human operator	Chinese strike	Human-operated	728	70.74	0.99	68.80	72.68
Human operator	U.S. strike	Human-operated	718	68.67	0.98	66.75	70.60
AI weapons system	Chinese strike	Autonomous AI	723	65.41	1.17	63.12	67.70
AI weapons system	U.S. strike	Autonomous AI	712	66.94	1.16	64.67	69.21

ggplot(
  responsibility_profile_summary,
  aes(x = weapons_condition, y = mean, color = country_condition, group = country_condition)
) +
  geom_line(linewidth = 0.9, alpha = 0.45, na.rm = TRUE) +
  geom_point(size = 2.8) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.7) +
  facet_wrap(~ actor, ncol = 2) +
  scale_color_manual(
    values = c("Chinese strike" = "#A63A33", "U.S. strike" = "#235A8C"),
    name = "Country scenario"
  ) +
  coord_cartesian(ylim = c(50, 85)) +
  labs(
    title = "Responsibility Profiles by Actor and Condition",
    x = "Weapons-system condition",
    y = "Perceived responsibility",
    caption = "Error bars are 95% confidence intervals. Human-operator and AI-system items are condition-specific."
  ) +
  theme_minimal(base_size = 12) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank(),
    strip.text = element_text(face = "bold")
  )

## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?

Exploratory Moderation Models

These models extend the primary specification by interacting the country-scenario indicator, autonomous-AI indicator, and moderator. Continuous moderators are demeaned. Models include Verasight demographic controls; moderation models not focused on party identification also control for pid_republican.

moderation_ai_comfort_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * ai_comfort_c +
    pid_republican + age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_ai_comfort_vcov <- vcovHC(moderation_ai_comfort_model, type = "HC2")
moderation_ai_comfort_coeftest <- coeftest(moderation_ai_comfort_model, vcov. = moderation_ai_comfort_vcov)
moderation_ai_comfort_ci <- coefci(moderation_ai_comfort_model, vcov. = moderation_ai_comfort_vcov)

moderation_ai_comfort_results <- data.frame(
  term = rownames(moderation_ai_comfort_coeftest),
  estimate = moderation_ai_comfort_coeftest[, 1],
  hc2_se = moderation_ai_comfort_coeftest[, 2],
  statistic = moderation_ai_comfort_coeftest[, 3],
  p_value = moderation_ai_comfort_coeftest[, 4],
  ci_low = moderation_ai_comfort_ci[, 1],
  ci_high = moderation_ai_comfort_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:ai_comfort_c") %>%
  mutate(moderator = "AI comfort", .before = term)

moderation_us_ai_effectiveness_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * us_ai_effectiveness_c +
    pid_republican + age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_us_ai_effectiveness_vcov <- vcovHC(moderation_us_ai_effectiveness_model, type = "HC2")
moderation_us_ai_effectiveness_coeftest <- coeftest(moderation_us_ai_effectiveness_model, vcov. = moderation_us_ai_effectiveness_vcov)
moderation_us_ai_effectiveness_ci <- coefci(moderation_us_ai_effectiveness_model, vcov. = moderation_us_ai_effectiveness_vcov)

moderation_us_ai_effectiveness_results <- data.frame(
  term = rownames(moderation_us_ai_effectiveness_coeftest),
  estimate = moderation_us_ai_effectiveness_coeftest[, 1],
  hc2_se = moderation_us_ai_effectiveness_coeftest[, 2],
  statistic = moderation_us_ai_effectiveness_coeftest[, 3],
  p_value = moderation_us_ai_effectiveness_coeftest[, 4],
  ci_low = moderation_us_ai_effectiveness_ci[, 1],
  ci_high = moderation_us_ai_effectiveness_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:us_ai_effectiveness_c") %>%
  mutate(moderator = "U.S. AI effectiveness", .before = term)

moderation_china_ai_effectiveness_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * china_ai_effectiveness_c +
    pid_republican + age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_china_ai_effectiveness_vcov <- vcovHC(moderation_china_ai_effectiveness_model, type = "HC2")
moderation_china_ai_effectiveness_coeftest <- coeftest(moderation_china_ai_effectiveness_model, vcov. = moderation_china_ai_effectiveness_vcov)
moderation_china_ai_effectiveness_ci <- coefci(moderation_china_ai_effectiveness_model, vcov. = moderation_china_ai_effectiveness_vcov)

moderation_china_ai_effectiveness_results <- data.frame(
  term = rownames(moderation_china_ai_effectiveness_coeftest),
  estimate = moderation_china_ai_effectiveness_coeftest[, 1],
  hc2_se = moderation_china_ai_effectiveness_coeftest[, 2],
  statistic = moderation_china_ai_effectiveness_coeftest[, 3],
  p_value = moderation_china_ai_effectiveness_coeftest[, 4],
  ci_low = moderation_china_ai_effectiveness_ci[, 1],
  ci_high = moderation_china_ai_effectiveness_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:china_ai_effectiveness_c") %>%
  mutate(moderator = "Chinese AI effectiveness", .before = term)

moderation_southkorea_ai_effectiveness_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * southkorea_ai_effectiveness_c +
    pid_republican + age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_southkorea_ai_effectiveness_vcov <- vcovHC(moderation_southkorea_ai_effectiveness_model, type = "HC2")
moderation_southkorea_ai_effectiveness_coeftest <- coeftest(moderation_southkorea_ai_effectiveness_model, vcov. = moderation_southkorea_ai_effectiveness_vcov)
moderation_southkorea_ai_effectiveness_ci <- coefci(moderation_southkorea_ai_effectiveness_model, vcov. = moderation_southkorea_ai_effectiveness_vcov)

moderation_southkorea_ai_effectiveness_results <- data.frame(
  term = rownames(moderation_southkorea_ai_effectiveness_coeftest),
  estimate = moderation_southkorea_ai_effectiveness_coeftest[, 1],
  hc2_se = moderation_southkorea_ai_effectiveness_coeftest[, 2],
  statistic = moderation_southkorea_ai_effectiveness_coeftest[, 3],
  p_value = moderation_southkorea_ai_effectiveness_coeftest[, 4],
  ci_low = moderation_southkorea_ai_effectiveness_ci[, 1],
  ci_high = moderation_southkorea_ai_effectiveness_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:southkorea_ai_effectiveness_c") %>%
  mutate(moderator = "South Korean AI effectiveness", .before = term)

moderation_relative_china_us_ai_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * relative_china_us_ai_c +
    pid_republican + age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_relative_china_us_ai_vcov <- vcovHC(moderation_relative_china_us_ai_model, type = "HC2")
moderation_relative_china_us_ai_coeftest <- coeftest(moderation_relative_china_us_ai_model, vcov. = moderation_relative_china_us_ai_vcov)
moderation_relative_china_us_ai_ci <- coefci(moderation_relative_china_us_ai_model, vcov. = moderation_relative_china_us_ai_vcov)

moderation_relative_china_us_ai_results <- data.frame(
  term = rownames(moderation_relative_china_us_ai_coeftest),
  estimate = moderation_relative_china_us_ai_coeftest[, 1],
  hc2_se = moderation_relative_china_us_ai_coeftest[, 2],
  statistic = moderation_relative_china_us_ai_coeftest[, 3],
  p_value = moderation_relative_china_us_ai_coeftest[, 4],
  ci_low = moderation_relative_china_us_ai_ci[, 1],
  ci_high = moderation_relative_china_us_ai_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:relative_china_us_ai_c") %>%
  mutate(moderator = "Chinese minus U.S. AI effectiveness", .before = term)

moderation_pid_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * pid_republican +
    ai_comfort_c + relative_china_us_ai_c +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = prereg_primary_data
)

moderation_pid_vcov <- vcovHC(moderation_pid_model, type = "HC2")
moderation_pid_coeftest <- coeftest(moderation_pid_model, vcov. = moderation_pid_vcov)
moderation_pid_ci <- coefci(moderation_pid_model, vcov. = moderation_pid_vcov)

moderation_pid_results <- data.frame(
  term = rownames(moderation_pid_coeftest),
  estimate = moderation_pid_coeftest[, 1],
  hc2_se = moderation_pid_coeftest[, 2],
  statistic = moderation_pid_coeftest[, 3],
  p_value = moderation_pid_coeftest[, 4],
  ci_low = moderation_pid_ci[, 1],
  ci_high = moderation_pid_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "country_scenario:autonomous_ai:pid_republican") %>%
  mutate(moderator = "Party identification", .before = term)

moderation_results <- bind_rows(
  moderation_ai_comfort_results,
  moderation_us_ai_effectiveness_results,
  moderation_china_ai_effectiveness_results,
  moderation_southkorea_ai_effectiveness_results,
  moderation_relative_china_us_ai_results,
  moderation_pid_results
)

kable(
  moderation_results,
  digits = 3,
  caption = "Focal three-way interaction terms from exploratory moderation models."
)

Focal three-way interaction terms from exploratory moderation models.

moderator	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
AI comfort	country_scenario:autonomous_ai:ai_comfort_c	4.428	1.773	2.498	0.013	0.952	7.904
U.S. AI effectiveness	country_scenario:autonomous_ai:us_ai_effectiveness_c	0.086	0.081	1.063	0.288	-0.073	0.245
Chinese AI effectiveness	country_scenario:autonomous_ai:china_ai_effectiveness_c	0.231	0.077	3.000	0.003	0.080	0.382
South Korean AI effectiveness	country_scenario:autonomous_ai:southkorea_ai_effectiveness_c	0.223	0.083	2.677	0.007	0.060	0.386
Chinese minus U.S. AI effectiveness	country_scenario:autonomous_ai:relative_china_us_ai_c	0.166	0.080	2.084	0.037	0.010	0.322
Party identification	country_scenario:autonomous_ai:pid_republican	-1.276	3.728	-0.342	0.732	-8.585	6.033

if (nrow(moderation_results) > 0) {
  moderation_plot_data <- moderation_results %>%
    mutate(
      moderator_term = ifelse(
        term == "country_scenario:autonomous_ai:pid_republican",
        "Republican vs. Democrat/Independent",
        moderator
      ),
      moderator_term = factor(moderator_term, levels = rev(unique(moderator_term)))
    )

  ggplot(moderation_plot_data, aes(x = estimate, y = moderator_term)) +
    geom_vline(xintercept = 0, linewidth = 0.5, color = "gray55") +
    geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.14, linewidth = 0.8) +
    geom_point(size = 2.8) +
    labs(
      title = "Exploratory Moderation of the Treatment Interaction",
      x = "Estimated three-way interaction",
      y = NULL,
      caption = "Points are focal three-way interaction estimates; error bars are 95% CIs using HC2 standard errors."
    ) +
    theme_minimal(base_size = 13) +
    theme(
      plot.title = element_text(face = "bold", hjust = 0.5),
      panel.grid.major.y = element_blank()
    )
}

## `height` was translated to `width`.

Exploratory Partisan Polarization

This exploratory analysis uses the three-category party identification variable to show whether perceived intentionality differs across Democrats, Independents, and Republicans within each experimental cell. The model is exploratory because the primary preregistered specification uses the binary pid_republican control rather than estimating separate treatment patterns for each party group.

polarization_data <- prereg_primary_data %>%
  mutate(
    pid_three_category = factor(pid, levels = c("Democrat", "Independent", "Republican")),
    country_condition = factor(
      country_condition,
      levels = c("Chinese strike / U.S. victim", "U.S. strike / Chinese victim"),
      labels = c("Chinese strike", "U.S. strike")
    ),
    weapons_condition = factor(weapons_condition, levels = c("Human-operated", "Autonomous AI"))
  ) %>%
  filter(!is.na(pid_three_category))

polarization_model <- lm(
  perceived_intentionality ~ country_scenario * autonomous_ai * pid_three_category +
    ai_comfort_c + relative_china_us_ai_c +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = polarization_data
)

polarization_vcov <- vcovHC(polarization_model, type = "HC2")
polarization_coeftest <- coeftest(polarization_model, vcov. = polarization_vcov)
polarization_ci <- coefci(polarization_model, vcov. = polarization_vcov)

polarization_results <- data.frame(
  term = rownames(polarization_coeftest),
  estimate = polarization_coeftest[, 1],
  hc2_se = polarization_coeftest[, 2],
  statistic = polarization_coeftest[, 3],
  p_value = polarization_coeftest[, 4],
  ci_low = polarization_ci[, 1],
  ci_high = polarization_ci[, 2],
  row.names = NULL
) %>%
  filter(grepl("pid_three_category", term))

polarization_summary <- polarization_data %>%
  group_by(pid_three_category, country_condition, weapons_condition) %>%
  summarize(
    n = n(),
    mean = mean(perceived_intentionality),
    se = sd(perceived_intentionality) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

kable(
  polarization_results,
  digits = 3,
  caption = "Exploratory party-identification terms from a three-category partisan moderation model with HC2 standard errors."
)

Exploratory party-identification terms from a three-category partisan moderation model with HC2 standard errors.

term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
pid_three_categoryIndependent	-0.127	2.769	-0.046	0.963	-5.556	5.302
pid_three_categoryRepublican	-4.051	2.003	-2.022	0.043	-7.978	-0.123
country_scenario:pid_three_categoryIndependent	-5.558	3.791	-1.466	0.143	-12.992	1.876
country_scenario:pid_three_categoryRepublican	7.589	2.733	2.776	0.006	2.229	12.948
autonomous_ai:pid_three_categoryIndependent	-2.382	4.023	-0.592	0.554	-10.271	5.506
autonomous_ai:pid_three_categoryRepublican	3.263	2.849	1.145	0.252	-2.324	8.850
country_scenario:autonomous_ai:pid_three_categoryIndependent	11.439	5.418	2.111	0.035	0.815	22.062
country_scenario:autonomous_ai:pid_three_categoryRepublican	1.304	3.915	0.333	0.739	-6.373	8.981

ggplot(
  polarization_summary,
  aes(x = weapons_condition, y = mean, color = country_condition, group = country_condition)
) +
  geom_line(linewidth = 0.9, alpha = 0.45) +
  geom_point(size = 2.8) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.7) +
  facet_wrap(~ pid_three_category) +
  scale_color_manual(
    values = c("Chinese strike" = "#A63A33", "U.S. strike" = "#235A8C"),
    name = "Country scenario"
  ) +
  coord_cartesian(ylim = c(45, 85)) +
  labs(
    title = "Perceived Intentionality by Party Identification",
    x = "Weapons-system condition",
    y = "Perceived intentionality",
    caption = "Exploratory descriptive means by party group. Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 12) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank(),
    strip.text = element_text(face = "bold")
  )

Exploratory Downstream Association Models

The preregistration describes mediation-style analyses as exploratory and associational. The following models test whether perceived intentionality is associated with downstream escalation and responsibility judgments, controlling for treatment assignment, AI comfort, relative China-U.S. AI effectiveness, Republican identification, and Verasight demographic controls.

prereg_association_data <- prereg_primary_data %>%
  mutate(perceived_intentionality_c = perceived_intentionality - mean(perceived_intentionality, na.rm = TRUE))

downstream_escalation_likelihood_data <- prereg_association_data %>%
  filter(!is.na(escalation_likelihood))

downstream_escalation_likelihood_model <- lm(
  escalation_likelihood ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = downstream_escalation_likelihood_data
)

downstream_escalation_likelihood_vcov <- vcovHC(downstream_escalation_likelihood_model, type = "HC2")
downstream_escalation_likelihood_coeftest <- coeftest(downstream_escalation_likelihood_model, vcov. = downstream_escalation_likelihood_vcov)
downstream_escalation_likelihood_ci <- coefci(downstream_escalation_likelihood_model, vcov. = downstream_escalation_likelihood_vcov)

downstream_escalation_likelihood_results <- data.frame(
  term = rownames(downstream_escalation_likelihood_coeftest),
  estimate = downstream_escalation_likelihood_coeftest[, 1],
  hc2_se = downstream_escalation_likelihood_coeftest[, 2],
  statistic = downstream_escalation_likelihood_coeftest[, 3],
  p_value = downstream_escalation_likelihood_coeftest[, 4],
  ci_low = downstream_escalation_likelihood_ci[, 1],
  ci_high = downstream_escalation_likelihood_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Likelihood of escalation", .before = term)

downstream_escalation_justifiability_data <- prereg_association_data %>%
  filter(!is.na(escalation_justifiability))

downstream_escalation_justifiability_model <- lm(
  escalation_justifiability ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = downstream_escalation_justifiability_data
)

downstream_escalation_justifiability_vcov <- vcovHC(downstream_escalation_justifiability_model, type = "HC2")
downstream_escalation_justifiability_coeftest <- coeftest(downstream_escalation_justifiability_model, vcov. = downstream_escalation_justifiability_vcov)
downstream_escalation_justifiability_ci <- coefci(downstream_escalation_justifiability_model, vcov. = downstream_escalation_justifiability_vcov)

downstream_escalation_justifiability_results <- data.frame(
  term = rownames(downstream_escalation_justifiability_coeftest),
  estimate = downstream_escalation_justifiability_coeftest[, 1],
  hc2_se = downstream_escalation_justifiability_coeftest[, 2],
  statistic = downstream_escalation_justifiability_coeftest[, 3],
  p_value = downstream_escalation_justifiability_coeftest[, 4],
  ci_low = downstream_escalation_justifiability_ci[, 1],
  ci_high = downstream_escalation_justifiability_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Justifiability of escalation", .before = term)

downstream_results <- bind_rows(
  downstream_escalation_likelihood_results,
  downstream_escalation_justifiability_results
)

prereg_responsibility_association_data <- bind_rows(
  prereg_association_data %>%
    mutate(actor = "Senior political leaders", responsibility = as.numeric(resp_leader)),
  prereg_association_data %>%
    mutate(actor = "Military officials", responsibility = as.numeric(resp_military)),
  prereg_association_data %>%
    mutate(actor = "Human operator", responsibility = as.numeric(resp_pilot)),
  prereg_association_data %>%
    mutate(actor = "AI weapons system", responsibility = as.numeric(resp_ai))
) %>%
  filter(!is.na(responsibility))

responsibility_association_senior_political_leaders_data <- prereg_responsibility_association_data %>%
  filter(actor == "Senior political leaders",
         !is.na(perceived_intentionality_c))

responsibility_association_senior_political_leaders_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = responsibility_association_senior_political_leaders_data
)

responsibility_association_senior_political_leaders_vcov <- vcovHC(responsibility_association_senior_political_leaders_model, type = "HC2")
responsibility_association_senior_political_leaders_coeftest <- coeftest(responsibility_association_senior_political_leaders_model, vcov. = responsibility_association_senior_political_leaders_vcov)
responsibility_association_senior_political_leaders_ci <- coefci(responsibility_association_senior_political_leaders_model, vcov. = responsibility_association_senior_political_leaders_vcov)

responsibility_association_senior_political_leaders_results <- data.frame(
  term = rownames(responsibility_association_senior_political_leaders_coeftest),
  estimate = responsibility_association_senior_political_leaders_coeftest[, 1],
  hc2_se = responsibility_association_senior_political_leaders_coeftest[, 2],
  statistic = responsibility_association_senior_political_leaders_coeftest[, 3],
  p_value = responsibility_association_senior_political_leaders_coeftest[, 4],
  ci_low = responsibility_association_senior_political_leaders_ci[, 1],
  ci_high = responsibility_association_senior_political_leaders_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Responsibility: Senior political leaders", .before = term)

responsibility_association_military_officials_data <- prereg_responsibility_association_data %>%
  filter(actor == "Military officials",
         !is.na(perceived_intentionality_c))

responsibility_association_military_officials_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = responsibility_association_military_officials_data
)

responsibility_association_military_officials_vcov <- vcovHC(responsibility_association_military_officials_model, type = "HC2")
responsibility_association_military_officials_coeftest <- coeftest(responsibility_association_military_officials_model, vcov. = responsibility_association_military_officials_vcov)
responsibility_association_military_officials_ci <- coefci(responsibility_association_military_officials_model, vcov. = responsibility_association_military_officials_vcov)

responsibility_association_military_officials_results <- data.frame(
  term = rownames(responsibility_association_military_officials_coeftest),
  estimate = responsibility_association_military_officials_coeftest[, 1],
  hc2_se = responsibility_association_military_officials_coeftest[, 2],
  statistic = responsibility_association_military_officials_coeftest[, 3],
  p_value = responsibility_association_military_officials_coeftest[, 4],
  ci_low = responsibility_association_military_officials_ci[, 1],
  ci_high = responsibility_association_military_officials_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Responsibility: Military officials", .before = term)

responsibility_association_human_operator_data <- prereg_responsibility_association_data %>%
  filter(actor == "Human operator",
         !is.na(perceived_intentionality_c))

responsibility_association_human_operator_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = responsibility_association_human_operator_data
)

responsibility_association_human_operator_vcov <- vcovHC(responsibility_association_human_operator_model, type = "HC2")
responsibility_association_human_operator_coeftest <- coeftest(responsibility_association_human_operator_model, vcov. = responsibility_association_human_operator_vcov)
responsibility_association_human_operator_ci <- coefci(responsibility_association_human_operator_model, vcov. = responsibility_association_human_operator_vcov)

responsibility_association_human_operator_results <- data.frame(
  term = rownames(responsibility_association_human_operator_coeftest),
  estimate = responsibility_association_human_operator_coeftest[, 1],
  hc2_se = responsibility_association_human_operator_coeftest[, 2],
  statistic = responsibility_association_human_operator_coeftest[, 3],
  p_value = responsibility_association_human_operator_coeftest[, 4],
  ci_low = responsibility_association_human_operator_ci[, 1],
  ci_high = responsibility_association_human_operator_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Responsibility: Human operator", .before = term)

responsibility_association_ai_weapons_system_data <- prereg_responsibility_association_data %>%
  filter(actor == "AI weapons system",
         !is.na(perceived_intentionality_c))

responsibility_association_ai_weapons_system_model <- lm(
  responsibility ~ country_scenario * autonomous_ai +
    perceived_intentionality_c + ai_comfort_c + relative_china_us_ai_c + pid_republican +
    age_ordinal + gender_female + education_ordinal + income_ordinal +
    race_white + hispanic + metro,
  data = responsibility_association_ai_weapons_system_data
)

responsibility_association_ai_weapons_system_vcov <- vcovHC(responsibility_association_ai_weapons_system_model, type = "HC2")
responsibility_association_ai_weapons_system_coeftest <- coeftest(responsibility_association_ai_weapons_system_model, vcov. = responsibility_association_ai_weapons_system_vcov)
responsibility_association_ai_weapons_system_ci <- coefci(responsibility_association_ai_weapons_system_model, vcov. = responsibility_association_ai_weapons_system_vcov)

responsibility_association_ai_weapons_system_results <- data.frame(
  term = rownames(responsibility_association_ai_weapons_system_coeftest),
  estimate = responsibility_association_ai_weapons_system_coeftest[, 1],
  hc2_se = responsibility_association_ai_weapons_system_coeftest[, 2],
  statistic = responsibility_association_ai_weapons_system_coeftest[, 3],
  p_value = responsibility_association_ai_weapons_system_coeftest[, 4],
  ci_low = responsibility_association_ai_weapons_system_ci[, 1],
  ci_high = responsibility_association_ai_weapons_system_ci[, 2],
  row.names = NULL
) %>%
  filter(term == "perceived_intentionality_c") %>%
  mutate(outcome = "Responsibility: AI weapons system", .before = term)

responsibility_association_results <- bind_rows(
  responsibility_association_senior_political_leaders_results,
  responsibility_association_military_officials_results,
  responsibility_association_human_operator_results,
  responsibility_association_ai_weapons_system_results
)

kable(
  bind_rows(downstream_results, responsibility_association_results),
  digits = 3,
  caption = "Association between perceived intentionality and downstream judgments using HC2 standard errors."
)

Association between perceived intentionality and downstream judgments using HC2 standard errors.

outcome	term	estimate	hc2_se	statistic	p_value	ci_low	ci_high
Likelihood of escalation	perceived_intentionality_c	0.323	0.019	17.370	0.000	0.286	0.359
Justifiability of escalation	perceived_intentionality_c	0.405	0.021	18.872	0.000	0.363	0.448
Responsibility: Senior political leaders	perceived_intentionality_c	0.384	0.021	18.338	0.000	0.343	0.425
Responsibility: Military officials	perceived_intentionality_c	0.273	0.020	13.330	0.000	0.233	0.313
Responsibility: Human operator	perceived_intentionality_c	0.143	0.033	4.351	0.000	0.079	0.208
Responsibility: AI weapons system	perceived_intentionality_c	-0.012	0.037	-0.313	0.754	-0.085	0.062

prereg_downstream_plot_data <- bind_rows(downstream_results, responsibility_association_results) %>%
  mutate(outcome = factor(outcome, levels = rev(outcome)))

ggplot(prereg_downstream_plot_data, aes(x = estimate, y = outcome)) +
  geom_vline(xintercept = 0, linewidth = 0.5, color = "gray55") +
  geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.14, linewidth = 0.8) +
  geom_point(size = 2.8) +
  labs(
    title = "Association Between Intentionality and Downstream Judgments",
    x = "Estimated association with perceived intentionality",
    y = NULL,
    caption = "Associational models control for treatment assignment, preregistered covariates, and demographic controls. Error bars are 95% CIs using HC2 standard errors."
  ) +
  theme_minimal(base_size = 13) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.major.y = element_blank()
  )

## `height` was translated to `width`.

Exploratory Distributional Checks for Intentionality

The primary analyses focus on average treatment effects. These plots look at the full distribution of perceived intentionality by experimental condition. This helps diagnose whether differences in means reflect broad shifts across respondents or changes concentrated at particular parts of the 0-to-100 scale.

intentionality_distribution_data <- prereg_analysis_data %>%
  mutate(
    country_condition = factor(
      country_condition,
      levels = c("Chinese strike / U.S. victim", "U.S. strike / Chinese victim"),
      labels = c("Chinese strike", "U.S. strike")
    ),
    weapons_condition = factor(weapons_condition, levels = c("Human-operated", "Autonomous AI"))
  )

intentionality_quantiles <- intentionality_distribution_data %>%
  group_by(country_condition, weapons_condition) %>%
  summarize(
    n = n(),
    mean = mean(perceived_intentionality, na.rm = TRUE),
    p10 = quantile(perceived_intentionality, 0.10, na.rm = TRUE),
    p25 = quantile(perceived_intentionality, 0.25, na.rm = TRUE),
    median = median(perceived_intentionality, na.rm = TRUE),
    p75 = quantile(perceived_intentionality, 0.75, na.rm = TRUE),
    p90 = quantile(perceived_intentionality, 0.90, na.rm = TRUE),
    .groups = "drop"
  )

kable(
  intentionality_quantiles,
  digits = 2,
  caption = "Distributional summaries for perceived intentionality by experimental cell."
)

Distributional summaries for perceived intentionality by experimental cell.

country_condition	weapons_condition	n	mean	p10	p25	median	p75	p90
Chinese strike	Human-operated	737	72.11	46.0	51	75	94	100
Chinese strike	Autonomous AI	737	62.09	33.6	50	59	80	99
U.S. strike	Human-operated	733	62.83	36.2	50	59	84	99
U.S. strike	Autonomous AI	728	54.21	17.0	43	51	71	92

ggplot(
  intentionality_distribution_data,
  aes(x = perceived_intentionality, fill = country_condition, color = country_condition)
) +
  geom_density(alpha = 0.20, linewidth = 0.8, adjust = 1.1, na.rm = TRUE) +
  facet_wrap(~ weapons_condition) +
  scale_fill_manual(
    values = c("Chinese strike" = "#A63A33", "U.S. strike" = "#235A8C"),
    name = "Country scenario"
  ) +
  scale_color_manual(
    values = c("Chinese strike" = "#A63A33", "U.S. strike" = "#235A8C"),
    name = "Country scenario"
  ) +
  scale_x_continuous(limits = c(0, 100), breaks = seq(0, 100, 20)) +
  labs(
    title = "Distribution of Perceived Intentionality by Condition",
    x = "Perceived intentionality",
    y = "Density",
    caption = "Density plots show the full outcome distribution rather than only the mean."
  ) +
  theme_minimal(base_size = 12) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank(),
    strip.text = element_text(face = "bold")
  )

Main Outcome Plots

Intentionality

intentionality_summary <- survey %>%
  filter(!is.na(condition_agressor),
         !is.na(condition_autonomy),
         !is.na(intentionality)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomy = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    intentionality = as.numeric(intentionality)
  ) %>%
  group_by(country, autonomy) %>%
  summarize(
    n = n(),
    mean = mean(intentionality),
    se = sd(intentionality) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

intentionality_summary$country <- factor(intentionality_summary$country, levels = c("China", "U.S."))
intentionality_summary$autonomy <- factor(intentionality_summary$autonomy, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  intentionality_summary,
  aes(x = autonomy, y = mean, color = country, group = country)
) +
  geom_line(linewidth = 1, alpha = 0.45) +
  geom_point(size = 3.5) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.8) +
  geom_text(aes(label = paste0(round(mean, 1), "%")),
            vjust = -1.2,
            show.legend = FALSE) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_y_continuous(
    limits = c(50, 80),
    breaks = seq(50, 80, 5),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by Country and Weapons System",
    x = "Weapons-system condition",
    y = "Probability of intentionality",
    caption = "Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 14) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank()
  )

Escalation

escalation_summary <- survey %>%
  filter(!is.na(condition_agressor),
         !is.na(condition_autonomy),
         !is.na(escalation)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomy = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    escalation = as.numeric(escalation)
  ) %>%
  group_by(country, autonomy) %>%
  summarize(
    n = n(),
    mean = mean(escalation),
    se = sd(escalation) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  ) %>%
  mutate(
    label_y = ifelse(country == "U.S.", ci_low - 2, ci_high + 2)
  )

escalation_summary$country <- factor(escalation_summary$country, levels = c("China", "U.S."))
escalation_summary$autonomy <- factor(escalation_summary$autonomy, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  escalation_summary,
  aes(x = autonomy, y = mean, color = country, group = country)
) +
  geom_line(linewidth = 1, alpha = 0.45) +
  geom_point(size = 3.5) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.8) +
  geom_text(aes(y = label_y, label = paste0(round(mean, 1), "%")),
            show.legend = FALSE) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_y_continuous(
    limits = c(50, 70),
    breaks = seq(50, 70, 5),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Likelihood of Escalation by Country and Weapons System",
    x = "Weapons-system condition",
    y = "Likelihood of escalation",
    caption = "Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 14) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank()
  )

Justification

justified_summary <- survey %>%
  filter(!is.na(condition_agressor),
         !is.na(condition_autonomy),
         !is.na(justified)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomy = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    justified = as.numeric(justified)
  ) %>%
  group_by(country, autonomy) %>%
  summarize(
    n = n(),
    mean = mean(justified),
    se = sd(justified) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  ) %>%
  mutate(
    label_y = ifelse(country == "U.S.", ci_low - 2, ci_high + 2)
  )

justified_summary$country <- factor(justified_summary$country, levels = c("China", "U.S."))
justified_summary$autonomy <- factor(justified_summary$autonomy, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  justified_summary,
  aes(x = autonomy, y = mean, color = country, group = country)
) +
  geom_line(linewidth = 1, alpha = 0.45) +
  geom_point(size = 3.5) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.8) +
  geom_text(aes(y = label_y, label = paste0(round(mean, 1))),
            show.legend = FALSE) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_y_continuous(
    limits = c(50, 70),
    breaks = seq(50, 70, 5),
    labels = function(x) paste0(x)
  ) +
  labs(
    title = "Justifiability of Escalation by Country and Weapons System",
    x = "Weapons-system condition",
    y = "Level of Justification",
    caption = "Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 14) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.major.x = element_blank()
  )

Responsibility

responsibility_data <- bind_rows(
  survey %>%
    transmute(
      condition_agressor,
      condition_autonomy,
      pid,
      actor = "Senior political leaders",
      responsibility = as.numeric(resp_leader)
    ),
  survey %>%
    transmute(
      condition_agressor,
      condition_autonomy,
      pid,
      actor = "Military officials",
      responsibility = as.numeric(resp_military)
    ),
  survey %>%
    transmute(
      condition_agressor,
      condition_autonomy,
      pid,
      actor = "Human operator",
      responsibility = as.numeric(resp_pilot)
    ),
  survey %>%
    transmute(
      condition_agressor,
      condition_autonomy,
      pid,
      actor = "AI weapons system",
      responsibility = as.numeric(resp_ai)
    )
)

responsibility_summary <- responsibility_data %>%
  filter(!is.na(condition_agressor),
         !is.na(condition_autonomy),
         !is.na(pid),
         !is.na(responsibility)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomy = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    pid = factor(pid, levels = c("Democrat", "Independent", "Republican"))
  ) %>%
  group_by(pid, actor, country, autonomy) %>%
  summarize(
    n = n(),
    mean = mean(responsibility),
    se = sd(responsibility) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

responsibility_summary$country <- factor(responsibility_summary$country, levels = c("China", "U.S."))
responsibility_summary$autonomy <- factor(responsibility_summary$autonomy, levels = c("Human-operated", "Autonomous AI"))
responsibility_summary$pid <- factor(
  responsibility_summary$pid,
  levels = c("Democrat", "Independent", "Republican")
)
responsibility_summary$actor <- factor(
  responsibility_summary$actor,
  levels = c("Senior political leaders", "Military officials", "Human operator", "AI weapons system")
)

ggplot(
  responsibility_summary,
  aes(x = autonomy, y = mean, color = country, group = country)
) +
  geom_line(linewidth = 1, alpha = 0.45) +
  geom_point(size = 2.5) +
  geom_errorbar(aes(ymin = ci_low, ymax = ci_high), width = 0.08, linewidth = 0.8) +
  facet_grid(pid ~ actor) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_y_continuous(
    breaks = seq(50, 90, 10),
    labels = function(x) paste0(x)
  ) +
  coord_cartesian(ylim = c(50, 90)) +
  labs(
    title = "Perceived Responsibility by Party Identification",
    x = "Weapons-system condition",
    y = "Perceived responsibility",
    caption = "Error bars are 95% confidence intervals."
  ) +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    axis.text.x = element_text(size = 6),
    panel.grid.major.x = element_blank(),
    strip.text = element_text(face = "bold", size = 7)
  )

## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?
## `geom_line()`: Each group consists of only one observation.
## ℹ Do you need to adjust the group aesthetic?

Covariate and Moderator Plots

Covariate Distributions

distribution_data <- bind_rows(
  intentionality_regression_data %>%
    transmute(variable = "Relative AI effectiveness", value = relative_ai),
  intentionality_regression_data %>%
    transmute(variable = "AI comfort", value = ai_comfort)
)

distribution_data$variable <- factor(
  distribution_data$variable,
  levels = c("Relative AI effectiveness", "AI comfort")
)

ggplot(distribution_data, aes(x = value)) +
  geom_histogram(bins = 30, fill = "#235A8C", color = "white") +
  facet_wrap(~ variable, scales = "free_x", ncol = 2) +
  labs(
    title = "Distributions of Relative AI Effectiveness and AI Comfort",
    x = "Value",
    y = "Number of respondents"
  ) +
  theme_minimal(base_size = 12) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    panel.grid.minor = element_blank(),
    strip.text = element_text(face = "bold")
  )

AI Comfort and Intentionality

comfort_intentionality_data <- intentionality_regression_data %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomous = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI")
  )

comfort_intentionality_data$country <- factor(comfort_intentionality_data$country, levels = c("China", "U.S."))
comfort_intentionality_data$autonomous <- factor(comfort_intentionality_data$autonomous, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  comfort_intentionality_data,
  aes(x = ai_comfort, y = perceived_intentionality, color = country, shape = autonomous)
) +
  geom_jitter(width = 0.06, height = 0.4, alpha = 0.15, size = 1.2) +
  geom_smooth(aes(linetype = autonomous), method = "lm", se = TRUE, linewidth = 0.9) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_linetype_manual(
    values = c("Human-operated" = "solid", "Autonomous AI" = "dashed"),
    name = "Weapons-system condition"
  ) +
  scale_x_continuous(
    limits = c(-2, 2),
    breaks = seq(-2, 2, 1)
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by AI Comfort",
    x = "AI comfort (-2 to 2 average)",
    y = "Perceived intentionality",
    caption = "Points are individual respondents; lines are linear fits by country and weapons-system condition."
  ) +
  guides(linetype = "none") +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank()
  )

## `geom_smooth()` using formula = 'y ~ x'

## Warning: Removed 355 rows containing missing values or values outside the scale range
## (`geom_point()`).

Relative AI Effectiveness and Intentionality

relative_ai_intentionality_data <- intentionality_regression_data %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomous = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI")
  )

relative_ai_intentionality_data$country <- factor(relative_ai_intentionality_data$country, levels = c("China", "U.S."))
relative_ai_intentionality_data$autonomous <- factor(relative_ai_intentionality_data$autonomous, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  relative_ai_intentionality_data,
  aes(x = relative_ai, y = perceived_intentionality, color = country, shape = autonomous)
) +
  geom_jitter(width = 0.6, height = 0.4, alpha = 0.15, size = 1.2) +
  geom_smooth(aes(linetype = autonomous), method = "lm", se = TRUE, linewidth = 0.9) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_linetype_manual(
    values = c("Human-operated" = "solid", "Autonomous AI" = "dashed"),
    name = "Weapons-system condition"
  ) +
  scale_x_continuous(
    limits = c(-100, 100),
    breaks = seq(-100, 100, 50)
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by Relative AI Effectiveness",
    x = "Relative U.S.-China AI effectiveness",
    y = "Perceived intentionality",
    caption = "Points are individual respondents; lines are linear fits by country and weapons-system condition."
  ) +
  guides(linetype = "none") +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank()
  )

## `geom_smooth()` using formula = 'y ~ x'

## Warning: Removed 151 rows containing missing values or values outside the scale range
## (`geom_point()`).

Party Identification and Intentionality

pid_intentionality_data <- intentionality_regression_data %>%
  filter(!is.na(pid)) %>%
  mutate(
    country = factor(ifelse(condition_agressor == "US", "U.S.", condition_agressor), levels = c("China", "U.S.")),
    autonomous = factor(ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"), levels = c("Human-operated", "Autonomous AI")),
    pid = factor(pid, levels = c("Democrat", "Independent", "Republican")),
    plot_group = interaction(country, autonomous, drop = TRUE)
  )

pid_intentionality_summary <- pid_intentionality_data %>%
  group_by(pid, country, autonomous, plot_group) %>%
  summarize(
    n = n(),
    mean = mean(perceived_intentionality),
    se = sd(perceived_intentionality) / sqrt(n),
    ci_low = mean - 1.96 * se,
    ci_high = mean + 1.96 * se,
    .groups = "drop"
  )

ggplot(
  pid_intentionality_data,
  aes(x = pid, y = perceived_intentionality, color = country, shape = autonomous, group = plot_group)
) +
  geom_point(
    position = position_jitterdodge(jitter.width = 0.16, jitter.height = 0.4, dodge.width = 0.75),
    alpha = 0.12,
    size = 1.1
  ) +
  geom_errorbar(
    data = pid_intentionality_summary,
    aes(y = mean, ymin = ci_low, ymax = ci_high),
    position = position_dodge(width = 0.75),
    width = 0.18,
    linewidth = 0.7
  ) +
  geom_point(
    data = pid_intentionality_summary,
    aes(y = mean),
    position = position_dodge(width = 0.75),
    size = 2.8
  ) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_x_discrete(
    labels = c("Democrat", "Independent", "Republican", "Other/none")
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by Party Identification",
    x = "Party identification",
    y = "Perceived intentionality",
    caption = "Faded points are individual respondents; larger points are group means with 95% confidence intervals."
  ) +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank(),
    axis.text.x = element_text(size = 7)
  )

## Warning: Removed 155 rows containing missing values or values outside the scale range
## (`geom_point()`).

U.S. AI Effectiveness and Intentionality

us_ai_effectiveness_data <- intentionality_regression_data %>%
  filter(!is.na(ai_us)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomous = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    us_ai_effectiveness = as.numeric(ai_us)
  ) %>%
  filter(!is.na(us_ai_effectiveness))

us_ai_effectiveness_data$country <- factor(us_ai_effectiveness_data$country, levels = c("China", "U.S."))
us_ai_effectiveness_data$autonomous <- factor(us_ai_effectiveness_data$autonomous, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  us_ai_effectiveness_data,
  aes(x = us_ai_effectiveness, y = perceived_intentionality, color = country, shape = autonomous)
) +
  geom_jitter(width = 0.6, height = 0.4, alpha = 0.15, size = 1.2) +
  geom_smooth(aes(linetype = autonomous), method = "lm", se = TRUE, linewidth = 0.9) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_linetype_manual(
    values = c("Human-operated" = "solid", "Autonomous AI" = "dashed"),
    name = "Weapons-system condition"
  ) +
  scale_x_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20)
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by U.S. AI Effectiveness",
    x = "U.S. AI effectiveness rating",
    y = "Perceived intentionality",
    caption = "Points are individual respondents; lines are linear fits by country and weapons-system condition."
  ) +
  guides(linetype = "none") +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank()
  )

## `geom_smooth()` using formula = 'y ~ x'

## Warning: Removed 265 rows containing missing values or values outside the scale range
## (`geom_point()`).

Chinese AI Effectiveness and Intentionality

chinese_ai_effectiveness_data <- intentionality_regression_data %>%
  filter(!is.na(ai_china)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomous = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    chinese_ai_effectiveness = as.numeric(ai_china)
  ) %>%
  filter(!is.na(chinese_ai_effectiveness))

chinese_ai_effectiveness_data$country <- factor(chinese_ai_effectiveness_data$country, levels = c("China", "U.S."))
chinese_ai_effectiveness_data$autonomous <- factor(chinese_ai_effectiveness_data$autonomous, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  chinese_ai_effectiveness_data,
  aes(x = chinese_ai_effectiveness, y = perceived_intentionality, color = country, shape = autonomous)
) +
  geom_jitter(width = 0.6, height = 0.4, alpha = 0.15, size = 1.2) +
  geom_smooth(aes(linetype = autonomous), method = "lm", se = TRUE, linewidth = 0.9) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_linetype_manual(
    values = c("Human-operated" = "solid", "Autonomous AI" = "dashed"),
    name = "Weapons-system condition"
  ) +
  scale_x_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20)
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by Chinese AI Effectiveness",
    x = "Chinese AI effectiveness rating",
    y = "Perceived intentionality",
    caption = "Points are individual respondents; lines are linear fits by country and weapons-system condition."
  ) +
  guides(linetype = "none") +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank()
  )

## `geom_smooth()` using formula = 'y ~ x'

## Warning: Removed 234 rows containing missing values or values outside the scale range
## (`geom_point()`).

South Korean AI Effectiveness and Intentionality

southkorea_ai_effectiveness_data <- intentionality_regression_data %>%
  filter(!is.na(ai_sk)) %>%
  mutate(
    country = ifelse(condition_agressor == "US", "U.S.", condition_agressor),
    autonomous = ifelse(condition_autonomy == "human", "Human-operated", "Autonomous AI"),
    southkorea_ai_effectiveness = as.numeric(ai_sk)
  ) %>%
  filter(!is.na(southkorea_ai_effectiveness))

southkorea_ai_effectiveness_data$country <- factor(southkorea_ai_effectiveness_data$country, levels = c("China", "U.S."))
southkorea_ai_effectiveness_data$autonomous <- factor(southkorea_ai_effectiveness_data$autonomous, levels = c("Human-operated", "Autonomous AI"))

ggplot(
  southkorea_ai_effectiveness_data,
  aes(x = southkorea_ai_effectiveness, y = perceived_intentionality, color = country, shape = autonomous)
) +
  geom_jitter(width = 0.6, height = 0.4, alpha = 0.15, size = 1.2) +
  geom_smooth(aes(linetype = autonomous), method = "lm", se = TRUE, linewidth = 0.9) +
  scale_color_manual(
    values = c("China" = "#A63A33", "U.S." = "#235A8C"),
    name = "Country condition"
  ) +
  scale_shape_manual(
    values = c("Human-operated" = 16, "Autonomous AI" = 17),
    name = "Weapons-system condition"
  ) +
  scale_linetype_manual(
    values = c("Human-operated" = "solid", "Autonomous AI" = "dashed"),
    name = "Weapons-system condition"
  ) +
  scale_x_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20)
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 20),
    labels = function(x) paste0(x, "%")
  ) +
  labs(
    title = "Perceived Intentionality by South Korean AI Effectiveness",
    x = "South Korea AI effectiveness rating",
    y = "Perceived intentionality",
    caption = "Points are individual respondents; lines are linear fits by country and weapons-system condition."
  ) +
  guides(linetype = "none") +
  theme_minimal(base_size = 8) +
  theme(
    plot.title = element_text(face = "bold", hjust = 0.5),
    legend.position = "top",
    panel.grid.minor = element_blank()
  )

## `geom_smooth()` using formula = 'y ~ x'

## Warning: Removed 225 rows containing missing values or values outside the scale range
## (`geom_point()`).