Draft 2025
Introduction
Party identification is one of the most important and most studied concepts in political science today. The use of party ID, as it is commonly known, as an explanatory variable is ubiquitous. With that said, there has been relatively little recent scholarship on the subject of how party id may change over time. That is largely as a result of a current consensus that party id is near-immutable and unavoidable, i.e. that almost no voters begin to identify with one party after identifying with the other and that there are very few true independents, politically engaged individuals who do not meaningfully identify with either party.
That consensus remains in spite of the fact that there has not been a significant attempt to re-evaluate the nature of party id in the era of “big data.” While there have been recent efforts to estimate the stability of party id (Tucker, Montgomery, and Smith, 2019) these do not fully leverage the technical advantages that have developed over the course of recent years.
Data
Data for this project comes from YouGov’s Profile Data Library (PDL) from 2004 to 2024. The PDL is how YouGov tracks core questions among respondents regardless of what surveys they opt into answering. When a respondent agrees to answer a survey, they may also be shown a number of other questions that are assumed to be broadly relevant, such as demographics, ideology, party id, and more, if they haven’t recently answered them. That updates their PDL entry. PDL questions can also be updated when a respondent answers a question from a contracted survey that is identical to a PDL question.
It is not the case the list of questions in the PDL are static. Questions can be added or removed over time. Throughout the years, however, party id in both three category varieties and seven category varieties have remained. The appendix contains a full listing of the variables included in the PDL, but in the main text we include PID as the primary variable and ideology, year of birth, religiosity, and presidential approval for comparison and context.
Party ID can have either be measured with three categories (‘Republican,’ ‘Democrat,’ or ‘Independent’) or seven categories (‘Strong Republican,’ ‘Weak Republican,’ ‘Lean Republican,’ ‘Independent,’ ‘Lean Democrat,’ ‘Weak Democrat,’ or ‘Strong Democrat’) where the seven categories are nested within the three. Thematically, we group party id with ideology which has five categories (‘Very conservative,’ ‘Conservative,’ ‘Moderate,’ ‘Liberal,’ or ‘Very liberal’) as they are often linked in the literature due to their thematic similarities.
Some respondents are removed because they likely falsified their answers. As we show in the attached appendix, respondents who shift their answers on their gender are much more likely to shift their answer on what their birth year is, which is something that cannot change. Excluding respondents who give different answers for their gender at different times is supported empirically and also corresponds to institutional knowledge at YouGov.
We do not have the same amount of data across all variables and all years. The core variables of party ID and ideology have hundreds of thousands of observations in 2004 and 2006-2024, which is more than enough data to support analysis. Demographic questions, which we include largely as a baseline for understanding the context for the stability of other variables, have large number of observations as well, though religious questions only started to be asked in 2008.
Presidential Elections
The central question of this analysis is simply “how stable is party id?” The data has a panel structure, with the same respondents giving multiple responses across time, but the observations are not at even intervals. Due to the nature of the PDL, responses come when the respondent is answering a separate survey and normally when it has been a sufficient period of time since they were last asked.
To get around the issue of inconsistent intervals between respondents’ answers to a PDL question, we start by subsetting the data to consistent intervals. We compare respondents’ answers between presidential election years. For example, a respondents’ last answer to the party id question in 2012 is compared to their last answer to party id in 2008. To begin with, we avoid the issue of multiple observations by the same respondent in the same year by solely selecting the final observation. In total, this ensures we are analyzing respondents over the course of identical four year periods, at the cost of eliminating some of the data from our analysis. Our overall N, however, is sufficiently large so as to allow this analysis regardless.
Between 2004 and 2008, 73.7 percent of 118951 respondents who answered the three category party id question in both years gave the same answer both times. For 2008 - 2012, that number is 77.8 of 73105 respondents, for 2012 - 2016, that number is 82.6 of 55647 respondents, for 2016 - 2020 that number is 82.3 of 58606 respondents, and for 2020 - 2024 that is 85.3 of 60571 respondents. That compares to 60.5, 68.4, 71.2, 70.5, and 70.5 percent for 7 category party ID. We see that overall these stability numbers are very high and have only been increasing over time. Considering the possibility of measurement error, which would attenuate stability numbers, the 85 percent stability for three category PID that respondents show in the four year window of 2020 to 2024 is very high. Interestingly, stability does not jump in the 2016-2020 interval, compared to the 2012-2016 interval. The possible substantive implication of this is that changes around Donald Trump’s first election were not immediately “locked-in.”
In Figure 1 we show what percentage of respondents ever change their partisan identity on the Y axis and on the X axis we show the number of presidential years in which the respondent gave a party ID response. The more a respondent answers their party id, the more likely they are to not select the same answer every single time, so a negative trend line if is guaranteed. While respondents are always more likely to be perfectly stable than not, around 35% of respondents who gave a response in every presidential election year from 2004 to 2024 changed their party id at least once. This is a fairly substantial group, though the persence of measurement error will increase the percentage of respondents who give multiple different answers for party id over the years.
Once we’ve shown how stable each partisan identity is, the natural follow-up question is where do the voters who change their partisanship go? Our understanding of how Americans change partisanship is very different if extreme partisans switch back and forth between political parties versus if they go back and forth from an extreme position to a moderate one.
In Table 1, Table 2, Table 3, Table 4, and Table 5 we present tables that compare the partisanship of respondents in one presidential election year with their partisanship four years later in the next presidential election year.
In these tables, the rows represent a respondent’s party identification in the earlier year, while the columns indicate a respondent’s party identification four years later. The diagonal values therefore show average stability for each value of three category party identification across each four year span in the dataset, weighted to Current Population Survey (CPS) estimates for gender, race, education, age, and census division. The off diagonal rows show the percentage of respondents in each party (row) who switched to each other partisanship (column).
|
2008 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2004 PID3 | dem | 91% | 7% | 2% |
| ind | 21% | 63% | 17% | |
| rep | 2% | 17% | 81% | |
|
2012 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2008 PID3 | dem | 90% | 8% | 2% |
| ind | 9% | 79% | 13% | |
| rep | 2% | 16% | 83% | |
|
2016 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2012 PID3 | dem | 90% | 7% | 2% |
| ind | 7% | 83% | 10% | |
| rep | 2% | 9% | 89% | |
|
2020 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2016 PID3 | dem | 91% | 6% | 3% |
| ind | 9% | 80% | 11% | |
| rep | 2% | 8% | 90% | |
|
2020 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2016 PID3 | dem | 91% | 6% | 3% |
| ind | 9% | 80% | 11% | |
| rep | 2% | 8% | 90% | |
We can see that with regards to three party party id, Democratic respondents are much more stable than either Republican or Independent respondents, until approximately the 2012-2016 period and onwards, where Republicans catch up. Possibly of note for Republican respondents, this is the first period in the data entirely post-Tea Party revolution in the Republican Party. Independent respondents also become more stable in the same period, though they remain behind the other respondents in stability. “Don’t know” responses are also excluded from this analysis.
As before, we compare the seven category party identification results in an election year with seven category party identification in the next election year. We again provide stability for each category, with the earlier party identification on the row and the later party identification on the column.
|
2008 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2004 PID7 | strong dem | 87% | 6% | 5% | 1% | 0% | 0% | 0% |
| weak dem | 45% | 33% | 11% | 4% | 3% | 2% | 2% | |
| lean dem | 28% | 13% | 48% | 9% | 1% | 1% | 0% | |
| ind | 7% | 5% | 22% | 43% | 15% | 4% | 4% | |
| lean rep | 1% | 1% | 2% | 13% | 50% | 12% | 21% | |
| weak rep | 2% | 3% | 4% | 6% | 17% | 37% | 32% | |
| strong rep | 0% | 0% | 0% | 2% | 10% | 10% | 77% | |
|
2012 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2008 PID7 | strong dem | 83% | 10% | 4% | 2% | 0% | 1% | 1% |
| weak dem | 20% | 54% | 10% | 7% | 3% | 3% | 2% | |
| lean dem | 14% | 8% | 60% | 14% | 3% | 1% | 0% | |
| ind | 2% | 2% | 11% | 68% | 13% | 2% | 3% | |
| lean rep | 0% | 1% | 2% | 19% | 59% | 6% | 12% | |
| weak rep | 1% | 2% | 2% | 6% | 17% | 53% | 19% | |
| strong rep | 0% | 0% | 0% | 3% | 13% | 9% | 75% | |
|
2016 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2012 PID7 | strong dem | 84% | 9% | 4% | 2% | 0% | 0% | 1% |
| weak dem | 16% | 60% | 8% | 7% | 2% | 4% | 2% | |
| lean dem | 11% | 9% | 59% | 17% | 2% | 1% | 0% | |
| ind | 1% | 2% | 7% | 72% | 12% | 3% | 3% | |
| lean rep | 0% | 0% | 1% | 17% | 60% | 9% | 12% | |
| weak rep | 1% | 2% | 1% | 5% | 12% | 63% | 16% | |
| strong rep | 1% | 0% | 0% | 2% | 9% | 11% | 76% | |
|
2020 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2016 PID7 | strong dem | 87% | 7% | 3% | 1% | 0% | 0% | 1% |
| weak dem | 21% | 57% | 8% | 5% | 2% | 3% | 4% | |
| lean dem | 14% | 9% | 63% | 11% | 1% | 1% | 1% | |
| ind | 3% | 3% | 10% | 64% | 13% | 3% | 5% | |
| lean rep | 1% | 1% | 2% | 13% | 60% | 8% | 17% | |
| weak rep | 1% | 3% | 2% | 5% | 9% | 53% | 27% | |
| strong rep | 1% | 0% | 0% | 1% | 4% | 7% | 87% | |
|
2024 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2020 PID7 | strong dem | 89% | 6% | 2% | 2% | 0% | 0% | 0% |
| weak dem | 17% | 66% | 7% | 4% | 2% | 2% | 1% | |
| lean dem | 11% | 9% | 67% | 11% | 1% | 1% | 1% | |
| ind | 2% | 2% | 10% | 73% | 10% | 2% | 2% | |
| lean rep | 0% | 1% | 2% | 14% | 66% | 7% | 10% | |
| weak rep | 1% | 3% | 2% | 5% | 11% | 64% | 14% | |
| strong rep | 1% | 0% | 0% | 2% | 6% | 9% | 83% | |
Table 6, Table 7, Table 8, Table 9, and Table 10 all show the rate of stability between each presidential election year for respondents of each seven category party id. Unsurprisingly, strong partisans are the most stable of any category. These respondents may be the highest in knowledge and are the most active people in partisan politics. Consistent with some previous results, partisan leans are actually more stable than weak partisans. This may be the case if a certain number of high knowledge respondents are influenced by normative bias in favor of self-identifying as an independent regardless of actual partisanship. The two question approach to seven category party id may even accentuate any such bias, as respondents are first asked if they’re a democrat, republican, or independent, before clarifying if they lean towards a particular party if they answered independent, and whether they’re a strong or weak partisan otherwise. Respondents may first be expressing a pro-Independent bias, before expressing their true partisan lean in the second question. If respondents who are more engaged in politics feel this pressure more highly, these respondents are likely also more stable in their choices.
Over time, stability increases among Republicans and Independents towards the end of the survey period. In comparison to, and unlike, three category party id, we do not see a dramatic difference in Republican or Independent respondents stability moving from 2008-2012 to 2012-2016. For this to be possible, there was likely a change in the distribution of Republican respondents across the weak Republican and strong Republican categories. Also notable is a dramatic amount of movement from lean and weak partisans to strong partisans in the 2004-2008 period.
Congressional Elections
To gain greater insight into the trends discussed in the presidential year to presidential year analysis, we break down the four year intervals into two year intervals - the same as the time between Congressional elections.
| 2006 PID3_3 | 2008 PID3_3 | ||||||
|---|---|---|---|---|---|---|---|
|
|
As expected, Table 11 shows that stability across two years is higher than across four. In Table 1, we saw that from 2004 to 2008, stability among Democrats was 91%, while stability among Republicans was 81%. Previous research shows that the party of the president loses vote share in the midterms relative to the previous general election. Interestingly, we find that the differential in partisan stability increases for outparty respondents in the second half of a presidential term relative to the first, as Democratic respondents grow more stable, while Republicans do not experience the same increase. One possible explanation for this pattern, which continues throughout our data to some degree is that the second half of presidential terms may be less popular or harden opposition even more than the first.
| 2010 PID3_3 | 2012 PID3_3 | ||||||
|---|---|---|---|---|---|---|---|
|
|
In Table 12, both Democratic and Republican respondents grow more stable in the 2010-2012 period compared to the 2008-2010 period, though the growth is much higher for Republican respondents. This brings Democrats and Republicans closer in line to each other when it comes to stability. For historical context, the Tea Party movement came into the forefront of American politics during the 2010 midterm election. There is likely some underlying historical reason that is increasing stability across all partisans.
| 2014 PID3_3 | 2016 PID3_3 | ||||||
|---|---|---|---|---|---|---|---|
|
|
In Table 13 we replicate the shift data for the next four years. In the last term of Obama’s presidency, Democrats remained stable, though at 93% stable in both of the two years, it would be difficult to increase this number much higher. The stability of Republican respondents remains stable at the 91% mark through the 2014-2016 period. It’s also worth noting that in two out of the three periods shown so far, independents decrease in stability in the second two year period compared to the first. This is compatible with a model in which Independents fall into a holding pattern of sorts between the first presidential election and the midterm election, though this is less stable than partisans. Then, Independents further reevaluate at the next presidential election. The particularities of year-to-year changes are also possible through what is solely a three interval sample.
| 2018 PID3_3 | 2020 PID3_3 | ||||||
|---|---|---|---|---|---|---|---|
|
|
Table 14 follows the patterns we have largely seen so far. Across Donald Trump’s first term, Democrats become marginally more stable and Independents less so. It is also the case that at this point the growth in partisan stability among non-Democrats has largely leveled off. Democrats, Republicans, and Independents all appear to max out at an inter-election stability of 85 to 93 or 94 percent. Measurement error caps this number below 100%, but this also indicates that there is some amount of partisan switching in every partisanship between every election.
|
2022 PID3
|
||||
|---|---|---|---|---|
| dem | ind | rep | ||
| 2020 PID3 | dem | 91% | 7% | 3% |
| ind | 10% | 84% | 7% | |
| rep | 3% | 9% | 89% | |
Unfortunately, there is insufficient data in 2024 to compare the 2022 - 2024 period to the 2020 - 2022 period. Compared to either the 2016 - 2018 or 2018 - 2020 period, however, stability is down largely across the board, though not by too much. Overall, transitions into, and out of, Independent partisanship is more common than any other partisanship. Again, this is consistent with the typical ordinal understanding of partisanship, which implies that any movement from Democrat to Republican or the inverse moves through Independent first.
To get more insight into the detailed mechanics of how respondents move through partisanship, we also compare the seven category party ID measure for congressional elections.
| 2006 PID7 | 2008 PID7 | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
|
Table 16, Table 17, Table 18, Table 19, and Table 20 replicate the 3 category party ID comparisons with 7 category party ID instead.
From Table 16 we can see a few things that are generally expected from the literature. Strong partisans are much more stable than weak or leaning partisans. Additionally, leaning partisans are indeed again more stable than weak partisans. Finally, true independents are significantly less stable than strong partisans, but less stable than weak or leaning partisans. This suggests that independent is an identity unto itself, but it is much less strong than Democratic or Republican partisanship.
As expected from the 2004-2006 and 2006-2008 3 category party ID cases, strong democrats are much more stable than strong republicans in the 2006 - 2008 period and marginally less so in the 2004 - 2006 period. Interestingly, as strong democrats become much more stable in the 2006 - 2008 period, weak democrats and lean democrats move towards the strong democrat category in large numbers. A full 40% of respondents who identified as a weak democrat in 2006 identified as a strong democrat in 2008. This also strongly suggests that movements across the partisan spectrum move in parallel, at least on each half of the partisan divide. Stability in strong democrats increases by 12% from the first period to the second, movement from weak democrat to increased by 16%, and movement from leaning democrat to strong democrat increased by 13%. There are no corresponding changes on the Republican side.
| 2010 PID7 | 2012 PID7 | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
|
Table 17 shows a similar trend. Strong democrat stability increases from the 2008-2010 period to the 2010-2012, and so does movement from weak democrat to strong democrat and lean democrat to strong democrat. In this instance, however, there is a 9 percentage point increase in strong republican stability, but without the corresponding increases in weak or leaning republican to strong republican movement. We also see overall increases in stability among non strong partisans, especially independents compared to the 2004-2006 or 2006-2008 periods.
| 2014 PID7 | 2016 PID7 | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
|
Table 18 shows a continuation of the previous trends. Stability on the high end has largely settled into what appears to be the approximate maximum for this measure in this time period. The non strong partisans have come up to a higher average stability as well, with levels in the 60% range, a large increase when compared to values in the 40-50% range earlier in the observation period. It is possible that this increase in stability for rarer or more specific responses is influenced by respondents being asked this question multiple times as YouGov respondents. That said, the beginning of the PDL dataset does not necessarily reflect the first time respondents saw these questions. Additionally, respondents enter into the PDL after it starts, so many of the respondents at any given time could have been added recently and may not have seen partisanship questions before.
| 2018 PID7 | 2020 PID7 | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
|
In Table 19, stability across 2016-2018 and 2018-2020 are nearly identical. Also, both republican and democratic respondents have similar stability to each other within years. Within Donald Trump’s first term, partisan identification is remarkably stable, which is somewhat in contrast to the popular discussion of this period in the news media.
|
2022 PID7
|
||||||||
|---|---|---|---|---|---|---|---|---|
| strong dem | weak dem | lean dem | ind | lean rep | weak rep | strong rep | ||
| 2020 PID7 | strong dem | 83% | 9% | 5% | 1% | 0% | 0% | 1% |
| weak dem | 17% | 56% | 13% | 7% | 2% | 3% | 1% | |
| lean dem | 12% | 12% | 58% | 15% | 2% | 1% | 1% | |
| ind | 2% | 3% | 11% | 70% | 10% | 3% | 2% | |
| lean rep | 0% | 1% | 3% | 19% | 58% | 9% | 9% | |
| weak rep | 1% | 4% | 2% | 7% | 11% | 56% | 18% | |
| strong rep | 2% | 1% | 0% | 2% | 8% | 9% | 78% | |
In Table 20 there is an overall decrease in the stability of 7 category party ID across the board. One of the more notable trends in stability is increased movement out of identifying as a “true” independent, though mostly towards the “leaning” partisanship categories.
Comparison
While we have measured the stability of party id between elections, we need to contextualize these stability numbers by comparing them to another data source. The ANES has the largest amount of historical data on voters partisan identification over time. Although much of the available panel data from the ANES comes from before the PDL data was measured, both the ANES and the PDL have panel data for 2016 to 2020. While the comparison of those years is the primary quantity of interest, we include all available ANES panels and all presidential year panels
In Table 22 and Table 21 we compare the stability of the three and seven category party id questions in our PDL data and in the ANES data. Included in these tables are the stabilities of each of the four year ANES panels, as well as the four year consistencies found in the PDL data. Both the ANES and PDL columns are weighted. The provided end of panel weights are used for the ANES, while the PDL is weighted to the CPS estimates for age, race, gender, education, and census division. The only four year period that there is both an ANES and PDL panel for is 2016-2020, which are the key variables for comparison and the last two columns in each table. In Table 21, the PDL respondents are 5 to 15 points more stable across different partisanships. There is more variance in the the difference between the ANES 2016-2020 and PDL 2016-2020 columns in Table 22 compared to Table 21. This makes logical sense given that there will always be fewer observations in the seven category PID cells than the three category cells. What this does not explain, however, is that the strong dem and strong rep categories are as stable or more stable than the dem and rep categories. This is explained by strong partisans being more committed to their partisanship and more educated as well, meaning they are easier to measure across surveys.
It is important to note that while the 1956-1960 and 1972-1976 panels are calculated in the same manner as the other panels, the political world in which they take place is one dramatically different from the other panel periods. It is completely expected, given the political context at the time, that the inter-election stability of these much earlier periods is lower. These periods are included not because they are expected to have similar stabilities to the PDL data. Instead, it is included to give a sense of just how much individual partisanship has changed over time. Even outside of any changes in the electorate, survey response rates have decreased over time, especially with the switch to opt-in, online polling. As response rates have decreased, respondents who are sampled are more attentive and more stable than the rest of the electorate (Converse, 2006).
| PID 3 | 56-60, ANES | 72 - 76, ANES | 04-08 | 08-12 | 12-16 | 16-20, ANES | 16-20 |
|---|---|---|---|---|---|---|---|
| dem | 86.2 | 85.6 | 88.1 | 86.8 | 85.7 | 87.2 | 86.7 |
| ind | 79.5 | 76.5 | 54.8 | 66.4 | 73.7 | 75.4 | 80.2 |
| rep | 84.1 | 84.8 | 77.4 | 74.0 | 78.2 | 85.1 | 82.9 |
| PID 7 | 56-60, ANES | 72 - 76, ANES | 04-08 | 08-12 | 12-16 | 16-20, ANES | 16-20 |
|---|---|---|---|---|---|---|---|
| strong dem | 81.0 | 81.4 | 85.8 | 81.0 | 80.2 | 83.7 | 85.0 |
| weak dem | 59.2 | 54.6 | 34.1 | 50.9 | 53.4 | 56.9 | 55.4 |
| lean dem | 55.8 | 56.8 | 44.6 | 42.6 | 55.1 | 60.5 | 58.0 |
| ind | 71.1 | 59.5 | 44.4 | 52.7 | 64.6 | 64.6 | 68.2 |
| lean rep | 57.8 | 56.9 | 45.5 | 52.4 | 56.7 | 57.2 | 60.5 |
| weak rep | 62.9 | 55.6 | 36.2 | 42.4 | 54.5 | 53.7 | 62.1 |
| strong rep | 74.9 | 81.0 | 76.7 | 68.1 | 74.9 | 83.0 | 74.9 |
###########
It is also important to have a wider base for comparisons. We know that 77.8 percent of respondents were stable between 2008 and 2012 with regards to three category party id, but is that a high number? Is it a low number? It’s more stable than 7 category ideology, but given the structure of how the question is asked - first respondents are asked their three category party id, then they are asked a follow-up question about whether they’re a strong or weak partisan or if they lean towards one of the two major parties as an independent - that must be the case. To better understand the context of how stable different attributes are, we examine additional variables.
We now include variables that both have sufficient PDL data to compare multiple election years and have importance in the literature on party id. By increasing the number of presidential year to presidential year comparisons, we account for the fact that some four year periods would be expected to have more or less stability in party id. Additional variables give the aforementioned context.
| Years | pid3 | pid7 | ideo5 | birthyear | relig | bornagain | prezapproval |
|---|---|---|---|---|---|---|---|
| 2004 - 2006 | 78.5 | 63.9 | 82.9 | 98.9 | 0.0 | 0.0 | 0.0 |
| 2006 - 2008 | 76.9 | 64.9 | 62.9 | 86.2 | 0.0 | 0.0 | 71.8 |
| 2008 - 2010 | 78.2 | 67.9 | 62.7 | 86.8 | 27.0 | 91.5 | 13.2 |
| 2010 - 2012 | 82.6 | 73.3 | 66.7 | 95.9 | 79.7 | 90.5 | 71.5 |
| 2012 - 2014 | 83.3 | 72.9 | 70.0 | 95.6 | 81.1 | 92.0 | 69.4 |
| 2014 - 2016 | 80.8 | 69.7 | 69.6 | 94.0 | 79.1 | 91.3 | 66.3 |
| 2016 - 2018 | 83.6 | 72.7 | 70.3 | 93.0 | 80.6 | 92.8 | 10.4 |
| 2018 - 2020 | 82.1 | 72.0 | 68.5 | 92.6 | 76.8 | 91.8 | 75.9 |
The results of our comparisons are in Table 23. We include respondent year of birth as a sort of theoretical maximum possible consistency. Measurement error is a common issue for survey research and our approach may exacerbate some issues given that we are solely selecting the last observation of an election year as the “true” response for respondents with multiple responses to a single question in a single year.
Partisan id is quite stable across all shown presidential election to presidential election panel years. The comparison between party id and religion is likely well made given the similar consistency results between three category party id and the PDL question measuring respondent religion. That said, born again Christianity seem to be stable at a rate comparable to birth year. At its peak, birth year is more consistent, but there are also periods in which birth year stability drops to the mid 80s.
As we would expect, approval of the president is incredibly stable between multiple years in which the president is the same, such as 2012 and 2016. In years in which the party of the president switches, consistency is incredibly low, lower in fact than simply subtracting consistency between years in which the president is the same from 100.
When Does Change Happen?
Consistent with smaller ANES panels earlier in time, there is clear evidence of partisan change on the scale of four year presidential election periods. Having shown that change in partisan identification does occur over time, the question of what causes change is of paramount importance.
A common thread discussed in the literature is the rationalist perspective that voters might change party id when their party no longer represents some truly held beliefs or values. An implication of the rationalist theory is that those whose expressed ideology does not match their expressed partisanship should be likely to change one or the other, as one of the two presumably cannot represent the respondent’s truly held beliefs or values. There are a few assumptions inherent in this approach, firstly that there exists “true” preferences to begin with, and secondly that one of expressed ideology or expressed partisanship reflects said preferences.
We treat ideological liberals with any level of Republican partisanship and ideological conservatives with any level of Democratic partisanship as having conflicting ideology and partisanship to test the implications of the rationalist perspective. Though historically there have liberal Republicans and conservative Democrats, in modern politics Democrat and liberal, as well as Republican and conservative, are essentially synonyms.
For the purposes of this project, partisanship matches ideology when the partisanship of the respondent is democratic and their ideology is very liberal or liberal, partisanship is conservative and ideology is conservative or very conservative, or when partisanship is independent, not sure, or other and ideology is moderate or not sure. We start by checking what percentage of three category observations have a non-matching ideology observation by the same respondent from the previous time they were asked. Overall, we find 5.4 percent of the time the previous ideology observation does not match a respondent’s current three category party id.
| Dem | Ind | Rep | |
|---|---|---|---|
| Lib | 92.5 | 78.0 | 45.4 |
| Mod | 82.1 | 82.7 | 65.1 |
| Con | 68.8 | 74.9 | 80.7 |
Ultimately, what we are interested in is if respondents who have potentially contravening party identity and ideology are more likely to change their party ideology in the future. Table 24 shows this by showing what proportion of respondents who have inconsistent party id and ideology measurements in 2008 then go on to change their party id in 2012. The row and the column are a respondent’s 2008 ideology and partisanship. The value in each cell is the percentage of respondents for that 2008 party id - ideology pair who had the same party id in 2012.
This makes it clear that conservative Democrats and liberal Republicans are much more likely to change partisanship by the next election cycle. In fact, ideologically moderate partisans are also more likely to change their partisanship by 2012 than liberal Democrats and conservative Republicans are. We do also adopt a similar simplifying technique as we did in Table 23 where the final observation in each respondent-year is selected for comparison.
We also reverse the analysis to compare the degree to which party id - ideology inconsistency is correlated with a change in party id versus a change in ideology. As our ideology variable is 5 category, while our party id is 3 category, we collapse the ideology variable into three categories to avoid a difference in question structure biasing our results.
`summarise()` has grouped output by 'ideo3_2008'. You can override using the
`.groups` argument.| Dem | Ind | Rep | |
|---|---|---|---|
| Lib | 88.8 | 76.8 | 46.2 |
| Mod | 64.4 | 71.0 | 54.8 |
| Con | 56.6 | 86.3 | 93.9 |
In Table 25 we can see what percentage of respondents with each combination of three category party id and ideology were stable in their ideology between 2008 and 2012. Where Table 24 measured party id stability, this table measures ideological stability. Liberal respondents were equally stable in both party id and ideology. Moderate Democrats were very unstable in ideology but stable in party id, while moderate Republicans were very unstable in both ideology and party id. conservative Democrats are slightly less stable in ideology than party id, while moderate and liberal Democrats are actually slightly more stable in ideology than party id.
That moderates are less stable in ideology than party id may indicate that moderate functions as less of an identity than liberal or conservative ideology or partisanship.
For liberal respondents, party id and ideology have relatively similar 2008-2012 stabilities across all partisanships. Conservative respondents, on the other hand, are quite a bit more stable with regards to ideology compared to party id. It is not completely clear why this is the case, but it the rise of the Tea Party movement during this period may be affecting things.
| Dem | Ind+ | Rep | |
|---|---|---|---|
| Lib | 1.8 | 0.9 | -0.3 |
| Mod | -3.5 | 0.6 | -1.5 |
| Con | -0.4 | 2.2 | 0.2 |
Though stability is a primary quantity of interest, context is extremely useful here. Knowing in which direction things are changing contextualizes Table 25 and Table 24. Table 26 shows the direction each party id - ideology category is moving in from 2008 to 2012. The value in each cell is the overall percentage of respondents with that specific party id - ideology combination in 2012 minus the percentage with that combination in 2008.
We can see that the respondents became more Independent in partisanship, yet less moderate in ideology between 2008 and 2012. These overall net changes are much smaller than the rate at which individuals were changing, regardless of ideology - party id consistency in Table 25 and Table 24. This does suggest that although on the aggregate respondents are becoming less likely to identify as having cross cutting ideology and party id, some of those who are becoming more consistent are being “replaced” by respondents becoming less consistent.
Momentum
Understanding the dynamics of how respondents change their party identification and ideology are important, but there may also be underlying traits that cause some Americans to be more or less amenable to shifting their party id. Those traits could come in many forms - demographic, behavioral, policy oriented, or other - but so long as they are stable over time, they would all imply one thing: that those who have changed their partisanship once are more likely to do so again in the future. This may also be explained by respondents who make dramatic changes in party id slowly, in multiple steps. The behavioral theory that some respondents are simply less stable in party id is directly testable using our data.
| Did Pid3 Change 2008 - 2012 | N | % Changing Pid3 2012 - 2016 |
|---|---|---|
| FALSE | 23656 | 10.2 |
| TRUE | 4992 | 27.4 |
In Table 27 we subset the sample into respondents with three category pid measurements in 2008, 2012, and 2016. The two rows in the data represent those whose last party id observation in 2008 was the same as their last party id observation in 2012 and those with differing party id observations in 2008 and 2012.
The sample is then split into respondents whose last observation in 2008 was different from its last observation in 2012 versus those whose final 2008 and 2012 observations were the same. Splitting the sample this way allows us to analyze the stability of respondents who have shifted in the past versus those that have not.
For each of those two categories, we calculate the proportion of respondents who saw a change in their final three category pid observation between 2012 and 2016. If there exists some underlying personal quality that affects a person’s likelihood or openness to switching parties, respondents who swapped party id between 2008 and 2012 are also more likely to swap between 2012 and 2016. Ultimately, this is what we observe. Respondents who are stable between 2008 and 2012 swap in 2016 at only a rate of 10.3%, while respondents who are unstable, swap at a rate of 27.6%. Stability, however, is much more common than instability when comparing presidential election cycles.
Party identification is measured in both three category and seven category responses however. The benefit of seven category response is that it elucidates smaller changes in party identification. These smaller changes might herald oncoming changes between the larger categories.
| Pid7 Change 2008 - 2012 | N | % Pid7 Change 2012 - 2016 |
|---|---|---|
| FALSE | 20245 | 19.5 |
| TRUE | 8358 | 42.2 |
As previously discussed, the rate of instability in seven category party id must be greater than that of three category party id, which is reflected in Table 28. The rate of instability among respondents who did not change between 2008 and 2012 is ten percentage points higher in seven category party id when compared to three category party id. When comparing respondents who DID change between 2008 and 2012, there is a fifteen percentage point difference. The fact that this is five points larger than the previous category may indicate that respondents move through categories with “momentum.”
| PID72008_2012 | dem | ind/+ | rep |
|---|---|---|---|
| dem - dem | 91% | 7% | 2% |
| dem - ind/+ | 20% | 73% | 7% |
| dem - rep | 19% | 16% | 65% |
| ind/+ - dem | 73% | 26% | 1% |
| ind/+ - ind/+ | 5% | 88% | 6% |
| ind/+ - rep | 3% | 20% | 78% |
| rep - dem | 79% | 15% | 6% |
| rep - ind/+ | 5% | 71% | 24% |
| rep - rep | 1% | 11% | 88% |
Uniquely, this dataset allows us to examine how respondents’ party id evolves over multiple election cycles. We show the evolution of party id over three election cycles in Table 29. In Table 2 we visualized 2012 three category party id as a function of 2008 seven category party id. We now extend this approach by visualizing 2016 seven category party id as a function of the combination of 2012 and 2008 three category party id.
The rows represent the combination of 3 category party id in 2008 and 2012. The column is the 3 category party id in 2016, and the cell is the mean stability from 2012 to 2016. We can see that there exists a certain amount of “memory” when it comes to 2012 to 2016 stability. Respondents who identified as Democrats in 2008 and 2012 were more stable than respondents who identified as Democrats only in 2012 and not in 2008. This is also true of Independents and Republicans. The magnitude of this difference is between 10 and 15 points.
| strong dem | weak dem | lean dem | ind/+ | lean rep | weak rep | strong rep | |
|---|---|---|---|---|---|---|---|
| strong dem - strong dem | 88% | 7% | 2% | 2% | 0% | 0% | 0% |
| strong dem - weak dem | 23% | 57% | 5% | 10% | 1% | 3% | 2% |
| strong dem - lean dem | 27% | 5% | 41% | 26% | 1% | 0% | 0% |
| strong dem - ind/+ | 6% | 18% | 2% | 59% | 9% | 4% | 2% |
| strong dem - lean rep | 2% | 9% | 2% | 9% | 43% | 13% | 23% |
| strong dem - weak rep | 14% | 14% | 5% | 16% | 12% | 32% | 7% |
| strong dem - strong rep | 26% | 2% | 4% | 8% | 8% | 1% | 51% |
| weak dem - strong dem | 75% | 19% | 3% | 2% | NA | 0% | NA |
| weak dem - weak dem | 11% | 70% | 4% | 7% | 1% | 4% | 1% |
| weak dem - lean dem | 4% | 15% | 60% | 16% | 3% | 1% | 0% |
| weak dem - ind/+ | 0% | 6% | 11% | 65% | 9% | 6% | 3% |
| weak dem - lean rep | NA | 1% | 1% | 10% | 51% | 20% | 17% |
| weak dem - weak rep | 2% | 8% | NA | 4% | 5% | 61% | 20% |
| weak dem - strong rep | 4% | 9% | NA | NA | 15% | 3% | 70% |
| lean dem - strong dem | 72% | 7% | 15% | 6% | NA | NA | 0% |
| lean dem - weak dem | 16% | 61% | 18% | 4% | 0% | 0% | 1% |
| lean dem - lean dem | 8% | 7% | 69% | 15% | 1% | 0% | NA |
| lean dem - ind/+ | 12% | 2% | 15% | 61% | 8% | 1% | 1% |
| lean dem - lean rep | NA | 0% | 5% | 24% | 42% | 18% | 11% |
| lean dem - weak rep | 0% | 2% | NA | 4% | 3% | 64% | 27% |
| lean dem - strong rep | 27% | NA | NA | 20% | 5% | 11% | 36% |
| ind/+ - strong dem | 58% | 11% | 12% | 16% | 2% | NA | 1% |
| ind/+ - weak dem | 10% | 38% | 13% | 34% | 3% | 2% | NA |
| ind/+ - lean dem | 10% | 8% | 57% | 23% | 1% | 0% | 0% |
| ind/+ - ind/+ | 0% | 2% | 6% | 80% | 7% | 2% | 3% |
| ind/+ - lean rep | NA | 0% | 1% | 27% | 51% | 7% | 14% |
| ind/+ - weak rep | NA | 0% | NA | 14% | 11% | 59% | 15% |
| ind/+ - strong rep | 18% | 1% | NA | 7% | 11% | 3% | 60% |
| lean rep - strong dem | 71% | 19% | NA | 10% | NA | NA | NA |
| lean rep - weak dem | 23% | 53% | 10% | 7% | 6% | 2% | NA |
| lean rep - lean dem | 6% | 12% | 48% | 31% | 3% | NA | NA |
| lean rep - ind/+ | 0% | 1% | 12% | 61% | 21% | 2% | 3% |
| lean rep - lean rep | 0% | 0% | 0% | 17% | 70% | 7% | 6% |
| lean rep - weak rep | NA | NA | 2% | 8% | 27% | 50% | 14% |
| lean rep - strong rep | 0% | NA | 0% | 2% | 19% | 17% | 62% |
| weak rep - strong dem | 49% | 22% | 3% | 5% | 11% | 5% | 4% |
| weak rep - weak dem | 21% | 52% | 6% | 4% | 11% | 7% | NA |
| weak rep - lean dem | 5% | 27% | 32% | 28% | 2% | 7% | NA |
| weak rep - ind/+ | 7% | 3% | 12% | 47% | 14% | 14% | 2% |
| weak rep - lean rep | NA | NA | 1% | 16% | 60% | 17% | 6% |
| weak rep - weak rep | 1% | 1% | 2% | 5% | 9% | 72% | 10% |
| weak rep - strong rep | NA | 1% | NA | 3% | 10% | 22% | 65% |
| strong rep - strong dem | 79% | 1% | NA | NA | 8% | 4% | 7% |
| strong rep - weak dem | 36% | 48% | 3% | 2% | 2% | 8% | NA |
| strong rep - lean dem | 25% | 1% | 46% | 14% | 4% | 10% | NA |
| strong rep - ind/+ | 2% | NA | 6% | 50% | 18% | 7% | 18% |
| strong rep - lean rep | 0% | 0% | 0% | 14% | 47% | 8% | 31% |
| strong rep - weak rep | 1% | 1% | 0% | 5% | 10% | 57% | 26% |
| strong rep - strong rep | 0% | 0% | 0% | 2% | 7% | 10% | 81% |
In Table 30, we repeat the previous table using seven category party id instead of three. As expected, and similarly to the three category case, once a respondent has selected the same party id two election cycles in a row it becomes even more likely that they choose it a third time. A respondent whose last observation was “strong democrat” in both 2008 and 2012 selects strong democrat as their last observation in 2016 88 percent of the time. This is substantially higher than respondents who were strong democrats in 2012 but something else in 2008. Respondents who were weak democrats in 2008 and strong democrats in 2012 remained strong democrats in 2016 at a 75 percent clip, which is nearly 15 percentage points less than those who were strong democrats in 2008.
The same pattern can be seen among Republicans. Respondents who were strong republicans in 2008 and 2012 remained strong Republicans at an 81 percent rate. Respondents who were weak republicans in 2008 but strong Republicans in 2012 remained strong Republicans at a 66 percent rate, which is an identical 15 percentage point difference as with Democrats.
Additionally, neither Table 7 nor Table 30 show substantial evidence that respondents move erratically around the party id space. Movement from democrat to republican is itself rare. Movement through independent categories is a more likely way to go from one end to the other. Also, the rare respondents who do go from one side to another can sometimes exhibit backsliding, returning to their original positions at the end of the three election cycle period. This can be seen in strong democrats in 2008 who were republican in 2012, especially weak republicans.
There are also numerous three year combinations of seven category party id with zero data in our dataset, for instance there are no respondents who were weak Republicans in 2008, strong Republicans in 2012, and strong Democrats in 2016.