Counter-Intuitive Findings on Affect and Ideology Likely Reflect Collider Bias: Commentary on Turner-Zwinkels et al., 2025

A large cross-disciplinary literature on affective polarization, i.e., mutual disliking between political groups (Iyengar et al., 2012, 2019), has developed in recent years, propelled by the observation that affective polarization has increased in the US in recent decades, a trend not widely seen in other developed nations (Boxell et al., 2022). Disagreement persists over the origins of affective polarization. Some scholars emphasize the role of social identity (e.g., Dias & Lelkes, 2022), arguing that partisans cherish their identities as supporters of a particular party, and this identification “triggers both positive feelings for the in group and negative evaluations of the out group” (Iyengar et al., 2019, p.130), in line with classic Social Identity Theory (e.g., Tajfel & Turner, 1979). From this perspective, US affective polarization has increased due to sorting, with liberals becoming more likely to identify as Democrats and conservatives as Republicans (Mason, 2015, 2018a, 2018b).

An alternative view emphasises the role of substantive concerns, in particular policy disagreements; from this perspective, partisans dislike each other because they sincerely believe the policies their opponents support transgress moral principles or have undesirable practical consequences (Bougher, 2017; Orr et al., 2023; Orr & Huber, 2020). Indeed, many studies find that people prefer citizens, politicians, and parties when they perceive them to be similar to themselves in terms of ideology and policy preferences (Algara & Zur, 2023; Bougher, 2017; Dias & Lelkes, 2022; Druckman et al., 2022; Lelkes, 2021; Orr et al., 2023; Orr & Huber, 2020; Rogowski & Sutherland, 2016; Webster & Abramowitz, 2017). Notably, supporters of the identity account do not dispute these effects, rather, they suggest they occur because sharing policy preferences implies shared partisanship (Dias & Lelkes, 2022).

However, recent findings from Turner-Zwinkels et al. (2025) conflict with these results¹. They compared, in two large-N cross-national cross-sectional studies, how ideologically similar people are to other citizens who support the same party as them, and how much they like that party². In one study of eight European nations (N = 4,246), they found that similarity and liking are negatively correlated – the more similar people are to their co-partisans the less they like the party – and in a second study of 42 elections from 36 nations across the World (N = 30,053), they found no relationship.

These results are extremely surprising given prior literature, as experimental studies have shown that manipulating perceptions of the ideology of in-party politicians and citizens causes people to like them more when they are perceived to be more similar to themselves (Lelkes, 2021; Orr et al., 2023; Orr & Huber, 2020). In fact, these effects are quite large – Lelkes (2021) found that people with ideologically “extreme” views rated in-party politicians whose policy views were as “extreme” as their own 36 points higher than moderate politicians on a 0-100 feelings thermometer scale, and Orr et al. (2023) found that partisans rate co-partisans who share the same policy position as them 16 points higher than co-partisans who have positions they disagree with. Indeed, Turner-Zwinkels et al. originally hypothesized that they too would find positive correlations.

In their original article, Turner-Zwinkels et al. tentatively suggested that optimal distinctiveness theory (Brewer, 1991; Leonardelli et al., 2010), which proposes that people seek to balance cohering with a group against being distinct from other in-group members, might explain their negative correlation:

However, while optimal distinctiveness theory might suggest that people like the state of affairs of being dissimilar to the rest of their party, it does not seem to follow that it would affect their liking of the party itself. Moreover, if this explanation were true, it is unclear why it would not have negatively biased affective ratings in the experimental literature where participants prefer more-similar in-party targets (Dias & Lelkes, 2022; Lelkes, 2021; Orr et al., 2023; Orr & Huber, 2020). Overall, this explanation seems unsatisfactory.

In this article, I argue that Turner-Zwinkels et al.’s findings can be best reconciled with prior experimental findings not by positing that there is some genuine difference in the causal mechanism that relates liking to ideological similarity when judging co-partisans compared to out-partisans, nor that prior methodologies were flawed or prior theorising mistaken. Rather, I suggest Turner-Zwinkels et al.’s findings are likely caused by collider bias. I support my argument with a re-analysis of the original data and through simulations (all code is available via https://osf.io/skqub).

Collider Bias

Numerous guides to collider bias exist elsewhere—see Pearl and Mackenzie (2018), Cinelli et al. (2024), and Lee et al. (2019) for brief overviews, Schneider (2020) for a review of examples from economic history, and numerous cases from medical science (Akimova et al., 2021; Cole et al., 2010; Hernán & Monge, 2023; Weiskopf et al., 2023). Additionally, a blogpost by Julia Rohrer (2017) discusses several highly accessible examples. Here, I will try to offer a brief, but intuitive explanation of what collider bias is and why it might matter in this case.

A collider is a variable whose value is influenced by two causally antecedent variables—see Figure 1 for a diagram. The basic problem caused by collider bias is that if we control for a collider while attempting to measure the relationship between these two causally antecedent variables (or any variables they causally influence or are influenced by), we get a distorted estimate. Colliders can be contrasted with confounders, which are variables that also causally influence our dependent variable when we want to estimate the effect of a different independent variable on the dependent variable (i.e., in Figure 1, Y is a confounder for our estimate of the effect of X on C), and mediators, which are intermediate variables on the causal path between our independent and dependent variables. Depending on our goals and the causal graph, controlling for confounders and mediators may or may not be appropriate (see Cinelli et al., 2024), but, except perhaps in very unusual circumstances, controlling for a collider is something we want to avoid.

Click to enlarge

Figure 1

Causal Graph Where C is a Collider for X and Y.

For example, the collider’s value may be the sum of the other two variables. Let us keep calling the collider C and the other variables X and Y. Crucially, when C is a collider of X and Y, the value of X is not independent of the value of Y when the value of C is fixed, even if X and Y are otherwise independent. Therefore, any aspect of our data collection or analytical procedure which restricts the value of C will bias estimates of their relationship, as well as estimates of the relationships between variables that correlate with X and Y – generally speaking, collider bias can occur for any pair of variables which have a path between them that contains a collider (Cinelli et al., 2024). Such a restriction occurs if we control for C in a regression (Cinelli et al., 2024), or if inclusion within our dataset is conditioned on the value for C (e.g., Schneider, 2020). This latter case is known as conditioning on a collider (also known as collider stratification bias, ascertainment bias, and Berkson’s paradox), and is the case of relevance for this article, because Turner-Zwinkels et al.’s analysis of whether the relationship between liking a party and being ideologically close to their supporters is moderated by supporting the party in question comprises two separate analyses, one where inclusion in the dataset is conditioned on the person supporting the relevant party, and one conditioned on them not.

To illustrate why conditioning on a collider is a problem, suppose C = X + Y, and we measure X, Y, and C across many cases, where X and Y are independent and have no correlation, with values in the range 0-100. If we restrict our dataset to only contain cases where C = 100, this will induce an artificial negative correlation between X and Y, because as X increases, Y must decrease in order for X and Y to sum to C: if X is 40, Y must be 60, if X is 80, Y must be 20, and so on. Thus, in this case, an artificial negative correlation occurs (and if C = X – Y, and we conditioned on, e.g., C = 50, an artificial positive correlation would be induced).

A similar bias emerges in less restrictive cases. Suppose we only specify C > 80. Again, this induces a negative bias in estimates of the correlation: if X = 20, Y must have values of 60-100, but if X = 50, Y must have values of 30-100, which will therefore tend to be lower on average, and if X = 80, Y must be in the even lower-bounded range 0-100, with an even lower average value. The lower bound on Y decreases as X increases, so as X goes up, the average value of Y will go down.

Note that the same also occurs when C is a weighted sum of X and Y, e.g., $C=w_{X}X+w_{Y}Y$. Suppose we condition on C > k. Then, for a given value of X, the lower bound on Y is $\frac{k-w_{X}X}{w_Y}$ (as $w_{X}X+w_{Y}Y>k$, so $w_{Y}Y>k-w_{X}X$, hence $Y>\frac{k-w_{X}X}{w_Y}$), which is a negative function of X. Therefore, as values of X increase, the average accompanying value of Y will tend to decrease, biasing the apparent correlation negatively.

Collider Bias in the Case of Turner-Zwinkels et al. (2025)

Conditioning on a collider may explain Turner-Zwinkels et al.’s negative and null correlations. Recall that these results come from Turner-Zwinkels et al.’s analysis of the relationship between the average similarity of a person’s political belief to the supporters of a party (‘Similarity’) and how much they like that party on a 0–10 scale (‘Liking’). To be a little more specific, Turner-Zwinkels et al. calculated ‘Similarity’ by finding the absolute mean distance between every pair of participants in each country/election-level sample across 10 (Study 1) or 9 (Study 2) items which measured policy preferences or ideological positioning (after first normalizing all variables to range from 0-1), then calculated how far each person was from each party’s supporters on average, and rescaled so that high scores indicated greater proximity and low scores further distance.

Crucially, the surprising negative and null correlations are found using cases where the participant supports the party that is the target of the Similarity and Liking ratings. Yet, supporting a party is almost certainly a collider on the path between Similarity and Liking, as the more a person likes and is ideologically close to the supporters of a party, the more likely it is they will support that party. This occurs even if the decision to support a party is not directly influenced by ideological similarity to the supporters of that party, but rather by ideological similarity to the party’s policies and/or elites, as this kind of similarity will tend to positively correlate with similarity to the party’s supporters. Furthermore, note that the door remains open to collider bias even if we think there are other causal pathways at work, e.g., from party identification to liking and similarity, or from similarity to liking directly (see Cinelli et al., 2024).

I suggest a simple model can capture this scenario, whereby people are more likely to support a party if their summed Liking and Similarity score exceeds a relatively high threshold. This would be true if people prize both Liking and Similarity (either to the party’s supporters, or to the party’s elites and policy positions) when deciding who to support, but being high on one can compensate for being a little lower on the other, i.e., I can support a party even if there is some ideological distance between us if I like them for some other reason, or overlook some reasons to dislike the party if they are ideologically close to me. This sum must be high enough to warrant them choosing that party to support over others (as participants could only indicate support for one party), and high enough for them to warrant saying they support any party at all.

This creates a similar scenario to the example above where C = X + Y, with C being a continuous latent measure of the person’s attraction towards the party, and X and Y being Similarity and Liking, but people only entering into the dataset if their C value is above the high threshold required for them to actually say they support them. Consequently, those high in Liking can ‘get away’ with having lower Similarity scores and yet still support the party (and vice versa), leading higher Liking scores to be found alongside lower Similarity scores, on average. Even if Similarity and Liking normally have a positive correlation, this induces a negative bias, pushing it downwards.

Simulations and Re-Analysis

To test whether this proposal does provide a plausible counter-explanation for Turner-Zwinkels et al.’s surprising results, I conducted proof-of-concept simulations, supported by some re-analysis of both of Turner-Zwinkels et al.’s studies, which I performed using Turner-Zwinkels et al.’s description of their methodology, following it entirely, and successfully replicating their correlation estimates (from the Correction to the original article), as well as performing some new analyses. All scripts are available in the OSF project for this article.

My first piece of re-analysis supports the general idea that this simple model is plausible, as participants’ (equally-weighted) summed Liking and Similarity scores do exceed relatively high thresholds for their in-party. In Study 1, 88.2% of participants support the party for which they have the highest summed liking and similarity, with 84.1% doing so in Study 2. Additionally, the summed liking and similarity scores for supported parties fall, relative to people’s summed liking and similarity scores for all rated parties, at the 96^th percentile on average in Study 1 and the 95^th percentile in Study 2, indicating that they exceed a high threshold.

My simulations show that with this model, a positive correlation between Similarity and Liking overall can turn null or negative when only in-party scores are analysed. I simulated 4050 electoral samples of 1000 agents each. Each agent has Similarity and Liking scores for 5 parties, and chooses to support the party with the highest sum for their Similarity score, Liking score, and the score from a random noise variable to capture, to some degree, the influence of unmodelled complexities in how people choose parties to support. Within each electoral sample, the 5000 Similarity scores are drawn at random from a normal distribution with a mean of 0.5 and a standard deviation of 1, which is also the case for the Liking scores; across simulations, I vary the correlation between Similarity and Liking from 0 to 0.4 in increments of 0.05. The noise variable is also normally distributed, with a mean of 0, and a standard deviation varied across simulations (as the ‘noise σ’ parameter), from 0 to 2 in increments of 0.25—a higher standard deviation means the agent’s choice of which party to support will be less influenced by Similarity and Liking. This creates 9 x 9 = 81 unique combinations of parameters, for each of which I generate 50 random samples, giving the 4050 electoral samples. For each sample, I calculate an “overall” correlation between Similarity and Liking across the agents’ scores for all 5 parties, which should be equal to the programmed correlation plus random error. I also calculate an “in-party” correlation between Similarity and Liking using only the scores the agents have for the party they support.

Figure 2 shows the relationship between these correlations. Each point represents the pair of correlations obtained for one electoral sample. The x-axis shows the “overall” correlation between Similarity and Liking across scores for all parties for that simulation, and the y-axis shows the “in-party” correlation across the scores the agents give to the party they support. The points fall below the diagonal, demonstrating a downward bias of in-party correlations relative to overall correlations. This bias is stronger with less noise, but equal across different levels of overall correlation.

Click to enlarge

Figure 2

The Correlation Between Liking and Similarity Across All Parties (X-Axis) and for the Parties the Agents Support (Y-Axis) Across Simulations

Note. Linear regression lines are shown for each level of noise.

In concurrence with the results of these simulations, my analysis of Turner-Zwinkels et al.’s data finds that estimates of the correlation between Liking and Similarity are substantially biased downwards when only “in-party” data is used. For Study 1, an “overall” correlation of 0.376 [0.363. 0.389] (p < .001) drops to -0.096 [-0.126, -0.066] (p < .001) for the “in-party” correlation, and for Study 2, 0.216 [0.211, 0.221] (p < .001) drops to 0.005 [-0.007, 0.016] (p = .424). Further simulations in the Supplementary Materials show that when liking and similarity scores are additionally influenced by which party the agents support, this worsens the bias.

Conclusion

The surprising negative and null relationships observed by Turner-Zwinkels et al. between liking the party you support and being ideologically similar to your co-partisans may be a statistical artefact, caused by conditioning on a collider. There is therefore probably no need to re-think existing findings, theories, or methodologies within the literature showing that people prefer those who concur with their ideology and policy positions (Algara & Zur, 2023; Bougher, 2017; Dias & Lelkes, 2022; Druckman et al., 2022; Lelkes, 2021; Orr et al., 2023; Orr & Huber, 2020; Rogowski & Sutherland, 2016; Webster & Abramowitz, 2017).

While I cannot definitively prove that collider bias explains Turner-Zwinkels et al.’s surprising results, the confluence of a) simulations which demonstrate the plausibility of this explanation in principle, and b) supporting statistical evidence from re-analyses of Turner-Zwinkels et al.’s data, makes a compelling case. Moreover, positing collider bias as the explanation provides a parsimonious account of how these results could arise when previous experimental studies suggest that ideological proximity to co-partisans boosts liking (Dias & Lelkes, 2022; Orr et al., 2023; Orr & Huber, 2020), as it avoids the need to theorise new causal mechanisms or identify methodological flaws in these studies.

This article highlights the particular danger collider bias poses for the interpretation of cross-sectional moderation analysis. Because collider bias can lead the estimated relationship between two variables to change drastically when a third variable is controlled for, even when that variable does not causally affect their relationship, it can give a misleading impression that the third variable acts as a moderator.³ This is plausibly what happened in Turner-Zwinkels et al., where the third variable was party identification, which appeared to moderate the relationship between liking a party and being ideologically close to their supporters, as the observed relationship was much stronger for parties the participants did not support than for those they did. Since the problem of collider bias is not widely appreciated in the context of moderation—indeed, Hayes’ textbook companion to the widely-used PROCESS software does not mention colliders at all (Hayes, 2022)—researchers may waste resources exploring non-existent causal mechanisms without better awareness of collider bias. Researchers should therefore determine whether their moderators could be colliders (i.e., are causally influenced by both the IV and DV), consulting existing guides for doing so (Elwert & Winship, 2014; MacKinnon & Lamp, 2021; Rohrer, 2018; Rohrer et al., 2022), and communicate explicitly about the limitations caused by possible collider bias where this cannot be ruled out, or better yet, use experimental methods where moderators can be manipulated, if possible. Reviewers too should be wary of the potential for cross-sectional moderation analyses to yield misleading results due to collider bias.

Overall, social and political psychologists should beware collider bias; it can cause artificial correlations to appear from nowhere, and for real correlations to be exaggerated, eliminated, or reversed. Therefore, when studying relationships between pairs of variables within sub-samples, such as partisans, if membership of the sub-sample could be conditional on both variables, it should be expected that cross-sectional estimates of the relationships will be distorted.