Partisan Discrimination Without Explicit Partisan Cues

Much research has demonstrated that Democrats and Republicans use information about party affiliation to discriminate against one another. However, we know little about how people gain the necessary information about other people’s partisanship to engage in discriminatory behavior. We explore whether people perceive partisanship when shown only images of faces, and whether they then use these perceptions to engage in partisan discrimination. We find that they do. Using two studies we show that the partisan perceptions people derive from seeing images of faces influence discrimination of job applicants, and propensities to engage is a wide range of social interactions. People appear to be making judgements about partisanship using only facial appearance, and are willing act on that perception. The implication of this finding is that partisan discrimination is likely widespread, and does not require the explicit communication of partisan affiliations. the prevalence of this behavior outside of the experimental setting. We are careful to note that we are not able to say what the rate of partisan assignment is in daily life, just that when prompted people appear to assign partisanship to faces. It would be quite challenging to determine the actual prevalence of these perceptions without directly asking people, but future research should explore this question.

racial discrimination (Iyengar & Westwood, 2015), the ability to conceal party identification much more easily than race suggests that the former could be a rarity compared to the latter.
Importantly, most social interactions with strangers are fleeting -we decide to let someone cut in a line, hold the door for them, or pick up an item they dropped quickly, and likely without much cognitive effort. However, if partisan discrimination is pervasive in society, it is possible our judgments about these people's partisanship influences these decisions. Of course, in these interactions, we are highly unlikely to have information about these people's partisan identities. We can, however, assess their appearance. Is partisan discrimination relegated to the subset of interactions where we are aware of others' partisanship, or do individual infer the partisanship of others quickly based on their appearance?
The question raised by this is how do people gain information about other's partisanship? Certainly, some do advertise their political affiliations through bumper stickers or yard signs, but these individuals are not the norm. If the average person goes about their day and remains agnostic about the partisanship of others unless they are explicitly informed, it suggests that people are not usually looking at those around them through the filter of partisanship. By extension, this would mean that partisan discrimination is somewhat of a rarity. On the other hand, if people do not need to be explicitly told that others are Democrats or Republicans, but they draw inferences from heuristics or stereotypes, then the door is opened to partisan-based discrimination being a more widespread phenomenon.
The focus of this paper is whether Americans use appearance to draw conclusions about the partisanship of others, and then use this perception to discriminate against them. Our first study is a face rating exercise where participants were shown a host of faces and asked whether they believed them to be Democrats, Republicans, or neither. We find that perceiving a face to be a Democrat or Republican is common, with most faces being identified as a partisan of one persuasion or the other. Knowing this, we turn to our primary focus which seeks to understand whether the partisan perceptions that people have from facial appearance matter for their discriminatory behaviors. We show that the perceived partisanship that participants have from faces are associated with discriminatory behavior in evaluating job applicants and across a number of different social interactions. The implications that follow from these findings are that partisanship is likely being perceived frequently, even from a relatively non-political stimuli, and these perceptions are used to discriminate. Reflecting on the literature demonstrating the sizeable magnitude of partisan-motivated discrimination, these findings suggest that such discrimination could be a common occurrence as people are making partisan judgements from information that is commonly on display.

Partisan Discrimination Based on Facial Appearance?
Do people ascribe partisanship to one another based on their appearance? There is good reason to suspect they do. Brunswik's Lens Model (Brunswik, 1952(Brunswik, , 1956) offers a view in which people do not require actual information to make judgements about others in interpersonal settings, but rather they use a host "cues" that they perceive in order to form judgements. A sizeable body of research has demonstrated that these cue-based judgements are used in a wide array of social settings and interactions, and with relatively high degrees of "judgmental accuracy" (see Karelaia & Hogarth, 2008) for an extensive meta-analysis). Facial cues such as Afrocentricity (Blair, Judd, Sadler, & Jenkins, 2002) and emotional expression (e.g., Tskhay & Rule, 2015) serve to drive important judgments of others.
First impressions based on facial appearance happen very quickly -in a matter of milliseconds -and are formative for a wide range of perceptions (Bar, Neta, & Linz, 2006;Todorov, Olivola, Dotsch, & Mende-Siedlecki, 2015;Willis & Todorov, 2006;Zebrowitz & Montepare, 2008). These judgements are likely to be automatic (Hassin & Trope, 2000), and are observed in children as young as 3 years old (Cogsdill, Todorov, Spelke, & Banaji, 2014). People infer a wide range of traits from these impressions -an individual's trustworthiness (Eckel, 2011;Ewing, Caulfield, Read, & Rhodes, 2015;Tingley, 2014) and reliability (Aksoy, Eckel, & Wilson, 2018) to their competence and aggressiveness (Willis & Todorov, 2006), all based on the impressions they get from faces. While many have explored different cues to see which ones are influential, our primary focus is on the outcomes produced by facial judgements. Perceived facial traits are consequential for a wide range of behaviors from promotion and hiring decisions (Ruffle & Shtudiner, 2015;Rule & Ambady, 2009) to criminal sentencing (Blair, Judd, & Chapleau, 2004;Eberhardt, Davies, Purdie-Vaughns, & Johnson, 2006;Shoemaker, South, & Lowe, 1973). To be sure, these judge ments are complex, and facial perceptions interact with other impressions that people have about the individual, but the overall takeaway is that facial cues are an important part of the calculus that people use to draw conclusions about others. Should we expect to see similar patterns emerge in the political domain? Considerable evidence suggests that we may. People use impressions drawn from faces to accurately predict who will win an election (Antonakis & Dalgas, 2009;Ballew & Todorov, 2007;Todorov, Mandisodza, Goren, & Hall, 2005), and to inform their own individual vote choices (Ahler, Citrin, Dougal, & Lenz, 2017;Berggren, Jordahl, & Poutvaara, 2017;Carpinella, Hehman, Freeman, & Johnson, 2016;Lenz & Lawson, 2011;Little, Burriss, Jones, & Roberts, 2007). These effects likely exist because people ascribe desirable leadership traits like competence to candidates based on their faces (Laustsen, 2014), or because of per ceived attractiveness (Jäckle, Metz, Wenzelburger, & König, 2020), and they may be more pronounced for low-knowledge voters (Lenz & Lawson, 2011). These patterns appear to extend from ascribing traits to political candidates, to ascribing traits to one's fellow citizens, even going as far as to view attractiveness differently depending on party identification (Easton & Holbein, 2021;Nicholson, Coe, Emory, & Song, 2016).
Especially relevant for our purposes, people appear to be able to guess the party identification of one another based on their facial appearance (Rule & Ambady, 2010;Samochowiec, Wänke, & Fiedler, 2010), especially related to emotion expression and babyfacedness. There appear to be stereotypes that people associate with individuals of each party -i.e. there is a stereotypically Republican look -that are consequential for how people evaluate others (Olivola, Sussman, Tsetsos, Kang, & Todorov, 2012;Olivola, Tingley, & Todorov, 2018). For example, those who are perceived as having a "baby face" are seen as more liberal, while those seen as having mature faces are perceived as more conservative (Tal-Or, Bivas, & Sagron, 2019), faces viewed as being powerful are more likely to be seen as Republican while those seen as warm are perceived as being Democratic (Rule & Ambady, 2010), and faces seen as happy are more likely to be categorized as liberal and those seen as angry being categorized as conservative (Tskhay & Rule, 2015). While the accuracy of these partisan guesses is somewhat limited (Herrmann & Shikano, 2016;Olivola & Todorov, 2010), the fact that the behavior is widespread is most important for our purposes.
The view we are left with is one in which individuals draw a host of conclusions from people's faces, and these con clusions include perceptions of others' political beliefs. Seeing someone that the individual perceives to be a Republican is likely to foster feelings of in-group favoritism for Republicans, or out-group distain for Democrats (Iyengar, Sood, & Lelkes, 2012). Given that, we expect that these political perceptions derived from faces will be correlated with partisan discriminatory behavior in a host of settings, both political and apolitical (e.g., Engelhardt & Utych, 2020;Iyengar & Westwood, 2015;Lee, 2020;Shafranek, 2021). We focus on two apolitical venues -employment and casual interpersonal encounters. The employment arena has been previously studied by those explicitly priming partisanship (e.g., Gift & Gift, 2015), and is one where highly consequential decisions are made.
H 1 : Job applicants who are perceived as being co-partisans should be evaluated more favorably than job applicants who are perceived as being non-co-partisans.
Less studied are casual daily encounters. Should we expect people to be thinking about partisanship, even in these interactions? There is reason to suspect so. Given that partisan assessments of the world are likely to be quick and subconscious, and are applied in settings that are far-removed from real life such as dictator and trust games (Iyengar & Westwood, 2015), they may always be a filter that is being applied to one's surroundings. These identities are pervasive and powerful enough that we see extensive social distancing along party in and out-group lines, even among those who are otherwise politically conflicted (Mason, 2018b). Further, as geographic polarization increases (Johnston, Manley, Jones, & Rohla, 2020) and the red state-blue state understanding of the political landscape has become widespread, it is likely that people are using this lens to understand (and make guesses about) the people they encounter in these spaces. Simply knowing where one is located should inform the individual about whether they are in friendly or hostile terrain with respect to partisanship, and likely serves to influence the perceptions we form. In sum, if people use partisan identity as a subconscious filter to understand the world around them, and they are increasingly accustomed to living in spaces that have a well-known political character, it is likely that a partisan screening is being frequently used by individuals even during mundane daily interactions.
H 2 : People who are perceived as being co-partisans should be treated more favorably in casual interper sonal interactions compared to those who are perceived to be non-co-partisans.
These expectations do not rest on the perceptions being accurate, just on whether the individual ascribes partisanship to a face, or not. Further, we don't expect that all individuals will engage in this behavior equally. People who have stronger party identification are more likely to engage in discriminatory behavior than people with weaker party identification (Iyengar & Westwood, 2015), so we expect that they will be the most likely to use faces to ascribe party identification to people and engage in in-group favoritism or out-group punishment.
H 3 : Strong partisans should engage in discriminatory behaviors based on perceived partisanship more than weaker partisans.

Study 1: Do People Infer Partisanship From Faces?
First, we examine the extent to which individuals infer partisanship from simply seeing someone's face. To examine this, we collected data from a sample of 245 participants using Lucid Theorem survey sampling in July, 2020. 1 The Lucid Theorem is an online platform for delivering opt-in surveys that use quota sampling to ensure that the sample is generally representative on demographic factors. Lucid samples also appear to be representative of the public at large on attitudinal metrics as well, making them advantageous compared to other venues such as Amazon Mechanical Turk (Coppock & McClellan, 2019). 2 The sample in Study 1 is 51.3% female, 69.0% white, 49.0% have a Bachelor's Degree or higher, and has a mean age of 43.7. Of those who identified as partisans (Independent leaners are coded as partisans), 41.7% were Democrats, 8.5% were Independents, and 49.8% as Republicans. Respondents were required to be living in the US. In this study, participants were asked to rate 40 different AI generated faces 3 , selected by the authors. We are not trying to understand what facial features predict certain partisan categorizations, but rather are trying to understand the extent to which people ascribe partisanship across a wide range of faces. As a result, faces were selected to encompass a wide range of features such as hair length, hair color, facial hair, and facial structure. Participants were randomly assigned to rate faces of either White or Hispanic/Latino individuals 4 , and each participant rated 20 male faces, and 20 female faces, presented in random order. With this approach, each face is rated by roughly 100 participants -since participants only rated the 40 white faces, or the 40 Hispanic/Latino faces. As such, a total of 80 faces were rated in this study. Prior to the set of face ratings, participants were given a practice rating task, and then told the following instructions: "Thank you for practicing a facial rating! Please work rapidly but carefully on the following ratings. We want to know your first reactions to these individuals. " In order to disguise the true purpose of the study, they were asked to rate faces on 3 dimensions -attractiveness, trustworthiness, and, key for our purposes, partisanship. Partisanship of each face was rated on a scale from 1 (Very likely a Republican) to 7 (Very likely a Democrat), with a neutral midpoint option of Equally likely Republican and Democrat. 5 1) Research follows the APSA principles for human subjects research. Please see the Appendix F in the Supplementary Materials for greater detail.
2) Because Lucid contracts directly with participants we do not know how much they were paid for their participation.
3) Faces were retrieved from https://generated.photos/faces. While these faces are AI generated, they resemble human faces well, to the point where we have minimal concerns that the "uncanny valley" effect would make individuals realize these faces are not actual people. We have presented these faces in the Appendix (see Supplementary Materials). Additionally, perhaps a strong design would have varied the same face using software -this is a limitation of this study, and we caution that results reflect inferences about partisanship from appearance generally, rather than facial features. 4) Based on classification provided by https://generated.photos/faces. We chose to not have participants rate Black faces, as we believe that racial cue would provide a very strong signal that someone is a Democrat. While a majority of Latinos identify as Democratic, the group is much more heterogeneous than those who are Black, with polls generally showing the percent of Latinos who identify as Democrats in the low to mid-60% range (Pew Research Center, 2020), with an even smaller share of Latino males identifying as Democratic. This partisan heterogeneity means that the racial cue is weaker and people could be using pictures of faces in more nuanced ways to draw conclusions.
5) It is possible that asking participants to rate faces on partisanship overstates the prevalence of this behavior outside of the experimental setting. We are careful to note that we are not able to say what the rate of partisan assignment is in daily life, just that when prompted people appear to assign partisanship to faces. It would be quite challenging to determine the actual prevalence of these perceptions without directly asking people, but future research should explore this question.
We first examine average responses of partisanship across all 10 faces of each gender and ethnicity. These results are presented in Table 1 below. Generally, we see that a large portion of the sample was willing to infer partisanship from faces -while a bit over 1/3 of all responses rated the partisanship of the faces at the midpoint, nearly 2/3 rated individuals as either Republican or Democratic. 6 These patterns were quite similar for White male and female faces, and Latino male faces, which were all slightly more likely to be rated as Republicans rather than Democrats. However, we see that, for Latina female faces, this pattern flips, with individuals more likely to rate them as Democrats. There is also variation between faces within groups, though this variation is quite small. The most Republican White male face was rated, on average, 3.44 on the seven-point scale, while the most Democratic White male face was rated 4.13. For White female faces, the most Republican face was rated as 3.49, while the most Democratic face was rated as 4.09. The most Republican Latino male was rated at 3.49, and the most Democratic at 4.15. The most Republican Latina female was rated at 3.89, and the most Democratic at 4.28. These face ratings are also correlated with ratings of trustworthiness and attractiveness. Both Republicans and Democrats are more likely to rate someone they view as co-partisan as more trustworthy. Democrats are more likely to rate co-partisans as more attractive, but Republicans are not. These results are presented in Appendix E (see Supplementary Materials).
This suggests that perceptions of partisanship through facial appearance are nuanced -there is not necessarily a "Republican" or "Democratic" face, but these perceptions vary between individuals. Looking across the categories, it is clear that factors such as race and gender are used to some extent to infer partisanship (especially for Latinas), but race and gender are not determinative. Among whites, we see a very similar distribution of perceptions between men and women, and the Latino male perceptions are also strikingly similar to the white male and female perceptions. Identifying someone as a white male, white female, or Latino male tells us relatively little about how people will perceive their party affiliation. The suggestion that follows is that partisan perceptions are a function of much more than just race and gender. Given that individuals can assign partisanship to another person with relatively limited information, we next examine whether or not individuals discriminate against others based on these assumptions. 6) There is generally strong consensus among these face ratings, with relatively high Cronbach's α (0.94 for white men, 0.94 for white women, 0.93 for Latino men, and 0.94 for Latina women).

Method (Study 2a)
2000 participants 7 were recruited from Lucid Theorem survey sampling in August, 2020 to complete this study. 8 Lucid recruits a nationally diverse sample, and this sample was indeed diverse, with participant ages ranging from 18 to 90, with a mean of 44.25. Half the sample identified as female, and half as male, while 65.8% identified as white. 37% of the sample has obtained a Bachelor's degree or higher, while 44.5% identify as Democrats, 38.5% as Republicans, and 17% as pure independents. Respondents were required to be living in the US.
After answering a series of demographic questions, participants were randomly assigned to receive a picture of one of four faces -two of which were men, and two of which were women. We caution, of course, that this is an observational, rather than experimental, study. Because the faces are not consistently rated as Democratic or Republican by respondents, we believe this manipulation is not sufficient to cue partisanship, based on results from Study 1. We chose to vary the faces to avoid any idiosyncratic factors about a particular face from driving partisan judgmentswhile we don't have expectations that this would be the case, varying the face that individuals are rating helps to alleviate that concern. 9 We still caution that these findings are observational and correlational in nature. 10 All faces were of White individuals that had the highest proportion rated as Republican or Democratic from each gender in Study 1. They were also presented a resume 11 , which did not vary between the four conditions, and told they were the manager of a firm hiring for the position of assistant manager. Importantly, they were asked to categorize what political party they thought this person was a member of -Republican, Democratic, or Neither/Independent. 12 This leads to the creation of our key variable of interest -co-partisanship -which is coded as 1 if Republican respondents said the individual was a Republican, or if Democratic respondents said the individual was a Democrat, and 0 otherwise. Since pure independents could not possibly code someone as co-partisan, they are excluded from analysis. However, we do include leaning partisans (those who said they were independent, but felt closer to either the Democratic or Republican party) as partisans in these analyses. Importantly, this partisan rating was conducted after the respondents answered all dependent variable questions. This was done to avoid "cueing" individuals to partisanship prior to their evaluations of the individual.
Of course, this strategy (nor any strategy we can think of, given the variation in ratings of partisanship of faces leads to a general lack of ability to run a truly randomized experiment) does not rule out alternative explanations. As such, we see this as an important but correlational step in determining how partisanship inferences, without explicit partisan cues, are associated with attitudes. In this instance, individuals could be spontaneously inferring partisanship, and using that inference to inform their judgments of a person. They could also be making simple judgments about the face of the person, and using that feeling of positivity or negativity towards the face to increase their likelihood of rating them more favorably and rating them as a co-partisan. Between these possibilities, we believe it is more likely that perceived 7) By excluding pure independents, we end up with a sample size of 1638. We chose a large sample to have confidence that any null results were a product of actual null effects, rather than small effects. This sample size allows us to detect a correlation of roughly .07, considerably smaller than effects typically found in the literature. 8) Although some have raised concerns about rising inattentiveness from Lucid participants during the COVID-19 pandemic (Aronow, Kalla, Orr, & Ternovski, 2020), this inattentiveness is not likely to change how participants are reacting to experimental treatments, though it may reduce the size of effects that we are able to detect (Peyton, Huber, & Coppock, 2020). 9) Of course, it may have made more sense to make the face presentation more random -rather than focusing on only 4 faces, we could have included all faces from Study 1, or even more faces available from the database. In retrospect, the decision to limit this study to 4 faces provides a limitation, but we have no reason to believe, based on the results from Study 1, that individuals are more or less likely to assign partisanship to these particular faces than any face generally.
10) Note that this design was not pre-registered. However, our a prioi expectations require moderation by partisan identity -that is, we expected that Demo crats would favor individuals they perceive as Democrats, and vice versa for Republicans, a claim that is not especially controversial in the existing literature on partisan discrimination. Analyses using all dependent variables collected in the studies are presented in text or in the Appendix (see Supplementary Materials), for greater research transparency. 11) Images of the faces and the text of the resume are available in the Appendix (see Supplementary Materials). 12) Note that the "Neither / Independent" category is included to not force a choice on partisanship. partisanship is being used to inform traits, than traits being used to form perceptions of partisanship. As a social identity and perceptual screen through which people evaluate the world around them, partisanship is likely to be a much more formative factor in driving perceptions than vice versa. This is similar to our understanding that partisanship forming issue attitudes is far more common than issue attitudes shaping partisanship. Further, much research has demonstrated that other social identities such as race are used to from trait perceptions. While we do not have evidence to adjudicate between these possibilities (it would be very difficult to do in any research setting), our theoretical understanding of partisanship and social identities suggests that the most likely explanation is that perceptions of party affiliation are the driving factor.
We see variation in ratings based on this measure. 27.1% categorized the individual as Republican, 38.3% as Demo cratic, and 34.6% as neither party. These ratings did vary a bit by treatment, as both the male and female Democratic faces were less likely to be categorized Republican than the Republican male face, but there remains considerable variation, with the lowest partisan rating coming for the male Democratic face being categorized as Republican by 20.2% of respondents. Further, there does appear to be a slight tendency of people to project their own partisanship onto the faces they are presented with. 49.9% judged the person to be a co-partisan, with 18.1% believing that they were a member of the out-party. This is not surprising, given that previous research finds that one's own partisanship influences how they perceive the partisanship of others (Wilson & Rule, 2014).
For the key dependent variables in this study, participants were asked, on a scale of 0-100, how likely they were to hire this person for the position, and what percentile of all applicants they would rate this person. 13 They were also asked to rate the person, on a seven-point scale, on multiple traits -competence, caring, experience, leadership, honesty, trustworthiness, skill, and niceness. 14 In all analyses, we use OLS regression to predict the effects of the key independent variable, perceived copartisanship, on each dependent variable. We include demographic controls for age, education, race, gender, and strength of partisanship in all analyses, along with dummy variables indicating which face they were presented with. 15

Results and Discussion
First, we examine how perceived co-partisanship influences attitudes toward quality of applicants. These results are presented in Table 2, and show that co-partisanship does impact these evaluations, with individuals rating perceived co-partisans over 2.5 points higher on = likelihood of hiring, and 3 points on applicant rank. Note that applicants are rated generally quite highly, with a mean of 65.37 for likelihood of hiring, and 65.55 for percentile.
Additionally, we analyze the effects of co-partisanship on trait evaluations. We scale together trait evaluations for competence traits (competent, experienced, skilled, good leader) and personal traits (caring, nice, trustworthy, honest). Each trait scale shows high levels of reliability (Cronbach's α = 0.88 for competence traits, 0.90 for personal traits), and are rescaled to range from 0 (lowest) to 1 (highest). 16 We present these analyses in Table 3 below. Here, similar patterns persist -individuals rate perceived co-partisans about 2.9 percentage points higher on competence traits, and 4.3 percentage points higher on personal traits, compared to non-copartisans. 17 13) While participants only see one face rather than a series of faces, asking for a ranking allows them to express an expectation about the face. We find that this measure largely performs in the same fashion as the hiring decision measure. 14) Full question wording is available in the Supplementary Materials, as are descriptive statistics for main variables. 15) We include these fixed effects for incidental variation in responses caused by traits of particular faces. For example, these faces may vary on a host of factors, such as hairstyle, perceptions of clothing, perceived age, perceived socio-economic status, and background of the image. Because these may influence evaluations as well, we attempt to control for these factors. However, note that the fact that these do vary across images is a limitation of this study, as we can determine that these differences are only caused by appearance broadly, rather than specifically facial appearance. Table C1    Of course, these results might be masking heterogeneity by strength of partisan identification. Given the rise of affective polarization in American politics (see Mason, 2018a), one could expect that those with the strongest affective attachment to their own party, strong partisans, would show stronger levels of discrimination than those with less strong partisan attachments. To evaluate this, we turn to an examination of co-partisanship's effect on applicant evaluations conditional on strength of partisanship (which is coded as 1 for leaning partisans, 2 for not very strong partisans, and 3 for strong partisans), which is presented in Table 4, and graphically in Figures 1 and 2. 18 18) Results are robust to analyses of each indicator individually, for trait evaluations, and are presented in Table C2 of the Appendix (see Supplementary Materials). As we see from Figures 1 and 2, these effects are driven primarily by strong partisans, who consistently rate perceived co-partisans higher on all four dependent variables. Not very strong partisans still tend to rate perceived co-partisans more favorably, but in only one case out of four (personal traits) does this difference reach statistical significance at the 90% confidence level. Interestingly, leaning partisans seem to, if anything, rate non-copartisans a bit higher, though this difference never achieves statistical significance. This provides evidence that, when someone perceives an individual as a co-partisan, they rate them more favorably on a host of traits, but this effect is heightened among the strongest partisans.

Effect of Co-Partisanship and Partisan Strength on Competence and Personal Traits
Discrimination Without Partisan Cues 298

Method (Study 2b)
Next, we had participants view a different face, and answer a different series of questions, after completing a distractor task. 19 Here, participants were randomly assigned one of the faces from Study 2a, with a requirement that they not receive the exact same image they saw previously. Again, we see variation in partisan ratings, with 32.2% rating a face as Republican, 33.7% as Democratic, and 34.2% as neither party, across all four images. There was once again variation by face, this time with every face less likely to be rated as Republican than the male Republican face, but again there was significant variation, with the lowest frequency of ratings for the male Republican face as Democrat, by 23.1% of respondents. As we saw in the resume task, there is a tendency to project one's own partisanship onto the faces presented -42.0% perceived the faces to be co-partisans while 25.4% perceived of them to be from the out-party. Then, participants were asked a series of questions about their likelihood of engaging in small interactions with the pictured person. Here, participants were asked their willingness to say hello on the sidewalk, let them cut in line, talk about your weekend, ask for directions, sit next to at a bar, talk about politics, give your phone number, return a dropped dollar, talk about your families, and help pick up dropped bags for the person they saw pictured. Here, participants rate their willingness on a five-point scale, ranging from very unlikely (1) to very likely (5). As with Study 2a, participants were asked to rate partisanship of these faces after their ratings of interpersonal interactions.
These were selected specifically because they are relatively low stakes interactions -while we found evidence of discrimination in a formal, higher stakes setting (hiring a job applicant), we wonder if this would translate to more mundane interactions. These also mirror minor interactions many people have in their day to day lives. If partisan discrimination, even without partisan cues, based entirely on our perceptions, is pervasive in American society, we would expect it to manifest itself in these small interactions. These ten items scale together reliably (Cronbach's α = 0.86), and are combined together in an additive index for analysis, rescaled to range from 0 (least likely) to 1 (most likely). We analyze these results in a similar fashion to Study 2a, creating a variable for perceived co-partisanship and also interacting that with strength of partisanship. Once again, pure independents are excluded from analysis, but partisan leaners are included. These results are presented in Table 5

20)
Main results are mostly robust to using each indicator individually, with the exception of a negative effect of perceived co-partisanship on returning a dropped dollar, and no effect on helping pick up dropped bags. These results are available in Table C3 of the Appendix (see Supplementary Materials).

21)
Results for the interactive models are mostly robust to using each indicator individually, with the exception of a significant interaction between partisan strength and perceived co-partisanship on giving someone your phone number. These results are available in

Effect of Co-Partisanship and Partisan Strength on Interpersonal Interactions
Here, we see a slightly different effect emerge, compared to Study 2a. Generally, perceived co-partisanship has an effect on willingness to engage in these interpersonal interactions, with individuals about 8.2 percentage points of the scale more likely to engage in these interactions with someone they perceive as a co-partisan, compared to someone they do not. However, we do not see a significant interaction emerge based on partisan strength. While there is a larger difference for stronger partisans, we see that all partisans prefer a co-partisan individual to one who is not co-partisan. This suggests that, even in these relatively small and trivial interactions, individuals are willing to engage in partisan discrimination based only upon perceived partisanship, and this is true even for the weakest partisans. While the variable for co-partisanship is not directly statistically significant in the interactive model, this is an artifact of how interactive models should be interpreted, with statistical significance often not telling us the whole story (Kam & Franzese, 2007). The preferred method to interpret these interactive effects is via marginal effects or predicted probability models -we present this in Figure 3, which shows us that individuals, regardless of strength of partisanship, prefer co-partisans to non-copartisans in this model. Note here that studies 2a and 2b differ in how strength of partisanship influences decisions (where only the strongest partisans discriminate in Study 2a, but all individuals show discrimination in Study 2b). We have no theoretical reason to explain why there is such a difference between employment evaluations and interpersonal interactions, but can speculate as to these differences. Perhaps the most compelling speculation is that individuals have more information in Study 2a -they are presented an entire resume, whereas in Study 2b they are presented with only a single image of a person. Given more information to make decisions may help to mute partisan discrimination, at least among weaker partisans. Note also that Study 2a features an unusual result -we see discrimination by Republicans, but not Democrats. Due to space limitations, and since we do not have a theoretical prediction here, we present those analyses and discussion in Appendix D (see Supplementary Materials).

Conclusion
While previous research has (separately) identified that people infer partisanship from faces (Rule & Ambady, 2010;Samochowiec, Wänke, & Fiedler, 2010), and that people discriminate based on partisanship (Gift & Gift, 2015;Iyengar & Westwood, 2015) we have explored the possible connection between these two ideas. Simply inferring partisanship from faces may not be of consequence with respect to discrimination because there is considerable uncertainty around these perceptions. That is, people may believe that someone is a Republican or Democrat based on their facial appearance, but they may be unwilling to act on this belief because they recognize that they could be wrong. The research that has found evidence of partisan discrimination has demonstrated these effects by explicitly informing people of one's partisanship. What we have shown here is that people are still willing to discriminate, even when they could be wrong in their perceptions.
Not only does this finding amplify the significance of previous research on partisan facial judgements by showing that people will act upon them, but it also says suggests that partisan discrimination could be fairly widespread. While some people advertise their party affiliation through bumper stickers, yard signs, or in conversation, it is far more common to lack these explicit cues, and to form judgements based on other criteria such as appearance. When given no concrete information about an individual's partisanship, most people make assumptions (we believe from facial appearance), and then use these perceptions to discriminate in a range of different ways. While partisanship could be inferred in the resume experiment from the contents of the resume, despite its apolitical nature, our second study provides no information about the individual other than their picture, suggesting that facial appearance is driving these evaluations.
Further, the magnitudes of these effects are sizeable, suggesting that these perceptions play an important role in shaping how people view and interact with one another. While these appearance-based partisan perceptions are quite important, we do not find that people agree on whether a given face is a Democrat or Republican. Put differently, there is not a broad consensus about the faces we presented -many will see a face and decide the person is a Republican, and many others will see the same face and decide that the person is a Democrat. What is notable however, is that for most faces a sizeable majority report a partisan perception. These perceptions appear to be consequential. Many people are undoubtedly being misperceived (see Olivola & Todorov, 2010 for an exploration of the lack of accuracy of partisan perceptions from appearance), and receiving differential treatment based on an inaccurate perception.
We have only looked at the role of facial appearance, and do not explore the ways in which other visual cues like clothing or a car could shape partisan perceptions and discriminatory behavior. It seems likely that providing additional heuristics would only amplify the number of the people who perceive partisanship from appearance. Put differently, the effects shown here are likely to be conservative estimates since we have supplied people with only a face to draw their inferences from. While we cannot definitively speak to how prevalent these behaviors are in daily interactions with this work, they are undoubtedly more common than situations where one's partisanship is explicitly communicated. This work is not without limitations, and raises some questions for future study. The results we show are not experimental treatment effects -we cannot randomly assign people to a perception of a face. While our results are observational and we cannot make causal claims, they have the advantage of more closely mirroring how partisan information is encountered and processed in daily life. Another consequence is that we are unable to confidently say whether the findings here are a story where people are making inferences from faces that are leading to partisan perceptions, or whether people are using their own partisanship to guide their perceptions. As noted previously, there is a tendency for people to perceive the faces as co-partisans (49.9% in the first experiment and 42.0% in the second experiment) as opposed to members of the out-party (18.1% in the first experiment and 25.4% in the second experiment), suggesting that there is likely some partisan projection taking place. However, it is not the case that partisan projection of perceptions explains the entire picture -in both experiments a majority of participants either view the person as being from neither party, or the out-party. For some people the story is likely one of partisan projection, and for others it is inferences from faces driving partisan perceptions. While our findings are correlational in nature, it is clear that individuals are making some judgment about faces that influences their behavior. They are either inferring partisanship and discriminating based on that, or they are inferring something else about the faces, and discriminating and inferring partisanship based on that.
Further, we only explore partisan perceptions and discrimination related to white faces. We do this in an effort to minimize confounding factors that would arise if we were comparing across races, but it does raise questions about the extent to which these results travel to other racial groups. It may be the case that race becomes the deciding factor that drives impressions and overwhelms any partisan perceptions, or it that racial discrimination and partisan discrimination could interact with one another. Finally, our focus is in the United States, where partisanship as a powerful social identity is a pervasive aspect of our society. Although partisan identity differs cross-nationally, evidence suggests that it may be a powerful social identity in a range of different democracies (Bankert, Huddy, & Rosema, 2017), suggesting that similar dynamics could be at play outside of the U.S. Future research should explore these questions.