https://doi.org/10.1177/0956797618788882

Psychological Science 2018, Vol. 29(11) 1807 –1823 © The Author(s) 2018 Article reuse guidelines: sagepub.com/journals-permissions DOI: 10.1177/0956797618788882 www.psychologicalscience.org/PS

ASSOCIATION FOR PSYCHOLOGICAL SCIENCEResearch Article

Faces are rich in information: They provide clues about gender, race, age, and trait attributes, which are inferred spontaneously and ubiquitously (Engell, Haxby, & Todorov, 2007; Todorov, 2017). Moreover, such infer- ences often guide our social behavior—for instance, we decide whom to trust on the basis of how trustworthy a face looks (Rezlescu, Duchaine, Olivola, & Chater, 2012; Van’t Wout & Sanfey, 2008). Many trait judgments made by participants across generations and cultures show consensus (Cogsdill, Todorov, Spelke, & Banaji, 2014; Lin, Adolphs, & Alvarez, 2017; Rule et al., 2010). But are trait judgments from faces accurate?

Previous research has shown that trait judgments from faces can be associated with important real-world social outcomes, such as dating and mating (Olivola et al., 2014; Valentine, Li, Penke, & Perrett, 2014), earn- ings and fundraising (Genevsky & Knutson, 2015; Hamermesh, 2011; Ravina, 2012), science communica- tion (Gheorghiu, Callan, & Skylark, 2017), sentencing decisions (Berry & Zebrowitz-McArthur, 1988; Blair, Judd, & Chapleau, 2004; Wilson & Rule, 2015; Zebrowitz

& McDonald, 1991), and leader selection (Todorov, Mandisodza, Goren, & Hall, 2005; for reviews, see Antonakis & Eubanks, 2017; Todorov, Olivola, Dotsch, & Mende-Siedlecki, 2015). Yet this prior research on the association between trait judgments from faces and real- world outcomes leaves open two important questions. First, most associations have focused on prosocial out- comes (e.g., correlations between competence judg- ments and election success; Todorov et  al., 2005). Second, most associations are plausibly driven not by the behavior of the targets whose face is being judged but by the interests of the perceivers who are making the judgments (e.g., correlations between interesting- looking scientists and the perceiver’s interest in their work). Here, we investigated an antisocial judgment that

788882 PSSXXX10.1177/0956797618788882Lin et al.Facial Inferences and Corruption research-article2018

Corresponding Author: Chujun Lin, California Institute of Technology, Division of Humanities and Social Sciences, HSS 228-77, 1200 E. California Blvd., Pasadena, CA 91125 E-mail: clin7@caltech.edu

Inferring Whether Officials Are Corruptible From Looking at Their Faces

Chujun Lin, Ralph Adolphs, and R. Michael Alvarez Division of Humanities and Social Sciences, California Institute of Technology

Abstract While inferences of traits from unfamiliar faces prominently reveal stereotypes, some facial inferences also correlate with real-world outcomes. We investigated whether facial inferences are associated with an important real-world outcome closely linked to the face bearer’s behavior: political corruption. In four preregistered studies (N = 325), participants made trait judgments of unfamiliar government officials on the basis of their photos. Relative to peers with clean records, federal and state officials convicted of political corruption (Study 1) and local officials who violated campaign finance laws (Study 2) were perceived as more corruptible, dishonest, selfish, and aggressive but similarly competent, ambitious, and masculine (Study 3). Mediation analyses and experiments in which the photos were digitally manipulated showed that participants’ judgments of how corruptible an official looked were causally influenced by the face width of the stimuli (Study 4). The findings shed new light on the complex causal mechanisms linking facial appearances with social behavior.

Keywords face perception, corruption, social attribution, stereotyping, political psychology, open data, open materials, preregistered

Received 9/25/17; Revision accepted 6/3/18

1808 Lin et al.

may offer a clearer insight into associations with the judged person’s own behavior: political corruption.

Political corruption has been a major cause of regime change and an important subject of much study in political science and economics (Rose-Ackerman, 2013). The possibility that corruptibility inferences from faces might be associated with real-world measures of cor- ruption is raised by three areas of previous research. First, theories of self-fulfilling prophecy argue that the impressions and expectations a face creates (e.g., how corruptible an official looks) influence how other peo- ple interact with the face bearer (e.g., how likely others would be to bribe the official) and that those recurrent interactions in turn shape the face bearer’s behavior so as to confirm other people’s impressions and expecta- tions (Haselhuhn, Wong, & Ormiston, 2013; Jussim, 1986; Slepian & Ames, 2016). Second, analyses of sen- tencing decisions show that evaluations of guilt and recommendations of punishment are influenced by the defendant’s facial appearance (Berry & Zebrowitz- McArthur, 1988; Blair et al., 2004; Wilson & Rule, 2015; Zebrowitz & McDonald, 1991). These findings suggest that officials who look more corruptible might be more likely to be accused, prosecuted, and convicted. Third, some studies have argued that the face contains a kernel of truth about a person’s nature—such as personality and criminal inclinations (Penton-Voak, Pound, Little, & Perrett, 2006; Valla, Ceci, & Williams, 2011)—even though the diag- nostic validity and the causal mechanisms remain obscure.

Given past research, we hypothesized that elected officials’ corruption records would be associated with traits, such as corruptibility, inferred from their facial appearances. We examined this association in three preregistered studies, where participants made trait inferences on the basis of the photos of unfamiliar government officials. To account for the possibility that this association might depend on the severity of corrup- tion and the level of office, we inspected both serious violations (i.e., cases considered political corruption) and minor violations (i.e., cases meriting a fine) and included officials at different levels of government (fed- eral, state, and local). In a fourth preregistered study, we explored which facial features might be causally mediating the impression of how corruptible an official was, using mediation analyses as well as experimental manipulations of the face stimuli. In this fourth study, we focused on metrics of facial structures—in particu- lar, facial width (relative to facial height) because it has been reported that men with wider faces are judged as less trustworthy (Haselhuhn et al., 2013; Stirrat & Perrett, 2010), more threatening (Geniole, Denson, Dixson, Carré, & McCormick, 2015), and less than fully human (Deska, Lloyd, & Hugenberg, 2018), although it remains unknown whether facial width-to-height ratio associates with actual behavior.

We have reported all measures, all conditions, all data exclusions, and how sample sizes were determined in this article and on the Open Science Framework (https://osf.io/k4mds/). All materials, data, and analysis codes for the present research can be accessed at this link.

Study 1

Our first study focused on federal and state officials and compared those who had clean records with those who were convicted of political corruption.

Method

Participants. This study was preregistered before data collection began (https://osf.io/mge8r/). A sample size of 100 participants was predetermined on the basis of two pilot studies—one carried out in the lab in May 2016 and the other via Amazon’s Mechanical Turk (MTurk) in October 2016. The lab study included 32 participants recruited from the general public of Southern California, and the MTurk study had 18 participants. For the hypoth- esis that elected officials’ corruption records would be associated with face-based inferences of corruptibility, the laboratory pilot study yielded an estimated effect size of 1.06, and the MTurk pilot study yielded an estimated effect size of 1.05, justifying a minimum sample size of 16 participants. Given these results and to ensure sufficient power even with dropout, we recruited 100 MTurk par- ticipants in November 2016. We selected participants who were native English speakers, located in the United States, and 18 years old or older. In addition, they had to have normal or corrected-to-normal vision, an educa- tional attainment of high school or above, a good MTurk participation history (a human-intelligence-task, or HIT, approval rate ≥ 95% and ≥ 1,000 HITs approved), and no prior participation in our pilot studies.

Eighteen individuals were excluded in total, 2 for not being native English speakers, 6 for pressing the same response key for all trials in a block, and 10 for failing to input valid responses for more than 10% of the trials in a block (responses were considered not valid if missing or entered within 100 ms—the minimum time needed for visual exploration of the face; Olivola & Todorov, 2010). After exclusion, there were 82 par- ticipants in our final sample (42 female; age: M = 39 years, SD = 12; 84% White, 10% Black, 5% Asian).

Stimuli. Stimuli were photos of 72 real elected officials. All were Caucasian males who have held federal or state legislative offices in the United States. Photos were offi- cial headshots obtained from government websites and personal campaign websites (63%), news articles (23%), and Wikipedia (14%). All photos were converted to gray-scale

Facial Inferences and Corruption 1809

images on a plain gray background and cropped to a uni- form size. All faces were frontal, smiling, in clear focus, and centered in the middle of the image.

Among the 72 officials, half were convicted of political corruption (corrupt officials), and the other half had clean records (noncorrupt officials). The cor- rupt officials were from two Wikipedia data sets (list of American state and local politicians convicted of crimes: https://en.wikipedia.org/wiki/List_of_American_state_ and_local_politicians_convicted_of_crimes; list of American federal politicians convicted of crimes: https://en.wikipedia.org/wiki/List_of_American_fed eral_politicians_convicted_of_crimes). To reduce sources of variability, we included only officials who were Caucasian, were male, held federal or state legisla- tive offices, and were convicted between 2001 and 2016 of political corruption conducted while in office (brib- ery, money laundering, embezzlement, mail fraud, wire fraud, tax fraud, conflict of interest, misusing funds, misusing office, or falsifying records). In addition, age information for these officials had to be publicly avail- able, as did frontal photographs of acceptable clarity in which the official was smiling. All photographs had been taken while officials were in office. Most photos of the corrupt officials had a known creation date, and we confirmed that the photos were taken before their conviction (72%); for the rest of the photos (28%), the creation date was unknown (analyses were also per- formed when excluding data for these stimuli; the pat- tern of results did not change). The noncorrupt officials were randomly matched from the list of incumbents who had clean records, were holding the same office in the same state, and were of the same gender, the same race, and similar age (±12 years) as the corrupt officials during the period of their misconduct. For instance, if the stimuli contained a Caucasian male cor- rupt official who was a member of the Arizona House

of Representatives during his misconduct at the age of 55, then a noncorrupt official would be randomly selected from our available stimulus set from the list of Arizona House of Representatives incumbents who had a clean record and who was a Caucasian male between the ages of 43 and 67.

Procedures. Participants were not informed of the pur- pose of the study or the sampling of the stimuli. In par- ticular, they were not given any information about the percentage of politicians in our stimulus set who might be corrupt in real life. They were told only that they would view a series of politician photos and that they should judge how corruptible, dishonest, selfish, trust- worthy, and generous these politicians looked to them (experiment instructions are available at https://osf.io/ k4mds/). Participants completed five blocks of experi- ments, with each block corresponding to judging one trait for all faces. The ordering of the faces within each block as well as the ordering of the blocks were randomized.

Each block started with an instruction screen that specified the trait to be judged (e.g., corruptibility). Participants were instructed to make their decisions as quickly and precisely as possible. Six practice trials familiarized participants with the task. Participants viewed photos of officials one at a time in randomized order and made judgments. Each trial began with a fixation cross, followed by the photo (1 s) with a 5-point Likert scale below it. Scales were anchored with bipolar adjectives (Fig. 1). Participants could make a decision as soon as the photo appeared and within 4 s after the photo disappeared. The orientation of the scale was randomized across blocks, and scores were reverse- coded as needed.

After completing all five blocks of ratings, partici- pants were asked whether they had recognized any of the officials and filled out a short survey questionnaire

0.5–1.5 s 1 s Maximum of 4 s 0.5 s

Time

Fig. 1. An example trial in the corruptibility-judgment block. Each trial began with a fixation cross. Then a photo of an official appeared for 1 s. The orientation of the scale was randomly flipped for each block and each participant. Participants made a decision by pressing one of the number keys from “1” to “5” on their keyboard. As soon as a valid key was pressed (or 4 s after the photo disappeared if no valid key was pressed), the trial ended, and there was a blank interstimulus interval.

1810 Lin et al.

on demographic characteristics, political attitudes, and personality.

Results

Reliability of face-based trait inferences. Following our preregistered plan, we excluded from further analy- sis any responses faster than 100 ms and data for officials who were recognized. Among the 82 participants, 7 rec- ognized one official (in total, four officials were recog- nized). The percentages of participants who used the full scale to rate the faces ranged from 59% to 68% across the five traits, and over 90% of the participants used scores on both sides of the midpoint to rate the faces (see Fig. S1 in the Supplemental Material available online).

First, we checked whether participants gave consis- tent judgments for a face across different traits. We expected consistent ratings for a face on traits with the same valence to be positively correlated and ratings on traits with opposite valences to be negatively corre- lated. Although this was not planned in our preregistra- tion, we computed repeated measures correlations (the R function rmcorr) to determine the common within- individuals correlations for ratings between each pair of traits, to handle the issue of nonindependence in repeated measures. Results (see Table S1 in the Supple- mental Material for coefficients and 95% confidence intervals, or CIs) showed that at an individual level, judgments of a face for traits with the same valence were positively correlated (repeated measures rs rang- ing from .24 to .31, ps < .001), and judgments of a face for traits with opposite valences were negatively cor- related (repeated measures rs ranging from –.30 to –.21, ps < .001). Following our preregistered plan, we also analyzed the consistency of these ratings at an aggre- gate level. Ratings for each face were first averaged over participants, and (tie-corrected) Spearman correlation coefficients were calculated for each pair of traits with those averaged ratings. Aggregate-level judgments for a face were highly consistent across traits because they averaged out the measurement noise inherent in the individual-level correlations (|r| ≥ .75; see Table S2 in the Supplemental Material).