Visualizing threat and trustworthiness prior beliefs in face perception in high versus low paranoia

Predictive processing accounts of psychosis conceptualize delusions as overly strong learned expectations (prior beliefs) that shape cognition and perception. Paranoia, the most prevalent form of delusions, involves threat prior beliefs that are inherently social. Here, we investigated whether paranoia is related to overly strong threat prior beliefs in face perception. Participants with subclinical levels of high (n = 109) versus low (n = 111) paranoia viewed face stimuli paired with written descriptions of threatening versus trustworthy behaviors, thereby activating their threat versus trustworthiness prior beliefs. Subsequently, they completed an established social-psychological reverse correlation image classification (RCIC) paradigm. This paradigm used participants’ responses to randomly varying face stimuli to generate individual classification images (ICIs) that intend to visualize either facial prior belief (threat vs. trust). An independent sample (n = 76) rated these ICIs as more threatening in the threat compared to the trust condition, validating the causal effect of prior beliefs on face perception. Contrary to expectations derived from predictive processing accounts, there was no evidence for a main effect of paranoia. This finding suggests that paranoia was not related to stronger threat prior beliefs that directly affected face perception, challenging the assumption that paranoid beliefs operate on a perceptual level.

The resulting 15 face stimuli per group were presented along with unique behavioral descriptions, ten of which either implied the trait threatening (e.g., "This member of Group X spies on you") or trustworthy (e.g., "This member of Group Y keeps a secret you told him") and five of which depicted neutral behaviors (e.g., "This member of Group X happens to wait with you at the train station").Generation of the threat-implying behavioral descriptions was guided by items typically used in scales assessing paranoid beliefs (e.g., Paranoia Checklist 3 ), while the trustworthiness-implying behavior descriptions were phrased complementarily in order to offset the threatening behaviors (similar to 4 ).The neutral items were used as distractors.The assignment of base faces to Group X and Group Y, the order of block-wise group presentation (i.e., Group X vs. Group Y first), as well as the order of face-behavior pairs within blocks were randomized.Face-behavior pairs were random but fixed across subjects.Stimulus pairs were implemented via Qualtrics (www.qualtrics.com), and task completion was self-paced (Mdn = 8.53 min, SD = 8.08) with a minimum stimulus presentation duration of 5 sec.

Table S1-1
Comparisons of the face stimuli used to create the group-specific base faces for prior activation.

Reverse Correlation Image Classification Paradigm Stimulus Generation
We created one group-ambiguous base face by using Adobe Photoshop to morph the groupspecific base faces used during prior activation.Next, we generated 400 stimulus pairs by superimposing this base face with both unique random sinusoidal noise patterns and their mathematical inverses (a white pixel in the original noise pattern is black in its inversion and vice versa, see Figure S1-3) using the generateStimuli2IFC() from the rcicr package (random seed set to 10) 2 .Stimulus pairs were identical for all participants, but presentation order as well as position on the screen (left vs. right) were randomized.The experiment was implemented by adapting the Processing (www.processing.org)code provided by Anton Gollwitzer (www.github.com/AntonGollwitzer/ReverseCorrelationRunningOnline) to the present research focus.In each trial, one pair of stimuli was presented side-by-side against a black background (512 × 512 pixels), with the categorization item "Wer gehört zu Gruppe X?" ("Who belongs to Group X?") displayed above and a progress percentage displayed below the stimuli (white font color).

Figure S1-3
Sample RCIC stimulus pair, superimosed with random noise and the inverted noise pattern

S2: RCIC Strategy use
We coded participants' open responses regarding their strategy use during RCIC task completion in categories.In total, 136 participants (62%) reported having used a specific strategy.Most frequent responses included focusing on facial features (mainly the eye and mouth region; n = 59, 27%), identifying specific facial features signaling positive or negative valence (e.g., 'aggressive gaze'; n = 27, 12%), and looking for positive or negative valence in the faces without naming specific features (e.g., 'looking trustworthy'; n = 36, 16%).9 participants (4%) reported having focused on other physical attributes (e.g., masculinity) or similarity to known persons.Responses of 5 participants (2%) did not apply to any category and were coded 'other'; 84 participants (38%) did not report a strategy.

S3: Deviations from Preregistered Procedure
• Explicit evaluation of Group X and Group Y: Given the empirical non-normality of the explicit group evaluation difference score, we deviated from our pre-registered analysis plan in that we used a Wilcoxon rank-sum test instead of a one-way ANOVA to test for differences between conditions.Moreover, we added a Wilcoxon rank-sum test to test for differences between HP and LP samples, which we had not preregistered, to probe whether the prior manipulation was equally effective across paranoia levels.
• Exclusion due to RCIC responses: In addition to the pre-registered exclusion criterion (i.e., click one stimulus in 95% of the trials; did not apply to any participant), we applied three additional exclusion criteria we had not preregistered and excluded n = 4 participants based on these criteria.(1) Due to a programming error of the RCIC paradigm we became aware of after data collection was complete, participants could also proceed to the next trial by clicking on the margins (i.e., not on one of two stimuli, but on the black background left, right, up, down, and in between the stimuli).We excluded two participants (both LP sample, threat condition) who clicked the margin in between the stimuli in a significant number (i.e., 25% and 42%) of the trials, given that we could not infer which stimulus they had intended to select for these trials.(2) One participant (LP sample, trust condition) was excluded because of a RCIC completion duration of >19 h given that we did not expect this participant to maintain the associations learned during the prior activation phase across this time.(3) Due to a programming error of the ICI rating task, one participant's ICI was not rated, so that we had to exclude them from further analyses (HP sample, trust condition).All exclusion criteria were applied before any analyses related to the main outcome variables have been conducted.

S4: Bayesian Analysis Results
We complemented our manipulation check and main analyses with Bayes Factors (BF) obtained from Bayesian analysis counterparts performed with JASP 5 .These analyses were not preregistered.BF hypothesis testing differs from hypothesis testing using the p-value in that it directly and continuously compares the predictive adequacy of two competing statistical models (i.e., the null hypothesis and the alternative hypothesis), thus quantifying the relative evidence for and change brought about each of these models after seeing the data (for an overview, see 6 ).Specifically, BF10 (and its inverse BF01 = 1/BF10) quantifies the intensity of evidence for the alternative hypothesis H1 versus the null hypothesis H0 (and vice versa).According to a rule of thumb guideline 7 , a BF10 between 1 and 3 can be interpreted as weak evidence, a BF10 between 3 and 10 as moderate, and a BF10 greater than 10 as strong evidence for the alternative as compared to the null hypothesis.

Manipulation check
As expected, the results of a two-sided Bayesian Mann-Whitney U test (also referred to as Wilcoxon rank sum test) across conditions indicated that the data were 3.48×10 +8 times more likely under H1 than H0 (see Table S4-1), yielding strong evidence for the hypothesis that participants in the threat condition rated Group X more negatively than Group Y (whereas the opposite was true for the trust condition, see Table S4-2).By contrast, a two-sided Mann-Whitney U test across paranoia levels suggested the data were approximately 6 times more likely under H0 than under H1, yielding moderate evidence for the hypothesis that explicit group evaluations did not differ across paranoia levels (see Figure S4-1).

Main analyses
We submitted the ICI threat scores to a Bayesian two-way ANOVA with Condition and Paranoia level as between-subjects factors.As can be seen in Table S4-3, the data were most likely under a model considering only Condition as predictor.More specifically, they were approximately 1/0.36 = 2.78 times more likely than under a model also including participants' paranoia level (additively) and 1/0.08 = 12.5 times more likely than under a model additionally including the interaction between both factors.The model-averaged analysis of effects (see Table S4-4) suggested that there was strong evidence for including Condition as a predictor (BFincl = 2.73×10 +14 ), whereas the evidence in the data for including Paranoia level as well as the interaction between both factors was anecdotal.Table B-5 shows a summary of the marginal model-average posterior distributions.In sum, the results of our Bayesian analysis counterparts converged with the results of the frequentist approach.1.17×10 -4 Note.Models = predictors included in each model; P(M) = prior model probability; P(M|data) = posterior model probability; BFM = posterior model odds; BF10 = Bayes factors of all models compared to the best model; error % = estimate of the numerical error in the computation of the Bayes factor, with errors below 20% being acceptable in many cases.Note.P(incl) = prior inclusion probability; P(excl) = Prior exclusion probability; P(incl|data) = posterior includion probability; P(excl|data) = posterior exclusion probability; BFincl = inclusion Bayes factor which can be interpreted as evidence in the data for including a predictor.

S10: Correlations among Raters' Paranoia and ICI Ratings
Given the association between paranoia and a bias in face ratings (i.e., individuals with high paranoia tend to rate faces as angrier and more dominant as well as less trustworthy), we tested whether raters' self-reported paranoia correlated with their mean ratings of the individual classification images (ICIs).However, raters' R-GPTS sum scores were not correlated with ICI trustworthiness ratings (r = .09,95% CI [-.14, .31],t(74) = 0.76, p = .448)or ICI threat ratings (r = .06,95% CI [-.17, .28],t(74) = 0.53, p = .594).Thus, we averaged ICI trustworthiness and threat ratings across raters to calculate the outcome variable.

Note.
Positive explicit group evaluations indicate that Group X was rated more positively than Group Y, whereas negative values indicate the opposite.LP = low paranoia, HP = high paranoia, SD = standard deviation, SE = standard error.

Figure S4- 1
Figure S4-1 Inferential plots of Bayesian Mann-Whitney U Tests comparing explicit group evaluation between conditions (a) and paranoia levels (b)

Table S6 - 1
Results of ANCOVA on ICI threat score including demographic covariates, completion duration, and individual informational values of the ICIs.

Table S8 -1
Paranoia level differences as well as means and standard deviations of the face ratings.