1 Code Availability

Code used to perform all modeling, simulation, and analysis is available on OSF and GitHub:

Methods Supplement

2 Study Sample

Participants were recruited from the University of Michigan Prechter Bipolar Longitudinal Study, community advertisements, and local clinics. During recruitment, distributions of HC subsamples were age- and sex-matched to those of the SZ and BD sub-samples. Participants had no history of medical conditions with neurological sequelae, visual acuity of 20/30 or better on Snellen chart, and no recent substance use disorder (patients had no substance abuse or dependence in past year, HC in last five years). BD met criteria for bipolar I and SZ met criteria for schizophrenia or schizoaffective disorder. HC had no history of axis I disorders and no first-degree relatives with bipolar or psychotic disorders.

3 Data Collection

Participants gave written informed consent and were compensated for participation. The study received approval from the Institutional Review Board at the University of Michigan Medical School. Study procedures included: diagnostic assessments and clinical ratings by trained assessors, neuropsychological tests of general/social cognition, self-reports, and a behavioral gaze discrimination task. The task was completed during the acquisition of electroencephalography (EEG) data, but EEG analysis is outside the scope of this paper.

4 Assessments

Diagnoses were confirmed by trained assessors using the Structured Clinical Interview for DSM-IV-TR or Diagnostic Interview for Genetic Studies. In the SZ sub-study, SZ symptoms (Scale for Assessment of Positive Symptoms [SAPS]; Scale for Assessment of Negative Symptoms [SANS]), depressive symptoms (Beck Depression Inventory-II [BDI-II 53]), general cognition (Brief Assessment of Cognition in Schizophrenia [BACS]), emotion-related social cognition (Mayer-Salovey-Caruso Emotional Intelligence Test [MSCEIT]), and social functioning (Social Adjustment Scale Self-Report, Social/Leisure scale [SASSR-Social]; inverse-coded) were assessed. In the BD sub-study, mania (Altman Self-Rated Mania scale [ASRM 52]) and depressive symptoms (BDI-II) were assessed. Inter-rater reliability was > 80% for diagnoses and clinician-assessed ratings.

For analyses in the main text, the measures of our clinical correlates were scored as follows:

  • BACS scoring: BACS subtests were z-scored relative to published age- and gender- norms of a normative HC sample and averaged to obtain a composite of general cognition (Keefe, 2008).
  • MSCEIT scoring: MSCEIT subscales were converted to age- and gender-corrected standard scores (Mayer, 2002), z-scored relative to the full sample, and averaged to obtain a composite of emotion-based social cognition.
  • SANS-Amotivation scoring (Negative symptom-related amotivation factor): The amotivation factor was calculated by applying published factor loadings (Sayers, 1996) to SANS avolition/apathy and asociality/anhedonia items and summing those to obtain a composite.
  • SAPS-Delusion scoring: Delusional symptoms were scored at the total summed score of delusional items on the SAPS.
  • SAPS-Hallucination scoring: Hallucination symptoms were scored at the total summed score of hallucination items on the SAPS.
  • BDI-Depression scoring: Depressive symptoms were scored as the total summed score on the BDI-II.
  • ASRM-Mania scoring: Mania symptoms were scored as the total summed score on the ASRM scale.

Modeling Supplement

5 Defining/Refining Model Space

The steps below detail the full process of defining and refining the full model space explored in the current study. After completing these steps, we were left with 8 models that underwent more comprehensive testing: 1,2,5,6,7,8,9, and 10. Full specification of these models is provided in the next section.

Defining Model Space
Step / Model Type Model Tested Rationale Pars Vary by GAZE Pars Vary by HEAD Pars Vary by EMO
Step 1: Define baseline model
Baseline
1 yes We assumed drift rate would need to vary by the condition (GAZE) that determines accuracy and this model doesn’t allow for that. So, while this is an implausible model, it does serve as our baseline – – –
Step 2: Define model space for models in which 1 single parameter varies by one or several task conditions
Draft Rate Vary Only
2 yes Plausible account of task behavior (drift rate is sensitive to stimulus properties so it is plausible that it could vary based upon any of our task conditions) drift rate – –
9 yes Plausible account of behavior (drift rate is sensitive to stimulus properties so it could plausibly vary based upon any task condition) drift rate drift rate –
10 yes Plausible account of behavior (drift rate is sensitive to stimulus properties so it could plausibly vary based upon any task condition) drift rate drift rate drift rate
5 yes Plausible account of behavior (drift rate is sensitive to stimulus properties so it could plausibly vary based upon any task condition) drift rate – drift rate
– no Plausible account of behavior (drift rate is sensitive to stimulus properties so it could plausibly vary based upon any task condition) – drift rate –
– no Exclusion Reason #1: (same as above) – drift rate drift rate
– no Exclusion Reason #1: (same as above) – – drift rate
Threshold Separation Vary Only
– no Exclusion Reason #1: (same as above) threshold separation – –
– no Exclusion Reason #1: (same as above) threshold separation threshold separation –
– no Exclusion Reason #1: (same as above) threshold separation threshold separation threshold separation
– no Exclusion Reason #1: (same as above) threshold separation – threshold separation
– no Exclusion Reason #1: (same as above) – threshold separation –
– no Exclusion Reason #1: (same as above) – threshold separation threshold separation
– no Exclusion Reason #1: (same as above) – – threshold separation
Start Point Vary Only
– no Exclusion Reason #1: (same as above) start point – –
– no Exclusion Reason #1: (same as above) start point start point –
– no Exclusion Reason #1: (same as above) start point start point start point
– no Exclusion Reason #1: (same as above) start point – start point
– no Exclusion Reason #1: (same as above) – start point –
– no Exclusion Reason #1: (same as above) – start point start point
– no Exclusion Reason #1: (same as above) – – start point
Step 3: For plausible models defined above, build upon those by letting one additional parameter vary by task conditions. Begin with simplest model (model 2)
Building Upon Model 2
3 partially Tested intially, but later removed because it violated a key assumption of this version of the DDM. Exclusion Reason #2: threshold separation and start point are assumed to be set before the onset of a stimulus and, therefore, cannot vary based upon stimulus properties that aren’t predictable based upon task design. Here, the only ‘predictable’ task condition was EMO because condition were blocked by stimulus emotions. As such, we do not pursue this model further. drift rate, threshold separation – –
4 partially Tested intially, but later removed because it violated a key assumption of this version of the DDM. Exclusion Reason #2: (same as above) drift rate, threshold separation, start point – –
– no Never tested because it violated a key assumption of this version of the DDM. Exclusion Reason #2: (same as above) drift rate, start point – –
6 yes Retained because it allowed drift rate to vary by the condition that determines accuracy (GAZE) AND the condition that threshold separation/start point parameters are allowed to vary by doesn’t violate assumptions of this version of the DDM drift rate – drift rate, threshold separation
7 yes Retained because it allowed drift rate to vary by the condition that determines accuracy (GAZE) AND the condition that threshold separation/start point parameters are allowed to vary by doesn’t violate assumptions of this version of the DDM drift rate – drift rate, start point
8 yes Retained because it allowed drift rate to vary by the condition that determines accuracy (GAZE) AND the condition that threshold separation/start point parameters are allowed to vary by doesn’t violate assumptions of this version of the DDM drift rate – drift rate, threshold separation, start point
Step 4: Initial model comparisons models made 2 things clear: 1) the models ran into difficulties when start point parameters varied by emotion condition (model 7 and 8); AND 2) fit was better fit when drift rate varied by gaze AND head (model 9 and 10 vs. others). Thus, the only additional models that made sense to consider building upon were model 9 and 10. Thus 2 more complex models were considered but not tested for reasons specified below
Building Upon Model 9
– no Exclusion Reason #3: This is a plausible account of behavior that does not violate assumptions laid out thus far. Although allowing threshold separation to vary by emotion condition in simpler models (Model 6) did outperform other models where it did not vary by emotion (Model 5 and 7), the LOO difference between models was well within the SE of the LOOIC (LOO diff = 145 to 166; LOO SE ~469). So – considering that allowing threshold separation to vary by emotion offered very little to improve the model account – we decided not to pursue testing this model further and, instead, opted to continue with earlier, more parsimonious models. drift rate drift rate threshold separation
Building Upon Model 10
– no Exclusion Reason #3: (same as above) drift rate drift rate drift rate, threshold separation

6 Model Specification

After refining the model space, we were left with 8 models that underwent more comprehensive testing: 1,2,5,6,7,8,9, and 10. We provide graphical and descriptive specification of these models is in the sections that follow.

For all groups h (from 1 to 3, where 1=hc, 2=bd, 3=sz), subjects i (from 1 to 100), and trials j (from 1 to max of 512), Cij indicates the choice (1=yes/upper bound, 2=no/lower bound) and RTij indicates the reaction time in seconds. RTij is Wiener distributed W(αi,βi,δi,τi,), where αi is the subject-level threshold separation (0.1<αi<3.9), βi is the subject-level start point (0<βi<1), δi is the subject-level drift rate (-4<δi<4), and τi is the subject-level NDT (0<τi<scaled minimum reaction time [minRTi] for subject i). Because the Wiener distribution in Stan allows upper-boundary responses only, lower boundary choices were modeled using -δi for drift rate and 1-βi for start point.

To facilitate sampling, a non-centered parameterization is used. This means that parameters are sampled in a ‘standardized’ space and then transformed into target distributions. For a given parameter (say, threshold separation), the diagnostic group- (αμh) and subject-level means (αi‘) were sampled from a normal N(0,1), while group-level variances (ασh) were sampled from a positive-bound normal distribution N(0,.2)+. Then, the subject’s threshold separation parameter (αi) was transformed via φ(αμh+ασh x αi) x 3.9+0.1, such that the αi is informed by subject- (αi’) and group-level information (αμh and ασh), transformed to a 0<x<1 scale using the φ standard cumulative normal distribution function, and scaled* to the desired range.

*NDT is estimated as a proportion (0<x<1) and scaled by each subject’s minRTi to bring the values into units of seconds. To help with issues sampling the NDT (i.e., the models had difficulty sampling on the trial in which each subject’s minimum RT occurred), we scaled the NDT by 0.9 of the subject’s actual minRTi.

For models 2, 5,6,7,8,9, and 10, gaze direction conditions, k (from 1 to 2, where 1=direct,2=indirect) were also accounted for by different parameters as indicated below. For models 5,6,7,8,9, and 10, head orientation conditions, m (from 1 to 2, where 1=forward,2=deviated), and/or emotion conditions, l (from 1 to 2, where=neutral,2=fearful), are also accounted for by different parameters as indicated below.

6.1 Model 1

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). (tau; τ).

6.2 Model 2

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), NDT (tau; τ). Gaze condition effect (k) for drift rate (delta; δ).

6.3 Model 5

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k) and emotion (l) condition effect for drift rate (delta; δ).

6.4 Model 6

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k) and emotion condition (l) effect for drift rate (delta; δ). Emotion condition (l) effect for start point (beta; β).

6.5 Model 7

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k) and emotion (l) condition effect for drift rate (delta; δ). Emotion condition (l) effect for threshold separation (alpha; α).

6.6 Model 8

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k) and emotion (l) condition effect for drift rate (delta; δ). Emotion condition (l) effect for start point (beta; β) and threshold separation alpha; α).

6.7 Model 9

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k) and head (m) condition effect for drift rate (delta; δ).

6.8 Model 10 (Winning Model)

Description: Group (h) and subject (i) effects for drift rate (delta; δ), threshold separation (alpha; α), start point (beta; β), ndt (tau; τ). Gaze (k), emotion (l), and head (m) condition effect for drift rate (delta; δ).

7 Convergence Checks

Basic convergence diagnostics were performed for all models. This included checking for 0 divergences, ensuring that all Rhat values were <1.1, checking that trace plots were well-mixed, verifying that autocorrelation was low by lag of ~30, and checking that the effective sample size (ESS) was sufficient. Models generally performed well by these standards, suggesting no problems with convergence. Though model 7 and model 8 (which allowed start point to vary by emotion condition) has some trouble with sampling efficiency. We do not include full diagnostics here for brevity. Instead, we only present these complete diagnostics below for the final winning model (Model 10).

8 Parameter Recovery

Parameter recovery analyses were performed at the start of model testing (for our baseline model) and prior to analyses (for our winning model) to ensure we could recover parameters for this task and DDM parameterization. Details of parameter recovery from both stages of modeling testing are detailed below.

8.1 Parameter Recovery (Model 1)

We fixed NDT at 0.2 here 1) to save on storage/computing time, and 2) because we were not interested in individual differences in NDT based on subject-level parameter estimates.

To begin, we selected plausible values for group-level “generating parameter values” (threshold separation=2, start point=.5, drift rate=-0.5). Group-level posteriors for each parameter were then defined as normal distributions with means equal to the aforementioned values and each with a SD of 0.1. Using these group-level parameter distributions, we randomly sampled subject-level parameter values for 25 subjects. These “subject-level generating values” were then fed into the “rwiener” function of the RWiener package (Wabersich & Vandekerckhove, 2014) to simulate choice and RTs for 512 trials—the number of trials on the current task. We fit simulated data in Stan using 1000 warmup draws and 4000 postwarmup draws sampled over multiple chains. Convergence checks (described previously) indicated that all models had converged. After fitting models, we calculated the means and 95% highest density intervals (HDIs) for the estimated posteriors (from the model fits) for group- and subject-level parameters. These steps were repeated for each of 50 separate simulations. We then evaluated parameter recovery in 3 ways, described below in the following 3 sections.

8.1.1 Group-Level Parameters (Averaged over simulations)

First, we examined how well the group-level generating values were recovered by simulated models fits on average. We averaged over the posterior means and 95% HDIs for group-level parameters for all 50 simulations; plotted generating group-level values (red X’s in plots below) superimposed over the averaged mean and 95% HDI (points and error bars in figure below); and assessed how well, on average, the group-level generating values were captured by simulated model fits.

Results below suggest that group level parameters were well-recovered on average because 1) the generating values (red X’s in plots below) were contained within the average 95% HDI for all parameters (error bars in plots below), and 2) the generating values are closely aligned with the group-level posterior mean (data points in plots below) of simulated fits on average for all parameters.

8.1.2 Group-Level Parameters (Separate for each simulation)

Second, we examined group-level parameters for the individual simulations to determine what % of the 50 simulations successfully recovered the group-level generating parameter values (vertical black line in plots below). Recovery was “successful” if the 95% HDI of the fitted group-level posterior (error bars in plots below) for a given simulation contained the original subject-level generating value. Data in green in the plots below indicate the group-level parameters were recovered and data in red indicates they were not.

Results below suggest that group level generating parameters were well-recovered because the generating values (vertical black line in plots below) were contained within the average 95% HDI (error bars in plots below) in 94-98% of simulations for all parameters.

8.1.3 Subject-Level Parameters

Third, we examined how well subject-level generating values were captured by simulated model fits. We determined what % of subject-level generating parameter values were successfully recovered across the 50 simulations. Recovery was “successful” if the 95% HDI of the fitted subject-level posterior contained the original subject-level generating value. Data in green in the plots below indicate the subject-level generating values were recovered and data in red indicates they were not.

Results below suggest that subject level generating parameters were well-recovered because the generating values (x-axis) were contained within the average 95% HDI of fitted values (error bars in plots below) in 94-96% of simulated participants.

8.2 Parameter Recovery (Model 10)

Parameter recovery was also performed for the winning model: Model 10. We used the same simulation and fitting procedure for Model 10 as we described above for Model 1, with a few exceptions: 1) simulated data and model fits accounted for the effects of gaze direction, head orientation, and facial emotion on drift rates; 2) parameter recovery was performed separately for HC, SZ, and BD; 3) group-level generating values were defined as normal distributions with means and SDs matching those of each diagnostic group’s group level mean posteriors from the “final model fit” of Model 10 described in section 10 below; and 4) for each simulation, the simulated number of subjects matched the actual number of subjects for each group. We evaluated parameter recovery in 3 ways, as before, which is described in the following 3 sections (broken down by by diagnostic group and task conditions).

8.2.1 Group-Level Parameters (Averaged over simulations)

First, we examined how well the group-level generating values were recovered by simulated models fits on average. We averaged over the posterior means and 95% HDIs for group-level parameters for all 50 simulations; plotted generating group-level values (red X’s in plots below) superimposed over the averaged mean and 95% HDI (points and error bars in figure below); and assessed how well, on average, the group-level generating values were captured by simulated model fits. This was done separately for each diagnostic group and task condition.

Results below suggest that group level parameters were well-recovered on average for SZ, BD, and HC because 1) the generating values (red X’s in plots below) were contained within the average 95% HDI for all parameters (error bars in plots below), and 2) the generating values are closely aligned with the group-level posterior mean (data points in plots below) of simulated fits on average for all parameters.

8.2.2 Group-Level Parameters (Separate for each simulation)

Second, we examined group-level parameters for the individual simulations to determine what % of the 50 simulations successfully recovered the group-level generating parameter values (vertical black line in plots below). Recovery was “successful” if the 95% HDI of the fitted group-level posterior (error bars in plots below) for a given simulation contained the original subject-level generating value. Data in green in the plots below indicate the group-level parameters were recovered and data in red indicates they were not. This was done separately for each diagnostic group and task condition.

Results below suggest that group level generating parameters were well—recovered because–across all diagnostic groups and task conditions—the generating values (vertical black line in plots below) were contained within the average 95% HDI (error bars in plots below) in 88-100% of simulations for all parameters.

8.2.3 Subject-Level Parameters

Third, we examined how well subject-level generating values were captured by simulated model fits. We determined what % of subject-level generating parameter values were successfully recovered across the 50 simulations. Recovery was “successful” if the 95% HDI of the fitted subject-level posterior contained the original subject-level generating value. Data in green in the plots below indicate the subject-level generating values were recovered and data in red indicates they were not. This was done separately for each diagnostic group and task condition.

Results below suggest that subject level generating parameters were well-recovered because the generating values (x-axis) were contained within the average 95% HDI of fitted values (error bars in plots below) in 92-97% of simulated participants.

9 Model Comparison

We performed model comparisons for models 1, 2, 5, 6, 7, 8, 9, and 10 using leave-one-out (LOO) cross-validation (Vehtari, 2017). LOO model comparisons were performed in the full sample and within diagnostic groups to ensure that the full sample winning model was also the winning model within diagnostic groups.

In the tables below, models are presented in order from best- to worst-fitting based on the expected log pointwise predictive density, where higher ELPD values indicate better model fit. Differences in out-of-sample predictive accuracy were assessed for all models relative to the best fitting model based on changes in the ELPD (∆ELPD-LOO; ‘delta_elpd’ below) relative to the best fitting model. Uncertainty around ∆ELPD-LOO is captured through the standard error (SE; ‘delta_elpd_se’ below) of the estimated pointwise ELPD differences. We consider an absolute change in ELPD < 1 SE as weak evidence for improved predictive accuracy. Below, “pareto_k” columns indicate the number of observations in acceptable (“good” and “ok”) versus problematic ranges (“bad” and “very bad”; Vehtari, 2022). Pareto K diagnostics flag issues with the integrity of the LOO approximation and can also signal problems with the model itself.

For all groups, the winning model (“Model 10” hereafter) was one in which all parameters varied by diagnostic group and evidence accumulation (drift rate) was influenced by gaze direction, head orientation, and emotion expression of stimuli. It assumed that response caution (threshold separation), start point (expectancy bias), and NDT operated as trait-level processes that did not vary in response to stimulus changes. This is a reasonable model account because the drift rate is influenced by the physical qualities of the stimulus. Model 10 showed improvements in predictive accuracy relative to the other models (i.e., ∆ELPD-LOO) that far exceeded the uncertainty of those estimates (i.e., |∆ELPD-LOO| was 5 to 105 times the SE).

9.1 Model Comparison (Full Sample)

First, we performed model comparisons on the full sample N=100 combined. Results showed that models perform better when the drift rate is allowed to vary by gaze direction (e.g., Model 2 vs 1) and head orientation (e.g., Model 9,10 vs others). The models also perform better when drift rate varies by gaze direction, head orientation, AND emotion (Model 10), versus when drift rate only varies by gaze direction and head orientation (Model 9). The difference between the best fitting model (Model 10) and the next best fitting model (Model 9) based on the difference in ELPD LOO (“delta_elpd” below) was over 5x greater than the SE of the estimate (“delta_elpd_se” below). So, the improvement from Model 9 to Model 10 was considerable.

Additionally, we noted some difficulties when start point is allowed to vary by emotion (Model 7 and 8), which have some high Pareto K values.

Model Comparison in the Full Sample
model looic looic_se elpd_loo elpd_loo_se delta_elpd delta_elpd_se pareto_k_good pareto_k_ok pareto_k_bad pareto_k_verybad
m10 10970.44 507.24 -5485.22 253.62 – – 50058 4 0 0
m9 11344.58 505.15 -5672.29 252.57 -187.07 36.01 50059 3 0 0
m6 20253.94 469.87 -10126.97 234.93 -4641.75 102.97 50059 3 0 0
m8 20256.08 469.77 -10128.04 234.89 -4642.82 103.74 50059 1 2 0
m7 20398.59 468.85 -10199.29 234.42 -4714.07 102.28 50060 1 1 0
m5 20419.81 469.32 -10209.90 234.66 -4724.68 101.49 50060 2 0 0
m2 20697.94 468.45 -10348.97 234.23 -4863.75 104.38 50061 1 0 0
m1 48026.44 339.75 -24013.22 169.88 -18528 176.86 50062 0 0 0

9.2 Model Comparison (by Group)

Second, we ensure that the model comparison results from the full sample hold within diagnostic groups. Results within groups show the same pattern of results: Model 10 is the best fitting model within SZ, BD, and HC. The improvement from Model 9 to Model 10 based on changes in LOO ELPD exceeded the SE of the differences for all groups, indicating Model 10 offered a considerably better account of the data. Based on these results, it was clear that Model 10 offered the best fit of the data based on LOO model comparisons.

Model Comparison by Group
group model looic looic_se elpd_loo elpd_loo_se delta_elpd delta_elpd_se pareto_k_good pareto_k_ok pareto_k_bad pareto_k_verybad
HC
m10 -4054.07 305.99 2027.04 153.00 – – 17162 2 0 0
m9 -3935.78 305.50 1967.89 152.75 -59.15 20.42 17163 1 0 0
m6 -203.35 278.45 101.68 139.22 -1925.36 66.61 17163 1 0 0
m8 -195.64 278.30 97.82 139.15 -1929.22 67.04 17163 0 1 0
m5 -168.42 278.44 84.21 139.22 -1942.83 66.13 17163 1 0 0
m7 -165.45 278.30 82.73 139.15 -1944.31 66.57 17163 1 0 0
m2 -72.85 278.26 36.43 139.13 -1990.61 67.42 17163 1 0 0
m1 11339.41 197.14 -5669.71 98.57 -7696.74 111.63 17164 0 0 0
BD
m10 8999.78 292.73 -4499.89 146.37 – – 18964 1 0 0
m9 9153.42 292.11 -4576.71 146.06 -76.82 21.92 18964 1 0 0
m6 12301.55 271.33 -6150.77 135.67 -1650.88 60.23 18964 1 0 0
m8 12303.47 271.41 -6151.74 135.71 -1651.84 60.55 18965 0 0 0
m7 12356.91 270.38 -6178.45 135.19 -1678.56 59.36 18965 0 0 0
m5 12363.14 270.57 -6181.57 135.28 -1681.68 59.04 18965 0 0 0
m2 12473.79 270.26 -6236.90 135.13 -1737 60.95 18965 0 0 0
m1 20918.42 197.45 -10459.21 98.72 -5959.32 101.71 18965 0 0 0
SZ
m10 6024.73 269.36 -3012.36 134.68 – – 13932 1 0 0
m9 6126.95 266.61 -3063.47 133.30 -51.11 19.98 13932 1 0 0
m8 8148.24 255.00 -4074.12 127.50 -1061.76 50.9 13931 1 1 0
m6 8155.74 255.08 -4077.87 127.54 -1065.51 50.28 13932 1 0 0
m7 8207.13 254.35 -4103.57 127.18 -1091.2 49.96 13932 0 1 0
m5 8225.09 254.86 -4112.54 127.43 -1100.18 49.31 13932 1 0 0
m2 8297.00 253.78 -4148.50 126.89 -1136.13 51.25 13933 0 0 0
m1 15767.32 187.69 -7883.66 93.84 -4871.3 91.14 13933 0 0 0

10 Confusion Matrix

We further evaluated whether it was possible to arbitrate between these 8 models (1, 2, 5, 6, 7, 8, 9, and 10) by generating a confusion matrix. This allowed us to determine whether data generated by each model could be identified as being best fit by the model that generated it.

We simulated 50 data sets using each of these 8 models, fit all 8 models to each simulated data set, generated LOO values for each model fit, and identified the best fitting model for each simulated data set (i.e., model with the lowest LOO value). Simulated data was generated using the same procedure described above for Parameter Recovery. Because of the computational resources this process demands, we took steps to reduce the overall run time: 1) we simulated only N=20 subjects for each data set; 2) we reduced the warm up/post warmup samples (500 warm up, 1000 post warm up over multiple chains); 3) we fixed NDT at 0.2; and 4) we performed this process in HC only (i.e., HC posterior values were used to generate simulated data).

Results are summarized below. This confusion matrix indicates the proportion of the N=50 data sets simulated using a given model (rows) were best fit by each of the 8 models (columns), where values in rows sum to 1. Ideally, the diagonal elements should equal 1 and off-diagonal elements should equal 0, meaning that the model that generated the data always fits the data best. Results show that data generated by Model 10—the best fitting model based on LOO model comparisons—was best fit by Model 10 in 98% of simulations. The same was true of Model 9—the second best-fitting model based on LOO model comparisons. There was more difficulty arbitrating between simpler models, especially Models 5, 7 and 8. It is noteworthy that for Models 7 and 8 (which allowed start point to vary based on emotion) we also noted sampling difficulties at other model evaluation steps. This further corroborates our decision to not explore those models further. Together, these results indicate that it was possible to arbitrate between these two best-fitting models, including Model 10–the winning model based on LOO model comparisons.

11 Posterior Predictions

After model comparisons and confusion matrices above, it Model 10 was the best performing model. Before selecting it as our winning model to be used for analyses, we subjected Model 10 to additional posterior predictive checks.

As such, Models 10 was run in cmdstanr with a larger number of samples to use to perform posterior predictive checks (i.e., 36 chains, 2500 warmup, and 2000 post warmup draws, resulting in 72,000 post warmup samples). Using means and SD’s of posteriors for all parameters, we used an estimation-based method to generate predicted RT quantiles and predicted choice proportions for all task conditions, for all diagnostic groups.

11.1 Predicted RT Quantiles (Model 10)

Results show that the predicted RT quantiles (data in pink below) for Model 10 matched the observed data well (data in teal below). This was true for all 3 groups in all conditions (left panel below) and when we marginalized over the emotion condition (right panel below). The exception was cases in which there were a lower number of trials (e.g., indrect/forward/neutral/acc=0). This indicates the model is capable of making accurate predictions about RT Quantiles.

Note: Posterior predicted RT quantiles were similarly accurate when we marginalized over the emotion condition. However, we do not include these below for brevity.

11.2 Predicted Choice % (Model 10)

Results show that the predicted proportion of “yes” responses (data in pink below) for Model 10 matched the observed data well (data in teal below). This was true for all 3 groups in all conditions (left panel below) and when we marginalized over the emotion condition (right panel below). This indicates the model is capable of making accurate predictions about choice proportions.

12 Winning Model Selection

To summarize the previous model evaluation steps: * Both in the full sample and separate diagnostic groups, Model 10 was the winning model based on LOO model comparisons. It fit the data considerably better than the next best-fitting model (Model 9) based on differences in ELPD. * A confusion matrix indicated that it was possible to arbitrate between Model 10 and other models. * Model 10 showed good recovery of group- and subject-level parameters. * Model 10 made accurate posterior predictions of RT quantiles and choice proportions across task conditions and diagnostic groups.

Considering all of the previous model evaluation steps, Model 10 was selected as the winning model. This model allowed all parameters to vary by diagnostic group and also allowed drift rate to vary in response to changes in gaze direction, head orientation, and facial emotion of stimuli. This is a plausible model account of behavior as the drift rate is known to be sensitive to the physical features of stimuli.

13 Final Fit of Winning Model

After selecting Model 10 as our winning model, we ran a “final fit” of this model using a larger number of samples. We did this to achieve an effective sample size (ESS) of >10,000 for all parameters, to ensure parameters we would be interpreting were sufficiently stable. The final fit of Model 10 was run in cmdstanr with 2,500 warm up samples and a total of 216,000 post warm up draws (obtained over 36 chains). Convergence checks are reported below and indicate that chains converged to target distributions.

13.1 Convergence Check (Model 10)

In the final fit of the winning model (Model 10), there were 0 divergent transitions, the ESS was >10,000 for all variables, trace plots were well-mixed, Rhat values were close to 1 (and all >1.1), and autocorrelation was low but a lag of ~30. Together this indicated that posteriors were stable and chains had converged to target distributions.

13.1.1 Rhat/ESS

13.1.2 Trace Plots

13.1.3 Autocorrelation

Note: We originally examined all 36 chains for autocorrelation. But here we randomly select 6 and display those (for brevity)

Analysis Supplement

14 Group/Condition Effects

We examined group and condition effects using group-level posteriors for all parameters. Between group differences were tested in a pairwise manner by subtracting the posterior for given parameter for one group from another. If the 90% highest density interval (HDI) of the posterior difference did not contain zero, then the difference was considered ‘credible’. The same process was followed to evaluate the credibility of task condition effects within groups. This approach was appropriate because the parameterization of the models inherently accounted for the shared group-level variability of within-group effects.

Results presented below and detail group differences in threshold separation, start point, NDT, and drift rate parameters. Because the winning model (Model 10) accounted for within-group condition effects on drift rates–for different gaze, head, and emotion conditions–we also present task condition effects on drift rate below (both in isolation [marginalized across group] and in combination with group effects).

14.1 Threshold Separation, Start Point, NDT (Group Effects)

Results show that threshold separation is credibly higher in SZ than HC, but no other group differences are credible for threshold separation. The three groups exhibited comparable values for start point and NDT parameters.

14.2 Drift Rate (Group/Gaze/Head/Emo Effects)

We looked at group differences in drift rates broken down by gaze and head conditions. This was done separately by neutral and fearful emotions. Results showed credible group differences in drift rates within gaze and head conditions. None of the groups exhibited group differences when gaze was direct and heads were deviated (this was true for both fearful and neutral emotions). When heads were forward and gaze was direct, HC had credibly higher drift rates than BD in both emotion conditions. When heads were forward and gaze was direct, HC had credibly higher drift rates than SZ only in the fearful emotion condition.

14.3 Drift Rate (Group/Gaze/Head Effects)

These results are reported in the main text. Here, we look at group differences in group-level drift rate parameters in a pairwise manner. This is done separately for drift rates in gaze and head conditions, after marginalizing over the emotion conditions.

14.4 Drift Rate (Group Effects)

We marginalized over all task conditions (emo, then head, then gaze) for drift rates to obtain a measure of overall evidence accumulation efficiency for all three groups. Results showed that HC had credibly higher drift rates than SZ and BD when samples were marginalized over all task conditions.

14.5 Drift Rate (Emo Effects)

We examined the influence of emotion on drift rates within gaze/head conditions and diagnostic groups to get an overall indicator of the effect of emotion on drift rates. Results show that there are not credible differences in drift rates between emotion conditions (i.e., all of the 90% HDI’s below contain zero) in any of the groups or head/gaze conditions. Because of this, we chose to marginalize over emotions conditions for drift rates and drift bias to simplify the analyses in the main text. We chose not to marginalize over gaze and head orientation because we did find at least some credible influence of gaze/head cues (reported below) on drift rates within groups. Additionally, for correlation analyses, we chose to further marginalize over gaze direction for drift rates to reduce the number of comparisons.

Marginalizing over emotion conditions: Marginalizing over emotion conditions may appear counterintuitive when the winning model—accounting for the influences of gaze, head, and emotion on drift rates—outperformed other models that did not also account for the influence of emotion on drift rates. This is an open question, but it is likely that Model 10 captured subtle nuances in how emotion influenced evidence accumulation that helped improve out-of-sample predictions but were not sufficiently large to yield credible condition-level effects, when tested at the group-level as we did.

One explanation is that Model 10 better captured variability at the subject-level and not the group-level as our tests of condition-level effects examined. In other words, the lack of within-group differences in drift rates based on emotion may not have occurred universally within the subjects of each group. Rather, condition-level emotions effects may have occurred for a subset of individuals within each group. In this case, the added variability introduced by emotion effects for some—but not all—subjects would have been best captured by Model 10. However, if this occurred in just a subset of subjects, these effects would not be sufficiently influential to produce group-level emotion effects in the group-level hyperparameters for the drift rates.

Although we chose to marginalize over samples in the analyses in the main text, for completeness and transparency we also report the tests of group differences in drift rates before marginalizing over the emotion condition in the supplement section named “Drift Rate (Group/Gaze/Head/Emo Effects)”.

Posterior Differences in Drift Rates Between Emotion Conditions (by Dx Group, Gaze Direction, Head Orientation)
group cond_gaze cond_head emo_contrast HDI_lo HDI_hi mean
hc direct forward neutral-fearful -0.379 0.271 -0.057
bd direct forward neutral-fearful -0.266 0.246 -0.014
sz direct forward neutral-fearful -0.327 0.463 0.076
hc indirect forward neutral-fearful -0.251 0.368 0.054
bd indirect forward neutral-fearful -0.264 0.280 0.015
sz indirect forward neutral-fearful -0.492 0.334 -0.079
hc direct deviated neutral-fearful -0.333 0.449 -0.333
bd direct deviated neutral-fearful -0.108 0.455 0.178
sz direct deviated neutral-fearful -0.168 0.695 0.264
hc indirect deviated neutral-fearful -0.229 0.163 -0.031
bd indirect deviated neutral-fearful -0.367 0.084 -0.142
sz indirect deviated neutral-fearful -0.549 0.131 -0.208

14.6 Drift Rate (Gaze Effects)

After marginalizing over emotion conditions, we examined the influence of gaze direction on drift rates in all three groups. Results show facilitation effects of gaze direction in all three groups for deviated heads (higher drift rates for indirect gaze when heads are deviated). However, we don’t see this as consistently for forwards heads.

Posterior Differences in Drift Rates Between Gaze Direction Conditions (by Dx Group, Head Orientation)
group cond_head gaze_contrast HDI_lo HDI_hi mean
hc forward direct-indirect -0.511 0.130 -0.186
bd forward direct-indirect 0.154 0.689 0.422
sz forward direct-indirect -0.120 0.690 0.281
hc deviated direct-indirect -2.895 -2.295 -2.597
bd deviated direct-indirect -1.814 -1.291 -1.551
sz deviated direct-indirect -1.890 -1.105 -1.500

14.7 Drift Rate (Head Effects)

After marginalizing over emotion conditions, we examined the influence of head orientation on drift rates in all three groups. Results show credible differences in drift rate based on the head orientation direction of stimuli, such that individuals show more efficient evidence accumulation for forward vs deviated heads. This is found in all three groups.

Posterior Differences in Drift Rates Between Head Orientation Conditions (by Dx Group, Gaze Direction)
group cond_gaze head_contrast HDI_lo HDI_hi mean
hc direct forward-deviated 1.163 1.669 1.422
bd direct forward-deviated 0.836 1.222 1.027
sz direct forward-deviated 0.639 1.224 0.934
hc indirect forward-deviated -1.177 -0.806 -0.989
bd indirect forward-deviated -1.126 -0.768 -0.947
sz indirect forward-deviated -1.115 -0.581 -0.847

14.8 Drift Bias (Group/Head Effects)

These results are presented in the main text. After marginalizing over emotion conditions, we examined the influence of diagnostic group on drift bias within forward and deviated head orientations. Results showed that HC had credibly lower drift bias than BD in forward and deviated head conditions. HC only showed credibly lower drift bias than SZ in deviated head conditions.

14.9 Drift Bias (Group Effects)

We examined the influence of diagnostic group on drift bias after marginalizing over emotion and head conditions to get an an overall indicator of the effect of group on drift bias. Results showed that, in general, HC had credibly lower drift bias than BD and SZ.

14.10 Drift Bias (Head Effects)

We calculated 90% HDIs for drift biases within both head orientation conditions after marginalizing over emotion conditions and groups. Results showed that biases toward self referential choices tended to be higher when heads were forward and lower when heads were deviated.

Drift Bias by Head Orientation Marginalized Over Group
head hdi_lo hdi_hi mean
forward -0.011 0.183 0.086
deviated -1.035 -0.849 -0.941

15 Preprocessing Correlates

Before conducting analyses on individual differences (correlations, regressions), we performed necessary preprocessing steps on our variables of interest.

  1. For subject-level drift rates and drift biases, we marginalized over emotion conditions (as was done for group-level parameters). To reduce the number of tests/comparisons correlations and regressions, we also marginalized over gaze conditions for drift rates. This was done by flipping the sign of drift rates for indirect gaze (which were originally negative-going) and then averaging over subject-level posteriors for direct and indirect gaze. This was done separately for both forward and deviated head conditions. As a result, for all subejcts, we had measures of drift rates and drift biases in both forward and deviated head conditions.

  2. To retain as much data as possible, we also winsorized the outermost .01 of data for variables that contained outliers.

16 SDT Analysis

We performed additional signal detection analyses to extract measures of sensitivity (discriminability parameter) and bias (criterion parameter) from participants’ choices on the gaze task. To do so, we programmed an equal variance gaussian SDT model in Stan that accounted for effects of head orientation on participant’s discriminability and criterion values. The SDT model was a hierarchical Bayesian model with weakly informative priors. These were fit separately to participants in SZ, BD, and HC groups (equivalent to how between-group effects were programmed in DDM’s). Models were run with 1,000 warmup samples and a total of 40,000 postwarmup samples obtained over several chains. As with DDM’s convergence checks were performed to ensure models had converged. There were no divergences for any model fits. Moreover, for all parameters of all SDT models fits, Rhat values were close to 1 (all were <1.1), trace plots were well-mixed, autocorrelation was low by a lag of ~30, and bulk and tail ESS were >10,000. Together, this suggested that chains had converged to their target distributions.

Note: For brevity we do not include outputs of convergence checks for SDT models here.

17 Interpreting Bayes Factors

For all statistical tests reported below, we calculate Bayes Factors (BF) and use those as an additional piece of information indicating the strength of an observed result. BF values index the evidence for the alternative hypothesis compared against that of the null hypothesis. BF > 1 favors the alternative hypothesis and BF < 1 favors the null hypothesis.

  • For demographic group differences: alternative hypothesis = difference between groups is not zero; null hypothesis = difference between groups is zero.
  • For correlations: alternative hypothesis = association is not zero; null hypothesis = association is zero.
  • For regression model comparisons: alternative hypothesis = favors the full model; null hypothsis = favors the reduced model.
  • For regression predictors: alternative hypothesis = value of predictor is not zero; null hypothesis = value of predictor is zero.

We then interpret each BF using the ranges below, from Lee and Wagenmakers (2014):

Bayes Factor Interpretation Scheme from Lee and Wagenmakers (2014)
BF Value Evidence Favors Strength of Evidence
>100 Alternative Extreme
30-100 Alternative Very Strong
10-30 Alternative Strong
3-10 Alternative Moderate
1-3 Alternative Anecdotal
1 No Evidence
0.33-1 Null Anecdotal
0.1-0.33 Null Moderate
0.03-0.1 Null Strong
0.01-0.03 Null Very Strong
<0.01 Null Extreme

18 Sample Demographics

Tests below are Bayesian t-tests, ANOVA, and proportion analyses run using ttestBF, anovaBF, and proportion BF in the ‘BayesFactor’ R package (Morey, 2022).

Sample Characteristics
HC (M) HC (SD) HC (N) BD (M) BD (SD) BD (N) SZ (M) SZ (SD) HC (N) GroupDiff PostHoc
Demographic
Age 41.56 12.92 34 41.16 11.51 37 41.61 13.3 28 BF=0.09 (Null/Strong) BD≈HC; BD≈SZ; HC≈SZ
Sex (% Female) 0.38 – 34 0.46 – 37 0.29 – 28 BF=0.2 (Null/Moderate) BD≈HC; BD≈SZ; HC≈SZ
Education (Years) 16.72 2.26 32 15.39 2.36 36 13.54 1.88 28 BF=11993.66 (Alternative/Extreme) BD<HC; BD>SZ; HC>SZ
Parental Education (Years) 14.67 2.75 30 15.55 2.7 22 14.54 3.68 28 BF=0.19 (Null/Moderate) BD≈HC; BD≈SZ; HC≈SZ
Race
White (%) 0.72 – 32 0.89 – 36 0.67 – 27 BF=0.75 (Null/Anecdotal) BD≈HC; BD>SZ; HC≈SZ
Black or African American (%) 0.12 – 32 0.06 – 36 0.26 – 27 BF=0.49 (Null/Anecdotal) BD≈HC; BD<SZ; HC≈SZ
Multiracial (%) 0.06 – 32 0.03 – 36 0.04 – 27 BF=0.02 (Null/Very Strong) BD≈HC; BD≈SZ; HC≈SZ
American Indian or Alaska Native (%) 0.03 – 32 0.03 – 36 0 – 27 BF=0.01 (Null/Very Strong) BD≈HC; BD≈SZ; HC≈SZ
Asian (%) 0.06 – 32 0 – 36 0 – 27 – –
Hispanic (%) 0 – 32 0 – 36 0.04 – 27 – –
Clinical
Illness Duration – – 0 24.32 12.27 37 21.07 12.97 28 BF=0.4 (Null/Anecdotal) BD≈SZ
Diagnosis
Schizophrenia (%) – – 0 0 – 37 0.75 – 28 – –
Schizoaffective (%) – – 0 0 – 37 0.25 – 28 – –
Bipolar I (%) – – 0 1 – 37 0 – 28 – –
Symptoms
ASRM 1.21 1.93 14 3.44 3.43 36 – – 0 BF=2.33 (Alternative/Anecdotal) BD>HC
BDI 1.26 1.68 27 11.83 10.21 36 10.37 8.47 27 BF=5937.23 (Alternative/Extreme) BD>HC; BD≈SZ; HC<SZ
SAPS-Hallucination – – 0 – – 0 5.79 6.64 28 – –
SAPS-Delusion – – 0 – – 0 8.86 9.58 28 – –
SANS-Motivation – – 0 – – 0 0.66 0.52 27 – –
SANS-Expressive – – 0 – – 0 0.69 0.69 27 – –
Medications
Antipsychotic User (%) – – 0 0.44 – 36 0.93 – 28 – –
Antidepressant User (%) – – 0 0.56 – 36 0.32 – 28 – –
Mood Stabilizer User (%) – – 0 0.72 – 36 0.18 – 28 – –
Stimulant User (%) – – 0 0.18 – 33 0.04 – 28 – –
Hypnotic User (%) – – 0 0.09 – 33 0.04 – 28 – –
Anxiolytic User (%) – – 0 0.24 – 33 0.21 – 28 – –
Anticholinergic User (%) – – 0 0 – 33 0.18 – 28 – –
CPZeq – – 0 85 136.44 33 489.42 397.68 27 BF=13339.77 (Alternative/Extreme) BD<SZ
General/Social Cognition
Cognition-General: BACS 0.52 0.44 18 – – 0 -0.33 0.71 28 BF=465.13 (Alternative/Extreme) HC>SZ
Cognition-Social: MSCEIT 0.47 0.64 18 – – 0 -0.34 0.7 27 BF=97.08 (Alternative/Very Strong) HC>SZ
Gaze Task Performance
Gaze Task: RT (ms) 681.46 102.05 34 750.01 143.67 38 807.82 166.63 28 BF=14.8 (Alternative/Strong) BD>HC; BD≈SZ; HC<SZ
Gaze Task: Accuracy 0.83 0.07 34 0.77 0.08 38 0.8 0.09 28 BF=2.96 (Alternative/Anecdotal) BD<HC; BD≈SZ; HC≈SZ
Gaze Task: Criterion 0.49 0.48 34 0.23 0.46 38 0.29 0.49 28 BF=0.92 (Null/Anecdotal) BD<HC; BD≈SZ; HC≈SZ
Gaze Task: Discriminability 2.63 0.6 34 1.95 0.76 38 2.17 0.82 28 BF=54.55 (Alternative/Very Strong) BD<HC; BD≈SZ; HC>SZ
Social Functioning
Social Functioning: SAS-SR 4.2 0.46 18 – – 0 3.61 0.46 27 BF=184.57 (Alternative/Extreme) HC>SZ

19 Correlations

We ran exploratory Correlations Between DDM Parameters, traditional performance metrics (Gaze accuracy, RT, signal detection discriminability and criterion), general cognition measures (BACS), emotion-based social cognition measures (MSCEIT), SZ symptoms (SANS/SAPS), and mood symptoms (BDI/ASRM). This was done using functions from the BayesFactor R package (Morey, 2022). For each correlation, we calculate the mean and 90% HDI of the correlation coefficient, and the Bayes factor (BF). We interpret associations using the 90% HDI: intervals that do not contain zero are considered credible. For each credible association, we also consider – as an added source of information – the strength of the evidence using its BF.

The sections that follow present a correlation matrix plot (to illustrate the nature of relationships between all measures); sensitivity analyses to test the robustness of observed relationships; and a series of post hoc follow-up tests, including correlations with SZ and BD, as well as associations looking at paranoia in SZ specifically.

19.1 Correlations

Below is a full correlation matrix plot with correlations to illustrate the nature of relationships between DDM parameters (in SZ, BD, HC), traditional metrics (in SZ, BD, HC), BACS/MSCEIT (in SZ, HC), SZ symptoms (SAPS/SANS; in SZ), depressive symptoms (BDI; in SZ, BD, HC), and mania symptoms (ASRM; in BD, HC).

In the plot below, large text displayed in cells below represent the posterior means of the correlation coefficient. This value is equivalent to the Pearson R value. The magnitude and direction of the posterior mean estimates control the color scheme of the correlation matrix (positive-going relationships are in red and negative-going are in blue). 90% HDI’s that do not contain zero are considered credible (marked with a ’*’ below).

Correlations Between DDM Parameters, Traditional Metrics, and Clinical Metrics (Full Sample)
Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
Drift Rate (Forward) BACS 0.63 [0.5,0.77]* 103614.43 Extreme Alternative 46
Drift Rate (Forward) MSCEIT 0.38 [0.19,0.57]* 17.79 Strong Alternative 45
Drift Rate (Forward) ASRM Mania -0.14 [-0.36,0.08] 0.56 Anecdotal Null 50
Drift Rate (Forward) BDI Depression -0.18 [-0.35,-0.03]* 1.29 Anecdotal Alternative 90
Drift Rate (Forward) SANS Amotivation 0.08 [-0.2,0.37] 0.46 Anecdotal Null 27
Drift Rate (Forward) SAPS Delusion 0.19 [-0.05,0.48] 0.79 Anecdotal Null 28
Drift Rate (Forward) SAPS Hallucination 0.08 [-0.21,0.35] 0.46 Anecdotal Null 28
Drift Rate (Deviated) BACS 0.57 [0.42,0.72]* 6314.00 Extreme Alternative 46
Drift Rate (Deviated) MSCEIT 0.21 [-0.01,0.42] 1.02 Anecdotal Alternative 45
Drift Rate (Deviated) ASRM Mania -0.10 [-0.3,0.13] 0.42 Anecdotal Null 50
Drift Rate (Deviated) BDI Depression -0.23 [-0.39,-0.07]* 3.21 Moderate Alternative 90
Drift Rate (Deviated) SANS Amotivation -0.06 [-0.36,0.22] 0.45 Anecdotal Null 27
Drift Rate (Deviated) SAPS Delusion 0.05 [-0.22,0.34] 0.43 Anecdotal Null 28
Drift Rate (Deviated) SAPS Hallucination -0.03 [-0.31,0.26] 0.42 Anecdotal Null 28
Threshold Separation BACS -0.12 [-0.35,0.1] 0.47 Anecdotal Null 46
Threshold Separation MSCEIT -0.10 [-0.33,0.13] 0.44 Anecdotal Null 45
Threshold Separation ASRM Mania -0.02 [-0.24,0.2] 0.32 Moderate Null 50
Threshold Separation BDI Depression 0.06 [-0.1,0.23] 0.29 Moderate Null 90
Threshold Separation SANS Amotivation -0.14 [-0.44,0.12] 0.59 Anecdotal Null 27
Threshold Separation SAPS Delusion -0.10 [-0.37,0.2] 0.49 Anecdotal Null 28
Threshold Separation SAPS Hallucination -0.07 [-0.37,0.2] 0.45 Anecdotal Null 28
Start Point BACS 0.32 [0.11,0.51]* 5.65 Moderate Alternative 46
Start Point MSCEIT 0.12 [-0.12,0.33] 0.48 Anecdotal Null 45
Start Point ASRM Mania -0.01 [-0.23,0.21] 0.32 Moderate Null 50
Start Point BDI Depression 0.08 [-0.08,0.25] 0.34 Anecdotal Null 90
Start Point SANS Amotivation 0.05 [-0.24,0.33] 0.44 Anecdotal Null 27
Start Point SAPS Delusion 0.31 [0.05,0.56]* 2.26 Anecdotal Alternative 28
Start Point SAPS Hallucination 0.17 [-0.1,0.44] 0.67 Anecdotal Null 28
Drift Bias (Forward) BACS -0.01 [-0.25,0.22] 0.33 Anecdotal Null 46
Drift Bias (Forward) MSCEIT -0.13 [-0.37,0.08] 0.51 Anecdotal Null 45
Drift Bias (Forward) ASRM Mania 0.20 [-0.01,0.4] 1.00 Anecdotal Alternative 50
Drift Bias (Forward) BDI Depression 0.10 [-0.07,0.27] 0.39 Anecdotal Null 90
Drift Bias (Forward) SANS Amotivation -0.29 [-0.54,-0.01]* 1.87 Anecdotal Alternative 27
Drift Bias (Forward) SAPS Delusion -0.25 [-0.49,0.03] 1.22 Anecdotal Alternative 28
Drift Bias (Forward) SAPS Hallucination -0.38 [-0.65,-0.14]* 6.08 Moderate Alternative 28
Drift Bias (Deviated) BACS -0.09 [-0.3,0.15] 0.41 Anecdotal Null 46
Drift Bias (Deviated) MSCEIT -0.19 [-0.41,0.03] 0.84 Anecdotal Null 45
Drift Bias (Deviated) ASRM Mania 0.11 [-0.12,0.31] 0.44 Anecdotal Null 50
Drift Bias (Deviated) BDI Depression 0.08 [-0.09,0.24] 0.34 Anecdotal Null 90
Drift Bias (Deviated) SANS Amotivation -0.16 [-0.43,0.12] 0.64 Anecdotal Null 27
Drift Bias (Deviated) SAPS Delusion -0.24 [-0.51,0.03] 1.16 Anecdotal Alternative 28
Drift Bias (Deviated) SAPS Hallucination -0.30 [-0.57,-0.03]* 2.29 Anecdotal Alternative 28
Accuracy BACS 0.50 [0.32,0.67]* 477.78 Extreme Alternative 46
Accuracy MSCEIT 0.16 [-0.05,0.39] 0.69 Anecdotal Null 45
Accuracy ASRM Mania -0.07 [-0.27,0.17] 0.37 Anecdotal Null 50
Accuracy BDI Depression -0.09 [-0.27,0.07] 0.37 Anecdotal Null 90
Accuracy SANS Amotivation -0.04 [-0.34,0.23] 0.43 Anecdotal Null 27
Accuracy SAPS Delusion 0.08 [-0.2,0.36] 0.45 Anecdotal Null 28
Accuracy SAPS Hallucination -0.04 [-0.32,0.24] 0.43 Anecdotal Null 28
RT BACS -0.18 [-0.4,0.04] 0.79 Anecdotal Null 46
RT MSCEIT -0.23 [-0.45,-0.02]* 1.40 Anecdotal Alternative 45
RT ASRM Mania 0.12 [-0.1,0.34] 0.48 Anecdotal Null 50
RT BDI Depression 0.09 [-0.07,0.28] 0.37 Anecdotal Null 90
RT SANS Amotivation -0.14 [-0.43,0.13] 0.58 Anecdotal Null 27
RT SAPS Delusion -0.17 [-0.44,0.09] 0.70 Anecdotal Null 28
RT SAPS Hallucination -0.18 [-0.45,0.09] 0.75 Anecdotal Null 28
SDT-Discriminability BACS 0.58 [0.43,0.74]* 10708.55 Extreme Alternative 46
SDT-Discriminability MSCEIT 0.33 [0.13,0.54]* 6.19 Moderate Alternative 45
SDT-Discriminability ASRM Mania -0.16 [-0.37,0.04] 0.66 Anecdotal Null 50
SDT-Discriminability BDI Depression -0.18 [-0.34,-0.01]* 1.07 Anecdotal Alternative 90
SDT-Discriminability SANS Amotivation 0.03 [-0.26,0.31] 0.42 Anecdotal Null 27
SDT-Discriminability SAPS Delusion 0.14 [-0.15,0.41] 0.56 Anecdotal Null 28
SDT-Discriminability SAPS Hallucination 0.08 [-0.21,0.35] 0.46 Anecdotal Null 28
SDT-Criterion BACS -0.07 [-0.29,0.15] 0.37 Anecdotal Null 46
SDT-Criterion MSCEIT 0.10 [-0.13,0.32] 0.44 Anecdotal Null 45
SDT-Criterion ASRM Mania -0.16 [-0.37,0.06] 0.67 Anecdotal Null 50
SDT-Criterion BDI Depression -0.08 [-0.24,0.09] 0.32 Moderate Null 90
SDT-Criterion SANS Amotivation 0.18 [-0.09,0.46] 0.74 Anecdotal Null 27
SDT-Criterion SAPS Delusion 0.18 [-0.1,0.45] 0.71 Anecdotal Null 28
SDT-Criterion SAPS Hallucination 0.35 [0.12,0.6]* 3.97 Moderate Alternative 28
Note:
* = Credible association (i.e., 90% HDI does not contain zero); Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); BF_Evidence_Favors = whether evidence favors the null hypothesis (i.e., the association between variable 1 and 2 is 0) or the alternative hypothesis (i.e., an association between variable 1 and 2 is not 0); N = Number of subjects with complete data included in correlation.

19.2 Sensitivity Analyses

Because some of the clinical metrics (SAPS, SANS) were zero-inflated and clinical metrics and/or performance metrics contained potentially influential observations, we ran sensitivity analyses to see if credible relationships identified above held when 1) potentially influential cases were removed, and 2) subjects with ‘0s’ for those Hallucination/Delusion symptom dimensions were removed.

Results of sensitivity analyses show that the correlation between hallucinations and drift bias for forward heads remains credible when outliers and subjects with hallucinations scores = 0 are removed. In fact, the correlation is strengthened somewhat. This suggests the correlation is not merely the result of influential cases. However, the correlations between hallucinations and drift bias for deviated heads–as well as SDT-Criterion–were no longer credible after removing potentially influential observations. As such, we will only interpret the correlation between hallucinations and drift bias (forward). Results also show that the correlation between delusions and start point is stronger when potentially influential cases and subject with delusions = ‘0’ are removed, suggesting the relationship is not the by-product of influential observations.

Correlations Between DDM Parameters, Traditional Metrics, and Clinical Metrics After Removing Potentially Influential and Zero-Inflated Cases
Sensitivity_Analysis Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
Halluc*Drift Bias (Forward) - Influential Cases Removed SAPS Hallucination Drift Bias (Forward) -0.48 [-0.69,-0.25]* 27.81 Strong Alternative 27
Halluc*Drift Bias (Deviated) - Influential Cases Removed SAPS Hallucination Drift Bias (Deviated) -0.13 [-0.4,0.16] 0.55 Anecdotal Null 27
Halluc*SDT-Criterion - Influential Cases Removed SAPS Hallucination SDT-Criterion 0.23 [-0.07,0.48] 1.07 Anecdotal Alternative 27
Halluc*Drift Bias (Forward) - SAPS Halluc=0 Cases Removed SAPS Hallucination Drift Bias (Forward) -0.41 [-0.7,-0.11]* 3.83 Moderate Alternative 17
Delusion*Start Point - Influential Cases Removed SAPS Delusion Start Point 0.40 [0.15,0.65]* 7.27 Moderate Alternative 26
Delusion*Start Point - SAPS Delusion=0 Cases Removed SAPS Delusion Start Point 0.33 [0.03,0.64]* 2.44 Anecdotal Alternative 21
Amotiv*Drift Bias (Forward) - Influential Cases Removed SANS Amotivation Drift Bias (Forward) -0.22 [-0.49,0.07] 1.06 Anecdotal Alternative 26
Note:
* = Credible association (i.e., 90% HDI does not contain zero); Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); N = Number of subjects with complete data included in correlation.

19.3 Post-Hoc Correlations

Following the primary correlations above, we performed a series of post-hoc correlation analyses.

19.3.1 Within SZ Only

We repeated the same key correlations of interest (from above) within the SZ group only. We exclude SAPS hallucinations and delusions below because the correlations above were already within SZ only. Credible associations between drift rates and BACS and drift rate and SDT-discriminability remain credible in SZ. The same is true for the association between drift rate for forward heads and MSCEIT. Several other credible associations also emerge but they are supported by only anecdotal evidence and we chose not to interpret those.

Correlations (Within SZ Only) Between DDM Parameters, Traditional Metrics, Clinical Metrics
Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
Drift Rate (Forward) BACS 0.43 [0.2,0.66]* 14.14 Strong Alternative 28
Drift Rate (Forward) MSCEIT 0.30 [0.05,0.58]* 1.96 Anecdotal Alternative 27
Drift Rate (Forward) BDI Depression 0.07 [-0.21,0.34] 0.45 Anecdotal Null 27
Drift Rate (Deviated) BACS 0.36 [0.12,0.61]* 4.19 Moderate Alternative 28
Drift Rate (Deviated) MSCEIT 0.17 [-0.11,0.45] 0.68 Anecdotal Null 27
Drift Rate (Deviated) BDI Depression -0.08 [-0.37,0.18] 0.48 Anecdotal Null 27
Threshold Separation BACS -0.10 [-0.37,0.18] 0.50 Anecdotal Null 28
Threshold Separation MSCEIT -0.03 [-0.34,0.23] 0.42 Anecdotal Null 27
Threshold Separation BDI Depression -0.15 [-0.43,0.13] 0.62 Anecdotal Null 27
Start Point BACS 0.25 [0,0.53] 1.19 Anecdotal Alternative 28
Start Point MSCEIT 0.24 [-0.03,0.49] 1.19 Anecdotal Alternative 27
Start Point BDI Depression 0.29 [0.04,0.57]* 1.92 Anecdotal Alternative 27
Drift Bias (Forward) BACS -0.05 [-0.34,0.22] 0.43 Anecdotal Null 28
Drift Bias (Forward) MSCEIT -0.29 [-0.55,-0.03]* 1.86 Anecdotal Alternative 27
Drift Bias (Forward) BDI Depression -0.20 [-0.48,0.07] 0.85 Anecdotal Null 27
Drift Bias (Deviated) BACS -0.03 [-0.29,0.26] 0.42 Anecdotal Null 28
Drift Bias (Deviated) MSCEIT -0.33 [-0.58,-0.08]* 2.95 Anecdotal Alternative 27
Drift Bias (Deviated) BDI Depression -0.30 [-0.55,-0.04]* 2.08 Anecdotal Alternative 27
Accuracy BACS 0.31 [0.07,0.58]* 2.48 Anecdotal Alternative 28
Accuracy MSCEIT 0.09 [-0.18,0.37] 0.49 Anecdotal Null 27
Accuracy BDI Depression -0.06 [-0.32,0.25] 0.44 Anecdotal Null 27
RT BACS -0.08 [-0.34,0.21] 0.46 Anecdotal Null 28
RT MSCEIT -0.16 [-0.45,0.11] 0.66 Anecdotal Null 27
RT BDI Depression -0.14 [-0.42,0.15] 0.60 Anecdotal Null 27
SDT-Discriminability BACS 0.37 [0.15,0.63]* 5.14 Moderate Alternative 28
SDT-Discriminability MSCEIT 0.25 [-0.01,0.53] 1.25 Anecdotal Alternative 27
SDT-Discriminability BDI Depression -0.02 [-0.31,0.27] 0.42 Anecdotal Null 27
SDT-Criterion BACS -0.02 [-0.31,0.24] 0.41 Anecdotal Null 28
SDT-Criterion MSCEIT 0.29 [0.03,0.55]* 1.87 Anecdotal Alternative 27
SDT-Criterion BDI Depression 0.24 [-0.03,0.51] 1.11 Anecdotal Alternative 27
Note:
Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); N = Number of subjects with complete data included in correlation.

19.3.2 Within BD Only

We repeated the same key correlations of interest (from above) within the BD group only. We only include ASRM and BDI because the other measures were not collected in BD. Results showed that neither ASRM or BDI showed credible associations with DDM parameters or traditional metrics in the BD sample.

Correlations (Within BD Only) Between DDM Parameters, Traditional Metrics, Clinical Metrics
Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
Drift Rate (Forward) ASRM Mania -0.07 [-0.31,0.18] 0.41 Anecdotal Null 36
Drift Rate (Forward) BDI Depression 0.07 [-0.18,0.31] 0.42 Anecdotal Null 36
Drift Rate (Deviated) ASRM Mania 0.05 [-0.2,0.3] 0.39 Anecdotal Null 36
Drift Rate (Deviated) BDI Depression 0.01 [-0.22,0.28] 0.37 Anecdotal Null 36
Threshold Separation ASRM Mania -0.04 [-0.27,0.23] 0.38 Anecdotal Null 36
Threshold Separation BDI Depression 0.11 [-0.12,0.37] 0.48 Anecdotal Null 36
Start Point ASRM Mania 0.06 [-0.2,0.31] 0.39 Anecdotal Null 36
Start Point BDI Depression 0.07 [-0.19,0.31] 0.42 Anecdotal Null 36
Drift Bias (Forward) ASRM Mania 0.10 [-0.14,0.35] 0.46 Anecdotal Null 36
Drift Bias (Forward) BDI Depression 0.16 [-0.1,0.41] 0.63 Anecdotal Null 36
Drift Bias (Deviated) ASRM Mania -0.01 [-0.25,0.25] 0.37 Anecdotal Null 36
Drift Bias (Deviated) BDI Depression 0.10 [-0.15,0.35] 0.46 Anecdotal Null 36
Accuracy ASRM Mania 0.00 [-0.28,0.23] 0.37 Anecdotal Null 36
Accuracy BDI Depression 0.13 [-0.13,0.38] 0.53 Anecdotal Null 36
RT ASRM Mania 0.04 [-0.22,0.28] 0.38 Anecdotal Null 36
RT BDI Depression 0.06 [-0.21,0.3] 0.40 Anecdotal Null 36
SDT-Discriminability ASRM Mania -0.09 [-0.33,0.17] 0.43 Anecdotal Null 36
SDT-Discriminability BDI Depression 0.04 [-0.2,0.33] 0.38 Anecdotal Null 36
SDT-Criterion ASRM Mania -0.08 [-0.31,0.19] 0.42 Anecdotal Null 36
SDT-Criterion BDI Depression -0.13 [-0.39,0.12] 0.54 Anecdotal Null 36
Note:
Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); N = Number of subjects with complete data included in correlation.

19.3.3 Start Point and Paranoia

Given the relationship we observed between start point and SAPS Delusions, we ran a post hoc follow-up test to examine whether this relationship was evident within paranoia symptoms specifically. To achieve this, we calculated a paranoia factor using results of a factor analysis on SAPS items from Peralta (1999). That study identified that two items from the SAPS – delusions of persecution and delusions of reference – loaded onto a ‘paranoia’ factor. In our data, we generated a paranoia factor by scaling participants’ scores on these two items by the factor loadings of Peralta (1999) and summing them. Then we ran a correlation (using a Bayesian approach) on this paranoia factor and the DDM start point parameter.

Results revealed a credible positive correlation, such that SZ patients with more paranoia also had higher start points (i.e., greater initial self-referential biases) on the gaze task.

Correlation Between Start Point and SAPS Paranoia Factor
Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
Start Point SAPS Paranoia Factor 0.31 [0.08,0.59]* 2.37 Anecdotal Alternative 28
Note:
Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); N = Number of subjects with complete data included in correlation.

19.3.4 Antipsychotics and Performance

To assess whether antipsychotic doses (CPZeq) were related to measures that tap processing speed, we ran correlations between CPZeq and measures that were sensitive to processing speed. This was done exclusively within participants in the SZ and BD groups taking antipsychotic medicaitons. This included BACS, DDM parameters, and traditional performance metrics on the gaze task. MSCEIT was not included as it does not impose a time limit on participants.

Results did not show credible associations between CPZ dose and any of the DDM parameters, BACS scores, or other gaze task performance metrics. This suggested that observed results were not merely the result of the influence of antipsychotic dosing.

Correlations Between Antipsychotic Dose and Performance
Variable1 Variable2 Mean HDI BF BF_Evidence_Strength BF_Evidence_Favors N
CPZeq BACS -0.03 [-0.32,0.25] 0.43 Anecdotal Null 27
CPZeq Threshold_Separation 0.14 [-0.05,0.35] 0.58 Anecdotal Null 60
CPZeq Start_Point -0.06 [-0.27,0.13] 0.33 Moderate Null 60
CPZeq Drift_Rate_Forward -0.14 [-0.35,0.05] 0.57 Anecdotal Null 60
CPZeq Drift_Rate_Deviated -0.07 [-0.27,0.14] 0.34 Anecdotal Null 60
CPZeq Drift_Bias_Forward -0.13 [-0.33,0.06] 0.54 Anecdotal Null 60
CPZeq Drift_Bias_Deviated 0.11 [-0.09,0.32] 0.44 Anecdotal Null 60
CPZeq Accuracy -0.13 [-0.35,0.05] 0.54 Anecdotal Null 60
CPZeq RT 0.09 [-0.12,0.28] 0.37 Anecdotal Null 60
CPZeq Criterion 0.01 [-0.18,0.22] 0.29 Moderate Null 60
CPZeq Discriminability -0.12 [-0.32,0.08] 0.51 Anecdotal Null 60
Note:
Mean = mean posterior estimate of correlation coefficient (equivalent to Pearson R); HDI = 90% HDI of the correlation coefficient; BF = Bayes Factor; BF_Evidence_Strength = Bayes Factor interpretation scheme based on Lee and Wagenmakers (2014); N = Number of subjects with complete data included in correlation.

20 Regressions

We ran separate hierarchical linear regressions to assess whether any of the DDM parameters and/or traditional gaze task performance metrics (accuracy, RT, SDT Criterion, SDT discriminability) could predict social functioning (Social Adjustment Scale-SR, Social/Leisure Sub scale) across SZ and HC, above and beyond diagnosis and common measures of general (BACS) and emotion-based social cognition (MSCEIT). We assessed these predictors on whether: 1) They were credible predictors of social functioning. Credible predictors were those in which the 90% HDI of the predictor coefficient did not contain zero. 2) They improved the out-of-sample predictive accuracy. This was done using LOO model comparisons.

These regression models were run in Stan via brms (Burkner, 2017) using standardized predictors and weakly informative priors (i.e., Normal(0,1) for each predictor). Models were sampled using 1000 warmup samples and 4000 postwarmup draws for each of 4 chains, resulting in 16000 total post warmup samples. The same procedure for assessing convergence (described above) indicated that parameters of all models had converged to their target distributions.

20.1 DDM Parameters Predicting Social Functioning

Start point predicts social functioning above and beyond diagnosis, general cognition, and social cognition. The model including start point showed increases in predictive accuracy relative to the null model, but the strength of evidence for both the full model (relative to the null) and strength of evidence for the start point predictor are both only anecdotal.

Bayesian Regression: Predicting Social Functioning from DDM Parameters
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -59.16 3.94 – – – –/–
Intercept 0.13 Null/Moderate 0 [-0.21, 0.23]
SZ_Dummy 12.99 Alternative/Strong -0.5 [-0.77, -0.22]*
BACS 0.19 Null/Moderate 0.1 [-0.17, 0.37]
MSCEIT 0.16 Null/Moderate -0.01 [-0.27, 0.25]
2 1 -59.66 4.08 -0.5 0.85 0.27 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.21, 0.22]
SZ_Dummy 15.06 Alternative/Strong -0.53 [-0.81, -0.24]*
BACS 0.31 Null/Moderate 0.18 [-0.12, 0.51]
MSCEIT 0.16 Null/Moderate 0 [-0.28, 0.25]
Drift_Rate_Forward 0.27 Null/Moderate -0.16 [-0.47, 0.15]
3 1 -59.12 3.73 0.03 1.15 0.36 Null/Anecdotal
Intercept 0.13 Null/Moderate 0 [-0.23, 0.21]
SZ_Dummy 21.49 Alternative/Strong -0.54 [-0.82, -0.27]*
BACS 0.38 Null/Anecdotal 0.22 [-0.1, 0.55]
MSCEIT 0.17 Null/Moderate -0.05 [-0.31, 0.22]
Drift_Rate_Deviated 0.36 Null/Anecdotal -0.21 [-0.51, 0.07]
4 1 -60.03 4.02 -0.88 0.65 0.16 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.21, 0.22]
SZ_Dummy 14.06 Alternative/Strong -0.51 [-0.77, -0.22]*
BACS 0.19 Null/Moderate 0.08 [-0.2, 0.36]
MSCEIT 0.16 Null/Moderate 0 [-0.28, 0.27]
Drift_Bias_Forward 0.16 Null/Moderate 0.08 [-0.14, 0.31]
5 1 -58.06 4.24 1.1 1.42 0.75 Null/Anecdotal
Intercept 0.13 Null/Moderate 0 [-0.22, 0.2]
SZ_Dummy 22.74 Alternative/Strong -0.54 [-0.83, -0.29]*
BACS 0.18 Null/Moderate 0.08 [-0.2, 0.34]
MSCEIT 0.16 Null/Moderate 0.02 [-0.24, 0.28]
Drift_Bias_Deviated 0.76 Null/Anecdotal 0.25 [0.03, 0.46]*
6 1 -57.14 3.39 2.02 1.99 1.55 Alternative/Anecdotal
Intercept 0.12 Null/Moderate 0 [-0.2, 0.2]
SZ_Dummy 17.98 Alternative/Strong -0.51 [-0.78, -0.25]*
BACS 0.36 Null/Anecdotal 0.2 [-0.07, 0.47]
MSCEIT 0.15 Null/Moderate -0.03 [-0.28, 0.22]
Start_Point 1.54 Alternative/Anecdotal -0.3 [-0.52, -0.07]*
7 1 -60.26 3.9 -1.1 0.26 0.14 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.21, 0.22]
SZ_Dummy 9.62 Alternative/Moderate -0.5 [-0.79, -0.23]*
BACS 0.2 Null/Moderate 0.1 [-0.19, 0.36]
MSCEIT 0.16 Null/Moderate -0.01 [-0.3, 0.25]
Threshold_Separation 0.13 Null/Moderate -0.01 [-0.24, 0.22]

20.2 Sensitivity Analysis (Control for CPZeq)

Results are the same when we control for antipsychotic dose.

Bayesian Regression (Sensitivity Analysis): Predicting Social Functioning from DDM Parameters after Controlling for Antipsychotic Dose
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -60.3 4.14 – – – –/–
Intercept 0.13 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 8.67 Alternative/Moderate -0.53 [-0.84, -0.23]*
BACS 0.2 Null/Moderate 0.1 [-0.18, 0.38]
MSCEIT 0.16 Null/Moderate -0.01 [-0.29, 0.26]
CPZeq 0.18 Null/Moderate 0.06 [-0.2, 0.34]
2 1 -61 4.15 -0.7 0.73 0.27 Null/Moderate
Intercept 0.14 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 8.79 Alternative/Moderate -0.54 [-0.84, -0.22]*
BACS 0.3 Null/Moderate 0.17 [-0.16, 0.5]
MSCEIT 0.16 Null/Moderate 0 [-0.29, 0.27]
CPZeq 0.18 Null/Moderate 0.01 [-0.28, 0.31]
Drift_Rate_Forward 0.27 Null/Moderate -0.15 [-0.5, 0.19]
3 1 -60.44 3.83 -0.14 1.1 0.35 Null/Anecdotal
Intercept 0.13 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 10.79 Alternative/Strong -0.55 [-0.85, -0.24]*
BACS 0.37 Null/Anecdotal 0.22 [-0.11, 0.55]
MSCEIT 0.18 Null/Moderate -0.04 [-0.32, 0.23]
CPZeq 0.16 Null/Moderate 0.02 [-0.25, 0.3]
Drift_Rate_Deviated 0.34 Null/Anecdotal -0.2 [-0.51, 0.09]
4 1 -61.23 4.34 -0.93 0.74 0.17 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.23, 0.21]
SZ_Dummy 11.49 Alternative/Strong -0.55 [-0.87, -0.24]*
BACS 0.19 Null/Moderate 0.08 [-0.19, 0.37]
MSCEIT 0.16 Null/Moderate 0.01 [-0.27, 0.27]
CPZeq 0.18 Null/Moderate 0.07 [-0.21, 0.34]
Drift_Bias_Forward 0.16 Null/Moderate 0.09 [-0.15, 0.31]
5 1 -59.53 4.42 0.77 1.55 0.69 Null/Anecdotal
Intercept 0.13 Null/Moderate 0 [-0.21, 0.22]
SZ_Dummy 14.38 Alternative/Strong -0.56 [-0.86, -0.25]*
BACS 0.18 Null/Moderate 0.07 [-0.18, 0.35]
MSCEIT 0.16 Null/Moderate 0.02 [-0.25, 0.28]
CPZeq 0.16 Null/Moderate 0.03 [-0.24, 0.29]
Drift_Bias_Deviated 0.74 Null/Anecdotal 0.24 [0.01, 0.46]*
6 1 -58.85 3.66 1.45 2.07 1.41 Alternative/Anecdotal
Intercept 0.13 Null/Moderate 0 [-0.21, 0.22]
SZ_Dummy 10.92 Alternative/Strong -0.53 [-0.82, -0.23]*
BACS 0.36 Null/Anecdotal 0.21 [-0.06, 0.5]
MSCEIT 0.16 Null/Moderate -0.02 [-0.28, 0.23]
CPZeq 0.16 Null/Moderate 0.03 [-0.22, 0.31]
Start_Point 1.35 Alternative/Anecdotal -0.3 [-0.53, -0.07]*
7 1 -61.42 4.14 -1.12 0.27 0.14 Null/Moderate
Intercept 0.14 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 9.36 Alternative/Moderate -0.53 [-0.84, -0.21]*
BACS 0.2 Null/Moderate 0.1 [-0.19, 0.37]
MSCEIT 0.16 Null/Moderate 0 [-0.27, 0.27]
CPZeq 0.18 Null/Moderate 0.06 [-0.22, 0.33]
Threshold_Separation 0.14 Null/Moderate -0.01 [-0.24, 0.21]

20.3 Post-Hoc Analyses

Results look similar in SZ. The direction of the relationship between the start point predictor and outcome is similar, but no longer credible. This isn’t surprising given the loss of small sample size from doing this within groups.

20.3.1 DDM Parameters as Predictors (SZ Only)

Bayesian Regression: Predicting Social Functioning from DDM Parameters (Within SZ Only)
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -38.23 3.74 – – – –/–
Intercept 0.17 Null/Moderate 0 [-0.31, 0.31]
BACS 0.72 Null/Anecdotal 0.32 [-0.03, 0.65]
MSCEIT 1.46 Alternative/Anecdotal -0.41 [-0.76, -0.08]*
2 1 -39.22 3.9 -0.99 0.73 0.27 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.33, 0.31]
BACS 0.9 Null/Anecdotal 0.38 [0, 0.77]
MSCEIT 1.24 Alternative/Anecdotal -0.39 [-0.75, -0.04]*
Drift_Rate_Forward 0.27 Null/Moderate -0.13 [-0.51, 0.25]
3 1 -38.67 3.66 -0.44 0.86 0.33 Null/Anecdotal
Intercept 0.19 Null/Moderate 0 [-0.31, 0.32]
BACS 1.21 Alternative/Anecdotal 0.41 [0.03, 0.8]*
MSCEIT 1.47 Alternative/Anecdotal -0.41 [-0.76, -0.05]*
Drift_Rate_Deviated 0.33 Null/Anecdotal -0.19 [-0.55, 0.16]
4 1 -39.48 3.71 -1.25 0.52 0.21 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.33, 0.3]
BACS 0.7 Null/Anecdotal 0.32 [-0.04, 0.67]
MSCEIT 1.22 Alternative/Anecdotal -0.42 [-0.79, -0.03]*
Drift_Bias_Forward 0.21 Null/Moderate -0.02 [-0.36, 0.33]
5 1 -39.18 3.61 -0.95 0.27 0.21 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.32, 0.32]
BACS 0.68 Null/Anecdotal 0.32 [-0.04, 0.68]
MSCEIT 1.08 Alternative/Anecdotal -0.41 [-0.79, -0.02]*
Drift_Bias_Deviated 0.22 Null/Moderate 0.01 [-0.35, 0.36]
6 1 -38.06 3.23 0.16 1.27 0.57 Null/Anecdotal
Intercept 0.18 Null/Moderate 0 [-0.31, 0.31]
BACS 1.07 Alternative/Anecdotal 0.38 [0.03, 0.73]*
MSCEIT 0.86 Null/Anecdotal -0.35 [-0.69, 0]
Start_Point 0.58 Null/Anecdotal -0.29 [-0.63, 0.05]
7 1 -38.92 3.52 -0.69 0.48 0.22 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.3, 0.32]
BACS 0.75 Null/Anecdotal 0.33 [-0.02, 0.69]
MSCEIT 1.5 Alternative/Anecdotal -0.42 [-0.77, -0.06]*
Threshold_Separation 0.22 Null/Moderate 0.09 [-0.23, 0.42]

20.3.2 DDM Parameters as Predictors (HC Only)

Bayesian Regression: Predicting Social Functioning from DDM Parameters (Within HC Only)
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -26.22 2.11 – – – –/–
Intercept 0.22 Null/Moderate 0 [-0.39, 0.35]
BACS 0.24 Null/Moderate -0.09 [-0.47, 0.3]
MSCEIT 2.19 Alternative/Anecdotal 0.48 [0.1, 0.86]*
2 1 -27.54 2.43 -1.33 1.38 0.34 Null/Anecdotal
Intercept 0.23 Null/Moderate 0 [-0.39, 0.37]
BACS 0.36 Null/Anecdotal -0.17 [-0.67, 0.33]
MSCEIT 2.07 Alternative/Anecdotal 0.5 [0.1, 0.91]*
Drift_Rate_Forward 0.35 Null/Anecdotal 0.14 [-0.36, 0.65]
3 1 -27.02 2.21 -0.8 0.32 0.4 Null/Anecdotal
Intercept 0.22 Null/Moderate 0 [-0.39, 0.36]
BACS 0.42 Null/Anecdotal -0.19 [-0.76, 0.4]
MSCEIT 2.04 Alternative/Anecdotal 0.52 [0.09, 0.94]*
Drift_Rate_Deviated 0.39 Null/Anecdotal 0.15 [-0.46, 0.74]
4 1 -27.29 1.97 -1.08 0.39 0.26 Null/Moderate
Intercept 0.23 Null/Moderate 0 [-0.38, 0.4]
BACS 0.27 Null/Moderate -0.08 [-0.52, 0.35]
MSCEIT 1.63 Alternative/Anecdotal 0.48 [0.05, 0.88]*
Drift_Bias_Forward 0.28 Null/Moderate -0.02 [-0.46, 0.44]
5 1 -25.81 2.15 0.41 0.83 0.85 Null/Anecdotal
Intercept 0.22 Null/Moderate 0 [-0.37, 0.34]
BACS 0.27 Null/Moderate -0.14 [-0.51, 0.24]
MSCEIT 0.69 Null/Anecdotal 0.34 [-0.06, 0.73]
Drift_Bias_Deviated 0.85 Null/Anecdotal 0.37 [-0.03, 0.76]
6 1 -27 1.92 -0.78 0.29 0.29 Null/Moderate
Intercept 0.23 Null/Moderate 0 [-0.38, 0.38]
BACS 0.25 Null/Moderate -0.06 [-0.48, 0.35]
MSCEIT 1.09 Alternative/Anecdotal 0.44 [-0.01, 0.88]
Start_Point 0.28 Null/Moderate -0.07 [-0.54, 0.39]
7 1 -27.52 2.18 -1.31 0.74 0.27 Null/Moderate
Intercept 0.23 Null/Moderate 0 [-0.4, 0.38]
BACS 0.24 Null/Moderate -0.06 [-0.46, 0.34]
MSCEIT 1.84 Alternative/Anecdotal 0.47 [0.09, 0.87]*
Threshold_Separation 0.29 Null/Moderate -0.11 [-0.52, 0.28]

20.3.3 Traditional Metrics as Predictors (Full Sample)

None of the traditional metrics can predict social functioning above and beyond diagnosis, BACS, and MSCEIT.

Bayesian Regression: Predicting Social Functioning from Traditional Metrics (Full Sample)
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -59.16 3.94 – – – –/–
Intercept 0.13 Null/Moderate 0 [-0.21, 0.23]
SZ_Dummy 12.99 Alternative/Strong -0.5 [-0.77, -0.22]*
BACS 0.19 Null/Moderate 0.1 [-0.17, 0.37]
MSCEIT 0.16 Null/Moderate -0.01 [-0.27, 0.25]
2 1 -59.87 3.86 -0.72 0.38 0.19 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 13.41 Alternative/Strong -0.5 [-0.77, -0.21]*
BACS 0.27 Null/Moderate 0.15 [-0.17, 0.48]
MSCEIT 0.16 Null/Moderate -0.02 [-0.3, 0.25]
Accuracy 0.18 Null/Moderate -0.09 [-0.37, 0.18]
3 1 -60.33 4.47 -1.18 1.4 0.18 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 12.68 Alternative/Strong -0.53 [-0.82, -0.25]*
BACS 0.19 Null/Moderate 0.09 [-0.18, 0.37]
MSCEIT 0.16 Null/Moderate 0 [-0.27, 0.26]
RT 0.17 Null/Moderate 0.09 [-0.13, 0.33]
4 1 -59.58 4.34 -0.42 1.03 0.24 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.22, 0.21]
SZ_Dummy 15.47 Alternative/Strong -0.51 [-0.78, -0.23]*
BACS 0.18 Null/Moderate 0.06 [-0.21, 0.34]
MSCEIT 0.16 Null/Moderate 0.01 [-0.25, 0.28]
Criterion 0.23 Null/Moderate -0.14 [-0.36, 0.08]
5 1 -59.65 4.2 -0.49 0.8 0.28 Null/Moderate
Intercept 0.13 Null/Moderate 0 [-0.22, 0.22]
SZ_Dummy 11.57 Alternative/Strong -0.51 [-0.79, -0.24]*
BACS 0.32 Null/Moderate 0.2 [-0.12, 0.53]
MSCEIT 0.17 Null/Moderate -0.01 [-0.28, 0.26]
Discriminability 0.28 Null/Moderate -0.16 [-0.46, 0.13]

20.3.4 Traditional Metrics as Predictors (SZ Only)

Bayesian Regression: Predicting Social Functioning from Traditional Metrics (Within SZ Only)
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -38.23 3.74 – – – –/–
Intercept 0.17 Null/Moderate 0 [-0.31, 0.31]
BACS 0.72 Null/Anecdotal 0.32 [-0.03, 0.65]
MSCEIT 1.46 Alternative/Anecdotal -0.41 [-0.76, -0.08]*
2 1 -38.82 3.89 -0.59 0.54 0.29 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.31, 0.31]
BACS 1.01 Alternative/Anecdotal 0.39 [0.02, 0.78]*
MSCEIT 1.42 Alternative/Anecdotal -0.42 [-0.78, -0.08]*
Accuracy 0.29 Null/Moderate -0.17 [-0.51, 0.19]
3 1 -39.76 4.08 -1.53 1.04 0.21 Null/Moderate
Intercept 0.2 Null/Moderate 0 [-0.32, 0.31]
BACS 0.66 Null/Anecdotal 0.32 [-0.03, 0.67]
MSCEIT 1.28 Alternative/Anecdotal -0.4 [-0.76, -0.05]*
RT 0.21 Null/Moderate 0.07 [-0.27, 0.38]
4 1 -39.27 3.59 -1.05 0.67 0.23 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.32, 0.32]
BACS 0.76 Null/Anecdotal 0.34 [-0.04, 0.69]
MSCEIT 1.52 Alternative/Anecdotal -0.45 [-0.83, -0.06]*
Criterion 0.23 Null/Moderate 0.08 [-0.27, 0.43]
5 1 -39.31 4.11 -1.08 0.6 0.26 Null/Moderate
Intercept 0.19 Null/Moderate 0 [-0.3, 0.33]
BACS 0.88 Null/Anecdotal 0.37 [-0.01, 0.76]
MSCEIT 1.33 Alternative/Anecdotal -0.4 [-0.75, -0.04]*
Discriminability 0.25 Null/Moderate -0.12 [-0.49, 0.25]

20.3.5 Traditional Metrics as Predictors (HC Only)

Bayesian Regression: Predicting Social Functioning from Traditional Metrics (Within HC Only)
Model Ref. Model LOO ELPD LOO ELPD SE ΔELPD ΔELPD SE Model BF Model BF Strength/ Direction Predictor Pred BF Pred BF Strength/ Direction Pred Mean [90% HDI]
1 1 -26.22 2.11 – – – –/–
Intercept 0.22 Null/Moderate 0 [-0.39, 0.35]
BACS 0.24 Null/Moderate -0.09 [-0.47, 0.3]
MSCEIT 2.19 Alternative/Anecdotal 0.48 [0.1, 0.86]*
2 1 -25.47 2.38 0.75 1.01 1.12 Alternative/Anecdotal
Intercept 0.2 Null/Moderate 0 [-0.35, 0.36]
BACS 1.02 Alternative/Anecdotal -0.53 [-1.16, 0.11]
MSCEIT 4.32 Alternative/Moderate 0.55 [0.16, 0.91]*
Accuracy 1.12 Alternative/Anecdotal 0.55 [-0.13, 1.16]
3 1 -27.53 2.08 -1.31 0.62 0.26 Null/Moderate
Intercept 0.22 Null/Moderate 0 [-0.39, 0.38]
BACS 0.27 Null/Moderate -0.11 [-0.51, 0.3]
MSCEIT 1.76 Alternative/Anecdotal 0.47 [0.06, 0.85]*
RT 0.27 Null/Moderate 0.09 [-0.31, 0.51]
4 1 -26.57 2.1 -0.36 0.72 0.44 Null/Anecdotal
Intercept 0.22 Null/Moderate 0 [-0.37, 0.36]
BACS 0.33 Null/Anecdotal -0.19 [-0.61, 0.23]
MSCEIT 1.01 Alternative/Anecdotal 0.4 [-0.01, 0.79]
Criterion 0.43 Null/Anecdotal -0.26 [-0.7, 0.16]
5 1 -27.23 2.23 -1.01 0.3 0.41 Null/Anecdotal
Intercept 0.22 Null/Moderate 0 [-0.39, 0.37]
BACS 0.42 Null/Anecdotal -0.06 [-0.76, 0.63]
MSCEIT 1.78 Alternative/Anecdotal 0.47 [0.06, 0.86]*
Discriminability 0.42 Null/Anecdotal -0.03 [-0.7, 0.68]

References

Bürkner PC. brms: An R Package for Bayesian Multilevel Models Using Stan. J Stat Softw. 2017;80:1-28. doi:10.18637/jss.v080.i01

Keefe R, Harvey P, Goldberg T, et al. Norms and standardization of the Brief Assessment of Cognition in Schizophrenia (BACS). Schizophr Res. 2008;102(1-3):108-115. doi:10.1016/j.schres.2008.03.024

Mayer J, Salovey P, Caruso D. Mayer-Salovey-Caruso Emotional Intelligence Test (MSCEIT) User’s Manual. MHS Publishers; 2002.

Lee MD, Wagenmakers E-J. Bayesian cognitive modeling: a practical course. Cambridge: Cambridge University Press; 2014.

Morey R, Rouder J (2022). BayesFactor: Computation of Bayes Factors for Common Designs. R package version 0.9.12-4.4, https://CRAN.R-project.org/package=BayesFactor.

Peralta, V., & Cuesta, M. J. (1999). Dimensional structure of psychotic symptoms: an item-level analysis of SAPS and SANS symptoms in psychotic disorders. Schizophrenia research, 38(1), 13-26.

Sayers SL, Curran PJ, Mueser KT. Factor structure and construct validity of the Scale for the Assessment of Negative Symptoms. Psychol Assess. 1996;8(3):269-280. doi:10.1037/1040-3590.8.3.269

Vehtari, A., Gelman, A., and Gabry, J. (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing. 27(5), 1413--1432. doi:10.1007/s11222-016-9696-4.

Vehtari, A., Simpson, D., Gelman, A., Yao, Y., and Gabry, J. (2022). Pareto smoothed importance sampling. preprint arXiv:1507.02646

Wabersich, D., & Vandekerckhove, J. (2014). The RWiener Package: an R Package Providing Distribution Functions for the Wiener Diffusion Model. R Journal, 6(1).

Wagenmakers, Eric-Jan, Tom Lodewyckx, Himanshu Kuriyal, and Raoul Grasman. 2010. “Bayesian Hypothesis Testing for Psychologists: A Tutorial on the Savage–Dickey Method.” Cognitive Psychology 60 (3): 158–89. https://doi.org/10.1016/j.cogpsych.2009.12.001.