Explorations of the Normal Neural Behavioral and Pathological Bases of Metacognition

SAMPLING PLAN

1. Existing data Registration before the creation of data: As of the date of submission of this research plan for preregistration, the data have not yet been collected, created, or realized.
2. Data collection procedures. Healthy volunteers will be recruited from the general population. Individuals with schizophrenia will be recruited from community mental health centers and outpatient clinics in the Versailles area and among the FACE-SZ (FondaMental Academic Centers of Expertise for Schizophrenia) cohort in Versailles. All participants will be naive to the purpose of the study, give informed consent in accordance with institutional guidelines and the Declaration of Helsinki, and receive a monetary compensation (10€ / h).
3. Sample size Maximum of 50 healthy controls vs. 50 individuals with schizophrenia.
4. Sample size rationale The estimated sample sizes allow testing effects of medium size between individuals with schizophrenia and healthy controls with a power of 0.8, based on one-sided two-sample t-test power calculation with Cohen's d = 0.5, α = 0.05. They allow measuring medium correlations within groups with a power of 0.7, based on approximate correlation-power calculation with r = 0.3, α = 0.05. Sample sizes for electrophysiological recordings are based on previous a study, with 20 patients vs. 20 controls, resulting in 13 vs. 13 after outlier exclusion.
5. Stopping rule Optional stopping will be avoided by using sequential Bayes factor analyses. Data collection will stop whenever a critical comparison reaches the threshold of BF = 3 or BF = 1/3. DESIGN PLAN
6. Study design The investigators will ask participants to discriminate the motion direction of a random dot kinetogram (type 1 task). They will use a mouse to indicate whether the dots were mostly moving rightward or leftward, by clicking on the side they think corresponds to a correct answer (red and blue circles, see Figure 1). The mouse trajectory corresponding to the type 1 task will be recorded and analyzed. Motion variance will be adapted for each subject before the experiment using a 1up/2down staircase, so to reach an average performance of 71%. An auditory feedback will be played if participants answer in more than 6s. On each trial, participants will then indicate on a visual analog scale the confidence in their response (type 2 task). The scale will range from 0% ("Certain my response is right") to 100% ("Certain my response is wrong"). The initial position of the cursor will always correspond to 50% confidence ("Uncertain of my response)". The experiment will consist in 10 blocks of 30 trials and last about 1h.
7. Randomization Motion direction (left or right) will be pseudo-randomized, with no more than 4 successive trials with the same direction. ANALYSIS PLAN
8. Statistical models 8.1. Behavioral data All analyses will be performed with R, using notably the afex, BayesFactor, ggplot2, lme4, lmerTest, and effects packages. In all ANOVAs, degrees of freedom will be corrected using the Greenhouse-Geisser method. The groups' socio-demographic (age, sex, education), cognitive (premorbid and current IQ, and executive performance with planning and working memory) and mood (depression) characteristics will be compared using the Student t test or Χ² tests when appropriate. Only variables that significantly differ between the two groups will be included as covariates in the following analyses. The metacognitive performance will be primarily analyzed with binomial mixed-effects models between accuracy and confidence, with group (patient vs. control) and several covariates (premorbid and current IQ, depression and executive performance with planning and working memory) as between-subject factors. Regression slope will be taken as an indicator of metacognitive performance and asymptotes as a marker of confidence bias, i.e. the tendency to report high or low confidence ratings independent of task performance. Likelihood ratio tests will assess significance. Predecisional behavioral variables (reaction times, mouse trajectory parameters) will be added to the model in a secondary analysis after main differences between patients and controls are established. Geometric features of mouse trajectories (motion entropy on the x-axis) will be quantified using the EMOT and Mousetrap packages. Correlations between motion entropy and confidence will be quantified by R², adjusted for the number of dependent variables relative to the number of data points. 8.2. Correlation between metacognitive performance and clinical characteristics in schizophrenia The investigators will run correlation analyses between metacognitive performance (regression slope between metacognitive judgments and accuracy of the first order task) and several clinical variables. The clinical variables will be:

• The positive and disorganization scores for the Positive and Negative Syndrome Scale (PANSS), according to the 5-factor model proposed by van der Gaag et al.
• The total score on the Birchwood Insight Scale (BIS) for insight into illness
• The total score on the Beck Cognitive Insight Scale (BCIS) for cognitive insight
• The total score on the Personal and Social Performance Scale (PSP) for social functioning The investigators will use a Spearman rank-order correlation test with false discovery rate to correct for multiple comparisons. 8.3. Electrophysiological data Preprocessing: continuous EEG will be acquired at 1200 Hz with a 64-channels Gtec HIamp system. Signal preprocessing will be performed using custom Matlab (Mathworks) scripts using functions from the EEGLAB toolbox. Following visual inspection, artifact-contaminated electrodes will be removed for each participant, and epoching will be performed at type 1 response onset. For each epoch, the signal from each electrode will be centered to zero and average-referenced. Following visual inspection and rejection of epochs containing artifactual signals, an independent component analysis will be applied to individual data sets, followed by a semi-automatic detection of artifactual components based on measures of autocorrelation, focal channel topography, and generic discontinuity. After artifacts rejection, artifact-contaminated electrodes will be interpolated using spherical splines. Statistical analysis: voltage amplitude will be averaged within temporal windows (e.g., 20ms), and analyzed with linear mixed effects models using R together with the lme4 and lmerTest packages. This method allows analyzing single trial data, with no averaging across condition or participants, and no discretization of confidence ratings. Models will be performed on each latency and electrode for individual trials, including raw confidence rating and accuracy as fixed effects, and random intercepts for subjects. Statistical significance for electrophysiological data within regions of interest (e.g., frontocentral and left parietal scalp regions) will be assessed after correction for false discovery rate. When possible, cluster-based permutation test will be used.

1. Transformations Data will be transformed in case they violate the assumption of normality (e.g., inverse reaction times).
2. Follow-up analyses Besides mixed logistic regressions, metacognitive performance will be analyzed using second-order signal detection theory: meta-d' will reflect the amount of perceptual evidence available when performing confidence judgments. Confidence biases will also be computed with receiver operating characteristic curves (ROC): the area between the ROC and major diagonal will be divided by the minor diagonal, and confidence bias will be defined as the log ratio of the lower and upper area. An ANOVA with group and appropriate covariates as between-subject factors will test for a decrease in metacognitive efficiency and an increase in confidence bias in patient vs. control participants. Drift-diffusion modeling will allow us to determine which aspects of reaction times during the type 1 task differ between schizophrenic patients and healthy controls (e.g., drift rate and boundary separation), and assess how such differences might determine confidence judgments, thereby allowing testing the existence of metacognitive deficits at a decisional-locus.
3. Inference criteria Two-tailed tests with group as the between-subject factor will be used. The threshold for significance will be set to alpha = 5%. When possible, Bayes factors will be computed to support null findings and set stopping rules (see above).
4. Data exclusion The first trials of each condition will be excluded from analysis if they contain large variations of the perceptual signal. Only trials with reaction times between 100 ms and 6 s for the type 1 task will be kept. Participants will be excluded in case they cannot reach 71% accuracy on the type 1 task, respond in more than 6 s in a majority of trials, or in case they do not use the confidence scale properly (e.g., no variance in confidence reports).
5. Missing data The use of mixed models applied to behavioral and electrophysiological data will allow dealing with unbalanced datasets so that data imputation will not be needed.
6. Exploratory analysis (optional) 14.1. Correlation between metacognitive bias and clinical characteristics in schizophrenia The investigators will run exploratory Spearman rank-order correlation analyses between metacognitive bias (asymptotes of the regression line between metacognitive judgments and accuracy of the first order task) and several clinical variables (positive and disorganization scores for the PANSS, the total score for BIS, BCIS, and PSP).

14.2. Heart rate Heart rate will be measured with a Gtec plethysmographic pulse sensor and quantified as a function of type 2 performance. Based on previous findings in healthy participants, the investigators expect greater confidence to be associated with faster heart rate between stimulus onset and type 2 response. The investigators will attempt to replicate these findings following the same methods as Allen and colleagues and extend it to patients.

14.3. galvanic skin response (GSR) As for heart rate, GSR will be measured with a Gtec dedicated sensor and quantified as a function of type 2 performance using the Ledalab toolbox under Matlab. To our knowledge, no study has quantified the link between GSR and metacognition so that the investigators will conduct exploratory analyses.