Model Fitting

Preprocessed response time and accuracy data were analyzed using hierarchical Bayesian parameter estimation in the Drift Diffusion Model [HDDM (9)]. We particularly focus on two DDM parameters – decision threshold, a, and the drift rate, v, (Table 1, Figure 1) – and on how these parameters respond to task demands (perceptual vs. value-based decisions, easy vs. difficult choices), and how these adjustments are modulated by the OCD diagnosis and gender.

Demographic and clinical characteristics of the study participants.

In bold are significant differences between groups.

To improve the quality of parameter estimation, we employed the basic 4-parameter model (58) and allowed these parameters to vary across trial types (PDM vs. VDM) and choice difficulty (easy vs. difficult). Next, to examine effects of interest, we allowed three of the parameters (the decision threshold, the drift rate, and the non-response time) to depend on the subject's diagnosis (Dx: OCD or HC) and gender. Finally, we included covariates that have been shown to affect the decision threshold and the drift rate in prior studies and that potentially could confound our estimates of effects of OCD diagnosis and gender. This approach produces the following models:

Model 0: a ~ trial type, difficulty; v ~ trial type, difficulty; τ ~ trial type, difficulty; z;

Model 1: a ~ trial type, difficulty, Dx; v ~ trial type, difficulty, Dx; τ ~ trial type, difficulty, Dx; z;

Model 2: a ~ trial type, difficulty, gender; v ~ trial type, difficulty, gender; τ ~ trial type, difficulty, gender; z;

Model 3: a ~ trial type, difficulty, Dx, gender; v ~ trial type, difficulty, Dx, gender; τ ~ trial type, difficulty, Dx, gender; z;

Next, in Models 4–6, we included age and IQ as covariates, since they have been previously to affect the decision threshold and the drift rate and excluding them could potentially confound our results (12, 16). We did not include age, IQ, and other variables as covariate for τ in Models 4–8 since we did not have a priori hypothesis and to avoid overfitting the model.

Model 4: a ~ trial type, choice difficulty, Dx, gender, age; v ~ trial type, choice difficulty, Dx, gender, age; τ ~ trial type, choice difficulty, Dx, gender; z;

Model 5: a ~ trial type, choice difficulty, Dx, gender, IQ; v ~ trial type, choice difficulty, Dx, gender, IQ; τ ~ trial type, choice difficulty, Dx, gender; z;

Model 6: a ~ trial type, choice difficulty, Dx, gender, age, IQ; v ~ trial type, choice difficulty, Dx, gender, age, IQ; τ ~ trial type, choice difficulty, Dx, gender; z.

In Model 7, we examined whether including severity of depression as covariate changes our estimates of effects of OCD diagnosis and gender on the decision threshold and the drift rate. Several prior studies reported that depression may affect a process of evidence accumulation, specifically, by making the decision thresholds wider (59, 60). Since individuals with OCD tend to report more of depressive symptoms than healthy individuals, not including severity of depression may potentially confound estimates of the effect of OCD diagnosis.

Model 7: a ~ trial type, choice difficulty, Dx, gender, age, IQ, BDI; v ~ trial type, choice difficulty, Dx, gender, age, IQ, BDI; τ ~ trial type, choice difficulty, Dx, gender; z.

Finally, in Model 8, we examined whether including self-reported impulsivity [measured by Barat Impulsivity Scale, BIS-11 (61)], as covariate changes our estimates of effects of OCD diagnosis and gender on the decision threshold and the drift rate. Note that impulsivity is a complex, multifaceted concept, and BIS-11 and decision threshold are likely to quantify different components of impulsivity (62). Still, not including a measure of impulsivity may confound estimates of the effect of OCD diagnosis on DDM parameters.

Model 8: a ~ trial type, choice difficulty, Dx, gender, age, IQ, BIS; v ~ trial type, choice difficulty, Dx, gender; τ ~ trial type, choice difficulty, Dx, gender; z.

Selection of the final model was based on deviance information criteria [DIC (63)] and on the comparison of posterior predictive probability density plots with the data-based normalized RT distribution for each condition.