Statistical Analysis

Following the CONSORT (Consolidated Standards of Reporting Trials) and CHEERS (Consolidated Health Economic Evaluation Reporting Standards) guidelines, all our analyses adhered to the intention-to-treat principle. To that end, missing values were imputed using multiply imputed chained equations (MICE) for nonparametric data with M of 100 bootstraps for each incomplete variable. Baseline variables predictive of effects (ie, QALYs and treatment response) were used for imputation, such as baseline data of the variable with missing values, treatment condition, ethnicity, education, sex, age, and safety behaviors at baseline. Safety behaviors, such as avoiding eye contact or escaping from social situations, were measured using the Safety Behaviour Questionnaire-Persecutory Delusions (SBQ-PD) [20]. Time spent with others showed a large difference at baseline despite randomization and was added as covariate in the CEA where time spent with others was used as the treatment response outcome of interest.

Both the CUA and CEA were conducted from the societal perspective. In each of these analyses, the incremental cost-effectiveness ratios (ICERs) were calculated as the between-group cost difference divided by the between-group effect difference. The ICER is interpreted as the additional costs per additional unit effect (ie, per additional treatment responder; per QALY gained). Cost and effect differences were obtained from seemingly unrelated regression equations of costs and effects, thus allowing for correlated residuals in the equations. The seemingly unrelated regression equations (SURE) models were bootstrapped 5000 times. In each bootstrap step, the mean cost differences and the mean outcome differences were computed and these were plotted on the cost-effectiveness plane. Finally, cost-effectiveness acceptability curves (CEACs) were graphed. CEACs inform decision makers about the likelihood that an intervention is deemed cost-effective, given a range of willingness-to-pay ceilings for gaining 1 QALY and gaining 1 treatment responder. All analyses were conducted in Stata, version 13.1 (StataCorp).

The following sensitivity analyses were carried out. First, a sensitivity analysis was done including safety behavior at baseline as a covariate because despite randomization there was a significant difference at baseline, and it was found to be the main mediator in reducing paranoid ideation [10]. Second, a sensitivity analysis was done including psychiatric admission costs at baseline as a covariate because there was a large difference between groups at baseline. Third, a sensitivity analysis was done including both safety behavior at baseline and psychiatric admission costs at baseline as covariates in the model.