A Computer-Assisted CBT Tool to Enhance Fidelity in CBOCs

Power analysis for primary outcome We aimed to recruit a total sample of 34 (17 per condition) providers to achieve a statistical power of 0.94 for analysis comparing our primary outcome of CBT fidelity between the two conditions. We used a general linear mixed model to account for patients clustered within the same providers and a type I error rate of .05. This initial sample size was also determined by assuming an effect size of 1.0 (1.3 point difference on a scale of CBT fidelity; 5.3 of 6 for the computer condition versus 4 of 6 for manual condition), a medium intraclass correlation of 0.5, and four patients per provider.

A total of 16 of 32 clinicians self-selected to provide audiotaped sessions that were assessed for fidelity. Assuming an effect size of 1.0 for condition on the fidelity outcome, an intraclass correlation of 0.5, four participants per provider, and a type I error rate of .05, our statistical power for comparing our primary outcome of fidelity between conditions using general linear mixed model is 0.65.

Statistical analysis For the primary outcome of providers' CBT fidelity, descriptive statistics were calculated for the entire sample and by session. The association between condition and the outcome of fidelity was examined using a general linear mixed model to account for Veteran participants clustered within providers. The treatment session and strata variables were also included in the model.

For the patient-level secondary outcomes, bivariate analysis was performed using generalized estimating equations due to the clustered structure of Veterans within the same providers. Associations between condition and potential covariates and between outcomes and potential covariates were examined. Covariates with p-values less than 0.10 were included in the multivariate models for associations between condition and outcomes over time. Generalized linear mixed models were used to account for the correlations for patients within providers as well as the correlations of multiple assessments within patients. Gamma distribution was specified for BSI-18 GSI scores after a small rescale for zero value due to its violation of normality and normal distribution was specified for the remaining outcomes as they were approximately normally distributed. All the models included the condition indicator variable, time (for the three interviews), strata, and covariates identified in the bivariate analysis. The covariates associated with condition (gender and primary diagnosis) were included in all the models with the exception of primary diagnosis not being included in the subgroup specific diagnosis group analysis. The covariates associated with the outcomes were also included in the corresponding outcome models. The interaction between condition and time was included in all of the models as hypothesized. General linear mixed models were also fit for disorder specific outcomes for subgroups of Veterans with the corresponding specific disorders as they were approximately normally distributed.

The LS mean differences (or ratios depending on the outcomes) between the two conditions and their corresponding 95% confidence intervals were calculated for evaluating the effect of condition. Similar differences (or ratios) between each follow-up and baseline by each condition and their corresponding 97.5% confidence intervals were also calculated for evaluating the effect of time. A narrower confidence interval (equivalent to a p-value of 0.025) was used to adjust for multiple comparisons. All the analyses were performed using SAS 9.4.