Background Acting on harmful command hallucinations is a major clinical concern. Our COMMAND CBT trial approximately halved the rate of harmful compliance (OR = 0.45, 95% CI 0.23-0.88, p = 0.021). The focus of the therapy was a single mechanism, the power dimension of voice appraisal, was also significantly reduced. We hypothesised that voice power differential (between voice and voice hearer) was the mediator of the treatment effect.Methods The trial sample (n = 197) was used. A logistic regression model predicting 18-month compliance was used to identify predictors, and an exploratory principal component analysis (PCA) of baseline variables used as potential predictors (confounders) in their own right. Stata's paramed command used to obtain estimates of the direct, indirect and total effects of treatment.Results Voice omnipotence was the best predictor although the PCA identified a highly predictive cognitive-affective dimension comprising: Voices' power, childhood trauma, depression and self-harm. In the mediation analysis, the indirect effect of treatment was fully explained by its effect on the hypothesised mediator: Voice power differential.Conclusion Voice power and treatment allocation were the best predictors of harmful compliance up to 18 months; post-treatment, voice power differential measured at nine months was the mediator of the effect of treatment on compliance at 18 months.