Repeated assessments of depressive symptoms in randomized psychosocial intervention trials: best practice for analyzing symptom change over time.

OBJECTIVE

Psychotherapy randomized trials rarely have tested for the best fitting model for time effects. We examined the fit of different statistical models for examining time when repeated assessments of depressive symptoms are the primary outcome.

METHOD

We used data from three studies comparing psychotherapy treatments for major depressive disorder. Outcome measures were self-report ratings for Study 1 (= 237) and Study 2 (= 100) and clinician ratings for Study 3 (= 120) of depressive symptoms measured at every session (Studies 1 and 2) or monthly (Study 3). We examined the fit of the following time patterns: linear, quadratic, cubic, log transformation of time, piece-wise linear, and unstructured.

RESULTS

In Study 1, a log-linear model had the best fit (Δ Akaike information criterion [AICc] = 7.5). In Study 2, all models had essentially no support (Δ AICcs > 10) in comparison to the best fitting model, which was the unstructured model. In Study 3, the cubic model had the best fit, but it was not significantly better than a log-linear (Δ AICc = 3.5) or unstructured model (Δ AICc = 2.5).

CONCLUSIONS

Trials should routinely compare different time models, including an unstructured model, when repeated measures of depressive symptoms are the primary outcome.