Causal estimation of time-varying treatment effects in observational studies: application to depressive disorder

Clinicians recognize three phases of the treatment of major depression: an acute phase to control disabling symptoms, a continuation phase to avoid relapses of a single episode, and a preventive phase to avoid recurrences of new episodes over time. With no directly measurable trace of the underlying pathological process, the distinction is based arbitrarily on the passage of time in remission. The clinician who has successfully treated a patient with antidepressant medications in the acute phase has a critical clinical decision to make for the continuation and preventive phases: whether to continue to prescribe the medication, for how long, and at what dose. This decision, like most clinical decisions in psychiatry, is not yet completely determined by the results of randomized clinical trials. Only a handful of such trials have been completed, covering just a fraction of the possible maintenance strategies (defined by treatment drop times). For many reasons, observational studies of the outcome of naturally occurring treatment choices play an important supporting role, helping to extend the reach of completed studies and to design new studies. Causal inference from observational studies has usually been considered in the context of a decision among a few fixed alternatives at a single time. The particular causal effect of interest in the maintenance of remission dictates that treatment be studied over remission time. This challenges the causal analysis of the observational study. We present issues arising from assessing temporal treatment effects due to nonrandomized treatment assignment over time. We use data from a large observational study of the course of affective illness, to illustrate an approach to this problem.