Balancing treatment comparisons in longitudinal studies.

Evaluation of treatment efficacy in psychiatric trials involves a comparison of outcomes for those who receive a treatment versus those who receive a control or comparison treatment. However, if the treatment and comparison groups are not comparable or “balanced,” observed differences in outcomes between treated and comparison groups may be due, in part, to these imbalances. In such settings, estimates of treatment efficacy may be biased unless some adjustments are made to make the groups more comparable. The preferred method of treatment assignment is randomization. Randomization ensures, on average, a balance of observed and unobserved baseline characteristics between those assigned to treatment and those assigned to the comparison group. In the absence of randomization, however, treatment groups could differ on the basis of both observed and non-observed characteristics. Longitudinal observational studies, studies that repeatedly measure outcomes on participants, are subject to additional analytic challenges. However, a) treatment groups may differ at baseline,1 and b) treatment groups may quickly become less comparable over the course of the study due to subject dropout, treatment switching, noncompliance, and missing data.2 Thus, estimates of treatment efficacy in longitudinal studies may result in over- or underestimates, unless comparisons can be balanced. In this article, we show how techniques used in observational cross-sectional studies can be used to balance comparisons in longitudinal studies. We build upon the methods presented in Stuart et al (to be published in an upcoming issue of Psychiatric Annals), which will describe methods for balancing comparisons in cross-sectional trials.3 We discuss when to adjust during the course of the study (eg, baseline, posttreatment, follow-up assessments, etc.). We also present methods for creating propensity scores (scalar summaries that measure how likely a subject is to receive the treatment rather than the control) for longitudinal studies such that the comparisons are balanced and inference is not compromised by adjusting for variables that have been affected by treatment.4 Throughout this article, we assume that the treatment is assigned once and it is not time-varying. Three longitudinal studies in which propensity scores are used to balance treatment comparisons illustrate our methods (see Table 1, page 807). These studies include the Runaway Youth Study, an intervention study aimed at preventing HIV transmission among runaway youths housed at shelters in New York City;1 the Gang-joining Study, a Montreal-based study that evaluated the effect of gang joining at age 14 on subsequent violence; and the MTA Follow-up Study, an observational follow-up to the randomized Multimodal Treatment Study of Children with Attention Deficit Hyperactivity Disorder (MTA).2 Each study has a different design and requires balancing at different points along the longitudinal course. PMC ID: PMC2722117 (December 2008)