Estimating the causal effect of dynamic treatment regimes in observational studies. Miguel Hernan, MD, Instructor in Epidemiology, Harvard School of Public Health. Joint work with James Robins and Stephen Cole Wednesday November 13 at noon Center for Basic Research in the Social Sciences 34 Kirkland Street, Room 22 Lunch will be served. Abstract: Large randomized experiments with full adherence and no losses to follow-up can be used to make causal inferences without further assumptions. These ideal experiments may be used to evaluate the effect of treatment regimes that a) fully describe each subject's treatment experience during the study even before the study starts (e.g., always treat, never treat, treat every other month), or b) do not fully describe each subject's treatment experience during the study before the study starts because the treatment actually received depends on the subject's time-varying characteristics (e.g., treat until side effects appear, treat only when some blood parameter reaches the value 200). We call the latter dynamic treatment regimes. The analysis of ideal experiments is straightforward: simply compare the mean outcome (or survival) in subjects under each (dynamic or non-dynamic) treatment regime. Observational studies can be used to make causal inferences only under the assumption that they are able to provide the same results as a (possibly hypothetical) randomized experiment. This assumption, also known as "the assumption of no unmeasured confounders" or "of sequential randomization", is not testable. But even if this assumption holds, and in the absence of model misspecification, the standard analysis of observational studies with time-varying treatments may yield estimates that cannot be endowed with a causal interpretation. This limitation of standard methods applies to the estimation of the effects of both dynamic and non-dynamic treatment regimes. In contrast, methods based on marginal structural models and structural nested models would yield valid causal estimates. We have previously applied a marginal structural Cox model to estimate the effects of non-dynamic treatments in HIV/AIDS epidemiology. The subject-matter question we addressed was "should HIV-infected individuals be treated with highly active antiretroviral therapy?" In this talk, I will review our previous results and describe the application of a structural nested accelerated failure time model for estimating the effect of dynamic treatments among HIV-infected individuals. The question is "should treatment with highly active antiretroviral therapy start when CD4 count drops to 200 or to 350?" Also, I will discuss computational issues and different methods to handle censoring. When the paper becomes available, a link will be sent to those requesting it (email corr(a)fas.harvard.edu) and to those on the workshop email list. Those unfamiliar with methods for causal inference from observational data may find it useful to read about estimating non-dynamic treatments as described in http://www.hsph.harvard.edu/causal/publications/joint-causal.pdf The Research Workshop in Applied Statistics is a forum for graduate students, faculty, and visiting scholars to present and discuss statistical innovations and applications in the social sciences. For more information or to subscribe to the permanent workshop list, email corr(a)fas.harvard.edu.