Inferring causal relationships from observational data
Most scientists know the maxim that association is not causation, but almost all users of statistical models and hypothesis tests are after causal relationships nevertheless. No statistical tool can inform our understanding of cause and effect without strong assumptions, some untestable, which are rarely articulated and often misunderstood. This talk will focus on statistical methods for distinguishing causal relationships from other kinds of associations, with examples from the health sciences. I will emphasize nonparametric identification, rather than estimation, of causal effects-that is, which causal effects we could learn about if we had infinite data and could thereby sidestep statistical models, which are required to make use of finite data in the real world.
Beyond the deceptively simple but already quite thorny question does X cause Y are a panoply of more complicated causal questions for which statistical methods have recently been developed. What is the effect of a longitudinal treatment on a longitudinal outcome? How can we study the causal effect of a treatment that affects not only the targeted individual but also that person's friends, family members, or contacts? Does the causal effect of X on Y operate through a mediating variable, or does X affect Y directly? How can a causal effect estimated in one population be generalized to a different population? I will discuss the conditions under which these effects can (and cannot) be identified and mention some open questions and new frontiers in causal inference methodology.