In most observational studies and in some randomized experiments, matching or adjusting for selection bias often leads to a more relevant summary of the data, particularly where causal hypotheses are concerned. For example, in evaluating the effects of drug and alcohol exposure in utero using a longitudinal cohort of mothers and infants, it would be typical to match exposed and unexposed infants on several maternal characteristics believed to effect both in utero exposure and the longitudinal outcome of interest, eg brain functioning. But what if one also wished to adjust for differential post-utero maternal environments? In this and many other scenarios, the hypothetical intervention of interest is longitudinal and thus so may be the selection bias. The goal of this talk is to offer and elicit perspectives on various methods in use for summarizing the effect of a longitudinal intervention with nonideal data. Additional examples to be considered are:
the evaluation of psychiatric care for treatment of depression
the etiology of cervical cancer
the 1997 faculty salary study
socio-temporal networks analysis
