ABSTRACT
In this chapter, we review methods for drawing causal inference from observational data. Of course, while this is an important goal in epidemiology in general and pharmacoepidemiology in particular, in practice, we rarely achieve it. Rather, we attempt to insulate our findings from bias to the greatest extent possible by conducting analyses with assumptions which when met support causal inferences. It is often difficult if not impossible to know with certainty that such assumptions are ever met. As an example, methods such as propensity score matching and marginal structural models eliminate static and dynamic confounding respectively under the assumption that all relevant confounders have been measured. Of course, this is an untestable assumption. However, we can determine the magnitude of the effect of an unmeasured confounder that would be required to change the conclusion of the analysis and in turn the plausibility of the existence of such an unmeasured confounder. In pharmacoepidemiology, one of the most important unmeasured confounders is severity of illness. We measure concomitant medications, comorbid diagnoses, hospitalizations, and prior experiences of the adverse effect before the drug exposure, but in general, we have no direct measure of the severity of illness of the patient at baseline or following the initial exposure. This limits our ability to draw causal inferences from these observational studies which form the basis for most work in pharmacoepidemiology. In the following sections, we review different design and analytic methods which minimize bias and take us down the path of causal inference.
