The notion of making causal inference is one that statisticians often view as mysterious or even dangerous ("correlation does not imply causation"). Despite the fact that the standard training of statisticians rarely includes formal discussion of this topic, the objective of most scientific inquiry is to infer "causation" or "cause and effect."In this talk, we will review the thinking of philosophers on defining "cause" and "effect." We will then introduce a widely used model for causal inference and show how the notions of "cause" and "effect" may be formalized within its framework. In the context of the model, special circumstances may be considered that lend insight into whether causal questions may be addressed and what assumptions might be necessary for causal inference to be made. In particular, this formulation makes explicit why the (properly executed) randomized experiment allows causal effects to be deduced and why an observational study may not. One key idea is what Rubin has termed the "stable-unit-treatment-value assumption" (SUTVA), whose definition and role in causal inference we will discuss.
The model formalizes a main basis for dispute on making causal interpretations from observational studies (or compromised randomized experiments, such as a clinical trial with noncompliance): the fact that experimental units are not assigned to the "treatments" or "exposures" of interest at random. Thus, there may be other confounding factors that may be associated both with the treatments and the response; consequently, epidemiological methods, which are focused on observational data, are concerned with avoiding or adjusting for such confounding. We will discuss the utility of this strategy by considering extension of the model to include covariates that may be potential confounding factors and examine assumptions under which such covariate adjustment may make causal inference possible. One key assumption is that of "strongly ignorable" treatment assignment. We will discuss this assumption as well as a particular strategy for covariate adjustment based on the so-called "propensity score."
This talk will provide background for two subsequent talks (Oct. 21, Nov. 11) covering statistical methodology for making causal inference.
Return to Biostatistics Working Group