The study of disease is carried out both on the basis of observational data, as in epidemiologic research to identify potential risk factors, and data from controlled clinical trials, where subjects are randomized to receive treatments that are to be compared. In both cases, interest focuses on estimation of causal effects and testing causal hypotheses; e.g. to infer a "cause and effect" relationship between risk factors or treatment interventions and disease. With observational data, problems with making causal inference are well-discussed; for example, the inference that smoking causes lung cancer may be called into question by the possibility of other risk factors, called confounding factors in epidemiology, that are associated with both. In clinical trials, despite the best efforts of investigators, assignment of treatments to participants may be flawed in ways that may affect the validity of causal statements; for example, subjects assigned to treatment A who refuse to take it or surreptitiously take B may be different from subjects assigned to placebo or to B via the randomization, respectively. Such failure of subjects to comply with randomization may compromise the ability to estimate the causal effect of treatment on disease.Although the objective of inferring causation is central to biomedical research and more generally to most areas of scientific inquiry, it is surprising that basic definitions and methods for making causal inference are still limited and open to debate and moreover have not been the focus of widespread discussion among biostatisticians. Statisticians often say that statistics can be used only to make statements of association, not causation, in observational studies; nonetheless, the scientific community is interested in causation. Failure to study this issue carefully only diminishes the role of the statistician. There is thus growing interest in elucidating the extent to which causal inferences may be drawn from observational data. Formalization of this notion has been cast in terms of so-called counterfactual analysis , and methods for making causal inference have been based on construction of causal diagrams or graphs . The goal of these procedures is to determine whether effects of interest are identifiable from the observational process.
This semester, we will introduce the issues associated with and formal study of causal inference in biomedical research.
Articles on this (nonexhaustive) list will form the basis for the material covered in presentations in October and November. If you would like to make copies of these articles, please see Sharon in Patterson 209
Return to Biostatistics Working Group
