Letters Section Editor: Robert M. Golub, MD, Senior Editor.
To the Editor: The study by Dr Stukel and colleagues1 is an empirical confirmation of a theoretical notion that has been described for decades: therapeutic benefits can almost never be assessed by observational studies.2,3 The exceptions are instances in which an “instrumental variable” can be identified as a pseudo-randomizer3 or when effect sizes are very large. Randomization is necessary because the application of therapy is aimed at intended beneficial effects and is guided by the perceived prognosis of the patient. The complexity of therapeutic decisions leads to incomparability of treated and untreated persons in ways that cannot be remedied by statistical analysis in observational studies because these decisions cannot be summarized in a number of crudely measured variables.2 This is demonstrated convincingly by Stukel et al. Neither analysis for confounding nor propensity score analysis was sufficient; only an instrumental variable analysis approached the results from randomized trials.1
Vandenbroucke JP. Analytic Approaches to Observational Studies With Treatment Selection Bias. JAMA. 2007;297(19):2077-2078. doi:10.1001/jama.297.19.2077-b