A Counterfactual Approach to Sensitivity Analysis for Unmeasured Confounding and Selection Bias

• Activity Number: 373
• Type: Invited
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Miguel A. Hernán*+
• Companies: Harvard School of Public Health
• Keywords: causal inference ; confounding ; sensitivity analysis

Abstract

Effect estimates from observational studies may be biased because of unmeasured confounding and selection bias. Neither the existence of these biases or their magnitude is identifiable from the joint distribution of the observables, but one can conduct a sensitivity analysis to quantify the variation in the estimates with the magnitude of the nonidentifiable bias. Even when many unmeasured confounders are hypothesized to exist, one can still specify a low-dimensional bias function that quantifies the magnitude of the bias by modeling the association of the counterfactual outcome variable with the treatment within levels of the measured confounders. Then, one can examine how the inferences change as the bias function is varied over a plausible range. This counterfactual approach to sensitivity analysis leads to computationally simple methods of estimation provided that the appropriate semiparametric methods are used. In this talk, we will present and discuss epidemiologic applications of this counterfactual approach for the estimation of the causal effects of dynamic and nondynamic treatment regimes.