Model and variable selection for causal inference

点击翻译按钮获取中文摘要

To unbiasedly estimate a causal effect utilizing observational data, one must adjust for potentially confounding covariates. Identifying confounding variables using only substantive prior knowledge can be a very difficult task, especially in new areas of research or those where current subject-matter knowledge is sparse. An intuitive strategy to circumvent this problem is to adjust for all potential confounders in a statistical model. This strategy can, however, produce causal exposure effect estimators whose variance is unnecessarily large. Instead, model selection can be attempted. Recent simulation results have shown that classical statistical approaches for model selection often fail to produce unbiased causal effect estimators. Hence, there is a need for the creation of new model selection approaches specifically tailored for causal inference purposes. I have recently proposed a novel model selection approach, the Bayesian causal effect estimation (BCEE) algorithm, for the estimation of the average causal effect of an exposure variable on a continuous outcome. The results from the simulation study I conducted were very promising since BCEE generally achieved at least some reduction of the mean squared error (MSE) of the causal exposure effect estimator as compared to estimators produced by either a fully adjusted model or by alternative model selection approaches. Notable MSE reduction was observed in some scenarios. Moreover, BCEE produced well calibrated confidence intervals. To facilitate the utilization of BCEE in practice, I also produced an R package that is freely available on the Internet. The main objective of my research program is to develop novel, easy to use, model selection approaches in a causal inference framework. My first focus will be to provide extensions of BCEE that would allow for binary, count and survival outcomes. I will also develop approaches that target the causal effect of the exposure in specific subpopulations. Afterward, I will produce methods that are robust to model misspecification (e.g., so-called double robustness property). Many other extensions are possible in the long term, such as developing methods for the estimation of the effect of an exposure regime in a longitudinal setup. Since model selection in a causal inference framework is a relatively novel area of research, much remains to be done. My research seeks to provide part of the answer to this need.