Causal Inference in Repeated Observational Studies

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): A major goal of many empirical studies in the health sciences is to evaluate the effect of treatments or policy changes. Frequently, random allocation of participants to treatments is not feasible due to practical and ethical reasons. Therefore, participants who choose a treatment may differ from those who choose the control condition. Lack of adequate controls for treated participants often leads to biased treatment effect estimation. Our proposed research is motivated by a repeated cross-sectional observational study on smoking cessation. The smoking cessation program has enrolled smokers every year since 2001 and participants voluntarily choose one of the two intervention arms. In January 2005, an indoor smoking ban was enacted in Italy, so the post-ban intervention effect is likely to be intertwined with the ban effect. Separating the effect due to this policy change from the intervention effect is of great interest to the scientific community. Several challenges are present in the analysis: 1) the program is repeated over time, thus participants are not only incomparable between different treatment arms, but also incomparable before and after the smoking ban. The analytical approach must take the time domain into consideration. 2) The unmeasured confounding is even a bigger issue in repeated observational studies, since it may influence participants' selection differently at different time points. 3) Some important outcomes, such as consumed cigarettes per day (CPD), have highly right-skewed distribution with a non-trivial portion of zeros. Thus standard regression approaches are not applicable and a distribution-free inference is desirable. Propensity score methodology is a popular approach to estimating a causal effect in observational studies. For cross-sectional data, matching or stratification based on propensity score can be used to balance the covariates distribution (Rosenbaum and Rubin, 1983). In longitudinal data, regression analysis incorporating propensity score weights is used to remove time-varying confounding provided all relevant confounders have been observed (Robins, et al. 2000). However, for repeated cross-sectional observational studies, little work has been published to address causal relationship. This project is an attempt to fill this gap by identifying assumptions for causal inference in repeated cross-sectional observational studies and establishing a new propensity score matching methodology to facilitate the estimation. The proposed propensity score matching estimators will be unbiased, distribution-free, and adapt to unknown time effects. Specifically, we plan to achieve two goals in this project: 1) Establishing a generalized potential outcome framework and extending the standard propensity score matching method to develop a difference-in-difference type of estimator for estimating the smoking cessation intervention effect, the policy change effect and their potential interaction. 2) Assessing the potential impact of unmeasured time-dependent covariates on the treatment effect estimate over time.