Causal Inference for Treatment Effect using Observational Healthcare Data with Unequal Sampling Weights



 DESCRIPTION (provided by applicant): A common goal in health outcome and policy research, as well as in other disciplines, is to evaluate the possible causal effect of an intervention, which is broadly defined as a policy change, program participation, or a medical treatment. In health outcome research, many datasets are observational since randomization is often not feasible or ethical. Also, many of these datasets are obtained from large probability surveys. This presents major challenges in inferring causal relationships due to the following facts: (1) the differences in outcomes for intervention groups could be due to the differences in covariates prior to the intervention; and (2) they always involve unequal weighting due to complex sampling designs. Propensity score-based adjustments are widely used to reduce the confounding bias of covariates in observational studies. But there is a critical methodological gap with regard to appropriately incorporating complex survey design in propensity score analysis, and there is a significant amount of confusion among researchers regarding the best practice for interpreting causal effect when using survey data. The overarching goal of this project is to develop a systematic and statistically valid approach for employing causal inference techniques with complex healthcare survey data. Because of the use of sampling weights, the proposed methods are particularly useful in the presence of heterogeneous treatment effects, i.e. different groups may respond differently to a policy change or the introduction of a certain medical treatment. The proposed study will achieve four specific aims: (1) Develop a potential-outcome-based theoretical framework to streamline causal inference in complex surveys; (2) Develop both propensity score and survey design adjusted estimators, including weighted, stratified and matched estimators; (3) Conduct extensive simulation studies to evaluate the performance of various estimators under different practical scenarios and develop a statistical software package for practitioners; and (4) Apply the proposed methodology to a real survey for comparative trauma care research. This study is expected to fill a critical gap in healthcare policy and treatment effect evaluation research by extending the commonly used propensity score adjustment for non-survey data to complex sampling designs. Findings of this study will help promote AHRQ's mission to produce more accurate evidence for health care program evaluation and to improve the current practice of comparative health outcome research. A significant contribution is this general purpose methodology which will be widely applicable and can benefit government agencies, policy makers, and social, political and health science researchers, in those situations where survey data are vital sources for comparative outcomes research and program policy evaluation.