Intellectual Merit:

Causal inference lies at the core of many scientific endeavors. For example, how effective will a new medication be for a specific group of patients, or how much can a new policy reduce the crime rate? These questions require moving beyond mere associations to address causality—if we intervene in a specific way, what will the outcome be?

Traditionally, statistical developments have relied heavily on randomized experiments. However, in the era of big data, many large datasets are either entirely observational or have a significant observational component. Causal inference for observational datasets is more challenging than for randomized experiments for several reasons: key covariates may be distributed differently between intervention groups; unmeasured confounders might exist and are correlated with both the treatment and the outcome; and treated and control groups may exhibit different functional patterns with respect to the response. These differences contribute to two widely recognized challenges of causal inference — the adjustment for confounders and the modeling of outcomes — which must be addressed simultaneously.  Moreover, the complexities of modern datasets necessitate flexible models to capture heterogeneity. However, both traditional design-based methods and many nonparametric regression-based approaches struggle to incorporate such information.

We addressed causal inference for large observational datasets by developing innovative Bayesian semiparametric methods that account for effect heterogeneity. These methods remove confounding factors by incorporating propensity scores and model potential outcomes using flexible semiparametric Gaussian process regressions.  They can accommodate complex subgroup or hierarchical structures within massive observational data and effectively integrate external information or expert knowledge about causal relationships.  Our causal estimators align with conventional propensity score matching estimators under certain prior specifications, while offering substantially improved efficiency and better identification of heterogeneous effects in general scenarios. Moreover, they provide clearer interpretations and more robust estimates compared to many machine learning estimators.

We also evaluated the impact of unmeasured confounding through Bayesian sensitivity analyses. In the presence of heterogeneity, the treatment effect is determined by a function of covariates rather than a single parameter, requiring more sophisticated evaluation strategies. We simulated and incorporated unmeasured confounders in both the response and treatment models, with coefficients serving as sensitivity parameters. The values of these sensitivity parameters were then compared with the marginal effects of observed covariates, providing researchers with a benchmark for determining problematic levels of confounding.

This project effectively combined the potential outcome framework with semiparametric Bayesian models, making significant contributions in formalizing and solving complex causal inference problems in observational studies.

Broader Impacts:

The project has had a significant impact on human resource development, fostering the academic growth of eleven doctoral students in the Statistics and Biostatistics PhD programs at Ohio State University.  Seven of these students are female, and four have since graduated and transitioned into industry or academic positions. Through their involvement in this project, they gained substantial experience in Bayesian causal inference, laying a strong foundation for their development into independent, analytically proficient statisticians with a comprehensive understanding of the field.

Also, the project has provided valuable opportunities for training and professional development. The knowledge gained from this research was integrated into a graduate-level course on Causal Inference, attended by both master’s and PhD students from diverse disciplines, including Statistics, Biostatistics, and Epidemiology. This course offered students firsthand exposure to cutting-edge research in causal inference.

The models developed through this project have a broad range of applications, spanning sectors such as medicine, education, and business. Through the PIs' proactive engagement with various communities, the project's results have been widely disseminated through presentations, scholarly publications, and collaborative research initiatives. Furthermore, our Bayesian nonparametric causal inference model has been implemented in R, with the computational code freely accessible to the public, promoting openness and collaboration.

Last Modified: 08/31/2024
Modified by: Xinyi Xu