CRII: RI: RUI: Robust Causal Inference Under Model Misspecification And Uncertainty

Williams College, Williamstown MA

Investigators

Abstract

Causal reasoning is central to human decision making. Enacting new economic policy, modifying high school syllabi, and approving new over-the-counter medications; these are just a few examples of human decision making that rely on rigorously assessing whether a new course of action is not just correlated with improved outcomes, but, in fact, causes them. For many economic, social, and scientific questions of interest however, running a controlled experiment to assess causality is typically infeasible or unethical. This project aims to develop new algorithms and software tools to enable data scientists to draw robust causal conclusions even when faced with messy observational sources of data, such as electronic health records, as well as significant uncertainty about the best technique to use to estimate the causal effect. Specifically, the project aims to develop statistical and machine learning methods that give data scientists the ability to put forward multiple different causal hypotheses, some of which may be incorrect, and automatically distill information from the subset of hypotheses that reflect the true causal relationships between variables while discarding information from all the rest. Model misspecification poses a fundamental challenge to reliable estimation of causal effects from observational data. In the context of causal inference, model misspecification can be divided into two broad categories: (i) Incorrect specification of a causal model that encodes substantive assumptions used to identify the target parameter, and (ii) Incorrect specification of statistical models, such as outcome regression or propensity score models, used to estimate the identified parameter from finite data. This project aims to build causal inference methods that display both forms of robustness. Concretely, the project aims to build algorithms and tools for data scientists that allow them to (i) Propose multiple causal models and combine effect estimates obtained by applying machine learning models to the identifying functionals of each model such that inferences made based on the aggregated estimate are valid if at least one of the proposed causal models is correct (without requiring any knowledge of which ones may be correct); and (ii) Perform model selection between partially specified candidate causal models such that the model selection procedure is robust to issues of unmeasured confounding and non-ignorable messiness in the data. Building such robust causal inference pipelines and software tools can greatly improve the reliability of decision making processes in a variety of fields even when a controlled experiment cannot be conducted in order to assess causality. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.