CAREER: Bayesian Tree Models for Next-Generation Studies in the Behavioral and Social Sciences

点击翻译按钮获取中文摘要

This research project will develop statistical methods and tools for inferring and understanding the implications of heterogenous effects of interventions in the behavioral and social sciences. Experimental research in the social and behavioral science is facing a crisis and an opportunity. Behavioral interventions that seemed promising based on initial studies have failed to replicate or have disappointed when implemented at scale. Ignoring heterogeneous effects across individuals and by contexts is a likely contributor of these results. This CAREER project will develop Bayesian methods for estimating heterogeneous treatment effects in large, complicated datasets. The methods to be developed will be applied to real-world behavioral interventions. The investigator will collaborate with the Texas Behavioral Science and Policy Institute and the University of Texas's OnRamps program to develop and evaluate interventions designed to promote growth mindset practices and beliefs among high school math teachers and college instructors. In addition, students will be mentored, and case studies and software will be developed and made freely available. This award is supported by the MMS program and the Education and Human Resources directorate's ECR program.This research project will develop Bayesian tree priors, models, and computational methods for complex study designs. Estimating heterogeneous effects of interventions is a challenging statistical problem, particularly when existing scientific knowledge (or even theories) about how effects vary by individuals and by context is lacking. Bayesian tree models, which combine Bayesian statistics and predictive methods from machine learning, haven proven to be some of the most effective methods for inferring heterogenous effects in large-scale empirical evaluations. However, current methods are limited to simple data structures, study designs, and outcomes, limiting their real-world applicability. It also can also be difficult to extract actionable insights from these sophisticated models, which is necessary for guiding the design of future interventions or studies and for understanding the implications of deploying an intervention at scale. This project will develop tree ensembles to encourage (partial) smoothing in estimated treatment effects with an eye to computational costs. The project also will adapt Bayesian tree models for heterogeneous effects to more complex data. The methods to be developed will be used to improve both the design and analysis of behavioral studies. The results of this project will advance knowledge in the fields of statistics, machine learning, data science, education, and behavioral science.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.