Jennifer Starling was presented with the Thomas R. Ten Have Award at the Atlantic Causal Inference Conference held at Carnegie Mellon University May 21- 23, 2018, for her poster titled “Functional BART for Causal Inference” in collaboration with Jared S. Murray, Patricia A. Lohr, Abigail R.A. Aiken, Carlos M. Carvalho, and James G. Scott. Congratulations to Jennifer on her achievement!
Jennifer will be presenting the funBART method with applications to stillbirth risk at the 2018 ISBA World Meeting this June.
About the application and method: We assessed the effectiveness of two early medical abortion regimens: simultaneous, where mifepristone and misoprostol are administered during a single clinic visit, and interval, where there is a 24 to 48 hour wait between administration. Simultaneous is about 3% less effective on average, approximately 97% versus 94%, but it was unknown if that gap changes over gestational age. The data is observational, with 85% of women self-assigning to simultaneous. Our goal is to model the probability of successful procedure as a function of gestational age, dependent on patient characteristics, without specifying forms of interactions between covariates.
We introduce a new method called funBART, short for Functional BART, which extends BART (Chipman, George, McCulloch, 2010) into the functional response regression (function-on-scalar) space by replacing the scalar leaf priors with smooth functions using Gaussian Processes. We then extended funBART into the Bayesian Causal Forest space (Hahn et al. 2017).
We find that funBART has excellent predictive performance, and the full posterior allows us to quantify uncertainty. We find that there is not a significant increase in the effectiveness gap between the early medical abortion regimens over time, and we designed a web app for doctors, to assist in offering personalized advice to patients.
Future Plans: We developed funBART with the aim to investigate perinatal mortality outcomes, in collaboration with Radek Bukowski in the Dell Medical School. Planned future work includes extending funBART to various other data structures, with applications including personalized medicine, women's health, and public policy.
