Primary Submission Category: Heterogeneous Treatment Effects
Bayesian Causal Forests for Ordinal Outcomes: Effects of A Synergistic Mindset Intervention on Adolescent Self Regard
Authors: Anna Morgan, Jared Murray, David Yeager,
Presenting Author: Anna Morgan*
We introduce a nonparametric Bayesian approach for estimating heterogeneous effects with ordered categorical outcomes, building upon Bayesian Causal Forests (BCF). Continuous regression methods have many notable weaknesses when used with ordinal outcome data, including sensitivity to researchers’ choice of outcome scores (which may be arbitrary) and susceptibility to detecting spurious interactions due to “ceiling” and “floor” effects. Because they may give misleading reports of effect moderation, continuous methods cannot take full advantage of BCF’s ability to capture treatment effect heterogeneity. Our method overcomes these concerns by modeling the ordinal outcome on the latent scale using BCF’s sums-of-trees and a probit link function. Furthermore, by accurately estimating posterior category probabilities, our model allows for partial identification of parameters directly related to the joint distribution of potential outcomes, such as the probability that the treatment is strictly beneficial. We propose reasonable default priors for parameters of the treatment effect function and discuss how these priors may be informed by beliefs about the potential scale of treatment effects, illustrated by simulation studies. We demonstrate the benefits of our method with a reanalysis of an experimental study evaluating the causal effects of a synergistic mindset intervention on reported self-regard among low-SES high-school students.