A1360
Title: A subjective Bayesian approach to address unchecked assumptions in a causal analysis
Authors: Jaiyool Kim - Florida State University (United States) [presenting]
Jonathan Bradley - Florida State University (United States)
Indrabati Bhattacharya - Florida State University (United States)
Abstract: Causal analyses typically rely on four standard assumptions, yet in many applications these assumptions may not hold. As a result, estimates may be biased and the uncertainty understated. The purpose is to introduce a subjective Bayesian framework that introduces a dichotomous parameter, which equals one when causal assumptions hold and equals zero otherwise. When the assumptions do not hold, the potential outcome is assumed to be unobserved and follows a semi-parametric statistical model, and the prior probability that the assumptions hold is pre-specified to a particular value p. This new perspective leads to a new inferential question. Namely, what is the smallest value of p that leads to a significant average treatment effect? In simulation studies, data is generated from an observational model that violates consistency through a spatially structured mediator, but intentionally fit a potential-outcome model that omits this mediator, mimicking an analysis with an unverifiable consistency assumption. Bias and interval coverage of the average treatment effect are evaluated. The method is also applied to county-level data on fine particulate matter exposure and COVID-19 mortality in the United States. The results demonstrate how to address unchecked subjective variability present due to unverifiable assumptions.
