Primary Submission Category: Machine Learning and Causal Inference
A Gaussian Process Framework for Survey-Based Event Studies
Authors: Soonhong Cho,
Presenting Author: Soonhong Cho*
Survey-based event studies exploit salient events that occur during the study period to estimate causal effects by comparing respondents interviewed before and after the event. This design faces methodological challenges mainly from (i) demographic differences between pre- and post-event respondent groups and (ii) time series complexities including trends, seasonality, and autocorrelation. Existing approaches use reweighting methods to address demographic imbalances, but these depend on researcher decisions about model specification, covariate binning, and balance metrics—choices that can substantially affect results. We propose a Gaussian Process (GP) regression framework that estimates counterfactuals by learning outcome-covariate relationships from pre-event data and projecting them to post-event periods. We simplify the GP hyperparameter structure to enable automated estimation with minimal user discretion. The framework flexibly learns time-varying relationships, provides principled uncertainty quantification, and automatically handles practical challenges like irregular response timing and missing data. Through simulation studies and political science applications, we demonstrate that our approach provides robust causal estimates while requiring fewer modeling assumptions than existing approaches.