Causal effect sensitivity across a multiverse of structural equation models
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Replication is a core principle of science, and psychological research is undergoing a rapid transformation since news of its own "replication crisis" spread in the early 2010s. Among the changes is a growing awareness of the nuanced requisites for a firm causal statement to be possible. Unfortunately, even with the aim of obtaining data to support a convincing causal effect estimate, unavoidable study design limitations can permit these data to support a plausible "multiverse" of models in which this effect can be estimated. In the current study, we develop a SEM multiverse path test that assesses the sensitivity of a causal effect estimate across sets of competing structural equation models (SEMs) that are used to estimate this effect, each representing a different theory about the data. Specifically, our test statistic assesses whether the estimation of a causal effect from a single dataset significantly varies across competing SEMs. In cases where a causal effect estimate does not significantly vary across competing models, the differing structures belonging to these competing models could be ignored, suggesting robust effect estimation despite the structural differences. Conversely, if a path does vary, it qualifies the results. To help situate the motivation and development of our test, we will first discuss various factors of the replication crisis in psychology and the resurgence of causal modeling techniques meant to improve replicability. We will then discuss the theoretical details underlying our test, followed by illustrations of test performance using both simulated and real data. This will be capped with a general discussion that includes future steps about automating the generation of plausible models, indirect path testing, and publication of our test.
Table of Contents
PubMed ID
Degree
Ph. D.