Poster #17 - Causal Machine Learning with Latent Variables

• In Event: Friday Poster Reception

Abstract

Latent covariates and outcomes (e.g., pre- and post-treatment mathematics achievement) are ubiquitous in policy research. Prior research has developed a deep and diverse set of methods specifically designed to accommodate such variables (e.g., structural equation models [SEMs]). A critical limitation of conventional methods such as SEM is that they typically operate under assumptions of linear effects and almost always assume correct model specification (e.g., all non-linearities and n-way interactions are known and included). Recent research in machine learning has relaxed such limitations by developing data adaptive methods that empirically construct these relationships. However, a significant weakness of machine learning methods is that they have been almost exclusively developed for observed, manifest or composite variables and do not accommodate latent variables. In this study, we develop a machine learning super learner algorithm to estimate the average treatment effect within the context of the targeted learning framework that supports causal inference with latent variables. The results suggest the proposed estimator substantially improves upon conventional linear estimators (e.g., SEM) and machine learning estimators that ignore measurement error and. Moreover, the results suggest the proposed estimator returns nearly unbiased estimates in finite samples and is consistent even when the functional relationships are unknown or misspecified. We demonstrate the methods with applications in new teacher mental health support and maternal health education and compare their performance with conventional SEMs and machine learning approaches.

Author

• Ben Kelcey, University of Cincinnati
  Presenting Author