Structural-before/after-measurement approaches for estimating treatment effects in latent outcomes with causal machine learning.

Prior research has developed a deep and diverse set of methods specifically designed to accommodate reflective latent factors (e.g., structural equation models). A critical limitation of conventional latent variable methods is that they typically operate under assumptions of linearity and almost always presume correct model specification (e.g., all nonlinearities and n-way interactions are known and included). Recent research in machine learning has relaxed such limitations by developing data-adaptive methods that identify treatment effects by empirically estimating the functional forms of relationships among covariates, treatment, and an outcome. A significant weakness of causal machine learning methods, however, is that they have been almost exclusively developed for observed outcome variables and have not considered reflective latent outcomes that are measured with error. In this study, we developed structural-after/before-measurement methods suitable for targeted machine learning with latent outcomes. The results suggested that the proposed estimators are nearly unbiased in finite samples even when the functional relationships are unknown and unspecified. We demonstrate the methods with an application to teachers' burnout study and compare their performance with conventional linear structural equation models. (PsycInfo Database Record (c) 2026 APA, all rights reserved).