Causal Machine Learning with Reflective Latent Variables: A Latent Deep & Targeted Learning (LDTL) Architecture

Amota Ataneka

Despite significant advances in machine learning (ML) and causal inference, integrating latent variables into these frameworks has lagged behind. Latent covariates and outcomes (e.g., pre- and post-treatment mathematics achievement) are ubiquitous across disciplines and 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 ML has relaxed such limitations by developing data adaptive methods that empirically construct these relationships. However, a significant weakness of ML methods is that they have been almost exclusively developed for observed, manifest or composite variables and do not accommodate latent variables. Given the centrality of latent constructs in policy, practice, and theory-development research, this gap represents a significant barrier in the widespread uptake and implementation of advanced causal and ML methods. In this study, I develop causal machine learning methods to estimate effects within the context of the targeted learning framework that supports causal inference with latent constructs. The results suggest the proposed methods substantially improve upon conventional linear estimators (e.g., SEM) and contemporary ML estimators that ignore measurement error. I apply the methods to evaluate the effect of mental-health components of teacher induction programs on beginning teachers' emotional exhaustion, depersonalization, depressive symptoms, and anxiety, controlling for baseline factors such as teacher prior mental-health.

About Amota Ataneka

Amota Ataneka is a doctoral candidate in Quantitative Methodologies at the College of Education, Criminal Justice and Human Services at the University of Cincinnati. His dissertation develops AI based methods to estimate cause-and-effect relationships when outcomes and predictors are latent variables. The methods build a new system of data adaptive algorithms that concurrently address measurement error and model uncertainty through innovative deep and targeted learning architectures. The resulting method integrates latent variables with modern machine learning methods to produce more robust, rigorous and credible conclusions about how interventions, programs and policies affect important but imperfectly measured constructs. Amota grew up on Nikunau, an outer island in Kiribati, a place with no doctors, no lawyers, and virtually no formal employment. As a boy, he paid three coconuts on weekends to watch Hollywood movies on a tv-set owned by the village, where he first heard the word professor and decided, with a child's confidence, that that is what he wanted to be. In 2007-2008, Amota earned the highest Year 11-Year 12 GPA in the country, securing a full-ride scholarship to attend university in Australia. He then moved to America for master's and doctoral training at the frontier of quantitative methodologies and AI. Amota is currently the only person, and just the second in his country's history, to pursue doctoral studies in America. His long journey from subsistence living on a remote coral atoll in the Pacific to the cutting edge of causal machine learning is a story about what education can do.