Data Adaptive Causal Learning with Latent Variables

• In Event: Survey Data Analysis, Issues in Measurement Error, and Latent Variable Modeling

Abstract

Purpose
Latent variables (e.g., achievement, efficacy, satisfaction) are a cornerstone to implementing empirical research across public policy and management areas. Prior research has developed a diverse set of methods to accommodate such variables (e.g., structural equation models [SEMs]) and address measurement error. A limitation of SEM and other methods is that they provide unbiased estimates and valid inferences only under correct model specification. Recent research in machine learning (ML) has developed data adaptive methods that relax or entirely remove such model specification requirements. However, a significant weakness of extant ML methods is that they do not accommodate reflective latent variables. We developed and integrated novel deep learning architectures and targeted learning methods for causal inference with latent variables (e.g., to empirically assess the effect of a treatment/policy on a latent outcome controlling for latent/observed covariates). The method relaxes model specification requirements while addressing measurement error to provide a flexible framework for double robust causal effect estimation with latent variables. Results suggested the method returns efficient and nearly unbiased estimates across diverse/adverse generating processes/conditions. 


Background
A critical limitation of structural equation models (SEMs) and other extant methods is that the consistency of parameter estimates depends on correct model specification. This assumption is restrictive when, for example, the functional forms of relationships among observed/latent variables are complex or unknown as they often (or arguably always) are in real-world evaluations. Conventional SEM methods lack a data-adaptive mechanism to detect and adjust for complex relationships without explicit (a priori) specification. As a result, parameter estimates are highly susceptible to model misspecification bias and this limits the utility/robustness of SEM estimates in real-world settings where functional forms are unknown. 


Method
We developed a framework that integrated structurally informed, measurement-aligned variational autoencoders with targeted learning for robust treatment effect estimation. Specifically, our approach builds on and synthesizes two frameworks: generative learning and targeted learning. On the generative side, our approach integrates SEM-style structures (e.g., measurement and structural models) with deep learning to model latent variables as (potentially unknown) functions of their indicators and their (potentially unknown) causally upstream structural relationships. Our architecture leverages domain knowledge about the factor structures and structural connections (similar to SEM) to develop a measurement-aligned, structurally-informed variational autoencoders (VAEs) that jointly learn the measurement relationships linking indicators to latent variables and the structural relationships connecting the variables. The result is that the generative model produces interpretable latent variables that are structurally consistent through design rather than regularization. On the targeted learning side, we construct an efficient influence function for the treatment effect with the SEM-VAE implied posterior distribution of the latent variables to estimate the average effect.


Illustration
We illustrate the method by applying it to several case studies including the effect of whole-school reform programs on student achievement, the impact of teacher mental health induction programs on teacher emotional exhaustion and attrition, and the impact of a nutrition education program on the health practices of new mothers.

Authors

• Ben Kelcey, University of Cincinnati
  Presenting Author

• Amota Ataneka, University of Cincinnati
  Non-Presenting Co-Author

• Jean Baptiste Habarurema, University of Cincinnati Main Campus
  Non-Presenting Co-Author