Identificação do efeito causal no modelo de mediação com variáveis latentes

Abstract

Causal mediation analysis, which is based on potential responses (counterfactuals), is a commonly used method for decomposing the causal effect of an intervention on an outcome in several applications. This method is widely used in various areas of knowledge, particularly in epidemiology and social sciences. The most well-known methods are described in terms of continuous variables, especially linear models, and are used in situations where the variables are measured without error. In certain scenarios, however, the mediator and/or outcome may not be directly observed but can be potentially defined through latent class models. The goal of this dissertation is to assess how estimates of natural direct and indirect effects behave under the identification criteria used in the causal mediation models that involve categorical latent variables, using latent class analysis (LCA), in situations that may include a latent mediator and/or outcome. The methods for computing the natural indirect effect (NIE) and the natural direct effect (NDE) are expanded to situations where the categorical latent variables have more than two classes. We also propose the use of propensity scores in structural marginal models with latent variables. To evaluate the effectiveness of our proposed methods, Monte Carlo simulation studies were conducted under different scenarios of violation of causal identification assumptions. We illustrate all methodologies for estimating the NIE and NDE in situations involving categorical latent variables through the analysis of real data. Our analysis evaluates the effects of an intersectoral health promotion intervention, related to diet and physical activity patterns, on obesity in adolescents, where lifestyle is the mediator. We also evaluate the impact of the municipal health management on the quality of child care by primary health care (PHC) teams, which is mediated by the quality of planning and organization of PHC services. The obtained results emphasize the significance of causal identification criteria to allow for the causal interpretation of mediated effects, which can provide valuable insights to advance knowledge. Additionally, the findings suggest potential areas for future research and underscore the importance of methodological rigor in estimating and identifying mediated causal effects.