Longitudinal mediation analysis with multilevel and latent growth models: a separable effects causal approach

Area of Science:

• Causal inference
• Biostatistics
• Longitudinal data analysis

Background:

• Causal mediation analysis is common in medical research, but estimating natural effects is challenging due to post-treatment confounding and latent variables in multilevel and latent growth models.
• Existing models struggle with accurate mediational effect estimation in longitudinal settings with complex data structures.

Purpose of the Study:

• To propose a novel causal interpretation for multilevel and latent growth models using separable mediational effects.
• To overcome limitations of natural effects in the presence of post-treatment confounding and latent variables.

Main Methods:

• Formal derivation of conditions for identifiability and analytical expressions for separable mediational effects using the g-formula.
• A simulation study to assess the impact of model misspecification and assumption violations on effect estimates.
• Application to real-world data.

Main Results:

• Model misspecification, especially severe cases, significantly impacts mediational effect estimates, with bias increasing over time.
• Violations of identifiability assumptions affect separable effect estimates differently in mixed-effect versus latent growth models.

Conclusions:

• The proposed separable effects provide a valid causal interpretation for multilevel and latent growth models.
• Emphasizes the critical importance of careful model selection due to the substantial impact of misspecification on effect estimates.