A Bayesian latent class approach to causal inference with longitudinal data.

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Bayesian methods are increasingly vital in comparative effectiveness research.
• Limited exploration exists for parametric Bayesian causal approaches with time-dependent treatments and covariates.
• Time-dependent confounding poses challenges in longitudinal studies.

Purpose of the Study:

• To propose a fully Bayesian causal approach for longitudinal data with time-dependent treatments and confounders.
• To implement this approach using latent confounder classes representing disease and health status.
• To reduce the dimensionality of time-dependent confounders in causal effect estimation.

Main Methods:

• Developed a Bayesian g-computation framework.
• Incorporated latent class analysis to model unobserved patient states.
• Formulated joint likelihoods for treatment, outcome, and latent class models.
• Utilized simulation studies to evaluate method performance.

Main Results:

• The proposed Bayesian latent class method demonstrated effective handling of time-dependent confounding.
• Dimension reduction of confounders was achieved through latent class representation.
• Performance was compared favorably against existing causal methods for longitudinal data.

Conclusions:

• The novel Bayesian causal approach with latent classes offers a robust solution for longitudinal comparative effectiveness research.
• This method effectively addresses complex confounding structures involving time-dependent variables.
• The approach was successfully illustrated in a study of juvenile dermatomyositis treatment.