Joint mixed-effects models for causal inference with longitudinal data.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational longitudinal data presents challenges for causal inference due to time-dependent and unmeasured confounding.
• Existing methods often require strong assumptions like no unmeasured confounders or missing at random, limiting their applicability.

Purpose of the Study:

• To propose a robust causal inference approach for observational longitudinal data with time-varying exposures.
• To address challenges of time-dependent confounding, unmeasured confounding, and missing data, including truncation by death.

Main Methods:

• A parametric joint mixed-effects model for outcome and exposure is combined with g-computation.
• The joint model, a shared parameter model, handles missing data and truncation by death, even when missingness is not at random.
• The approach estimates participant-specific or population-average causal effects.

Main Results:

• Simulation studies demonstrated the performance of the proposed method.
• The method was compared against linear mixed- and fixed-effects models with g-computation and targeted maximum likelihood estimation.
• The approach was applied to an epidemiologic study on vitamin D and depressive symptoms in older adults.

Conclusions:

• The proposed joint modeling with g-computation offers a flexible and powerful tool for causal inference in complex observational data.
• The method effectively handles time-dependent confounding, unmeasured confounding, and various missing data patterns.
• Availability of SAS code enhances the practical application of this advanced causal inference technique.