Estimating time-varying treatment effects in longitudinal studies.

Area of Science:

• Causal Inference
• Longitudinal Data Analysis
• Biostatistics

Background:

• Longitudinal studies assess treatment effects over time, but time-varying confounding poses significant analytical challenges.
• Standard regression adjustment for time-varying covariates can introduce bias, either by including post-treatment variables or omitting crucial confounders.

Purpose of the Study:

• To introduce g-estimation, a causal inference strategy for evaluating time-varying treatments amidst time-varying confounding.
• To provide a flexible and robust method for unbiased estimation of causal effects in longitudinal research.

Main Methods:

• G-estimation adjusts for confounding by utilizing only pre-treatment instances of all variables.
• This approach allows for effect modification analysis via covariate-treatment interactions.
• Mean models for treatment or outcome can be specified using standard regression functions.

Main Results:

• G-estimation effectively handles time-varying confounding, preventing bias from both inclusion and omission of time-varying covariates.
• The method accommodates continuous or non-continuous treatments and permits various mean models.
• Unbiased estimation requires correct specification of either the treatment or outcome model, not necessarily both.

Conclusions:

• G-estimation is an effective, flexible, and robust strategy for causal inference with time-varying treatments and confounders.
• The method is relatively straightforward to implement using standard statistical software.
• This approach enhances the validity of inferences in complex longitudinal study designs.