Comparing approaches to causal inference for longitudinal data: inverse probability weighting versus propensity...

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational studies lack randomization, leading to potential bias in treatment effect estimation.
• Standard methods like Inverse Probability of Treatment Weighting (IPTW) and Propensity Score (PS) are used for consistent causal effect estimation.
• Time-dependent covariates affected by prior treatment complicate longitudinal causal effect estimation.

Purpose of the Study:

• To compare the performance of IPTW and the Generalized Propensity Score (GPS) for causal inference.
• To extend the GPS approach to the longitudinal setting for time-dependent treatments and covariates.
• To evaluate which method, IPTW or GPS, yields estimators with lower Mean-Square Error (MSE).

Main Methods:

• Utilized Propensity Score (PS) methods, including Inverse Probability of Treatment Weighting (IPTW).
• Introduced and extended the Generalized Propensity Score (GPS) for longitudinal data analysis.
• Conducted three simulation studies and analyzed two real-world datasets to compare methods.

Main Results:

• Propensity Score (PS) methods generally produce estimators with lower Mean-Square Error (MSE) than IPTW in simple cases.
• The Generalized Propensity Score (GPS) extension to the longitudinal setting also demonstrated lower MSE in simulations.
• GPS appears to be a robust approach for causal inference in complex longitudinal settings.

Conclusions:

• The Generalized Propensity Score (GPS) method shows promise for improving causal effect estimation in observational studies, particularly in longitudinal settings.
• GPS may offer advantages over traditional IPTW in terms of Mean-Square Error (MSE).
• Further application and validation of GPS in diverse real-world scenarios are warranted.