Estimating inverse probability weights using super learner when weight-model specification is unknown in a marginal...

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Learning

Background:

• Accurate inverse probability weighting (IPW) is crucial for valid inference using marginal structural Cox models (MSCM).
• Parametric models for IPW often face misspecification issues in real-world applications.
• Data-adaptive statistical learning methods offer potential alternatives to traditional parametric approaches.

Purpose of the Study:

• To evaluate the effectiveness of Super Learner (SL) for estimating IPW in MSCMs.
• To compare SL's performance against common statistical learning methods under various weight model specifications.
• To assess SL's utility in a real-world cohort study on multiple sclerosis treatment.

Main Methods:

• Simulated four scenarios with varying true weight model specifications (linear and/or additive).
• Employed Super Learner (SL), a cross-validation-based method, to select optimal IPW estimators from a diverse candidate set.
• Applied the methodology to a multiple sclerosis cohort to estimate the effect of beta-interferon on disability progression.

Main Results:

• Super Learner (SL) demonstrated superior performance compared to commonly used methods when the weight model was misspecified.
• SL achieved lower Mean Squared Error (MSE) and improved coverage probabilities for estimated effects in MSCMs.
• The study successfully applied MSCM with SL-estimated IPW to analyze beta-interferon treatment impact in multiple sclerosis.

Conclusions:

• Super Learner (SL) provides a robust and adaptable approach for estimating inverse probability weighting (IPW) in marginal structural Cox models (MSCM).
• SL is particularly beneficial when the true weight model specification is uncertain or misspecified.
• The findings support the use of SL in observational studies for more reliable causal effect estimation.