Variance estimation in inverse probability weighted Cox models.

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Inverse probability weighted Cox models are crucial for estimating marginal hazard ratios in observational studies.
• The standard robust sandwich variance estimator often overestimates variance due to ignoring weight estimation uncertainty, leading to inefficient inference.

Purpose of the Study:

• To propose a novel variance estimator for inverse probability weighted Cox models that accounts for the uncertainty in weight estimation.
• To improve the efficiency and accuracy of hazard ratio estimation in observational studies.

Main Methods:

• Developed a new variance estimator using stacked estimating equations, integrating hazard ratio and weight estimation.
• Incorporated adjustments for non-independently and identically distributed terms in the Cox partial likelihood score equation.
• Extended the method to handle clustered data and proved its asymptotic equivalence to existing linearization methods.

Main Results:

• The proposed variance estimator is analytically shown to be less conservative than the robust sandwich estimator.
• Simulation studies demonstrate superior finite sample performance compared to alternative methods.
• The method was successfully applied to both independent and clustered data examples, including bariatric surgery and multiple readmission datasets.

Conclusions:

• The novel variance estimator provides more efficient and accurate inference for marginal hazard ratios from inverse probability weighted Cox models.
• The proposed method, implemented in the R package ipwCoxCSV, is suitable for both independent and clustered observational data.
• This advancement enhances the reliability of statistical analyses in epidemiological and clinical research.