Causal inference in survival analysis under deterministic missingness of confounders in register data

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Long-term register data offer valuable insights into causal effects on time-to-event outcomes.
• Methodological challenges arise from missing confounders and informative censoring in such datasets.
• The Swedish Renal Registry serves as a motivating example for analyzing renal replacement therapies.

Purpose of the Study:

• To investigate the impact of missing confounders and informative censoring on causal effect estimation.
• To evaluate different imputation models and estimation methods for population average survival.
• To identify optimal strategies for analyzing time-to-event data in healthcare registries.

Main Methods:

• Utilizing multiple imputation techniques for missing covariate data.
• Comparing various imputation models, including those with cumulative baseline hazard and interactions.
• Applying regression standardization and inverse probability of treatment weighting for causal effect estimation.
• Assessing sensitivity to censoring mechanisms and model misspecification.

Main Results:

• An imputation model incorporating cumulative baseline hazard, event indicator, covariates, and their interactions, followed by regression standardization, yielded the best simulation results.
• Regression standardization effectively addressed informative censoring by including the entry date in the outcome model.
• Standardization offered advantages over inverse probability of treatment weighting, including straightforward variance computation.

Conclusions:

• The proposed imputation and regression standardization approach enhances causal effect estimation from register data.
• Accounting for entry date is crucial for handling informative censoring in survival analyses.
• This methodology provides a robust framework for analyzing complex time-to-event data in large-scale health registries.