Inference procedures in sequential trial emulation with survival outcomes: Comparing confidence intervals based on...

Area of Science:

• Causal inference
• Biostatistics
• Epidemiology

Background:

• Sequential trial emulation (STE) estimates causal effects from observational data.
• Inverse probability weighting addresses time-varying confounding and dependent censoring in STE.
• Accurate confidence intervals (CIs) are crucial for robust causal effect estimation in STE.

Purpose of the Study:

• To evaluate and compare different methods for constructing CIs for marginal risk differences in STE with survival outcomes.
• To assess the performance of nonparametric bootstrap, linearised estimating function (LEF) bootstrap, jackknife, and sandwich variance estimators.
• To provide guidance on selecting appropriate CI methods for causal survival analysis in STE.

Main Methods:

• Simulations were conducted to compare CI coverage across various methods.
• Methods evaluated included nonparametric bootstrap, LEF bootstrap, jackknife, and sandwich variance estimators.
• The focus was on estimating marginal risk differences in the context of STE with survival data.

Main Results:

• LEF bootstrap CIs showed superior coverage compared to nonparametric bootstrap and sandwich variance estimators in scenarios with small/moderate sample sizes, low event rates, and low treatment prevalence.
• LEF bootstrap was less sensitive to treatment group imbalance and computationally faster than nonparametric bootstrap.
• For large sample sizes and medium/high event rates, sandwich variance estimator CIs provided the best coverage and fastest computation.

Conclusions:

• LEF bootstrap is a recommended method for constructing CIs in STE, particularly in challenging scenarios common to STE.
• Sandwich variance estimators are efficient and effective for large sample sizes and higher event rates.
• These findings aid researchers in choosing optimal CI construction methods for causal survival analysis using STE.