Confidence Interval Construction for Causally Generalized Estimates With Target Sample Summary Information.

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Generalizing causal findings like the average treatment effect (ATE) across populations is crucial in biomedical research.
• Covariate shift, or differences in treatment effect modifiers' distributions, can cause ATEs to vary between source and target populations.
• Existing weighting methods to estimate target ATEs using summary data face limitations in statistical inference due to reliance on individual-level data for variance calculation.

Purpose of the Study:

• To propose a novel resampling-based perturbation method for constructing confidence intervals for the estimated target ATE.
• To enable accurate statistical inference for the target ATE when only summary-level information is available.
• To address the limitations of previous methods that require individual-level data for variance estimation.

Main Methods:

• Developed a resampling-based perturbation technique for confidence interval construction.
• Utilized additional summary-level information from the target sample.
• Applied the method in both simulation studies and real-world data analyses.

Main Results:

• The proposed method effectively constructs confidence intervals for the target ATE using only summary-level data.
• Demonstrated the approach's validity and effectiveness in simulation settings.
• Validated the method's performance on real-world datasets, confirming its practical applicability.

Conclusions:

• The resampling-based perturbation method offers a viable solution for statistical inference of the target ATE under covariate shift when only summary data is accessible.
• This approach overcomes the limitations of prior methods by avoiding the need for individual-level target data.
• The findings enhance the ability to generalize causal findings in biomedical research using readily available summary statistics.