INTEGRATING INCOMPLETE DATA FOR MEDIATION ANALYSIS.

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Mediation analysis typically requires a single, complete dataset with exposure, mediator, and outcome variables.
• Existing methods are limited by the need for complete data, hindering analysis when data is fragmented or only summary statistics are available.

Purpose of the Study:

• To develop semiparametric methods for mediation analysis that can utilize incomplete datasets.
• To enable the estimation of direct and indirect causal effects by combining information from multiple data sources, including summary statistics.

Main Methods:

• Proposed semiparametric approach to estimate causal parameters (direct and indirect effects).
• Methodology designed to integrate data from several incomplete datasets, each containing only two of the three key variables (exposure, mediator, outcome).
• Capability to handle analyses using only summary statistics derived from these incomplete datasets.

Main Results:

• The developed methods provide asymptotically unbiased and normally distributed estimates of causal parameters.
• Simulations demonstrate the performance of the methods in finite samples and quantify efficiency loss due to incomplete data.
• Application to breast cancer risk data investigates mediation by terminal duct lobular units between polygenic risk scores and cancer risk.

Conclusions:

• Semiparametric methods offer a viable solution for mediation analysis when complete data is unavailable.
• The proposed approach enhances the utility of fragmented datasets and summary statistics for causal inference.
• The study successfully applies these methods to a relevant biomedical question in breast cancer etiology.