A synthetic data integration framework to leverage external summary-level information from heterogeneous populations.

Area of Science:

• Biostatistics
• Statistical Inference
• Risk Prediction Modeling

Background:

• Growing need for flexible frameworks integrating individual data with external summary information.
• External information for risk prediction can be in various forms (e.g., regression coefficients, predicted values).
• External models may differ in predictors, algorithms, and underlying populations compared to internal studies.

Purpose of the Study:

• To propose an imputation-based methodology for fitting target regression models using all internal study predictors.
• To leverage external summary information from models that may use only a subset of predictors.
• To address prostate cancer risk prediction with novel biomarkers measured only in the internal study.

Main Methods:

• Proposes an imputation-based methodology for statistical inference.
• Generates synthetic outcome data in external populations.
• Utilizes stacked multiple imputation for a long dataset with complete covariate information.
• Conducts final analysis via weighted regression on the stacked imputed data.

Main Results:

• Improves statistical efficiency of estimated coefficients in the internal study.
• Enhances predictions by incorporating partial information from external models.
• Enables statistical inference for external populations, accommodating potential covariate effect heterogeneity.

Conclusions:

• The proposed flexible and unified approach effectively integrates individual and external summary data.
• It improves statistical efficiency and prediction accuracy in risk modeling.
• The method provides robust statistical inference across different populations with varying covariate effects.