Accounting for random observation time in risk prediction with longitudinal markers: an imputation approach

Authors

• Y Han
• D Liu

Abstract

To predict the risk of clinical endpoints, it is important to incorporate longitudinally measured biomarkers, since subject-specific marker trajectory contains additional information on pathology and the critical windows. The work is motivated by the Scandinavian Fetal Growth Study, which aims at predicting pregnancy outcomes using repeated ultrasound measurements during pregnancy. While the measuring time of markers often varies across individuals, it is not well understood how the variations affect risk prediction. Existing methods of longitudinal risk prediction, such as shared random effects model (SREM) and pattern mixture model (PMM), construct a prediction implicitly as a function of the biomarkers and their observation time. Methods ignoring the longitudinal structure, such as sufficient dimension reduction (SDR) and logistic regression, have better interpretability regarding how a biomarker measured at specific time window contributes to the disease risk, but often result in reduced accuracy because of ignoring the observation time information. We propose a novel imputation approach to handle this problem: a longitudinal regression model is estimated to impute the marker value at the designed times, then the imputations are used to construct risk prediction using SDR or other cross-sectional methods. Through extensive simulation studies and analyses of Scandinavian Fetal Growth Study data, we systematically compare performances of the above methods in terms of their discrimination and risk calibration. The findings suggest that even small variations in the observation time can make considerable difference in risk prediction accuracy, and that our proposed method is more robust to model mis-specifications than SREM and PMM.

Scientific Focus Area: Epidemiology

This page was last updated on Friday, March 26, 2021