Statistical methods in biomarker evaluation and combination

Eunice Kennedy Shriver National Institute Of Child Health & Human Development

Investigators

Linked publications & trials

Abstract

A common practice in biomarker research is to take multiple measurements over time. The prediction accuracy could be substantially improved by accounting for the longitudinal trajectories of the biomarkers and their correlation structures. In addition, individual risk calculators could be developed based on the biomarker combination rule. The methods we developed can effectively combine longitudinal biomarkers in predicting pregnancy complications. Including more biomarkers into the combination does not always improve the prediction accuracy. Selection of markers is a crucial step, especially when the number of candidate markers is large. Marker selection involves several aspects: selecting markers with high classification power, selecting the important time points for making observations, and selecting subgroups of patients for enhanced prediction. An important objective of this project is developing the model selection techniques in the longitudinal biomarker combination framework. Incomplete observations are common in most longitudinal studies, but this problem has not been studied in the context of biomarker combination. Both the biomarker and the outcome could be missing due to loss of follow-up. The missingness might be informative if the dropout process depends on a subjects unobserved characteristics. For example, when predicting preterm birth, the subjects with early preterm tend to have a shorter sequence of biomarker evaluations. We will develop efficient and robust techniques to correct for the bias resulted from ignoring the missing data.