Logistic regression and CART in the analysis of multimarker studies

Abstract

Background: the rapid development of new biomarkers increasingly motivates multimarker studies to assess/compare the value of different markers for risk stratification or diagnostic prediction. Analysis of these studies is usually governed by logistic regression (LR) which is however often applied quite uncritically resulting in unclear or even deceptive results.

Methods and data: this methodological review is intended for the practical medical researcher and critically discusses in non-technical terms where possible LR models and CART (classification and regression trees) as analytical approaches for multimarker studies. Some in practice often neglected but vital aspects for a valid application and interpretation of LR models are presented and the basic principles of building and interpreting a CART tree are demonstrated. All approaches are applied to original data of a multimarker study which aimed to assess and compare the value of different laboratory parameters to predict the need for intensive care treatment in unselected emergency room patients.

Conclusions: when applied prudently, both LR and CART are suitable for the analysis of multimarker studies. For a valid interpretation of LR it is especially important that the format of the model fits the study design. In general LR and CART will identify very similar sets of significant markers. While LR models generally focus more on the relative statistical significance of the assessed markers, CART results emphasise the absolute effects; results come as observed risk groups. Thus the complementary use of both techniques seems to be a promising approach to perform the difficult task of analysing and interpreting the results of multimarker studies.