Statistical methods for evaluating diagnostic tests and biomarkers fall far behind existing methods for evaluating therapeutic drugs and etiologic epidemiology studies. This is a revised competitive renewal of a grant that in previous cycles has forged new fundamental methodology aimed at bringing biomarker evaluation to a level comparable with other areas of research. The 5-phase paradigm for biomarker development (Pepe et al JNCI2001) and the ROC-GLM regression framework (Pepe Biometrika 1997) are examples. In the next grant cycle we propose to again tackle basic issues in biomarker study design and analysis that have never been addressed before. These include:
in Aim 2 (a), evaluating the implications and advantages of matching cases to controls in regards to covariates, and selection of the optimal case-control ratio in designing a matched biomarker study;
in Aim 2 (b) developing methods for estimating the ROC derivative with implications for making inference in data analysis and for sample size and case control ratio calculations in study design. We propose a conceptually new and intuitive approach to biomarker evaluation in Aim 3 that uses the controls simply as a reference distribution for standardizing marker values. Traditionally difficult problems in biomarker evaluation such as covariate adjustment, examining equivalence of markers and event time outcome data are easily handled in this conceptual framework. A graphical display that describes the distribution of risk in the population is proposed in Aim 4 for the evaluation of risk prediction markers and models. This provides clinically meaningful descriptions of predictiveness that have not been attributes of standard predictiveness measures. Applications in cancer biomarker development provide a context for our research. Data from the Early Detection Research Network and from several large cohort studies including the Physicians'Health Study and the Prostate Cancer Prevention Trial will be analyzed. Relative to our previous submission, Aims 1 and 5 have been eliminated while aims 2, 3 and 4 have been elaborated upon and expanded.
Showing the most recent 10 out of 74 publications
