The statistical evaluation of medical tests for classification and prediction.

*(English)*Zbl 1039.62105
Oxford Statistical Science Series 28. Oxford: Oxford University Press (ISBN 0-19-850984-7/hbk). xvi, 302 p. (2003).

The book provides a structured summary of existing statistical methodology and practice of research studies aiming at evaluating medical tests and biomarkers for classification and prediction. It is organized in nine chapters.

The first chapter introduces the broader context of medical tests and study designs as well as seven data sets used to illustrate the application of the described statistical theory throughout the book. The next two chapters deal with medical tests yielding dichotomous results. Various measures for quantifying diagnostic accuracy and their estimates are described in chapter 2. Chapter 3 provides methods to compare diagnostic accuracy of medical tests and several regression modelling frameworks to evaluate factors potentially affecting test performance. The ROC curves and their summary indices as measures to quantify diagnostic accuracy of continuous and ordinal-valued tests are discussed in chapters 4 to 6. Empirical methods, distribution-free parametric methods, and parametric distribution methods for the estimation of ROC curves are introduced. Further issues are the comparison of ROC curves, their summary indices, and regression method approaches for evaluating covariate effects on the diagnostic accuracy of non-binary tests.

The impact and statistical methods to deal with the error arising through incompleteness in the data and the common issue of uncertain true classification are described in chapter 7. The concept of a series of studies for the development and evaluation of a medical test is introduced in chapter 8. The appropriate design and sample size calculations for each of these studies are discussed. The last chapter describes further extensions and applications of the statistical methodology, e.g., the analysis of event time data, the question of how to combine results of several predictors to define a better composite test, and meta-analysis procedures to combine the results of several studies.

In summary this book is to be recommended to both practicing and more academic research biostatisticians.

The first chapter introduces the broader context of medical tests and study designs as well as seven data sets used to illustrate the application of the described statistical theory throughout the book. The next two chapters deal with medical tests yielding dichotomous results. Various measures for quantifying diagnostic accuracy and their estimates are described in chapter 2. Chapter 3 provides methods to compare diagnostic accuracy of medical tests and several regression modelling frameworks to evaluate factors potentially affecting test performance. The ROC curves and their summary indices as measures to quantify diagnostic accuracy of continuous and ordinal-valued tests are discussed in chapters 4 to 6. Empirical methods, distribution-free parametric methods, and parametric distribution methods for the estimation of ROC curves are introduced. Further issues are the comparison of ROC curves, their summary indices, and regression method approaches for evaluating covariate effects on the diagnostic accuracy of non-binary tests.

The impact and statistical methods to deal with the error arising through incompleteness in the data and the common issue of uncertain true classification are described in chapter 7. The concept of a series of studies for the development and evaluation of a medical test is introduced in chapter 8. The appropriate design and sample size calculations for each of these studies are discussed. The last chapter describes further extensions and applications of the statistical methodology, e.g., the analysis of event time data, the question of how to combine results of several predictors to define a better composite test, and meta-analysis procedures to combine the results of several studies.

In summary this book is to be recommended to both practicing and more academic research biostatisticians.

Reviewer: Carina Ittrich (Heidelberg)