Multivariate techniques to assess laboratory tests in cancer patients.

In this chapter the application of multivariate techniques for the assessment of laboratory tests in cancer patients has been reviewed. We emphasize that the transformation of laboratory test values into just two categories (normal or abnormal) may entail a considerable loss of information. For instance, correlation between two laboratory tests that may be important for differentiating among various clinical categories of patients may disappear when this procedure is used. When only a single set of laboratory results measured in the same specimen is available for a given patient, we must compare these values to those obtained from other patients or healthy subjects to make inferences about the patient on the basis of the laboratory results. Thus, the analysis of the data must be group based. Discriminant analysis, logistic regression analysis, and survival analysis based on Cox's regression model are the techniques most often used in this situation. By contrast, when previous results are available from the same patient we may compare his or her present values to those previously obtained when we want to make inferences about the patient. Our objective is to make a prediction about the time that will elapse until some specified event (death or recurrence of disease) occurs. Two models that have been applied in this situation--the Markov chain and the autoregressive time series model--were reviewed and examples of specific medical applications presented.