Multivariate statistics book hits target – mostly

Abstract

Multivariable Analysis. A Practical Guide for Cliniciansby Mitchell H. KatzCambridge University Press, 2000. £18.95 pbk (xv + 192 pages) ISBN 0 521 59301 9Multivariable Analysis by Mitchell H. Katz hits its stated target by providing a book that should be of help to clinical researchers performing their first research project requiring multivariable analysis. However, the title seems to be a little misleading in that the author does not attempt to cover multivariable (or multivariate) methods thoroughly, but provides a primer on the most commonly used modeling methods for clinical research. Nevertheless, this book will make interesting reading for basic scientists also, who might be reticent to use these methods. Methods of multivariable analysis should be particularly valuable to immunologists, because it is becoming increasingly clear that endpoints of interest (e.g. resistance to a particular pathogen), generally, are not dependent on one mechanism, but on many. Thus, it would be reasonable to use multivariable methods to understand or predict these endpoints.The use of numerous examples from the clinical literature is one strong point of this book; investigators are likely to find an example similar enough to their planned study to select an appropriate method of analysis and avoid common pitfalls. Another strong point is the author's unintimidating question and answer approach. The reader is led quickly through a basic explanation of multivariable analyses and the experimental designs for which their use is applicable. Useful diagrams depicting key concepts, such as confounders and suppressors, are included. Selecting an appropriate method of multivariable analysis, setting up the analysis (including issues such as determining the sample size required) and performing the analysis are covered using questions that are likely to arise during this process. Interpreting the data, checking that the data do not violate the assumptions of the analysis (and dealing with data that do violate assumptions) and validating models are discussed in the final chapters. The last chapter, ‘Steps for constructing a multivariable model’, is an excellent, two-page summary of the book in a sequential, step-by-step format.However, a few issues are dealt with slightly superficially. For example, data transformation and dealing with outliers are matters of considerable discussion and varying opinion among statisticians. Multivariable Analysis gives the impression that handling these matters is straightforward, and readers who encounter differences of opinion among reviewers or consulting statisticians might be unpleasantly surprised. In addition, there are a few problems with terminology (e.g. coefficient of determination is referred to mistakenly as coefficient of variation on one occasion) and with examples that do not explain clearly how all the information was obtained.A quick reading of this book should give investigators sufficient background information on multivariable methods to formulate reasonable questions and understand the answers of a collaborating statistician. A more careful reading should prepare an investigator to use these methods independently, but not without another step. One must still be prepared to read the manual or seek the advice of an experienced user of one of several software packages to learn basics, such as the format of data entry and how missing data are handled. The discussion of software packages and their relative merits is quite limited in the book, but the advice offered is solid. However, the idea that this book and a software manual (or knowledgeable colleague) will allow one to perform these analyses independently might be slightly optimistic in most cases. With most software packages [particularly SAS (SAS Institute, Cary, NC, USA)], it would probably be wise to consult a statistician when interpreting the results of the diagnostic tests performed by the program.Many immunologists would do well to read this book and apply these methods. However, one note of caution is that all of these methods assume that data for dependent and independent variables will be obtained from each subject. In other words, generally, one cannot obtain data on immune function from one set of rodents and data on host resistance from another set and then model or analyze these two data sets using multivariable methods. Under some circumstances, it is possible to violate this assumption and use data from separate sets of mice, but a rather arduous validation is necessary to confirm that this violation has not affected substantially the outcome of the analysis 1xEvaluation of multivariate statistical methods for analysis and modeling of immunotoxicology data. Keil, D. et al. Toxicol. Sci. 1999; 51: 245–258Crossref | PubMed | Scopus (14)See all