Mathematical Geology, Vol. 20, No. 2, 1988
Book Review Applied Linear Regression, 2rid ed. by Sanford Weisberg John Wiley & Sons, New York, 1985, 324 p., $34.95 (U.S.). Regression is one of the more common methods of relating one or more variables to another in a predictive model. However, it is perhaps the most abused statistical technique because the model is so simple to explain in concept. Inherent in regression models are many assumptions that are seldom tested or fulfilled in practice. Any text which attempts to explain applied regression, therefore, must be examined carefully. This is not a book for the beginner, nor is it a book for the advanced practitioner. It really has no recognizable niche. In the preface to the first edition which follows a short preface to the second edition, the author states that this book " i f used as a textbook, is intended as a second or third course in statistics." I may have lost touch with modem teaching standards these days, but I don't think this statement is warranted. As a reference book, this is surely not the best, but it is not the worst either; as a textbook, I find it disastrous. For instance, let's take a look at Chapter 1, certainly a good place to begin. Here basic concepts and general simple models are introduced. On page 7 we are confronted with the statistical error: E ( e i ) = 0, var (eg) = o "2 and cov (e/, ej ) = 0 in equation 1.2. Now this is supposed to be a text for advanced folks, so we should know what these are by now. But the author does not bother to tell us that E means expectation, as we should know that. In the Appendix, and on page 277, E is expectation, hut the author did not suggest that we consult the appendix. On page 9, Table 1.2 definitions of variance and covariance, as well as other useful quantities, are given. Here, variance is SD~ and SDy2, not var (x) or var ( y ) and covariance is Sxy, not coy (x, y), as is used throughout the text. The same problem occurs with the use of corr (xy) and r (xy) which are never cross-used in the text as the correlation coefficient. We press onward in this first chapter and find a problem with a strange word "unbiasedness," (I should know this word as I am an advanced student by now.) on page 13. In the next section on the same page, without explanation, Appendix 1A.4 is referred to. On page 281 we find an explanation, which with 135 0882-8121/88/020o-0135506.00/I © 1988 International Association for Mathematical Geology
136
Book Review
our previous background, we should, and probably would, understand. However, we first need to read the two previous sections of this appendix in order to read the part to which we are referred. (If you really look into this, you actually should be instructed to read the appendix long before you read the first chapter.) Anyway, I can now find out what the author means by bias--I think. I can certainly understand his arguments if I try. But should I slip, I am led to the final statement which says if I rearrange the terms on page 281 in the final equation then I'll get the answer for equation 1.26 on page 22 of Chapter 1, which is only about 9 pages ahead of where we were in the text. In fairness, it doesn't get any worse, but it doesn't get any better either. This exercise hopefully is sufficient to give an idea of why I ' m confused about my opinion of this book. I certainly think the introductory chapter should be just that--an introduction--including the notation to be used, and I don't care care what level of expertise the reader might have at this point. Most areas of classical and modem regression are covered in 12 chapters, some quite well and others not so well. The chapter on multiple regression is tedious and breaks with the rest of the text (except for two later chapters) with the introduction of an elaborate and distracting type font. An Appendix is only for those with a basic to fair background in matrix algebra. Unfortunately, it is badly arranged and confused with the text, as previously indicated. In addition, a short Tables section gives only the standard tables of the Student t distribution, Chi-squared distribution, F distribution, and a Rankits test. Two indices are included: Subject and Author. The first is relatively complete, but the latter has problems. Of the 200 or so author references, over 50 % are 1975 or much older, and the newer ones are attributable to a handful of workers. I mean here that the references are somewhat out of date even though there are some as recent as 1984. However, most of the basic literature is covered. Numerous typographical errors are found. As usual, they are a nuisance, but I do not consider them to materially alter the content nor to be excessive. I do have a particular pet peeve. Problems are given at the end of many of the chapters, but answers to the problems are never given. In some cases, wording of the problems is vague, and in at least one case, I could think of several approaches to a valid solution. It would have been nice to find out what the author thought about the problem. My overall impression of this book is that it is a compilation of a series of lecture notes, refined over time, but never merged into an integrated text. I base my guess on the inconsistencies in the text and the appendix. Also the shortness and inadequate descriptions of more modem techniques in the latter chapters suggest that they were added as an afterthought.
Book Review
137
If you are a beginner or are advanced in regression, I cannot recommend this book. But I can recommend it if you have no other on the subject. 1 recently read a review of another book by another author on another subject. The reviewer stated that he felt privileged to review the subject book as he got a free copy of it. I feel somewhat that way myself, for this is an attractively bound book. However, I don't think that even at the moderate price of $34.95 I would have purchased it. In summary, this is an interesting book, but disorganized. All the facts are probably there in some way, but unfortunately it's a bit more effort to dig out the pieces and put them together than I ' m willing to spend. W. Brent Hempkins
Applied Mathematics & Statistics Group Engineering Technology Department Chevron Corporation P. O. Box 5030 San Ramon, California 94583-0930 U.S.A.