Statistical Inference: A Commentary for the Social and Behavioural Sciences.

Statistical Inference: A Commentary for the Social and Behavioural Sciences

By Oakes M. New York, NY 101 58, john Wiley & Sons, 1987, cloth, 185 pp, $41.95

What is the nature of statistical truth? The book's author answers this disturbing question using comparative and probabilistic statements; perhaps this is the only truthful way a statistical treatise can address this question. The text goes well beyond the usual treatment of statistics, however, challenging a variety of assumptions most of us hold without question. For example, one of the fundamental assumptions of statistical inferential tests is the use of one, and only one, "experiment-wise" inferential test. Theoretically, then, surveys and other multiple-inferential test research virtually always contain Type I errors. In a move that is characteristic of theoretical statisticians, the author disposes of this problem by asserting that no inferential statistics ought to be calculated; he thus avoids the question of what inferential statistical truth is for such "descriptive" work.

The critical acumen of this brief and well-written book is not only unsparingly penetrating but reasonably fair. For example, the author exposes the fallacies of Bayesian subjectivity along with classical, or Fisherian, pseudo-objectivity. He coherently reveals the often weak premises of r[.sup.2] and omega-squared as measures of variance explained; the usually ignored asymmetry of confidence intervals makes especially informative reading for researchers who thump the bible of statistical prowess in their zeal for converts.

The seven chapters in this exquisitely reasoned book are organized into three parts: a general commentary on the utility of statistical tests, a discussion of the origins and fundamental thinking of the schools of statistical inference, and a summary of the role and relevance of statistical tests in science today. The first part, on statistical tests, examines the conceptual underpinnings of power, Type I and Type II errors, sampling instability, and more esoteric topics that include the intuitive meaningfulness of correlation coefficients (which, after all, are descriptive and not inferential statistics). The second part of this book, "Schools of Statistical Inference," is a further bid by the author to exterminate the scourge of mindless reliance on p values, by exposing the historical origins and personal influences on development of statistical inferences. The author discusses, for example, why we should require an "alternative hypothesis" when Fisher himself "specifically eschews" their use. The third part of the text is perhaps most discomfiting.

The author casts aspersions on meta-analysis and other "nouveau-statistics," disparages the utility of new computer resampling techniques, and says that the future of covariate analysis, nonrandom samples, and nonrandom assignment is bleak.

One highlight of this book is its simplicity. Understanding the author's cogent arguments requires no more statistical sophistication than that required for understanding the assumptions of an independent-means t test. The main remedy the author proposes, however, that of greater homage to the fundamental postulates of statistics, is something of a straw man. No one disagrees that more rigorous treatment is preferable to sloppy scholarship but a better question to ask is, How closely must empirical data fit theoretical assumptions before such statistical tests may be applied? The author believes that, most often, inferential statistical analyses should not be applied, which makes this question moot, and he provides several excellent arguments bolstering the point. Still, the reader is left wishing for less vitriol and more verism in this epistemologically sound text. I enthusiastically recommend this book to persons who teach statistics to undergraduate or graduate students at any level. The work deserves to be widely read and discussed if only for its disquieting and thought-provoking exegesis of current statistical practice.

David E Krebs, PhD