Multivariate Analysis: Techniques for Educational and Psychological Research

This second edition of Tatsuoka's widely used text is, in my opinion, required reading for anyone who has the idealistic goal of devoting his life to instruction in an environment where demands, tools, standards, professional recognition, even hope of financial success, change rapidly. This book, like the first edition, is not a dogmatic textbook, of which we have so many in mathematical statistics but, as is evident in every chapter, shows the author's struggle to provide students with what they need (proficiency) rather than what they want (cookbook, phrases, ready-made tools). I fully agree with the author's desire to encourage students to write their own programs rather than applying ready-made packages. My experience, which has been very similar to Tatsuoka's, would dictate a different ordering of the chapters. In the review below, I will indicate this and attempt to compare the approach followed in the first (1971) edition with that in the current (1988) edition. The placement of chapter 2 (Matrix manipulations) at the beginning of this text seems to be an open invitation to instructors with little or no knowledge of either mathematics or educational research, to teach a course by being merely a chapter ahead of their students. As the author doubtless knows, these techniques can be taught only in a laboratory atmosphere, by careful supervision of active work. In this connection, I prefer the approach in the first edition, where the matrix manipulations were used to cast a univariate technique (multiple regression) into a notation that permits the introduction of multivariate techniques by analogy. In the beginning of chapter 4, the author refers to Multiple Regression as an exception to multivariate techniques. In all statistical curricula that I have seen in the past 35 years, it is considered a univariate technique, because it involves only one response variable, and thus chapter 3, at the very beginning, may discourage readers from continuing in this text. The inclusion of multiple correlation, clearly a multivariate technique, makes this all the more confusing. The shrinkage formula (p. 52) and the incremental R -test (p. 50), which are Beta statistics borrowed from univariate analysis, seem unnecessary in this context, because the multivariate noncentral distribution of the multiple correlation is as easily evalu-