Predictive Liminology: Methods for Predictive Modelling. - book reviews

Ecology, Jan, 1997 by John T. Lehman

This book is an intriguing and useful text, although it is daunting at first impression. The authors wrote that they intended it to be both a text for graduate courses and also a cookbook of recipes for predictive modeling. In fact, the large format volume (30 x 22 cm) is packed with hundreds of data tables, graphs, and regression equations, to such extent that your head can literally spin as you try to leaf between text references to items and the graphics which may be pages away. There are both good things and bad things to report, but on balance the good outweighs the bad.

First of all, the book certainly lives up to its title. It gives the reader a fine introduction and how-to course on the construction of regression-based statistical models and on blending these models with simple dynamic equations. There are plentiful and helpful prescriptions about inspecting the data before loading them into the regression hopper. Advice and examples about frequency distributions, transformations, dealing with outliers, and uncertainty analysis are kept simple and direct. The emphasis is on predicting quantitative outcomes, like the amount of mercury in fish tissue, rates of primary production or sedimentation in lakes, or the chemical contents of the water. The authors make a credible effort to develop an algorithm for measurement of "predictive power" of the models, an effort that requires a way to balance the collective statistical statements embodied in [r.sup.2] values with the propagated uncertainty of the model equations and their parameters. The result is a rationale by which to judge whether the increased statistical fit that can grow with model complexity reaches a diminished return in predictive value. Not surprisingly, we learn that the simplest models often make the most powerful quantitative predictions, a lesson that recalls a remark first credited to Einstein: "Everything should be as simple as possible, but no simpler."

Balanced against the evident good content of the volume is the fact that technical editing and layout leaves a lot to be desired. Numerous sentences are simply ungrammatical, and the prose is less flowing than readers had grown accustomed to from the second author. It is also possible to become distracted by the authors' habit of informing the reader what the text is NOT intended to do, or how NOT to interpret a result, or what is NOT the aim of the chapter, with capitalized exuberance that seems to shout from print as many as four times on a single page. The reader is continually reminded that the data are drawn "from 25 Swedish lakes," from 95 Swedish lakes, or "from 926 Swedish lakes" and that model results must NOT be applied to other lake types, sizes, or ranges of attributes. In that sense, the work is intensely focused, and it would seem that the models must best be treated as analogies to the kinds of statistical treatments that could be applied to lakes in other regions. The restricted regional focus may prove a drawback to its adoption as a text outside northern Europe. It is evident that the first author had at his command a vast matrix of lake survey data for secchi depths, total P concentration, lake water color, alkalinity, conductivity, pH, as well as catchment basin maps, and the book is replete with statistical interactions among these properties, or more commonly their logarithmic, square root, or exponential transformations. Clearly Hakanson has achieved success with efforts to model environmental properties of practical necessity, such as mercury and radiocesium contamination of the lakes and landscape, and this book teaches how to do so, step by step.

My deepest sadness, however, was to learn while reading this book that the second author, Rob Peters, had passed away after a painful illness. This is a tragedy for our science. This text gives us a glimpse of a maturing face of the discipline that Peters had promoted so eloquently as "predictive limnology." Here we find efforts to develop predictions that are based on underlying causation, and to forge a union of mechanistic processes with empirical patterns. The emerging models promised to be ever more useful, instructive, and illuminating. We mourn the loss of a great intellect, but one who has enriched us with a legacy of philosophical and scientific work that will endure and grow.

JOHN T. LEHMAN University of Michigan Department of Biology Natural Science Building Ann Arbor, Michigan 48109