Fitting and interpreting loglinear interactions in cross-classifications from health policy and medicine

A nontechnical exposition is presented of current statistical techniques for the analysis of multidimensional tables of counted data. Performing an original analysis of a data set of interest to researchers in health policy and medicine, the paper considers what kinds of questions an analysis by loglinear modeling can address, and what kinds of answers it can obtain and how they may be sought. Unlike most previous expository accounts seeking to provide introductions to this field, this paper does not require a background from the reader in either regression or the analysis of variance. By a thoroughgoing use of odds ratios and higher-order odds ratios, it nevertheless provides a technically accurate account of the key concepts of higher-order interactions among variables, and of models being hierarchical. Statistically more advanced readers are provided with a means of effectively expositing their loglinear modeling methods and conclusions to nonstatisticians; a number of footnotes are directed toward such readers.