Some hazards of using nonstandard log-linear models, and how to avoid them.

A. von Eye and J. Brandtst?dter (1998) proposed 3 new concepts for log-linear modeling of categorical data for causal dependency: the wedge, the fork, and the chain. These concepts are difficult to operationalize uniquely and therefore difficult to translate unambiguously into statistical models. In addition, the statistical models proposed by von Eye and Brandtst?dter do not correspond to the verbal descriptions given in their article. For some of their data sets, simpler standard log-linear or logit models fit the data well and can be interpreted unambiguously. Nonstandard log-linear models offer the possibility of testing important substantive theories, but caution is required so that the statistical model appropriately captures the substantive theory. The present article demonstrates some techniques for determining the proper interpretation of a nonstandard model. (PsycINFO Database Record (c) 2010 APA, all rights reserved)