The purpose of this study was to evaluate the use of log-linear models in detecting test item bias. The evaluation of the log-linear approach in this regard will focussed on the following: (1) the stability of the procedure across replications--the effect of sampling error across random samples of ethnic groups; (2) the amount of artifactual bias--analyzed by comparing results for different ethnic groups with the results from two random equivalent white groups and with the results from two white groups which differ in measurability; and (3) the interpretability of the items identified as deviant. Two sets of items were selected for this study: one from the American College Testing program and one from a high school minimum competency test. Although the possible effect of sample size needs to be examined further, the results suggest that methods based on log-linear models could reasonably be used to examine sets of items for potential bias. (BW)