Accounting for aberrant test response patterns using multilevel models

Hypotheses about aberrant test-response behavior and hence invalid person-measurement have hitherto included factors like ability, gender, language, test-anxiety, and motivation, but these have not previously been collectively investigated with real data, or with multilevel models. This study analyzes the effect of these factors on person aberrance using a real mathematics assessment data set under the framework of a two-level (person and classroom) hierarchical model. The results suggest that higher-scoring pupils, and, to a lesser extent, second-language learners are significantly more often aberrant. But more importantly, we find that the classroom makes a significant contribution to person aberrance and conclude that studies that investigate the sources of person aberrance with real data should model the classroom as well as individual levels. © 2007 by the National Council on Measurement in Education.