Gessaroli, M.,Breithaupt, Krista.2009-03-252009-03-2519951995Source: Masters Abstracts International, Volume: 34-05, page: 1737.9780612078376http://hdl.handle.net/10393/9704http://dx.doi.org/10.20381/ruor-7926The purpose of this study was to compare two methods to assess the dimensionality underlying a set of responses to test items; Stout's T statistic (Stout, 1987) and a form of non-linear factor analysis, the approximate $\chi\sp2$ statistic (De Champlain, 1992). Response data was simulated to compare the number of rejections of unidimensionality made using the approximate $\chi\sp2$ statistic and using Stout's T statistic. Two conditions have been identified as possibly affecting the accuracy of dimensionality assessment techniques; the degree of correlation among abilities and the amount of guessing present in the data. These two variables were manipulated in the generation of 100 different data sets, each with scores for 1000 examinees under each experimental condition for both two-dimensional and unidimensional response data. The dimensionality underlying each data set was assessed by the program NOHARM II (Fraser & McDonald, 1988) for the approximate $\chi\sp2$ and by DIMTEST (Stout, Douglas, Junker & Roussos, 1993) fot Stout's T statistic. Both statistics had higher Type I error rates when unidimensional data was simulated with pseudo-guessing present. As has been found in earlier studies, longer tests were also associated with lower Type I error rates for these statistics in the unidimensional data condition. Both statistics performed with acceptable levels of Type I errors (less than 5% in most conditions). Logit-linear analysis of the number of rejections of unidimensionality made by both statistics showed main effects of pseudo-guessing, test length and the level of correlation of the latent factors in the two-dimensional simulation condition. Longer tests had a greater number of (correct) rejections of unidimensionality. Rejections of unidimensionality were less frequent when the latent traits were highly correlated (r =.7). Both statistics performed well when latent traits were correlated at r =.5. There were more correct rejections of unidimensionality made with response data which contained no pseudo-guessing. (Abstract shortened by UMI.)50 p.Education, Educational Psychology.Thesis