Abstract
The effects of violating the unidimensionality assumption when applying item response logistic models were studied. Using a multidimensional two parameter logistic model, two different tests that are common in practice were simulated: a) a test composed of two equally relevant dimensions, and b) a test with a dominant dimension and a secundary one. Each test was composed of 40 items, 25 corresponding to the first dimension, 15 to the second. Two sample sizes (N= 300 and N= 1000) and five levels of correlation between the dimensions (0.05, 0.30, 0.60, 0.90, 0.95) were used to generate the data. The unidimensional two-parameter logistic model was used to estimate the item parameters and the ability of the examinees. The results indicate that the unidimensional estimates are consistently robust. Estimates of the item difficulty parameter are less affected by the violation of the unidimensionality assumption than the other item parameter estimates. The item discrimination parameter and the ability estimates are influenced by the size of the correlation between the dimensions and by the type of multidimensionality displayed by the data.
