DescriptionUnidimensionality -- a condition, under which only one dominant construct is being measured by the test, is a fundamental assumption of most modern day psychometric models. However, some tests are multidimensional by design. A test, for instance, might measure physics, biology and chemistry subscales combined to measure a "general science" composite. The relative magnitudes of those subscales sometimes shift from administration to administration, which results in an altered composite. This study examined the conditions under which two different forms of a multidimensional test measure the same composite construct to a degree that allows them to be equated, i.e. used interchangeably.
IRT true-score equating was used in a simulation study to assess the closeness of the scores on the forms. Conditions examined included the correlations between subscales, varying number of items per subscale form to form, and different subpopulation ability estimates on the subscales. Differences in the equating errors due to generating model (1PL or 3PL) were also examined. A way of calculating a unidimensional composite from a two-dimensional ability was devised and compared to the unidimensional composite obtained from Parscale.
It was found that in general, the errors increase with decreasing correlation between traits and increased divergence of the two forms to be equated, with the later being the main predictor of the equating errors. However, the magnitude of those errors was small for the population as a whole especially when all examinee abilities are drawn from the same distribution. It was concluded that IRT true score equating is relatively robust to multidimensionality for the conditions examined, especially if the overall population score is desired. However, when accurate estimate of the equated score for individuals at the extremes of the population is needed, or whenever population abilities are drawn from more than one distribution, the unidimensional true score equating functions well only for very similar forms and with high correlations between traits.