Author/Authors :
SÜNBÜL, Önder Mersin Üniversitesi - Eğitim Fakültesi - Eğitim Bilimleri Bölümü, Eğitimde Ölçme ve Değerlendirme Anabilim Dalı, Turkey , ERKUŞ, Adnan
Abstract :
In this study, item parameter invariance in constructs which have characteristics of various dimensionality was examined within the framework of Unidimensional Item Response Theory (UIRT).For this purpose, two-dimensional data simulation was realized depending on compensatory MIRT model. For the simulation, the number of the items in the first dimension (24) and the mean of a1 parameters (1.00) were kept constant in all constructs. Manipulations were applied to the items in the second dimension and correlation between dimensions. The numbers of the items in the second dimension were set to 4, 8, 12, 16, 20, 24; mean a2 parameters were set to 0.25, 0.50, 0.75,1.00, and correlations between dimensions were set to 0.00, 0.30, 0.60, 0.80. Simulation criteria were fully crossed and 96 experimental cells were obtained. Dichotomous responses of 90 000 subjects were simulated for each cell. By this way, 96 populations were established. In order to determine how parameter invariance is affected by sample size, 100 samples for 500 sample size, 100 samples for 1000 sample size and 100 samples for 3000 sample size were drawn from populations by using simple random sampling. Thus, 28800 samples were revealed. These samples were analyzed by using UIRT. In order to analyze the analysis results, average standard errors were used as the indicator of parameter invariance. Despite the fact that all datasets are treated as unidimensional for IRT analysis, avarage standard errors were computed separately for each dimension and how both dimensions have been affected by parameter invariance was determined in this manner.When the results for IRT were examined, parameter invariance were provided to a great extent for both a and b parameters but it cannot be provided for some constructs. Those constructs are empirical cells in which the first dimension is equivalent in terms of strength to the second dimension and in which the dimensions are orthogonal. Average standard errors which were computed for the particular empirical cells seem bigger than the acceptable level for especially b parameter. Empirical cells can produce invariance item parameters in spite of the violation of IRT’s unidimensionality assumption. As shown in literature, UIRT is quite robust against multidimensionality.
