Discrimination, difficulty, and information
We ran a graded-response polytomous model (which is analogous to a 2PLM for dichotomous outcomes) on the 12 Dirty Dozen items while allowing items to load on their respective latent subscales (dimensions). The discrimination (a) and difficulty (bs) parameters from these models appear in Table 2.
Across samples, the 12 items had as ranging from 0.74 to 2.34, which suggested the items adequately discriminated among people along their respective latent traits.
The eight (Sample 1) or six (Sample 2) difficulty parameters for each item (bs) suggested that DTDD items tended to be ‘‘difficult’’ to endorse on average, perhaps because each trait is considered socially undesirable.
Narcissism items were easier to endorse, given the negative-to-positive crossover point in the difficulty parameters was near the midpoint of the latent variable; however, psychopathy and Machiavellianism items were more difficult to endorse across both samples.
The corresponding scale information curves (SICs) for each subscale are overlaid in Fig. 1. Because SICs are an additive function of the number of items contributing to a given scale, they are impressive given that each subscale has only four items. Table 3 shows the total information (area under the curve—AUC) for each SIC for each subscale, and the total information per item. Recall that information in IRT is related to reliability in CTT. This relationship was confirmed in both samples: mean interitem correlations (MICs in Table 2) were correlated with SIC AUCs from the three subscales, rs(1) = .999 and .947 in Sample 1 and 2, respectively.