Innovations In Clinical Neuroscience

NOV-DEC 2017

A peer-reviewed, evidence-based journal for clinicians in the field of neuroscience

Issue link:

Contents of this Issue


Page 44 of 83

45 ICNS INNOVATIONS IN CLINICAL NEUROSCIENCE November-December 2017 • Volume 14 • Number 11–12 O R I G I N A L R E S E A R C H Finally, we also used the confirmatory bifactor solution to compute two important indices. The first is the explained common variance per item (ECVI), which is simply an item's squared loading on the general factor squared divided by the communality. ECVI values from 0.50 to 1.0 indicate that the item is a more "pure" univocal measure of the general factor (symptom severity here), and ECVI values less than 0.50 indicate that the item is a relatively better measure of a specific dimension. We also computed two model- based reliability coefficients, omega ω and omega hierarchical ω H . 33 ω values indicate the degree to which observed scores reflect all reliable sources of common variance (i.e., the general factor and the five specific factors). ω H reflects the degree to which variance in total scores reflects the general symptom severity factor. As ω H values approach 1, total scores are unambiguous indicators of relative standing on the common dimension, uncontaminated by specific dimensions. The difference in ω and ω H indicate the degree to which reliable variance is contaminated by the multidimensionality of the items. Estimating IRT bifactor model parameters. A bifactor IRT model was estimated using the mirt 34 library in R. The model specified one general factor and five specific domain factors, where each item was allowed to load on the general and only a single specific factor. For each item, the model estimated five discrimination parameters per item (one for the general, and five for the specific dimensions), and four intercept parameters (one for each between category boundary). These estimated parameters are called "conditional" parameters and reflect the relation between the item and each latent trait conditional on the other traits being zero (i.e., at the mean of the other dimensions). As such they are difficult to meaningfully interpret. 14 We thus transformed the conditional IRT parameters into so-called "marginal" parameters using the formula provided by Toland et al. 35 These marginal IRT parameters better reflect the relation between trait standing and the item responses. Finally, the marginal IRT parameters were used to construct a pseudo-IRT unidimensional model that included only the general factor (symptom severity) and excluded the specific factors using methods described by Toland et al. 35 In this final model, each item has a single discrimination parameter and five thresholds. In turn, this final model was used to derive CRCs and IICs for each item, as well as other derived indices. To judge the psychometric qualities of the remission set, two analyses were performed. First, we computed the remission status for each subject using the criterion defined for the Remission set (i.e., item scores ≤3). We then used the final IRT model to estimate each individual's standing on the latent trait using expected a posteriori scoring (EAP). 36 We then compared the distribution of symptom severity scores for the judged remitted versus non- remitted groups. Second, we computed TICs based on just the eight Remission set items to discern how discriminating this item set is with respect to symptom severity and where along the latent trait the Remission set provides the best discrimination. For comparative purposes, we derived TICs for three alternative eight-item sets and compared them to the Remission set. Specifically, we formed TICs based on eight items that A) had the highest IRT discrimination parameter (i.e., the most discriminating items), B) had the highest ECVI (i.e., the most univocal items), and C) provided the most information in the low trait/ symptom relief range, where low was judged as trait standing of θ=( -1,-4). This trait level value was selected based on the item rating anchors of the PANSS (1=absent, 2=minimal, 3=mild, 4=moderate, 5=moderate/severe, 6=severe, 7=extreme). The theta value was selected based on inspection of an expected average item response that was based on linking the item score metric to the latent trait metric using a test response curve, which is basically the weighted sum of CRCs divided by 30 (items). This curve is shown in Figure 3. We used this curve to discern the value of the latent variable that predicts an item score of 2 (minimal) or less. At trait level= -1, the expected item score across all 30 items is 1.83, which lies between the score anchors of either "Absent=1" or "Minimal=2" on the PANSS scale. FIGURE 3. Expected item score as a function of symptom severity level—Setting item thresholds ≤2 typically map to the trait level below a group-mean.

Articles in this issue

Archives of this issue

view archives of Innovations In Clinical Neuroscience - NOV-DEC 2017