Purchase this article with an account.
Micaela Gobeille, Christopher Bradley, Judith E Goldstein, Kyoko Fujiwara, Robert W Massof; Recalibration and fit statistics of the Activity Inventory item bank. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1934.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Visual function questionnaires are important tools to measure patient-reported outcomes for visually impaired patients. Rigorous estimation of these measures is critical, and Rasch analysis using the Andrich rating scale model has been the preferred method. This model produces disordered thresholds which require categories to be collapsed, decreasing precision. The Method of Successive Dichotimizations (MSD) overcomes this limitation by employing a Rasch difference model. Here, we apply MSD to Activity Inventory (AI) patient responses to recalibrate the existing item bank and refine analytics.
MSD is applied to a large AI database pooled from 5 studies (n=3739). Item measures and threshold values are calibrated and serve as anchors. Person measures and fit statistics for item and person measures are calculated at the task, goal, and domain (i.e. reading, mobility, visual information (VI), visual motor (VM)) levels.
Person infit frequency distribution is compared to the expected chi-square function based on true degrees of freedom for each person (i.e. number of item responses). Infit is platykurtic with greater variance in the right tail. Person infit z-scores are calculated(mean=0.16, median=0.32, range=-6.86 to 4.21). Most values fall within ±2 standard deviations from 0.Item infit z-scores are calculated(mean=1.98, median=1.14, range=-27.41 to 30.01) and compared to item measure and cumulative frequency. Goals and VI behave as predicted by MSD while reading overfits the model (less variance than expected) and VM and mobility underfit (more variance than expected).Across domains, goals and tasks correlate most strongly (r=0.75). Factor analysis with varimax rotation shows that two factors are necessary and sufficient to explain 64% of all variance. Reading loads more heavily on one factor while mobility loads more heavily on the other. VI, VM, and overall goals fall close to the principal axis. The amount of variance explained in each domain is: reading(48%), mobility(66%), VM(69%), and VI(72%).
Recalibration of the AI provides more accurate estimates of 510 item measures and 4 category thresholds to anchor the scale for adaptive testing of visually impaired patients. Consistent with past studies, person measures have a two dimensional structure. The calibrated AI is a powerful tool to measure outcomes centered on patient preferences with the same precision and accuracy for any sample size.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only