Rasch analysis of the Ocular Surface Disease Index (OSDI)

Abstract

Purpose: The Ocular Surface Disease Index (OSDI) is a 12-item scale for the assessment of symptoms related to dry eye disease and their effect on vision. Its reliability and validity have been investigated within the classical test theory framework and, more recently, using Rasch analysis. The purpose of the present analysis was to more completely investigate the functioning of its response category structure, the validity of its three subscales, and the unidimensionality of the latent construct it is intended to assess.

Methods: Responses to the OSDI from 172 females participating in the Dry Eye in Postmenopause (DEiM) study who had previously been diagnosed with dry eye or reported significant ocular irritation and dryness were analyzed. Response category structure and item fit statistics were evaluated for assessment of model fit. Person separation statistics were used to examine the validity of the subscales. Unidimensionality was assessed by principal component analysis (PCA) of model residuals.

Results: The recommended five-category response structure resulted in disordered response thresholds. A four-category structure resulted in ordered thresholds. Item infit statistics were acceptable for all 12 items. Person separation with this category structure was adequate, with a person separation index of 2.16. None of the three subscales demonstrated adequate person separation. PCA showed one other significant factor onto which the three environmental items loaded significantly.

Conclusions: All items demonstrated acceptable fit to the model after collapsing categories to order the response thresholds. The original subscales did not prove valid, and there is some evidence of multidimensionality and poor targeting.

Figure 1.

Category probability curves for the five-category instrument showing underutilization of category 3 and disordered thresholds. Curves for items 1–5, 6–9, and 10–12 are shown in (a), (b), and (c), respectively. Blue: none of the time; red: some of the time; green: half of the time; purple: most of the time; light blue: all the time.

Figure 2.

Category probability curves for the four-category instrument show ordered thresholds. Curves for items 1–5, 6–9, and 10–12 are shown in (a), (b), and (c), respectively. Blue: none of the time; red: some of the time; green: half of the time or most of the time; purple: all the time.

Figure 3.

Person-item chart, which demonstrates nonideal targeting. The x-axis is the measure in logits, and the y-axis is the number of subjects or items located at that measurement location. Ideally, the average person measure would be closer to the average item measure, which is set to 0. The subjects represented by the bars to the left of 0 are not well-targeted by the instrument.