Application of stochastic measurement models to visual function rating scale questionnaires

Purpose: To test hypotheses that low vision patient responses to visual function rating scale questionnaires conform to an additive conjoint structure and that the Likert score is a sufficient statistic for the latent patient trait; to compare results for two competing stochastic measurement models; and to determine if different questionnaires measure the same construct in low vision patients. Methods: Visual function rating scale questionnaires were administered to 284 low vision subjects by telephone. Each subject was administered two of four questionnaires: ADVS, NEI VFQ-25 plus supplement, expanded VAQ, and VF-14. Results: Data were analyzed with the Muraki item response model and the Andrich measurement model. The estimates of latent person, item, and response threshold measures from the two models are linearly related. The Muraki model produced a better overall fit to the item response data, the Andrich model produced a better fit to the average ratings for each person and item. Fit statistics for the Andrich model were proportional to the item-dependent discrimination parameter in the Muraki model. The ADVS was the most accurate measure and the NEI VFQ was the least. Reliability was similar for all four instruments. Person measures for each pair of instruments were linearly related indicating that all four instruments measured the same construct. The person measure estimate from the Andrich model is monotonic with the average rating. That relationship suggests a transformation of the Likert score that can correct the floor and ceiling effects in rating scale data. Conclusions: Patient responses to all four questionnaires conform to varying degrees to an additive conjoint structure. The Likert score is a sufficient statistic for the ADVS and the VAQ, but not for the NEI VFQ_or VF-14. All four instruments measure the same construct in the low vision population, but they differ in measurement accuracy and precision.