Item response theory analyses of physical functioning items in the medical outcomes study

Constructing and evaluating item pools that measure a single domain of health-related quality of life (HRQOL) is one of the fundamental objectives of the Patient Reported Outcome Measurement and Information System (PROMIS) project. The initial focus of PROMIS is on 5 HRQOL domains: physical function, fatigue, pain, emotional distress, and social/role participation. Analysis of related, available data can inform construction of these new banks.

We sought to inform the building of the PROMIS physical function item bank by evaluating physical functioning items administered to participants in the Medical Outcomes Study.

Secondary analyses of physical functioning items administered at baseline (Fall, 1986) to 3223 individuals in the Medical Outcomes Study.

The mean age of participants was 54 years old; 15% had less than a high school education, 30% had a high school education, 27% had some but <4 years of college, and 28% had 4 or more years of college. Sixty-one percent were female; 80% were non-Hispanic white, 14% non-Hispanic black, 3% Hispanic, 1% Asian, and 2% other race/ethnicity.

Fifteen physical functioning items were selected for analyses including the 10 SF-36 physical functioning items, an item assessing limitations in everyday physical activities because of health problems, an item measuring satisfaction with physical ability, and 3 mobility items.

Cronbach's alpha was 0.94 for the 15-item scale and item-total correlations ranged from 0.51 (limited in bathing or dressing) to 0.80 (limited walking several blocks). A one-factor confirmatory categorical model fit the data well according to practical fit indices (comparative fit index = 0.95), but the Root Mean Square Error of Approximation was 0.12. Estimated factor loadings were all large (0.72 to 0.96) and statistically significant. We fit the graded response model with freely estimated and fixed slope estimates to compare 2-PL and 1-PL models. The difference between these 2 goodness of fit statistics for the 2 models was statistically significant: chi (df = 14) = 772.9, P < 0.001. Slope estimates for some of the items in the 2-PL model varied from the 1-PL common slope to a noteworthy extent (eg, items 2, 5, 7, 8, and 9). Category threshold estimates generally were similar for the 2 models. Threshold estimates for the 2-PL model ranged from -2.90 (item 14: between no, never, and yes, occasionally in bed or in a chair most or all of the day because of health) to 1.55 (item 12: between very satisfied and completely satisfied with physical ability to do what you want to do). Information peaked at -1.0 theta (information = 33.60) for the 2-PL model. Information was only 7.43 at 0.8 theta (2-PL) and decreased with increasing theta.

The 15 physical functioning items provide reasonably good fit to a unidimensional item response theory model that provides satisfactory coverage of the lower levels of physical functioning but does not provide very much information at higher levels of functioning.