377 – Contributed Oral Poster Presentations: Section on Statistical Learning and Data Mining

Are Provider Communication Constructs and Structural Relationships the Same Accross English and Spanish?

at American Board of Internal Medicine

We compare patient ratings of primary-care physicians on three composite measures related to provider communication skills (6 items), shared decision making (2 items), and overall satisfaction (11-point rating scale). Measures are adapted from practice-based Clinician and Group (CG) CAHPS® Surveys in English or Spanish; surveys are completed anonymously (minimum of 25 responses) using an automated phone or Web system. These surveys satisfy self-evaluation of practice performance requirements for physicians engaged in maintaining board certification in internal medicine. Responses may differ among patients because measurement constructs may not translate well between languages. If latent constructs are interpreted by Spanish- and English-speaking patients similarly, then covariance structures for the same physicians would be indistinguishable after controlling for differences unrelated to language (e.g., health status, education). We compared patient ratings for 79 physicians who administered surveys to both Spanish- (301 cases) and English- (1,708 controls) speaking patients. A second control group included 2,048 ratings (English only) of another 79 physicians who were propensity-matched to each study physician based on 31 practice, demographic, training, and achievement characteristics and 9 patient demographic variables. Groups were divided into 49 training pairs and 30 testing pairs for validation. A three-group, MIMIC, structural equation model was adapted to the Spanish cases and the two English control groups. Covariance and mean structures were not equivalent (p<0.001), and a series of hypothesis tests suggests that language factor structures were the same for Spanish and English patients (p=.061), factor loadings for items differed (p=.003), and differences in propensity score regressions between groups were not due to selection alone (p<.001). This study shows that patients' ratings on three provider constructs had similar structural relationships among latent constructs but showed differential item functioning by language. There was a significant propensity score interaction with group membership. The study suggests that structural equation models may be useful in assessing whether different language groups vary with respect to what constitutes good communication skills in their health care providers, the quality of the items used to measure constructs, and whether propensity score matching works uniformly across groups.