Modeling determinants of satisfaction with health care in youth with inflammatory bowel disease part 2: semiparametric distributional regression.

Abstract

Background: Methodological challenges arise with the analysis of patient satisfaction as a measure of health care quality. One of them is the necessity to adjust for differences in patient characteristics or other variables. A combination of several helpful extensions to regression analysis is shown based on patients with inflammatory bowel disease (IBD) to help identify important covariates associated with the distribution of satisfaction.Patients and methods: Analyses were based on cross-sectional data from a postal survey on the health care of patients with IBD aged 15–25, with satisfaction assessed using a 32-item validated questionnaire weighing experience by perceived relevance. The weighted summary score was modeled using a Beta distribution in a generalized additive model for location, scale and shape. Covariates were distinguished in 3 groups and the model was entered in separate, consecutive analyses. First, demographic and disease-related variables were included. Next, information about the IBD specialist was added. The third step added care quality indicators. Results are presented as OR with 95% CI.Results: In the survey, 619 questionnaires were returned and the data set had 453 complete cases for analysis. Satisfaction appeared increased for patients working (OR 1.59, 95% CI: 1.19–2.11) or studying (1.25, 1.00–1.56) as compared to those still at school or in non-academic job training. High anxiety scores and an older age of onset were associated with lower satisfaction. The variation of satisfaction is higher for patients with Crohn’s disease or who have statutory insurance (1.19, 1.01–1.40 and 1.22, 1.06–1.40).Conclusion: Modeling the entire distribution of the response uncovered additional influences on the variance of patient satisfaction not previously identified by classical regression. It also resulted in a richer model for the mean. The construction of a combined model for different features of the distribution also helped to improve the control of confounding.