Small area estimation of regional diabetes prevalence from biased health insurance data


Poster Session

Poster no.04

Jan Pablo Burgard (University of Trier, Faculty IV - Economics, Economics and Social Statistics Department)
Joscha Krause (University of Trier) (Speaker)
Ralf Münnich (University of Trier)

Medical studies suggest that assessing disease distributions on regional levels is very important for planning and providing healthcare programmes. Data from national health surveys is frequently used in this context to obtain regional prevalence estimates. But these surveys often lack sufficient local observations due to limited ressources. Subsequently, regional prevalence estimates can be unreliable because of large variance.

Health insurance companies represent alternative data sources on this behalf. Their data encloses large insurance populations with rich health-related information, even at regional levels. It allows the researcher to analyze company-specific disease distributions in very high resolution. But a major issue occurs when extrapolating company-specific prevalence patterns to the total population. Health insurance membership is often informative for disease risk due to national insurance market regulations. The insurance population of an individual company cannot be treated as a random sample of the total population. Accordingly, prevalence estimates from individual insurance data may be heavily biased.

We propose a small area approach to simultaneously reduce the company-specific bias and quantify regional disease distributions in high resolution. This is achieved by combining health insurance data and administrative information within an area-level model. The structural health differences between the populations are quantified, extracted and used to predict regional prevalences of some disease of interest. An empirical application of our methodology is provided on the example of diabetes mellitus type 2 in Germany. We use data of the German Public Health Insurance Company to predict cohort-referenced diabetes prevalences for all German districts. The results suggest that our approach can reduce the bias associated with informativity and provide significant efficiency gains