OBJECTIVES

Claims databases represent a powerful resource to study routine clinical care with regard to a wide range of research questions. However, nationally or even internationally representative databases are rarely available, which might limit generalisability of results. As an example, statutory health insurance (SHI) in Germany is divided into >100 different funds, with insured populations that might differ in their characteristics from the overall German population. The aim of this study is to propose a methodology to increase representativeness of analyses of such data.

METHODS

We used a German claims dataset (AOK PLUS) and identified main characteristics of insured persons in that database alive on the 01/01/2017. Age, gender and occurrence of diabetes, as an exemplary comorbidity, were assessed for these patients. Subsequentially, respective characteristics of the overall German SHI population (publicly available statistics) were used to conduct a model-based standardization. Individuals in the AOK PLUS were re-weighted by a propensity score methodology, based on a logistic model. The propensity score was defined as the probability that an individual from the German SHI belongs to the specific dataset, given the investigated characteristics.

RESULTS

In total, the investigated dataset included 3,396,276 persons alive on the 01/01/2017. Their mean age was 45.5 years (25.2% aged ≥65 years, 51.8% female, 12.9% affected by diabetes). In the whole SHI population, proportion of persons aged ≥65 years was 21.9%; 52.1% were female and 9.3% had diabetes. After conducting the model-based standardization, the effective sample size was 3,343,829 persons. Following the adjustment, mean age was 43.5 years (21.9% aged ≥65 years, 52.1% females, diabetes prevalence 9.3%).

CONCLUSIONS

Weighting based on correct specified propensity score models can adjust imbalances in the distribution of characteristics between an analyzed claims sample and the target population. We recommend to use such an approach, as this increases the generalizability of results.