This thesis focuses on including modern statistical machine learning methods into the analytical tools used for health insurance. To do so, a study implemented by the company Munich Health, dating back to 2013 was chosen to be reworked. This study had examined the profitability of disease management programs (DMP) in which chronically ill patients are supported with a special treatment. To eliminate any biases, a creatively enhanced approach is necessary to come up with an appropriate control group. For that purpose, in the thesis, the matched pair approach is used for selecting the patients from a potential control group as the most similar to the DMP patients. The matched pair method relies on having a comparable criterion for each participant and non-participant. In the original study, the authors used predicted costs, obtained by a (log)-linear regression and factor analysis. These methods are reworked, extended and additionally replaced by machine learning methods. In particular, penalty regression, linear mixed regression, tree boosting, different approaches with neural networks and K-mean clustering are used. Furthermore, the original study is extended with five years of new, additional data, which facilitates an update to the old results. It is shown that the prediction of costs with machine learning is much superior to the original methods. It leads to more precise and more stable results, in terms of robustness for outliers and the number of assumptions to the model.     «

This thesis focuses on including modern statistical machine learning methods into the analytical tools used for health insurance. To do so, a study implemented by the company Munich Health, dating back to 2013 was chosen to be reworked. This study had examined the profitability of disease management programs (DMP) in which chronically ill patients are supported with a special treatment. To eliminate any biases, a creatively enhanced approach is necessary to come up with an appropriate control gr...     »