Public Health Brown Bag Seminar - Xander Koolman
|When:||Th 17-10-2019 12:00 - 13:00|
|Where:||De Skibril, Duisenberg, Zernike|
Xander Koolman is a health economist with a PhD in health econometrics applied to the measurement of health inequality, and inequity in health care use. Following his PhD he specialized in the measurement of health care provider performance and started a consulting firm SiRM (now SiRM and Equalis). He continued to work in academia where he set up the Talma Institute, an interdisciplinary health care research institute. His research interest now includes affordability of care, which completes the public health system aims: quality, accessibility and affordability. From 2017 onwards Xander chairs the Health Economics section in the Faculty of Science at the Vrije Universiteit Amsterdam and is a member of advisory board of the Dutch Health Care Authority (NZa).
Risk equalization (RE) models are developed to mitigate risk selection incentives in regulated health insurance markets. Inaccurate financial compensation for insurers still leaves incentives for risk selection however. Machine learning may improve on OLS regression-based RE models through the automatized modeling of interaction effects. This study investigates whether Random Forests (RFs) and Gradient Boosted Machines (GBMs) improve the prediction of the somatic medical expenses of plan holders compared to the Dutch RE model of 2018, while holding the set of risk classes constant. First, we used population-wide data (n ≈ 17 million) to simulate the Dutch RE model using OLS regression. Second, we developed two alternative models using the RF and GBM algorithms. The performance of these models was evaluated using the R-squared, Cummings Prediction Measure (CPM) and the Mean Absolute Predictive Error (MAPE) on an individual level and the Mean Equalization Result (MER) on a subgroup level. On an individual level, the RF outperformed the OLS regression model on all measures (R-squared, CPM and MAPE), while the GBM only improved the R-squared. On a subgroup level, the RF reduced the mean under- or overcompensation – as expressed by the MER – for all but the first two population deciles, while the GBM improved the MER specifically in the deciles with relatively low or high expenses. Moreover, the GBM significantly reduced the mean overcompensation of individuals with low expenses in the past (year t-3) from €110 to €3.5 compared to the OLS model. We conclude that machine learning can indeed improve on the performance of the Dutch RE model and that this model may significantly benefit from the incorporation of additional interaction terms.