Exploring the predictive power of interaction terms in a sophisticated risk equalization model using regression...

Area of Science:

• Health Economics
• Health Services Research
• Biostatistics

Background:

• The Dutch risk equalization (RE) model aims to predict healthcare expenses.
• The 2014 RE model is complex, with millions of potential interaction terms between risk adjusters.
• Existing models may not fully capture expense variations due to unaddressed interactions.

Purpose of the Study:

• To explore the predictive power of interaction terms in the 2014 Dutch RE model.
• To identify statistically significant interaction terms that explain residual variation in healthcare expenses.
• To assess the impact of incorporating these interaction terms on expense prediction and risk selection incentives.

Main Methods:

• Utilized regression tree modeling, a novel approach in RE research, to identify significant interaction terms.
• Analyzed a large dataset of approximately 16.7 million individuals.
• Integrated identified interaction terms as additional risk adjusters into the RE model.

Main Results:

• Interaction terms were found to significantly explain variation in observed expenses beyond existing risk adjusters.
• Incorporating interaction terms improved overall expense prediction and prediction for specific population subgroups.
• However, expense prediction deteriorated for certain selective groups, potentially altering risk selection incentives.

Conclusions:

• Interaction terms offer potential for enhancing the accuracy of risk equalization models.
• The impact of interactions on risk selection incentives is mixed, reducing them for some groups while increasing them for others.
• Regression tree models require additional criteria (e.g., incentive structure, expert opinion) for robust selection of interaction terms for practical implementation.