If You Build It, Will They Come? Predicting Consumer Responses to Insurance Benefit Designs

In 2012, the California Public Employees’ Retirement System (CalPERS) implemented a reference pricing program for colonoscopy outpatient surgical services. The program uses non-linear cost-sharing to incentivize consumers to receive care from less expensive providers. Consumers who receive a colonoscopy at a hospital outpatient department (HOPD) with a price above $1,500 are responsible for the full marginal costs above $1,500. Patients who receive care at a less expensive HOPD or an ambulatory surgical center (ASC) are only responsible for standard cost-sharing. Using control group data from a non-CalPERS population within the same insurer, previous work demonstrates that the program leads to a 15% shift in consumer demand from high-priced to low-priced providers and a 21% reduction in procedure costs (Robinson et al, 2015).

This paper estimates the potential effects if other markets implemented reference pricing programs. We start by estimating the heterogeneous treatment effects by California market and then use a machine learning model to identify the underlying market characteristics that drive the heterogeneity in treatment effects.

We then use nationwide medical claims data from the Health Care Cost Institute (HCCI), along with companies providing data to it—Aetna, Humana, and UnitedHealthcare. From the HCCI data, we identify 3.3 million colonoscopies over the 2009-2013 period. We first use the HCCI data to create market-level price indices used in the machine learning model. We find substantial variation in both within and between market variation in colonoscopy prices and market structure. We then apply the predicted parameters to the nationwide data sample to predict the potential effects of reference pricing programs in each market.

The machine learning model identifies market-level prices and price variation, gastroenterologist concentration, and ASC penetration as the largest predictors of the heterogeneity in savings due to reference pricing. We estimate that if other the program was implemented nationwide, then the per-procedure spending on colonoscopies would decrease by approximately 7%. The per-procedure savings translate to an approximately $100 million reduction in medical spending if applied to all colonoscopies in the HCCI data. Using a particular example of reference pricing for colonoscopy procedures, this paper demonstrates how machine learning methods can be used to rigorously make out-of-sample predictions.