This was an observational study. The owners of the data refused to give us the data so we could not proceed with the analyses.

Several employers in the US have introduced a program where their employees receive a financial incentive to receive lower cost care. Under this "Rewards" program, patients are free to choose providers but if they visit a pre-determined low-cost laboratory or radiology facility (called a "rewards provider"), they receive a financial incentive. The financial incentive is typically in the form of a Health Savings Account (HSA) contribution. The dollar amount varies by employer. This study will use medical claims data to examine if this program leads to an increase in the volume of services performed by low-cost providers and decreased health care spending.

This will be an observational study using Differences in Differences and Regression Discontinuity designs.

In designing our analytic methods for this observation study, we had to consider two key potential sources of bias. First, rewards providers might differ from non-rewards providers in observed and unobserved ways such as such as quality or convenience. This might confound the true effect of the rewards program on service volume. That is, an analysis of the rewards program that simply compares service volume of rewards and non-rewards providers after launch of the rewards program will capture both the effects of the rewards program as well as the effects of other differences between rewards and non-rewards providers. Second, service volume of rewards program might change due to other factors coincident with the launch of the rewards program. Our proposed statistical methods attempt to address both these sources of confounding. We will examine the data in a series of way to test the robustness of our findings.

Difference-in-differences (DD) linear regression We will use a difference-in-differences regression to examine within-employer changes in provider utilization following the implementation of the rewards program.

In the first set of regressions we will use data from employers who have implemented the rewards program. These regressions will compare changes in provider volume following the launch of the rewards program for rewards providers (first difference) to change in service volume for non-rewards providers during the same time period. We hypothesize that rewards providers will experience a greater increase in volume than non-rewards providers. This analysis uses non-rewards providers from the same employer as a control group and assumes that reward providers would have experienced the same change in volume as non-reward providers in the absence of the rewards program.

In another set of regressions we will use data from both employers who have launched the rewards program and employers who have not launched the rewards program. This analysis will be only feasible for providers who have a unique id across employers and who see patients from both rewards and non-rewards employers. These regressions will compare changes in provider volume for rewards providers coming from rewards employers (first difference) to changes in service volume for the same reward providers coming from non-reward employers during the same time period. We hypothesize that rewards providers will experience a greater increase in volume coming from rewards employers than non-rewards employers. This analysis uses non-reward employers as the control group and assumes that reward providers would have experienced similar increase in service volume from rewards and non-rewards employers in the absence of the rewards program. This analysis helps to address the potential bias that rewards providers had an increase in volume because of other factors such as quality or convenience.

Regression discontinuity regressions Our second study design is a regression discontinuity design. Providers are designated as rewards providers based on their relative cost within a geographic market. Providers are ranked based on an index of prices and providers below a pre-specified ranking or threshold on this index are designated as rewards providers. The regression discontinuity model will compare changes in volumes between providers that are just above this threshold with providers that are just below this threshold.

Instrumental Variables Analysis Our third study design uses an instrumental variable analysis. The dollar amount of the financial incentive varies between the employers who have introduced the program. In this analysis we will exploit that difference. We will assess whether the effects of rewards on service volume vary by the size and nature of the rewards.

Independent variables As independent variables, we will use employer, month, year, geography, year X geography, and provider fixed effects. If there are changes in the employee population before and after the introduction of the Rewards program, we will control for those differences.

Inclusion Criteria:

- Employee or dependent of an intervention or control employer

Exclusion criteria:

- Not continuously enrolled in health plan and therefore some claims may be missing

The study population is composed of employees of both the intervention and control employers. We will use medical health plan claims from 2012 to 2014 for laboratory and imaging services across the study population. Health plan claims data will be provided by Castlight Health. The approximate number of providers of laboratory and radiology services who will provide care to this population will be 475,000.