Data from existing administrative, clinical, and Y sources will be evaluated. A dedicated electronic tracking and billing database, developed by UHG to help the Y administer the DPP, will allow us to analyze attendance and weight loss for program participants. Medical, pharmacy, and laboratory claims, available for all UHG enrollees, will allow us to compare changes in total and sector-specific (eg, inpatient) health care expenditures among different groups of enrollees regardless of DPP participation. Pharmacy claims will enable us to assess changes in treatment intensity for conditions linked to obesity and high metabolic risk (high blood pressure, high cholesterol, diabetes) (14). For subgroups of UHG enrollees with available test results associated with laboratory claims (approximately 20% of claims submitted nationally), we will evaluate changes in total cholesterol, low-density lipoprotein cholesterol, and hemoglobin A1c levels. Finally, data provided by the Y will enable us to explore whether regional variation in the structure and process of program implementation can help explain geographic differences in key program outcomes.

One strategy used by UHG to identify prediabetes and to enroll high-risk adults in a Y-based DPP intervention involves large-scale blood glucose or hemoglobin A1c testing as part of a workplace wellness or risk assessment initiative. Because these initiatives are deployed at an employer level, we have designed a cluster-randomized encouragement trial (CRET) as an innovative component of our research design (15,16). The CRET design will allow for comparisons with a randomized control group without interfering with the natural implementation of the program.

In select regions, UHG will provide the research team with a list of large employers whom it would aim to engage prospectively for on-site testing. We will randomly assign these employers to sequential waves of outreach and release them back to UHG to initiate encouragement procedures at a rate that matches the capacity of the UHG outreach team. As some employers are released into the active encouragement arm, others will remain in abeyance to serve as controls until being released at a future date (Figure). By using an intent-to-treat framework, clients who are actively encouraged (and far more likely to participate in the DPP) can be compared with clients who are not encouraged until a later time (17).

Although people cannot be randomized to attending the DPP, the CRET design will also enable us to analyze the treatment response among those who do elect to participate (18). Typically, comparisons of such self-selected subgroups can introduce selection bias. However, through the use of instrumental variables analysis (18–21), the CRET design will enable us to use the status of random assignment (ie, encouragement now vs encouragement later) as an instrumental variable to construct a robust estimate of the treatment effect of the program while minimizing the threat of selection bias.

The evaluation will focus on comparable samples of high-risk UHG clients who have prior claims-based evidence of diagnosed prediabetes (International Classification of Diseases, Ninth Revision, Clinical Modification [ICD-9-CM] code 790.21 or 790.22) or other indications of high metabolic risk (eg, diagnosed metabolic syndrome [ICD-9-CM code 277.7] or multiple metabolic traits such as overweight and obesity; high blood pressure; abnormal blood cholesterol). In this sample, the study will compare differences between randomized groups in the changes in laboratory tests for blood cholesterol and hemoglobin A1c, medication treatment intensities, and patterns and overall expenditures for health care. Total per-person-per-month (PPPM) health care expenditures will be compared by using 24 months of baseline and 12 month of follow-up data. We expect the cluster-randomization to yield a sample of approximately 60 employers and a minimum of 12,500 high-risk employees in each arm. Under reasonable assumptions that fewer than 30% of the observations may be missing (ie, as clients withdraw from the health plan) and that the intracluster correlation coefficient (ie, within employer) will be no more than 3%, we should have more than 80% power to detect mean differences in total health care costs as small as $300 during the 12 months of follow-up.

One possible challenge of using a CRET design is that it may prove more powerful for evaluating the effects of UHG’s encouragement efforts than for evaluating direct effects of exposure to the DPP. If, for example, the number of employers randomized to the “active encouragement” arm exceeds the capacity of UHG to successfully engage them to identify high-risk employees, then the overall “dose” of DPP exposure in the active arm will be lower than anticipated, and statistical power of the study could be reduced.