The primary study groups were employees who participated in either the Weight Management (WM) or the Weight Management Plus (WM+) study arm as part of the Steps to Health study.(27) Obese (body mass index ≥30) employees at Duke were randomized 1:1 to the two weight management programs. The WM program is primarily educational, while WM+ also incorporates behavioral coaching sessions. Both programs were delivered to the participants by trained health coaches. The programs were each approximately 12 months in length, but the number and method of contacts between the two programs varied. In WM employees received a face-to-face meeting with their coach in the first month, had follow-up coaching calls in months 6 and 12, and received monthly mailings of health education materials. In WM+ employees received monthly counseling sessions (face-to-face in months 1, 4, 8, and 12, and the rest via telephone), as well as meetings with an exercise physiologist in months 2 and 5. They also received quarterly biometric feedback and targeted health education materials. Trained staff measured heights and weights using a Seca portable stadiometer (baseline only) and a Tanita BWB-800 scale (at all time points), and these measurements were used to calculate body mass index (BMI). To be eligible for the current analysis, participants also must have been enrolled in one of the health insurance programs offered through Duke for a minimum of 9 months in the 12 months prior to randomization and for a minimum of one month following their scheduled completion of the study intervention.

The study design also included a third group: an observational control group of employees who completed a health risk appraisal (HRA) between the dates of the first and last randomization of study participants (January 2011 and July 2012) with a BMI ≥30, but who did not participate in one of the weight management programs.

Employees were excluded from these analyses if they had bariatric surgery, an organ transplant, end stage renal disease, HIV/AIDS, or hemophilia.(29–31)

The overall study design is illustrated in Figure 1.

In addition to data collected specifically through the Steps to Health Study, most data used for the current analyses were obtained through the Duke Health and Safety Surveillance System (DHSSS).(32) The extracted DHSSS data included all medical claims (inpatient and outpatient), pharmacy claims, and employee demographic and work-related variables.

Health care utilization measures include mean monthly rates for emergency room visits, hospital admissions, and days of stay in the hospital. Utilization rates are expressed here as mean events per 1000 months of health plan participation. For each study period (pre- and post-intervention) individual utilization rates were calculated as the sum of events, divided by the number of months of insurance plan participation during the time period.

Analyses included all claims and costs incurred including costs paid by the insurance plan as well as deductibles and copays incurred by the two study group participants and the observational controls. Claims for illnesses unlikely to benefit from study participation during the study time window (i.e. s, fertility treatments,, trauma, and burns) were excluded based on Healthcare Cost and Utilization Project (HCUP) Clinical Classification System (CCS) diagnosis groupings.(29) All costs were adjusted to 2013 dollars using the BLS Medical Care Consumer Price Index (CPI).(33) Categories of expenditures included mean monthly inpatient and outpatient medical inpatient costs, pharmacy costs, and total costs (pharmacy plus medical). Medical costs were further categorized by Major Diagnostic Category (MDC) and drug costs were further categorized by general therapeutic class based on the National Drug Code (NDC) using Redbook classifications. Classification procedures and the choice of MDC and drug categories for these sub-analyses were based on significant associations with BMI observed in our prior work.(8)

Medical and pharmacy claims during the 12 months of insurance coverage prior to the index date (randomization date for study subject or HRA date for controls) were used to calculate several measures of disease prevalence and co-morbidity. Individual prevalent health conditions at baseline commonly associated with obesity (coronary artery disease, congestive heart failure, diabetes, hypertension, depression, and peripheral vascular disease) were identified using the HCUP CCS(29), Elixhauser comorbid conditions(34), or the National Committee for Quality Assurance (NCQA) HEDIS classification system.(35, 36)

Two baseline summary measures of comorbidity were calculated: the combined Chronic Illness and Disability Payment System (CDPS) and the Medicaid Rx system (MRX)(37), and the Elixhauser comorbidity measure.(34, 38)

To compare the intervention groups to those not taking part in the interventions, we used propensity adjustment to balance risk factors and covariates.(39) Propensity adjustment using a regression model is designed to simulate results that might be obtained from a randomized control trial,(40) the assumption is that by conditioning on the propensity score systematic differences between treated and untreated subjects are eliminated. The propensity score includes demographic variables (age, gender, and race/ethnicity), health factors (BMI, smoking status), occupational group, and comorbidity scores. For a full list of the included variables in the propensity score see appendix 1.

Covariate balance achieved using the propensity score was assessed by calculating the standardized differences between the intervention arm (WM and WM+) and the controls, for each covariate.(41) We considered absolute standardized differences less than 0.10 as an indication that adequate covariate balance was achieved.(42)

Descriptive analyses were used to compare baseline characteristics of study participants and observational controls. Continuous variables were compared by ANOVA, and any variables that departed significantly from normal distribution were compared using the Wilcoxon rank-sum test. Dichotomous or categorical variables were compared using the chi-square test of general association.

We calculated mean utilization and costs by study arm and time period as an initial comparison (using an intent-to-treat approach). Effects of study participation on mean monthly health care utilization and costs relative to the observational controls were estimated using a difference-in-difference regression approach. Our initial model compared those randomized to the WM+ or WM study arms and did not include a propensity score as this was a randomized trial. The comparison of the combined study arm (WM and WM+) to the observational controls used similar models with propensity score adjustment for baseline covariates.

The parameter for the interaction of study arm with follow-up period is the desired difference-in-difference parameter estimate. We used a repeated measures negative binomial model with a log link function based our prior analyses of obesity related costs among this employee population.(28) In order to account for potentially correlated outcome data, all models were based on use of generalized estimated equations (GEE) assuming an exchangeable correlation structure, also based on our prior analyses.

We also compared the pre and post-intervention utilization and costs among WM+ and WM participants based on change in BMI during the intervention. Individuals in the WM+ or WM were classified as those losing weight and those not losing or even gaining weight during the intervention. Effects of weight loss on mean monthly health care utilization and costs relative to the observational controls were estimated using an approach similar to that used for the primary analyses. These analyses employed repeated measures negative binomial models with a log link function with direct adjustment for age, gender, race/ethnicity, smoking status at baseline, BMI at baseline, and baseline comorbidities. Some cost and utilization categories were combined for these sub-analyses for greater model stability.

All statistical analyses were conducted using SAS 9.3.(43)

The strengths of the Steps to Health study include the relatively large sample size, the racial diversity of the sample, and the use of a randomized control study design. Although all data are from this one employer, this employer is the second largest in North Carolina, and the results should be broadly generalizable to the adult working population given the breadth and diversity of the workers in terms of gender, age, and race/ethnicity. The use of both medical and pharmacy claims data further strengthens the findings. Another strength of this study is that it evaluated two established programs in the workplace that had support from management.

While the study attempted to capture a broad spectrum of factors that may evidence change, this resulted in potentially high participant burden. The participants were invited into the study – i.e. their motivation was likely not as strong as for those who seek out weigh loss programs on their own accord. We also conducted a number of statistical comparisons.

The modest positive effects of weight loss and behaviors and the lack of impact of assignment to two weight management programs on health care costs and utilization observed in this study, and in the earlier published analysis of this data,(8) indicate that to achieve weight loss through the work place, and in particular morbidity reductions strong enough to impact employer health costs, more extensive and intensive interventions are required, with more attention to motivation and compliance.