The DASH for Health website was developed by a team at Boston University School of Medicine (BUSM). The website provides tools for enrollees to record and follow their body weight, blood pressure, eating habits, and exercise habits as well as providing a variety of healthy recipes based on the DASH diet. These tools include easy-to-use entry fields for entering weight, blood pressure, 24-hour food recall, and exercise habits. Although enrollees were encouraged to use these data “progress reports” as a means of encouraging their adherence to the program, there were no requirements or set expectations of how often enrollees would enter their own data. The website also provides two new articles each week on improving nutritional habits (based on the DASH diet) or healthy exercise. Everyone who enrolls in the program gets a reminder email each time a new article is posted on the website. The website is hosted at SignalZ Corporation (Montpelier, VT, USA). Self-reported data on demographics (age, gender, and employee/dependent status) and website visits during each year were collected by SignalZ and transferred to the research team at BUSM. Because enrollees had to enter the site using their username and password, standard website monitoring software was used to track how many times a particular enrollee visited the site. We define a “visit” as a unique instance of an enrollee logging on to the site.

When the program was about to launch, EMC Corporation informed employees that DASH for Health would be available at no cost to them and their adult household members. EMC had no other role in the program. EMC management had no access to the identities of the subjects or to their medical data. Once the program began, the employer took no role in encouraging website usage. The announcing emails from EMC leadership clearly stated that the employer would not know any enrollee’s individual data or the identities of people who did and did not enroll.

EMC’s clinical data warehouse, Ingenix, provided demographic and medical and pharmacy cost data for employees and their adult dependents. These data included date of birth, gender, and employee status. D2Hawkeye, a medical analytics firm specializing in data warehousing and health claims analysis, provided data cleansing services and summarized study subjects’ demographic information, condition specific diagnoses, and total health care costs. D2Hawkeye calculated a risk adjustment factor for each study subject, using a proprietary methodology that combines diagnostic, procedural, and pharmacy information in health claims data with pattern recognition and times series analysis. The risk index provides a single representation of an individual’s disease burden and has been shown to correctly distinguish high-cost and low-cost persons 77% of the time. Additionally, the likelihood that an individual will be in a high-cost group increases as the risk index value increases. We used an extension of the risk index, the adjusted risk index (ARI), which incorporates assessments of gaps in care (medical or pharmaceutical) for various conditions [13].

We used individual subject-level linkages to merge data on DASH for Health participants’ website visit frequency with health care cost data and other variables. A subset of study subjects was identified with evidence of diabetes, hypertension, or hyperlipidemia, conditions that could be expected to benefit from the improved dietary behaviors that are encouraged by the DASH for Health program. This subset is referred to as the cardiovascular (CV) risk group. Data were managed and analyzed using SAS, version 9.1.3 (SAS Institute Inc, Cary, NC, USA). Approval for this study was obtained from the Boston University Institutional Review Board.

Study subjects with missing data for key fields such as age, employee status, or gender were excluded, as were DASH participants with no matching entries in the health care cost files. These enrollees were either not covered by the EMC insurance plan or were not affiliated with EMC at the end of the study period. The study was limited to study subjects with health care costs in both study years to ensure that those with bad claims data reported as negative or zero amounts were excluded from the analysis and to increase the likelihood that the analysis was restricted to those with primary insurance coverage through EMC. Information on job classification was not included since baseline year health care costs and risk assessment are proximal for our analysis.

The total number of EMC employees and their spouses was 29,675, of which approximately 75% (n = 22,354) were reported to have received health coverage through EMC and had health care costs in the baseline or study year. Sixty-eight percent (n = 15,237) of these subjects had costs in both years. Additionally, 3797 EMC employees or spouses enrolled in the DASH for Health program and used the DASH program website at least once in the study year (DASH participants). Seventy-three percent (n = 2756) of these DASH participants were found in the health claims summary file, and 71% (n = 1967) of these 2756 had health care costs in both years. The final study sample (N = 15,237, DASH = 1967, non-DASH = 13,270) also reflected the restrictions described above.

In univariate analyses to describe our sample, we generated means and standard deviations for continuous variables and counts with percentages for categorical variables. We used bivariate analyses to examine the relationships among demographic, annual health care cost, and website usage data for 15,237 study subjects over the 2-year baseline and study year period. In these analyses, we employed chi-square tests for analyses of categorical data and two-sample t tests for comparisons of continuous by categorical variables. Multiple linear regression analysis examined the relationship of DASH program enrollment with study year health care costs, controlling for other salient factors. To control for potential bias of self-selection of the DASH participants, we used a logistic regression model to construct a propensity score that represented the likelihood of participation in the DASH program. Studies that use observational data, as ours does, are subject to bias because the “treatment” and “control groups” are not randomly assigned and may differ in ways that affect the outcome of interest. A propensity score, which represents the conditional probability of receiving a given treatment, given a vector of measured covariates, is frequently used in such studies to adjust for differences in the observed characteristics between the treatment and the control groups. Our use of logistic regression is one of several recommended approaches for determining the propensity score [14].

All available study variables were included in the propensity score model (baseline year risk index and ARI, baseline year costs, age, gender, and employee status). We then used a linear regression model to compare study year costs among DASH and non-DASH participants, controlling for age, gender, ARI, baseline year costs, and likelihood of DASH participation (propensity score). In this model, study year total health care costs was the dependent (outcome) variable; independent (predictor) variables were DASH participation, age, gender, employee status, baseline year ARI, and baseline year total health care costs. Next, because stronger effects of an intervention are often observed in the most compliant or frequent participants, we examined whether more frequent use of the website than is typical is more strongly associated with reduced costs among DASH participants than across the full range of website use overall. For this, we used a similar linear regression analysis to evaluate the relationship of website usage intensity at or above the median to health care costs in the study year. Study year total health care costs was the dependent variable; independent variables were number of website visits, age, gender, employee status, baseline year ARI, and baseline year total health care costs. We repeated these analyses using the 735 DASH CV risk group and non-DASH CV risk group study subjects who showed evidence in both study years of diabetes, hyperlipidemia, or hypertension.

To address the skewed distribution of the cost data, baseline and study year costs in all analyses were top-coded at US$25,000, which represents the 99th percentile of annual health care costs in our study population. Website usage was top-coded at 75 website visits to address skewness. The analyses were repeated using trimming to remove high-cost outliers, using the entire study population and using top-coding at other thresholds. As top-coding is recommended for reducing the effects of high-cost outliers on model results while retaining useable observations, we report results of those analyses here. Our choice of US$25,000, or the 99th percentile, is also consistent with others [15-17].

As part of model development, we added interaction terms as covariates and conducted additional analyses where we limited study subjects to employees, males, and male employees, reflecting the larger number of employees and males in the study group. To account for nonlinear effects of DASH participation and website use intensity, we grouped study year costs into quartiles; we also grouped website visits and study year costs into quartiles. There were no differences suggesting that the interaction terms, study population restrictions or groupings should be included in the final model as none of these modifications affected the results. We repeated the analyses we report here using square root and log transformations for all cost variables because of their good performance with heteroscedastic health care cost data [18]. Our results based on the transformed data were similar to those observed using the untransformed data. Thus, to allow for ease of interpretation, we present results here for the untransformed data.

Demographic, ARI, and baseline medical and pharmacy cost information were gathered for all study subjects (N = 15,237) and for the subgroup of subjects with CV risk conditions (N = 735). These measures were also examined for DASH (N = 1967) and non-DASH (N = 13,270) participants overall and in the CV risk group (DASH: N = 134; non-DASH: N = 601). These results are shown in Table 1.

Among the 15,237 study subjects, 55% (n = 8384) were employees, while 45% (n = 6853) were spouses. The overall study sample was 46% male (n = 7041), with an average age of 40.2 years. Average total baseline year health care costs were US$2684. A slightly higher proportion of the DASH participants was male, compared to the non-DASH participants (56% vs 45%). The average age of the DASH participants was 40.7 years, slightly higher than the non-DASH participants (40.1 years). DASH participants were mostly employees (84%; n = 1659). DASH participants had a mean ARI of 3.52, which was slightly lower than the mean ARI for the non-DASH participants (3.72). Average total costs among DASH participants in the baseline year were lower than among the non-DASH participants (US$2181 vs US$2758).

Table 1 also shows demographic, ARI, and baseline year health care costs for the DASH and non-DASH CV risk groups. This subgroup was 65% male (n = 476) and 64% employees (n = 469), with an average age of 47.5 years, older than the overall study sample. The mean baseline year ARI of 12.14 (DASH: 10.10, non-DASH: 12.59) was also higher than in the general study sample, as were mean total baseline health care costs (overall: US$5663; DASH participants: US$4239, non-DASH: US$5980).

The results of the linear regression analysis of study year costs for DASH vs non-DASH in the full sample are shown in Table 2. Among all study subjects, DASH participation was associated with increased health care costs, although this result was not statistically significant (difference in mean costs, DASH vs non-DASH = US$85.14; P = .35). Model covariates associated with significantly higher study year costs were older age (P = .01), being female (P = .004), higher baseline ARI (P < .001), and higher baseline year costs (P < .001).

Results of the linear regression analysis of study year costs for DASH and non-DASH CV risk group study subjects are also shown in Table 2. DASH CV risk group members’ study year health care costs were US$827 lower, on average, than those of the non-DASH CV risk group members (P = .05). Higher baseline year ARI was significantly associated with higher study year costs, with each additional unit of risk associated with an increase, on average, of US$134 (P < .001). Each additional dollar in baseline year costs was associated with an additional US$0.34, on average, in study year costs (P < .001).

Results of the analyses of intensity of website usage and study year costs are shown in Table 3. Among 1967 DASH participants, each additional website visit was associated with almost a US$7 decrease in study year health care costs, but this result was not statistically significant (P = .21). Among participants whose website usage was at or above the nine-visit study year median (n = 1028), each additional website visit was associated with a US$14 decrease in health care costs on average (P = .04).

Figure 3

CV risk group change in unadjusted total costs from baseline to DASH study year in DASH participants vs nonparticipants

A number of study limitations should be noted. Our analysis used observational data and is vulnerable to selection effects, which represent the largest threat to validity in observational studies. Our use of a control group and a pre-test, coupled with our use of a propensity score, are standard mechanisms for addressing selection bias in quasi-experimental study designs. However, our model would not control for the possibility that study subjects who were motivated to manage their CV risk conditions were more likely to enroll in DASH for Health and would visit the website more often. However, it is unlikely that our finding that even moderate use of the DASH for Health program is associated with lower health care costs is attributable to motivation and not to participation in the program. This finding is consistent with clinically oriented studies of the DASH diet program indicating that the DASH diet is associated with improved blood pressure, lower cholesterol levels, and increased insulin sensitivity. Such clinical improvements would be expected to be associated with lower health care costs.

Several other limitations should be noted. First, the study sample was limited to persons with health care costs in both years. The proportion of study subjects with costs in only one year (costs in baseline year only: 90%; costs in study year only: 79%) is consistent with the national Medical Expenditure and Panel Survey (MEPS) data indicating that 11% of commercially insured persons do not use health services in a given year. D2Hawkeye received and cleaned the source claims data and reported that the proportion of study subjects with zero costs was higher than expected. Because the data could not be sent back to the data warehouse for review, we chose to restrict the study to persons with costs in both years. In analyses with all study subjects, including those with no costs in either or both years, costs for all DASH participants were significantly lower than for non-DASH participants, and the cost savings were higher among DASH participants with CV risk conditions. (Results available on request.) Restricting study subjects to those with costs understates the effects of the DASH program on costs. Second, our analysis compares study subjects who visited the DASH website at least once with those who did not visit it at all. However, subjects who visited the website once during the study year would not be expected to benefit from the nutrition and exercise education that the DASH for Health website program offers. Our finding of a dose-response effect would be strengthened by the inclusion of subjects who signed up for the program but never visited the website. From the perspective of benefit design, this effect would be more useful for guiding planning or coaching efforts. Third, our propensity score is based only on available data and therefore may not completely address potential selection bias. It is possible that DASH participants with CV risk conditions would have had decreased health care costs without participation in DASH. However, the clinically observable effects of the DASH diet on CV risk conditions make this result less likely. Fourth, our results are based on all health care costs for all conditions. Costs for conditions targeted by the DASH diet cannot be distinguished from costs for medical conditions that are not likely to benefit from improved diet, such as pregnancy and childbirth and back pain. However, the dose-response effect among participants who use the DASH website more often and the increasing significance of additional website usage in the DASH CV risk group suggest that participation in the DASH program is related to decreases in health care costs among persons at higher risk for health care expenses. We performed additional analyses of the effects of DASH participation on study year costs for subjects with different baseline year risk levels. Based on the regression model, at higher levels of baseline year risk, DASH participants had significantly lower study year costs than nonparticipants. (Results available on request.) These results are of particular interest given that higher levels of baseline year risk are predictive of higher costs [13]. Finally, study year costs include health care costs incurred during the initial enrollment period. The effects of DASH for Health program participation would not expected to be evident during the initial launch of the program or during the first few months of the year. However, this limitation suggests that our study results understate the effects of DASH for Health program participation since they include this initial time frame.

Our focus on short-term cost savings does not examine whether these savings are offset by increases in other costs [40]. Although the DASH for Health program is relatively inexpensive, further study to evaluate its cost-effectiveness as a corporate benefit would be beneficial. Finally, the study sample was limited to employees at one technology firm and their dependents. Our study results may not be generalizable to persons without employer-based insurance or to adults who are less comfortable using the Internet for information about diet, exercise, and health management. Research into the effectiveness of DASH for Healthin other populations is warranted.

Use of an Internet-based program that targets changes in diet and exercise to reduce weight, cholesterol, and blood pressure shows evidence of reducing short-term health care costs among persons at high risk for health care expenditures with conditions targeted by the diet. Offering access to a website with diet and exercise information appears to have promise as a low-cost, employer-sponsored benefit that contributes to lower health care costs among persons at higher risk for above-average health costs and utilization.