Our study was based at Group Health Cooperative of Puget Sound (GHC), a consumer-governed, mixed-model health maintenance organization with more than 500,000 members. We received administrative and claims data on 8473 members aged 65 or older who enrolled in the GHC MA plan, were continuously enrolled at GHC for at least 1 year before joining the program, and participated in SS from January 1, 1998, through December 31, 2003. Up to three GHC MA enrollees (n = 24,331) who never used the program were matched by age and sex to serve as controls for each SS participant. Participants and their matched controls were each assigned an index date representing the month that the participant first enrolled in SS. We excluded members who had less than 2 years of continuous enrollment after their index date, had missing cost data in any of the 3 years, had long-term care costs at baseline, or were unmatched SS participants or controls, which left 4766 SS participants and 9035 matched controls in our study for analysis.

The SS program provided the GHC MA enrollees access to selected local fitness centers in an unstructured format. Participants had access to conditioning classes designed for older adults, exercise equipment, a pool, a sauna, and other amenities that varied across facilities. A subcontractor administered the program and worked with the fitness centers. The GHC MA enrollees learned about the SS program from targeted mailings, a member benefits Web site, or their health care providers during routine preventive visits.

GHC administrative data, which have been used extensively in prior research (17,18), were the source of all utilization, cost, patient demographics, and other covariates. Cost variables were derived from the GHC cost accounting system, which integrates clinical information, units of service, and actual costs from the general ledger for 15 separate feeder systems. GHC identified all costs as either direct patient care costs or overhead costs. All overhead costs are fully allocated to individual patient care departments. Departments captured in the database included medical staff, nursing, pharmacy, laboratory, radiology, hospital inpatient, and community health services. Units of service were weighted by relative value units for ancillary departments, by technical relative value units for radiology, by College of Anatomical Pathology units for laboratory, and by visit length for outpatient visits. From this process, the precise cost for each unit of service delivered was then calculated, and costs were assigned to patients on the basis of units of service used. For example, primary care costs included all direct and indirect costs associated with visits or telephone calls by primary care or preventive medicine personnel that were related to direct patient care, preventive services, or risk factor reduction counseling.

The utilization outcomes we examined were for inpatient admissions, primary care visits, and specialty care visits (defined as obstetrics and gynecology services, cardiac diagnostics, diagnostic pathology, alternative medicine, and rehabilitative services). The cost outcomes we examined were for primary and specialty care costs, inpatient admission costs, and total health care costs. We selected primary care visits and costs because a more general outpatient cost summary was not available. Total health care costs were examined to provide an overall summary of the impact of SS participation on costs.

In the analysis, we controlled for covariates that might influence the economic outcomes that were available in GHC administrative data. These covariates included age, sex, baseline utilization or cost (as appropriate), inclusion on the GHC diabetes or heart registries, indication of arthritis on the outpatient visit problem list, patient risk, and a preventive services index. Comorbid conditions (arthritis, coronary artery disease, congestive heart failure, hypertension, depression, and diabetes) were ascertained from problem lists for outpatient visits according to International Classification of Diseases, Ninth Revision, Clinical Modification diagnostic codes (19). Patient risk was measured using RxRisk (20), a measure of chronic disease burden and comorbidity that was previously shown to have good predictive power for explaining odds of hospitalization (21) and total health care costs (15,16,22). RxRisk was calculated for each member on the basis of age, sex, and pharmacy utilization data for a 6-month period before the index date (20). Because members who use other preventive services may be more likely to participate in SS than are members who do not, we calculated a preventive services index to adjust for self-selection of health-oriented individuals into SS participation. The preventive services index was derived from the sum of the number of times a person received a fecal occult blood test, a flexible sigmoidoscopy, a screening mammogram, prostate cancer screening, an influenza vaccine, or a pneumococcal vaccine during the 2 years preceding the index date.

We included all SS participants in the main analyses regardless of whether they made any visits to a health club over 2 years. We used two-tailed t tests and chi-square tests for unadjusted comparisons between SS participants and controls. We used multivariate ordinary least squares (OLS) regressions to determine differences in health care costs between SS participants and controls for 1 and 2 years after the index date while adjusting for age, sex, RxRisk, preventive services index, arthritis visits, inclusion in the health plan's heart or diabetes registries, and baseline use and costs. The results were similar to those obtained using generalized linear models with a gamma distribution and log-link function (23), so we present OLS results. Previous work suggests that using OLS regressions and large samples (≥500 observations) would yield unbiased estimates of absolute differences in use and cost data even when assumptions about normality and homoscedasticity are not met (24).

We performed exploratory dose-response analyses using OLS on the basis of the average number of health club visits during 2 years to determine incremental differences in total health care costs in members whose visits averaged fewer than 1 visit per week, 1 to fewer than 2 visits per week, 2 to fewer than 3 visits per week, and 3 or more visits per week. Average attendance was calculated by adding all health club visits during the 2 years and dividing by 104 weeks. SS participants who persisted with their visits to the health clubs for 2 years were compared with those who stopped using their physical activity benefit after the first year of SS enrollment. Because this study was interested primarily in differences in total health care costs between SS participants and controls and because subgroup analyses were purely exploratory, statistical tests were not adjusted for multiple comparisons.

To improve balance in observed covariates, we used propensity score (PS) adjustments in a sensitivity analysis (25,26) We estimated a logit model to generate each member's propensity of joining SS and entered PS as an additional covariate in our models. The inclusion of PS did not change the results of any of the models, so we present results from the simpler multivariate models.

All cost data were adjusted to 2003 dollars. Robust standard errors were used in all regressions. All statistical procedures were performed with Stata 9.0 (Stata Corporation, College Station, Texas). Institutional review boards at GHC and the University of Washington approved the study protocol.

Compared with controls, SS participants were slightly older, more likely to be male, had a lower chronic disease burden, used more preventive services, and had higher total health expenditures at baseline (Tables 1 and 2). A small percentage of members (2%) who signed up for the SS program never made a visit to a health club during the 2 years; another 2% did not visit a health club until the second year. The number of health club visits made by SS participants was 75 visits (median, 49; interquartile range, 11–120) in year 1 and 55 visits (median, 12; interquartile range, 0–89) in year 2.

The follow-up interval for all SS participants and controls was 2 years. In year 1, unadjusted total, inpatient admission, and specialty care costs were not different between SS participants and controls (Table 2). However, SS participants had more primary and specialty care visits (both, P < .05) and slightly fewer inpatient admissions than did controls (P = .02) in year 1. In year 2, SS participants had lower unadjusted total health care costs and fewer inpatient admissions than did controls (both, P < .01); unadjusted outpatient primary and specialty care utilization and costs were higher among SS participants (all, P < .01).

In year 1, adjusted total health care costs were similar for SS participants and controls (+$2; 95% confidence interval [CI], −$341 to $344; P = .99) (Table 2). We observed a modest difference between SS participants and controls in inpatient admissions in the adjusted model (−1.0%; 95% CI, −2.1% to −0.1%; P = .05). SS participants made more primary and specialty care visits than did controls (both, P < .001).

By year 2, total health care costs were significantly lower for SS participants compared with controls (–$500; 95% CI, –$892 to –$106; P = .01). This difference in costs was mainly due to the fewer inpatient admissions among SS participants compared with controls (–2.3%, 95% CI, –3.3% to –1.2%; P < .001) and the slightly lower inpatient costs (–$270; 95% CI, –$533 to –$6; P = .05). SS participants made more primary and specialty care visits and incurred greater costs associated with primary care than did controls (all, P < .001).

SS participants were categorized according to the mean number of health club visits per week over 2 years: fewer than 1 visit per week, 1 to fewer than 2 visits per week, 2 to fewer than 3 visits per week, and 3 or more visits per week. We observed graded baseline differences in the proportion of women, RxRisk, mean preventive services index, and health care costs across the visit categories (results not shown). Adjusted models showed a significant threshold dose effect on total health care costs at year 2 (Figure). Compared with SS participants who averaged less than one visit per week, those who averaged 2 to fewer than 3 visits per week or 3 or more visits per week had similar reductions in total health care costs at year 2 (2 to <3 visits, –$1252; ≥3 visits, –$1309).

Approximately 61% (n = 2902) of the SS participants continued to use their health club membership in the second year. The regular attendance of these continuers is reflected in their total number (SD) of health club visits. In year 1, continuers made 109 (84) visits, and in year 2, they made 89 (86) visits, higher than the average number of health club visits for the SS group as a whole. Although total health care costs at baseline were similar for both subgroups, members who did not continue health club attendance in year 2 (n = 1659), or noncontinuers, had significantly greater health service use in year 1. For example, more noncontinuers (11%) had an inpatient admission for all causes than did continuers (8%). In addition, noncontinuers had a mean (SD) of 5.7 (4.5) primary care visits and 3.5 (3.4) specialty care visits in year 1, compared with continuers, who had 5.1 (4.2) primary care visits and 3.2 (3.3) specialty care visits during the same year. Because we did not have access to data on whether disease burden increased in year 1, we compared the proportion of outpatient visits with new diagnostic codes between these two subgroups. In year 1, noncontinuers were more likely to receive new diagnostic codes on their problem list for arthritis (12.5%, noncontinuers vs 10.3%, continuers), cardiovascular disease (13.6%, noncontinuers vs 12.0%, continuers), diabetes (2.6%, noncontinuers vs 2.0%, continuers), and depression (8.1%, noncontinuers vs 5.6%, continuers).