The study was conducted among 50–70 year old members of the HealthPartners health plan in the Minneapolis/St. Paul metropolitan area.
Eligibility, Sampling, and Recruitment
We have discussed eligibility, sampling and recruitment in detail elsewhere (Martinson, et al., 2008
). Briefly, we used health plan administrative data to identify age-eligible members who had been enrolled in the health plan for at least 11 of the 12 months prior to eligibility screening. Recruitment was initiated through direct mailings to random samples of individuals not
meeting study exclusion criteria based on initial clinical records review. We supplemented this direct mail approach with study advertisements to facilitate “self referrals” to the study. Following an initial phone-based eligibility screening, an institutionally approved consent form was mailed to interested individuals. When completed consent forms were received, a baseline telephone interview was scheduled, upon completion of which the subject was randomized to either the treatment or control arm. Recruitment occurred over 15 months from July to August 2004 and December 2004 through December 2005.
We considered as study-eligible those who reported accumulating at least 30 minutes of moderate or vigorous PA a day at least 2 days per week on average over the past four weeks, and who reported that this represented an increase in PA within the past 12 months. Individuals were excluded who had a modified Charlson comorbidity score > 3, a standard index of comorbidity calculated using prior year diagnoses of a broad range of serious medical conditions (Charlson, et al., 2008
, Charlson, et al., 1994
, Deyo, et al., 1992
, Rush, et al., 2000
), or had diagnoses of coronary heart disease (CHD), congestive heart failure (CHF), atrial or ventricular arrhythmias, cardiac arrest, or had an implantable defibrillator.
The study coordinator randomized 1,049 subjects to either the PA treatment condition (KAM) or a usual care control group (UC). Subjects were allocated equally in blocks of 20 according to a schedule prepared by the study statistician using a random numbers table and unobservable to the study coordinator. All participants self-reported their PA levels at baseline and at 6, 12, and 24 month follow-ups. Primary study outcomes were PA expressed as estimated kcal/wk of energy expenditure and maintenance of PA levels relative to baseline. Telephone interviewers collecting self-report data were blind to study condition.
KAM Intervention Description
Participants randomized into the intervention were offered a 24-month interactive telephone and mail-based PA support program, based primarily on the principles of Social Cognitive Theory (SCT) (Bandura, 2004
, Bandura, 1986
) and relapse prevention theory (Marlatt and Gordon, 1985
). Intervention strategies were weighted toward maintenance focused self-management, including cognitive (goal setting, identification of barriers, and problem solving), behavioral (self-monitoring using pedometers and log-books), and environmental (telephone coaching support, leveraging participants’ social support networks) strategies. The core component of the intervention was a seven session course delivered approximately bi-weekly over the telephone by PA coaches with exercise science backgrounds and training in behavior change theory. Participants then received monthly follow-up calls for the remainder of the first year and bi-monthly calls for the second year. Additional intervention components included motivational campaigns and a lending library of PA resources. Details of the intervention are provided in with further details available elsewhere (Sherwood, et al., 2008
Overview of all Keep Active Minnesota intervention contacts and topics addressed
Usual Care Description
Participants randomized to the UC arm received information about the 10,000 steps PA program offered by the health plan and 4 newsletters focused on general health and wellness during their two years of study participation.
All outcome measures were collected during a 45-minute telephone interview administered prior to randomization (baseline) and 6, 12, and 24 months later.
Outcomes of Interest
The outcome variables of interest were kilocalories expended per week in a range of physical activity (AllPA kcal), and specifically in moderate and vigorous intensity activities (MVPA kcal), calculated at baseline and 6, 12, and 24 months, and maintenance of PA at 6, 12, and 24 months relative to baseline (maintenance). Both kcal expenditure measures were computed using the CHAMPS instrument, designed to quantify relative kcal expenditure in adult populations based on self-reported frequency and duration of a range of common physical activities (Stewart, et al., 2001
). CHAMPS-calculated kcal expenditure has demonstrated acceptable reliability, with ICCs for moderate intensity activities of 0.67, 0.76, and 0.81–0.88 at six months, two weeks, and one week, respectively (Stewart, et al., 2001
, Cyarto, et al., 2006
). The instrument has also demonstrated adequate discriminant and construct validity, correlates well with other measures of PA, and is sensitive to change (Stewart, et al., 2001
). Two types of routine activities, “Do heavy work around the house (washing windows, cleaning gutters, shoveling snow),” “Do heavy gardening (spading, raking, pushing a lawnmower)” were reported at unrealistically high levels consistently over time and across study groups. Similar high reporting on these two items has been found by the instrument authors in other studies of mid-life and older adults (Castro, et al., 2008
, King, et al., 2009
). Similar to these studies, we removed the two over-reported items from all kcal expenditure calculations, while two items pertaining to light housework and gardening remained. We identified participants as maintaining PA if their MVPA kcal expenditure at a follow-up measurement was at least 80% of their baseline expenditure and at least 1500kcal/wk. These lower bounds ensured that participants classified as maintaining PA were engaging in about the same amount of activity as they had been at baseline, and that this activity level approximated the recommended 5 bouts of 30 minutes of moderate activity per week.
General linear mixed model (Laird and Ware, 1982
) regression models were estimated using SAS PROC MIXED specifying time within participant, random participant intercepts, unspecified covariance structure and restricted maximum likelihood estimation (Littell, et al., 1996
, Statistical Analysis System, 2002–2003
) to test the hypothesis that KAM participants maintained kcal expenditure from baseline through the three follow-up time points relative to UC participant expenditure levels. In the primary efficacy analyses, AllPA and MVPA kcal were predicted separately from the time at which kcal was measured (baseline, 6 months, 12 months, 24 months), which varied within participants, and randomized treatment arm (KAM, UC), which varied across participants. We chose the mixed model approach because, relative to a general linear model approach (e.g., repeated measures ANOVA or ANCOVA), it accommodates variation in the number of observations per participant without reliance on imputation to replace missing observations. Preliminary examination of the kcal expenditure measures at each time point revealed outlying observations. Observations greater than 5 standard deviations from the time-specific median were excluded from the analyses (AllPA kcal n=3 at baseline, n=4 at 6 months, n=1 at 12 months, n=1 at 24 months; MVPA kcal n=3 at baseline, n=4 at 6 months, n=1 at 12 months, n=4 at 24 months) to prevent unrealistically high kcal expenditure values from resulting in the (likely) over-estimation of KAM efficacy. Additionally, kcal and maintenance outcomes were missing for time points at which participants did not complete the CHAMPS (6 months: n= 34 UC, n=28 KAM, p=.45; 12 months: n=39 UC, n=28 KAM, p=.17; 24 months: n=51 UC, n=32 KAM, p=.03). Maximum likelihood estimation ensured that all available kcal observations, excluding those greater than 5 SD above median, from all randomized participants were used to estimate model parameters. We estimated models including the outlying observations to ensure that their omission would not affect conclusions drawn from the analyses. There were no substantive differences in any of the omnibus tests or estimated model parameters, and the planned contrasts at 6, 12, and 24 months estimated the between groups differences to be inconsequentially larger than what is reported below.
The UC arm and baseline measurement were treated as reference categories for the treatment and time effects in the mixed regression models. Thus, the KAM parameter tested whether the KAM and UC groups were different at baseline. Because a separate random intercept was estimated for each participant, the 6-month, 12-month and 24-month parameters estimated how much, on average, participants in the UC arm increased or decreased kcal expenditure at each follow-up relative to their own baseline kcal expenditure. The 6-month*KAM, 12-month*KAM, and 24-month*KAM interaction parameters estimated the difference in the average change in expended kcal/wk from baseline to each follow-up among KAM compared to UC participants. Planned comparisons of kcal expenditure among KAM versus UC participants at each time point assessed whether KAM successfully helped participants maintain PA at 6, 12, and 24 months.
We estimated a generalized linear mixed model regression (GLMM) (Breslow and Clayton, 1993
) using subject-specific pseudo-likelihood estimation (Wolfinger and O’Connell, 1993
) in SAS PROC GLIMMIX to test the hypothesis that PA maintenance (0/1 outcome) would be higher among KAM participants than UC participants at 6, 12, and 24 months. In this model, the binary outcome was normalized using a logit link function and binary distribution, and the 6 month measurement served as the referent time point.
Sample size was based on that needed to detect a time (24 month vs. baseline) by treatment (KAM vs. UC) interaction, in which the standardized between groups difference at baseline was Cohen’s d = .00 and d = .25 at 24 months, at .80 power (two-tailed, alpha = 0.05) on AllPA kcal in a two group repeated measures ANOVA. The use of GLMM meant that more observations were included in the primary analyses than was assumed in the power analysis, so that they were better powered and more generalizable than a GLM/ANOVA approach. We assumed a common standard deviation of 1500 kcals at each of 4 time points and a first order autoregressive residual covariance structure. These parameters suggested that N=349 per study arm would be needed. Assuming non-differential 70% retention across study groups at 24 months, we targeted n=500 per arm for recruitment.
Outcome analyses were conducted after study recruitment was completed and intervention staff were blinded to results during the intervention delivery period so that neither the sample size, group assignment, nor intervention delivery could be influenced by knowledge of the impact of the intervention on PA.