Procedure
Data were taken from the Weigh-to-Be project, a collaboration between the University of Minnesota and HealthPartners, a large Minnesota managed care organization (MCO), designed to evaluate phone and mail-based weight loss interventions. The University of Minnesota and HealthPartners Institutional Review Board committees approved the study protocol. The sample was comprised of 508 men and 1293 women who provided baseline data and were enrolled in the study. Participants were randomized to one of three groups: a mail-based weight intervention, a phone-based weight intervention, or a usual care group. Weight loss protocols were designed as potentially cost-effective delivery modes for weight loss within a large health plan population. The intervention modules were offered continuously over the 24 months of the study. After randomization, intervention participants were asked to contact study personnel to activate their assigned module (phone or mail). Following activation, participants spent an average of 5.3 months on phone course activities and an average of 8.8 months on mail-based activities. Participants in the usual care group were not offered weight intervention beyond health promotion programs in the health plan. Details of the study are described elsewhere [5, 6].
Measures
Age, gender, educational attainment, ethnicity, marital status, and smoking status were measured by self-report questionnaire at baseline. During the baseline visit, trained research staff measured height using a wall-mounted stadiometer. Baseline and 24-month weights were measured using a calibrated digital scale. Self-reported weights were collected from participant questionnaires at 6, 12, and 18 months, and were adjusted by +1.5 kg for men and +1.7 kg for women to account for self-report bias [7]. Weight change in kilograms was assessed at 6, 12, 18, and 24 months, and body mass index (kg/m2) was computed.
At 6, 12, 18, and 24 months, participants were asked the following item to assess engagement in weight loss strategies: "Indicate whether you did this during the past six months: reduce number of calories eaten, increase exercise levels, increase fruits and vegetables, decrease fat intake, cut out sweets and junk food, reduce amount of food eaten." Duration was assessed by asking participants: "For those items you did in the past six months, also write in the total number of weeks that you did it."
Data analysis
Statistical analyses were conducted using Statistical Analysis System Version 8.2 (SAS) [8] and Mplus Version 3.12 [9]. Frequencies, indicating engagement in a given behavior during any of the four time intervals (between baseline and 6 months, 6–12 months, 12–18 months, and 18–24 months), were calculated to determine prevalence of weight loss strategy use. Mean durations of strategy engagement (in weeks) were summed across the four intervals, for a possible total of 104 weeks. SAS general linear models were used to examine associations of strategy endorsement or duration with 24-month weight change; models controlled for baseline weight (in kilograms). To assist in interpretation of the significance of these findings, effect sizes are reported in tables using Cohen's d statistic (small = .20, medium = .50, large = .80) [10].
Structural equation methodology (SEM) was used to examine the temporal relationship between weight loss strategy use and weight change. SEM has four distinct advantages over the general linear model in this analysis: 1) a latent variable is used to measure the use of weight loss strategies, 2) a model-testing approach directly compares alternative models, 3) temporal effects of the intervention are modeled and compared, and 4) all cases, including those with missing data, can be included in the model if data meet conditions for either missing at random (MAR) or missing completely at random (MCAR). A latent variable representing weight loss strategies was estimated for each of the four time points following baseline based on six manifest indicators. Each manifest indicator was the number of weeks during the previous 6-month period in which individuals employed each of the specific weight loss strategies noted above.
After establishing the fit of the measurement model, a set of nested structural models were estimated and compared. All structural models included a set of baseline exogenous covariates and predictors that were allowed to covary. All post-baseline weight loss strategy latent variables and weight outcomes were regressed on the full set of covariates and predictors. The first structural model included autoregressive paths of lag 1 for both the weight loss strategy latent variables and weight. The next structural model added cross-lagged paths from the weight loss strategy latent variable for a 6-month interval and weight at the end of that interval. The next structural model replaced the cross-lagged paths from weight loss strategy to weight with cross-lagged paths from weight to weight loss strategy for the subsequent 6-month period. The final structural model included both sets of cross-lagged paths. Two relative fit indices, the Akaike Information Criterion (AIC) and the sample-size adjusted Bayesian Information Criterion (SBIC) were used to determine the best-fitting model. Once a final structural model was retained, further constraints were used to compare the magnitude of specific structural paths [11, 12].
Missing data were primarily due to attrition. Examination of missing data found a relationship between missingness and a number of variables, including age, marital status, education, ethnicity, and weight at baseline. In order to justify the use of maximum likelihood for missing data in the current analysis, the covariates shown to be associated with missingness were included in the structural model. All weight loss strategy latent variables and weight following baseline were regressed on all covariates, except weight at baseline, which was associated only with weight at 6 months.