Research design and methods
NAVIGATOR and study participants
NAVIGATOR (ClinicalTrials.gov: NCT00097786) was a double-blind, randomized, placebo-controlled clinical trial with a 2-by-2 factorial design. It evaluated whether valsartan or nateglinide, in addition to lifestyle modification, could reduce the risk of diabetes and cardiovascular events in people with impaired glucose tolerance who either had CVD or risk factors for CVD.14 15 Details of the trial, including the protocol, results, and design, have been published.12 14 15 A total of 9306 participants from 40 countries participated in the trial for a median of 5.0 years for the diabetes outcome.
Lifestyle modification intervention and identification of incident diabetes
All participants participated in the lifestyle program designed to achieve and maintain a 5% weight loss; reduce dietary saturated and total fats; and increase physical activity to 150 min/week. Study staff provided information at each visit in the first year consisting of written materials and videos with reinforcement via telephone contacts. The co-primary outcome of incident diabetes was defined as the development of a fasting plasma glucose ≥7.0 mmol/L and/or 2-hour glucose (measured 2 hours after oral glucose load) ≥200 mg/dL (≥11.1 mmol/L), confirmed on oral glucose tolerance test performed within 12 weeks. Diagnoses of diabetes made outside of the study were adjudicated by an independent committee blinded to treatment allocation.
Physical activity measures
Habitual ambulatory activity was assessed objectively using research-grade pedometers (Accusplit AE120, San Jose, California, USA) in all NAVIGATOR Study centers. The pedometers measured purposeful steps taken through a horizontal, spring-suspended lever arm moving vertically with each step, activating an electric circuit, and registering a step. Two weeks after the initial baseline, clinical measurements were taken, participants were fitted with a pedometer and instructed to wear it during waking hours for seven consecutive days. Participants were given a log book and instructed to write down their daily step counts at the end of each day and then return the log book to the study team.
Statistical analysis
Continuous variables are reported as medians (25th, 75th percentiles); categorical variables are reported by proportions. Patient characteristics were compared using the Pearson's χ2 test or Fisher's exact test for categorical variables and with the Kruskal-Wallis test for continuous variables. Unadjusted and adjusted Cox proportional hazards regression models were used to assess the associations of physical activity (pedometer steps) with the clinical outcomes. Covariate adjustment was implemented in three stages. First, we report an unadjusted model. Second, we adjusted for age, sex, and region, avoiding other patient characteristics that might be on the causal pathway between exercise and outcome. Third, a fully adjusted model added all variables significantly contributing to the prediction of diabetes in a baseline model in the original study.16 These factors or variables are listed in the footnotes of the respective tables.
Using a backward selection technique, the adjustment model was a priori selected as only those factors which at baseline were associated with the incidence of type 2 diabetes in this study.
We tested the proportional hazards assumption and linearity assumption for pedometer steps. As the linearity assumption was not satisfied we tested two spline pieces in our models. Since the second piece was distant from being significant, the pedometer data were truncated at 10 000 steps per day. In order to preserve important information on non-missing covariates and improve our multivariable models, pedometer data for the 23% of missing baseline data were handled by multiple imputation. The final models used five imputed data sets, and the results were combined using valid statistical methods that account for imputation variation. In order to investigate whether imputation affected the strength of results, the analysis was repeated for the subset of 7118 participants with complete data. For all outcomes, two parallel analyses were performed: (1) All patients with pedometer measurements (n=7118). (2) Five imputed data sets. The point estimates of the HRs were very similar for these analyses. Since the results from the complete case and imputed (n=9308) data sets were similar, we report here only the results from the imputed data. Note, the linearity assumption was not satisfied for pedometer steps; two spline pieces were tested in the model and the final model only included the first piece, which is equivalent to truncating the pedometer steps at 10 000 steps per day.
A two-sided P value<0.05 was considered statistically significant for main effects. A P value <0.1 was considered statistically significant for interactions. Statistical analyses were done using SAS V.9.2 (SAS Institute, Cary, North Carolina, USA).