Statistical Analysis

A priori power calculations used a 2-phased testing procedure, which has an overall familywise error rate that is controlled at 0.05.²⁹ In the first phase, the 3 intervention arms were compared with the control arm. We estimated that a sample of 600 participants (150 per arm) would provide at least 80% power to detect an 800-step difference between each intervention arm and the control arm. This approach assumed that the control group had mean daily steps of 6000, an SD of 2000 steps, a 10% dropout rate, and a conservative Bonferroni adjustment of the type I error rate with a 2-sided α of .017. In the second phase, only intervention arms that were significantly different from the control arm were compared with each other using a conservative Bonferroni adjustment of the type I error rate with a 2-sided α of .017 to adjust for up to 3 comparisons.

All randomly assigned participants were included in the intention-to-treat analysis. For each patient on each day of the study (participant-day level), the number of steps achieved was obtained as a continuous variable. These data were dichotomized at the participant-day level to create a binary variable that indicated whether each participant achieved his or her step goal, and this variable was used to estimate the proportion of participant-days that step goals were achieved.

Data are missing for any day that the participant did not use the wearable device or did not upload data. During the intervention period, missing step data rates ranged from 19% to 29% (eTable 1 in Supplement 2), which is similar to previous physical activity interventions.^{14,17,18,19,20,21} For the prespecified main analysis, we used multiple imputation for step values that were missing or for values less than 1000 steps per day. This method has been used in prior work^{14,17,18,19,20,21} and in this study because evidence indicates that daily step values less than 1000 may not represent full data capture.^23,24 Five imputations were conducted using the mice package in R, version 3.4.0 (R Foundation for Statistical Computing), which allows for participant random effects with this data structure.³⁰ The following determinants of missing data were included: study arm, week of study, calendar month, baseline step count, age, sex, race/ethnicity, educational level, marital status, annual household income, self-reported health, prior use of a smartphone or wearable device to track step counts, level of experience with wireless technology to track activity, regular travel to another city for work, and BMI. Results were combined using Rubin’s standard rules.³¹ This imputation approach has been used in prior work.^14,17 Sensitivity analyses were conducted using collected data without multiple imputation, both with and without step values less than 1000.

Unadjusted analyses estimated mean daily steps for each month of the study by arm. Similar to prior work,^14,17 adjusted analyses used PROC GLIMMIX in SAS, version 9.4 (SAS Institute Inc) to fit generalized mixed-effects models with a random intercept, to adjust for participant random effects, and to account for the repeated measures of daily step counts. In the main adjusted model, we included baseline step count and fixed effects for calendar month and study arm. To test the robustness of our findings, we fit a fully adjusted model that also included age, sex, race/ethnicity, marital status, annual household income, and BMI (calculated from self-reported height and weight). For change in steps, we assumed a normal distribution and obtained difference in steps between arms for the intervention and follow-up periods using the least squared means command. To estimate the adjusted difference in the proportion of participant-days that step goals were achieved between arms, we used the bootstrap method, resampling participants from within each arm 500 times. Determinants of success or failure in the interventions will be reported in the future.