Comparing Survival Analysis Models for Assessing the Persistence of Use of Activity Trackers for Increasing Sedentary Adults Physical Activity

This study compares parametric and semiparametric survival models for analyzing persistence in using an activity tracker to increase daily activity levels. Most studies employ semiparametric Cox regression to avoid distributional assumptions. This overlooks the potential benefits of parametric estimation for situations where it is well suited such as randomized trials.

Participants were randomized to either Activity Tracker Only (ATO, N = 474) or Activity Tracker plus Coaching (AT+C, N = 473). Exercise Readiness for Change was a measured continuous covariate. Males composed 16% of the sample and average age was 43 (22 to 68). Average BMI was 36 and ranged from 28.1 to 68.75; average weight was 222.5 pounds.

The primary study outcome was the number of months continuously using the activity tracker during one year (recorded electronically). Comparisons among models were made using log-likelihood values along with survival curves and SE’s. Robust standard errors were estimated for all models. Total time at risk was 4,616.8 months.

Fully 25% of participants discontinued using the activity tracker within one week and 50% of participants discontinued by 6-months. AT+C participants remained active users 6-months on average compared to 4.1months for AT participants. Because of the rapid discontinuation, the accelerated failure time models performed best in terms of –LL values (Gamma regression = -1800.24; Log-logistic regression = -1820.84) with Weibull regression (-1851.25) and Exponential regression performing well (-2148.37). The Cox model performed relatively less well (-4566.64). In this context, the Exponential and Weibull models produced the smallest SE’s.