Modeling Weight Change in a Lifestyle Program to Prevent Type 2 Diabetes (303222)

Keywords: missing-not-at-random, Bayesian, modeling, prediabetes, nonignorable

Participants (n=13,273) in 250 CDC-recognized diabetes prevention programs attended weekly then monthly sessions during a year-long lifestyle intervention. Body weight was recorded at each session. Our objective was to model weight change as a function of number of sessions attended and physical activity minutes reported, adjusting for sex, age, and race/ethnicity. The main challenge was due to the fact that scheduling intervals and session attendance varied by program and individual, resulting in missing data. We considered 3 models. Model 1 used a participant’s % weight change at the end of the intervention period as the outcome variable; models 2 and 3 were repeated measures models which used cumulative % weight change calculated at monthly and weekly intervals, respectively. To account for nonignorable missing data, a Bayesian selection model was used which factors the joint distribution of weight change and missingness into the marginal distribution of weight change and the conditional distribution of missingness given weight change. We compared estimates of % weight change and covariate effects from the three models to assess the effect of missing body weight measurements.