A Tailored Multiple Imputation Approach To Handle Arbitrarily-Missing Accelerometer Data In A Randomized Controlled Trial

Abstract

Common applications of Multiple Imputation (MI) are too generic, yielding highly variable and nonreplicable results. PURPOSE: Evaluate a tailored MI approach for handling missing physical behavior (PB) outcome summaries (e.g., sitting time) due to accelerometer non-wear in an RCT and its impact in estimating time spent in PBs. METHODS: A missing data simulation was conducted from a complete subsample (N=39) of accelerometer data collected for 7-days at the start and end of a yearlong RCT. Data from 3 PB variables (sitting, standing and stepping time) were randomly deleted for 3 study groups at each timepoint to generate 10 datasets per group × timepoint with arbitrarily missing data (8-77%) in increments of 8%. A tailored MI approach was used for missing data where: i) each variable was imputed separately using unique correlated auxiliary variables, and ii) the number of imputations necessary to produce replicable and stable parameter estimates and standard errors were computed for each imputation model. Statistical differences in parameter estimates from univariate timepoint and repeated measures mixed model analyses between imputed and complete datasets were tested with paired sample T-tests and two-tailed Z scores, respectively. Errors (%) in parameter estimates relative to the complete dataset were calculated to quantify the magnitude and variability of the bias. RESULTS: The tailored MI approach produced unbiased parameter estimates and standard errors in both univariate timepoint and change analyses in sample sizes as small as N=13 with up to 54% missing data. Error and variability in parameter estimates increased exponentially above the 54% threshold in both univariate (mean % error ± SD: above threshold = 31 ± 10%, below threshold = 11 ± 4%) and change analyses (mean % error ± SD: above threshold =465 ± 98%, below threshold = 175 ± 55%). CONCLUSIONS: To our knowledge, this tailored approach is the most robust MI methodology to date for imputing incrementally missing accelerometer-based summary PB data in an RCT. Prior PB MI simulations yielded lower acceptable missing data thresholds (≤ 30%) in larger sample sizes (N ≥ 20), and did not test the impact on analyzing change between repeated measures. Tailoring MI to restore lost statistical power may prevent conservative estimates of the treatment effects in PB RCTs.