Background: Wearable devices are increasingly utilised to estimate physical activity (PA) in free-living subjects. These monitors facilitate long-term, associative research and generate extremely large datasets, providing new opportunities for research. With these new opportunities comes new considerations for researchers.
Based on the results of preliminary autocorrelation analyses, we developed a novel framework which utilises local, hourly PA data to account for missing data and therefore minimise the extent to which missing data can bias conclusions. This study compared the framework to alternative strategies used in accelerometer research.
Methods: A simulation study was conducted using the 14-days of minute-level, Fitbit charge 2 data collected in the NoHoW trial (ISRCTN88405328). Participants were selected based on amount of non-wear time (<2%). Next, PA data were deleted at random to produce datasets with 1315% missing data, occurring at random time points. Relative to the true data, we compared the bias introduced by the framework, the removal of missing data, mean imputation and multiple imputation.
Results: Comparisons were made using true and imputed data for 53 participants (minutes=1,068,480, hours=17,808, days=742). Using the proposed framework, agreement with the true data was superior to alternative strategies, with the root mean squared error in average steps/day for the framework being 313, compared with 527 for multiple imputation, 536 for mean imputation and 1393 for removal.
Conclusion: The proposed framework produces excellent agreement between true and imputed data. This novel method has applications for the maximisation of data utilisation and the minimisation of bias in PA research using commercial activity monitors.
12 Sep 2019 - 13 Sep 2019