National Institute on Aging
Comparing methods for creating an overall physical activity estimate from multiple accelerometer days
- Published on Jun 21, 2017
Background: Accelerometers provide an objective assessment of physical activity and are frequently deployed for at least multiple days. However, it is unclear how to summarize multiple days into one overall physical activity estimate, particularly in the presence of missing data. This study compared summarization method prediction and precision, and the effect of summarization choice on the association of physical activity and mortality.
Methods: Data were from NHANES adults (2003-2006) with mortality follow-up through 2013. Overall activity estimates were calculated under different random day, median, mean (natural scale), mean (log transformed), and mixed models adjusting for day of the week and observation day number. To evaluate the summary method prediction and precision, activity estimates were compared to a left-out day using an adjusted R² for each summarization method. Summary methods were also evaluated by number of days of observation. To compare the effect of the summarization method on the association of physical activity and mortality, hazard ratios (95% confidence intervals) were calculated using Cox survival analysis.
Result: Random day was the poorest predictor of the left-out day (R² 23.6% – 48.9%). Mixed models, adjusting for day of the week and day order, performed the best (74.6% – 85.4%), particularly among few observation days. However, among 6 or 7 days, medians (66.0% – 68.6%) or means (60.9% – 69.5%) perform adequately. All summary methods showed similar inverse, curvilinear trends between physical activity and mortality (Figure 1).
Conclusion: While summarization methods differ in their predictive ability of the left-out day, particularly among few observed days, all methods result in similar conclusions in terms of the association between physical activity and mortality. This provides some evidence that physical activity and mortality studies using various summarization techniques may be able to be directly compared.
- Eric Shiroma 1
- Osorio Meirelles 1
- Lenore Launer 1
- Tamara Harris 1
ICAMPAM 2017 Abstract Booklet