Impact Of Accelerometer Data Processing Decisions On Data From Large Cohort Studies
- Presented on May 30, 2014
Background: Accelerometers objectively assess physical activity (PA) and are increasingly used in epidemiologic studies. However, processing techniques are not standardized and limit data comparability across studies.
Purpose: To compare the impact of wear-time assessment method and ﬁlter choice on accelerometer output in a large cohort.
Methods: Participants (7,650 women, mean age 71.4y (SD = 5.8)) were mailed an accelerometer instructed to wear it for 7 days, use a log to record dates and times the monitor was on/off, and return the monitor and log via mail. Data were processed using three wear-time methods (log, Troiano and Choi algorithms) and two ﬁlters (normal vs. low-frequency extension (LFE)). Available sample size, wear time and estimated sedentary time, light and moderate-to-vigorous (MV) PA were compared across methods.
Results: Logs had little missing data on dates (<2.5%), but a sizeable proportion had missing on/off times, leading to 9% fewer available subjects compared to algorithms. Using algorithms resulted in “mail-days” incorrectly identiﬁed as “wear-days” (27-87% of subjects); therefore a record of dates (not on/off times) that the monitor was worn is required. Filter choice and wear-time algorithm impacted sedentary time (~40 min higher for normal vs. LFE; 60 min lower for Troiano vs. Choi). Light PA (~40 min lower for normal vs. LFE), and steps per day (~7000 lower for normal vs. LFE) were impacted by ﬁlter choice but not wear-time algorithm, and MVPA was not substantially impacted by either.
Conclusion: Combining recorded dates (but not on/off times) and the Choi algorithm minimized missing data, participant and researcher burden. These data are useful for informing the development of standardized procedures in large-studies using accelerometers and a mail-based protocol.