Reliability of accelerometer-determined physical activity (PA), and thus the required length of a monitoring period, appears to depend on the analytic approach used for its calculation. We compared reliability of objectively measured PA using different resolution of data in a sample of 221 Norwegian 2-6-year-old children providing 2-3 valid 14-day periods of accelerometer monitoring (ActiGraph GT3X+) during September-October, January-February, and May-June 2015-2016. Reliability (intra-class correlation [ICC]) was measured for 1-14 days of monitoring across the measurement periods using linear mixed effect modelling. These results were compared to reliability estimated using different resolution of data using the Spearman-Brown formula. The measured reliability improved only marginally with increased monitoring length and levelled off after 5-6 days. Estimated reliability differed substantially when derived from different resolution of data: 3.9-5.4, 6.7-9.2, 13.4-26.7 and 26.3-87.7 days of monitoring was required to achieve an ICC = 0.80 using an hour-by-hour, a day-by-day, a week-by-week and a period-by-period approach, respectively. Reliability could not be correctly estimated from any single resolution of data. We conclude that reconsideration is needed with regard to how reproducibility of objectively measured PA is analysed and interpreted.
Keywords: Test–retest; accelerometry; intra-class correlation; measurement error; reliability.