Focused on the accelerometer-determined physical activity and sedentary time metrics in 9–11-year-old children, we sought to determine the following: (i) number of days that are necessary to achieve reliable estimates (Ggreater than or equal to0.8); (ii) proportion of variance attributed to different facets (participants and days) of reliability estimates; and (iii) actual reliability of data as collected in The International Study of Childhood Obesity, Lifestyle and Environment (ISCOLE).
The analytical sample consisted of 6025 children (55% girls) from sites in 12 countries. Physical activity and sedentary time metrics measures were assessed for up to 7 consecutive days for 24 h per day with a waist-worn ActiGraph GT3X+. Generalizability theory using R software was used to investigate the objectives i and ii. Intra-class correlation coefficients (ICC) were computed using SAS PROC GLM to inform objective iii.
The estimated minimum number of days required to achieve a reliability estimate of Ggreater than or equal to0.8 ranged from 5 to 9 for boys and 3 to 11 for girls for light physical activity (LPA); 5 to 9 and 3 to 10, for moderate-to-vigorous physical activity (MVPA); 5 to 10 and 4 to 10 for total activity counts; and 7 to 11 and 6 to 11 for sedentary time, respectively. For all variables investigated, the ‘participant’ facet accounted for 30–50% of the variability, whereas the ‘days’ facet accounted for less than or equal to5%, and the interaction (P × D) accounted for 50–70% of the variability. The actual reliability for boys in ISCOLE ranged from ICCs of 0.78 to 0.86, 0.73 to 0.85 and 0.72 to 0.86 for LPA, MVPA and total activity counts, respectively, and 0.67 to 0.79 for sedentary time. The corresponding values for girls were 0.80–0.88, 0.70–0.89, 0.74–0.86 and 0.64–0.80.
It was rare that only 4 days from all participants would be enough to achieve desirable reliability estimates. However, asking participants to wear the device for 7 days and requiring greater than or equal to4 days of data to include the participant in the analysis might be an appropriate approach to achieve reliable estimates for most accelerometer-derived metrics.