Purpose: The spontaneous physical activity of free-living humans is influenced by many factors, both endogenous and exogenous. Thus, knowing the extent of intraindividual variance is critical to appropriate data sampling, whether monitoring individual patterns of physical activity or exploring the determinants of physical activity. We have recently had the opportunity to collect detailed physical activity measurements on a substantial healthy elderly population over an entire year. We here apply the techniques of variance analysis and Fourier transformation to examine patterns of variability of daily step counts within the individual and within our test population; this has allowed us to make mathematical estimates of the number of days of sequenced, randomly timed and seasonally-timed observations needed to predict the annual habitual physical activity in healthy elderly people over a 365-day period.
Methods: Subjects were 37 males (aged 71 ± 4 years, height 1.57 ± 0.06 m, body mass 58 ± 9 kg) and 44 females (aged 71 ± 4 years, height 1.47 ± 0.06 m, body mass 51 ± 8 kg), healthy Japanese. All gave their written informed consent to participate in this institutionally approved study after the protocol, stresses, and possible risks had been fully described to them. A uniaxial accelerometer/pedometer was worn on the waistband throughout each 24-hr period for an entire year, accumulating information on the individual’s daily step count. A computer program randomly resequenced the original day-by-day data from each subject, thus giving information equivalent to the sampling of step counts on randomly selected days. The original day-by-day data from each subject were also regularly resequenced based on season and day of the week. Total variances for the original, randomly resequenced and regularly resequenced data were each separated into between-subjects and within-subject components. We then applied a modification of the classical Spearman-Brown calculation, estimating the intraclass reliability coefficient as the proportion of total variance attributable to between-subjects variance.
Results: The step count spectrum for the original data showed peaks with periods of 2.3, 3.5, and 7.0 days and an aperiodic component that had a greater power at low frequencies (i.e., nonwhite noise). These characteristics were absent in randomly resequenced data. To ensure that 80% of total variance was attributable to between-subjects variance, 25 and 8 consecutive days of observation were needed in male and female subjects, respectively. To achieve 90% on this same measure of reliability, 105 and 37 consecutive days of observation were required. In contrast, 4 days of randomly timed observations yielded 80% reliability for both men and women, and 11 and 9 days gave 90% reliability in men and women, respectively. If sampling also took account of season and day of the week, the respective observation periods for men and women were reduced to 8 and 4 days (i.e., 2 and 1 consecutive days of sampling every 89 days) for 80%, and to 16 and 12 days (i.e., 4 and 3 consecutive days every 89 days) for 90% reliability.
Conclusion: When estimating annual step counts, seasonal and/or random sampling of data allows collection of reliable data over substantially fewer days than would be needed for consecutive observations.