Background. Sedentary behaviour studies often describe the extent of sitting by a time proportion; typically per cent time spent sitting. Proportions are examples of so-called “compositional data,” since they add up to a constrained sum (i.e. 100%). Compositional data differ from non-compositional data in aspects of essential importance to their analy-sis and interpretation, including how to address variability. Compositional data analysis (CDA) acts in a space of logarithmically transformed ratios of proportions, rather than on the proportions per se. We compared the statistical properties of confidence intervals (CI) of group mean values of sitting-time proportions obtained using standard procedures and CDA, exemplified by sample sizes required to obtain a specified precision.

Methods. Sitting and non-sitting time proportions calculated from whole-day accelerom-eter recordings in 25 office workers were used as a heuristic example. Variability between subjects was assessed using standard statistics and CDA. In both cases, the size and shape of a 95% CI on the estimated mean sitting-time proportion of n subjects was assessed for different sizes of the mean and values of n.

Results. While standard CIs at a specific n are independent of the mean value and sym-metric, CDA-derived CIs are asymmetric, except at a mean of 50%, and wider at “medium” than at “extreme” mean values. In the example, a 95% CI of ±5% around the mean was ob-tained using n=26 subjects according to standard procedures. However, using CDA, upper 95% CI limits of +5% were obtained with n=5 for a mean value of 90%, but required n=58 when the mean value was 60%. Similar-sized lower 95% CI limits of -5% were obtained with n=13 and n=63 at 90% and 60% means, respectively.

Discussion. CDA-based estimates of sample sizes differed markedly from estimates based on standard statistics. Properties and implications of CDA in sedentary behaviour research deserve further consideration.