This cross-sectional study used baseline data from two Danish studies: the Danish PHysical ACTivity cohort with Objective measurements (DPhacto) and the New Method for Objective Measurements of Physical Activity in Daily Living (NOMAD) study. DPhacto and NOMAD were identical in data procedures and collection, which facilitated merging the data. Details of the studies have been described previously [23, 24].
The study population consisted of workers with low SES recruited from Danish workplaces within cleaning, transportation, manufacturing, construction, road maintenance, garbage disposal, assembly, mobile plant operator, and health care[23, 24]. Eligible workers were employed in one of the mentioned sectors for at least 20 h/week; between 18–65 years old; and had given voluntary consent to participate. Workers were excluded if they were pregnant, had fever on the day of data collection, or band-aid allergy.
The DPhacto and NOMAD studies were approved by the local Ethics Committee (file number H-2-2012-011  and file number H-2-2011-047 , respectively). Both studies were conducted according to the Helsinki declaration and all data were anonymized in relation to individuals and workplaces.
Data were collected included questionnaires, health checks, and accelerometer-based measurements [23, 24]. Eligible workers were invited to complete a questionnaire and to participate in a health check, which consisted of anthropometric measurements and a physical health examination. Moreover, participants were asked to wear accelerometers for a minimum of two consecutive workdays and to complete a diary reporting time at work, time in bed at night and non-wear time.
Accelerometer Measurements of Physical Activity and Sedentary Behaviour
Physical activity at work and leisure time was assessed using data from one tri-axial ActiGraph GT3X + accelerometer (Actigraph, Pensacola, FL, USA). The accelerometer was fixed using double-sided adhesive tape (3 M, Hair-Set, St. Paul, MN, USA) and Fixomull (Fixomull BSN medical GmbH, Hamburg, Germany) and placed on the right thigh. Accelerometer data were downloaded using Actilife Software version 5.5 (Actigraph, Pensacola, FL, USA)  and analyzed using the custom-made MATLAB program Acti4 (The National Research Centre for the Working Environment, Copenhagen, Denmark) . The Acti4 program has been shown to separate physical activity types with high sensitivity and specificity under semi-standardized and non-standardized conditions. Classification of physical behaviors using Acti4 has been described previously. In brief, physical behaviors (i.e., cycling, stair climbing, running, walking, standing, sitting and lying) were classified based on an algorithm using angles from the accelerometers axis and standard deviation of mean acceleration.
Day of the week, daily work hours, leisure time and time-in-bed were defined from the participants’ self-reported diary information. Only workers with at least one day of valid accelerometer measurements of work and leisure time periods were included. A valid day consisted of ≥ 4 h of accelerometer-derived work and leisure time or ≥ 75% of the individual’s average work and leisure time. Leisure-time periods before work and time-in-bed were not considered in this study. Figure 1 show the flowchart of the study population. A total of 1200 eligible workers answered the questionnaire and/or participated in the physical health check. Of these workers, 37 were excluded due to being department leaders or students, were on holiday, pregnant or did not want to participate. Two-hundred workers were excluded from the study because they did not valid have leisure time accelerometer measurements on at least one weekday. Therefore, a total of 963 workers were included in the study.
(INSERT FIGURE 1 HERE)
Sex and age of the workers were determined from each worker’s unique Danish civil registration number. BMI (Body Mass Index) was calculated as weight (kg) divided by height (m) squared (kg/m2). Information on smoking-status was obtained by the question: “Do you smoke” with four response categories: daily smoking, occasionally smoking, formerly smoked, and never smoked. The variable was dichotomized into smokers and non-smokers (including formerly smokers). Work duration was calculated as the log of total accelerometer-derived work time. Information on shift work was assessed using the question: “At what time(s) of the day do you usually work in your main occupation?” with three response categories: fixed day work, night/varying work hours with night, and other. The variable was dichotomized into workers with and without fixed day work.
Time use of daily work and leisure time behaviours was treated as two compositions of activities performed within a 24-hour day. Work and leisure time were defined as a 3-part composition, consisting of time spent on sedentary (i.e. sitting or lying), standing and active (i.e. walking, running, stair climbing or cycling).
Compositional means were used to describe the day-to-day pattern of work and leisure time physical behaviours[29, 30]. They were obtained by computing the geometric mean of each individual behaviour of the respective compositions and then normalising (closing) these vectors of geometric means to workers’ average daily work and leisure time (i.e. 450 min and 450 min, respectively). On non-workdays, the leisure time composition consisted of daily waking time, normalised to the workers’ average daily time spent awake (i.e. 960 minutes).
Daily work and leisure time-use compositions were expressed using pivot isometric log-ratio (ilr) coordinates. The first pivot-coordinate was calculated as the normalised log ratio of the first compositional part, relative to the geometric mean of the remaining parts within each of the work and leisure time compositions. The work and leisure time behaviours were sequentially rearranged to place each behaviour in the first position once, and the corresponding ilr-coordinate sets were then computed. In this way, the relative importance of each behaviour was sequentially represented in the first ilr-coordinate (ilr1) of a set for subsequent statistical significance testing through regression analysis. A detailed description of how the pivot-coordinates were calculated is provided in Additional file 1.
Using the pivot-coordinates to express the leisure time-use composition as the outcome, we assessed associations with 1) day of the week, 2) type of day, (i.e. workday/non-workday), 3) work duration and 4) the work time-use composition (expressed as pivot-coordinates). The analysis was performed in multiple steps, using multivariate multilevel models. This way, the repeated measurements as well as multiple outcomes (i.e. the two pivot-coordinates to express the leisure time-use composition) for each worker were taken into account. In model 1, day of the week and an interaction between day of the week and type of day (reference = non-workday) were entered as a level-2 predictors. In model 2, the following level-2 predictors were entered: work duration, the work time-use composition, and interaction terms between day of the week and work duration and the work time-use composition, respectively. Both models were adjusted by including the following level-2 predictors (reference in parenthesis for categorical variables): sex (men), smoking-status (smoker), BMI, and age. These covariates were chosen as potential confounders based on previous literature and theoretical assumptions concerning their possible influence on day-to-day pattern of leisure time and work physical behaviours and work duration[13, 17]. A detailed description of model development is provided in Additional file 1.
Model 1 was fitted three times. This was done to isolate the association with one of the leisure time behaviours with respect to the others in the first ilr-coordinate (denoted by ilr1). Moreover, model 2 was fitted six times to investigate the association between each part of the leisure time and work compositions, respectively. Of note, only results of the association between the relative work time spent active and standing (as a proxies of physical work demands) and leisure time physical behaviours are shown. Missing data was not imputed. Results on the association between relative work time spent sedentary and leisure time spent sedentary, standing and active, respectively are shown in Additional file 2.
All analyses were performed in R version 1.1.3, using the compositions and MCMCglmm packages. We used the MCMCglmm package to conduct the multivariate multilevel analysis, following the guide provided by Baldwin et al, by which a Bayesian approach with uninformative priors were used. The assumptions of normality and homoscedasticity of the residuals were assessed for all models by visual inspection of residuals versus predicted values and quantile-quantile plots.
A sensitivity analysis was conducted in which only workers with at least two days of measurements were included (N = 831).