Does high occupational physical activity hamper the beneficial health effects of leisure time physical activity? Evidence of the physical activity health paradox from a prospective study on compositional accelerometry data and long-term sickness absence

Background: The ‘physical activity health paradox’ advocates that leisure physical activity (PA) promotes health while occupational PA impairs health. However, this paradox can be explained by methodological limitations of the previous studies—self-reported PA measures, socioeconomic confounding or not addressing the compositional nature of PA. Therefore, this study investigated the association between compositions of accelerometer-based moderate to vigorous PA (MVPA) time at work and leisure and onset of long-term sickness absence (LTSA). Methods: Time spent on MVPA and remaining physical behaviours (sedentary behaviour, standing, light PA and time in bed) at work and leisure was measured for 929 workers using thigh accelerometry and expressed as isometric log-ratios (ilrs). LTSA was register-based events of ≥6 consecutive weeks during 4-year follow-up. The association between ilrs and LTSA was analysed using a Cox proportional hazards model adjusted for remaining physical behaviours and potential confounders, then separately adjusting for and stratifying on education and type of work. Results: During the follow-up, 21% workers experienced LTSA. During leisure, more relative MVPA time was negatively associated with LTSA (20% lower risk with 20 minutes higher MVPA, P=0.02). At work, more relative MVPA time was positively associated with LTSA (15% higher risk with 20 minutes higher MVPA, P=0.02). Beneficial association between MVPA at leisure and LTSA was only observed for the lowest tertile of MVPA at work (P=0.03). Results remained unchanged when adjusted for or stratified on education and type of work. Conclusion: These findings provide further support to the ‘PA health paradox’.


Background
Physical activity (PA) prevents chronic diseases and mortality (1). However, research indicating the health benefits of PA is predominantly limited to the leisure domain (2).
Adults engage in PA at work -a domain where individuals spend half of their awake time.
However, there is no consistent documentation of a beneficial health effect of occupational PA (OPA) (3)(4)(5)(6). Rather, a recent meta-analysis of almost 200,000 participants observed an increased risk of all-cause mortality among males with high OPA (7). This potential contrasting health effect of PA during leisure and work domains is termed 'the physical activity paradox', which has recently received extensive attention in the field of PA and health (8,9).
In particular, researchers have suggested that the PA paradox is merely a result of methodological limitations of existing studies (9). One such limitation lies in the measurements of PA, like the use of questionnaires that has been found to be imprecise and potentially biased (10,11). Besides this, existing prospective studies on the PA paradox have disregarded the compositional nature of time-use data like PA (12)(13)(14)(15). The compositional nature of PA data means that the longer time spent on a specific PA, such as moderate-to-vigorous PA (MVPA), will consequentially require less time spent on other physical behaviors, such as light PA (LIPA), sedentary behavior or sleep. To counter this challenge, the time-use data on PA should be analysed using a Compositional Analysis (CoDA) approach (12)(13)(14)16). Another limitation is the potentially inadequate adjustments for socioeconomic status (SES) confounding, where analyses of homogeneous groups with respect to socioeconomic characteristics are preferable (9).
The PA paradox has been shown to be associated with long-term sickness absence (LTSA) -an established predictor of all-cause mortality (17), chronic disease (18), and early exit from the labour market (19)(20)(21)(22) with considerable economic burdens on companies and society (23,24). Studies have shown that high levels of OPA increases risk of prospective LTSA (3,25) while high levels of leisure time PA decreases this risk (3). However, the present study addresses, for the first time, the previous limitations of these studies by using better measurement methods of PA at work and leisure, addressing the compositional nature of PA data and adjusting for SES confounding.

Data and study population
The present study is based on the prospective data from the 'technically measured compositional Physical wOrk DEmands and Prospective register-based Sickness Absence study (PODESA) cohort (15). This cohort was formed by harmonizing two cohorts, the 'Danish Physical ACTivity cohort with Objective measurements' (DPhacto) (26) and the 'New method for Objective Measurements of physical Activity in Daily living' (NOMAD) cohort (27). Both cohorts used similar procedures of 24-hour time accelerometry and comprised mainly blue-collar workers in Denmark, enabling the harmonization. More details on the setting, locations, recruitment, and inclusion and exclusion criteria in these cohorts and on the harmonizing procedures can be found elsewhere (15). 6 Accelerometry at work and leisure Workers wore a thigh-based triaxial ActiGraph accelerometer (GT3X+, Florida, U.S.A) for 24 hours for up to five workdays (27,28). Simultaneously, during those five days, workers also filled-in a diary reporting their time of starting and ending work and going to and out of the bed each day. The accelerometry data were downloaded using Actilife Software version 5.5 (29) and further processed using a MATLAB program Acti4 (30,31). Acti4 has previously shown a high sensitivity and specificity in detecting PA at work and leisure (32). Acti4 was used to determine time spent sedentary (sitting and/or lying), standing still, moving (standing with slight movements), walking slow (< 100 steps per min) and fast (≥ 100 steps per min), running, cycling and stair climbing (32). For the analysis, time spent moving and slow walking was merged to calculate low-intensity physical activity (LIPA) while time spent on fast walking, stair climbing and running was merged to calculate moderate to vigorous physical activity (MVPA) (33). Leisure MVPA also included cycling time (33). Diary-based information was used to determine time in bed -a period between going to and out of the bed that were further visually checked for verification in the Acti4. A work period was defined as self-reported working hours spent on primary occupation while leisure period was defined as non-work periods, excluding time in bed.
All non-work days and accelerometry non-wear periods were excluded. Workers who had at least one day with valid work and leisure period were involved in further analyses. A work or a leisure period was considered valid if it comprised ≥ 4 h of wear time or ≥ 75% of the average wear-time across days, respectively (16,26,34,35). Prospective register-based long-term sickness absence Four-year prospective data on LTSA was retrieved from the DREAM register (37,38). This register contains weekly information on granted subsidised sickness absence for each individual in Denmark. The sickness absence compensation is given to the employer who can claim a refund from the state after 30 days of sickness absence. Therefore, DREAM contains information on sickness absence periods of ≥ 5 consecutive weeks. LTSA was defined as the occurrence of the first (if any) ≥ 6 consecutive weeks of sickness absence period during the 4-year follow-up from the date of completing the baseline measurements. We selected this cut off point based on previous research (39). Previous research has shown the accuracy of DREAM register-based LTSA (40).

Potential confounders
Potential confounders included age, sex, Body Mass Index (BMI), smoking status, duration of occupational lifting and carrying, and education and type of work as proxy indicators of socioeconomic status (SES). Age was determined using workers' unique civil registration number. Sex of the workers was determined using single item 'are you male or female'?
Workers' height and weight were objectively measured to determine their BMI (kg/m 2 ).
Smoking status was determined using a single item with response categories summarized to smokers (smoking daily or sometimes) and non-smokers (ex-smokers and never smoked). Occupational lifting and carrying duration was determined using a single item with 6 responses ranging from 'almost all the time' to 'never' (33). The information on workers' education and type of work was included as indicators of SES (41)(42)(43). The education of the workers was determined using a single item 'are you skilled or unskilled?' The information on type of work was collected using single item 'are you a worker engaged in administrative work tasks [white collar] or in production [blue-collar]'. Later, the information on these two measures was summarized in three categories -white-collar, blue-collar skilled, and blue-collar unskilled.
The data were analyzed in the CODA approach (46). First, the four-part time composition of work (MVPA, sedentary, standing, and LIPA) and five-part time composition of leisure (MVPA, sedentary, standing, LIPA, and time in bed) were expressed as isometric log-ratios (ilrs). The first ilr coordinate for the work and leisure composition represents time spent on MVPA relative to the geometric mean of remaining behaviors. In subsequent ilrs, the numerator of the first ilr was further split to create remaining ilrs (47) The Cox proportional hazards regression model was then fitted by maximizing the partial likelihood function to the ilrs (i.e. the log-transformed work and leisure compositions) and the onset of LTSA (48). We censored 47 workers who were pregnant or going on maternity leave within 8 months during the study period, or either emigrated, became deceased, entered retirement or early retirement. The Cox regression model was adjusted for age, sex, BMI, smoking status, occupational lifting/carrying duration, and MVPA and other physical behaviors in the mutual domain (set of ilrs at work and leisure were entered together in the model). The assumptions of proportional hazards were met when tested by visual inspection and using the Grambsch-Therneau test (49). The model coefficient associated with ilrs were assessed using Wald test statistics (z) and associated significance (P), considering P < 0.05 statistically significant.
Due to many missing in SES data (three categories: white-collar, blue-collar-skilled, and blue-collar-unskilled; missing = 118), we performed separate analyses to test if the main results were independent of SES. Therefore, we performed two separate analyses without and with additional adjustment for SES on 811 workers, who had data on SES. We also stratified the analyses on the three categories of SES.

Effect size interpretation
To interpret the strength of the association, procedures explained in previous studies was used (33). First, sample compositional mean of all physical behaviours at work and leisure was calculated (Table 1) To test the combined association of relative MVPA at work and leisure and sickness absence, we stratified the population based on the tertiles (rounded to the closest bound) of the work MVPA. The Cox proportional hazards regression model, as explained above, was re-fitted separately on each of these sub-groups. Thereafter predicted new HRs corresponding to 20, 40, 60 minutes of relative MVPA at leisure were plotted for the subgroups with low, moderate and high MVPA at work.
The sample size in this manuscript was derived from the a priori calculation made for the PODESA project. Based on a statistical significance of 0.05, LTSA prevalence of 20%, and an effect size corresponding to a 7% (or higher) change in HR per 10 minutes of MVPA, we require 850 workers for the analyses to achieve a statistical power of 80%.

Results
Out of the 2,498 eligible participants, 929 (37%) workers had sufficient data to be involved in the analyses. A detailed flow chart is shown in Fig. 1. Table 1 shows the descriptive of the workers involved in the analysis. The participants were on average 45 years old with a BMI of 27 kg/m 2 , 55% were men, and 30% smoked.   Figure 2 shows that, for example, reallocating 20 minutes to MVPA at work from the remaining work behaviors was associated with ≈ 15% higher risk of LTSA while reallocating 20 minutes to MVPA at leisure from remaining leisure behaviors was associated with ≈ 20% lower risk of LTSA.  Nevertheless, the overall finding of an increased risk for future LTSA with higher levels of work PA is in line with the some studies based on self-reports (3,7). The potential mechanism behind our finding could be that work MVPA is influenced by different constraints and comprise different characteristics than leisure PA (56). Work PA is performed mainly to complete working tasks and compared with leisure, there is limited possibility of tailoring the duration, intensity, and variation of the PA according to the individual needs and preferences. Because of these constraints, the work PA can lead to excessive exertion and fatigue without sufficient possibility for recovery (57), which over time can increase risk of impaired health and LTSA (58,59).
We also observed that our results were robust when the analyses were adjusted for SES indicators. Studies testing the PA paradox have been criticized for not adjusting for or removing the socioeconomic confounding (9). To address this limitation, we performed the analyses without and with adjustment for a proxy measure of SES (three categories: whitecollar, blue-collar-skilled, and blue-collar-unskilled) and even stratified the analyses on these categories. We still observed the PA paradox even after these adjustments and stratifications based on SES, confirming that PA paradox exist independent of SES of workers.
When we stratified the workers in three groups, low, moderate and high work MVPA, the negative association between relative leisure MVPA and LTSA was clear among those with low work MVPA. This negative association was weak and non-significant among those with low and moderate work MVPA. These results show that leisure MVPA is of particular importance among those with low work MVPA, for example administrative workers or the drivers. Conversely, too much work MVPA time seem to reduce the beneficial effect of leisure MVPA.

Strengths and limitations
The main strengths of the study are the thigh-worn accelerometery-based physical behaviors data, shown to be highly reliable and valid (32,60). Another strength is the CoDA approach applied for the analyses of the study, handling the compositional structure of time-use data of PA (46,61). Additionally, this study adjusted for remaining physical behaviors (sedentary behaviours, standing and LIPA and time in bed) within 24 hours.
Another strength was the usage of national register data with valid prospective measures of LTSA (37). Finally, the opportunity to remove/adjust for possible SES confounding when testing the PA paradox supports our main findings.  (27). All participants in the NOMAD and DPhacto studies received written and oral information about the projects, the practicalities of participating, potential risks of participating and having the possibility of withdrawal from the project without giving a specific reason with sufficient time for considerations of their participation. The persons agreeing to participate gave a written consent to participate in the study and the use of the data for research studies.

Availability of data and materials
The fully anonymized data from the baseline in NOMAD and DPHACTO from each participant involved in the main analysis of this study are available in a Danish public The fully anonymized data on prospective long-term sickness absence is available upon request from statistics Denmark (A Central Authority on Danish Statistics) (62).

Competing interests
The Danish Working Environment Research Fund had no role in study design, collection, analysis, data interpretation, manuscript writing, or decision to submit the manuscript for publication. The authors declare no conflicts of interest.

Funding
The PODESA study received funding from the Danish Working Environment Research Fund

Supplementary Files
This is a list of supplementary files associated with the primary manuscript. Click to download.