Protective Effect on Mortality of Active Commuting to Work: A Systematic Review and Meta-analysis

Sedentary behaviour is a major risk of mortality. However, data are contradictory regarding the effects of active commuting on mortality. To perform a systematic review and meta-analysis on the effects of active commuting on mortality. The PubMed, Cochrane Library, Embase, and Science Direct databases were searched for studies reporting mortality data and active commuting (walking or cycling) to or from work. We computed meta-analysis stratified on type of mortality, type of commuting, and level of commuting, each with two models (based on fully adjusted estimates of risks, and on crude or less adjusted estimates). 17 studies representing 829,098 workers were included. Using the fully adjusted estimates of risks, active commuting decreased all-cause mortality by 9% (95% confidence intervals 3–15%), and cardiovascular mortality by 15% (3–27%) (p < 0.001). For stratification by type of commuting, walking decreased significantly all-cause mortality by 13% (1–25%), and cycling decreased significantly both all-cause mortality by 21% (11–31%) and cardiovascular mortality by 33% (10–55%) (p < 0.001). For stratification by level of active commuting, only high level decreased all-cause mortality by 11% (3–19%) and both intermediate and high level decreased cardiovascular mortality. Low level did not decrease any type of mortality. Cancer mortality did not decrease with walking or cycling, and the level of active commuting had no effect. Low level walking did not decrease any type of mortality, intermediate level of walking decreased only all-cause mortality by 15% (2–28%), and high level of walking decreased both all-cause and cardiovascular mortality by 19% (8–30%) and by 31% (9–52%), respectively. Both low, intermediate and high intensities of cycling decreased all-cause mortality. Meta-analysis based on crude or less fully adjusted estimates retrieved similar results, with also significant reductions of cancer mortality with cycling (23%, 5–42%), high level of active commuting (14%, 4–24%), and high level of active commuting by walking (16%, 0–32%). Active commuting decreases mainly all-cause and cardiovascular mortality, with a dose–response relationship, especially for walking. Preventive strategies should focus on the benefits of active commuting.


Introduction
Sedentary lifestyle is a major public health problem [1][2][3]. Physical inactivity is one of the leading causes of death in developed countries [4] increasing notably all-cause, cardiovascular and cancer mortality [5,6]. In particular, sedentary transport modes have been described as a factor increasing mortality [7][8][9][10][11][12], despite conflicting results [13,14] and relatively few studies focusing solely on workplace displacement [14][15][16]. Nearly half of the population is working worldwide, i.e., more than 3 billion workers [17]. For available data in some developed countries, less than 10% of workers go to work using active commuting like walking or cycling [14], whereas 50% of workers live within 8 km of their workplace [18,19]. Demonstrating benefits of active commuting on mortality can be of major importance to promote effective preventive strategies based on changing behavior when going to work [20][21][22][23]. Moreover, the benefits of leisure-time physical activity on mortality have been largely demonstrated [2,13,24], without clear comparisons between walking and cycling. There is also a clear dose-response relationship between level of leisure-time physical activity and mortality reduction [25][26][27][28][29]. These relationships remain to be determined for active commuting to or from work, particularly in light of limitations observed in a previous meta-analyses (see Sect. 4.2) [15]. Therefore, we aimed to conduct a systematic review and meta-analysis on the effect of active commuting to or from work on mortality, stratified by type of mortality (all-cause, cardiovascular, cancer), type of active commuting (walking or cycling) and level (low, moderate and high). Moreover, articles quantifying risks report very frequently different models of calculation of the same risk (depending on the method of calculation or adjustment models such as crude or adjusted hazard ratio). Thus, we used two models for each meta-analysis: a model using the most adjusted risks listed in included articles, and a model using only crude or less adjusted risks.

Literature Search
We reviewed all studies reporting a risk of mortality depending on the type of commuting to go to work. The PubMed, Cochrane Library, ScienceDirect and Embase databases were searched on April 30th 2020, using the following keywords: "commuting" or "commute" and "mortality" (details of search strategy are presented in Electronic Supplementary Material Appendix S1). The search was not limited to specific years and no language restrictions were applied. To be included, articles needed to describe our primary outcome variables, i.e., mortality data in relation to active commuting. In addition, reference lists from all publications meeting the inclusion criteria, and from reviews retrieved with our keywords, were manually searched to identify any further studies that were not found using the electronic search. The search strategy is presented in Fig. 1. Two authors (SP and FD) independently conducted all literature searches, collated and reviewed the abstracts and based on the selection criteria, decided the suitability of the articles for inclusion. A third author (BP) was asked to review the articles where consensus on suitability was debated. All authors then reviewed the eligible articles. We followed the guidelines outlined by PRISMA (Electronic Supplementary Material Appendix S2).

Data Collection
The data collected included first author's name, publication year, study design, aims, outcomes of included articles, sample size, number of deaths, mean age, sex (percentage of males), type of mortality (all-cause, cardiovascular, cancer), type of active commuting (walking or cycling), level (low, moderate, high), risks data, type of risk (depending on method of calculation or adjustment model such as crude or adjusted odd ratio), methods of evaluation of active commuting within included studies (such as self-reported questionnaires or interviews), methods of evaluation of mortality within included studies (such as details on death registry), and putative adjustment/explaining factors (such as geographical zone, BMI, smoking, or leisure physical activity).

Quality of Assessment
The Newcastle-Ottawa Scale (NOS) was used to check the quality of included articles [30]. Two authors (SP and VN) evaluated the methodological quality of each included study using eight items with a maximum score of 9 stars as points. The following eight items were assessed in all cohort studies: four items on selection bias (representativeness of the exposed and non-exposed cohort, ascertainment of exposure), one item divided in two sub-items on comparability bias (design and analysis) and three items on outcome bias (assessment of outcome, length and adequacy of followup). Similar items were used to evaluate case control studies. Each item was assigned a judgment of "Yes", "No", "Unclear", or "Not applicable". One point was given for each item within the selection and exposure categories, and a maximum of two points was given for the comparability item involving the control of the most important factor and any specific control for other important factors. Disagreements were addressed by obtaining a consensus with a third author (FD).

Statistical Considerations
Statistical analysis was conducted using Stata software (v15, StataCorp, College Station, US). Baseline characteristics were summarized for each study sample and reported as means ± standard deviations and number (%) for continuous and categorical variables, respectively. Heterogeneity of the study results was evaluated by examining forest plots, confidence intervals (CI) and using formal tests for homogeneity based on the I-squared (I 2 ) statistic. I 2 is easily interpretable and the most common metric to measure the magnitude of between-study heterogeneity. I 2 values range between 0 and 100% and are typically considered low for < 25%, modest for 25-50%, and high for > 50% [31]. For example, a significant heterogeneity may be due to the variability between the characteristics of the studies such as those of workers (age, sex, etc.), type of mortality, type of commuting, or level of commuting. Random effects meta-analyses (DerSimonian and Laird approach) were conducted when data could be pooled [32]. p values less than 0.05 were considered statistically significant. We conducted meta-analysis on the effects of active commuting to or from work on mortality. We stratified these meta-analyses on the type of mortality (all-cause, cardiovascular, and cancer), the type of active commuting (cycling and walking), and the level of active commuting (low, intermediate, and high). Moreover, we took particular care to avoid the inclusion of the same cohort several times in the same analysis, to avoid one cohort being accorded more weight than others. Therefore, for each meta-analysis, we computed two models (a model using the most adjusted risks listed in included articles, and a model using only crude or less adjusted risks). Similarly, for articles that reported hazard ratios for both walking and cycling, using the same control group of non-active commuters, we arbitrarily chose to present data for cycling only, and to compute sensitivity analyses with walking only, to verify the absence of difference for stratification by type of mortality and by level of active commuting. Hazard ratios were centered at one if the mortality of workers with active commuting did not differ from the mortality of workers without active commuting. A hazard ratio > 1 denoted an increased risk of mortality, and a hazard ratio < 1 reflected a decreased risk of mortality with active commuting to or from work. For rigor, funnel plots of these meta-analyses were used to search for potential publication bias. Visual inspection of funnel plot asymmetry was performed to address for possible small-study effect and Egger's test was used to assess publication bias [33]. A value of p < 0.05 was considered statistically significant. In cases of significant publication bias, Duval and Tweedie's trim and fill method was used to allow for the estimation of an adjusted effect size [34]. To verify the strength of the results, further meta-analyses were then conducted excluding studies that were not evenly distributed around the base of the funnel [35]. Similarly, when there may have been overlap of some workers in two studies, we also performed sensitivity analyses using only one cohort. Where possible (sufficient sample size), meta-regressions were proposed to study the relationship between active commuting and mortality and clinically relevant parameters such as the workers' age (continuous data) and sex, the type of mortality, and the type and level of active commuting to or from work (categorical data). Results were expressed as regression coefficients and 95% CI.

Results
An initial search produced 96,124 articles ( Fig. 1). Removal of duplicates and use of the selection criteria reduced the number of articles reporting mortality and data to 17 articles [7-14, 16, 22, 36-42]. All articles were written in English. All studies reported that ethical approval had been obtained.

Quality of Articles
The assessment of the quality of the 17 included studies was performed using the Newcastle Ottawa quality scale, varying from 55.6 [38] to 100% [11], with a mean score of 75.2 ± 10.1. There was a low risk of bias for all data except for the assessment of exposure and outcomes (mostly self-reported) (Fig. 2). To further elucidate risk of bias, we provide details in Sect. 3.2 relating to the representativeness of the population for all studies except three that included only patients with type 2 diabetes [36],  [37], or elderly workers [12], together with study characteristics.

Population
Description of the sociodemographic of the workers included in our meta-analyses was lacking for several of the included articles because these data were part of a larger body of data on physical activity and active commuting in general, and most often only the sociodemographic of the entire cohort were described.
Sample size Population sizes ranged from 2538 [13] to 263,450 [14]. In total, at least 829,098 workers were included in this meta-analysis. Only one study did not provide a detailed number of workers [11]. The prospective cohorts were representative of the general population, with or without diseases, except one study on workers with type 2 diabetes [36], one on workers with hypertension [37], and one on elderly workers [12].
Sex Similarly, sex was described sometimes only in workers, sometimes only for the general population, and sometimes for none of the study participants. Within the ten studies reporting the sex of workers [7,8,13,14,16,37,40], a total of 342 891 men were included compared to 339,350 women workers. Two studies described sex only for the general population [11,13] and five studies [9,12,36,38,39] did not specify the proportion of men or women both in workers or in the general population. All studies included both men and women except one study which included only women [11] and one only men [16].

Aims and Outcomes of Included Articles
In our meta-analyses, we included only mortality data in relationship to active commuting (walking, cycling) to or from work. However, the aim of the majority of the included studies was to examine the relationship between mortality and domain-specific physical activity (i.e., leisure time physical activity [7-9, 12, 13, 36, 37, 40], activity at work [7,9,13,40], occupational physical activity [8,12,36,37,39], exercise or sports [7,[9][10][11]40], active commuting in general [10-13, 22, 38, 39]; data on mortality and transportation to and from work were also reported [7-14, 16, 36-40]. Only three studies focused on the mode of commuting to or from work [14,16,41]. One study aimed to examine the individual, combined and isolated effects of movement-based behavior [38], and one study aimed to determine whether bicycle commuting was associated with a risk of injury [42].

Evaluation of Type of Mortality within Included Studies
The majority of the studies described the three types of mortality [9, 11-14, 16, 22, 39, 40]. Three studies described only all-cause and cardiovascular mortality [8,10,36], one described only data for cardiovascular mortality [37], and two studies described only data for all-cause mortality [41,42].

Evaluation of Mortality within Included Studies
All studies retrieved mortality from the death registry obtained from health authorities, mostly from nationwide databases: the nationwide death register of Finland [8], National Health Service Central Registry of the UK [16], National Patient Registry of Denmark [9,41], UK Office for National Statistics (ONS) [10,39], National Health Service (NHS) of England, Wales and Scotland [14,22,42], Statistics Finland [36,37], Social Security Administration and Medicare/Medicaid of USA [38], National Vital Statistics Database from the Ministry of Health, Labour and Welfare of Japan [12], and the Swiss National Cohort (SNC) [40]. Local health authorities databases were used in one study [13], and the imprecise "vital statistics and cancer registry" in one study [11]. All the databases used the International Classification of Diseases (ICD) for reporting causes of mortality: the 8th revision [16], the 9th revision [11,13], the 10th revision [12,14,22,42], the 8th and 9th revision [9], the 9th and 10th revision [39], or the 8th, 9th, and 10th revisions [8,36,37,40]. Even if not mentioned within the articles, the three other studies also used the ICD for type of mortality-data from ONS, data from health services of the USA, and data from The National Patient Register and National Diabetes Registry of Denmark [10,38,41]. One study did not state how mortality and cause of mortality were retrieved [7].  Fig  S10-S12). Despite some significant results on the effects of active commuting using cycling, the number of studies included precluded robust conclusions for cardiovascular and cancer mortality.

Publication Bias and Sensitivity Analyses
Funnel plots (metafunnels) of meta-analyses (Sect. 3.8) used for analysing potential publication bias are presented in Electronic Supplementary Material Fig S13. Metaanalyses reperformed after the exclusion of studies that were not evenly distributed around the base of the funnel showed similar results (data not shown). Egger's test was non-significant. To further address sensitivity analyses, we also reperformed the aforementioned meta-analyses using crude or less adjusted hazard ratio (meta-analysis stratified by type of mortality: Electronic Supplementary Material Figs S14 and S15; meta-analysis stratified by type of commuting: Electronic Supplementary Material Fig S14, and S16-S18; meta-analysis stratified by level of active commuting: Electronic Supplementary Material  Fig S28). In addition to all previous significant results using fully adjusted measures of risks, meta-analysis based on crude or less fully adjusted estimates retrieved similar results, with also a significant reduction of cancer mortality with cycling (hazard ratio = 0.77, 95% CI 0.56-0.95), high level of active commuting (0.86, 95% CI 0.76-0.96), and high level of active commuting by walking (0.84, 95% CI 0.68-1.00). Furthermore, sensitivity analyses using only walking data for studies reporting both hazard ratio for walking and hazard ratio for cycling, using the same control group of non-active commuters, showed similar results (data not shown). Lastly, excluding alternatively one of the two studies [14,42] that may had some degree of overlapping between their populations showed similar results (data not shown). Fig. 3 Summary of meta-analysis showing the effects of active commuting by type of mortality (all cause, cardiovascular and cancer mortality), by type of active commuting (walking, cycling or both), by level of active commuting, by level of active commuting by walking, and by level of active commuting by cycling, based on data from adjusted hazard ratio

Metaregressions
There were no effects of age or sex on the reduction of mortality from active commuting. Metaregressions also failed to demonstrate best outcomes depending on type of mortality (all-cause, cardiovascular, or cancer mortality) or level of active commuting (

Discussion
The main findings were that active commuting to or from work decrease mainly all-cause and cardiovascular mortality, with a dose-response relationship, especially for walking.

Benefits of Active Commuting on Mortality
Sedentary lifestyle is a major public health problem [3]. Sedentary behavior in general increases the risk of sleep disturbance [47] and poor mental health [48], cardiovascular and metabolic diseases [5,6,49], incidence of colon, breast, and endometrial cancer [50], and mortality (all-cause, cardiovascular, cancer) [3,6,51,52]. To promote health benefits, national and international campaigns encourage a decrease in overall sedentary behavior by incorporating more physical activity in daily life, especially during commuting [4].
In 1995, the US Centers for Disease Control and Prevention and the American College of Sports Medicine recommended a moderate amount of physical activity (e.g., 30 min of brisk  walking) preferably every day [53]. Despite existing data regarding causes of mortality due to sedentary behavior [6], the putative effects of active commuting to and from work on mortality have been poorly studied [14][15][16]. We demonstrated strong benefits on mortality of active commuting. Considering that most workers used their car for short distances to go to work [18,19], active commuting could provide a significant change with strong public health benefits [20,23]. Furthermore, commuting by walking or cycling has a negligible cost compared to driving or public transport. Moreover, active commuting could partly reduce the most problematic global warming issue. Considering a mean petrol consumption of 5 L per 100 km by car, there are 120-130 g of CO 2 emitted in the atmosphere per kilometer [54]. Taking into account the number of workers using a car for less than 5 km to commute to and from work this practice can produce billions of tons of CO 2 emitted per year [55]. Therefore, in addition to health benefits, active commuting should also be promoted as an ecological target [56,57].

All-Cause and Cardiovascular Mortality are decreased by Active Commuting
The positive effect of leisure time physical activity on allcause, cardiovascular and cancer mortality was largely demonstrated [1,24,40], with conflicting results for active commuting to and from work. We demonstrated that active commuting to or from work decreased mainly all-cause and cardiovascular mortality whatever model was used, and with significant benefits for cancer mortality with crude or less adjusted meta-analyses. Our results do not agree with the previous meta-analysis which reported that active commuting only decreased all-cause mortality but not cardiovascular and cancer mortality [15]. A possible explanation for this may relate to that our search strategy being more rigorous, i.e., we included from 22 to 45% more articles (16, 11 and 7 articles instead of 11, 9 and 5 for all-cause mortality, cardiovascular and cancer mortality, respectively). Moreover, our meta-analysis was more rigorous by computing two models: one based on fully adjusted hazard ratio, and one based on crude or less adjusted hazard ratio. All models gave significant findings for all-cause and cardiovascular mortality. Finally, we also detailed a dose response-relationship considering the level of physical activity during active commuting.

Level of Active Commuting: A Dose Response Relationship?
A large body of evidence suggests an independent dose-response relationship between the level of leisure time, physical activity and mortality [9,58,59]. We also demonstrated that the highest intensities of active commuting in general had the strongest benefits, with also a dose-response relationship for active commuting by walking. Low, intermediate and high active commuting had a reduction of 4, 5, and 11%, respectively, for all-cause mortality. Interestingly, for cardiovascular mortality, we showed that even a low level tended to decrease mortality, in line with the literature showing that small amounts of physical activity (min) can also reduce mortality [26]. Consequently, going to work actively, even at a low level of activity during short periods, may have strong health benefits. Secondly, it can be noted that the decrease in cardiovascular mortality seemed to follow a U-shape, with stronger benefits for intermediate intensities.
Although physical activity at work is confounded by socioeconomic status, this could be consistent with the literature showing that physical activity at work can be detrimental to life expectancy [60,61]. Similarly, intense physical activity at work can produce anxiety for workers, despite leisure time physical activity having benefits on mental health which also display a dose-response relationship [62]. Whatever the mode of active commuting to and from work, it might follow the same logic with high intensities not necessarily having the strongest benefits-even in the model using crude or less adjusted hazard ratio.

Walking and Cycling for Active Commuting
Walking at a brisk/fast pace demonstrated a stronger reduction of all-cause, cardiovascular or cancer mortality compared with walking at a slower pace [63], and we report for the first-time similar benefits of walking (of up to 41%, depending on level of active commuting and type of mortality) during active commuting to or from work. Being bipedal is an intrinsic characteristic within the evolution of human beings [64]. Furthermore, walking is free [65] and can provide social connections [66]. Moreover, walking has strong psychological and mental health benefits [67]. Being more relaxed when arriving at work may also have economic advantages for companies [68]. Reduction of mortality is also related to active commuting by cycling (up to 33%). We also demonstrated a dose-response relationship for walking, but the reduction of mortality provided by active commuting by cycling follows more of a U shape. However, we can note that when both walking and cycling were assessed in the same study [14], adding cycling to walking seemed more protective. Since there is a 'win-win' relationship for workers and companies whatever the type of active commuting (a healthy worker is a more profitable worker) [69], it would be necessary to encourage the installation of showers in the workplace to promote active commuting in general, and to permit active commuting at a high level for those who are willing to cycle or walk to work. Infrastructure should also include other facilities such as changing rooms, lockers or bicycle garages. Some companies and government departments have experimented with allocation of incentives/bonuses to workers actively commuting to and from the workplace [70].

Limitations and Strengths
Our review has some limitations. A limited number of studies were included. However, we included a large sample size of 829,098 workers, promoting the generalisability of our results. Our meta-analysis inherited the limitations of the 17 individual studies that contributed to the study and met the inclusion criteria. Therefore, the review was subject to the bias of included studies such as the bias of exposure ascertainment and the bias of outcome assessment. Even if all individual studies included in our meta-analyses were excellent and published in very well-known high-quality journals, description of active commuting was mainly selfreported, limiting the quality of the reported data. Conditions of fulfilment (e.g., at home or at work, in occupational medicine departments) were not always reported, which has been shown to influence results from self-reported questionnaires [71]. Similarly, the level of physical activity was also self-reported by questionnaires. Ideally, a rigorous analysis of intensity should have been adjusted for fitness [72][73][74], but measuring heart rate during exercise and performing maximal oxygen uptake (VO 2max ) tests is deemed impossible for such large-scale cohort studies. While most studies attempted to assess dose response effects, our groups of levels of physical activity were heterogeneous because of the variability of the parameters retrieved within each included study (intensity [13,22], duration per day [8,9,16,36,37] or week [10,39], MET-hour/week [11,41,42], or frequency [12]). However, mortality was assessed mainly with the use of national mortality databases guaranteeing accuracy of data. All included studies had a prospective cohort design. Though there were similarities between the populations' characteristics within each included article, they were not identical. Although data were collected in several countries, emerging countries are not represented. Moreover, the majority of studies were conducted in Europe [7-10, 13, 14, 16, 36, 37, 39, 40], with only one study from the USA [38], one from Japan [12] and one from China [11]. However, all ethnicities were included [14]. The generalisability of our results may also suffer from some single-site studies [7,9,10,12,16,39,40], even if half of the included studies were multi-site [8,11,13,14,[36][37][38]. We computed sensitivity analyses (data not shown) demonstrating similar results whatever the study design. Even if having more effect sizes within each meta-analysis than constituent studies that can be found in several articles published in high ranked journals [75,76], our meta-analysis followed a rigorous methodology and only once did we include the same population in our meta-analysis. We manipulated the different data for the same outcome depending on the adjustment model by computing a meta-analysis using only fully adjusted estimates of risks, and we computed a sensitivity analyses based on crude or less adjusted risks. Therefore, the weight of studies did not require further consideration [77]. Moreover, Egger's test was non-significant highlighting the absence of major publication bias. A major limitation of our meta-analysis is the lack of data on characteristics of workers because several studies were part of larger cohorts evaluating the effects of physical activity in general. However, the fact that most included studies assessed leisure physical activity to adjust risk calculation, limits bias of confusion. This was the case even if leisure physical activity was assessed on self-reported questionnaires-as well as for active commuting. Most studies adjusted risk calculations based on adequate models considering several putative confounding factors and included workers representative of the general population. However, we noted that no article adjusted risk calculation for pre-existing illness. Causality should also be discussed as a limitation in the context of observational studies. Another limitation is the absence of significant results from meta-regressions.

Conclusion
We demonstrated that active commuting decreases mainly all-cause and cardiovascular mortality, with a dose response relationship, especially for walking. Preventive strategies should focus on the benefits of active commuting.