Research question 1

What is the reduced risk of all-cause mortality from an exposure of (i) walking and (ii) cycling equivalent to physical activity recommendations (11.25 MET.hours per week)? (Population: identified cohort participants, healthy at baseline; Intervention: exposure to regular walking or cycling above referent level; Comparison: exposure to regular walking or cycling at referent (lowest) level; Outcome: all-cause mortality during follow up (controlled for other PA); Study design: prospective cohort studies).

Research question 2

What is the shape of the dose-response relationship for reduction in risk of all-cause mortality across the reported ranges of (i) walking and (ii) cycling exposures?

Systematic review

Meta-analysis

Dose-response relationship

We chose to investigate the dose-response relationship through meta-analysis of pooled relative risks within three exposure intervals. A similar approach has been previously demonstrated for all PA and mortality [13]. Selection of the interval ranges (exposure categories) was based on the distribution of the reported walking and cycling exposures to give sufficient results in each category to allow meta-analysis. For walking, the categories for exposures in MET.hours per week were [1 to 16], [16 to 30] and [30 to 70]. For cycling the categories were [1 to 24], [24 to 40] and [40 to 90].

We then conducted separate random effects meta-analyses for each category using the risk reductions for exposures within the category range. When there was more than one exposure group from the same study within a single category we pooled the RR using the method proposed by Hamling et al. [14]. This method requires group size and number of events for each category. Included study authors were contacted by email to request missing data; studies were excluded from this analysis when these data were not available. When assessing the dose response relationship we were able to include results from 11/14 walking studies (79%) and 6/7 cycling studies (86%).

From these results we calculated the incremental change in risk for each MET.hour per week increase in either walking or cycling for the different regions of the exposure range. Finally the results of the separate meta-analyses were plotted against the midpoint of the respective exposure category. Confidence intervals were also plotted for each data point.

The analyses approaches are summarised in the analysis flow chart (Additional file 1: Figure S1).

Study protocol

The full study protocol was registered with PROSPERO (International database of prospectively registered systematic reviews in health and social care) and is available at http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID=CRD42013004266.

Ethical approval

Ethical approval was not required for this secondary analysis of cohort studies.

Systematic review

A total of 8,901 titles were identified from databases, and 427 were retained after first appraisal with 4 further records identified from other sources. After second appraisal 49 full texts were retained for detailed evaluation with 18 meeting all inclusion criteria.

A number of identified walking studies were excluded from analysis after detailed evaluation: 18 did not adjust or control for other PA, 3 did not present ACM risk reductions (e.g. reported survival function), 3 did not report “walking only” on closer inspection (e.g. exposure was walking and standing), 3 reported walking by pace only preventing assessment of dose and 3 analysed the same participants as an already included higher quality study. We also excluded a number of studies that used walking as a dimension of PA (e.g. Low = no PA, Medium = Walking at least 3 hours per week, High = Sports and exercise for 3 hours per week) as these did not allow assessment of the adjusted effect of walking. One small cycling study in older adults from the Netherlands matched the criteria for cycling and mortality, but was excluded as it did not control for other PA [15].

From the 18 included studies, 14 studies presented a total of 18 results on walking and ACM, and 7 studies presented a total of 8 results on cycling and ACM. These data are displayed in the study flow chart (Figure 1). Additional file 2: Table S1 shows the full data extraction for all included studies. Additional file 3: Table S2 shows the NOS quality scores.

Description of studies

Exposure conversions and point estimates

Once converted to point estimates at an exposure of 11.25 MET.hours per week, 3 of the 18 walking results (17%) were statistically significant based on our extracted confidence intervals. For cycling, 4 of the 8 results (50%) were significant at 11.25 MET.hours per week. There were no obvious patterns or differences between the results that were statistically significant and non-significant in terms of mean age, follow-up length, exposure ranges or study quality.

Meta-analysis

Random effects meta-analysis of the point estimates for the 18 results from 14 walking studies found a reduced risk of 0.89 (95% CI = 0.83 to 0.96) corresponding to an exposure of 11.25 MET.hours per week (Figure 2). Random effects meta-analysis of point estimates for the 8 results from 7 cycling studies found a risk-reduction of 0.90 (95% CI = 0.87-0.94) corresponding to an exposure to 11.25 MET.hours per week (Figure 3).

Sensitivity analyses on the selection of intensity for conversion of walking and cycling exposures were carried out and the calculations were found to be robust. For example, conducting the meta-analyses using an average intensity of 3.0 METs (instead of 4.0 METs) for walking or 8.0 METs (instead of 6.8 METs) for cycling for the conversions made a 1 percentage point difference to the final meta-analyses results.

Heterogeneity in study results was assessed through examination of the forest plots and quantified using the I² statistic to determine whether the observed variation in the study results was compatible with the variation expected by chance alone [12]. We used a random-effects model as the default to incorporate heterogeneity between studies. I² for walking was 82.6% (P < 0.05) and for cycling was 20.5% (p = 0.27) suggesting considerable heterogeneity.

We plotted trial effect against standard error and presented it as a funnel plot (Additional file 4: Figure S2). Asymmetry could be caused by a relationship between effect size and sample size or by publication bias, and we also examined any observed effect for heterogeneity [34]. We found no evidence of publication bias.

The study by Bath et al. [20] was a considerable outlier on the forest and funnel plots. By a substantial amount, this study had the greatest risk reduction at 11.25 MET.hours per week (0.40 (95% CI = 0.34 to 0.48); see Table 2). Repeating the analysis with Bath excluded did not change the pooled estimate for walking.

Dose-response relationship

Figures 4 and 5 show these same data as dose-response plots for walking and cycling across the ranges of reported exposures. Walking shows a slowing of the rate of benefit after 8 MET.hours per week (equivalent to 120 minutes per week at an intensity of 4.0 METs). The rate may increase slightly after 22.5 MET.hours per week, but the large confidence interval on the final data point suggests large uncertainty. The dose-response relationship for cycling shows the rate of received benefits drops in each subsequent category.

Strengths and limitations

The estimates for walking are based on 280000 participants and 2.6 million person-years and for cycling they are based on 187000 individuals and 2.1 million person-years. Further, the included studies were all high quality as assessed by the Newcastle-Ottawa scale. We only included studies that controlled for other domains of PA. This means the risk reductions calculated are more likely to be the result of walking or cycling exposure, independent of a confounding effect of other PA. However, the different ways and extents to which the studies assessed other PA prior to adjustment is a limitation to this method. The broad range of reported exposures also allowed us to investigate the dose-response relationships for walking and cycling across the distributions typically observed in general populations.

Included studies were predominantly from developed countries. It is not clear how generalizable the results would be to low and middle income countries where environmental or social factors might be different (e.g. road safety, air pollution, or prevalence of PA from other sources). While there was no evidence of publication bias in the funnel plots, we cannot completely rule out this possibility. Many studies were from large cohorts not specifically designed to test walking or cycling; it is possible other similar cohort studies finding non-significant or non-beneficial results did not publish their findings. The I² statistic suggested considerable heterogeneity, especially for walking. This could be due to differences in exposure categorisation or sample bias, or due to residual confounding. The Bath study was a considerable outlier. It also had the one of the crudest categorisations of walking exposure (more or less than 10 minutes per day) which led to a large risk reduction when extrapolated to 11.25 MET.hours per week.

To convert the exposures to the same metric we had to make certain assumptions about intensity of behaviours and width of category ranges. These assumptions are needed as epidemiological studies on walking and cycling and ACM rarely capture distributions of speeds and intensities [19],[31]. The intensity assumptions were reasonable and transparent; they are based on the widely used Compendium of Activities and the assumptions about mid-points of open ended ranges were necessary for our chosen approach. The results were shown to be robust to intensity selections. However, we must acknowledge that while required for our method, these assumptions and conversions may introduce some error. Further, at a population level walking and cycling will be described by a wide range of intensities based on individual fitness, journey purpose and terrain. For the meta-analysis of pooled relative risks, the choice of exposure categories is to some extent subjective and this is a known limitation of this approach.

The conversion of exposures to MET.hours per week and calculation of point estimates for risk is justified as this approach quantifies the exposure required for a given effect. This allows the reader to consider how much walking or cycling is required for a given effect and is a strength of our study.

As with many epidemiological studies, the measures of walking and cycling were often varied and crudely defined. In all cases they were self-reported. Our assumptions are unlikely to have added substantial further uncertainty; the exposure data allow us to estimate broad exposure to walking or cycling and the assumptions do not detract from this ability. Recent studies have also suggested that self-reported walking and cycling may be relatively valid at the group level [35]. We are confident that the findings are robust with regards to the methodological parameters available and selected for this analysis. The consistency of results across the studies, despite using different exposure levels and taking place in different walking and cycling environments support the overall conclusion of a beneficial effect of these forms of physical activity.

Comparison to previous literature

For walking a previous meta-analysis found a relative risk for ACM of 0.68 (95% CI = 0.59 to 0.78) [5]. However this study did not specify or treat exposure in quantitative terms and thus it is not known how much walking confers this benefit. A meta-analysis of 13 studies by Samitz et al., reported a smaller reduced risk of 0.97 (95% CI = 0.95 to 0.99) for an increment one hour per week of walking [2]. This is a lower exposure than we have used to calculate our risk reduction; extrapolated to our exposure this risk would be 0.92 which is 27% smaller than the magnitude of our finding. Neither analysis considered whether included studies had controlled for other PA.

A more recent meta-analysis of walking fitted a 0.375 power transformation to the exposure and standardised at 11 MET.hours per week. The authors reported a relative risk of 0.89 (95% CI = 0.82 to 0.96) however they were only able to include five cohort studies (1.6 million person years) in this analysis [3]. They also included one study in which the exposure was a combination of walking and standing time [36]. Our results confirm this finding in a more robust analysis with new studies, more studies (all graded high quality) and covering 2.6 million person years.

A meta-analysis of cycling has not previously been possible due to the low number of published studies though Oja et al., and Saunders et al., have previously reviewed benefits for reduced morbidity from cycling [4],[8]. This meta-analysis is helpful as individual studies have reported relative risks as low as 0.44 (95% CI = 0.40 to 1.07) [31] and as high as 1.02 (95% CI = 0.77 to 1.35) [32]. A meta-analysis of leisure-time physical activity (which might be considered similar to cycling) by Samitz et al., reported a risk reduction of 0.94 (95% CI = 0.92-0.97) for an increment one hour of physical activity per week [2]. This analysis was based on seven studies, which did not all adjust for other PA. Extrapolated to our exposure this risk would be 0.90 which agrees with our finding.

It is now widely theorized that there is a non-linear dose-response relationship between various physical activities and reduced risk for all-cause mortality [3]. Woodcock presented a 0.375 power transform for walking and a 0.25 transform for non-vigorous PA. Oja et al. found evidence of an inverse step-wise decrease and a “ceiling effect” for general health benefit through increased cycling [4]. We chose to analyse the possible non-linear dose-response relationship through meta-analysis of pooled relative risks within three exposure intervals. This allowed us to better reflect the reported physiological effects at low and high exposures and the non-linear nature of the relationship between walking or cycling and ACM across the broad range of exposures. Another approach for modelling the quantitative exposure-response relationship is the use of cubic splines [37],[38] but many of our studies had dichotomous exposure variables, meaning they could not be included in such analysis and precluding this more advanced tool [37].

Implications

Previous walking analyses presented researchers with a wide range (3 to 32%) of possible risk reductions. With our results, researchers can now model effects of walking and cycling, and compare to other risk factors, with greater confidence. Previous cycling modelling had to rely on results from single studies [39] and this new analysis will provide a better (and more conservative) estimate for researchers and modellers. We also present dose-response relationships for considering benefits across walking and cycling activity levels when before and after data are available to model change from different baseline exposures. Combined the results presented in this paper will provide a stronger information base for health impact assessment, economic analysis and active travel policy.

These results confirm that public health policy and recommendations should include walking and cycling. Walking in particular is an important behaviour as it is free, accessible and requires no equipment or training; it has been described as the nearest thing to “perfect exercise” [40]. Walking and cycling promotion strategies should also consider safety aspects to ensure that the positive health effects are not diminished by accidents in these more vulnerable road users [41]. The dose-response results suggest the greatest impact is in the first exposure category of walking and cycling per week. This is the first 120 minutes per week for walking and the first 100 minutes per week for cycling. Public health population approaches to PA promotion should focus on promotion and maintenance in this range for the greatest impact on ACM.

It may be considered unexpected that cycling, traditionally a more intense activity than walking had a very similar risk reduction (1 percentage point difference). We note that this is for the same energy expenditure, but as cycling is more intense, would be achieved in a shorter time (100 minutes per week for cycling; 170 minutes per week for walking).

Our methods have demonstrated a robust way to combine multiple studies with heterogeneous exposures in meta-analysis and estimate (1) the effect at PA recommendations and (2) the dose-response across the exposure range. Our paper also shows the need for improvement in exposure assessment. We have provided outputs that will be suitable for simple and complex modelling such as the Integrated Transport and Health Impact Modelling Tool (ITHIM) [42], the Health Economic Assessment Tool (HEAT) for walking and cycling [39],[43] or integrated models for preventable risk [44]. We have confirmed the beneficial effects and the greater importance of increasing the behaviours at low levels.

Future research

Systematic reviews generally report the need to update their results once new studies are published. We reserve judgement on this; the advent of cohort studies using objective exposure assessments such as accelerometer, pedometer or GPS (global positioning system) devices may indicate that subsequent reviews start afresh if exposures are non-comparable.

As more evidence does become available we recommend investigation of benefits and dose-response patterns by age, gender and geographic region. It is also important that studies of walking and cycling are conducted in low- and middle-income settings and included in any future iterations of this review to make the global utility of the findings more robust.