Sedentary behaviour and risk of all-cause, cardiovascular and cancer mortality, and incident type 2 diabetes: a systematic review and dose response meta-analysis

Studies had to have assessed the association between total daily sitting/sedentary, TV viewing or leisure sitting time, and at least one of the outcomes of interest: all-cause, CVD or cancer mortality, incident (fatal and non-fatal) CVD and incident T2D. Time spent sitting/sedentary could be self-reported or objectively measured. Only primary research studies with a prospective design, with at least an abstract in English and investigating non-diseased adults (≥ 18 years) in the general population were included.

Sources included:

Titles and abstracts were screened by one author (RP) using the inclusion criteria, full reports were assessed where these were met or where there was uncertainty, allowing a final decision on eligibility. Where different articles formed part of the same cohort study, only data from the most recent publication for any given exposure-outcome combination was used. A minimum of 4 different eligible cohorts were required in order to carry out an analysis. Authors were contacted for additional information where needed.

Using a pre-designed data extraction spreadsheet, two authors carried out independent extractions and disagreement was resolved through discussion (RP and EM).

The quality of each study was assessed using these criteria: size of cohort, length of follow-up, description of inclusion criteria and sampling strategy and sample representativeness, based on the Newcastle–Ottawa scale [22]. No overall quality score was assigned for use in analyses, to prevent the scale itself becoming a source of bias [23].

Extracted data were harmonized, converting each measure into one of: total sedentary, TV viewing or leisure sedentary time, quantified in h/day. Categories of sedentary time were assigned a dose, either the mid-point, or, in case of open-ended categories, half the width of the adjacent interval from the boundary (Online Appendix Tables 2 and 3). Where the lowest exposure was not the referent category, hazards were recalculated [24, 25].

Estimates of the linear association for each contributing study were calculated using Generalized Least-Squares regression [26, 27]. These were used to perform a random effects meta-analysis within each exposure-outcome combination, for both the most adjusted model without adjustment for PA and the least adjusted model with adjustment for PA [28]. These provided the summary RR per additional hour/day of exposure. Statistical heterogeneity of contributing studies was assessed with the I² statistic, which was considered low if < 25%, and high if > 75% [29]. To examine publication bias and small study effect, Funnel plots were used, Egger’s tests were derived for each exposure-outcome combination with ≥ 5 contributing studies [30].

Following the estimation of a linear association, a restricted cubic spline transformation was carried out to investigate the shape of the dose–response relationship. Knots were placed at the 10th, 50th and 90th percentiles [31]. A random effects meta-analysis was then carried out to estimate the non-linear relationship between sedentary time and the respective health outcome. Where a change in strength of association was seen at a certain level on the exposure continuum, the RR (per h/day increment in exposure) on either side of this exposure level was calculated as the difference in log(RR) divided by difference in hours of exposure. Results were presented back on the original scale and were based on PA-adjusted analyses.

Sensitivity analyses were carried out to investigate the influence of study characteristics which might lead to risk of bias. Where sufficient data were available these were carried out on the linear PA adjusted associations. Factors investigated include: adiposity adjustment, sex, length of follow-up, age of cohort at baseline and representativeness of cohort.

PAF estimates were calculated for TV viewing, indicating the proportional reduction in incidence of the respective outcome if prevalent TV viewing levels were reduced to zero, assuming causality. As we did not have access to TV viewing data worldwide, TV viewing data from the 2012 Health Survey for England (HSE), a nationally representative sample of the English population, were used to carry out a Monte-Carlo micro-simulation as an illustration of the potential magnitude of the impact. Each 17+ year old participant in HSE was probabilistically assigned an RR based on the RR and associated uncertainty from the PA-adjusted dose–response analysis which corresponded to their TV viewing category (0, 0 to < 2, 2 to < 4, 4 to < 6 and 6+ h/day of TV viewing). These assigned RRs, along with the proportional contribution of each individual, according to HSE survey weights, were used to calculate an attributable fraction for each participant, according to the formula below [32].

Equation 1—Population attributable fractionP_i = proportion of population at exposure level i, P_i′ = proportion of population at counterfactual exposure level, i.e. zero exposure, RR = the relative risk at exposure level i. n = the number of exposure levels.

This procedure was then repeated 5000 times and the final PAF estimate was the median value across the 5000 simulations. A 95% CI was calculated using the 2.5th and 97.5th percentiles of the simulated distribution.

Stata version 14.2, StataCorp, USA was used for the meta-analysis. The PAF calculations were carried out using Analytica Free 101 edition, Lumina Decision Systems Inc., USA.

This work was supported by the British Heart Foundation, the Medical Research Council, Cancer Research UK, Economic and Social Research Council, National Institute for Health Research, and the Wellcome Trust. The funders had no role in study design, conduct, or reporting of the results.

The literature search provided 2201 potential articles. Following screening of titles and abstracts, full text was retrieved for 124 publications, rendering 39 studies for which inclusion criteria were met (Fig. 1). For 5 studies there was an insufficient number of comparator cohorts within the same exposure-outcome combination [33‐37], leaving 34 studies, across 8 exposure-outcome combinations, to be included in the analysis [38‐71]. An insufficient number of cohorts was identified to allow investigation of associations between leisure sedentary time with any outcome. There were also insufficient studies investigating incident CVD with any exposure. We were therefore unable to carry out the planned analysis on these exposures and outcomes. Additional data were successfully obtained from 11 studies [38‐48].

Data from a total of 1,331,468 unique participants was included. Table 1 and Online Appendix Table 4 summarize the characteristics of the 34 included studies (with additional data from publications describing cohort characteristics [72‐85]). The size of included studies ranged from 208 [51] to 240,819 [61] participants with a mean of 39,161. Follow-up was on average 8.9 years and varied from 2 [69] to 31 [46] years. The numbers of cases and participants by outcome are shown in Table 2. Of the 34 studies, 17 were from North America, 9 from Europe, 4 from Australia and 4 from Asia. The dates of publication spanned 2001–2016, although 65% were published in 2013 or later.

The quality of the included studies was generally good (see Table 1 and Online Appendix Table 4). Three of the five smallest cohorts (presenting 32 cases/208 participants [51], 112/1677 [67], and 409/2918 [42]) measured sedentary time objectively. Of 34 studies, 22 had ≥ 10,000 participants [39‐41, 45, 48, 50, 52‐62, 64, 65, 68, 70, 71]. There were 3 all-male studies [42, 47, 52] and 5 all-female studies [53, 58, 59, 63, 68] (providing data from 4 cohorts). Some articles presented results for multiple exposures and/or outcomes, with 34 publications presenting 57 analyses; the numbers of contributing articles for each exposure/outcome combination are presented in Table 2.

Most studies assessed sedentary behaviour by questionnaire (31 out of 34) with three studies measuring sedentary time objectively using accelerometers worn on the participant’s hip, waist or lower back for up to 7 days (Online Appendix Table 4) [42, 51, 67].

Outcomes were assessed objectively in 27 of the 34 studies (Online Appendix Table 4). This represents those studies with a mortality outcome assessed using death registries, in addition to four studies which used an objective measure to define T2D status [38, 44, 50, 65], the remainder being self-reported T2D.

All included studies reported adjusted effect estimates, with adjustment for PA present in all but four studies [44, 46, 51, 70]. One study presented results for some outcomes with PA adjustment and some without [42]. Adjustment for PA varied in detail; from simply meeting the PA guideline or not, to calculating weight-adjusted energy expenditure across multiple domains of PA (Online Appendix Table 4).

The association between total sedentary behaviour and all-cause mortality appeared to be non-linear, both with and without adjustment for PA (Fig. 2). Testing for non-linearity supported this finding. At lower levels of exposure, there were small increases in risk associated with increasing sedentary behaviour; above approximately 8 h/day of sedentary behaviour, the risk increased more rapidly. In PA adjusted analyses this resulted in an estimated RR of 1.01 (1.00–1.01) for each additional hour of exposure below 8 h/day and 1.04 (1.03–1.05) for each hour above 8 h/day (Online Appendix Table 5).

For TV viewing, the association also appeared to be non-linear (Fig. 2 and Table 2) with a change in gradient at approximately 3.5 h/day in the PA adjusted analyses (Fig. 2). Below this level the RR was 1.03 (1.01–1.04) per hour/day, and greater increases in risk were seen above this level (RR = 1.06 (1.05–1.08) per h/day; Online Appendix Table 5). Due to insufficient availability of data, investigation of the associations between leisure-time SB and mortality could not be undertaken.

For total sedentary behaviour, non-linearity was seen again for both the non-PA adjusted and PA adjusted models (Fig. 2 and Table 2). In the PA adjusted analysis, the threshold was in the region of 6 h/day of exposure, below which each additional hour was associated with an estimated RR of 1.01 (0.99–1.02) and above which each additional hour was associated with an RR of 1.04 (1.03–1.04) (Online Appendix Table 5).

The PA adjusted non-linear association between CVD mortality and TV viewing showed greater risk increases with every hour above a threshold of approximately 4 h/day (Fig. 2). The estimated RR for each additional hour of TV viewing was 1.02 (0.99–1.04) below 4 h and 1.08 (1.05–1.12) above.

The linear association between total sedentary behaviour and cancer mortality was marginally non-significant and unaffected by PA adjustment; the PA adjusted estimate was 1.01 (1.00–1.02). There was no evidence for non-linearity (Table 2).

The linear association between TV viewing and cancer mortality was estimated to be 1.03 (1.02–1.04) in non-PA adjusted models and 1.02 (1.01–1.03) when adjusted for PA. There was no evidence for non-linearity for either PA adjusted or unadjusted associations (Fig. 2 and Table 2).

The linear association between total sedentary behaviour and T2D was estimated to be 1.01 (1.00, 1.01), which could only be estimated with PA adjustment as there were insufficient studies without PA adjustment (Table 2 and Fig. 2).

PA adjusted analysis of the association between TV viewing and T2D shows some deviation from linearity (Fig. 2) although statistical evidence was equivocal (p = 0.066; Table 2). The PA adjusted linear association was estimated to be 1.09 (1.07, 1.12) However, larger increases in risk were seen with increasing TV viewing below approximately 4 h (1.12 (1.08–1.15) in PA adjusted analysis), above this level increasing TV viewing was associated with lower increases in risk (1.05 (1.03–1.07)). In non PA-adjusted analyses the association appeared linear with an estimated RR of 1.12 (1.08, 1.16) associated with each additional hour of TV viewing.

Across all combinations, PA adjustment appeared to attenuate the effect estimates. The difference between the effect sizes with and without PA adjustment was relatively small when total sedentary time was the exposure, but somewhat greater when TV viewing was examined. Substantial heterogeneity was observed for the pooled effect estimates, ranging from I² values of 0% for both total sedentary behaviour and CVD, and total sedentary behaviour and cancer mortality, to I² of 90.1% for TV viewing and CVD mortality. Funnel plots and Egger’s tests, showed no definitive evidence for publication bias (Online Appendix Fig. 5). However, the low numbers of contributing studies for some associations made it difficult to rule out these biases.

Sensitivity analyses are presented in Online Appendix Table 6. The RR of CVD mortality associated with each additional hour of TV viewing was greater when studies of younger participants were excluded, 1.08 (1.06–1.10) compared with 1.04 (1.01–1.08) in the main analysis. All other sensitivity analyses showed no substantive change from the main findings.

For all-cause mortality, 8% (6–10%) was associated with TV-viewing in the English population, when using the PAF method. This estimate was 5% (1–8%) for CVD and 5% (2–7%) for cancer mortality. For T2D 29% (26–32%) of incidence was estimated to be related to TV-viewing.

Biological mechanisms have been suggested to explain the independent associations of sedentary behaviour, in particular for cardio-metabolic diseases, through independent effects of prolonged sitting on lipid and glucose metabolism in the large skeletal muscles involved in posture (legs and core) and on hemodynamic vascular signalling potentially causing atherogenesis [86‐89]. Associations for TV viewing were generally stronger than those for total sitting time with the same outcome which could be explained by several factors. Firstly, TV viewing has been linked to higher intakes of energy and macronutrients along with greater energy from snacks [90]. Poor diet quality and increased total calorie intake have been associated with increased risk of mortality and is a strong determinant of T2D, suggesting an important mediating role for dietary intake which is likely less relevant for total sitting time [91, 92]. Second, a potentially different confounding structure for TV viewing may be more difficult to fully account for. Third, criterion validity of self-reported TV viewing estimates tends to be stronger than those for total sitting time estimates [93, 94]. Lastly, the typical timing of TV viewing, i.e. in the evening following the main meal of the day [95], may exacerbate the repetitive cardiovascular effects of postprandial glucose and lipid excursions following this meal, especially if TV viewing is predominantly accumulated in prolonged bouts of sitting [96, 97].

The methods used to measure exposure varied; measurement of sedentary behaviour is still primarily reliant on self-report questionnaires. Heterogeneity in question phrasing, the time period considered and whether a question is single or multipart can all influence validity [93]. Misclassification of sedentary exposure would potentially dilute the association in our analysis, resulting in possible underestimation of effect size. The use of accelerometer measured sedentary time addresses some of the limitations of questionnaires, however this data has its own limitations. For example, some accelerometer methods cannot detect cycling or swimming, or fail to distinguish between sitting/lying and standing still [87]. This substantial heterogeneity in exposure measurement contributed to the high heterogeneity indices (I²) for the pooled estimates which may have influenced our overall findings. It is possible that only some of the constituents of total sitting are detrimental to health, for example sitting while reading is potentially advantageous [98]. That we were unable to investigate leisure sitting time in this meta-analysis due to insufficient studies would indicate that more research is required on the effect of different domains of sedentary behaviour. In addition to the exposure measure, the quality of the measurement of important covariates, such as PA, diet and socio-economic position, varied greatly between studies, if included at all, potentially leading to residual confounding. The low number of studies for some combinations meant that investigation of leisure time sitting could not be carried out. It also led to a lack of statistical power for subgroup or sensitivity analysis and bias assessment. This also meant that meta-regression techniques to investigate the impact of the potential sources of heterogeneity were precluded.

This meta-analysis considered total sedentary behaviour and TV viewing time as separate exposures. This is important as they have different associated socio-demographic and/or behavioural patterns (e.g. dietary intake) and therefore different confounding/mediating patterns [6, 90]. Inclusion of emerging research using objectively measured sedentary time is another strength. In addition, we investigated the shape of the dose–response curves, to identify where the greatest risk/benefits lie along the spectrum of exposure for all exposure—outcome combinations. Moreover, to our knowledge, this is the first study to calculate PAF estimates for TV viewing time and all considered health outcomes based on meta-analytical risk estimates when potential non-linearity of associations were taken into account.

However, our meta-analysis also has certain limitations. The use of summary data means heterogeneity of used statistical methods may influence comparability of included studies [29]. In order to investigate the effect of PA adjustment we had to select models which were as similar as possible except for adjustment for PA. However, in some studies additional differences in covariates were seen between these models and this may have resulted in residual confounding of the considered study-specific risk estimates. It was also necessary to make several assumptions during the dose-assignment procedure. Whilst these assumptions may have been crude in studies reporting little detail on the exposure, this approach allowed us to consider the totality of the currently published evidence. Treating the many and heterogeneous conditions that make up cancer as one outcome may have contributed to our mixed findings with these analyses. Investigating separate cancer types may be more informative, where there is enough data [99]. Attempting to reduce reverse causality by only including prospective studies may not have been entirely effective, especially in the case of T2D. An estimated 27% of those with the condition have no formal diagnosis, therefore having the condition may have preceded ascertainment of exposure data [100]. Finally, the calculation of PAFs rests on the assumption of causality, and the use of unbiased estimates with no measurement error.

To calculate the PAF estimates we have used the exposure profile representative of the population of England. These might not be representative for other countries. However, average US TV viewing levels, for example, are similar, 2.6 h on a weekday and 3.3 h on a weekend day compared with 2.7 h and 3.1 h/day respectively in England [101]. The estimated 29% of T2D in England in 2012 that could be prevented or postponed by eliminating TV viewing, assuming causality, reflects the high RRs seen for this association. The linear association and relatively high RR even at lower exposure levels are important contributors, as a large proportion of HSE participants report lower TV viewing levels (75% of participants report < 4 h/day), but only a small proportion reports no TV viewing (3%). The PAF estimates for all-cause mortality (8%) and CVD and cancer mortality (both 5%) also suggest a potentially important burden of preventable disease from current population levels of TV viewing.

The differing nature of the relationships between TV viewing and different outcomes may complicate any prevention strategy. The prevention of T2D would perhaps be best served by reducing TV viewing among the whole population, however, to prevent other outcomes targeting those with highest exposure levels may be more appropriate as these are the individuals for whom any reduction would confer the greatest benefit. Furthermore, the effect of any behaviour change will also be influenced by the nature of the replacement activity [13]. For example, greater reductions in risk may occur when replacing sedentary time with strenuous exercise compared with walking for pleasure. Replacing some sedentary behaviours may confer greater benefits than others, e.g. replacing TV viewing may be more beneficial than replacing general screen time [13].

Another potentially important determinant of the health effects of sedentary behaviour is the extent to which breaks are taken in extended periods of sitting. None of the studies included in this meta-analysis took into account accumulation pattern of sitting and therefore this falls outwith the scope of this study.