Information sources and search strategy

This meta-analysis was conducted following the checklist of the Meta-Analyses of Observational Studies in Epidemiology[19]. We used a three-part search strategy to identify potential studies for this meta-analysis: 1) we searched the reference lists of the first systematic reviews that covered the published literature on a range of sedentary behaviors in adults (e.g., sitting, TV-viewing) and multiple health outcomes (e.g., overweight or obesity, cardiovascular disease, diabetes, cancer, mortality) covering a period from 1989 to January 2011 [8,9]; 2) we searched Medline, Pubmed, Embase, and Web of Science for studies published between January 1, 2011 and January 31, 2013 with the following search terms: (("sitting time" OR "sedentary behavior" OR "sedentary behaviour") AND (mortality OR mortalities OR death OR fatal)) AND (risk OR Cox OR hazard OR survival analysis OR odds); and 3) we further searched the reference lists of included papers and later review articles, as well as all authors’ personal literature databases for relevant studies, including publications in press.

Eligibility criteria

Inclusion criteria were: prospective design; published in English; studies of healthy adult populations at baseline; measured total sitting time as an exposure variable; had all-cause mortality as an outcome variable (i.e., mortality was ascertained without consideration of specific cause of death); provided estimates of relative risk (RR) or odds ratios (OR) or hazard ratios (HR) with 95% confidence intervals (CIs) or reported data for their calculation[13]. Conference abstracts were not included.

Data extraction

The following data were extracted from retrieved articles: author(s), study name, year of publication, total sitting time measure, sample size, age at baseline, follow-up time, person years, confounding variables that were adjusted for in the analysis, and the HR, RR or OR estimates with corresponding 95% CIs for models with and without adjustment for physical activity. Two authors independently extracted the data from each study and compared the data for consistency. Any discrepancies between the two reviewers were settled through discussion and a third reviewer’s help was sought for resolving disagreements. We contacted corresponding authors to confirm or request missing data and incorporated responses into the analyses.

Study appraisal

Two authors independently appraised the methodological quality of all included studies using a quality rating list based on previous checklists[20-23]. This rating list consisted of 15 criteria and each criterion was assessed as ‘yes’ (=1) or ‘no’ (=0), with each assigned equal weighting, so that a quality score ranging from 0 to 15 could be calculated for each study. Any disagreements in quality ratings between the two reviewers were resolved in a consensus meeting between them, with a third reviewer consulted in the event of disagreement.

Analysis

“Dose” was assigned using the midpoint between the lower and upper boundary of each sitting category for which a HR, RR or OR was provided. For open ended categories we assumed the same magnitude of dose as for the neighboring category. For example, the categories <4, 4-<8, 8-<11, ≥11 h/day were assigned doses of 2, 5, 9 and 12 h/day, respectively.

First, for data with and without adjustment for physical activity, we used spaghetti plots to graphically illustrate the shape and direction of the dose-response association between sitting hours and risk of death across all studies[24]. Using the generalized least squares for trend (GLST) procedure we then estimated the trend for each study sample and pooled the estimates to produce a forest plot and random effects trend[25]. A funnel plot was made for visual inspection of publication bias.

We then used the method of pool-first to estimate the dose-response trend for all study samples combined[25]. Dose-response linear and multiple non-linear models were fitted to determine the model of best fit for the pooled data. Non-linear responses were tested with piece-wise spline regressions using inflection point(s), or “knots”, identified from the model of best fit[26]. Model adequacy was determined by the model chi-square and goodness-of-fit statistics produced by the GLST procedure.

Statistical heterogeneity was tested by calculating the I² statistic and its interpretation was based on the Cochrane Collaboration interpretation whereby 50% or greater represents substantial heterogeneity[27].

We calculated the population attributable fraction (PAF) for each included study following methods outlined by Katzmarzyk and Lee[15]. An overall weighted PAF estimate was computed based on the weights from the meta-analysis. All analyses were carried out with STATA version 11 (STATA Corporation, College Station, TX, USA).

For the purposes of this paper, we refer to daily sedentary time and sitting time as ‘daily total sitting time’, and moderate-to-vigorous physical activity as ‘physical activity’, unless explicitly stated otherwise.

Study selection

The literature search yielded a total of 173 abstracts (Figure 1). One hundred and sixty three did not meet the core inclusion criteria and were excluded initially, and ten full-text articles were further considered. Four studies were subsequently excluded because they did not measure daily total sitting time[16,17,28,29]. Only one study, by Koster et al, objectively assessed total sedentary time using accelerometers[30]. As accelerometers have been shown to measure sitting time in a different manner to questionnaires[31], we ran models with this study included and excluded, to see if the results were changed with inclusion of this paper. Two papers were included from authors’ personal literature databases[32,33]. Therefore, a total of six studies was included in the meta-analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Selection of studies for inclusion in the meta-analysis.

https://doi.org/10.1371/journal.pone.0080000.g001

Study characteristics

The characteristics and main outcomes of the prospective cohort studies included in this meta-analysis are shown in Table 1.

Data from 595,086 people were included in the meta-analysis. There were 29,162 deaths during 3,565,569 person-years of follow-up. Study sample sizes varied, ranging from 1,906 [30] to 240,819 [11] with follow-up periods ranging from a mean of 2.8 years [10,30] to over 8 years[11,34]. Two studies involved participants aged at least 45 years old[10,34], two involved adults aged at least 50 years[11,30], one study involved adults aged 18 years or more[33], and one study had participants aged 76-81 years old[32]. One study involved women only[32], and over 50% of participants in four of the five other studies were female[10,30,33,34]. One study reported results for men and women separately[34]. All studies ascertained participants’ mortality status through linkage with a regional or national death registry.

Five studies assessed daily total sitting time by self-report[10,11,32-34]. One study used objective measurement of daily sedentary time[30]. All studies assessed daily total sitting time in hours per day grouped into categories. However, the cut-points for the categories were not consistent across the studies (see Table 1). For example, Koster et al [30] divided daily total sedentary time into quartiles, whereas two studies categorized total sitting time as <4, 4-<8, 8-<11, ≥11 h/day[10,32].

All studies adjusted for multiple potential confounding factors including moderate-to-vigorous physical activity. Two studies operationalized physical activity as walking and other moderate-to-vigorous physical activity[10,32], one measured multiple indicators including heavy physical work, walking, exercise and sports[34], one study only assessed moderate-to-vigorous physical activity in leisure time[11], and one study categorized participants as meeting or not meeting physical activity guidelines[33]. Five of the six studies also presented data for models without adjustment for physical activity[10,11,30,32,33], which allowed us to examine associations between daily total sitting time and mortality risk, with and without adjustment for physical activity.

Quality assessment

The study appraisal criteria and number of studies scoring a point for each item are presented in Table 2. Study quality scores ranged from 10/15 to 12/15 with a mean percentage agreement of 85.6% on quality ratings between the reviewers. All studies provided information about their objectives, study design, participant sampling and recruitment, measures of total sitting and mortality, data sources and statistical methods.

Associations of daily total sitting time with all-cause mortality risk

The analyses involving data with and without adjusting for physical activity involved seven samples from six studies (one study reported results for men and women separately)[34] and five samples from five studies, respectively. The multivariable-adjusted hazard ratios, with adjustment for physical activity, are shown in Figure 2. Figure 2A presents the spaghetti plot of the raw HRs for all-cause mortality from each sample, by dose of daily total sitting time, with the pooled estimate from the fitted spline model, with 95% confidence limits. We observed similar effect sizes for all study samples across dose of sitting, except in the accelerometer study[30], which showed considerably larger hazards of dying from all causes among respondents whose total daily sitting time exceeded the first quartile of the sample (7.6 h/day for men, 7.2 h/day for women). The forest plot of the hazard per-hour increase in sitting, with multivariable adjustment including physical activity, is shown in Figure 2B. For each study the HRs, with 95% CIs, are shown, along with the percent weighting. We found a pooled hazard ratio of 1.02 (95%CI: 1.01-1.03). Heterogeneity was high (I²=82.7%) and statistically significant (p<0.001). A sensitivity analysis excluding the results of the Koster et al [30] study did not change the pooled estimate (HR=1.02; 95% CI: 1.01-1.03), nor improve heterogeneity (I²=81.0%, p <0.001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Associations of daily total sitting time with all-cause mortality risk with adjustment for physical activity.

A: Spaghetti plot of the raw Hazard Ratios (HR) for all-cause mortality from each study sample, by dose of daily total sitting time, with multivariable adjustment including for physical activity. The bold solid line denotes the pooled HR estimate from the fitted spline model and the dotted lines the 95% confidence limits. B: Forest plot of the HR per-hour increase in sitting with multivariable adjustment including for physical activity. (n=7 samples from 6 studies).

https://doi.org/10.1371/journal.pone.0080000.g002

Figure 3 presents the results for the analyses involving data with no correction for physical activity. We observed similar patterns to those seen for data with adjustment for physical activity (Figure 2). The pooled HR of 1.04 (95% CI: 1.02-1.05) was higher than that for the physical activity adjusted analyses, and heterogeneity was also high (I²=85.1%, p<0.001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Associations of daily total sitting time with all-cause mortality risk with no adjustment for physical activity.

A: Spaghetti plot of the raw Hazard Ratios (HR) for all-cause mortality from each study sample, by dose of daily total sitting time, with multivariable adjustment but not for physical activity. The bold solid line denotes the pooled HR estimate from the fitted spline model and the dotted lines the 95% confidence limits. B: Forest plot of the HR per-hour increase in sitting with multivariable adjustment but not for physical activity. (n=5 samples from 5 studies).

https://doi.org/10.1371/journal.pone.0080000.g003

Evidence of dose-response

Initial linear random effects analyses suggested that the association between daily total sitting time and all-cause mortality was not linear in models with adjustment for physical activity (HR=1.03; 95% CI: 0.98-1.08) and without physical activity adjustment (HR=1.05; 95% CI:0.98-1.13). The goodness-of-fit test suggested that the linear dose-response models did not fit the data well with adjustment for physical activity (χ²₍₁₉₎ = 71.02, p<0.001) and without physical activity adjustment (χ²₍₁₅₎ = 91.37, p<0.001).

Multiple non-linear spline models were subsequently fitted and are shown graphically by the bold solid lines in Figures 2A and 3A. Estimates at specific doses take into account the trends at preceding “knots”, and it is best to read estimates from the figures provided. For the physical activity adjusted data, the best fitted spline model was with knots at >3 and >7 h/day (χ²₍₁₇₎ =22.67, p=0.160) (Figure 2A). For every 1-hour increment of sitting in intervals 0-3, >3-7 and >7 h/day total sitting time, the HRs were 1.00 (95% CI: 0.98-1.03), 1.02 (95% CI: 0.99-1.05) and 1.05 (95% CI: 1.02-1.08) respectively. For data without adjustment for physical activity level, the trend estimates were similar but with slightly steeper slopes. The best fitted spline model was with knots at >4 and >8 h/day (χ²₍₁₃₎=17.85, p=0.163) (Figure 3A). For every 1-hour increment of sitting in intervals 0-4, >4-8 and >8h/day total sitting time, the HRs were 1.02 (95% CI: 0.99-1.04), 1.02 (95% CI: 1.00-1.04), and 1.08 (95% CI: 1.05-1.11), respectively. Spline modeling with knots at >3 and >7 h/day did not fit the data without adjustment for physical activity well (χ²₍₁₃₎ =24.27, p=0.029). For example, based on the dose-response spline models (Figures 2A and 3A), an adult sitting for 10 h/day would have 34% (HR=1.34, 95% CI: 1.28-1.40) and 52% (HR=1.52, 95% CI: 1.46-1.58) increased all-cause mortality risk with and without adjusting for physical activity, respectively.

The overall weighted PAF for all-cause mortality attributable to total daily sitting time, after taking physical activity into account, was estimated to be 5.9%. Individual study PAFs ranged from 1.9% for the largest study with self-report data [11] to 58% for the smaller objective measure study by Koster and colleagues[30]. When the Koster et al study was excluded, the overall weighted PAF for all-cause mortality attributable to total daily sitting time was 4.4% (range: 1.9% - 15.6%), after accounting for physical activity.

Evidence of publication bias

The small number of studies and presence of significant heterogeneity precluded formal testing for publication bias[35]. However, visual inspection of the forest and funnel plots (Figure S1) suggest that there may be some publication bias present, with larger cohort studies reporting the smaller effects[36].