Introduction

Coffee is a complex mixture of over a thousand bioactive compounds including caffeine, chlorogenic acids, and diterpenes [1]. As coffee is one of the most commonly consumed beverages around the world, its potential effects on human health could be large on a population scale. Coffee was considered as potentially harmful to human health because of caffeine which may raise blood pressure [2], and the possible carcinogenicity of coffee had been suggested regarding certain cancers such as urinary bladder cancer [3, 4]. However, recent summary results from cumulative evidence show that moderate coffee consumption is associated with decreased risk of type 2 diabetes, cardiovascular disease (CVD), mortality, and several types of cancers, including liver and endometrial cancers, and possibly colorectal, breast, and prostate cancers [5, 6]. These findings suggest that coffee may be included as part of healthy diet.

Many observational studies have investigated the association between coffee consumption and mortality from all-causes, CVD, and cancers in the general population [7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45]. Some of the studies have also reported the estimates stratified by various factors including age, BMI, alcohol drinking or smoking status [8, 9, 12, 13, 22, 25, 27, 31, 33,34,35, 37, 43,44,45]. Several clinical studies suggested that the effect of coffee consumption could be different by age [46, 47] or obesity status [48, 49]. Elderly were more sensitive to the pressor effects of caffeine [47] and obese people had smaller thermogenesis induced by coffee than lean people [48]. Also coffee consumption has been found positively associated with lifestyle factors such as smoking and alcohol drinking [10, 43]. Considering that aging, obesity, alcohol drinking and smoking are closely linked to the incidence of chronic disease, it is of interest to examine possible variations in the association between coffee intake and mortality by these factors. Several meta-analyses to determine the relation of coffee intake and risk of mortality [50, 51] indicated that high coffee consumption was associated with lower risk of premature mortality, but to our knowledge, no meta-analysis has stratified by age, BMI, or alcohol drinking. There was only one meta-analysis that provided the results by smoking status and found an inverse association between coffee consumption and mortality in non-smokers [50]. Furthermore, no meta-analysis has reported pooled RR of mortality from respiratory disease, diabetes, and non-CVD, non-cancer causes.

Therefore, this meta-analysis examined the association between coffee consumption and all-cause mortality, including results stratified by age, BMI, alcohol consumption, smoking status, sex, and geographical region, and separately by caffeine content of the coffee. In addition, we conducted a meta-analysis of coffee consumption in relation to mortality from CVD, cancer, respiratory disease, diabetes, and all non-CVD, non-cancer causes.

Methods

Literature search and study selection

A systematic literature search on PubMed and ISI Web of Science databases for studies of coffee and mortality published through March 8, 2019 was conducted. The search terms were as follows: “(coffee OR caffeine OR hot beverages)” combined with “(mortality OR death OR survival OR fatal)”. Manual search was also performed to identify additional relevant studies by reviewing the reference list of review and retrieved articles. The searches were limited to articles published as full-length and in English. Inclusion criteria were as follows: (1) had a prospective design; (2) the exposure of interest was coffee consumption; (3) the outcome of interest was mortality from all-causes, CVD, or cancer; (4) reported relative risks (RR) and confidence intervals (CI) or sufficient data to calculate them. Studies that evaluated risk of mortality in people with disease were excluded. When more than one study reported the results from the same cohort, we selected the study which included longer follow-up times and larger sample sizes.

Data extraction

Two independent authors (Y. K. and Y. J.) extracted data according to the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines [52]. The following information was collected from each study: year of publication; first author’s surname; country and study cohort name; baseline age and sex of participants; number of deaths; number of participants or person-years; follow-up period; each category of coffee consumption; RR and 95% CI for all categories of coffee consumption; adjustment factors. Any discrepancies between the authors in this process were addressed by discussion.

Quality assessment

Quality of the original studies which were included in meta-analysis was evaluated using the Newcastle–Ottawa Scale [53]. The quality assessment scale awards 0–13 points based on three perspectives, as follows: selection of study population; comparability; outcome assessment. We considered studies with a total score of ≥ 9 points to represent high quality.

Statistical analysis

The random-effects models by Dersimonian and Laird [54], which incorporated variations both within and between studies, were used to calculate pooled RR of all-causes, CVD, and cancer mortality for the highest versus lowest coffee consumption and for 1 cup a day increment. When studies had not used the lowest category as a reference, we recalculated the RRs and their 95% CI relative to the lowest category. For any study which provided results separately by sex, we combined the RRs and then included the pooled RR in the meta-analysis.

Linear dose–response relationships were evaluated using the method developed by Greenland and Longnecker [55,56,57] to estimate the study specific slope lines. Studies that provided RRs for only 2 exposure categories [17, 34, 39] or did not report the number of deaths and subjects for each coffee consumption category [18, 24, 25, 28] could not be included in dose–response analysis. We used the median value of coffee consumption for each exposure category. If the upper category was open-ended, we assumed the same interval as the adjacent category. A potential non-linear dose–response relationship was also assessed between coffee consumption and mortality using restricted cubic splines with 4 knots at fixed percentiles (5, 35, 65, and 95%) of the aggregated exposure. We computed the P value for non-linearity by testing the null hypothesis in which the coefficient of the second spline is equal to zero [58].

To investigate whether the association between coffee consumption and all-cause mortality differed by age (< 60/≥ 60 years), sex, BMI (< 25/≥ 25 kg/m2), alcohol consumption (low/high), smoking status (non-smoker/smoker), geographical region (US/Europe/Asia) or caffeine content of coffee (decaffeinated/caffeinated), we conducted stratified analysis when separated data were available.

Heterogeneity among the studies was evaluated using the Q statistic, and inconsistency was quantified through the I2 statistic. To investigate the robustness of the pooled RR, sensitivity analysis excluding one study at a time was conducted. Publication bias was assessed with Begg’s [59] and Egger’s tests [60]. A two-tailed P < 0.05 was assumed to be statistically significant. The Stata/SE software (version 14.2; Stata Corp LP, College Station, Texas) was used for all the statistical analyses.

Results

Study characteristics

A total of forty prospective cohort studies from 39 articles (1 article [8] provided results from 3 cohorts and 2 articles [41, 44] reported results for mortality from different causes from the same cohort) involving 3,852,651 subjects and 323,120 deaths from all-causes, 229,884 deaths from cancer, and 81,188 deaths from CVD (2 studies [24, 28] did not report the number of deaths) were included in this meta-analysis [7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45]. Details of the study selection are shown in Fig. 1. Table 1 shows the main characteristics of studies included in the meta-analysis. The follow-up periods ranged from 6 to 28 years. Studies were conducted in Europe (n = 17) [10, 12,13,14,15, 17, 18, 21, 26, 30, 34, 38,39,40, 42, 43, 45], US (n = 15) [7,8,9, 11, 19, 20, 22, 24, 25, 27, 28, 32, 33]. Japan (n = 6) [16, 29, 31, 35,36,37, 41, 44], Singapore (n = 1) [31], and Australia (n = 1) [23]. Several studies provide separated results by age (n = 11) [8, 9, 25, 27, 33, 35, 37, 44, 45], BMI (n = 7) [8, 9, 13, 27, 43], alcohol consumption (n = 4) [9, 12, 13, 27], smoking status (n = 15) [8, 9, 12, 13, 22, 27, 31, 33,34,35, 43,44,45], and type of coffee (n = 12) [7,8,9, 11, 13, 27, 32, 33, 42, 43]. The results of quality assessment showed that all studies had high qualities indicating nine or higher scores.

Fig. 1
figure 1

Flow chart of study selection

Full size image
Table 1 Characteristics of the prospective cohort studies included in the meta-analysis
Full size table

All-cause and cause-specific mortality

A total of thirty six prospective cohort studies from 34 articles including 323,120 deaths and 2,837,526 subjects examined the association between coffee consumption and all-cause mortality [7,8,9, 11,12,13,14, 16,17,18,19,20,21,22,23, 25,26,27,28, 30,31,32,33,34,35,36,37,38,39,40, 42,43,44,45]. The pooled RR for highest versus lowest consumption was 0.88 (95% CI 0.84–0.92) with significant heterogeneity (P < 0.001, I2 = 76.6%) (Table 2, Supplementary Figure 1). The heterogeneity was slightly reduced when two outlying studies [19, 25] were excluded (P < 0.001, I2 = 67.0%). A significant non-linear association between coffee intake and all-cause mortality was found (P for non-linearity < 0.0001) (Fig. 2). The largest reduction in RR was observed with the consumption of 3.5 cups/day (RR = 0.85, 95% CI 0.82–0.89), compared with no coffee consumption.

Table 2 Summary of pooled relative risks (RR) of all cause-and cause-specific mortality for coffee consumption
Full size table
Fig. 2
figure 2

Pooled dose-response association between coffee consumption and all-cause mortality. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image

A total of thirty one prospective cohort studies from 29 articles with 81,188 deaths and 2,631,398 subjects were included in the analysis of coffee consumption and CVD mortality [7,8,9, 11,12,13,14,15, 17, 18, 21,22,23, 25,26,27, 29,30,31, 33,34,35,36, 38, 40, 42,43,44,45]. The pooled RR for highest versus lowest consumption was 0.87 (95% CI 0.82–0.94) with moderate heterogeneity (P = 0.001, I2 = 49.5%) (Table 2, Supplementary Figure 2). The heterogeneity was decreased after excluding two outlying studies [22, 29] (P = 0.04, I2 = 33.9%). We found some evidence of nonlinearity for the association between coffee intake and CVD mortality (P for non-linearity < 0.0001) (Fig. 3). The largest reduction in RR for CVD mortality was observed in consumption of 2.5 cups/day (RR = 0.83, 95% CI 0.80–0.87), compared with no coffee consumption.

Fig. 3
figure 3

Pooled dose-response association between coffee consumption and CVD mortality. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image

A total of twenty six prospective cohort studies from 24 articles including 229,884 deaths and 3,419,278 subjects investigated the association between coffee consumption and cancer mortality [7,8,9,10,11,12,13,14, 16, 17, 24, 26, 27, 31, 33,34,35,36,37,38, 41,42,43, 45]. The pooled RR for highest versus lowest consumption was 0.99 (95% CI 0.94–1.04) with moderate heterogeneity (P < 0.001, I2 = 57.9%) (Table 2, Supplementary Figure 3). The observed heterogeneity was decreased after excluding two outlying studies [8, 26] (P = 0.004, I2 = 51.0%). In the dose–response analysis, a significant linear relationship was not observed (Table 2), and some evidence of a non-linear association between coffee intake and cancer mortality was found (P for non-linearity < 0.0001) (Fig. 4). Compared with no coffee consumption, the pooled RR of cancer mortality for 2–2.5 cups/day of coffee was 0.96 (95% CI 0.94–0.99), and no further cancer mortality reduction was observed with further coffee consumption, with no significant positive association observed at high levels of coffee consumption.

Fig. 4
figure 4

Pooled dose-response association between coffee consumption and cancer mortality. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image

A total of nine studies including 14,694 deaths and 2,101,160 subjects reported the association between coffee consumption and respiratory disease mortality [8, 9, 13, 27, 31, 33, 35, 38, 43]. The pooled RR for highest versus lowest consumption was 0.90 (95% CI 0.75–1.07) (Table 2). There was some evidence of non-linearity in the analyses of mortality from respiratory disease (P for non-linearity < 0.0001) (Fig. 5). Despite significant non-linearity, RRs were reduced sequentially from 1 through to 6.5 cups/day by 28%. A total of four studies including 4010 deaths and 886,933 subjects provided the estimates for the association between coffee consumption and diabetes mortality [8, 9, 27, 33]. The pooled RR for highest versus lowest consumption was 0.76 (95% CI 0.65–0.90) (Table 2). A non-linear inverse association between coffee consumption and diabetes mortality (P for non-linearity = 0.006) was observed, showing most of the reduction in risk for consumption of 2.5 cups/day (RR = 0.70, 95% CI 0.60–0.81) (Fig. 6). A total of six studies including 3330 deaths and 141,783 subjects reported the association between coffee consumption and mortality from non-CVD and non-cancer causes [7, 14, 26, 34, 36, 45]. The pooled RR for highest versus lowest consumption was 0.65 (95% CI 0.51–0.83) (Table 2). The pooled RR for 1 cup/day increment of coffee consumption was 0.93 (95% CI 0.91–0.96), and there was little evidence of non-linearity (P for non-linearity = 0.054) (Fig. 7).

Fig. 5
figure 5

Pooled dose-response association between coffee consumption and respiratory disease mortality. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image
Fig. 6
figure 6

Pooled dose-response association between coffee consumption and diabetes mortality. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image
Fig. 7
figure 7

Pooled dose-response association between coffee consumption and mortality from non-CVD, noncancer causes. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image

Stratified analyses of all-cause mortality for coffee consumption

When we conducted a stratified analysis by alcohol consumption, the inverse association was suggestively stronger among heavy drinkers (RR = 0.80, 95% CI 0.71–0.91) compared with low alcohol drinkers (RR = 0.87, 95% CI 0.79–0.95) (P for low vs. high alcohol consumption = 0.46) (Table 3). By smoking status, the results from non-smokers (RR = 0.85, 95% CI 0.80–0.90) showed a suggestively stronger inverse association than those from smokers (RR = 0.88, 95% CI 0.81–0.96) (P for non-smokers vs. smokers = 0.42). There was some evidence of difference in RRs by geographic region. A significant inverse association was observed among studies in Europe (RR = 0.81, 95% CI 0.74–0.88) and Asia (RR = 0.83, 95% CI 0.76–0.91), while a non-significant inverse association was shown among studies in the US (RR = 0.96, 95% CI 0.89–1.03). No significant differences were found by baseline age, overweight status, sex, and caffeine content of coffee (P for difference > 0.6 in all comparisons).

Table 3 Summary of pooled relative risks (RR) of all-cause mortality for coffee consumption
Full size table

The stratified dose–response meta-analysis by age showed similar results for younger people (< 60 years) (RR = 0.96, 95% CI 0.94–0.98) and older people (≥ 60 years) (RR = 0.96, 95% CI 0.95–0.97) for 1 cup/day increment of coffee consumption (Table 3). However, we found evidence of a non-linear association in both younger people (P for non-linearity < 0.001) and older people (P for non-linearity = 0.003) (Fig. 8). The pooled RR showed the greatest reduction at consumption of 6.5 cups/day in younger people (RR = 0.78, 95% CI 0.75–0.82) and 4.5 cups/day in older people (RR = 0.85, 95% CI 0.80–0.91), compared with no coffee consumption. The inverse association was slightly stronger in younger people (RR = 0.80, 95% CI 0.77–0.82) than older people (RR = 0.86, 95% CI 0.82–0.91) at consumption of 3.5 cups/day. By smoking status, the inverse association was slightly stronger in non-smokers than smokers (P for non-smokers vs. smokers = 0.33) (Table 3).

Fig. 8
figure 8

Pooled dose-response association between coffee consumption and all-cause mortality stratified by age. Solid lines represent relative risk (RR), dashed lines represent 95% confidence intervals

Full size image

Publication bias

There was no evidence of publication bias with Begg’s test for mortality from all-causes (P = 0.31), CVD (P = 0.39), and cancer (P = 0.32), and Egger’s test for mortality from all-causes (P = 0.25), CVD (P = 0.75), and cancer (P = 0.14). In addition, we found no indication of publication bias for mortality from other deaths (Begg’s P > 0.4 and Egger’s P > 0.2 for the all analyses).

Discussion

Findings from our meta-analysis of 40 prospective cohort studies indicate lower mortality from all-causes (12% lower risk) and CVD (13% lower risk) in the comparison of highest versus lowest coffee consumption categories, which was similar by categories of age, sex, overweight status, alcohol drinking, smoking status, and by caffeine content of coffee. Further dose–response analyses showed nonlinear associations of coffee consumption and all-cause and CVD mortality. The strongest association was observed in 3.5 cups/day of coffee consumption for all-cause mortality (15% lower risk) and 2.5 cups/day of coffee consumption for CVD mortality (17% lower risk), compared with no consumption. Regarding cancer mortality, a non-linear association with the lowest risk in 2–2.5 cups/day of coffee consumption (4% lower risk) was observed, compared with no consumption. We found 24% and 35% lower risks of mortality from diabetes and non-CVD, non-cancer causes in the comparison of highest versus lowest coffee consumption categories. The non-linear relationships were found for mortality from respiratory disease and diabetes, and a linear relationship was found for mortality from non-CVD, non-cancer causes.

The results of the present meta-analysis were consistent with previous results [61] in that the largest decrease in RR of mortality from all causes and CVD was observed at moderate coffee consumption of 2–4 cups/day. However, the results of analysis for cancer mortality differed from those of previous meta-analyses, which reported no significant association between coffee drinking and cancer mortality [61] or found a significant inverse association only in non-smokers [50]. In contrast, we included more studies and found a significant non-linear relationship between coffee consumption and cancer mortality. The strongest inverse association was shown at coffee consumption of 2–2.5 cups/day, but an association was not apparent at higher coffee consumption. One reason why the lower risk disappeared with higher intake could be residual cofounding by smoking because smoking is a very strong risk factor for cancer and is correlated with coffee drinking.

We found some evidence of differences by age in the non-linear dose–response analysis, though not in the analysis of highest versus lowest coffee consumption. Younger people (< 60 years) showed a lower risk (20%) of all-cause mortality than older people (≥ 60 years) (14%) at coffee consumption of 3.5 cups/day. The stronger inverse association between coffee drinking and all-cause mortality in younger people than older people was also found in previous studies [8, 33]. Older people are more susceptible to the blood pressure-raising effect of caffeine compared to younger people [47], which may attenuate the inverse association. A stronger inverse association for CVD mortality was shown in younger people compared to older people. However, further study conducting an analysis by age is warranted due to small number of studies included in the analysis by age.

Previous meta-analyses investigating the association between coffee consumption and risk of chronic disease provided the results by some potential modifiers. A stronger inverse association was observed among non-smokers than smokers for type 2 diabetes [62] and CVD [63]. This is similar to our result showing slightly lower mortality in non-smokers. For the results by BMI, on the other hand, we observed a slightly lower risk of death in overweight people, but previous studies have reported higher risks of type 2 diabetes in overweight [62] or obese people [64]. Our results by caffeine content of coffee showed little difference in the risk of death. Unlike our results, some studies suggested lower risk of type 2 diabetes [62] and CVD [63] in people consuming caffeinated coffee than those drinking decaffeinated coffee.

We observed the presence of heterogeneity among the studies. When we excluded outlying studies, the observed heterogeneity reduced, but the inverse association between coffee consumption and risk of mortality did not substantially change. A potential source of heterogeneity may be related to geographic region. We found stronger inverse associations from studies performed in Europe and Asia than those performed in the US. The type of coffee powder, roasting, brewing methods, beverage preparation, sugar and cream added to coffee and cup sizes are different by geographic region and this may affect the association between coffee consumption and mortality. In addition, genetic variants may explain heterogeneities across the studies. The genotype such as cytochrome P-450 1A2 metabolizes caffeine, and thus, is associated with variability in effects of caffeine and coffee consumption [65]. Other possible reasons of heterogeneity could be different distributions of confounding factors and biological differences that Asians are more sensitive to insulin resistance than other races [66, 67].

In the current meta-analysis, we showed the results on mortality from specific causes such as respiratory disease, diabetes, and non-CVD, non-cancer causes. We observed inverse associations between coffee consumption and respiratory disease or diabetes-specific mortality, and these are consistent with the results from previous epidemiological studies examining the association between coffee consumption and disease. Recent review showed that coffee consumption was associated with decreased prevalence of asthma and could be used as a treatment for persistent cough [68]. Regarding diabetes, a recent meta-analysis from 30 cohort studies found that high coffee consumption is associated with 29% lower risk of type 2 diabetes [69]. For mortality from non-CVD, non-cancer causes, we observed a strong inverse association, but, the results should be interpreted with caution because the number of studies included was relatively small. One study included in the analysis of non-CVD, non-cancer causes showed that suicide and respiratory disease were the major causes of other causes of death [26]. Recent studies reported that coffee intake was associated with lower risk of depression [70] and suicide [71] may be through increasing turnover of serotonin [71]. Also, as mentioned before, we found an inverse association between moderate coffee consumption and respiratory disease-specific mortality. The inverse association between coffee and depression or respiratory disease-specific mortality may have contributed to a strong inverse association between coffee consumption and mortality from non-CVD, non-cancer causes.

Although it is still unclear, several potential mechanisms could explain a beneficial effect of coffee consumption on health. Increased oxidative stress and prolonged inflammation may contribute to premature death by increasing the risk of chronic disease [72, 73]. Coffee contains various antioxidant components such as caffeine, chlorogenic acid, melanoidins, cafestol, kahweol, and trigonelline [74]. Previous experimental studies reported a significant increased level of serum antioxidant enzymes (i.e. glutathione peroxidase and glutathione-S-transferase) [75, 76] and decreased levels of lipid peroxidation [75] and oxidative DNA damage [77, 78] among the subjects who consumed coffee. Many human studies also have shown that coffee intake may be associated with the levels of pro-inflammatory biomarkers including tumor necrosis factor alpha, [79] C-reactive protein [80, 81], and interleukin 18 [82], and increase the levels of anti-inflammatory biomarkers such as adiponectin [81, 83]. These antioxidant and anti-inflammatory properties of coffee compounds may lead to a decreased risk of mortality through slowing the development of some major chronic diseases including diabetes, CVD, and cancers.

There are several strengths in the current meta-analysis. To the best of our knowledge, this is the first meta-analysis to examine the association between coffee drinking and mortality using both high versus low analyses and linear and non-linear dose–response analyses. Compared to previous meta-analyses [50, 51, 84] we added recent studies increasing the number of subjects [10, 13, 23, 28, 30, 38, 41,42,43,44,45]. Due to the large number of studies, we could investigate the association of coffee intake on all-cause mortality in various subpopulations by age, sex, geographic region, overweight status, alcohol consumption, and smoking status. Unlike previous meta-analyses, we performed meta-analyses of coffee consumption and mortality from less common causes, including respiratory disease, diabetes, and non-CVD, non-cancer causes besides all-cause, CVD, and cancer mortality.

The present meta-analysis has several limitations. First, because of the observational design of the included studies, unmeasured or residual confounding is possible. However, most of the studies included in meta-analysis provided estimates that adjusted for various mortality risk factors, and results were observed in stratified meta-analyses by major potential confounders. Second, misclassification of actual coffee amounts consumed might occur because coffee consumption was assessed by self-reported questionnaires and the size of coffee cup varied. However, any misclassification in coffee intake categories is likely to be non-differential and would have probably led to dilute the association rather than to strengthen it, pushing estimates towards the null value in the high versus low analysis. Third, the highest categories of coffee consumption differed across the studies. However, we examined the dose–response relation between coffee consumption and risk of mortality by conducting linear and non-linear dose–response analyses as well as high versus low analysis. Lastly, the quality assessment indicates high qualities of included studies, but most of studies did not provide the details for type of coffee, brewing methods, or preparation which could help to understand the potential effect of coffee intake on risk of mortality.

In conclusion, our findings provide further evidence for a beneficial effect of moderate coffee consumption (e.g. 2–4 cups/day) on the risk of mortality from respiratory disease, diabetes, and all non-CVD, non-cancer causes as well as mortality from all causes, CVD, and cancer. The inverse association between coffee drinking and all-cause mortality was consistent in various subpopulations by overweight status, alcohol consumption, smoking status and by caffeine content of coffee. Future large prospective studies with detailed information of coffee preparation, sugar and cream added to coffee, or genotype of population could provide more definitive conclusion on the potential effects of coffee intake on risk of mortality.