Introduction

A large body of literature points to the health benefits of long-term coffee consumption in relation to diabetes, cardiovascular disease (CVD), and certain cancers [1]. A recent report of the IARC monographs Working Group regarding the possible carcinogenicity of coffee found no conclusive evidence for a carcinogenic effect of drinking coffee, focusing the attention on drinking very hot beverages (as probable cause of esophageal cancer) rather than the drinks themselves [2]. In contrast, coffee contains several biochemically active compounds that may exert beneficial effects on human health [3]. Although the mechanisms of action are not entirely understood, the effects of antioxidant compounds may explain the potential benefit of coffee against those conditions associated with a chronic state of subclinical inflammation, such as CVD and cancer [4–7]. Indeed, coffee has been reported to exert the main anti-oxidant effect among the most commonly consumed foods in the diet [8]. Among the components recently studied as responsible for such action, polyphenols demonstrated anti-oxidant and anti-inflammatory capacity in laboratory studies and their consumption has been associated in prospective cohort studies with reduced risk of diabetes, CVD, and some cancers [9–11].

Studies exploring the relation between coffee consumption and mortality have produced puzzling results. Relatively recent summary analyses of cohort studies provided quantitative evidence that coffee intake might be inversely related to all cause and, probably, CVD mortality [12, 13]. Other studies reported a non-linear dose-response association, pointing out the possibility of unfavorable effects of coffee intake at high concentrations [14, 15]. However, findings across studies are not consistent. Older studies may be affected by methodological issues, such as lack of adjustment for important confounding factors as smoking status. Moreover, some meta-analyses on specific health outcomes for which smoking habit represent a known risk factor, suggested that adjustment by smoking status may not be sufficient to overcome the strong association between coffee drinking and smoking, and reduced the confounding effect by stratifying the analysis for smokers and non-smokers [16, 17]. A major limitation of previous meta-analyses on coffee consumption and mortality risk is lack of stratification by smoking status. The aim of the present study was to perform a dose-response meta-analysis of prospective cohort studies exploring the association of coffee consumption with mortality risk. We also examined the dose-response mortality risk by smoking status in order to compare the shape of association between smokers and non-smokers. Additional analyses were performed to evaluate whether other potential confounders exist.

Methods

Study selection and data extraction

A systematic search on PubMed (http://www.ncbi.nlm.nih.gov/pubmed/) and EMBASE (http://www.embase.com/) databases of studies published up to December 2015 was performed with the search terms “coffee” and “mortality”. Inclusion criteria were: (1) had a prospective design; (2) evaluated association between coffee intake and risk of mortality; (3) assessed and reported hazard ratios (HRs) and the corresponding 95 % CI for mortality for ≥3 exposure categories; and (4) provided defined amount of coffee consumption (i.e., cups per day) per category of exposure. Exclusion criteria were the following: (1) reported insufficient statistics and (2) assessed composite outcome from which was not possible to derive mortality risk (i.e., incidence of CVD event, including cardiovascular death). Reference lists of studies of interest were also examined for any additional study not previously identified. If more than one study was conducted on the same cohort, only the study including the entire cohort or the longest follow-up was included.

Data were abstracted from each identified study by using a standardized extraction form. The following information was collected: (1) first author name; (2) year of publication; (3) study cohort name and country; (4) number, gender, and age (mean or range) of participants; (5) follow-up period; (6) endpoints and cases; (7) distributions of cases and person-years, HRs and 95 % CIs for all categories of exposure; (8) median intake of coffee per each category of exposure; (9) covariates used in adjustments. This process was independently performed by two authors (G.G. and A.M.) and discrepancies were discussed and resolved by consensus.

Statistical analysis

Outcomes evaluated in the analyses included all-cause, CVD, and cancer mortality. When coffee consumption was reported by ranges of intake, the midpoint of the range was used. When the highest category was open-ended, we assumed the width of the category to be the same as the adjacent category. When the lowest category was open-ended, we set the lower boundary to zero. Two-stage random-effects dose-response meta-analysis was performed to examine linear and non-linear relationship between coffee intake and all-cause, CVD, and cancer mortality. In the first stage the method reported by Greenland and Orsini (generalized least-squares, GLS) was used to calculate study-specific coefficients on the basis of results across categories of coffee intake taking into account the correlation within each set of retrieved HRs [18, 19]. Non-linear dose–response analysis was modeled using restricted cubic splines with 3 knots at fixed percentiles (25, 50, and 75 %) of the distribution [20]. We combined the coefficients that had been estimated within each study by performing random-effects meta-analysis. In linear dose-response meta-analysis the method of DerSimonian and Laird was used and in non-linear dose-response meta-analysis the multivariate extension of the method of moments was used to estimate the relative risks (RRs). We calculated an overall P value by testing that the 2 regression coefficients were simultaneously equal to zero. We then calculated a P value for non-linearity by testing that the coefficient of the second spline was equal to zero. A number of sensitivity analyses were conducted to test stability of results: (a) by grouping studies according the level of adjustment for smoking-related variables; (b) by excluding one study at the time; (c) by excluding studies that did not report the number of cases, individuals, and person-years for each category of exposure; and (d) by converting the volume of cups of coffee into a homogeneous measure (1 cup = 150 ml coffee) in those studies in which volume was given. To test for potential confounders/effect modifiers, dose-response analyses for all-cause, CVD, and cancer mortality were performed according to some variables of interest for which stratified data was available, such as smoking status [smokers (current and former)/non-smokers (never)] and other factors, including gender (men/women), year of publication, geographical area, and type of coffee. Publication bias was assessed with Egger’s regression test. Statistical heterogeneity between studies was assessed using the Chi square test (defined as a P value less than 0.10) and quantified through the multivariate generalization of the I 2 statistic: no, low, medium, and high heterogeneity were defined by I 2 values <25, <50, <75, and ≥75 %, respectively. All analyses were performed with R software version 3.0.3, dosresmeta and mvmeta packages (Development Core Team, Vienna, Austria).

Results

Study characteristics

The selection process of studies potentially relevant for the meta-analyses is presented in Fig. 1. Out of 470 screened studies, a total of 31 studies [21–51] involving 1,610,543 individuals and 183,991 cases of all-cause, 34,574 of CVD, and 40,991 of cancer deaths (2 studies [23, 26] did not provide number of cases) were included in the meta-analyses to test the association between coffee consumption and mortality risk. A description of the studies and cohorts included is presented in Table 1. Eleven studies were conducted in US, 14 were settled in Europe, and 7 in Asia (6 in Japan and 1 in Singapore). Follow-up periods ranged between 6 and 28 years, providing reasonable time to observe the outcomes studied. Sixteen studies [25, 26, 28–30, 36, 37, 39–43, 45–47, 50] provided gender-specific risk estimates and 6 studies [43, 46, 48–51] reported stratified analysis for smokers and non-smokers. Three studies [22, 24, 25] did not provide 95 % CIs and were excluded from the analyses because they could only be included by deriving missing information as crude risk estimates, which is poorly informative for the purpose of the present study. Moreover, the same studies considered “non-smokers” as “never and former smokers”, which did not fit our inclusion criteria for the analyses. All other studies adjusted analyses for potential confounding factors associated with the outcomes of interest, including age, gender, body mass index, and smoking status. Regarding the latter, 8 studies provided adjustment only for current/non-smokers [21, 23, 25, 26, 28, 35, 45, 51], 9 adjusted for current/former/never smokers [27, 29, 30, 34, 36–38, 42, 47], while 12 further adjusted for further information (i.e., number of cigarettes) [22, 24, 31–33, 40, 41, 43, 44, 46, 48, 50]. Adjustment for other covariates was not equal across studies: 8 included evaluation of health parameters (i.e., blood pressure, lipids, etc.) [23–25, 27–30, 32], 18 included prevalence of chronic non-communicable diseases [21, 25, 26, 30, 31, 34–36, 40–46, 48, 50, 51], and 17 included information on other diet-related factors [29, 33, 37–46, 48–51].

Fig. 1
figure 1

Process of study selection for inclusion in the meta-analysis on all-cause, CVD, and cancer mortality

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

All-cause and cause-specific mortality

The dose-response meta-analyses for all-cause (including 24 studies [21, 23, 26, 27, 29–36, 38, 40, 42–51)], CVD (including 23 studies [26–30, 32, 33, 35–41, 43–51]), CHD (including 12 studies [28–30, 36, 38–41, 43, 46, 48, 50]), stroke (including 9 studies [36, 38–41, 43, 46, 48, 50]), and cancer mortality (including 15 studies [26, 27, 31, 33, 35, 40, 42–44, 46–51]) are showed in Fig. 2. Compared with no coffee consumption, the summary RR of all-cause mortality for 4 cups/day of coffee was 0.84 (95 % CI = 0.81, 0.88; I 2 = 83 %, P heterogeneity  < 0.001) and RR of CVD mortality was 0.83 (95 % CI = 0.75, 0.92; I 2 = 92 %, P heterogeneity  < 0.001), while increased intake was associated with no further lower risk (Table 2). Similar risk estimates were found for CHD mortality while risk of stroke was slightly lower (RR = 0.84, 95 % CI = 0.71, 0.99; I 2 = 95 %, P heterogeneity  < 0.001; and RR = 0.72, 95 % CI = 0.6, 0.87; I 2 = 89 %, P heterogeneity  < 0.001). No significant association was found between coffee consumption and cancer mortality risk. Egger’s regression test provided no evidence of substantial publication bias. A sensitivity analysis by level of adjustment by smoking status did not change previous findings besides resulting in a stronger association between coffee and CVD mortality in the model including studies adjustment for additional variables related to smoking status (Supplementary Table 1). Additional sensitivity analyses by excluding one study at the time, by excluding studies due to lack of number of individuals/cases for each category of exposure, and by uniformed converted doses of cup of coffee for those studies reporting the exact amount of coffee per cup did not show differences from main analyses (data not shown).

Fig. 2
figure 2

Dose-response association between coffee consumption and all-cause, CVD, and cancer mortality. Solid lines represent relative risk, dashed lines represent 95 % confidence intervals

Full size image
Table 2 Dose-response meta-analysis of prospective studies on coffee consumption and all-cause, cardiovascular, coronary heart disease, stroke, and cancer mortality
Full size table

Stratified dose-response analyses

When performing a dose-response analysis on population stratified by smoking status, heterogeneity was reduced in all models (Table 3). No differences were found between smokers and non-smokers for all-cause and CVD mortality risk (Fig. 3), both significantly reduced for higher compared to no consumption of coffee, with no/small evidence of heterogeneity or publication bias (Table 3); in contrast, cancer mortality was significantly decreased only when considering non-smokers, while increased in smokers (Table 3 and Fig. 3). A sensitivity analysis by level of adjustment for smoking-related variables conducted in all models on smokers showed stable results for all-cause mortality (despite with evidence of heterogeneity in the most adjusted studies), significant decreased risk of CVD mortality when considering the most adjusted studies, and non-significant results for cancer mortality (Supplementary Table 2). When considering non-smokers, all models showed significant decreased risk with no/small evidence of heterogeneity or publication bias: a linear dose-response analysis showed a significant decreased risk by 6 % for each additional cup of coffee per day consumed for all-cause and CVD mortality (RR = 0.94, 95 % CI = 0.93, 0.96 and RR = 0.94, 95 % CI = 0.91, 0.97, respectively) and significant decreased risk of 2 % for cancer mortality (RR = 0.98, 95 % CI = 0.96, 1.00).

Table 3 Dose-response meta-analysis of prospective studies on coffee consumption and all-cause, cardiovascular, coronary heart disease, stroke, and cancer mortality stratified by smoking status
Full size table
Fig. 3
figure 3

Dose-response association between coffee consumption and all-cause, CVD, and cancer mortality stratified by smoking status. Solid lines represent relative risk, dashed lines represent 95 % confidence intervals

Full size image

Stratified dose-response analyses by gender, year of publication, type of coffee, and geographical area showed less remarkable differences in the association between coffee consumption and the outcomes explored between the strata (Supplementary Table 3). Summary risk estimates for men and women were similar between genders (Supplementary Figure 1). Also the results stratified by geographical area did not show differences for most of the analyses, despite coffee consumption was associated with risk of CVD mortality in a U-shaped, rather than J-shaped manner (Supplementary Figure 2). In contrast, older studies (publication year prior 2010) and consumption of caffeinated, rather than decaffeinated coffee were mostly contributing to the increased risk of cancer mortality associated with increased intake of coffee (Supplementary Figure 3 and Supplementary Figure 4, respectively).

Discussion

After analyzing results from existing cohorts on coffee consumption and stratifying analyses by smoking status, we found evidence of a decreased risk of all-cause, CVD, and cancer mortality among non-smokers. Compared with previous meta-analyses [14, 47], we included a higher number of individuals and we performed dose-response analyses for a number of strata to test for potential confounders/effect modifiers, including smoking status, but also gender, year of publication, geographical area, and type of coffee. In smokers, a J-shaped association with all outcomes was found. The slope for benefit was steeper at the lower range for CVD mortality, but appeared linear for all outcomes. Among other variables of interest, there were no particular findings to be noted.

Previous meta-analyses reported that the association of coffee drinking with mortality risk may not be linear. Potential reasons for any adverse effect of coffee have frequently been attributed to a trigger of coronary events induced acutely by high doses or caffeine [52, 53]. However, this hypothesis is not confirmed in pooled analyses of prospective cohort studies on coffee and caffeine intake and risk of atrial fibrillation and heart failure [54, 55]. While caffeine intake may have some acute effects, habitual, rather than occasional intake of coffee has been demonstrated to reduce inflammatory and glycemic markers [7, 56, 57]. Moreover, habitual coffee intake may also lead to desensitization to the acute effects of caffeine [58], and previous investigations showed potential beneficial effects of coffee not related with caffeine intake [59]. Findings from meta-analyses on CVD and mortality risk were in accord on the decreased risk associated with coffee consumption. We confirmed the association, reporting that coffee was associated with decreased risk of all-cause and CVD mortality, with no effect modification from any factor except smoking, which weakens the strength of the association when considering smokers. For cancer mortality, previous studies reported increased risk associated with higher coffee intake. In contrast, some recent studies suggest that coffee may be associated with decreased risk of some cancers, including colorectal, oral, endometrial, prostate, and liver cancers [60–63]. In the present meta-analysis we observed a different pattern for risk of cancer mortality when stratified by smoking: while a suggestive increased risk was observed in smokers, we observed a linear decreased risk of cancer mortality when the analysis was restricted to non-smokers. While it is hardly plausible that any biological effect of coffee causally differs by smoking status, given that coffee drinking and smoking are correlated, and that smoking is the strongest risk factor for cancer, we believe that residual confounding by smoking is the most likely the explanation for such increased risk and that the association between coffee and cancer mortality is difficult to isolate when considering smokers. These findings argue against any adverse effects of coffee on cancer risk, at least among non-smokers, and in contrast are consistent with potential beneficial effects.

The polyphenol content of coffee has been considered as the main biological explanation for coffee’s benefits on human health. Coffee is one of the main contributors to polyphenol content in the diet of European individuals, accounting for up to 40 % of the total polyphenol intake [64–68]. The polyphenol mainly represented in coffee are phenolic acids, which have been only recently considered in epidemiological studies showing potential benefits toward metabolic disorders [69]. Chlorogenic acids, the most abundant group of phenolic acids contained in coffee, showed improvements in blood pressure alterations, and the ability to affect some metabolic pathways, for instance by improving glucose metabolism and decreasing inflammation and endothelial dysfunction [70, 71]. This family of compounds also demonstrated antitumor properties by inhibiting key enzymes involved in tumor genesis and metastasis [72]. Among other components, kaweol and cafestol have been proposed as possible antitumor agents due to their capacity of regulating angiogenesis, apoptosis and inflammation process [73, 74].

Results of the present study should be considered in light of some limitations. First, some analyses reported moderate yet significant heterogeneity. As previously suggested, a number of factors may explain differences across studies, including type of coffee powder (Arabica or Robusta), roasting, and beverage preparation. Moreover, genetic variants associated with caffeine metabolism are not considered in prospective cohort studies included in the meta-analysis and may explain part of the heterogeneity [75]. Second, although most of the included studies reported adjusted measures and we further stratified by the main potential confounding factors, the observational design did not exclude the presence of residual or unmeasured confounding from other mortality risk factors. Because coffee was assessed before outcome, recall bias is unlikely. However, misclassification of the actual amounts consumed may have affected the dose-response relation. Reverse causation may have affected the results if individuals changed coffee intake due to a diagnosed medical condition or disease; however, any such effects would be muted in studies with long duration. Finally, time-related variables, such as period of evaluation (baseline, continuous, etc.) and duration of coffee consumption have not been investigated.

In conclusion, coffee consumption is associated with decreased risk of all-cause, CVD, and cancer mortality when considering non-smokers. The contrasting results for higher intake of coffee among smokers depend most likely on the effect modification of smoking habit and more realistic association with mortality risk is provided in non-smokers. Future studies should also distinguish between smoking-related cancer mortality, in order to add further details to the associations retrieved and confirm those on cancer mortality as well as to explore other potential confounding factors that may justify previous findings (i.e., drinking very hot beverage, including coffee). Moreover, further details related to coffee type and preparation could also provide future insights on the potential effects of coffee consumption on human health. However, from a public health point of view, there is no evidence of harmful effects of coffee drinking with regard to the outcomes investigated, rather a potential beneficial effect evidenced in non-smokers.