Universidade do Porto, Porto, Portugal. Correspondence

We systematically reviewed the literature on the association between coffee consumption and gastric cancer and performed a meta-analysis of the results. Published cohort and case-control studies were identified in PubMed and reference lists. Random effects meta-analysis was used to pool effects from 23 studies, and heterogeneity was explored by stratification and meta-regression. The odds ratio (OR) for the overall association between coffee and gastric cancer (highest vs. lowest category of exposure) was 0.97 (95%CI: 0.86-1.09), similar for cohort (OR = 1.02; 95%CI: 0.76-1.37) and case-control studies (population-based: OR = 0.90; 95%CI: 0.70-1.15; hospital-based: OR = 0.97; 95%CI: 0.83-1.13). The OR was 1.26 (95%CI: 1.02-1.57) when considering five studies conducted in the USA, 0.97 (95%CI: 0.82-1.14) for the five Japanese studies, 0.98 (95%CI: 0.81-1.17) for the six studies from Europe, and 0.64 (95%CI: 0.47-0.86) for the two studies from South America. In this meta-analysis we found no adverse effect of coffee associated with gastric cancer. Knowledge on the level of exposure to different coffee constituents may provide a deeper understanding of this reassuring result and the real role of coffee on cancer risk.


Introduction
Coffee is one of the most popular beverages, with a yearly world average consumption of 1.1kg per capita, which reaches 4.5kg in industrialized countries 1 .
Roasted coffee is a complex mixture of more than a thousand chemicals.These constituents have been described as having genotoxic and mutagenic properties, but also antimutagenic and antioxidant activities and the capacity to inhibit cancer-promoting agents 2,3 .Caffeine appears to disturb cell cycle checkpoint integrity, alter mechanisms of DNA repair, modify the apoptotic response, and potentiate a variety of DNA-damaging agents 2,3 .However, it can also inhibit carcinogenesis in vivo and alter carcinogen metabolism, decreasing the cytotoxic, cytostatic, or mutagenic activity of aromatic DNA-damaging compounds through a decrease in the concentration of free aromatic procarcinogens available for cytochrome activation 4,5 .The protective effects of coffee were partially ascribed to the potential for kahweol and cafestol palmitates to convert rapid acetylators into a slow acetylator phenotype 6 .These diterpenes may also have anti-inflammatory and anti-carcinogenic properties by interfering with nitric oxide 7 , prostaglandin E2 production, and cyclooxigenase-2 expression 8 .The antioxidant attributes of coffee may contribute to a protective role against cancer 9,10 .
Cad. Saúde Pública, Rio de Janeiro, 22 (5):889-900, mai, 2006 These contrasting effects parallel the results of previous epidemiological studies that disclosed no clear-cut effect of coffee consumption on cancer.Reviews and meta-analyses of available studies showed a 20% increased urinary tract cancer risk in coffee consumers 11 , no effect on pancreatic cancer risk 12 , and a 25% reduction in colorectal cancer risk 13,14 .
Investigation of other cancer sites has been less extensive.An earlier review of the association between coffee and gastric cancer identified two cohort studies and nine case-control studies published until 1996 15 and underscored the contradictory nature of the results.Coffee is such a common exposure that any small effect can result in a large population impact.Therefore, we systematically reviewed and performed a meta-analysis of the published data addressing the association between coffee consumption and gastric cancer.

Material and methods
We performed a systematic review of case-control and cohort studies evaluating the association between coffee consumption and gastric cancer.We used PubMed (http://www.ncbi.nlm.nih.gov/entrez) to identify studies published through December 2004, under the searching expression "(stomach cancer OR gastric cancer OR cardia cancer) AND (coffee OR lifestyle OR tea)".The reference lists provided by the identified papers was additionally hand-searched.We evaluated full papers published in English, Spanish, French, Portuguese, and Italian, and English abstracts of full papers written in other languages.
Two reviewers extracted information from each study following a previously defined data collection procedure.Discrepancies in the evaluation of the articles were resolved by consensus, involving a third researcher.The protocol for data extraction covered: study design (case-control, population-based or hospital-based, or cohort, with respective follow-up time; number of subjects); histological confirmation of cases; histological type and location of gastric neoplasias; evaluation of exposure (interviewer or self-administered questionnaire as measuring instrument; type of coffee, reference class, exposure strata, timing of exposure); risk estimates for the association between coffee consumption and gastric cancer, considering two exposure levels (lowest vs. highest); precision estimates (confidence intervals, P values, and number of participants in each exposure category); control of confounding factors; and country of origin.
In one study 16 there were only one case and one control in the upper exposure category and we opted for the results based on the immediate class of consumption.
We excluded from meta-analysis studies that did not provide relative risk estimates and respective variance, or the information needed to calculate it.When a study provided more than one estimate, we selected the one adjusted for the largest set of variables, and when results were available according to gender or ethnicity we included all estimates in the final analysis as if obtained from different studies.
Combined risk estimates and 95% confidence intervals were computed using the random effects method, and statistical tests for homogeneity 17 were performed.Heterogeneity was investigated by subgroup analysis, looking at the magnitude of the combined risk estimates in each stratum as well as to the respective tests of heterogeneity, and meta-regression, to assess the independent contribution of each variable to explain heterogeneity.Publication bias was examined through funnel plot visual analysis, the Begg adjusted rank correlation test 18 , and the Egger's regression asymmetry test 19 .A 0.1 level of significance was used in the statistical tests.The software STATA version 8.0 (Stata Corporation, College Station, USA) was used in all analyses.

Results
We identified 40 publications.The English abstracts of two articles published in Chinese 20,21 , two in Russian 22,23 , one in Japanese 24 and one in Serbian 25 provided insufficient information to be included in the final analysis.One of these studies 25 described a negative correlation and another stated that in large amounts coffee increased the risk 23 .Three publications 26,27,28 provided no quantitative risk estimates but stated that no statistically significant association was found.One study 29 only stated the finding of a non-statistically significant association in males, and a significant association in females (hazard ratio = 2.54) without providing any further statistical estimates.Two studies 30,31 were later reported in more detail 32,33 , and the more recent publications were used.Four reports were excluded from analysis because tea and coffee were considered combined (OR = 3.2) 34 , coffee assessed as a preference (OR = 1.02) 35 , only cardia cancers (OR = 1.3) 36 were evaluated, or results were presented separately for intestinal (OR = 0.50) and diffuse (OR = 1.98) type gastric cancer using the same control group 37 .
One case-control study was not considered for meta-analysis because no precision estimates were available 38 .Table 1 briefly characterizes the 17 studies excluded from the final analysis.
Fourteen studies used only histologically confirmed cases, two 40,50 stated that respectively "most" and 90.6% of cases had histological confirmation, five 46,52,54,56,57 only indicated having used cancer registries to identify cases, and two 16,58 did not provide this information.
Six studies 45,47,52,54,55,57 used a self-administered questionnaire, one 16 did not specify the measuring instrument, and the remaining 16 collected information by interview.Among the 16 case-control studies, three evaluated coffee exposure five or more years prior to the interview 44,48,51 , two did not specify the timing of exposure 16,42 , and the rest assessed coffee consumption within five years before onset of disease.In 14 studies, the reference category was non-coffee drinkers, but 9 included subjects drinking more than one cup of coffee per week in the reference group.Nine studies presented results considering only two groups of exposure, and 14 defined three or more categories of consumption.Only two reports specified the type of coffee consumed as being Turkish coffee 16 and caffeine-containing coffee 30 .One study evaluated the consumption of hot coffee 32 .
From 9 studies only crude risk estimates could be used, 10 controlled for the potential confounding effect of tobacco smoking, and 4 provided results adjusted for the consumption of fruit or vegetables.
The studies that adjusted coffee effect for smoking resulted in a combined odds ratio of COFFEE AND GASTRIC CANCER: A META-ANALYSIS 891 0.91 (95%CI: 0.75-1.11,heterogeneity test: p = 0.02), and 0.94 (95%CI: 0.79-1.12,heterogeneity test: p = 0.99) for the studies adjusting for fruit or vegetable consumption.The combined risk estimates including only the studies with all cases histologically confirmed was 0.97 (95%CI: 0.84-1.12,heterogeneity test: p = 0.06).
A multivariate analysis including all the above variables showed that risk estimates differed significantly according to country of origin, with North America studies presenting a significantly higher risk (coefficient = 0.45, p = 0.003), number of exposure strata (coefficient = -0.21,p = 0.06, for 3 or more groups compared with two), and reference category (coefficient = -0.27,p = 0.02, for some drinking in reference group compared to non-exposed to coffee).
Neither visual inspection nor tests of statistical significance resulted in funnel plot asymmetry (Figure 2).
As described above, most of the studies excluded from the meta-analysis were not considered simply because they did not provide the necessary risk and precision estimates, were duplicate publications, or analyzed exposures  30 Hospital based La Vecchia et al. 32 Jarebinski Yugoslavia Case-control -"coffee in large -Article in Russian et al. 23 Hospital based amounts... …increases the risk" Jarebinski Yugoslavia Case-control -"negative correlation" -Article in Serbian et al. 25 Hoshiyama & Japan Case-control ≥ 10 cups/week vs. 0.8 (0. to coffee and tea together.Relaxing the inclusion criteria to accommodate the remaining studies 35,36,37 the summary risk estimate was 1.05 (95%CI: 0.90-1.23,heterogeneity test: p < 0.001), similar to that obtained in our main analysis.

Discussion
This meta-analysis of studies published during the last three decades showed no overall effect of coffee consumption on gastric cancer risk.However, we observed substantial methodological differences between studies that have potential effect on the risk estimates.
Most studies presented results on the association between coffee and gastric cancer as secondary data analysis or part of confounder evaluation 17,33,38,39,40,41,42,46,47,49,50,51,56,57,58 , and it is unlikely that this specific result influenced publication.The studies excluded from the analysis presented non-significant associa-tions between coffee and gastric cancer, and after changing our inclusion criteria, enlarging the number of studies combined, the results remained virtually unchanged.Publication bias is an improbable cause of our findings, as supported by the funnel plot analysis and the result of the regression asymmetry test, and it is unlikely that unpublished results would change our conclusions, because the bias is due to the over-publication of positive findings 59 .
In this analysis, several sources of heterogeneity are likely, even if most risk estimates from individual reports were not significantly different as assessed by statistical tests.
In case-control studies, coffee consumption among controls may not represent the target population, and bias is even more probable with hospital controls.However, considering the results from hospital-and population-based studies, this appears unlikely.Symptoms, namely heartburn, are associated with coffee intake, and cases might avoid coffee.This differential information bias could lead to a misleading Risk estimates .1 .5 1 2 10
Meta-analysis of studies evaluating the association between coffee consumption and gastric cancer*.
The classification of exposure differs considerably across the reviewed studies, and the results were shown to be significantly different according to the number of consumption categories evaluated or the characteristics of the reference class regarding consumption.The lower risk estimates when consumption is evaluated in several categories could be explained if coffee had a protective effect that would be stronger when the difference between the groups of highest and lowest exposure was larger.On the other hand, in opposition to what is observed, if coffee had some effect on cancer risk we would expect the combined estimates to approach the null when the reference class includes coffee drinkers, and this is likely to be a chance finding.The method of questionnaire administration and the timing of exposure evaluation did not significantly influence our results, and our conclusions of lack of positive association between coffee and cancer do not appear to be affected by the specificities of exposure evaluation.
In the meta-analysis, we opted for individual risk estimates based on exposure categories with different cut-off points, and a dose-response analysis would have given more information on the underlying association.However, more than one-third of the studies allowed no such analysis, and we opted to include the 9 studies providing data for only two consumption categories.This option proved to be adequate, since the combined risk estimates differed across studies presenting results for two or more groups of exposure.
Coffee consumption tends to be associated with tobacco smoking 54,60 , but many studies did not account for this potential confounding in the data analysis.Moreover, adding to residual confounding by tobacco smoking, other factors may be influencing the estimates for the association between coffee and gastric cancer, since non-coffee drinkers may differ from the general population of coffee drinkers concerning other exposures such as tea, alcohol, or fruit and vegetable intake 61,62 .The studies included in our review rarely considered confounding or interaction from these variables.
We observed differences in risk estimates according to the geographical origin of the study.Methodological options in study design did not explain such differences, but none of the reviewed studies considered in detail the characteristics of the coffee consumed.It is known that the type of coffee beans 63,64,65 , roasting procedure 66,67 , and specific method of preparing the coffee can influence composition 68,69,70 .We can speculate on any potential adverse effect of specific constituents present in coffee consumed in North America.The underlying risk of gastric cancer in each population, international differences in the typical amount of coffee consumed, coffee type, or brewing method 70 may contribute to these differences.Studies presenting different risk estimates across strata of gender 48,51,58 , ethnicity 41 , or generation of Japanese migrants to Hawaii 40 favor this hypothesis of various exposures under the label coffee, but the alternate explanation of confounding seems sounder.
The complexity of coffee composition and the multiple social contexts underlying consumption make the evaluation of the effect of coffee on gastric cancer very difficult.Human experimental studies on such associations are unlikely, making observational studies the best available source of evidence on risk.Overall, it is reassuring that this meta-analysis showed no adverse effect of coffee associated with gastric cancer.Knowledge on the level of exposure to different coffee constituents may provide a deeper understanding of the real role of coffee on cancer risk and ultimately allow the design of safer beverages.

Figure 2 Funnel
Figure 2Funnel plot of studies evaluating the association between gastric cancer and coffee consumption*.

Table 1
Summary characteristics of studies excluded from the meta-analysis.

Table 2
Summary characteristics of studies included in the meta-analysis.