Methods

Study design and participants

This study was conducted within the NLCS, which started in September 1986 with the enrolment of Dutch men (n = 58 279) and women (n = 62 573) women aged 55–70 years.12

For reasons of efficiency, the case–cohort approach was used for data processing and analysis.13 Cases were derived from the entire cohort, and the number of person-years at risk for the entire cohort was estimated from a subcohort of 5000, who were randomly sampled from the total cohort at baseline. Person-years at risk were calculated from the start of the study until oesophageal or gastric cancer diagnosis, death, emigration, loss to follow-up or end of follow-up, whichever occurred first.

Follow-up for cancer incidence was performed by record linkage to the Netherlands Cancer Registry (NCR) and the nationwide network and registry of histopathology and cytopathology in The Netherlands (PALGA).14 The completeness of cancer follow-up is ⩾96%,15 and follow-up of the subcohort was nearly 100% complete (only one male subcohort member was lost to follow-up) after 16.3 years (from 17 September 1986 until 31 December 2002). The following numbers of incident, microscopically confirmed, primary carcinomas were identified: 130 OSCC16 (ICD-O-3 C15), 181 OAC16 (C15), 206 GCA (C16.0) and 594 GNCA (C16.1–C16.9) (fig 1). Excluded were non-microscopically confirmed cancers, which formed 1% of both oesophageal and gastric cancers. The group of GNCA included cancers with a lesion with overlapping subsites of the stomach (C16.8, n = 160) and some gastric not otherwise specified cancers (C16.9, n = 75), raising the possibility that some cardia cancers might be included in the non-cardia category. However, as we found risk estimates to be similar in separate analyses of gastric cancers of specified sites (C16.1–C16.5) and other gastric cancers (C16.6–C16.9) (data not shown), we combined the groups in the analysis.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Flow diagram of subcohort members and cases on whom the analyses were based. *PALGA: nationwide network and registry of histopathology and cytopathology in The Netherlands.

We excluded subcohort members who reported having prevalent cancer other than skin cancer at baseline. Also excluded were cases and subcohort members with incomplete or inconsistent dietary data17 or missing data on exposure or confounding variables. Figure 1 shows that 3962 subcohort members and 107 OSCC, 145 OAC, 164 GCA and 491 GNCA cases were available for analysis.

Exposure information

All cohort members completed a self-administered questionnaire at baseline. This questionnaire included a 150-item food frequency questionnaire (FFQ), with questions on alcohol consumption, and questions on other cancer risk factors, such as smoking habits, level of education, body mass index (BMI), physical activity, family history of cancer and use of medication.

We asked about the habitual consumption of alcohol during the year preceding the start of the study, and this was measured by six items: (1) beer; (2) red wine; (3) white wine; (4) sherry and other fortified wines; (5) liquor types containing on average 16% alcohol; and (6) (Dutch) gin, brandy and whisky. Questions were asked about the frequency of consumption and the number of glasses consumed on each drinking occasion. For analysis, we combined (2), (3) and (4) into “wine”, and (5) and (6) into “liquor”. Mean daily alcohol consumption was calculated using the Dutch food-composition table.18 Based on a pilot study, a glass of beer, wine and liquor was assumed to contain 200, 105 and 45 ml of the beverage, respectively (unpublished findings). For “beer” and “other alcoholic beverages”, participants could indicate whether 5 years ago they drunk (1) more than, (2) equal amounts of or (3) less than today. The fourth answer option was (4) “I never use this”. Using these questions, we selected those participants with stable alcohol consumption to perform a sensitivity analysis to investigate the robustness of our results.

Regarding cigarette smoking, questions were asked about the following: whether the subject was a smoker at baseline, age at which they started smoking, age at smoking cessation, the number of cigarettes smoked daily and the number of smoking years (excluding stopping periods). Based on these questions, the following variables were constructed: smoking status (never, former, current), current smoking (yes/no), frequency (number of cigarettes/day), duration (number of years), pack-years of cigarette smoking (number) and time since cessation (years).

The FFQ has been validated against a 9-day diet record, and the Spearman correlation coefficient between the alcohol intake assessed by the questionnaire and that estimated by the diet record was 0.89 for all subjects and 0.85 for users of alcoholic beverages.17 The reproducibility of the FFQ was established and the test–retest correlation was 0.90 for alcohol intake, and this correlation declined only 0.01–0.02 per year.19 This indicates that the single FFQ measurement was able to rank subjects according to alcohol intake and this ability dropped only slightly over time. The single FFQ measurement that is used in our cohort study can characterise dietary habits for a period of at least 5 years.19

Questionnaire data were key-entered and processed in a standardised manner, blinded with respect to case/subcohort status in order to minimise observer bias in coding and data interpretation.

Data analysis

The confounders considered were20 21 22: age (years), sex, level of education (primary, lower vocational, secondary and medium vocational, and university and higher vocational), BMI (kg/m²), non-occupational physical activity (<30, 30 to <60, 60 to <90, ⩾90 min/day), energy intake (kJ), tea consumption (cups/day), intake of fruit, vegetables, legumes, fish, red meat and meat products (all g/day), family history of oesophageal or gastric cancer (yes/no in first-degree relatives), reported long-term (>0.5 years) use of non-steroidal anti-inflammatory drugs (NSAIDs) or aspirin (ATC codes M01A, N02B) (yes/no) and reported long-term use (yes/no) of lower oesophageal sphincter (LOS)-relaxing medication.23 24 25 Alcohol consumption and cigarette smoking were mutually adjusted in the statistical models. A variable was considered a confounder if including it in the model changed the rate ratio (RR) for any of the cancer types by >5%. For all four cancers, the same confounders were used in analyses.

Incidence RRs and corresponding 95% CIs for alcohol consumption and cigarette smoking were estimated in age- and sex-adjusted and multivariable adjusted case–cohort analyses using Cox proportional hazards models.26 Analyses were done using the Stata 9.2 statistical software package (StataCorp, College Station, Texas, USA). Standard errors were estimated using the robust Huber–White sandwich estimator to account for additional variance introduced by sampling from the cohort. This method is equivalent to the variance–covariance estimator by Barlow.27 The proportional hazards assumption was tested using the scaled Schoenfeld residuals.28 If the assumption was violated for a confounder, a time-varying covariate was added to the model. In case the assumption was violated for the exposure variable, it was checked whether the time-varying covariate for this variable was statistically significant in the model. Tests for dose–response trends were assessed by fitting ordinal exposure variables as continuous terms. Two-sided p values are reported throughout the article.

To evaluate whether substances in alcoholic beverages, other than ethanol, have an effect on the risk of OSCC, OAC, GCA or GNCA, we also analysed beer, wine and liquor consumption, adjusted for ethanol intake.

In smoking analyses, the different aspects of smoking were mutually adjusted for, to investigate which aspect is most important in terms of oesophageal or gastric cancer risk.

Interaction

To evaluate possible interaction between alcohol and cigarette smoking status, we estimated RRs of all four cancers for combinations of these exposures. The p value for interaction was assessed by including a cross-product term in the model. In the interaction analyses it was necessary to change the alcohol categories to include subjects with low alcohol consumption in the reference category, besides abstainers, because of too low numbers of cases per stratum.

Results

Descriptives

Table 1 presents the characteristics of the subcohort members and cases, including all variables that were considered potential confounders. Oesophageal and gastric cancer cases were more often men, especially OAC and GCA cases. Alcohol consumption was higher in cases, especially OSCC cases, than in subcohort members, while the proportion of alcohol abstainers was higher in the subcohort. About 60% of subcohort members and cases were stable alcohol consumers. Cases were exposed more to cigarette smoke in terms of smoking status, frequency, duration and pack-years compared with the subcohort.

Cases also differed from subcohort members with respect to some other characteristics: OAC and GCA cases had a higher BMI than subcohort members. All cases consumed less fruit and slightly less vegetables than the subcohort, but cases more often reported a family history of oesophageal or gastric cancer and use of LOS-relaxing medication (table 1).

The proportional hazards assumption was violated only in a few analyses (see footnotes in table 2 and appendix 1 available online).

Alcohol consumption

A daily alcohol consumption of ⩾30 g, when compared with abstaining, was associated with a significantly increased risk of OSCC (multivariable adjusted RR = 4.61, 95% CI 2.24 to 9.50, p trend <0.001) (table 2). Women were at somewhat higher risk than men, and the interaction with sex was statistically significant in continuous analyses (p = 0.04), but not in categorical analyses (p = 0.68). No association was observed between alcohol consumption and OAC (RR for ⩾30 g/day = 1.04, 95% CI 0.54 to 2.02), GCA (RR = 0.90, 95% CI 0.50 to 1.64) or GNCA (RR = 1.00, 95% CI 0.68 to 1.47).

Multivariable RRs for alcohol consumption and OSCC were slightly attenuated when compared with age- and sex-adjusted RRs (see appendix 1 online). For OAC, GCA and GNCA, age- and sex-adjusted RRs were very similar to multivariable adjusted RRs.

When the analyses were restricted to stable alcohol consumers, the association with OSCC became somewhat stronger. The results for OAC, GCA and GNCA barely changed (table 2).

Results regarding alcoholic beverages are shown in table 2. After adjustment for total alcohol intake, beer consumption was associated with an increased risk of OSCC (p trend = 0.23) and GNCA (p trend = 0.06). Wine consumption was inversely associated with risk of OSCC (RR per 1 glass/day increment = 0.67, 95% CI 0.50 to 0.90), but was not associated with OAC, GCA or GNCA. Consumption of liquor was not significantly associated with risks of OSCC, OAC and GCA, and with a reduced risk of GNCA.

Cigarette smoking

Current and former cigarette smoking were associated with an increased risk of all four cancer types, when compared with never smokers (table 3). The strongest associations were found for OSCC and GNCA.

Frequency of cigarette smoking was found to be associated with increased risks of all four cancers, independent of other smoking aspects. The association was statistically significant for OSCC and OAC. The RRs for the gastric cancers were lower than for the oesophageal cancers (table 3). Duration of cigarette smoking was independently associated with increased risks of OSCC, GCA and GNCA, but the RR was only statistically significant for GNCA. Pack-years of smoking were associated with increased risks (mostly statistically significant) of all four cancers. The highest RRs were found in those who smoked ⩾40 pack-years.

We also looked into the effect of cigarette smoking cessation on the risk of oesophageal and gastric cancers. The risks of all four cancer types declined for smokers who had stopped since <10, 10 to <20 or ⩾20 years (p trend <0.05 for all four cancers), when compared with current smokers. However, when compared with never smokers, the risks were still elevated for OSCC, OAC and GNCA. For OSCC, OAC and GCA, the RRs of time since smoking cessation were attenuated when we adjusted for smoking duration (data not shown).

Multivariable RRs for cigarette smoking were attenuated when compared with age- and sex-adjusted RRs (appendix 2 online). This difference could largely be explained by the mutual adjustment of cigarette smoking aspects. For OSCC, confounding by alcohol consumption also caused attenuation of the associations.

Interaction

Table 4 shows RRs for several exposure combinations of alcohol consumption and cigarette smoking, using never smokers and low alcohol consumers as a reference. A graphical presentation of these results can be found in fig 2A–D.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Combined exposure to alcohol and cigarette smoking and risk of oesophageal and gastric cancer subtypes; Netherlands Cohort Study on diet and cancer, 1986–2002. (A) Oesophageal squamous cell carcinoma, (B) oesophageal adenocarcinoma, (C) gastric cardia adenocarcinoma, (D) gastric non-cardia adenocarcinoma. Multivariable RRs: adjusted for age, sex, body mass index, level of education, energy intake, consumption of fruits, vegetables, fish.

The RR of OSCC was 8.05 (95% CI 3.89 to 16.60) in the category of current smokers who consumed >15 g/day of ethanol. This RR is compatible with multiplicative interaction, but was not statistically significant (p interaction = 0.65). We found no evidence for interaction between alcohol consumption and cigarette smoking on risk of OAC, GCA or GNCA.

Discussion

Main findings

In this study, which is one of the first prospective cohort studies on the topic, we found the following results. Alcohol consumption was related to a strongly increased risk of OSCC, whereas it was unrelated to OAC, GCA or GNCA. Cigarette smoking status, frequency, duration and pack-years were associated with increased risks of all four cancer types. A multiplicative interaction was found between alcohol consumption and cigarette smoking on the risk of OSCC, but no such interaction existed for OAC, GCA or GNCA.

Strengths and limitations

A strength of our study includes the prospective character, which makes selection and information bias unlikely. Moreover, the division we made into histological subtypes of oesophageal cancer and localisations of gastric cancer is important, as this allows the evaluation of different risk factors for the cancer subtypes and more precise estimates of the strength of the associations. Our study is also one of the largest prospective cohort studies investigating this topic.

The ranges of the participants’ alcohol consumption and cigarette smoking habits were wide, which gave good contrasts between high and low exposures. Also, alcohol consumption and cigarette smoking habits were addressed extensively in the questionnaire. We lack data on smoking and alcohol consumption after baseline, but we did ask participants about their habits during a long period before baseline. As the development of a tumour probably takes several decades, we believe we asked about the exposure in a relevant time window.

Participants could have changed their alcohol consumption and smoking habits due to preclinical cancer, which may influence results. This is unlikely though, because the smoking data reflect the lifetime exposure of the participants. As for alcohol, results were robust for restriction of analyses to stable users.

In this epidemiological study, we analyse the possibility of residual confounding, by, for example, Helicobacter pylori infection. This infection is associated with increased risk of GNCA,29 and possibly with decreased risk of OAC,30 31 while it is not associated with GCA29 or OSCC risk.31 Unfortunately, we lack data on H pylori infection, but we estimate the prevalence to have been ∼50%.32 The infection might have confounded the associations if it is correlated with alcohol consumption or smoking. However, studies investigating this correlation have been inconsistent.33 34 35 Residual confounding by other variables cannot be completely ruled out, but the associations found were strong and residual confounding probably cannot explain these associations entirely. Moreover, the true associations may be stronger than observed, because we probably underestimated their strength due to random measurement error.

Comparison with previous research: alcohol

The strong positive association between alcohol consumption and OSCC risk we found is in agreement with associations found in the other prospective cohorts.9 10 36 The association we found showed a statistically significant trend (p<0.001), although low consumers of alcohol were found to be at slightly decreased OSCC risk, and this association remained when we restricted our analyses to the group of stable alcohol consumers. These findings are consistent with findings from two other cohorts,10 36 and an explanation was suggested by Freedman et al.10 Possibly, some of the subjects who were abstainers at baseline used to be heavy drinkers long (>5 years) before baseline. These drinkers remain at increased risk for a decade after cessation of alcohol consumption.37

Our findings that alcohol was not associated with OAC, GCA and GNCA confirm the null results found in two other Western cohorts.9 10

To investigate whether substances in alcoholic beverages other than ethanol are relevant for cancer risk, we adjusted analyses of these beverages for total ethanol intake. This approach has not been used for oesophageal or gastric cancers before and gave some new insights. For the beverages, OSCC risk was lower compared with the association with ethanol. This suggests that ethanol is probably the key substance in these beverages, and this suggestion is strengthened by the fact that associations have been found across the world, for many different alcoholic beverages.10 38 39 Still, other substances in alcoholic beverages might be relevant. For instance, flavonoids40 in wine might explain the inverse association we found with OSCC. This inverse association has also been found by a case–control study,41 but this association was unadjusted for ethanol intake. Regarding beer, N-nitroso compounds present in this beverage may be partly responsible for the positive associations with OSCC and GNCA.42

Comparison with previous research: cigarette smoking

We found that cigarette smoking was associated with an increased risk of all oesophageal and gastric cancer subtypes. For these cancer subtypes, we are the first cohort study to analyse the association with all kinds of smoking aspects. Our findings are consistent with previous reviews.43 44 45 It appeared that smoking frequency as well as duration were important in terms of increasing a person’s risk. Therefore, the number of pack-years smoked may be a good indicator of a person’s total exposure.

The relationships we found between smoking and oesophageal cancer types were weaker compared with a previous case–control45 and two cohort10 11 studies, but similar to the results of a third cohort study.9 There may be several reasons why other studies found stronger associations: under-reporting of smoking habits in case–control studies, no mutual adjustment for smoking aspects, or no or insufficient adjustment for confounding (eg, by alcohol consumption). Our RRs might be attenuated because some smokers may have stopped smoking after baseline, but this is unlikely because after smoking cessation people stay at risk for a long period.

Our results showed that smoking cessation decreased risks of all four cancers, but only after one to two decades, and these results are in agreement with two reviews.43 46

Interaction

This is the first prospective study to investigate the interaction between alcohol consumption and cigarette smoking and the risk of OSCC. An interaction for OSCC has been established by previous case–control studies (see, for example,4 41 47 48). We confirm this interaction, although this was not statistically significant, probably due to low power. Yet, for OAC, GCA and GNCA it was uncertain whether a similar interaction is present, as only one study examined this.10 According to our findings, no interaction was present.

The interaction between alcohol consumption and cigarette smoking can be explained biologically. Cigarette smoke contains many carcinogens, and the ethanol from alcoholic beverages is metabolised to acetaldehyde, which was classified as a human carcinogen by the International Agency for Research on Cancer (IARC).49 Acetaldehyde can circulate in the blood after formation in the liver, but can also be formed locally by oral bacteria.50 Ethanol itself can cause local irritation of the upper gastrointestinal tract51 and may facilitate the uptake of the carcinogens present in cigarette smoke.50

In conclusion, alcohol and cigarette smoke each have an individual effect on OSCC risk, but, when combined, they act synergistically. We found no interaction for OAC, GCA or GNCA.

Conclusions and recommendations

In summary, we found alcohol to be positively associated with OSCC risk, but not with OAC, GCA or GNCA. Cigarette smoking was positively associated with risk of all four cancer types. Alcohol consumption and smoking interacted in a multiplicative way on the risk of OSCC.

Alcohol consumption cannot explain the previously mentioned rising incidences of OAC and GCA, because it was not associated with risk of either. Cigarette smoking habits cannot explain the increasing trends either, because these habits have not increased over time in Western countries. Therefore, we suggest that further research should focus on other risk factors for oesophageal and gastric cancer subtypes, to search for explanations for these increases.