Estimates of alcohol-related oesophageal cancer burden in Japan: systematic review and meta-analyses

Abstract Objective To refine estimates of the burden of alcohol-related oesophageal cancer in Japan. Methods We searched PubMed for published reviews and original studies on alcohol intake, aldehyde dehydrogenase polymorphisms, and risk for oesophageal cancer in Japan, published before 2014. We conducted random-effects meta-analyses, including subgroup analyses by aldehyde dehydrogenase variants. We estimated deaths and loss of disability-adjusted life years (DALYs) from oesophageal cancer using exposure distributions for alcohol based on age, sex and relative risks per unit of exposure. Findings We identified 14 relevant studies. Three cohort studies and four case-control studies had dose–response data. Evidence from cohort studies showed that people who consumed the equivalent of 100 g/day of pure alcohol had an 11.71 fold, (95% confidence interval, CI: 2.67–51.32) risk of oesophageal cancer compared to those who never consumed alcohol. Evidence from case-control studies showed that the increase in risk was 33.11 fold (95% CI: 8.15–134.43) in the population at large. The difference by study design is explained by the 159 fold (95% CI: 27.2–938.2) risk among those with an inactive aldehyde dehydrogenase enzyme variant. Applying these dose–response estimates to the national profile of alcohol intake yielded 5279 oesophageal cancer deaths and 102 988 DALYs lost – almost double the estimates produced by the most recent global burden of disease exercise. Conclusion Use of global dose–response data results in an underestimate of the burden of disease from oesophageal cancer in Japan. Where possible, national burden of disease studies should use results from the population concerned.


Introduction
Alcohol consumption is a major contributor to the global burden of disease 1,2 and is a major risk factor for cancer. [3][4][5][6] Of all alcohol-related cancers, oesophageal has the highest alcohol-attributable fraction 6 -i.e. the highest proportion of these cancers would be prevented if no alcohol were consumed. [6][7][8] The global burden of disease (GBD) study estimates that in 2010 alcohol-attributable oesophageal cancer resulted in 76 700 deaths and 1 825 000 disability adjusted life years (DALYs) lost, globally. 9 A large portion of oesophageal cancers attributable to alcohol consumption occur in Asian countries -52.2% (40 000) of all alcohol-attributable oesophageal cancer deaths and 51.8% (945 000) of all alcohol-attributable oesophageal cancer DALYs. The alcohol-attributable portions for countries in this region have been calculated based on global meta-analyses. 10,11 However, this assumes that the alcohol-attributable risk for oesophageal cancer is the same in all regions. Preliminary evidence, on the other hand, shows that the risk for this cancer is different for people of Asian origin, because of genetic polymorphisms -most importantly the aldehyde dehydrogenase 2 (ALDH2) and alcohol dehydrogenase 1B (ADH1B) polymorphisms. [12][13][14][15] Thus, the real risk and burden of alcohol-attributable oesophageal cancer in Asia may have been underestimated.
In Japan in 2010, oesophageal cancer was among the top 20 causes of years of life lost (11 deaths and 181 DALYs per 100 000 people per year). 9 We did a systematic review and meta-analyses of studies conducted in the Japanese population to estimate the alcohol-attributable burden of oesopha-geal cancer. We then compared these estimates to the GBD 2010 estimates. 1 We also estimated risk functions according to ALDH2 subsets and investigated potential interactions between ALDH2 and ADH1B polymorphisms.

Data search and selection
We followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses guidelines. 16 We used the latest editions of the International Agency for Research on Cancer (IARC) monographs on alcohol 3,4 to identify potentially eligible studies. Additionally, we searched PubMed for publications published before 2014. We did two searches using the following search terms; Search 1: "cancer or neoplasm or carcinom*" and "ALDH2 or ADH1B or ADH2 or ADH3 or ADH1C or dehydrogenase*" and "alcohol or ethanol"; Search 2: "alcohol or ethanol" and "cohort" and "cancer" and "japan" and "review" and "mortality". Inclusion criteria for analyses investigating the relationship between alcohol consumption, ALDH2, and oesophageal cancer were: (i) prospective or historical cohort or case-control study design; (ii) a measure of risk and its corresponding measure of variability was reported or there were sufficient data for us to calculate these; (iii) oesophageal cancer was reported as a separate outcome; (iv) data on total alcohol intake for at least two exposure categories among current drinkers, or estimates for ALDH2 variants by alcohol intake were reported; (v) risk estimates were at least age-adjusted; and (vi) the study was conducted in Japan after 1980. In addition, we searched reference lists of Objective To refine estimates of the burden of alcohol-related oesophageal cancer in Japan. Methods We searched PubMed for published reviews and original studies on alcohol intake, aldehyde dehydrogenase polymorphisms, and risk for oesophageal cancer in Japan, published before 2014. We conducted random-effects meta-analyses, including subgroup analyses by aldehyde dehydrogenase variants. We estimated deaths and loss of disability-adjusted life years (DALYs) from oesophageal cancer using exposure distributions for alcohol based on age, sex and relative risks per unit of exposure. Findings We identified 14 relevant studies. Three cohort studies and four case-control studies had dose-response data. Evidence from cohort studies showed that people who consumed the equivalent of 100 g/day of pure alcohol had an 11.71 fold, (95% confidence interval, CI: 2.67-51.32) risk of oesophageal cancer compared to those who never consumed alcohol. Evidence from case-control studies showed that the increase in risk was 33.11 fold (95% CI: 8. 15-134.43) in the population at large. The difference by study design is explained by the 159 fold (95% CI: 27.2-938.2) risk among those with an inactive aldehyde dehydrogenase enzyme variant. Applying these dose-response estimates to the national profile of alcohol intake yielded 5279 oesophageal cancer deaths and 102 988 DALYs lost -almost double the estimates produced by the most recent global burden of disease exercise. Conclusion Use of global dose-response data results in an underestimate of the burden of disease from oesophageal cancer in Japan. Where possible, national burden of disease studies should use results from the population concerned.
identified articles for additional articles. No active filters or language restrictions were applied. We excluded measures of pure drinking frequency and qualitative characteristics -such as social or problem drinker. Oesophageal cancer cases (International Classification of Diseases [ICD] version 9: 150, ICD-10: C15) were defined as newly diagnosed at the first visit to a specialized clinic, through cancer registries or cause of death on death certificates.
Most quality scores for primary studies are tailored for meta-analyses of randomized trials of interventions [17][18][19] and many criteria for such scores do not apply to epidemiological studies examined in this study. Additionally, quality score use in meta-analyses remains controversial. 19,20 As a result, we included quality components in the inclusion and exclusion criteria of the systematic search and separate metaanalyses -such as study design and alcohol measurement -and conducted subgroup analyses based on study design and genetic polymorphisms.

Data extraction
From all relevant articles we extracted: authors' names, year of publication, country, calendar year(s) of baseline examination, follow-up period, setting, assessment of oesophageal cancer diagnosis, range of age at baseline, sex, number of observed oesophageal cancer cases among participants by alcohol exposure category, number of total participants by alcohol exposure category, adjustment for potential confounders and effect size with its standard error. We used the most fully adjusted effect size reported and selected estimates where lifetime abstainers were used as the risk reference group when those were available. Assessment of full-text articles with uncertain eligibility and data abstraction were conducted independently by two authors who discussed differences until consensus was reached. When there was not enough information presented in the article, we contacted the corresponding author.
We converted alcohol intake into grams of pure alcohol per day (g/day) using the midpoints (mean) of reported categories in the studies. For openended categories of alcohol intake, we added three-fourths of the previous category's range to the lower bound of the open-ended categories. We used reported conversion factors in the studies when standard drinks were the unit of measurement. Hazard ratios and odds ratios were assumed to be equivalent to relative risks (RR). We used fractional polynomials 21 to derive the best fitting function for average alcohol consumption in g/day using the pool-first approach described by Orsini et al. 22 Linear and first-degree models were estimated using the following range of powers: −2, −1, 0, 1, 2, 3. 21 Significant increases in deviance were determined by likelihood ratio tests with one degree of freedom.

Data analyses
We conducted several meta-analyses and used the most comprehensive data available separately for each analysis when multiple reports from the same cohort were published. For studies providing data on two or more alcohol intake categories among current drinkers, we pooled data from (i) cohort studies; (ii) case-control studies; (iii) case-control studies that provided stratified data by ALDH2 variants. We conducted sensitivity analyses on the interaction between variants of ADH1B within the genetic variants of ALDH2. In analyses investigating ALDH2 variants, studies were pooled separately for the active variant (ALDH2*1/*1) and inactive variants (ALDH2*1/*2 and ALDH2*2/*2). No cohort studies provided ALDH2 genotype data. Where possible, we avoided ALDH2*2/*2 vari-ants because of the low number of cases. No systematic information on the distribution of ALDH2 variants by drinking level was available and we therefore used the distribution of drinking by ALDH2 variants among controls in case-control studies to estimate this distribution at the population level. Finally, studies were pooled using DerSimonian-Laird random-effect models to allow for between-study heterogeneity. 23 Variation in the effect size other than chance because of heterogeneity between studies was quantified using the I 2 statistic. 24 We conducted meta-regression analyses to identify study characteristics that influenced the association between alcohol consumption and oesophageal cancer risk. Because of few available studies, we were only able to investigate study design in such meta-regression analyses. Examination of potential publication bias using Egger's regression-based test 25 was planned, but was not done because of the few studies included. All metaanalyses were conducted on the natural log scale in Stata statistical software, version 12.1 (StataCorp. LP, College Station, United States of America) and P < 0.05 (two-sided) was considered statistically significant.
We estimated deaths and DALYs lost from oesophageal cancer attributable to alcohol consumption in Japan applying a standardized alcohol-attributable fraction method 26 using the sta-

Results
After removal of duplicates, we evaluated 1333 records for inclusion in our study. Based on titles and abstracts, we excluded 1174 articles and screened 159 in full-text articles ( Fig. 1). After excluding duplicate reports of the same cohorts, we analysed 11 case-control studies 28-38 and 3 cohort studies. [39][40][41] Eight case-control and cohort studies [32][33][34][35][36][39][40][41] reported estimates for at least two alcohol intake categories in comparison to non-drinkers. These studies were used for a nonlinear dose-response analysis of oesophageal cancer risk, including stratified analyses by ALDH2 variants. Four case-control studies [28][29][30][31] provided indirect evidence for only one alcohol intake category Table 1 (available from: http://www.who.int/bulletin/ volumes/93/5/14-142141). In total five studies had data on ALDH2 and ADH1B variants stratified or adjusted by level of alcohol consumption. 29,31,[36][37][38] As shown in Fig. 2, the risk for oesophageal cancer identified in cohort studies from Japan 39-41 was higher compared with the most recent GBD estimate (RR: 11.71; 95% confidence interval, CI: 2.67-51.32 and RR: 3.59; 95% CI: 3.34-3.87, respectively at 100 g/ day of pure alcohol intake). The risk  43 at 100 g/day of pure alcohol intake) was much higher than the Japanese cohort studies or GBD estimates. In a meta-regression, the difference between case-control studies and cohort studies was significant (P = 0.014). We observed moderate heterogeneity among cohort studies (I 2 = 60%, P = 0.082), and high heterogeneity among case-control studies (I 2 = 89%, P < 0.001).
The risk curves by ALDH2 variants in Japan are displayed in Fig. 3. Three case-control studies 33,34,36 provided dose-response data for an investigation of ALDH2 polymorphisms in reference to non-drinkers: ALDH2*1/*2 (372 cases) and ALDH2*1/*1 (151 cases). Inactive variants of ALDH2 enzyme showed markedly higher risks with increasing alcohol consumption. The RR compared to non-drinkers was 36.15 (95% CI: 10.34-126.40) at 50 g/ day of pure alcohol and 159 (95% CI: 27.2-938.2) at 100 g/day of pure alcohol intake among people carrying the ALDH2*1/*2 variant. In comparison, the RR among those carrying the ALDH2*1/*1 variant was 2.99 (95% CI: 1.75-5.12) at 50 g/day of pure alcohol intake and 8.94 (95% CI: 3.05-26.23) at 100 g/day of pure alcohol. Based on two studies that included people with alcohol dependence (median 120 g/day of pure alcohol intake), people with the inactive variant of ALDH2 had an RR of 13.00 (95% CI: 8.99-18.80) compared to those with the active variant. 28,29 We interpolated this difference in risk in the curve for ALDH2*1/*2 in Fig. 3, and held the risk increase among people with this ALDH2 variant constant beyond 100 g/day of pure alcohol intake because there were insufficient data to reliably estimate this risk function. Once casecontrol studies were stratified by ALDH2 variant, there was little or no heterogeneity (ALDH2*1/*1, I 2 = 0%, P = 0.78; ALDH2*1/*2, I 2 = 44%, P = 0.17). Another two studies, 30,31 although they did not provide data in reference to nondrinkers, were in close agreement with our reported risk functions.
With regard to differences in risk curves by study design, Table 2 shows that among case-control studies with multiple alcohol intake categories, 72% (350/483) of oesophageal cancer cases among drinkers occurred in 32% (313/980) of the drinking population, namely individuals with the genetic variant ALDH2*1/*2. When the risk curves from case-control studies (Fig. 3) were combined (weighted by their distribution of alcohol consumption by ALDH2 variants at the population level) the risk functions from case-control and cohort studies almost entirely overlapped (Fig. 4). Combining adjusted casecontrol and cohort studies, at 100 g/day pure alcohol intake, the risk in Japan was markedly elevated (RR: 11.65, 95% CI: 4.16-32.62) compared to GBD estimates (RR: 3.55, 95% CI: 3.30-3.82) (Fig. 5).
Using our calculated risk relations for alcohol-attributable oesophageal cancer results in almost twofold higher estimates for deaths (5279) and DALYs lost (102 988) compared with the current GBD estimates (2749 and 53 826, respectively; Table 3). These results are irrespective of whether the estimates were based on cohort studies or on case-control studies, in each case Alcohol-related oesophageal cancer in Japan Michael Roerecke et al.
adjusted for population prevalence of genotypes.

Discussion
For Japan, we estimated a twofold higher mortality and burden of disease using risk functions derived from Japanese populations compared with the 2010 GBD estimates, which are based on global risk functions. We obtained separate estimates based on independent methods -either Japanese cohort data or adjusted Japanese case-control studies -and these estimates were comparable. This strengthens our conclusion that the current GBD method underestimates Japanese oesophageal cancer outcomes. Since we took into consideration genetic polymorphisms commonly observed in the Japanese population, we would predict a similar degree of underestimation for alcoholattributable oral, pharynx and larynx cancers in Japan. 34,42,43 Furthermore, the burden of other alcohol-attributable cancers where acetaldehyde plays an important role might be underestimated. 3,4 We found that the slow-acting AD-H1B variant also increased the risk for oesophageal cancer, regardless of ALDH2 variant. However, the slow-acting variant is only present in 6% of the Japanese population, 44 whereas 90% of Caucasians carry the variant. 45 As we had restricted all our analyses to Japanese individuals, the potential protective effect of the fastacting ADH1B*1/*2 or *2/*2 variants has been already included. Similarly, risk estimates from cohort studies should not be affected by the differential risk curves for ALDH2 and ADH1B variants and their combinations if the prevalence of each combination of polymorphisms is reflected in the sample.
The study has some limitations. First, any systematic review or metaanalysis is only as good as the literature it is based on. Although case-control studies initially showed high heterogeneity as measured by I 2 values indicating potential bias, there was little heterogeneity after these studies were stratified by ALDH2 variants. Second, while the procedures to estimate alcohol-attributable fractions are standard, 26,27,46 subsequent adjustments to survey results may bias consumption in either direction. [47][48][49] However, as the same method of triangulation of surveys and per capita estimation was applied to GBD and to national estimates, 1 the comparison between these estimates should be valid. Finally, the estimates of attributable risk and burden of disease for heavy alcohol intake are based on few studies and thus may be biased. 50 By including risk estimates of people with alcohol dependence, we attempted to minimize this bias.
While there may be some biases in our quantitative estimates of alcoholattributable burden for oesophageal cancer, they still show that global estimates underestimate the burden in Japan. This will likely be true for GBD estimates for China and the Republic of Korea as Fig. 6. Relationship between alcohol dehydrogenase 1B polymorphisms and oesophageal cancer risk in Japanese with inactive aldehyde dehydrogenase 2

Systematic reviews
Alcohol-related oesophageal cancer in Japan Michael Roerecke et al.
well, where a considerable proportion of the population also carry the inactive ALDH2 allele, (34% and 29%, respectively). 44 In populations with a high proportion of these polymorphisms, studies based on global dose response data are likely to underestimate many alcohol-attributable cancers.
Efforts should be made to estimate country-specific risks for diseases affected by genetic polymorphisms, especially in countries with higher proportions of such polymorphisms. The current stan-dard of applying global risk functions to local exposure data should be replaced by country-specific risk functions whenever possible. Country-specific risk functions should also be applied for other risk factors than alcohol. 1
Resultados Se identificaron 14 estudios pertinentes. Tres estudios de cohorte y cuatro estudios de casos y controles contenían datos sobre la respuesta en relación con la dosis. Las pruebas de los estudios de cohorte demostraron que el riesgo de cáncer esofágico de que quienes consumen el equivalente a 100 g/día de alcohol puro era 11 Alcohol-related oesophageal cancer in Japan Michael Roerecke et al.
Conclusión El uso de datos mundiales sobre la respuesta en relación con la dosis da lugar a una subestimación de la carga de enfermedad de cáncer esofágico en Japón. Siempre que sea posible, los estudios sobre la carga nacional de la enfermedad deben utilizar los resultados de la población afectada.