Association between alcohol intake and the risk of pancreatic cancer: a dose–response meta-analysis of cohort studies

This review was conducted and reported according to the criteria for conducting and reporting meta-analysis of observational studies in epidemiology (Additional file 1) [17]. Any prospective study that examined the association between alcohol intake and subsequent pancreatic cancer risk was eligible for inclusion in this study, with no restrictions placed on language or publication status.

Relevant studies were identified using the following procedures. We searched electronic databases including PubMed, Embase, Ovid, and the Cochrane Library for articles published up to Aug 2015. Search terms examining both medical subject headings and free-language searches for “ethanol” OR “alcohol” OR “alcoholic beverages” OR “drinking behavior” OR “alcohol drinking” OR “drink” OR “liquor” OR “ethanol intake” OR “alcohol drink” OR “ethanol drink” AND (”pancreas” OR “pancreatic”) AND (“cancer” OR “carcinoma” OR “neoplasm”) AND (“cohort” OR “cohort studies”) were used. Other sources included meeting abstracts, meta-analyses, or reviews already published on related topics. Authors were contacted for essential information from publications that were not available in full. The medical subject heading, methods, population, study design, exposure, and outcome variables of these articles were used to identify the relevant studies.

The literature search was independently undertaken by two investigators using a standardized approach. Any inconsistencies between these investigators were identified by the principal investigator and resolved by consensus. We restricted our meta-analysis to prospective cohort studies that were less likely to be subject to confounding variables and bias than traditional case control studies. A study was eligible for inclusion if the study had a prospective cohort design, the study investigated the association between alcohol intake and the risk of pancreatic cancer, and the authors reported effect estimates (risk ratio [RR] or hazard ratio [HR]) and 95 % confidence intervals (CIs) comparing different alcohol intake categories with the lowest alcohol intake category.

The information collected included the study group’s name, country, study design, sample size, age at baseline, follow-up duration, effect estimate, and covariates, all of which were included in the fully adjusted model. We also extracted the number of cases, persons, person-years, the effect of different exposure categories, and their 95 % CIs. For studies that reported several multivariable adjusted RRs, we selected the effect estimate that was maximally adjusted for potential confounders. The Newcastle-Ottawa Scale (NOS), which is comprehensive and has been partially validated for evaluating the quality of observational studies in meta-analyses, was used to evaluate methodological quality [18, 19]. The NOS is based on three subscales, selection consisting of four items, comparability consisting of one item, and outcome consisting of three items. A “star system” (range, 0–9) has been developed for assessment [18]. Data extraction and quality assessment were independently conducted by two authors. The data was then independently examined and adjudicated by an additional author, while referring to the original studies.

We examined the relationship between alcohol intake and risk of pancreatic cancer based on the effect estimate (RR or HR) and its 95 % CI as published in each study. We used a fixed-effect model to calculate summary RRs and 95 % CIs for different alcohol intake levels compared with the lowest alcohol intake level or no alcohol intake [20, 21]. We then used a random-effects model to calculate summary RRs and 95 % CIs for different alcohol intake levels compared with the lowest alcohol intake level or no alcohol intake [22, 23]. We converted all measurements into grams per day and defined one drink as 12 g of alcohol intake. Using a semi-parametric method, we evaluated the association between light (0–12 g per day), moderate (≥12-24 g per day), or heavy alcohol (≥24 g per day) intake and the risk of pancreatic cancer. The value assigned to each alcohol intake category was the mid-point for closed categories and the median for open categories. Furthermore, we constructed a dose response curve based on the correlated natural log of RRs or HRs across alcohol intake categories, and modeled alcohol intake by using restricted cubic splines with three knots at fixed percentiles of 10 %, 50 %, and 90 % of the distribution [24, 25]. Heterogeneity between studies was investigated using the I² statistic as a measure of the proportion of total variation between studies that is attributable to heterogeneity, where I² values of 25 %, 50 %, and 75 % were assigned as cut-off points for low, moderate, and high degrees of heterogeneity [26‐28]. Subgroup analyses were conducted based on country, duration of follow-up, adjustment of covariates (including smoking status, body mass index [BMI], diabetes mellitus, and energy intake [EI]), and study quality. We also performed a sensitivity analysis by eliminating individual studies from the meta-analysis [29]. Several methods were used to check for potential publication bias, including visually inspecting the Funnel plots for pancreatic cancer, and using the Egger [30] and Begg [31] tests for a statistical bias assessment. All reported P values are 2-sided, and P values <0.05 were considered statistically significant for all included studies. Statistical analyses were performed using STATA software (version 12.0; Stata Corporation, College Station, TX, USA).

The study-selection process is illustrated in Fig. 1. We identified 469 articles during our initial electronic search, of which, 425 were excluded as duplicates or irrelevant, leaving 44 potentially eligible studies to be selected. After detailed evaluations, 19 prospective studies consisting of 21 cohorts were selected for the final meta-analysis [9‐11, 13‐16, 32‐43]. A manual search of the reference lists from these studies did not yield any additional eligible studies. The general characteristics of the included studies are presented in Table 1.

In the included studies, follow-up periods for participants ranged from six to 30 years, and had from 7132 to 1,290,000 individuals included. Nine studies (ten cohorts) were conducted in the United States [11, 16, 32, 35, 36, 38‐40, 42], six (seven cohorts) in Europe [9, 13, 33, 34, 37, 43], and four in other countries [10, 14, 15, 41]. In total, the meta-analysis included 11,846 incident cases and more than 4,211,129 individuals. Study quality was assessed using the NOS, with studies receiving a score ≥8 considered to be high quality (Table 1). Overall, four cohorts had a score of 9 [14, 16, 33, 34], eight cohorts (six studies) had a score of 8 [9, 11, 13, 38, 39, 43], five cohorts had a score of 7 [10, 15, 35, 37, 41], and the remaining four cohorts had a score of 6 [32, 36, 40, 42].

In the pooled analysis (Fig. 2), low (RR, 0.97; 95 % CI, 0.89–1.05; P = 0.389; Additional file 2: Figure S1), moderate (RR, 0.98; 95 % CI: 0.93–1.03; P = 0.513; Additional file 3: Figure S2), and total alcohol intake (RR, 1.02; 95 % CI: 0.95–1.08; P = 0.634; Additional file 4: Figure S3) were not associated with pancreatic cancer risk, compared with the lowest alcohol intake level. However, high alcohol intake was associated with an increased risk of pancreatic cancer (RR, 1.15; 95 % CI: 1.06–1.25; P = 0.001; Additional file 5: Figure S4). Between-study heterogeneity was moderate for total alcohol intake (I² = 39.4 %) and low for low (I² = 0.0 %), moderate (I² = 0.0 %), and high alcohol intake (I² = 14.5 %). Analysis using the summary RR showed that low (RR, 0.98; 95 % CI, 0.84–1.15; P = 0.836), moderate (RR, 0.93; 95 % CI, 0.80–1.09; P = 0.372), and total alcohol intake (RR, 1.03; 95 % CI, 0.91–1.17; P = 0.664) were not associated with pancreatic cancer risk in men, compared with the lowest alcohol intake level. However, high alcohol intake was associated with an increased risk of pancreatic cancer in men (RR, 1.18; 95 % CI: 1.00–1.39; P = 0.045). Results from men exhibited substantial heterogeneity for total alcohol intake (I² = 48.7 %), moderate heterogeneity for low alcohol intake (I² = 21.2 %), and low heterogeneity for moderate (I² = 0.0 %) or high alcohol intake (I² = 12.9 %). No significant association was found between low, moderate, high, or total alcohol intake and pancreatic cancer risk in women, and there was no evidence of heterogeneity across studies in this population (low: I² = 0.0 %; moderate: I² = 0.0 %; high: I² = 0.0 %).

Analysis based on the type of alcohol showed that, high liquor intake was associated with an increased risk of pancreatic cancer in men (RR, 1.66; 95 % CI: 1.24–2.23; Fig. 3) and in the total cohort (RR, 1.43; 95 % CI: 1.17–1.74; Fig. 3). However, there was no significant association between any other types of alcohol intake and risk of pancreatic cancer.

A total of 13 cohorts (12 studies) were included in the restricted cubic splines analysis examining the association between alcohol intake and the incidence of pancreatic cancer. As shown in Fig. 4, we found no evidence for a potential nonlinear relationship between alcohol intake and the risk of pancreatic cancer (P = 0.0874), although alcohol intake greater than 15 g/day seemed to be associated with an increased risk of pancreatic cancer. A dose–response analysis examining the association between alcohol intake and pancreatic cancer risk in men was performed with seven cohorts, and found no significant relationship between alcohol intake and the risk of pancreatic cancer (P = 0.8450; Additional file 6: Figure S5A). Alcohol intake rates of 25.0–55.0 g/day seemed to be associated with an increased risk of pancreatic cancer, but alcohol intake rates greater than 55.0 g/day were not associated with the risk of pancreatic cancer. This analysis performed on data from women, as shown in Additional file 6: Figure S5B, found no evidence of a nonlinear relationship between alcohol intake and the risk of pancreatic cancer based on the P value for nonlinearity (P = 0.0524).

We conducted subgroup analyses to minimize heterogeneity among the included studies and evaluated the association between alcohol intake and risk of pancreatic cancer in specific subpopulations (Table 2). First, we noted that high alcohol intake was associated with an increased risk of pancreatic cancer in North America; when the duration of follow-up was greater than 10 years; in studies with adjustments for smoking status, BMI, diabetes mellitus, and EI; and in studies scored as high quality. Second, high alcohol intake was associated with an increased risk of pancreatic cancer in men if the duration of the follow-up was less than 10 years. Third, high alcohol intake was associated with an increased risk of pancreatic cancer in women if the follow-up duration was greater than 10 years and if the study adjusted for EI. Lastly, alcohol intake was associated with an increased risk of pancreatic cancer in men in studies scored as low quality.

After review of the funnel plots, we could not rule out the potential for publication bias (Fig. 5). However, the Egger [30] and Begg [31] tests showed no evidence of publication bias (Egger test, P = 0.199; Begg test, P = 0.928).