We followed the Meta-Analysis of Observational Studies in Epidemiology (MOOSE) guidelines in performing our systematic review. Relevant articles were identified by two reviewers (M.T. and J.A.) by systematically searching through MEDLINE (from 1950), PubMed (from 1946), EMBASE (from 1949), and Current Contents Connect (from 1998) through to July 26, 2013. The search was performed using the terms folate, folic acid, or vitamin B9, and esophageal, gastric, stomach, or pancreatic cancer, neoplasm, squamous cell carcinoma, or adenocarcinoma. The search terms used were searched as text word and as exploded medical subject headings where possible. The reference lists of relevant articles were also searched for appropriate studies. No language restrictions were used in either the search or study selection. A search for unpublished literature was not performed. Disagreement on article inclusion between the two reviewers was resolved via a third reviewer (G.E.).
In order to be included, eligible studies needed to: (i) have a study design of either a cohort or case control; (ii) report the risk of esophageal cancer, gastric cancer, or pancreatic cancer in association with total folate intake, dietary folate intake, or folate serum levels. Dietary folate was defined as all folate intake via foods, and total folate intake was defined as all folate intake via foods and supplements; (iii) report the risk point estimate as an odds ratio (OR), hazard ratio, or relative risk that compared a higher level of folate intake with a lower level of folate intake. When multiple levels of folate intake were presented, the ratio comparing the highest intake versus the lowest intake was chosen; (iv) report the 95% confidence interval (CI) for the point estimate; and (v) use an internal comparison when calculating the risk estimate. Studies were excluded if they were duplicates of other studies eligible for inclusion, or if they were meta-analyses of studies. In the case of duplicate studies, the most recently published study was chosen for inclusion.
One reviewer (M.T.) performed the data extraction via a standardized data extraction form. Information was extracted on the publication year, study design, number of cases, number of controls, total sample size, temporal direction, population type, country, continent, economic development, response rate, case–control matching, mean age, number of adjusted variables, years of follow up, lowest folate level, highest folate level, difference between highest and lowest folate levels, and the risk estimates on esophageal cancer, gastric cancer, and pancreatic cancer. Authors were not contacted if data were missing. Adjusted ratios were extracted in preference to non-adjusted ratios. When more than one adjusted ratio was reported, the ratio with the highest number of adjusted variables was chosen.
Pooled estimates of the OR and 95% CI for the risk of esophageal cancer, gastric cancer, and pancreatic cancer in association with folate were calculated using the random effects model of DerSimonian and Laird. Heterogeneity was assessed using the I2 statistic, which determines the proportion of variability across studies that is due to heterogeneity as opposed to sampling error. Sensitivity analyses were performed when statistically significantly heterogeneity was detected. Subgroup analyses stratified by either study design or histological subtype were also performed. We pre-specified analyses of meta-regression of log OR versus highest absolute folate intake level, lowest absolute folate intake level, and difference between highest and lowest absolute folate intake level.
Publication bias was assessed with Egger's regression model. If publication bias was detected, the additional publication bias methods consisting of the fail-safe number method and the trim-and-fill method were employed to quantify the effect of the bias. The fail-safe number method calculates the number of unpublished studies needed to convert the observed result to statistical non-significance at the alpha level of significance P < 0.05 level. Publication bias is considered to be an issue if the fail-safe number is less than 5n + 10, where n is the number of studies included in the meta-analysis. The trim-and-fill method simulates unpublished studies in the meta-analysis to calculate a new pooled OR, which is then compared with the original pooled OR. If the new pooled OR is similar to the original pooled OR, this indicates that publication bias has little effect on the meta-analysis results. Results were regarded as statistically significant if P < 0.05. All analyses were done with Comprehensive Meta-analysis (version 2.0; Biostat, Englwood, NJ, USA).
J Gastroenterol Hepatol. 2014;29(2):250-258. © 2014 Blackwell Publishing