Vitamin B6 and Cancer Risk: A Field Synopsis and Meta-analysis

Simone Mocellin; Marta Briarava; Pierluigi Pilati


J Natl Cancer Inst. 2016;109(3) 

In This Article


Search Strategy, Eligibility Criteria, and Data Extraction

Following the Meta-analysis Of Observational Studies in Epidemiology (MOOSE)[8] and Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA)[9] guidelines, we searched for studies dealing with the association between vitamin B6 intake or PLP peripheral blood (plasma or serum) levels and cancer risk in humans; to this aim, both prospective (cohort and nested case-control studies) and retrospective designs (case-control studies) were considered eligible. We also searched the international literature to find randomized controlled trials (RCTs) testing the hypothesis that pharmacological doses of vitamin B6 might reduce the incidence of cancer. No language restriction was applied. Exclusion criteria were data published in abstract form only and lack of risk estimates (or data necessary to calculate them).

A two-step search strategy was adopted. First, a systematic review was performed by querying Medline (via the PubMed gateway) and Web of Science databases until January 2016; the following search terms were used: ("vitamin B6" OR "pyridoxine" OR "PLP" OR "pyridoxal-5'-phosphate") AND ("cancer" OR "tumor" OR "carcinoma" OR "melanoma" OR "sarcoma" OR "lymphoma" OR "leukemia"). In the second phase, cited references from eligible articles were searched. In case of overlapping series, only the most updated version was included. Authors were contacted whenever unreported data were potentially useful to enable the inclusion of the study into the systematic review or to rule out data published in different articles but regarding overlapping series. When the same article reported findings separately from more than one data set (eg, men and women; premenopausal and postmenopausal women; colon and rectal cancer), we maintained and analyzed them separately.

The following data from eligible studies were independently extracted by two authors (SM and MB), disagreements being resolved by discussion and consensus: authors' names; country where the study was conducted; year of publication; numbers of cases (defined as patients diagnosed with cancer based on histological evaluation, cancer site being recorded whenever available) and participants; prevalent ethnicity (>80%); study design, tumor site, type of vitamin intake (dietary [food only] or total intake [dietary plus supplements]), adjustment for confounding factors, levels of exposure (quantiles of vitamin B6 intake and plasma levels), and measures of association (choosing those with the highest degree of control for potential confounding factors), along with their 95% confidence intervals (CIs) for the risk comparison between high and low categories as well as across more than two exposure levels (for both vitamin B6 intake and PLP blood concentration).

Statistical Analysis

Different measures of association quantifying the strength of association between exposure and outcome were expected: odds ratio for retrospective studies (such as case-control and nested case-control studies); rate ratio (incidence rate data) or risk ratio (cumulative incidence data) or hazard ratio (time to event data) for prospective studies (including RCTs). For simplicity, we used relative risk (RR) as a generic term to refer to all the above.

Summary relative risks (along with their corresponding 95% CIs) were calculated by performing random effects meta-analysis (using the DerSimonian and Laird inverse variance method).[10] This analysis was utilized for pooling data comparing high vs low categories of exposure, such as highest vs lowest quantiles of vitamin B6 intake or PLP blood concentrations (observational studies), or intervention vs placebo/observational arms (RCTs).

Moreover, a dose-response meta-analysis was carried out to assess a linear trend between different exposure levels and cancer risk using a random effects meta-regression.[11,12] The median level for each exposure category (ie, vitamin B6 daily intake or PLP blood level quantiles) was assigned to the corresponding RR for each study; then, we first pooled the study data and afterward estimated the dose-response model (taking into account the correlation between RRs within each study). A potential nonlinear relationship was investigated by means of a restricted cubic spline model with three knots at fixed percentiles (25th, 50th, and 75th); a P value for nonlinearity was calculated by testing the null hypothesis that the coefficient of the second spline is equal to 0.[12]

To maximize clarity of data interpretation and reporting, findings on dietary (food only) and total (food and supplements) vitamin B6 intake were analyzed and described separately. We performed evaluation of heterogeneity, subgroup analyses, and examination for bias. Between-study heterogeneity (true variance of effect size across studies) was quantified using the I-square statistic (which indicates the percentage of the variability in effect estimates due to true heterogeneity rather than within-study sampling error; low: <25%, intermediate: 25%–50%; high: >50%) and formally tested by means of the Cochran Q-test.[13]

Subgroup analysis by tumor site, study design (retrospective vs prospective), and ethnicity (Asian vs Caucasian/other) was performed, if data permitted; effect differences were formally tested by means of random effects meta-regression. In addition, the impact of number of cases examined, range of exposure (highest minus lowest category values), publication year, and study quality were verified by means of random effects meta-regression.

Publication and selection biases in meta-analysis are more likely to affect small studies, which also tend to be of lower methodological quality: This may lead to the so-called "small study effect," where the smaller studies show larger effects. A funnel plot was used to detect this effect, and asymmetry was formally investigated with the Egger linear regression approach (if at least 10 studies were available).[14]

The alpha level of statistical significance was set at less than .05, except for the Cochran Q-test and the Egger test, for which a P value of less than .10 was considered statistically significant. All statistical analyses were performed with STATA 11.2/SE (StataCorp LP, College Station, TX). All statistical tests were two-sided.

Quality Assessment and Evidence Grading

Quality of observational studies was assessed by means of the Newcastle-Ottawa Scale (NOS), which assigns up to nine stars to each study based on eight items grouped into three categories (selection, comparability, and exposure or outcome for case control or cohort studies, respectively).[15] For intervention randomized trials, the Cochrane Collaboration's quality assessment tool was adopted to assign low, unclear, or high risk of bias to each study based on issues addressed by six domains (sequence generation; allocation concealment; blinding of participants, personnel, and outcome assessors; incomplete data outcome; selective outcome reporting; and other sources of bias).[16]

The Grades of Recommendation, Assessment, Development and Evaluation (GRADE) system was employed to grade the quality of evidence into four levels: high, moderate, low, and very low.[17] Briefly, evidence from RCTs and observational studies start with a high and low rating, respectively. Then, quality can be downgraded by one (serious concern) or two levels (very serious concern) for the following reasons: study limitations (risk of bias), evidence for publication bias, indirectness (indirect population, intervention, control, outcomes), inconsistency (between-study statistical heterogeneity), and imprecision (confidence intervals crossing the unit, single trial). On the other side, quality can be upgraded for two main reasons: large beneficial effect (we considered risk reduction >30%) and dose-response relationship between treatment/exposure and risk.