To identify all studies on the association between the IL6 -174G/C polymorphism and cancer risk, we conducted a systematic search of the literature published before April 2011 using the MEDLINE database and the HuGE Published Literature database (HuGE Pub Lit) (19) with no restrictions. For MEDLINE, keywords ‘IL-6’ OR ‘IL 6’ OR ‘IL6’ OR ‘interleukin-6’ OR ‘interleukin 6’ AND ‘polymorphism’ AND ‘cancer’ were used; For HuGE Pub Lit, keywords ‘IL6’ AND ‘cancer’ were used for searching eligible studies. In addition, a manual review of references from primary or review articles was screened to trace additional relevant studies. Studies were included if they had a case-control design and the available frequency of three genotypes regarding the IL6 -174G/C polymorphism. Of the studies with overlapping data, we selected the ones with the largest number of subjects.

Three investigators independently extracted data and reached a consensus on all items. The following data were extracted from each study: the first author's last name, publication year, ethnicity of the subjects, cancer type, study design (retrospective case-control or prospective cohort study), and numbers of genotyped cases and controls with GG, GC or CC genotypes. Ethnic group was defined as African, Caucasian or ‘mixed’, including more than one ethnic category. Studies investigating more than one type of cancer with overlapping or same controls were regarded as individual data sets only in subgroup analyses by cancer type.

Hardy-Weinberg equilibrium (HWE) analysis for the frequencies of GG, GC and CC genotypes among controls in each study was assessed using Pearson's Chi-square test. The strength of the association between IL6 -174G/C polymorphism and cancer risk was measured by odds ratio (OR) with its 95% confidence interval (CI). The pooled ORs for IL6 -174G/C genotypes CC, GC and C allele carriers (CC or GC) against GG genotype were calculated, respectively. The significance of the pooled OR was determined by the Z-test and P<0.05 was considered statistically significant. Subgroup analysis was performed using stratification by study character, cancer type, ethnicity and study design, respectively. If a cancer type contained less than three independent individual studies, it was categorized into the ‘other types’ group.

Testing for heterogeneity among studies was performed by a Chi-square-based Q-test (20). Since Q-test is poor at detecting true heterogeneity, heterogeneity was considered significant for P<0.10 rather than P<0.05 (21). Additionally, the magnitude of the between-study heterogeneity was also assessed by I², which can be calculated from the basic results of a typical meta-analysis as I² = 100% x (Q-df)/Q, ranges form 0 to 100%, and is typically considered low for I²<25%, modest for 25–50% and large for >50% (22). Meta-regression was carried out to investigate whether statistical heterogeneity between the results of the multiple studies was related to one or more characteristics of the studies (23). To identify the studies that led to significant heterogeneity, sensitivity analysis for between-study heterogeneity was implemented by the sequential algorithm proposed by Patsopoulos et al (24). Whenever the P-value of the Q-test was >0.10, the summarized OR estimate of each study was calculated by the fixed-effects model (Mantel-Haenszel method) (25). Otherwise, the random-effects model (DerSimonian and Laird method) was used (26). Funnel plots were used to examine whether the results of a meta-analysis may have been affected by publication bias (27). A modified version of Egger's test proposed by Harbord, Egger and Sterne was implemented to test funnel plot asymmetry (28). All statistical analyses were performed using Stata statistical software (Stata/SE version 10.1 for Windows; Stata Corp, College Station, TX, USA).

The detailed steps of our literature search are described in Fig. 1. Eighty-three independent articles that met the inclusion criteria were included in the final analysis. Of these articles, one study provided data on breast and prostate cancer using independent controls (29), therefore each group in this article was treated as an independent study in our meta-analysis. The characteristics of the included studies are summarized in Table I. Overall, the present meta-analysis is based on a total of 105,482 participants, including 44,735 cancer patients and 60,747 controls. The studies were published between April 2000 and March 2011. Seventeen studies were conducted with a prospective cohort design, and 67 were conducted with a retrospective case-control design. Approximately two-thirds of cases and controls (29,019 cases and 42,120 controls) were from 73 studies involving Caucasian populations, a fraction of the data (1,138 cases and 1,299 controls) from seven studies involving African populations, and nearly one-third of the data (14,578 cases and 17,328 controls) from four studies involving ‘mixed’ populations. As shown in Table I, there were two studies from Dossus et al involving Caucasians, African-Americans, Asians, Latinos and native Hawaiians; one study from Ognjanovic et al involving Caucasians, Asians and Hawaiians; and one study from Bushley et al involving Caucasians, Asians and other populations. In the controls, the frequency of the rare C allele among controls varied considerably between Caucasians and Africans (0.417±0.052 and 0.207±0.097, respectively; P<0.001). A significant deviation from HWE was noted in two studies in Africans, nine in Caucasians and four in mixed populations (Table I).

There was a significant heterogeneity in overall comparison of the CC genotype vs. GG genotype (P<0.001 and I²=43.8%). The meta-regression showed that the strong heterogeneity could not be traditionally explained by cancer types, ethnicities or study designs (P=0.285, 0.129 and 0.306, respectively). Furthermore, the 15 studies that deviate from HWE showed similar heterogeneity with that of studies that were in HWE (Table II), suggesting that the remarkable heterogeneity among the overall analysis was not due to the variability of the control quality. Therefore, we carefully assessed the association of the IL6 -174G/C polymorphism with cancer risk in several subgroups, and carried out sensitivity analyses of between-study heterogeneity to detect studies that have remarkable influence on homogeneity.

For ethnic-specific subgroup analysis, no heterogeneity was detected within African population studies (P=0.294), but there were significant heterogeneity within Caucasian population studies (P<0.001). Sensitivity analysis of between-study heterogeneity revealed that five studies (30–34) mainly contributed to the heterogeneity within Caucasian population studies. After performing cancer type-specific analyses, we found no heterogeneity in studies of breast, prostate, lung, gastric cancer, lymphoma, multiple myeloma, melanoma and cervical cancer (Table II). However, there was strong heterogeneity for colorectal cancer, which was due to one Caucasian study (32).

Overall, the CC genotype was not significantly associated with cancer risk when compared to the GG genotype (OR=1.01, 95% CI 0.95–1.08, P=0.698). Ethnic-specific ORs showed that cancer risk was increased for individuals carrying the CC genotype compared to those with the GG genotype in African populations (OR=1.83, 95% CI 1.26–2.67, P=0.002; Fig. 2), but not in Caucasian populations (OR=1.00, 95% CI 0.92–1.08, P=0.938; Table II). After excluding the studies (30–34) responsible for heterogeneity, we found that Caucasian individuals carrying the CC genotype had no remarkable effect on risk of cancer compared to GG genotype individuals (OR=1.02, 95% CI 0.97–1.07, P=0.561) with no significant between-study heterogeneities (P=0.196, I²=12.6%). Although there were nine data sets in which the genotype distribution did not follow HWE, the corresponding meta-analysis was qualitatively similar with or without excluding them.

Subsequently, we stratified the association between the IL6 -174G/C polymorphism and cancer risk by cancer types. When compared to individuals with the GG genotype, those with the CC genotype were associated with increased risk of cervical cancer (OR=1.85, 95% CI 1.12–3.08, P=0.017), but not with that of other types of cancer (Table II). Furthermore, no significant association of the IL6 -174G/C polymorphism with risk of breast, colorectal and prostate cancer was observed in individuals with Caucasian ancestry (Fig. 3). A few studies involving African populations did not allow us to perform subgroup analysis in Africans (Table I).

Lastly, we also assessed the ORs of cancer for individuals with the GC genotype or C allele carriers (GC or CC genotype) compared to those with the GG genotype, and found no significant association in overall and subgroup analyses (Table II).

The shape of the funnel plot did not reveal any evidence of obvious asymmetry in comparison of the CC genotype vs. GG genotype (Fig. 4A). Then, we used Harbord's test to provide statistical evidence of funnel plot symmetry, and the result did not show evidence of publication bias (t=0.91, P=0.366; Fig. 4B). Subgroup analyses by ethnicity and cancer type did not provide any evidence of publication bias (t=−1.59, P=0.172 for African populations; t=0.27, P=0.786 for Caucasian populations; t=−0.58, P=0.575 for breast cancer and t=1.02, P=0.329 for colorectal cancer). Similarly, no publication bias was detected when comparing the GC genotype to the GG genotype (t=0.18, P=0.861), the GC/CC genotype to the GG genotype (t=0.42, P=0.677) and in any comparison of the corresponding subgroup analyses.