Data were independently gathered in duplicate by two investigators on the basis of a standard protocol. Discrepancies were settled by discussion until a consensus was reached. A literature search was conducted for studies that examined associations between the +104T/C polymorphism and knee osteoarthritis. A search was performed on PubMed, Google Scholar and the China National Knowledge Infrastructure database using the following keywords: ‘5′-UTR or GDF5’, ‘+104T/C polymorphism or genotype’ and ‘knee osteoarthritis’.

To be eligible for inclusion in this meta-analysis, the following criteria were established: i) case-control study that addressed knee osteoarthritis cases and healthy controls; ii) studies that evaluated the association between +104T/C polymorphism and knee osteoarthritis risk; iii) studies that included sufficient genotype data for extraction; and iv) the genotype distribution among the control population in Hardy-Weinberg equilibrium (HWE).

The following information was extracted from the eligible studies: i) first author; ii) year of publication; iii) country; iv) ethnicity; v) sample size of subjects with and without CAD; vi) genotype distribution of subjects with and without knee osteoarthritis; and vii) P-value for HWE test in subjects without knee osteoarthritis.

Crude odds ratios (ORs) with 95% confidence intervals (CIs) were calculated to estimate the strength of correlation between the +104T/C polymorphism and knee osteoarthritis risk. The pooled ORs were performed under a homozygote comparison (TT vs. CC), a heterozygote comparison (TT vs. TC), a dominant model (CC+TC vs. TT) and a recessive model (TT+TC vs. CC). In addition, we conducted a stratification analysis in Asian and European individuals to explore and explain diversity among the results of various studies (7). We tested whether genotype frequencies of controls were in HWE using the Chi-square test. Q- and I² tests were performed to evaluate whether the variation was due to heterogeneity or chance. If heterogeneity was identified among the studies (I²<50%), the pooled OR was estimated by the fixed-effects model. Otherwise, the random-effects model was used to estimate the pooled OR. Sensitivity analysis was performed through the random-effects model values compared with the fixed-effects to ensure the stability of the results. Publication bias was assessed by Begg's test (P<0.05 was considered to indicate a statistically significant difference). Analyses were performed using STATA version 12.0 (Stata Corporation, College Station, TX, USA). All P-values were two-sided.

A total of 33 studies were obtained by the literature search, among which 6 fit the inclusion criteria (8–13). The flowchart of reviews demonstrates the detailed process of selection (Fig. 1). Of the 25 excluded studies, 21 were not associated with +104T/C polymorphism, 1 did not have sufficient genotype data for extraction, 1 was a meta-analysis and 2 were reviews. Data were collected from 2,744 knee osteoarthritis patients and 4,518 controls. Genotype distributions among the controls of all included studies were consistent with HWE. The 6 included studies contained 8 population studies of 5 Asian and 3 Caucasian groups. Information from these studies and the number of cases and controls with TT, TC and CC genotypes reported in each are shown in Table I.

A summary of the meta-analysis findings of the association between +104T/C polymorphism and knee osteoarthritis risk is shown in Table II. Overall, we found a significant association of +104T/C polymorphism with knee osteoarthritis risk (TT vs. CC: OR 1.68, 95% CI=1.41–2.01; TT vs. TC: OR 1.18, 95% CI=1.01–1.38; dominant model: OR 0.72, 95% CI=0.61–0.86). In the subgroup analysis based on ethnicity, studies were divided into Caucasian and Asian populations. The same association was observed in the Caucasian (TT vs. CC: OR 1.45, 95% CI=1.13–1.85) and Asian (TT vs. CC: OR 1.99, 95% CI=1.53–2.60; TT vs. TC: OR 1.33, 95% CI=1.16–1.52; dominant model: OR 0.64, 95% CI=0.56–0.72; recessive model: OR 1.77, 95% CI=1.37–2.29) groups. Sensitivity analysis was performed through random-effects model values compared with the fixed-effects, the results of which indicated no statistically significant difference, rendering the values of meta-analysis valid (TT vs. CC: OR 1.68, 95% CI=1.40–2.01; TT vs. TC: OR 1.19, 95% CI=1.07–1.32; dominant model: OR 0.74, 95% CI=0.67–0.81; recessive model: OR 0.28, 95% CI=0.25–0.32). Publication bias of the studies was assessed using the funnel plot and Begg's test. Publication bias was not observed in the funnel plot (all P>0.05, Table II and Fig 2).