Introduction Rare genetic defects that cause extremely high plasma homocysteine levels also cause coronary heart disease (CHD) [1]–[3]. It was therefore hypothesised that, even within the normal range of plasma homocysteine concentrations, higher levels might appreciably increase CHD risk [3]. Retrospective studies originally suggested a strong relationship, but subsequent prospective observational studies suggested weaker associations [3],[4]. A meta-analysis of prospective studies found that, after adjusting for known risk factors, 25% lower usual homocysteine level (achievable in many populations by fortification of cereals with folic acid) was associated with only about 11% (95% CI 4%–17%, p 100 people (Figure S2; Table S4 in Text S1). Mean folate levels were estimated for each of these five categories on the basis of secular and geographic trends in folic acid fortification policies (Appendix S1 in Text S1). Randomized Folate Trials Finally, we updated a previous meta-analysis [5] of seven large-scale placebo-controlled trials assessing the effects on cardiovascular disease of lowering homocysteine with B-vitamins by adding three trials [22]–[24] that reported their results after publication of the meta-analysis (Table S5 in Text S1). The additional trials were identified by searching the electronic literature using search terms “cardiovascular disease,” “coronary heart disease,” “coronary stenosis,” “myocardial infarction” and “randomized controlled trial,” “clinical trial,” and “folic acid” or “B-vitamins.” As in the original meta-analysis [5], additional randomized trials were eligible if (i) they involved a double-blind randomized comparison of B-vitamin supplements containing folic acid versus placebo for the prevention of vascular disease; (ii) the relevant treatment arms differed only with respect to the homocysteine-lowering intervention; and (iii) the trial involved ≥1,000 participants with treatment duration of ≥1 y. Statistical Methods Mean folate levels and mean log homocysteine by genotype were estimated from individual participant data where available, or from published reports. In calculating these means we sought to give all individuals similar weight, so large studies contribute proportionally more than small ones. (Random effects models were not used, as they can give undue weight to individuals in smaller studies [25].) The homocysteine difference between TT and CC genotypes was estimated from linear regression (stratified by study) of log homocysteine on genotype in heterozygotes [26],[27]. The CHD OR for TT versus CC genotype (OR) was estimated by logistic regression, stratified by study; this yields an approximately inverse-variance-weighted average of the log OR in each study. In the PROCARDIS study, which included both related and unrelated cases and controls, allowance for familial clustering was made, which slightly increased the variance estimate [15]. In the LOLIPOP and PROMIS studies of South Asians also, the CHD OR for TT versus CC genotypes was estimated after correction for population admixture (to avoid false positive association due to population stratification) using adjustment for principal components involving the results of random genetic markers within that study [15], which was not possible in the published studies. Details of the methods used to estimate nonpublication bias are shown in Appendix S2 in Text S1. Heterogeneity was assessed using chi-squared tests [28], also citing I 2 = 100%(1−[degrees of freedom]/[chi-squared test statistic]) [29]. CIs are 95%, except where specified as 99% to allow for multiple comparisons. Analyses used SAS version 9.1. Results Figure 1 plots mean folate levels by calendar year in 81 population surveys (total 200,103 participants), categorising the surveys by study place (Asia, Europe or North America and Australasia [US & ANZ]) and time (before or after national folate supplementation began). Asian surveys were all in unsupplemented populations, so Figure 1 defines only five probable folate status categories. Table 1 gives the mean folate levels in each category. Although assay methods may have varied, there appeared to be similarly low folate levels in the Asian and unsupplemented European populations (11.0 and 11.9 nmol/l), intermediate folate levels in the supplemented European and unsupplemented US and ANZ populations (18.2 and 20.8 nmol/l), and high folate levels in supplemented US and ANZ populations (33.3 nmol/l). Thus, there are only two low-folate unsupplemented categories. 10.1371/journal.pmed.1001177.g001 Figure 1 Mean serum folate concentrations in 81 population surveys, by calendar year and region. White squares, no folate supplementation; black squares, after folate supplementation; broken vertical line, 1995–1996, when folate supplementation began in the United States, Canada, Australia, New Zealand (US & ANZ), and some but not all European countries. No Asian surveys were in supplemented populations. 10.1371/journal.pmed.1001177.t001 Table 1 Relevance in population surveys of study place and time to (i) the mean general population serum folate level, and (ii) the excess plasma homocysteine level in the TT versus CC MTHFR C677T genotype. Region, and Whether after Folate Supplementation Surveys of Folate Levels Studies of MTHFR C677T Genotype and Plasma Homocysteine Folate Surveys n people Mean (SE) Serum Folate Concentration, nmol/la Homocysteine MTHFR Studies n People Percent Higher Homocysteine, TT Versus CC (and 99% CI)b Asiac no supplementation 7 4,841 11.0 (0.014) 15 6,553 25 (21–30) Europe, presupplementation 21 31,767 11.9 (0.006) 14 24,199 21 (19–24) Europe, post-supplementation 30 13,504 18.2 (0.009) 25 8,702 18 (15–22) US & ANZ, presupplementation 13 57,104 20.8 (0.004) 8 26,853 13 (11–15) US & ANZ, post-supplementation 10 92,887 33.3 (0.003) 8 2,062 7 (2–13) All regions and time periods 81 200,103 24.8 (0.002) 70 68,369 18 (17–19) a Mean folate levels average all who were surveyed; SE denotes the standard error due only to within-survey variation. Between-survey variation in folate levels is illustrated in Figure 1. b From inverse-variance-weighted averages of within-study differences in log homocysteine; Figure S1, Table S2 in Text S1. c Mainly of Japanese, Chinese, or Korean populations; none of South Asians. Homocysteine differences by MTHFR genotype are also given in Table 1, based on 70 biochemical studies of MTHFR genotype and homocysteine in the general population (total 68,369 participants, mostly Caucasian or East Asian). These analyses of within-study percentage differences in homocysteine levels between TT and CC genotypes (Figure S3) should be little affected by any variation in homocysteine assay methods. The TT versus CC homocysteine difference appears to have been only moderately affected by folate supplementation, but was appreciably greater in Asia and Europe than in the US & ANZ (although the TT versus CC homocysteine difference in US & ANZ after folate supplementation had a wide CI and is not reliably known). Differences in homocysteine between the CT and CC genotypes were only about a quarter as great as those between the homozygous TT and CC genotypes (Figure S3). Tables S3 and S4 in Text S1 give separately each survey of folate levels and each study of MTHFR genotype and homocysteine, and Tables S1, S2 in Text S1 give separately each case-control study result. Among the controls there was substantial variation in genotype frequencies (ratio of TT to CC 0.03–0.04 in South Asians, 0.2–0.3 in northern Europe, 0.4 or more in Japan, and 0.7 in Italy), illustrating the potential for bias from population substructure. Our case-control analyses of MTHFR genotype and CHD risk compare TT versus CC homozygotes, as this is the comparison that involves the greatest homocysteine contrast. The findings for CHD risk in the unpublished datasets are given in Figure 2, subdivided by the number of variants examined (i.e., genotyping panel size). Overall, there would have been about a 20% excess homocysteine associated with the TT versus the CC genotype, but the excess CHD risk associated with the TT versus the CC genotype was only 2%, was not significant (OR = 1.02, 95% CI 0.98–1.07, p = 0.28), and was similar in the datasets with larger and small genotyping panel sizes. Any null bias from nonpublication would have biased the expected log OR in the aggregate of all unpublished studies downwards by only about 0.001 (0.003 in the small-panel studies and 0.0002 in the large-panel studies: Appendix S2 in Text S1), thereby multiplying the overall OR by 0.999, which is negligible. 10.1371/journal.pmed.1001177.g002 Figure 2 Homozygote CHD OR (TT versus CC MTHFR C677T genotype) in 19 unpublished datasets, yielding 24 parts that are classified by genotyping panel size. For these datasets, being unpublished introduces a negligible bias (less than 0.3% for each OR and about 0.1% for the overall OR: eAppendix 1). Black squares indicate OR (with areas inversely proportional to the variance of log OR), and horizontal lines indicate 99% CIs. The subtotals and their 99% CIs are indicated by black diamonds. The overall OR and its 95% CI is indicated by a white diamond. The weight (defined as the inverse of the variance of the maximum likelihood estimate of the log OR) and the product of the weight times OR indicates how much each study has contributed to the subtotals and totals. Because the weights and products are approximately additive, they can be used to estimate the effects of ignoring particular studies, or of grouping studies in different ways. Figure 3 categorizes these results by the probable folate status of the populations studied. Half the evidence was from low-folate unsupplemented populations in Asia or Europe. But, even if attention is restricted to these populations (where the excess of homocysteine associated with the TT versus CC genotype would have been somewhat greater than elsewhere), there was still no evidence that the TT genotype was associated with any excess risk of CHD (OR = 1.01: 1.03 in low-folate Asia, 0.99 in low-folate Europe; Figure 3). As the homocysteine difference between CT and CC genotypes is only about a quarter of that between TT and CC genotypes, inclusion of the CT results does not materially alter these findings (Figure S4). Thus, the aggregated results from the 19 unpublished datasets suggested little or no hazard, even in unsupplemented low-folate populations. 10.1371/journal.pmed.1001177.g003 Figure 3 Homozygote CHD OR (TT versus CC MTHFR C677T genotype) in each probable folate status category, from meta-analyses of 19 unpublished datasets (all large). Average homocysteine difference (in the non-CHD general population) for all areas and periods is weighted in proportion to the numbers of TT CHD cases in all 19 unpublished datasets. Nonpublication involves negligible bias: Appendix S2 in Text S1. In contrast, the aggregated TT versus CC results from the 86 published studies (total 28,617 cases and 41,857 controls: Figure 4; Figure S5) suggested a 15% excess risk of CHD (OR 1.15, 95% CI 1.09–1.21), which is significantly discrepant (p = 0.001) with the results from the unpublished datasets (Figure 2). Larger studies may be less prone than smaller ones to selective publication based on their findings and may also be less prone to other, less clearly recognizable, methodological problems (and, publication bias may involve not only random but also any systematic errors due to preferential publication of positive results) [30]. In Figure S5, the CHD ORs in each of the 86 published studies are therefore ordered by study size (as defined by the variance of the log OR). Figure 4 indicates that, although the 72 smaller published studies contributed most to the suggestion of increased risk (OR = 1.18), the 14 larger published studies, which typically had >250 cases and >250 controls, also contributed to some extent (OR = 1.12). 10.1371/journal.pmed.1001177.g004 Figure 4 Homozygote CHD OR (TT versus CC MTHFR C677T genotype) in each probable folate status category, from meta-analyses of 86 published studies, 14 large (i.e., variance of log OR less than 0.05) and 72 smaller studies. Black squares indicate OR (with areas inversely proportional to the variance of log OR in each subdivision), and horizontal lines indicate 99% CIs. The overall OR and its 95% CI are indicated by a black diamond. Average homocysteine difference (in the non-CHD general population) for all areas and periods is weighted in proportion to the numbers of TT CHD cases in all 86 studies. Of these large studies, only two, both from Japan, suggested significantly increased risk. None of the others did, including those from low-folate Europe, where TT versus CC homocysteine differences were probably similar to those in Japan (Figure S3). The large published and unpublished Japanese studies are described separately in Table S6 in Text S1; in these, there appeared to be substantial heterogeneity in the TT and CC genotype frequencies (the odds, TT/CC, varied from 0.23 to 0.68 in controls), which makes it difficult to interpret the findings. (All these studies were located in mainland Japan, where there is little ethnic heterogeneity, so no large differences in genotype frequency would be expected [31].) The large published studies in all other populations, like the unpublished datasets, indicated no material effect between homozygote genotype and CHD risk. Overall, almost half of the cases in the published CHD studies also had data on homocysteine, but those in the only two large studies with significantly increased risk did not. Hence, when analyses were restricted to the subset with homocysteine no significant association between TT versus CC genotype and CHD risk remained (unpublished data). For the ten large trials of B-vitamins for homocysteine reduction (Table S5 in Text S1), Figure 5 shows that folate supplementation (which reduces normal homocysteine levels by about 25%) had little or no effect on the 5-y incidence of CHD incidence (rate ratio, folate versus placebo, 1.02, 95% CI 0.96–1.08). 10.1371/journal.pmed.1001177.g005 Figure 5 Effects of folic acid on major coronary events (nonfatal myocardial infarction or coronary death) in a meta-analysis of the published results of all large randomized trials of homocysteine reduction. Data for the VITATOPS trial are for myocardial infarction only. Data for FAVORIT are for all cardiovascular disease outcomes. Symbols and conventions as in Figure 2. Discussion The present meta-analyses of unpublished datasets involving 48,175 cases and 67,961 controls finds no evidence of an increased risk of CHD in TT versus CC homozygotes for the MTHFR C677T polymorphism, either in all such datasets or in those from unsupplemented low-folate populations. This null result is not materially affected by publication bias and is significantly discrepant with the moderately positive association found in our meta-analysis of 86 published studies of this question, or, equivalently, in other recent meta-analyses of published studies [6],[10]–[12]. Although publication bias (involving not only random errors but also any systematic errors in particular studies) may well have appreciably affected the meta-analyses of published studies, nonpublication bias (i.e., failure to publish null results) should have had a negligible effect on the present meta-analyses of unpublished studies. For ORs of the magnitude that may be plausible for MTHFR (i.e., about 1.08), the probability of a result from a large SNP panel study reaching statistical significance after allowance for multiple testing can be shown to be negligible (i.e., biasing the overall log OR by only about 0.001; Appendix S2 in Text S1). In these datasets, the TT versus CC comparison involves a nonsignificant excess CHD risk of only about 2% in all populations and 1% in low-folate unsupplemented populations (both with upper confidence limit 7%). Consistent with the null results of the folate trials, the results of the present meta-analyses of unpublished MTHFR studies provide no evidence for an association of life-long moderate elevations in homocysteine levels with CHD risk and support the suggestion [12] that the associations observed in meta-analyses of previously published MTHFR studies may be an artefact of publication bias. The discrepancy between the overall results in the unpublished and the published datasets is too extreme to be plausibly dismissed as a chance finding (as is the discrepancy between the published results in low-folate Europe and Japan, which refutes the suggestion that differences in folate supplementation could explain the differences between Japanese and other published studies). Some studies, particularly if small, might have been prioritised for publication by investigators, referees, or editors according to the positivity of their results [30], and some may have been liable to other methodological problems that bias the average of all results. To avoid such biases, we chiefly emphasise the new results from the previously unpublished datasets. These show little or no hazard in Japan or elsewhere from moderate lifelong elevation of normal homocysteine levels. The magnitude of the effect of publication bias is substantial and in addition to distorting the association of MTHFR with CHD in published studies, publication bias may also help explain the discrepant findings recently reported for MTHFR and stroke [32]. Genetic epidemiology of the effects of common polymorphisms on common diseases is increasingly dominated by consortia of GWA studies with tens of thousands of cases and large panels of tens or hundreds of thousands of polymorphisms [15],[16]. Thus, GWA (or other large panel genotyping) studies offer the possibility of avoiding unduly data-dependent emphasis on particular studies or on particular genetic loci and of making sophisticated allowance for population admixture. (Such allowance was not possible in the published studies and was available to us from only some of the unpublished datasets [15].) Although there is little evidence of significant population admixture in mainland Japan [31], the control frequency of the T allele varied somewhat across the Japanese case-control studies (0.33–0.45, Table S6 in Text S1), perhaps because variation in genotyping methods can affect MTHFR C677T genotype calls. As these small differences in T-allele frequency correspond to substantial differences in the TT/CC odds (Table S5 in Text S1), they reinforce the potential importance of cases and controls being blindly genotyped (assayed, called, and quality-control filtered) together, particularly for a polymorphism such as MTHFR C677T that varies in frequency between populations and does not have a substantial effect on risk. The Mendelian randomization approach to assessing the effects of a particular biochemical factor such as homocysteine assumes no relevant pleiotropic effects of the genetic variant on other factors [33],[34] (whereas, for example, the TT genotype does also slightly affect folate levels) [35]. Large randomized trials of folate supplementation also provide an independent test of the causal relevance of homocysteine (assuming no material effects of folate on CHD except via homocysteine). A meta-analysis of 10 trials involving 50,378 participants had little or no effect on the 5-y incidence of CHD (rate ratio, folate versus placebo, 1.02, 95% CI 0.96–1.08). The null result from the folic acid trials is now directly reinforced by this Mendelian randomization meta-analysis of unpublished genetic epidemiology datasets, which is not materially affected by publication bias, involves large numbers of relevant outcomes, and shows no evidence that even a lifelong 20% difference in plasma homocysteine (within the normal range) meaningfully effects CHD risk. Supporting Information Figure S1 Screening and selection of articles for MTHFR and CHD risk and MTHFR and homocysteine levels. (TIF) Click here for additional data file. Figure S2 Screening and selection of population surveys of folate status. (TIF) Click here for additional data file. Figure S3 Percent higher homocysteine by MTHFR C677T genotype in 70 biochemical studies of non-CHD populations. Subtotal results are from inverse-variance-weighted averages of within-study differences in log homocysteine, so the 95% CIs for them (solid diamonds) reflect only the within-study variation; other CIs are 99% CIs. (TIF) Click here for additional data file. Figure S4 CHD OR (OR, TT versus CC MTHFR C677T genotype) from CC/CT/TT results in 19 unpublished datasets, yielding 24 parts that are classified by probable folate status category: maximum likelihood estimate, assuming that the underlying log OR for CT/CC is 0.25 times that for TT/CC. Black squares indicate OR, and horizontal lines indicate 99% CIs. The subtotals and their 99% CIs are indicated by black diamonds. The overall OR and its 95% CI is indicated by a white diamond. The weight (defined as the inverse of the variance of the maximum likelihood estimate of the log OR) and the product of the weight times OR indicates how much each study has contributed to the subtotals and totals. (TIF) Click here for additional data file. Figure S5 CHD OR for MTHFR TT versus CC genotype in 86 published studies, from Table S4, classified by probable folate status category and sorted by effective study size (i.e., variance of log OR, for which the cutoff 0.05 is indicated by dashed lines). Weight is the inverse of the variance of the maximum likelihood estimate of the log OR. Additivity of the weights is therefore only approximate. NB, presupplementation Europe subtotal allows for the common control group in Frederiksen-Prospective (P) and Frederiksen-Case-Control (CC). 95% CIs for total; other CIs are 99%. (TIF) Click here for additional data file. Text S1 Webmaterial for homocysteine and coronary heart disease: meta-analysis of MTHFR case-control studies, avoiding publication bias. (PDF) Click here for additional data file.
