IJE Advance Access originally published online on July 26, 2006
International Journal of Epidemiology 2006 35(4):944-947; doi:10.1093/ije/dyl149
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Commentary |
Commentary: Fibrinogen and coronary heart disease—test of causality by ‘Mendelian’ randomization by Keavney et al.
Department of Epidemiology and Population Health, London School of Hygiene and Tropical Medicine, Keppel Street, London WC1E 7HT, UK
* Corresponding author. E-mail: tom.meade{at}lshtm.ac.uk
The theory of Mendelian randomization (MR) represents a relatively new and promising way of determining whether associations between risk factors and important clinical events such as myocardial infarction (MI) are of causal significance or not. But how does it work out in practice? A good deal of attention has been given to the relation between fibrinogen and MI. In this issue of the IJE, Keavney et al.1 conclude that usual fibrinogen concentrations do not materially influence coronary disease incidence and that their analysis provides strong evidence (our italics) that fibrinogen is not a major determinant of coronary risk. They acknowledge that in this case, the contribution of a common polymorphism to the level of fibrinogen is small, so the contribution of any polymorphism to MI risk is also likely to be small. So far, so good. But there are a number of other questions to be answered about their analyses and the strength of their conclusions. Similar findings have also been recently reported by Davey Smith et al.2 based on a meta-analysis of published data and are also considered by Keavney et al. A concern of practical relevance is that by erroneously concluding that elevated fibrinogen levels are not to any degree or under any circumstances causal, no further work will be carried out to develop fibrinogen-lowering treatments. If so, patients at high fibrinogen-MI risk might be denied the benefit of a potentially useful therapeutic approach.
 |
Mendelian randomization |
|---|
 Top Mendelian randomization Estimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
MR requires four different estimates to be made. These are, first, the relationship between the level of the risk trait and the disease, second, the relationship between a selected genotype and levels of the risk trait, and third, the relationship between the same genotype and disease risk. Usually from the first two estimates a fourth, the predicted relationship between genotype and disease risk, is computed and compared with the third. At the core of these calculations is the expectation that the first estimate—that of the risk trait effect on disease—is biased by uncontrolled/uncontrollable confounding factors. If so, the fourth estimate—the predicted genetic effect on disease—is also expected to be biased. If, in the absence of biases, however, the observed and predicted gene–disease associations are similar and significant, the results are thought to be compatible with the risk trait being causally related to the disease. If they instead disagree, causality is refuted (more formally, an unbiased estimate of the causal effect of the risk trait can be estimated using instrumental variables methods3). Usually, however, the random variation affecting these estimates is such that the confidence intervals (CIs) for the observed and predicted gene–disease effects overlap, leaving room for uncertainty. Moreover, each of the four estimation steps is based on a number of underlying assumptions and is, therefore, associated with potential systematic errors and biases. These are now considered (see Figure 1).
|
|
Estimating the relationship between fibrinogen levels and MI risk |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
Keavney et al. estimate the fibrinogen-MI risk association using data from cases and controls in the ISIS study. Their estimated odds ratio of 2.99 per 1 g/l (95% CI 2.55–3.30) is based on 57% of the cases and 85% of the controls, because it is restricted to participants with complete data on fibrinogen, ß-fibrinogen genotype, and potential confounders. Given the considerably reduced data and the unequal coverage of cases and controls, selection bias may have affected the estimate. Further, the estimation included corrections for (i) ‘usual’ fibrinogen using the within-subject correlation (R = 0.59) obtained from repeated measures on a subset of the controls, and (ii) pre-MI fibrinogen measures of the cases by imputation from post-MI measures according to a constant adjustment factor of 0.4% per hour from pain onset. Several arguably questionable assumptions, therefore, underlie this element of the MR-based test of causality. First, it is likely that the post-MI rise in fibrinogen and the effects of anti-inflammatory agents after the event vary widely between individuals. This is surely a scenario where actual values, not estimates, are needed before the event—more difficult but still feasible. Consequently, the ‘usual fibrinogen’ correction method, which implies equal amounts of within-subject variability affecting controls and cases, may well be invalid. Second, the same correction factor was used for every logistic model fitted, even those that include potential confounders. This is incorrect,4 since not accounting for the correlation between fibrinogen and the other variables in the model—in particular smoking, BMI, and the apoB/apoA1 ratio—may have led to over-adjustment. Third, the adjustment depends heavily on the time gap between repeated measures, as shown by the Fibrinogen Studies Collaborative (FSC) group5. The longer the gap—as in the 2–3 years of Keavney et al.—the larger the correction, because the within-subject correlation decreases. Finally this may not be valid for fibrinogen: clot structure is clearly determined in part by fibrinogen concentration,6,7 and small changes in subjects with high levels may be causing a greater thrombotic risk than small changes in those with low levels.
The reported estimate of an OR of 2.99 per m/l increase in fibrinogen is, therefore, likely to be biased for several reasons. The recent paper from the Fibrinogen Studies Collaboration (FSC) group, based on 1.38 million person years of follow-up and 6944 fatal or non-fatal MI events,5 may provide a more accurate estimate of the effect of ‘usual’ fibrinogen on non-fatal MI [hazard ratio per g/l: 2.29 (2.13–2.45)]. Interestingly, however, the FSC paper also shows some evidence (albeit of borderline-significance) that the relationship between elevated fibrinogen and risk was stronger in studies where a ‘functional’ assay (clotting time) of fibrinogen was used rather than a non-clotting method.
|
Estimating the relationship between genotype and fibrinogen levels |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
The most widely studied fibrinogen polymorphism is the –455G>A change in the promoter region of the ß fibrinogen gene (FIBB). This shows essentially 100% allelic association with the nearby –148C>T change (complete linkage disequilibrium, LD) and this allows studies that have typed for either SNP to be combined. Based on the 1923 female and 1537 male ISIS control subjects, Keavney et al. estimate a fibrinogen-raising effect of 0.14 g/l per FIBB allele (standard error (SE) = 0.024; P < 0.0001), and a co-dominant effect, i.e. subjects with two raising alleles having, on average, levels 0.28 g/l higher than those with none. Surprisingly, they do not correct for ‘usual’ levels: the estimated effect of FIBB on ‘usual fibrinogen’ would still be 0.14 g/l but its SE would be larger. Davey Smith et al.,2 obtained a similar estimate of effect (0.117 g/l, (0.091–0.142)) from an overview of more than 10 000 subjects, but may have inappropriately combined results of studies that used different methods for measuring fibrinogen concentration, such as nephelometry and clot weight. Whether either estimate is affected by confounding due to factors such as smoking, age, gender, etc. is also an issue. Keavney et al. did not find any statistical evidence for this, but other larger studies (e.g. Tybjaerg Hansen et al.8) have reported effect modifications.
Despite these concerns, the FIBB genotype appears to be robustly associated with elevated fibrinogen measures. This is, however, only one condition for its use in MR, as the latter requires that the genotype is associated only with the quantity but not the quality of the cognate protein. This assumption is invalid for the SNPs used here. It is well known that there is strong LD between SNPs at the 
, ß, and 
fibrinogen loci, and that several common SNPs that alter amino acids in either the A gene (e.g. A312T) or the Bß chain (e.g. FIBB R448K) show varying degrees of LD with the two promoter SNPs (e.g. Mannila et al. 20059). Also, the chain has a variable length extension at the C-terminus (called ') and the control of its expression and its functional significance on fibrinogen and fibrin clot structure have unclear. Fibrinogen isolated from people carrying these different amino acid alleles has different properties with respect to the resulting fibrin gel formed by the clotting process.6 These affect fibre dimensions, clot porosity, and, thus alter fibrinolytic resistance. All these may influence the clinical importance (thrombotic potential) of the fibrinogen produced from the –455G>A and –148C>T genotypes in ways that are hard to predict. Thus, the real effect of FIBB genotype on the resulting clot structure may be inappropriately suggested. It is, therefore, possible that the use of the –455G>A or –148C>T SNPs are inappropriate for MR studies, and data using other SNPs will be required to examine the actual effect of fibrinogen. Interestingly, a recent study reported that a haplotype of several SNPs in the and fibrinogen loci was associated with elevated MI risk, while other haplotypes were associated with fibrinogen levels but not risk,9 so such SNPs may already be available.
|
Estimating the relationship between genotype and MI risk |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
It is not surprising that both Davey Smith et al. and Keavney et al. come up with similar estimates for the genotypic effect on MI risk, since Davey Smith et al. used published data for their estimate [0.98 (0.92–1.04)], while Keavney et al. used the same data with the addition of the ISIS study (1.06 (0.96–1.16)) where the estimated effect in ISIS alone is 1.00 (0.96–1.05). These estimates do not support a direct effect. However, the concern is whether these estimates are flawed by study design, population substructure, or other hidden confounders. The ISIS contribution to this estimate is based on a subset of its total population and suffers from the usual limitations of case–control studies relative to cohort studies, e.g. selection bias arising from the identification and participation of surviving cases and controls. If the mortality rates from all causes are affected by fibrinogen levels and the FSC group did find such an association5, then the genotype distribution among the cases would be distorted. Further, there might be a difference in allele frequency in different recruitment regions (say, higher in Scotland than in England or in Europe10) and this high allele group might make up a higher proportion of cases than the controls. If so, there could have been population stratification and confounding that cannot be identified and controlled.
As regards hidden confounders, the estimate of the genotype-MI risk is taken to represent the true unbiased value under other assumptions. It is assumed that there is no ‘canalisation’ or compensation to the altered gene expression, i.e. the higher fibrinogen produced from the T-allele during development. Although Davey Smith et al. state that they ‘know of no evidence to either support or refute’ such a concept, it appears highly plausible. The clotting system is adapted to be extremely homeostatic, and levels of many components of the system change markedly during different stages of human development, usually without major changes in risk of thrombosis or bleeding (e.g. in pregnancy). It is thus at least possible that compensation of the modest elevation found in carriers of the –455A allele could occur. Another major underpinning assumption, discussed by both Davey Smith and Ebrahim11 and Nitsch et al.12 is that other risk factors for MI should be distributed randomly with respect to the genotype under examination, so avoiding confounding. This assumption is not true for the ISIS study, where the ratio of apoB/apoA1 is significantly higher in –148TT subjects. The finding is dismissed by Keavney et al. as being of minor importance and of borderline statistical significance given the number of comparisons made, but nevertheless is a breach of the assumption. The higher apoB/apoA1 ratio in carriers might be a consequence of LD between the –148C/T polymorphism and other genes influencing these two factors, or some life-time ‘adaptation’ to the genotypically induced raised fibrinogen. In each of these scenarios, the association between genotype and non-fatal MI would not act exclusively via plasma fibrinogen.
|
Predicting the relationship between genotype and MI risk |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
Keavney et al. predict a per allele effect on MI risk of 1.17 (1.14–1.19), which they contrast with their observed effect of 1.00 (0.96–1.05). On the other hand, Davey Smith et al. use published data on the effect of fibrinogen on MI risk, not of the genotypic effect. They report an apparently protective effect (OR = 0.81 per g/l; 0.46–1.40), which they compare with the observed OR of 1.80 per g/l (1.6–2.0). Both results ignore the uncertainty associated with the genotype–fibrinogen association used to compute the predictions. As regards the predicted OR by Keavney et al., if this uncertainty were accounted for, the CI would be much wider. Using their reported CIs for the effect of fibrinogen per 1 g/l (and taking their estimate to be an odds ratio, not a risk ratio, given the study design), the
95% CI for the predicted genotype-MI risk would be (1.10–1.23), partly overlapping the CI for the observed OR of (0.88–1.13). The width of our corrected CI would be wider if we could additionally account for the ‘usual fibrinogen’ correction discussed in Section 2. The extent of the overlap between observed and predicted CIs is confirmed in simulations carried out using data structure and observed relations, similar to those in Keavney et al. While this overlap is not large, there are, as we have pointed out, several other considerations to be taken into account in drawing any conclusions. Thus, the data to refute causality with certainty do not yet exist. |
Conclusions |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
So, overall, how safe is the conclusion that ‘long term differences in fibrinogen concentration are not a major determinant of cardiovascular disease risk’? Given the current limited state of genetic knowledge and the biases we have considered, this seems premature. Carrying the rare FIBB promoter allele does result in a robust but modest elevation of blood fibrinogen measures, but does not appear to increase risk of non-fatal MI to any significant extent. Given the current uncertainties, however, its predicted impact on risk cannot be estimated with accuracy, and further genotyping with additional SNPs is urgently required. We also believe that the paper by Keavney et al. overlooks the probable importance of the pathways through which high fibrinogen levels may lead to MI, especially via its contribution to blood viscosity. Ultimately, if data support the view that elevated fibrinogen is an independent causal risk factor only in those with extreme levels, or indeed if elevated fibrinogen is both a marker of disease as well as making a small causal contribution,13 it is still useful as a test to identify those with elevated cardiovascular disease risk. It is worth pointing out that in the FSC group report, in the subset of 7000 subjects with both fibrinogen and CRP measures, adjustment for CRP (as a marker of inflammation) did not significantly reduce the predictive value of elevated fibrinogen, supporting the concept that elevated fibrinogen is not acting simply as an inflammatory marker.
We have intended to focus mainly on the theory and practice of MR using the paper by Keavney et al. as an example. Nevertheless, there are some particular shortcomings in their account to which we have drawn attention. The part played by fibrinogen in CHD—marker, cause, or both—remains uncertain and is likely to remain unresolved unless adequate trials of selective fibrinogen-lowering agents are carried out. The current debate is reminiscent of the controversies over cholesterol and CHD before the value of lipid-modifying agents was demonstrated. The ‘fibrinogen story’ is a much more recent one, by 20 years at least. As we indicated at the outset, it would be short-sighted and quite possibly counter-productive if premature conclusions based on MR in its current state of development were taken to have settled the contribution of fibrinogen to risk of MI. The conclusions of Keavney et al. and Davey Smith et al. both remain ‘not proven’, and for some traits and genes, the key underlying assumptions necessary for the valid application of MR for investigating causality may be hard to satisfy.
|
References |
|---|
|
TopMendelian randomizationEstimating the relationship...Estimating the relationship...Estimating the relationship...Predicting the relationship...ConclusionsReferences |
|---|
1 Keavney B, Danesh J, Parish S et al. Fibrinogen and coronary heart disease: test of causality by "Mendelian randomisation". Int J Epidemiol 2006;35:935–43.
2 Davey Smith G, Harbord R, Milton J, Ebrahim S, Sterne JA. Does elevated plasma fibrinogen increase the risk of coronary heart disease? Evidence from a meta-analysis of genetic association studies. Arterioscler Thromb Vasc Biol 2005;25:2228–33.
3 Greenland S. An introduction to instrumental variables for epidemiologists. Int J Epidemiol 2000;29:722–29.
4 Rosner B, Spiegelman D, Willett WC. Correction of logistic regression relative risk estimates and confidence intervals for measurement error: the case of multiple covariates measured with error. Am J Epidemiol 1990;132:734–45.
5 Fibrinogen Studies Collaboration Group. Plasma fibrinogen level and the risk of major cardiovascular diseases and nonvascular mortality: an individual participant meta-analysis. JAMA 2005;294:1799–809.
6 Lim BC, Ariens RA, Carter AM, Weisel JW, Grant PJ. Genetic regulation of fibrin structure and function: complex gene-environment interactions may modulate vascular risk. Lancet 2003;361:1424–31.[CrossRef][Web of Science][Medline]
7 Mannila MN, Eriksson P, Eriksson CG, Hamsten A, Silveira A. Epistatic and pleitropic effects of polymorphisms in the fibrinogen and coagulation factor XIII genes on plasma fibrinogen concentration, fibrin gel structure and risk of myocardial infarction. Thromb Haemost 2006;95:420–27.[Web of Science][Medline]
8 Tybjaerg-Hansen, Agerholm-Larsen B, Humphries SE, Abildgaard S, Schnohr P, Nordestgaard BG. A common mutation (G-455-A) in the beta-fibrinogen is an independent predictor of plasma fibrinogen, a risk factor for ischemic heart disease. J Clin Invest 1997;99:3034–39.[Web of Science][Medline]
9 Mannila MN, Eriksson P, Lundman P et al. Contribution of haplotypes across the fibrinogen gene cluster to variation in risk of myocardial infarction. Thromb Haemost 2005;93:570–77.[Web of Science][Medline]
10 Mannila MN, Silveira A, Hawe E et al. Plasma fibrinogen concentration predicts the risk of myocardial infarction differently in various parts of Europe: effects of beta-fibrinogen genotype and environmental factors. The HIFMECH Study. Thromb Haemost 2004;92:1240–49.[Web of Science][Medline]
11 Nitsch D, Molokhia M, Smeeth L, DeStavola BL, Whittaker JC, Leon DA. Limits to causal inference based on Mendelian randomization: a comparison with randomized controlled trials. Am J Epidemiol 2006;163:397–403.
12 Meade TW. Fibrinogen measurements to assess the risk of arterial thrombosis in individual patients: yes. J Thromb Haemost 2005;3:632–34.[CrossRef][Web of Science][Medline]
13 Danesh J. Collins R, App Peto R. Association of fibrinogen, C-reactive protein, albumin, or leukocyte count with coronary heart disease. J Am Med Assoc 1998;279:1477–82.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  G. D. Smith 'Something funny seems to happen': J.B.S. Haldane and our chaotic, complex but understandable world Int. J. Epidemiol., June 1, 2008; 37(3): 423 - 426. [Full Text] [PDF]
|
|
|
|
 |
|
 Associations of Plasma Fibrinogen Levels with Established Cardiovascular Disease Risk Factors, Inflammatory Markers, and Other Characteristics: Individual Participant Meta-Analysis of 154,211 Adults in 31 Prospective Studies: The Fibrinogen Studies Collaboration Am. J. Epidemiol., October 15, 2007; 166(8): 867 - 879. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. M. van Dam and F. B. Hu Lipocalins and Insulin Resistance: Etiological Role of Retinol-Binding Protein 4 and Lipocalin-2? Clin. Chem., January 1, 2007; 53(1): 5 - 7. [Full Text] [PDF]
|
|
|
|
|
|
G. D. SMITH Mensuration, Mendel, and a 19th century public health justification for US imperialism Int. J. Epidemiol., August 1, 2006; 35(4): 811 - 813. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||