Commentary: Development of Mendelian randomization: from hypothesis test to ‘Mendelian deconfounding’

Martin D Tobin, Cosetta Minelli, Paul R Burton and John R Thompson

University of Leicester, Department of Epidemiology and Public Health, 22–28 Princess Road West, Leicester LE1 6TP, UK. E-mail: mt47{at}leicester.ac.uk

In his letter to the Lancet in 1986, reprinted in this issue of the International Journal of Epidemiology (IJE), Katan described the idea of using data from genetic studies to test for a relationship between a quantitative intermediate phenotype and a disease in a way that is not distorted by confounding or reverse causality.1 Following the application of these ideas by other authors2–5 interest in the concept has grown, although it is still not widely understood. This important and novel method has the potential to improve the way that the quantitative phenotypes that underlie common diseases are investigated, so better informing public health interventions that alter the level of the phenotype in order to reduce the risk of disease.4

Katan described how evidence of the effect of the apolipoprotein E (APOE) genotype on cancer risk could be used to test the hypothesis that ‘a naturally low cholesterol favours tumour growth.’1 Given that ‘the gradient in serum cholesterol levels in the population is associated with a gradient in APOE [genotype]’, under the causal hypothesis we would expect to see a corresponding association between APOE and cancer. The absence of such a genetic association ‘would suggest that the association between low cholesterol and cancer is spurious’. Katan emphasized that APOE genotype is present since birth and is not disturbed by disease, so unlike conventional epidemiological methods the genetic test is not influenced by reverse causation or confounding.

The term ‘Mendelian randomization’ has been used by a number of researchers when applying this idea to investigate phenotype–disease associations2,4,5 but not always in exactly the same way. The term was first used in a completely different context to describe a method of pseudo-randomization in a particular clinical trial for which the randomization of treatment was not otherwise possible6 (see Wheatley and Gray's Commentary in this issue of IJE)7. At its most basic, ‘Mendelian randomization’ simply means that, according to Mendel's laws of segregation and independence,8 a subject's genotype is determined by an apparently random process at conception. So ‘Mendelian randomization’ is a fundamental biological process that should reasonably underpin the appropriate interpretation of any study in which genotype is related to an outcome. However, by common usage, the term ‘Mendelian randomization’ has also become attached to the epidemiological method that appears to be based upon Katan's ideas that generates indirect, and unconfounded, inferences about the association between a phenotype and a disease given direct information on the gene–disease and gene–phenotype associations.4 The confusion and ambiguity that this double meaning engenders is impeding the transmission of ideas about the value of this important epidemiological approach. In our view it would be better to give the epidemiological method an alternative name, such as ‘Mendelian deconfounding’, and to reserve the term ‘Mendelian randomization’ for the more fundamental biological process.

In applying the concept described by Katan, the emphasis originally was on hypothesis testing to confirm or refute the evidence for particular phenotype–disease associations found in observational studies. However, the method can be developed to provide an estimate of the size of the unconfounded effect of a phenotype on disease together with a measure of its uncertainty. In this commentary we describe how these ideas have been developed since Katan's paper and in particular we emphasize the benefits of estimating the size of the effect of phenotype on disease over simple hypothesis testing. Finally we consider possible future developments particularly in regards to meta-analysis.


    Applications of Mendelian randomization to learn about phenotype–disease relationships
 Top
 Applications of Mendelian...
 ORK/{Delta}IP
 Developing the concept: from...
 The future
 References
 
Youngman et al. and Keavney et al. were the first authors to use the term ‘Mendelian randomization’ in a similar epidemiological context to that described by Katan.2,3 Youngman et al. studied fibrinogen levels and beta-fibrogen genotype in premature myocardial infarction (MI) cases and related controls. From their dataset, three associations were observed. These are shown in Figure 1, where G represents the genotype (beta-fibrinogen HindIII), IP the intermediate phenotype (fibrinogen), and D the disease (MI). The association between G and D (shown as a broken line) is induced only through the causal effects of G on IP and of IP on D. The model assumes that there is no other pathway through which the gene exerts its effect on MI. Under the assumption of Mendelian randomization, the measurements of the G–D and G–IP associations are unconfounded, while the measured IP–D association is likely to be confounded and subject to reverse causation.



View larger version (7K):
[in this window]
[in a new window]
 
Figure 1 A pictorial representation of the model used to test for a causal association between intermediate phenotype and disease (IP-D) or to derive an unconfounded estimate of the size of that effect

 
The information available enabled Youngman et al. to assess whether fibrinogen (IP) had a causal link with MI (D). The authors obtained unconfounded estimates of the G–IP (0.12 g/l per allele, standard error 0.018, P < 0.00001) and G–D, odds ratio = 1.03 (95% CI: 0.96, 1.10) associations. As the G–IP association is clearly established, a causal IP–D relationship would have resulted in an observed G–D association. As this G–D link is not seen, their observed IP–D odds ratio of 1.20 (95% CI: 1.13, 1.26) is probably a result of confounding or reverse causation.

The non-significance of the G–D association does lead us to doubt the hypothesized causal pathway. However, this conclusion is subject to all of the reservations we would have about using P-values. An alternative interpretation would be that there were insufficient data to rule out a 10% increase in MI risk for a (modest) 0.12-g/l change in the level of fibrinogen. Thought of in this way, the analysis based on the assumption of Mendelian randomization suggests that any causal IP–D association is not large but it does not rule it out completely. The use of the data to estimate the potential size of the IP–D association is likely to be much more informative than relying on hypothesis tests of G–IP and G–D to rule in or rule out a causal link.

To estimate the size of the unconfounded IP–D odds ratio associated with a specified change, K, in the intermediate phenotype when the measured G–IP difference is {Delta}IP per allele and the odds ratio (per allele) of G–D is OR, we may calculate


    ORK/{Delta}IP
 Top
 Applications of Mendelian...
 ORK/{Delta}IP
 Developing the concept: from...
 The future
 References
 
If this estimation approach is used then it is important that uncertainty in the derived IP–D odds ratio accurately reflects the uncertainties in both the G–IP and G–D estimates.9 Some analyses have ignored the uncertainty in G–IP when calculating CI for the derived odds ratios, which may be very misleading when G–IP is inaccurately assessed.

Youngman and colleagues also used the assumption of Mendelian randomization to study the relationship between plasma apolipoproteins A1 (IP1) and B (IP2) and MI (D) using apolipoprotein E (G1) and cholesteryl ester transfer protein (G2) genotypes as illustrated in Figure 2.3



View larger version (9K):
[in this window]
[in a new window]
 
Figure 2 Modelling the effect of intermediate phenotypes on a complex disease

 
This model has the additional complications of more than one gene affecting each intermediate phenotype (genetic, specifically locus, heterogeneity) and each gene exerting its effect on more than one intermediate phenotype, a special case of pleiotropy. The development of approaches to derive unconfounded estimates of the effect of IP on D in such situations have not yet been fully developed but would be of enormous benefit, given that the phenotypes of greatest interest in public health terms are those that underlie common disorders where such complexity is the norm.10 Further examples of the use of Mendelian randomization can be found in the extensive review by Davey Smith and Ebrahim.4


    Developing the concept: from hypothesis testing to Mendelian deconfounding
 Top
 Applications of Mendelian...
 ORK/{Delta}IP
 Developing the concept: from...
 The future
 References
 
The key research question from a public health perspective is: what is the unconfounded effect of IP on D? Suitably designed genetic studies provide epidemiologists with a tool to derive unconfounded estimates of the size of the effect of the IP on D together with a measure of its uncertainty.9 As with most research involving human subjects, the purpose of such studies will usually be to determine the magnitude of the effect of a causal factor or an intervention aimed at preventing disease.11 In these situations, estimation (of the magnitude of the IP–D association) rather than hypothesis testing (of whether observational epidemiology studies have been subject to confounding) will be of greater utility. There is a substantial literature that stresses the advantages of estimation over hypothesis testing to inform decisions in health-related research.11,12 Clinically important effects may be statistically non-significant if the sample size is inadequate. On the other hand, clinically irrelevant effects may sometimes be statistically significant.

Unconfounded estimates of the IP–D association can readily be adjusted for a realistic reduction in the level of IP that one could expect from a public health intervention aimed at reducing D. Misleading inferences could result, however, if the intervention exerted its effect on D via causal pathways other than IP. This is the case with interventions aimed at reducing MI risk by lowering fibrinogen, which also appear to affect MI risk via other pathways.13–15

In order to produce tight CI for the IP–D odds ratio we need accurate estimates of both G–D and G–IP. Such information will often only be available from meta-analyses. Minelli et al. have described meta-analytical approaches for Mendelian deconfounding, addressing the important issue that G–IP and G–D associations may be correlated when both estimates are obtained from the same study and describing methods to allow for such correlation.16


    The future
 Top
 Applications of Mendelian...
 ORK/{Delta}IP
 Developing the concept: from...
 The future
 References
 
Precise estimates are needed to indirectly estimate the effect of IP on D from Mendelian randomized studies. As almost all current genetic studies are statistically underpowered to detect the relatively small effects of the frequent gene variants that underlie common, complex diseases,17 there has been an increasing emphasis on evidence synthesis and meta-analysis in genetic epidemiology. The Human Genome Epidemiology Network (HuGENet) now co-ordinates a series of reviews that integrate evidence from genetic association studies (http:// www.cdc.gov/genomics/hugenet/default.htm).18 However, only 2 of the 20 reviews published by April 2003 actually employed meta-analysis.19–21 We are optimistic about the possibility of larger studies in the near future because of the substantial reductions in genotyping costs, but study size still remains limited by the cost of proper phenotyping and Mendelian randomization is likely to be based on evidence from meta-analyses for the foreseeable future.

When undertaking meta-analyses of genetic studies to derive unconfounded estimates of IP–D association, researchers will need to be mindful of the limitations of Mendelian randomization studies described by Davey Smith and Ebrahim4 and the limitations of standard meta-analytical problems, such as publication bias.22 Furthermore, reporting and publication bias are so pervasive in genetic association studies17 that especial caution may be needed. Funnel plots should be visually inspected and the sensitivity of the results to methods which ‘adjust’ for the presence of publication bias could be tested.23 In addition, approaches that account for the correlation between G–D and G–IP associations may be required where these estimates are obtained from the same studies.16

The ambitious drive to understand aetiological pathways that underlie the so-called complex diseases, such as asthma and coronary heart disease, has gathered pace with the plethora of biological knowledge and data that have arisen from the human genome project.24 Given that such diseases are common, improved understanding of these pathways will probably be necessary for significant improvements in public health. However, complex diseases are also characterized by their multifactorial nature, an uncertain disease definition, pleiotropy, phenocopies, and genetic heterogeneity.10 In short, individual genetic effects are modest, difficult to detect, and likely to be strongly influenced by environment. The diseases where ‘Mendelian deconfounding’ has the greatest potential also throw up the greatest challenges, since the assumptions that relate to Mendelian randomization are least likely to hold for complex diseases. The most important assumption is the absence of an alternative pathway through which the gene exerts its effect on disease (a special case of pleiotropy), which will affect the validity of a phenotype–disease association derived by from genetic studies. Furthermore, the findings may be less generalizable where there is linkage disequilibrium to, or interaction with, a gene with functional effects on the phenotype and/or disease under study, or gene–environment interaction.

However, given that direct estimates of the effect of an intermediate phenotype on disease in traditional observational studies are highly prone to confounding and reverse causation,4 the derivation of phenotype–disease associations from genetic studies should be considered as a valuable alternative to observational studies and efforts should be directed towards developing methods to appropriately model this complexity. These methods have the potential to be particularly useful in the future as knowledge of biological pathways improves, more suitable polymorphisms can be used, high quality data from large genetic association studies become available, and methods to derive estimates from Mendelian randomised studies are refined.


    Acknowledgments
 
Martin Tobin is funded by a Medical Research Council Clinical Training Fellowship in Health of the Public Research.


    References
 Top
 Applications of Mendelian...
 ORK/{Delta}IP
 Developing the concept: from...
 The future
 References
 
1 Katan MB. Apolipoprotein E isoforms, serum cholesterol, and cancer. Lancet 1986;i:507–08. (Reprinted Int J Epidemiol 2004;33:.)

2 Youngman L, Keavney B, Palmer A et al. Plasma fibrinogen and fibrinogen genotypes in 4685 cases of myocardial infarction and in 6002 controls: test of causality by ‘Mendelian randomisation’. Circulation 2000;102(Suppl.II):31–32.

3 Keavney B, Youngman L, Palmer A et al. Large-scale test of hypothesised associations between polymorphisms of lipid-related genes and myocardial infarction in about 5000 cases and 6000 controls. Circulation 2000;102(Suppl.II):852.[Abstract/Free Full Text]

4 Davey Smith G, Ebrahim S. ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol 2003;32:1–22.[CrossRef][ISI][Medline]

5 Verhoef P, Klerk M, Katan MB. Use of ‘Mendelian randomisation’ for testing causality. J Inherit Metab Dis 2003;26(Suppl.1):39.

6 Gray R, Wheatley K. How to avoid bias when comparing bone marrow transplantation with chemotherapy. Bone Marrow Transplant 1991;7(Suppl.3):9–12.[ISI][Medline]

7 Wheatley K, Gray R. Commentary: Mendelian randomization—an update on its use to evaluate allogenic stem cell transplantation in leukaemia. Int J Epidemiol 2004;33:15–17.[Free Full Text]

8 Wijsman EM. Mendel's Laws. In: Elston R, Olson JM, Palmer LJ (eds). Biostatistical Genetics and Genetic Epidemiology. Chichester: Wiley, 2002, pp. 527–29.

9 Thompson JR, Tobin MD, Minelli C. On the Accuracy of Estimates of Phenotype on Disease Derived from Mendelian Randomised Studies. Technical report 2003/GE1. Leicester: University of Leicester, 2003 (http://www.prw.le.ac.uk/research/HCG/getechrep.html).

10 Palmer LJ. Complex Diseases. In: Elston R, Olson JM, Palmer LJ (eds). Biostatistical Genetics and Genetic Epidemiology. Chichester: Wiley, 2002, pp. 206–17.

11 Gardner MJ, Altman DG. Confidence intervals rather than P values: estimation rather than hypothesis testing. BMJ (Clin Res Ed) 1986;292:746–50.[ISI][Medline]

12 Rothman KJ. A show of confidence. N Engl J Med 1978;299: 1362–63.[ISI][Medline]

13 Rosenson RS, Tangney CC. Antiatherothrombotic properties of statins: implications for cardiovascular event reduction. [comment] JAMA 1998;279:1643–50.[Abstract/Free Full Text]

14 Danesh J, Collins R, Appleby P, Peto R. Association of fibrinogen, C-reactive protein, albumin, or leukocyte count with coronary heart disease: meta-analyses of prospective studies. JAMA 1998;279:1477–82.[Abstract/Free Full Text]

15 Lowe GD. Why do smokers have higher plasma fibrinogen levels than non-smokers? Clin Sci (Lond) 2001;101:209–10.[CrossRef][Medline]

16 Minelli C, Thompson JR, Tobin MD. Meta-Analytical Methods for the Synthesis of Genetic Studies Using Mendelian Randomisation. Technical report 2003/GE2. Leicester: University of Leicester, 2003 (http://www.prw.le.ac.uk/research/HCG/getechrep.html).

17 Cardon LR, Bell JI. Association study designs for complex diseases. Nat Rev Genet 2001;2:91–99.[CrossRef][ISI][Medline]

18 Khoury MJ, Little J. Human genome epidemiologic reviews: the beginning of something HuGE. Am J Epidemiol 2000;151:2–3.[ISI][Medline]

19 Little J, Khoury MJ, Bradley L et al. The human genome project is complete. How do we develop a handle for the pump? Am J Epidemiol 2003;157:667–73.[Free Full Text]

20 Botto LD, Yang Q. 5,10-Methylenetetrahydrofolate reductase gene variants and congenital anomalies: a HuGE review. Am J Epidemiol 2000;151:862–77.[Abstract]

21 Engel LS, Taioli E, Pfeiffer R et al. Pooled analysis and meta-analysis of glutathione S-transferase M1 and bladder cancer: a HuGE review. Am J Epidemiol 2002;156:95–109.[Abstract/Free Full Text]

22 Sterne JAC, Egger M, Davey Smith G. Investigating and dealing with publication and other biases. In: Egger M, Davey Smith G, Altman DG (eds). Systematic Reviews in Health Care: Meta-analysis in Context. 2nd Edn. London: BMJ Books, 2001, pp. 189–211.

23 Sutton AJ, Abrams KR, Jones DR. Generalized synthesis of evidence and the threat of dissemination bias. The example of electronic fetal heart rate monitoring (EFM). J Clin Epidemiol 2002;55:1013–24.[CrossRef][ISI][Medline]

24 Burton P, Tobin MD. Epidemiology and Genetic Epidemiology. In: Balding DJ, Bishop M, Cannings C (eds). Handbook of Statistical Genetics. 2nd Edn. Chichester: Wiley, 2003.