A cautionary note on the use of Mendelian randomization to infer causation in observational epidemiology

The concept of Mendelian randomization when used in the context of association studies refers to the random allocation of alleles at the time of gamete formation. This concept has received a lot of attention in recent years in observational epidemiology for its potential to reduce residual confounding and to protect against reverse causation.1 In brief, the random segregation of alleles at the time of gamete formation is a natural experiment that occurs before the outcome of interest, which protects against falsely reversing the inference about a cause-effect relationship. If a particular genetic variant is strongly associated with a risk factor for the outcome of interest, the association between this genetic variant and the outcome of interest may be used to infer causality. Table 1 indicates the commonly acknowledged necessary conditions (1–9) for Mendelian randomization to provide this protection in observational epidemiology.2,3 We list additional necessary conditions (10–12) that have been given little attention and may be of relevance for certain genetic variants. Necessary conditions for the use of Mendelian randomization to infer causality in observational epidemiology Adapted and expanded from Nitsch et al.2 and Davey Smith and Ebrahim3. Necessary conditions for the use of Mendelian randomization to infer causality in observational epidemiology Adapted and expanded from Nitsch et al.2 and Davey Smith and Ebrahim3. The first of these additional conditions is the absence of transmission ratio distortion (TRD). TRD occurs when the distribution of alleles at a particular locus differs in the surviving offspring from that expected on the basis of Mendelian proportions. TRD can occur either during or after meiosis. Segregation distortion or meiotic drive refers to the phenomenon of non-random assignment of alleles at the time of gamete formation and therefore describes TRD during meiosis.4 After meiosis, TRD may result from selective survival between conception and birth or later. In humans, a few examples of TRD have been reported.5–10 In particular, a recent report suggests that transmission distortion may be widespread in the human genome.6 Selective survival due to the genetic variant of interest is acknowledged as a potential source of bias analogous to differential loss of follow up in randomized controlled trials.2 With that perspective, selective survival may occur not only between conception and birth, but also between birth and entry into the study if the genetic variant of interest causes early mortality. The second additional condition is the absence of parent-of-origin effect. A parent-of-origin effect occurs whenever the effect of an allele on a phenotype of interest depends on whether the allele was passed on to the child from the father or from the mother. Some genes may be silent or functionally active depending on whether a particular variant was maternally or paternally inherited. This effect is mediated through an epigenetic modification called imprinting. As a consequence, a parent-of-origin effect implies that the effect on the phenotype conferred by a specific genetic variant is not homogeneous in the population. Segregation distortion and parent-of-origin effect have, to our knowledge, not been previously described as potential limitations to the use of Mendelian randomization for inferring causality in observational epidemiology. We consider here data on the relationship between homocysteine blood level and stroke to illustrate how these limitations may jeopardize the use of Mendelian randomization to infer causation. Homocysteine blood level is associated with increased risk of stroke11 and is determined, in part, by the activity of methylenetetrahydrofolate reductase (MTHFR). The C677T MTHFR variant is associated with both the risk of stroke and homocysteine blood level.12 The risk of stroke conferred by the C677T MTHFR variant is consistent with the risk expected from the increased homocysteine blood level associated with the C677T MTHFR variant.12 These findings strongly suggest a causal relationship between homocysteine and stroke. We now consider to what extent the earlier-mentioned conditions may not have been met for this particular example. Infante-Rivard and Weinberg13 assessed TRD for variants in five genes associated with thrombophilic disorders, including the MTHFR gene. They found some evidence of TRD for the MTHFR gene, which may be due to either segregation distortion at the time of gamete formation or due to post-meiotic events occurring between conception and birth. Furthermore, fetal C677T MTHFR status has been associated with fetal viability, i.e. selective survival between gamete formation and birth.14,15 Maternal C677T MTHFR status may influence pregnancy outcome. The C677T status of the mother has been associated with idiopathic recurrent pregnancy losses in some studies16–18 but not in others.19–22 Maternal hyperhomocysteinemia could be a risk factor for recurrent early pregnancy loss.23 Together, these observations may contribute to TRD at this particular locus and so invalidate the assumption of random segregation of alleles that underlies Mendelian randomization. Finally, although there is currently no evidence for a parent-of-origin effect, few studies have evaluated genomic imprinting at the C677T MTHFR locus13 so far. Due to the growing evidence for the role of imprinted genes in common complex diseases,24–28 assuming that parent-of-origin effects are absent may be disputable in many instances. Hence, for this specific example, not only may the process of randomization itself be flawed, but substantial selective survival may also occur between the time of randomization (i.e. gamete formation) and the time of birth. As a consequence, the genotype distribution at birth may differ from the distribution at the time of randomization. The influence of maternal C677T genotype and maternal homocysteine levels on the fetal genotype distribution may represent additional sources of bias. Currently, little is known about the mechanisms that distort the distribution of alleles from that expected from Mendel's first law between gamete formation, conception, birth and entry into a study. It is therefore difficult to estimate the extent to which these mechanisms may affect the results of Mendelian randomization studies. However, these mechanisms should not be assumed to be negligible until they have been properly evaluated. Potential for segregation distortion, selective survival and parent-of-origin effect should be kept in mind when conducting Mendelian randomization studies to draw causal inference in observational studies.