Causal inference in health and disease: a review of the principles and applications of Mendelian randomization

Abstract

Mendelian randomization (MR) is a genetic epidemiological technique that uses genetic variation to infer causal relationships between modifiable exposures and outcome variables. Conventional observational epidemiological studies are subject to bias from a range of sources; MR analyses can offer an advantage in that they are less prone to bias as they use genetic variants inherited at conception as "instrumental variables", which are proxies of an exposure. However, as with all research tools, MR studies must be carefully designed to yield valuable insights into causal relationships between exposures and outcomes, and to avoid biased or misleading results that undermine the validity of the causal inferences drawn from the study. In this review, we outline Mendel's laws of inheritance, the assumptions and principles that underlie MR, MR study designs and methods, and how MR analyses can be applied and reported. Using the example of serum phosphate concentrations on liability to kidney stone disease we illustrate how MR estimates may be visualized and, finally, we contextualize MR in bone and mineral research including exemplifying how this technique could be employed to inform clinical studies and future guidelines concerning BMD and fracture risk. This review provides a framework to enhance understanding of how MR may be used to triangulate evidence and progress research in bone and mineral metabolism as we strive to infer causal effects in health and disease.

Keywords: epidemiology; genetic association studies; human association studies; mendelian randomization; therapeutics.

Figure 1

Pitfalls in observational studies. Confounding variable affecting exposure and outcome: measured confounders can be accounted for in multivariable analyses; unmeasured confounders can cause unreliable results. Residual confounding: confounding that persists despite adjusting for measured confounders. Reverse causality: possible effect of outcome on exposure, often neglected. Regression dilution bias: random errors in measurement causing bias in estimated association with outcome.

Figure 2

Analogy of an RCT and MR. MR leverages naturally occurring genetic variations as “instrumental variables” to mimic the random assignment of exposures, like random allocation in an RCT. This genetic randomization helps address confounding factors and strengthens causal inference when investigating the relationships between exposures, outcomes, and potential interventions.

Figure 3

Conditional assumptions in MR. (A) Relevance: The genetic variant(s) acting as a proxy for the exposure variable are strongly associated with the exposure. (B) Independence: The IV is independent of confounding factors, thus genetic association with confounding variables is random, akin to randomization in an RCT. (C) Exclusion restriction: The IV affects the outcome solely through the exposure variable and not by external, horizontally-pleiotropic, pathways.

Figure 4

Mediation MR. Mediation analysis estimates the effect of an exposure on an outcome via mediating factors (indirect mediating effect) and via mechanisms independent of the mediating factor (direct casual effect). The proportion mediated effect can subsequently be derived.

Figure 5

Visual representations of MR analyses of the effect of albumin-adjusted serum calcium concentrations on risk of kidney stone disease- reanalysis of data from Lovegrove et al. MR, Mendelian randomization; SNP, single nucleotide polymorphism. (A) Scatter plot of genetic associations and causal estimates. (B) Forest plot of variant-specific causal estimates, IVW, and MR-Egger estimates. (C) Funnel plot of variant-specific causal estimates. The estimate effect is plotted against the precision (reciprocal of standard error) of the estimate for each genetic variant. The vertical lines correspond to the estimates for IVW and MR-Egger estimates. (D) Forest plot of leave one out analyses. For each variant, the effect of removing it from the IV is plotted. (E) Radial plot. The x-axis represents the weigh attributed to each genetic variant and the y-axis represents its Z-statistic (point estimate divided by standard error). The greater the vertical distance between a genetic variant and the overall causal estimate, the greater its contribution to Cochran’s Q statistic of heterogeneity.

Figure 6

Assessing assumptions in MR. (A) Assumption in MR: MR assumes that there is an IV comprised of genetic variant(s) which are a proxy of the effect of an exposure variable in an outcome measure. (B) The F-statistic can evaluate the strength of association between the IV and exposure variable. (C) Robust MR methods, for example MR-Egger regression, can identify when genetic variants have pleiotropic effects and produce a reliable effect estimate so long as the magnitude of pleiotropic effects is not correlated with the variant-exposure effects. (D) Steiger filtering selects genetic variants that explain more variance in the exposure variable than the outcome and excludes those that fail to meet this condition to mitigate against mis-specification of the primary exposure.