Use of Mendelian Randomization to assess the causal status of modifiable exposures for rheumatic diseases, 2024, Zhao et al

[forestglip]

Use of Mendelian Randomization to assess the causal status of modifiable exposures for rheumatic diseases

Sizheng Steven Zhao, Stephen Burgess

[Line breaks added]

Abstract
The explosion in Mendelian randomization (MR) publications is hard to ignore and shows no signs of slowing. Clinician readers, who may not be familiar with jargon-ridden methods, are expected to discern the good from the many low-quality studies that make overconfident claims of causality or stretch the plausibility of what MR can investigate.

We aim to equip readers with foundational concepts, contextualized using examples in rheumatology, to appraise the many MR papers that are or will appear in their journals. We highlight the importance of assessing whether exposures are under plausibly specific genetic influence, whether the hypothesized causal pathways make biological sense, and whether results stand up to replication and use of control outcomes.

Quality of research can vary substantially using MR as with any design, and all methods have inherent limitations. MR studies have provided and can still contribute valuable insights in the context of evidence triangulation.

Link | PDF (Best Practice & Research Clinical Rheumatology) [Open Access]

 

[forestglip]

This is an accessible overview of what limitations to be aware of with MR studies.

One of the main things they stress is that there should be a plausible connection between the genetic variant and the exposure. Ideally there is some direct mechanistic connection between the gene and the exposure of interest. Using genes that were obtained from a GWAS that are simply associated with an exposure have much more potential to violate the assumptions of MR, even when the typical statistical tests indicate no violations.

The first question a reader should ask is: Can genetic variation plausibly influence the exposure in ways analogous to a proposed intervention? MR only makes sense when there is gene-environment equivalence. For example, genetic variation in the SOST gene that encodes sclerostin leads to low or absent sclerostin levels and high bone mass [7], while therapeutic sclerostin inhibition using romosozumab increases bone mass. There is sufficient analogy between genetic and therapeutic perturbation in sclerostin to allow inferences about drug efficacy and safety to be drawn from genetic associations in this gene region. If genetically predicted sclerostin function is associated with adverse cardiovascular profile, then there is genetic support for the concern over cardiovascular risk when using romosozumab [8,9].

Gene-environment equivalence is an important sense checking step that should first take place, before appraising whether any statistical gymnastics can provide meaningful findings. As shown in the next section, authors can claim various assurances for the instrumental variable assumptions even when it is implausible that genetic variants mimic interventions on the exposure. When biologically plausible genetic variants are thoughtfully selected as instrumental variables, the MR analysis is more likely to fulfil key assumptions required for valid causal inference.

It seems, at first glance, reasonable to identify genetic variants strongly associated with use versus non-use of statins to study the effect of statins on IPF risk using MR; specifically, one could select uncorrelated variants associated with statin-use across the genome that reach genome-wide significance.

Readers may believe all three assumptions ‘satisfied’ by a two-sample MR paper that includes large F statistics, ancestrally homogenous populations, and ‘pleiotropy-robust’ methods such as MR-Egger that demonstrate no statistical evidence of directional pleiotropy, and conclude that statin-use affects IPF risk.

However, this analysis should not pass the first sense check, since it is implausible that whether an individual chooses to take a statin is under specific genetic influence. Liability to taking a statin is influenced by having any one of multiple cardiovascular risk factors, clinician/patient preferences, compliance, tolerance, and so forth; genetic variants that associate with increased statin usage may influence any of these traits. Such variants will likely influence multiple traits that violate the third assumption, yet show no statistical evidence of horizontal pleiotropy in sensitivity analyses [16]. Indeed, genetic variants in the PCSK9 gene region that are associated with increased statin usage, are associated with increased levels of LDL-cholesterol, not reduced levels. This is because the increased statin usage is in response to raised cholesterol levels, and not because statins raise cholesterol [17].

When MR was first proposed two decades ago, genetic variants with known biological function were used to mimic or proxy perturbations in the exposure. By contrast, many MR studies of non-molecular traits apply biology-agnostic selection of genetic variants (i.e., independent, genome-wide significant SNPs) that have no clear mechanistic role. The risk of including variants for alternate causal paths increases as GWAS of the exposure increases in size. For example, larger GWASs of GORD [gastro-oesophageal reflux disease] are likely to pick up more genetic variants that have direct effects on upstream risk factors that may influence IPF [idiopathic pulmonary fibrosis] susceptibility via pathways independent of GORD. Readers should be as cautious about confounding and overconfident causal claims in MR studies as they are with any other observational design. Combining insights from methods that make different assumptions (evidence triangulation) is essential for robust conclusions.

[Edit: changed wording] They say it is good to check if researchers used the MR-STROBE checklist.

However, it is important to note that authors can present readers with superficially impressive looking manuscripts including statistical assurances for all three assumptions even when the analysis is conceptually problematic. The MR-STROBE checklist [12] was devised to improve the standard of reporting in MR publications. It forces researchers to justify the plausibility of hypotheses, which should reduce more obvious examples of problematic exposures such as estimated exposure to air pollution [4], occupation [13], or specific dietary intakes [5,14]. However, there are seemingly sensible exposures that are nevertheless unsuitable for MR analyses.