Dr. Rachel Shulman @shulman_rachel obtained her medical degree at the University of Pennsylvania where she is currently a clinical and research fellow pursuing a Master of Science in Clinical Epidemiology. Her published research has focused on the epidemiology of essential hypertension, with particular reference to its complex relation to sleep patterns and obesity. Currently she is pursuing her interest in the pathophysiology and pharmacoepidemiology of non-proteinuric CKD with the hope of identifying targeted treatment strategies to prevent its progression. She has been a tutor and lecturer for medical students and residents and hopes to contribute to educational programs for both professionals and the public on vital issues of public health. Dr Shulman is a 2023-24 AJKD Editorial Intern.

In a study recently published in AJKD, Shan et al addressed the mortality risk of kidney insufficiency is a kind of showcase for the causal inference method of Mendelian randomization. This paper evaluated the causal link between reduced estimated glomerular filtration rate (eGFR) and mortality using individual data from nearly half a million individuals from the UK Biobank. Using mendelian randomization, Shan et al. found a casual association between low eGFR and cardiac mortality but, notably, not all-cause mortality.

Mendelian randomization is an advanced epidemiologic technique for observational studies which attempts to distinguish causality from correlation. Why is mendelian randomization better than traditional observational approaches? This approach has garnered much enthusiasm in the last 20 years, providing the statistical method for hundreds of research studies. Mendelian randomization relies on the genetic variation to ‘randomize’ individuals to understand the causality between a given risk factor and outcomes in observational data (Figure 1). Long employed in econometrics under the name “instrumental variable analysis,” mendelian randomization was pressed upon medical research as an antidote to the failures of observational studies. Such studies suffered significant reputational damage when it became clear that they produced either implausible results or associations which vanished in randomized controlled trials (RCTs).

Figure 1. Comparison of Randomized Controlled Trial, Mendelian Randomization Analysis, and Traditional Observational Trials. © Shulman

The problems with observational studies can be traced to at least two flaws which undermine the claim that they are able to identify causal relationships. First, there is the confounding role of measured and unmeasured covariates. If coffee consumption is high among college professors, we might imagine an observational study “proving” that caffeine causes IQs to rise. Second, observational studies ignore the epistemological imperative that a cause cannot precede an effect. An observational study showing that hypertension is associated with kidney disease could not be used to prove that kidney function is preserved if hypertension is eliminated. The true causal relationship may be the other way round. “Reverse causality” errors of this sort led to great concern when risk factors shown by observational studies to cause disease were therapeutically manipulated with no clinical benefit.

Shan et al. have drawn upon this background. They suggest that often-replicated observational results pointing to the lethal consequences of a decreased eGFR may represent a failure of the observational method. Consequently, they perform their study twice, comparing the observational results with those obtained by mendelian randomization. The name mendelian randomization  reveals the heart of the method, which is simply an application of Mendel’s law of independent assortment. Localized genetic variants–i.e., single nucleotide polymorphisms (SNPs) associated with eGFR–are tested for their association with death from various causes. The selected SNPs determine only a single phenotype (eGFR) that exercises no effect on traits and behaviors which would be typically treated as confounders. It is as if nature has performed its own RCT, relying on a wisely selected nucleotide to introduce a trait of interest (a low eGFR) in a population for which every covariate imaginable remains randomly distributed.

Broadly speaking, Shan et al. report that when both epidemiological methods are employed, the primary conclusion is unchanged: a low eGFR elevates the risk of death. A disparity arose only in the nature of the risk. The conventional observational approach found that the risk of all-cause death increased in exponential (ie, non-linear) fashion after eGFR fell below 90 ml/min/1.73 m². The mendelian randomization  analysis, in contrast, identified a linear increase in mortality risk with any fall in eGFR, but limited to cardiovascular deaths only. The authors place “greater confidence” in the mendelian randomization  results because they were derived from a method designed to reveal causal relationships. Still, it may be instructive to ask if the “greater confidence” expressed by the authors in mendelian randomization  can be challenged on the methodological grounds where mendelian randomization is, in fact, held to be clearly superior.

There is first of all the fundamental fact that mendelian randomization, like any study design, is built upon assumptions: 1) if a SNP is to serve as a proxy for a phenotypic risk factor, it cannot affect more than one phenotype or be linked to another SNP which does so, (2) it is desirable that the link between the genetic variant and its phenotype involve no intervening variables, which might also act as confounders; 3) finally, the variant cannot act on the outcome measure directly because to do so would reverse the role of cause-and-effect. The genetic variant is a proxy for an exposure. The variant can be treated as if it were a causal agent, but only if its influence on the outcome (death) is exercised exclusively via its influence on the exposure (eGFR).

How, then, is one to know in a particular instance if the above assumptions are justified? Shan et al. address this issue appropriately, applying both standard and advanced statistical methods which examine the effect on their results when the conditions for a causal relationship are not quite met. No major violations of the assumptions emerged, and their findings were therefore considered robust. Still, it is fair to say that mendelian randomization has its critics, and some of them have doubted whether one can ever rely on statistical methods to remove the effect of confounders one has not measured, or even thought of. Perhaps less controversial is the claim that our highest level of confidence should be reserved for negative mendelian randomization results. Such results serve an invaluable service of their own by constraining the launch of costly, time-consuming, or unethical RCTs destined to achieve nothing.

Another concern is the question of biological plausibility. Insights into molecular biology have encouraged an insistence that cause-and-effect relationships are proven only by the demonstration of mechanisms operating at the molecular level. Clearly, Shan et al. cannot use mendelian randomization to point to a mechanism that explains why persons with “mildly” impaired kidney function are at risk for cardiovascular death. Mechanistic explanations are appealing because they suggest targets for therapeutic intervention, but their theoretical significance rests with the restriction exclusion principle. What appears in the mendelian randomization analysis as an association between an eGFR and cardiovascular death might actually arise from a genetic variant that affects both kidney and cardiac function via an unknown pathway. This would violate the assumption that our genetic variant affects outcomes only via eGFR. An even more subtle question for clinicians is captured by the principle of gene-environment equivalence. This equivalence requires that a low eGFR associated with a genetic disposition present from birth is indistinguishable from a low eGFR identified in an unselected adult visitor to our clinic.

Finally, one must consider the actual advance over the classical observational literature: contrary to prior evidence, a reduced eGFR was found to carry a risk of cardiovascular death which did not generalize to all-cause death. No patients with advanced kidney failure were included in the study, an important limitation of the results, but the outcome “cardiovascular death” itself raises a measurement problem unrelated to analytic approach. A “cause” of death is not a measurement at all in the strict sense (Shan et al. call a “notoriously challenging” source of difficulty). It is instead a byproduct of physician behavior, as recent controversy over the true prevalence of “COVID-19-related” deaths makes clear. Thus, without prejudice to Shan et al.’s scrupulously executed study, it is worth remembering that no statistical analysis, however sophisticated, can escape the possibility of error altogether. Even our most advanced epidemiological models must derive their value to medical practice through compromise with the complexity of events on the human stage.

– Post prepared by Rachel Shulman @shulman_rachel

To view Shan et al (Open Access), please visit AJKD.org.
Title: Kidney Function Measures and Mortality: A Mendelian Randomization Study
Authors: Ying Shan, Jingwen Zhang, Yueqi Lu, Jinlan Liao, Yuyang Liu, Liang Dai, Jing Li, Congying Song, Guobin Su, Sara Hägg, Zuying Xiong, Dorothea Nitsch, Juan Jesus Carrero, and Xiaoyan Huang
DOI: 10.1053/j.ajkd.2023.10.014