Statistical hocus pocus

Recent article

A recent paper entitled “Causal Associations Between Socioeconomic Status, Intelligence, Cognition and Atrial Fibrillation - Evidence From Mendelian Randomization”(Kaisaier et al., n.d.) claims that the link between educational attainment and atrial fibrillation (AF) is causal and largely mediated through cardiometabolic traits like heart failure, BMI, and coronary artery disease. The authors reach these conclusions by applying a two-step Mendelian randomization (MR) approach and conclude, for example, that heart failure mediates 95% of the effect of education on AF. I believe the paper is seriously flawed from scientific, statistical and sociological perspectives.

While I am not a geneticist, I do have a background in statistics and epidemiology and have been impressed with Mendelian randomization (MR) studies. For example, MR was able to sort out the causal, or rather non causal, role of C-reactive protein (CRP) in the etiology of coronary artery disease(Collaboration et al. 2011). I must also preface my remarks by underscoring that it is infinitely easier to criticize a study than to perform it.

My main concern with this paper is that the authors appear to be using a Mendelian randomization approach to study a causal relationship that is not biologically well defined. They claim to be studying the causal relationship between socioeconomic status, intelligence, cognition and atrial fibrillation, but the definitions of these terms and in particular their genetic determinants lack clarity making it difficult to interpret their results and to understand the implications of their findings.

Socio-economic status (SES) is not a biological exposure — it is a complex, socially constructed phenomenon and using “years of education” or “income” as proxies ignores many of the complex associated social, cultural, and structural intricacies of SES. Ignoring these complexities can lead to oversimplified conclusions that do not accurately reflect the multifaceted nature of SES.

Mendelian randomization

MR proposes genetic variants as instrumental variables to infer causal relationships between exposures (like education) and outcomes (like atrial fibrillation). The idea is that if a genetic variant is associated with an exposure, and that exposure is causally related to an outcome, then the genetic variant should also be associated with the outcome.

In MR studies of education, polygenic scores are derived from GWAS of years-of-schooling often capturing tiny, noisy associations across thousands of SNPs. moreover these associations may be deeply confounded by environmental stratification: access to schooling, neighborhood quality, and social class. Using such instruments to make causal claims about how “genetically instrumented education” reduces atrial fibrillation implicitly reinforces the false view that educational differences are genetically fixed.

What is a good instrument?

• Something that is correlated with the exposure Relevance
  
• Something that does not directly cause the outcome, other than by the treatment Exclusion, or ignorability
  
• Something that is not correlated with the omitted variables Exogenity
  
  

Generic IV DAG

Z = instrument, X = exposure, Y = outcome, U = unmeasured (omitted) variable Relevance Correlated with exposure
Z → X   Cor(Z, X) ≠ 0 testable with stats

Excludability Correlated with outcome only through exposure
Z → X → Y   Z ↛ Y   Cor(Z, Y | X) = 0 testable with stats + story

Exogeneity Not correlated with omitted variables
U ↛ Z   Cor(Z, U) = 0
requires story, no stats

Essentially trades one implausible assumption (ignorability of treatment variable) by a more plausible assumption (ignorability of the instrument)

Mendelian randomization assumptions

Specifically, in the context of MR, the instrument (the genetic variants) must meet three key criteria:
1. Be strongly associated with the exposure (e.g., education)
2. Not be associated with confounders
3. Influence the outcome (AF) only through the exposure (no horizontal pleiotropy).

This paper makes the very strong assumption that education-associated SNPs meet these criteria — yet we know that pleiotropy is common, and gene–environment interactions complicate interpretation.

There is no proof for the exclusion (i.e., no alternate causal path from SNP to AF) and exogeneity (i.e., no confounding between SNP and AF) assumptions and a more likely DAG for this paper could be as follows:

Summary of DAG Paths in MR Analysis
Types of paths and assumption violations
Path	Type	Violation
Z → Education	Valid IV path	No
Z → AF	Pleiotropy	Yes (exclusion restriction)
G2 → Mediator → AF	Possibly pleiotropy	Possibly
G2 → AF	Pleiotropy	Yes (exclusion restriction)
U → Education	Confounding	Yes (independence assumption)
U → AF	Confounding	Yes

Other limitations

Another core issue is transparency what genetic variants were used as instruments for education or for any of the mediators is not disclosed. Rather the paper cites source GWAS studies. The authors do not provide the specific genetic markers or SNPs associated with education, income or cognitive function, making it difficult to assess the reproducibility and validity of their genetic-centric SES instrument. Moreover, GWAS hits for both educational attainment and the various mediators are known to be highly pleiotropic, often associated with broader behavioral and metabolic traits undermining the above excludability assumption.

These domains are mostly influenced by social environment and educational opportunity, not inherent genetic ability alone. Intelligence and education are bidirectionally related, and disentangling cause/effect even in MR frameworks is nontrivial, if not impossible. Even if a perfect genetic marker existed for education attainment, equating that with SES is problematic. SES is a complex construct influenced by many factors beyond education, including among other factors early-life adversity, wealth, racism, neighborhood environment and social capital. Using a single genetic marker to represent such a multifaceted concept oversimplifies the issue and risks gross misinterpretations.

Framing education and cognition as genetic risk factors for cardiovascular disease raises ethical concerns, particularly if interpreted as genetic determinism of social disadvantage. Such conclusions may inadvertently reinforce inequities by falsely suggesting that socially remediable conditions are immutable genetic traits. This risks stigmatizing individuals with lower educational attainment or cognitive function as “genetically inferior,” which is scientifically and ethically problematic. The authors’ conclusions could be misinterpreted as suggesting that educational disparities are biologically predetermined, undermining the importance of addressing structural inequalities in education and health.

The eight potential mediators are potentially correlated, are not independent, yet they seem to have been consider in isolation as witness by the abstract that report a “mediation proportion” > 300%. The two-step MR mediation assumes that the mediator and outcome are conditionally independent given the exposure (education). which is highly doubtful with correlated mediators such as BMI, waist circumference, body fat mass, and CHD. By performing univariate analyses, the authors ignore mediator interactions and collinearity, which can distort both the magnitude and attribution of mediation effects.

The mediation analysis leads to circularity reasoning in interpreting effects. Saying that higher education → lower AF via obesity and heart failure just rephrases long-known epidemiological associations without adding actionable insight as the causal claim rests on ill-defined genetic instruments that don’t meaningfully map to policy or intervention.

The authors state they have followed MR reporting standards (e.g., STROBE-MR), but this is not the case. The paper does not provide sufficient details on the genetic instruments used, such as specific SNPs or their associations with educational attainment. There is no mention of F-statistics or variance explained by the instruments, which are critical for assessing instrument strength. The methods for clumping and harmonization of SNPs are not described, making it impossible to evaluate the validity of the MR analysis.

There are additional statistical issues with the analysis, or at least with its reporting. The authors present mediation proportions (e.g., “heart failure explains 95% of the effect”) with apparent numerical precision — a case of false certainty dressed in the complex causal inference tools of Mendelian Randomization. They mention using MR-PRESSO to detect and correct for horizontal pleiotropy, but do not provide any details on the number of outliers removed, the distortion test results, or any before-and-after comparison. This lack of transparency makes it difficult to assess the robustness of their findings. They also their data synthesis as a “meta-analysis” but offer no information about whether a fixed or random-effects model was used, nor do they address between-study heterogeneity.

Conclusion

This study presents an elegant and complex statistical exercise, but the study overextends Mendelian randomization into domains — SES, cognition, education — where the assumptions of genetic instrumental variables are least likely to hold. In doing so, it risks lending a veneer of causal genetic certainty to deeply contingent, context-dependent social processes. MR here feels more like a tool in search of a question, rather than answering a coherent biological hypothesis.

A more modest interpretation of their findings would emphasize that educational attainment is consistently associated with AF risk — a result known from previous observational studies — and that addressing structural inequities in education remains a key intervention target, regardless of genetic predisposition.

The danger with this MR overreach is real and palpable, whereby complex social phenomena are force-fit into a biological framework, implying genetic determinism of social disadvantage, which is ethically and scientifically dangerous. The authors’ conclusions, while statistically sophisticated, risk reinforcing harmful stereotypes about intelligence and socioeconomic status as fixed biological traits rather than socially constructed phenomena shaped by a myriad of factors. This reflects a fundamental misunderstanding of SES as a biological trait and not the social and structural determinant that it is more likely to be. Educational attainment is also influenced by a host of environmental and policy factors, many of which are population-specific and cannot be reduced to genotypes.

This genetic determinism suggests that people are “programmed” by their DNA and promotes the fallacy that social inequalities have genetic causes, which is harmful and scientifically incoherent falsely implying implies inevitability and immutability. By treating SES as a biological trait, we risk reviving discredited eugenic tropes and obscuring the social and policy levers that truly shape education and health.

References

Collaboration, C. Reactive Protein Coronary Heart Disease Genetics, F. Wensley, P. Gao, S. Burgess, S. Kaptoge, E. Di Angelantonio, T. Shah, et al. 2011. “Association Between c Reactive Protein and Coronary Heart Disease: Mendelian Randomisation Analysis Based on Individual Participant Data.” Journal Article. BMJ 342: d548. https://doi.org/10.1136/bmj.d548.

Kaisaier, Wulamiding, Min Ye, Zexuan Wu, Chen Liu, Jiangui He, Gregory Y. H. Lip, Yili Chen, and Wengen Zhu. n.d. “Causal Associations Between Socioeconomic Status, Intelligence, Cognition and Atrial Fibrillation.” Journal Article. JACC: Asia 0 (0). https://doi.org/doi:10.1016/j.jacasi.2025.04.008.