Mendelian Randomisation in Longevity Research: Assumptions and Limitations

What the method can establish

Mendelian randomisation (MR) uses genetic variants associated with an exposure as instrumental variables. Researchers compare variant–exposure and variant–outcome associations to investigate whether the exposure influences the outcome. Inheritance precedes later disease, reducing some problems of reverse causation, but genetic allocation does not make an observational sample equivalent to a randomised trial.

Human MR provides evidence about genetically influenced exposure differences. Animal lifespan experiments and laboratory studies provide separate evidence about mechanisms; neither establishes that the corresponding intervention extends human life. See the guide to reading MR studies.

Three core instrumental-variable assumptions
Assumption	Meaning	Potential failure
Relevance	Variants predict the exposure.	Weak associations make estimates imprecise and potentially biased.
Independence	Variants are independent of exposure–outcome confounders.	Population structure, family effects or selection can create associations with confounders.
Exclusion restriction	Variants affect the outcome only through the exposure.	Horizontal pleiotropy creates an alternative pathway to the outcome.

Robustness checks have their own assumptions

Horizontal pleiotropy differs from vertical pleiotropy: downstream consequences of the exposure can belong to the causal pathway being estimated. Associations with several traits therefore require biological interpretation.

Compare estimates across suitable methods, inspect heterogeneity and influential variants, and report instrument strength. MR-Egger and weighted-median analyses tolerate different patterns of invalid instruments; agreement does not prove validity, and a non-significant pleiotropy test cannot exclude bias.

Two-sample MR also requires compatible populations, exposure definitions and allele alignment. Participant overlap can change weak-instrument bias. For molecular traits, colocalisation can help assess whether nearby association signals share a causal variant, but does not independently prove mediation. These checks are discussed in the MR investigation guidelines.

Survival to recruitment can distort associations

Older participants must survive long enough to enrol. If both the exposure and other outcome determinants affect survival, restricting analysis to survivors can associate the genetic instrument with those other determinants. This selection can undermine independence even when inheritance was initially unconfounded.

A simulation study showed that survival selection could weaken instruments and distort MR estimates, with bias depending on recruitment age and study design. These are methodological simulation findings, not measured treatment effects in humans or animals. The direction and magnitude should not be assumed identical in every longevity dataset. See Smit and colleagues on survival bias.

When interpreting an older cohort, ask who died before recruitment, who participated and whether the analysis addresses selection. Merely adjusting for age does not establish that survival bias has been removed.

A genetic estimate is not a treatment forecast

Genetic differences may influence exposure over many years. A drug started late in life may differ in timing, dose, tissue effects and unintended actions. An MR association therefore does not specify a treatment regimen or establish its safety.

For longevity interpretation, distinguish survival to an age threshold, lifespan, disease incidence and functional healthspan. Likewise, an association with an ageing biomarker does not automatically establish an effect on survival. This is an application of the guidelines’ emphasis on defining the causal question and interpreting the measured outcome. See the MR investigation guidelines.

Questions to bring to a paper

What exposure, population and outcome does the estimate actually describe?
How are instrument choice, strength and assumptions justified?
Are sample overlap, ancestry and selection addressed?
Are confidence intervals and sensitivity analyses reported alongside the main estimate?
Does the conclusion distinguish genetic evidence from intervention evidence?

These questions adapt the reporting priorities in STROBE-MR. Transparent reporting enables appraisal; it does not certify a causal conclusion. Read alongside correlation, confounding and causation and sensitivity analyses in ageing research.