Bayesian Evidence and Uncertainty in Longevity Research
Prior, likelihood and posterior
A prior distribution describes uncertainty about a quantity before analysing the current data. The likelihood describes how compatible those data are with possible values of that quantity. Combining them produces the posterior distribution: updated uncertainty conditional on the model and evidence.
A 95% credible interval contains 95% of the posterior probability for the quantity. Its interpretation depends on the chosen prior and model. A posterior probability of benefit also requires a defined outcome and threshold; it is not a general probability that an intervention slows ageing.
These principles are explained in the FDA’s Bayesian statistics guidance, whose regulatory scope is medical devices.
An illustrative update
Consider two specified, mutually exclusive models: H1 and H0. Suppose the prior probability of H1 is 10%, giving prior odds of 1:9. If the observed data are three times as probable under H1 as under H0, the Bayes factor is 3. Posterior odds become 3:9, giving a posterior probability of 25% for H1.
To convert odds to a probability, divide the weight for H1 by the combined weights for both models: 3 / (3 + 9) = 25%. The Bayes factor multiplies the prior odds, not the prior probability. Multiplying 10% by three would give 30%, which is not the posterior probability in this example.
With a 50% prior probability and the same Bayes factor, the posterior probability would instead be 75%. These are invented teaching numbers, not findings about any longevity intervention. For composite models, the Bayes factor itself depends on their parameter priors. Evidence favouring one model therefore does not establish a universal probability that its biological interpretation is correct.
Keep the biological question attached to the evidence
For longevity research, apply the update to the question actually studied:
- Laboratory evidence: a change in cultured cells concerns that experimental system; it does not directly measure whole-organism healthspan.
- Animal evidence: survival or function in a specified species concerns those animals and conditions; transferring the result to humans requires additional assumptions.
- Human evidence: a biomarker change concerns the measured marker. It does not by itself establish longer survival or improved daily functioning.
This distinction follows the separation between biological measurements and validated surrogate endpoints. Validation requires evidence that an intervention’s effect on the surrogate predicts a specific clinical benefit in a defined context. See the FDA–NIH definition of a validated surrogate endpoint.
Check how much the conclusion depends on assumptions
Inspect the justification for prior information and its relevance to the study population and endpoint. Compare results under plausible alternative priors and assumptions about missing data. Examine model fit and distinguish uncertainty about the estimated effect from uncertainty about future outcomes. These checks are discussed in the FDA guidance.
As a reading exercise, rewrite a study’s conclusion with its population, outcome, follow-up period and assumptions included. If that sentence describes a biomarker result in mice, replacing it with a claim about human lifespan changes the question rather than merely simplifying the language.
Related research literacy
Continue with sensitivity analyses and robustness checks, surrogate endpoints versus clinical outcomes, and external validity in ageing research.