Population Data and Individual Decisions: Bridging the Gap

Why a Strong Association Is a Weak Prediction

Consider a marker where the top quartile has twice the mortality of the bottom quartile over ten years. That is a large effect by epidemiological standards.

Now consider what it means for one person in the top quartile. If ten-year mortality in the bottom quartile is 4 per cent, the top quartile figure is 8 per cent, which means 92 per cent of people in the worse group are still alive. The relative difference is large and the absolute difference is four percentage points, and an individual cannot know which side of it they fall on.

This is the distinction between relative and absolute risk, and it is where most misinterpretation happens. A doubling of a small risk remains a small risk, and a 20 per cent relative reduction in a large risk is a bigger absolute gain than a doubling of a rare one.

The practical implication: population associations are appropriate for deciding what to prioritise, since they rank factors by expected effect, and inappropriate for predicting what will happen to you. Both uses are legitimate; only the second is unavailable.

The Confounding Problem

These are not hypothetical. Observational data supported beta-carotene, vitamin E and hormone therapy in configurations that randomised trials subsequently contradicted, in some cases showing harm. The observational studies were competently conducted, and the confounding was simply not removable by adjustment.

Mendelian randomisation partly addresses this by using inherited variants as a natural randomisation, which is why the lipoprotein(a) and apolipoprotein B evidence is more persuasive than an ordinary cohort association would be. It has its own assumptions and it is a genuine improvement.

What Population Data Are Genuinely Good For

The criticism above is not an argument for ignoring them, and their legitimate uses are important.

Ranking priorities. Population data establish that not smoking, blood pressure control, apoB reduction, fitness and adiposity matter more than any supplement. That ranking is robust and it is the single most useful output of this kind of evidence.

Identifying candidates. Associations generate hypotheses for trials, which is how the interventions that later proved effective were found.

Establishing reference distributions. Knowing where a value sits in a population is necessary context for interpreting it, even when it does not predict an individual outcome.

Detecting rare harms. Trials are too small and too short for rare adverse effects; surveillance and cohort data find them.

Answering questions trials cannot. Long-term exposures, rare outcomes and populations that cannot be randomised.

Calibrating expectations. Effect sizes from cohorts, even if inflated by confounding, indicate the order of magnitude available, which is useful for deciding whether an intervention is worth pursuing.

The Individual Layer That Population Data Cannot Supply

Three things are only knowable about a person, which is where individual observation earns its place.

Baseline status. Whether your vitamin D, iron or omega-3 status is low is not a population question, and it determines whether correction helps.

Your specific interactions and contraindications. A population recommendation cannot check your medications.

Your response. Whether a compound moved a marker in you is answerable only by measuring you before and after. No trial supplies it.

That third one is the substantive addition, and it is why the observation window matters. A trial average of no effect is compatible with some people responding and others not, and the only way to find out which you are is to change one thing and re-measure.

The honest limits of that too. A single-person before-and-after comparison is vulnerable to regression to the mean, seasonal variation, measurement noise and expectation. Which is why the observation window has to be long enough for real change, the marker has to be reliable, conditions have to be standardised, and only one thing can change at a time. Done carelessly, individual observation produces more confident wrong conclusions than population data does.

How the Two Combine

The structure that follows from all of this is a sequence rather than a choice.

Population data set the priorities. Do the things with large, robust effects first. This eliminates most of what people spend time on, since supplements rank below behaviour on every population ranking.

Trial evidence sets the candidate list. Only compounds with human evidence at achievable doses are worth considering, which is a short list.

Individual measurement selects from that list. Baseline status determines which corrections are relevant, and medications determine which candidates are excluded.

Individual observation determines whether it worked. One change, defined window, re-measurement, and a willingness to remove.

Population data set the expectations. If a cohort effect size is small, an individual should not expect a large change, and a large perceived change is more likely to be something else.

Each stage constrains the next, and skipping a stage is where the common errors happen: acting on a population association without trial evidence, taking a compound without measuring status, or judging a response without a controlled comparison.

The Honest Position on Personalisation

Given the above, personalisation is genuinely valuable and its value is narrower than the market's claims.

It cannot tell you what will happen to you. No amount of individual data overcomes the fact that outcomes are probabilistic and the distribution is wide.

It can tell you which population-recommended interventions are relevant to your status, which candidates are unsafe given your medications, and whether a change you made moved something measurable. Those three are real and none is trivial.

It is also worth noting what personalisation frequently substitutes for. A person with an elaborate personalised protocol who is not sleeping regularly, not training and not addressing blood pressure has personalised the small variable. The population data are unambiguous about the ordering, and personalisation applied before the universal advice is followed is optimising at the margin while the centre goes unattended.

The defensible framing: population data tell you what to do, trial evidence tells you what might work, individual measurement tells you which parts apply to you, and individual observation tells you whether they did. All four are needed and the first carries the most weight.