How a Surrogate Endpoint Is Validated, and Why So Few Qualify

A surrogate endpoint is a measurement, often a biomarker like blood pressure or LDL cholesterol, used in place of the outcome that actually matters to patients, such as a stroke or death. Validating a surrogate means showing that an effect on the surrogate reliably predicts the effect on the real outcome, which is a much stronger requirement than showing the two are merely correlated. The formal standard, the Prentice criteria, is demanding, and the more practical modern approach asks whether treatment effects on the surrogate track treatment effects on the outcome across many trials.

What a surrogate is doing for a trial

Many trials cannot afford to wait for the outcome that matters. Showing that a cholesterol drug prevents heart attacks may take years and thousands of patients; showing that it lowers LDL cholesterol takes weeks. A surrogate endpoint is a stand-in measurement, usually earlier and easier, that a trial measures instead of the clinical outcome, betting that moving the surrogate means moving the outcome.

The bet can pay off well or fail badly. LDL lowering has proven to be a reasonably dependable surrogate for cardiovascular events across many drug classes. Other surrogates have led whole fields astray. The difference is whether the surrogate was validated, and validation is a specific, hard-won property, not an assumption that follows from biology.

Correlation is not enough, and that is the trap

The intuitive test is the wrong one. It is easy to show that a biomarker is correlated with an outcome: people with higher blood sugar have more complications, people with more extra heartbeats have more sudden death. That correlation tempts everyone to assume a drug lowering the biomarker will lower the outcome. It does not follow.

The classic warning comes from antiarrhythmic drugs that suppressed the extra heartbeats associated with death after a heart attack. The drugs moved the surrogate in the desired direction and increased mortality. A biomarker can sit on the causal path to an outcome and still be a bad surrogate, because a drug can affect the outcome through pathways that have nothing to do with the biomarker, including harms. Fleming and DeMets described this failure in detail, and it reshaped how carefully surrogates are treated.

The Prentice criteria

Prentice gave the first formal definition. A valid surrogate, in his framing, is an endpoint such that a test of the treatment's effect on the surrogate is also a valid test of its effect on the true outcome. From that definition come the operational criteria: the treatment must affect the surrogate, the treatment must affect the true outcome, the surrogate must be associated with the true outcome, and, the demanding one, the effect of treatment on the true outcome must be fully captured by the surrogate.

That last requirement is the heart of it. Full capture means that once you know what happened to the surrogate, knowing the treatment adds no further information about the outcome. If it does add information, some of the treatment's effect is flowing around the surrogate, and the surrogate is not telling the whole story.

Why almost nothing meets the strict bar

The full-capture requirement is where most candidate surrogates fail. Real treatments rarely act through a single channel. A drug that lowers a biomarker may also have off-target effects, toxicities, or additional benefits the biomarker never sees. If any of the treatment's effect on the outcome flows around the surrogate, the strict criterion is not met, and the surrogate can mislead.

Proving full capture in a single trial is close to impossible, because it requires showing that a residual treatment effect is exactly zero, which no finite study can establish with confidence. Statisticians proposed patches, such as the proportion of the treatment effect explained by the surrogate, but that measure turned out to be unstable and hard to interpret. The single-trial approach ran into a wall.

The modern answer: validate across trials

The practical answer moved the question up a level. Instead of asking whether one trial's surrogate captures the effect, the meta-analytic approach asks whether, across many trials, the treatment effect on the surrogate predicts the treatment effect on the outcome. Buyse and colleagues formalized this into two kinds of agreement.

One is individual-level association: do patients with better surrogate values tend to have better outcomes. The more important one is trial-level association: do trials that produced a bigger effect on the surrogate also produce a bigger effect on the outcome. Trial-level surrogacy is the property that actually licenses using a surrogate to predict clinical benefit. It requires a body of randomized trials measuring both endpoints, and it yields a quantitative sense of how tightly the two move together and how wide the prediction's uncertainty is.

Reading a surrogate-based result

When a trial reports success on a surrogate, ask what tier of evidence stands behind that surrogate. Has it been validated at the trial level across multiple treatments, or is the field relying on biological plausibility and individual-level correlation. The same biomarker can be a strong surrogate for one drug class and unproven for another, because validation is specific to the mechanisms in play.

This is why regulators treat surrogate-based approvals as provisional in many cases, pairing them with a requirement to confirm the clinical outcome later. A surrogate buys speed by borrowing against a prediction. Reading the result well means knowing how good the collateral is.