Surrogate Endpoints Versus Outcomes That Matter

A clinical outcome describes how a person feels, functions, or survives. A surrogate endpoint is a laboratory measure, imaging feature, physiologic value, or intermediate event used in place of that outcome. Viral load, blood pressure, LDL cholesterol, tumor response, amyloid plaque, and glycated hemoglobin can each serve as surrogates in particular development programs.

The advantage is time. A biomarker may change in weeks while kidney failure, stroke, disability, or death takes years. The risk is substitution error: a treatment can improve the marker without improving the outcome, improve one outcome while worsening another, or change the marker through a pathway that does not carry the expected benefit.

Endpoint, biomarker, and surrogate are different roles#

An endpoint is the precisely defined variable used to answer a trial question. It includes how the measure is obtained, when it is assessed, and how results are summarized. A biomarker is a defined biological characteristic. It becomes a surrogate endpoint only when it substitutes for a clinical outcome in a specific use.

The FDA and NIH BEST resource defines a validated surrogate as one supported by a clear mechanistic rationale and clinical data showing that effects on the surrogate predict effects on the clinical outcome. The context is part of the definition. Validation for one outcome or intervention mechanism does not grant universal validity.

A patient-reported symptom score is generally a clinical outcome assessment, not a biomarker surrogate, because it directly records how a person feels or functions; hospitalization can be a clinical endpoint, although admission practices and composite definitions can still complicate interpretation. Time to an intermediate disease event may be closer to the patient than a blood test but still fail to capture survival or quality of life.

Correlation within people is not enough#

Suppose people with lower marker values have fewer events. That individual-level association can be strong because the marker reflects disease severity. It does not follow that forcing the marker down with any intervention will reproduce the lower risk.

The marker may be a bystander rather than a causal link. It may sit on one of several pathways. A drug may alter the laboratory result without altering the damaging process. Confounding can make the marker and outcome move together even when changing the marker is ineffective.

Surrogate validation therefore needs trial-level evidence. Across multiple randomized comparisons, do treatment effects on the marker predict treatment effects on the clinical outcome? A useful model examines both strength and consistency of that relation, uncertainty, treatment classes, and whether predictions hold in new trials.

Perfect individual correlation would still not guarantee trial-level surrogacy. Conversely, a marker can have imperfect individual prediction yet be useful for estimating a treatment effect at the trial level. The two levels answer different questions.

Causal pathways explain common failures#

Fleming and DeMets described why a surrogate can fail when it does not capture every important pathway through which disease affects outcome or treatment affects health. Four patterns recur.

First, the surrogate may not be in the causal pathway. Second, it may capture only one branch while disease harms people through other branches. Third, treatment may change the surrogate through a mechanism unrelated to benefit. Fourth, treatment may have harmful effects that bypass the surrogate entirely.

A causal diagram makes the problem visible. Disease can lead to the marker and the outcome through shared causes. Treatment can affect the marker, the target outcome, and harms through separate arrows. Measuring only the marker hides the rest, which is why biological plausibility cannot finish validation: a coherent mechanism raises confidence, but unforeseen pharmacology, compensatory physiology, and adverse effects remain possible.

The classic warning from arrhythmia suppression#

After myocardial infarction, premature ventricular beats were associated with higher mortality. It appeared reasonable that suppressing those beats would prevent death. The Cardiac Arrhythmia Suppression Trial tested that premise.

In CAST, active antiarrhythmic drugs suppressed the electrical marker yet increased arrhythmic death and overall mortality compared with placebo. The surrogate moved in the desired direction while the clinical outcome moved in the wrong direction.

CAST does not prove that all surrogates are unreliable. It shows why association plus mechanism cannot replace randomized outcome evidence when a treatment may create competing harm, and a good surrogate must capture the net effect relevant to the target clinical outcome, not only target activity.

Examples span a spectrum of confidence#

Blood pressure is widely accepted as a surrogate for several cardiovascular outcomes because randomized trials across interventions show that lowering it reduces stroke and other events, and the relation is still not identical for every medicine, population, target, or adverse effect. A drug that lowers pressure but causes another serious harm cannot be judged from pressure alone.

LDL cholesterol has strong causal, genetic, and randomized-treatment support for atherosclerotic cardiovascular risk. Yet benefit depends on absolute risk, duration, degree of lowering, mechanism, and adverse effects. A change in LDL is not a complete outcomes report, and you should not read it as one.

Sustained virologic response after hepatitis C treatment is considered a validated surrogate in its context because viral eradication predicts durable clinical benefit. HIV viral-load measures also support treatment evaluation within defined settings, alongside resistance, toxicity, and clinical outcomes.

Tumor shrinkage or response rate can support an earlier cancer-drug decision when a large, durable response is reasonably likely to predict benefit in a serious disease with unmet need. It may not predict longer survival or improved function, particularly when responses are brief, assessment is unblinded, or later-line treatments differ. The FDA surrogate table lists the endpoints used as the basis of approvals, but inclusion documents regulatory precedent in a stated use; it is not a universal validation certificate.

HbA1c is useful and incomplete#

Glycated hemoglobin summarizes average glycemia over preceding months. Higher HbA1c is associated with microvascular complications, and randomized evidence supports reducing sustained hyperglycemia to lower selected microvascular risk. It is therefore an efficient endpoint for glucose-lowering efficacy.

HbA1c cannot show the full treatment effect. Two medicines can produce the same mean change while differing in hypoglycemia, weight, heart failure, kidney outcomes, gastrointestinal effects, treatment burden, and mortality. A therapy can lower HbA1c while creating enough harm to make the net outcome unfavorable.

Measurement also has limits. Red-cell lifespan, hemoglobin variants, recent blood loss or transfusion, kidney disease, and assay factors can alter interpretation; a mean HbA1c does not show you glycemic variability or time below range. Modern diabetes programs therefore often pair glycemic endpoints with dedicated cardiovascular and kidney outcomes in relevant populations. What HbA1c measures explains the laboratory measure, while absolute versus relative risk helps translate event results.

Validated versus reasonably likely#

FDA's surrogate endpoint resources describe degrees of validation. A validated surrogate is known to predict clinical benefit based on strong evidence, and a reasonably likely surrogate has support but greater uncertainty and may be used under an accelerated pathway for serious conditions with unmet need.

Accelerated approval is not a claim that clinical benefit has already been fully verified. It allows earlier access while requiring confirmatory work to verify and describe benefit, and if the expected benefit is not confirmed, the indication can be withdrawn, sometimes through a formal process.

The confirmatory trial must answer the unresolved clinical question. Repeating the same surrogate alone may not be enough. Timing matters: delayed trials prolong uncertainty, while an adequately advanced confirmatory trial at approval can shorten it. And “reasonably likely” is a regulatory standard in a context, not ordinary-language certainty, so when you describe one of these approvals, say whether it was traditional or accelerated, which endpoint supported it, and where the confirmatory evidence stands.

Composite and intermediate clinical endpoints#

Not every endpoint fits neatly into surrogate or final-outcome boxes: disease progression, hospitalization, and composite endpoints may be clinically meaningful yet still differ in importance from death or irreversible disability.

A composite combines events to increase statistical efficiency. It is interpretable when components have similar importance, expected effects, and ascertainment. A favorable composite can be driven by the most frequent, least serious component while death is unchanged, and the solution is to report every component with uncertainty, not to discard composites entirely.

Progression-free survival in oncology captures time to progression or death and can matter to patients, but assessment timing, measurement error, symptoms, subsequent therapy, and survival remain important. A radiographic change without symptomatic or functional benefit may have different value than delaying a painful or disabling event.

A checklist for reading a surrogate trial#

Before you accept the substitution, ask six questions:

1. What clinical outcome is the surrogate intended to predict?
2. Is validation based on multiple randomized trials and treatment classes, or mainly on biological association?
3. Does the treatment act through the same mechanism represented in the validation evidence?
4. Which benefits and harms bypass the surrogate?
5. Is the claim based on a validated or reasonably likely surrogate, and what confirmation is pending?
6. Would the observed change be large and durable enough to predict a meaningful clinical effect within the studied population?

Then inspect the estimate yourself. A statistically precise biomarker change can coexist with wide uncertainty about patient benefit. Conversely, a modest surrogate change can matter when validation is strong and baseline clinical risk is high.

How to write the conclusion accurately#

If a trial lowered a surrogate, say exactly that. Do not replace “lowered LDL,” “reduced plaque,” or “improved HbA1c” with “prevented heart attacks,” “stopped dementia,” or “prevented complications” unless clinical-outcome evidence supports the stronger sentence.

When validation is strong, state the predicted outcome and context. When it is reasonably likely, state that clinical benefit remains to be confirmed. When it is unvalidated, call it a biomarker or intermediate endpoint and avoid implying substitution.

This precision does not diminish efficient drug development. It protects the bridge between an early signal and the outcome a patient actually values.

Validation can also age. Changes in background therapy, disease stage, measurement technology, or mechanism of action may weaken a relationship established in an earlier treatment era. Reassessment is especially important when a new intervention affects pathways beyond the biomarker or when off-target harms could offset the predicted benefit, and a surrogate is not a permanent property of a laboratory value. It is an evidence claim tied to a setting and a body of interventions.

References#

1. FDA and NIH BEST surrogate-endpoint resource
2. FDA surrogate endpoint resources
3. Institute of Medicine evaluation framework
4. Fleming and DeMets on surrogate endpoints
5. CAST arrhythmia-suppression findings
6. FDA table of surrogates used for approvals