Better Numbers, Better Patients? — An Introduction to Surrogate Endpoints

When we say a treatment worked, we almost always point to a single number. 「The values went down」, 「the imaging improved」, 「progression slowed」. But what patients actually experience is dying, breaking a bone, going to the hospital, or getting too short of breath to climb stairs. Clinical outcomes are things the patient directly experiences and feels, and surrogate endpoints are intermediate measurements taken in place of those outcomes. The numbers we point to when talking about the success or failure of a treatment are usually surrogate endpoints.

Surrogate endpoints are also frequently used as the basis for approval. Among 51 products approved between 2011 and 2018 through two expedited pathways by the European Medicines Agency, only 5 (10%) were approved based on clinical outcomes themselves, while 46 (90%) were approved based on surrogate endpoints. The authors searched separately for literature validating each surrogate endpoint, but could not find a single validation study confirming that the surrogate endpoint predicted clinical outcomes. Most remained at grades such as 「reasonably expected to predict clinical outcomes」 or 「biologically plausible」. The authors also note the limitation that they may have missed validation studies (PMID 31504034).

There are good reasons to use surrogate endpoints. They can be measured quickly, with fewer people, and less expensively. But this substitution comes with a cost, and when we pay that price is what this article is about. Numbers in diagnosis were read in "What a Test Actually Tells You — An Introduction to Sensitivity, Specificity, and Pretest Probability" and "Where Does Pretest Probability Come From? — History, Physical Examination, and Clinical Prediction Rules". This article moves on to the numbers that say a treatment was effective. It only deals with how to read indicators, and does not judge whether a specific treatment is good or bad. The judgment to start or adjust a treatment for an individual is something made by weighing that person's illness, risks, and alternatives, and is outside the scope of this article.

Why You Need to Know — Most of the Numbers We See Are Surrogates

The situation is similar in the United States. Accelerated approval is a system that allows drugs to be licensed first based on surrogate markers, with actual benefits confirmed through confirmatory trials after licensing. Across 156 drug-indication pairs granted accelerated approval between 2015 and 2022, 77% of the pivotal trials serving as the basis for approval used a single-arm design, meaning a design looking at only a single group receiving treatment without a control group. The median number of participants was 92. In post-approval confirmatory studies, the proportion using surrogate endpoints increased from 50% to 72% (PMID 39695642). However, the authors add that the strength of evidence appeared to improve in 2021–22 (single-arm design 91% → 69%, randomized confirmatory studies 42% → 75%).

Looking closely at a single field makes it even clearer. Nearly all glaucoma trials use intraocular pressure as an endpoint. Yet a review summarizing this field notes that for no class of intraocular pressure-lowering treatment has intraocular pressure ever been properly validated as a surrogate endpoint to date (PMID 25034049). The same review notes that evidence has been accumulating for imaging markers that measure optic nerve damage as surrogate endpoints.

So this is not a rare occurrence. A substantial portion of the language we use when talking about the success or failure of treatments is the language of surrogates.

The Big Picture — Where a Surrogate Marker Betrayed Us

The axis of this article is an event that actually happened. People who have ventricular arrhythmia after a myocardial infarction have a poor prognosis. There were drugs that suppressed arrhythmia. Thus, the deduction emerged that suppressing arrhythmia would reduce death. Built on this deduction, called the "suppression hypothesis," drugs were used and trials were designed. The logic seemed to have no flaws.

The result was the exact opposite. The drugs, which worked by slowing conduction, markedly suppressed ventricular arrhythmia while increasing death in a substantial subset of patients with heart disease. This harmful effect was confirmed both in placebo-controlled randomized trials and in a meta-analysis combining small randomized trials (PMID 8780325).

This event changed the grammar of antiarrhythmic drug trials. Afterward, the endpoint of trials in this field shifted from surrogate markers like "was arrhythmia suppressed to a set degree?" to total mortality (PMID 8780325). A contemporary overview notes that this trial showed two things: how important a placebo control is in judging treatment efficacy, and what pitfalls exist when a surrogate marker is substituted for death. New prescriptions for this class of drugs steadily decreased thereafter, and the "suppression hypothesis" was no longer used as the theoretical basis for that treatment (PMID 7654403).

The lesson of this event lies in the pathway. Arrhythmia appeared together with bad outcomes, but getting rid of arrhythmia did not make bad outcomes go away. And the way of getting rid of arrhythmia itself could create other harms. This is one of the typical ways a surrogate marker breaks down.

Conditions for a Valid Surrogate Marker — Predicting and Fully Capturing the Effect

Clinical outcomes are also called true outcomes. Death, fractures, hospitalization, symptoms, and function fall under this category. Surrogate markers include test values, imaging findings, pathology findings, and time to progression.

The glaucoma primer writes the conditions a surrogate marker must meet in the briefest terms: A valid surrogate marker must be able to predict clinically important outcomes and fully capture the effect of an intervention on those outcomes (PMID 25034049).

The latter condition is difficult. Predicting the outcome alone is not enough. The effect of treatment on the outcome must entirely travel through that surrogate marker. If the treatment also acts through other paths that do not pass through the surrogate marker, whether those paths are beneficial or harmful, you cannot know the outcome by looking at the surrogate marker alone.

The Individual Level and the Trial Level Are Different — The Most Important Distinction in This Article

The relationship between a surrogate marker and an outcome can be viewed at two layers.

• The relationship at the individual level is the relationship that "if this person's metric is bad, this person's prognosis is bad." This is a property possessed by prognostic factors, which are characteristics that inform future progression in advance.
• The relationship at the trial level is the relationship that "if a treatment changes this metric by this much, that treatment changes the outcome by this much as well." This is the property required of a surrogate marker. Multiple trials are gathered to see whether the treatment effect shown in the surrogate marker and the treatment effect shown in the outcome move together.

Even if the former is true, the latter can be false. Examples of that are below.

Data from an integrated analysis of 13 lung cancer immunotherapy trials (9,285 people) by a U.S. regulatory agency illustrate this difference like a picture. The endpoints used here are clarified first. Overall survival is the duration until death from any cause, and progression-free survival is the duration until the disease progresses or death occurs.

Response refers to the tumor shrinking by a defined standard on imaging, and response rate is the proportion of such patients. At the individual level, people who responded to treatment survived much longer than those who did not. The hazard ratio, which is the value dividing the speed at which events occur in the two groups, was 0.28 (95% CI 0.26–0.30).

The coefficient of determination is a proportion indicating how much the treatment effect seen in the surrogate marker across trials explains the difference in the treatment effect seen in the clinical outcome, and a larger value means that results are predicted well just by looking at the surrogate marker. The relationship at the trial level is measured with this value. In this data, the coefficients of determination for response rate and overall survival stayed at a moderate level of 0.61, and for progression-free survival and overall survival at 0.70 (PMID 38458207). The fact that responders live longer alone could not guarantee that a treatment producing more responses increases survival by that much.

In osteoporosis data, the answer to the same question varied depending on the analysis method. Individual patient data analysis gathers the data of each and every trial participant and recalculates. Meta-regression is a method that gathers summary values published for each trial and draws a line to see whether two effects move together across trials.

According to a review addressing the relationship between bone mineral density and fractures, analyses using individual patient data reported that the portion of the fracture prevention effect of antiresorptive agents explained by increased bone mineral density was limited. Analyses using meta-regression on summary statistics reported that most of the fracture prevention effect was thanks to bone mineral density improvement. For the same question and the same treatment, the answer differed depending on the analysis method, and the review notes that caution is needed in interpretation when results conflict (PMID 15615079). The same review notes that the bone mineral density value itself correlates with fracture risk and is useful for screening patients needing treatment. Even so, at the time the review was published, the evidence that bone mineral density raised by antiresorptive agents reliably substituted for fracture risk reduction was judged to be limited. A metric useful at the individual level was still awaiting validation at the trial level.

Thresholds and Composite Endpoints — How Much It Must Move and What Got Bundled

Just because there is a trial-level association does not mean it can be used right away. We need to know how much a surrogate endpoint must move to predict an improvement in the true outcome. The surrogate threshold effect (STE) is the minimum treatment effect required in a surrogate endpoint to predict a non-zero effect on the true outcome. It was introduced to solve the problem that a single coefficient of determination value is hard to interpret, and its advantage is that it reads naturally in a clinical context (PMID 17080751).

Something to look at alongside surrogate endpoints is composite endpoints. A composite endpoint is a metric that bundles different events, such as death, myocardial infarction, and stroke, and counts them as a single outcome. It shares the same advantage as surrogate endpoints in that events happen more frequently, which reduces the required sample size or follow-up duration. A commentary summarizing this notes that when reading a study reporting a composite endpoint, one must check ① the clinical importance of each component, ② the frequency of each, and ③ whether the effect of the intervention on each is similar (PMID 32739473). If even one of the three is way off, the result that the composite endpoint decreased may actually be the result of the most common and mildest component moving.

Why They Disagree — Three Ways

The first is when the surrogate marker captures only part of the treatment effect. The share of the treatment effect that passes through the surrogate marker can actually be calculated.

In an analysis pooling individual patient data from 23 osteoporosis trials (91,779 people), there was a clear link between changes in bone mineral density and reductions in fractures (r² of 0.73 for hip bone mineral density regarding vertebral fractures). In the same analysis, the share of the treatment effect explained by changes in bone mineral density was 44% to 67% (PMID 32707115). The authors consider this link strong enough to support using bone mineral density as a surrogate outcome, concluding accordingly. The same numbers also mean that one-third to half of the treatment effect remains outside bone mineral density.

The numbers on the pulmonary arterial hypertension side are smaller. In an analysis pooling patient-level data from 10 randomized placebo-controlled trials submitted to regulatory agencies (2,404 people), the 6-minute walk distance change at week 12 explained 22.1% (95% CI 12.1–31.1) of the treatment effect. The authors conclude that this marker has limited validity as a surrogate marker and may not be a sufficient surrogate marker (PMID 22696079).

The second is when the treatment helps or harms through paths outside the surrogate marker. When a treatment acts via paths that do not pass through the surrogate marker, the surrogate marker misses that effect. This happens in two directions.

One direction is missing good news. A class of diabetes drugs that increases glucose excretion in the kidneys was originally developed to control blood sugar. However, in large cardiovascular outcome trials, this class improved cardiovascular and renal outcomes, and that benefit extended even to patients with heart failure with reduced ejection fraction without diabetes. A review summarizing this class notes that this benefit is unlikely to be related to improved blood sugar control (PMID 32665641). If you had looked only at blood sugar as a surrogate marker, a large part of this benefit would have been invisible. Surrogate markers blind you to the good side, too.

The other direction is missing harm, and the arrhythmia suppression from earlier is an example. The surrogate marker of arrhythmia frequency moved in the intended direction, but deaths increased. A contemporary review cites the property of slowing conduction itself as one of the reasons that can cause arrhythmias (PMID 8780325, PMID 7654403). The reason this direction is frightening is that in trials looking only at surrogate markers, it is invisible in principle. Because the marker is improving, the trial appears successful.

The third is when even the same marker changes if the conditions change. The validity of a surrogate marker is tied to what disease, what treatment class, and what outcome it has been validated for.

Bone mineral density shows this most cleanly. In antiresorptive trials, changes in bone mineral density predicted fracture reductions. However, in a meta-regression separately pooling 15 trials (47,365 people) supplementing calcium alone or with vitamin D, greater increases in lumbar spine bone mineral density did not lead to greater reductions in fractures. Hip bone mineral density showed a relationship with fracture reduction, but after excluding the two largest studies that measured bone mineral density in only a very small subset of participants, it fell just short of the significance level, and the authors note that this may have skewed the results. The authors read the results to mean that changes in bone mineral density cannot be used as a surrogate marker for anti-fracture effects in this context, and add that calcium and vitamin D may reduce fractures through mechanisms unrelated to bone mineral density (PMID 21060990).

The same property repeats in oncology. A study analyzing 101 randomized trials of advanced colorectal cancer notes that surrogate relationships observed in selected situations may not apply to other classes or other lines of therapy (PMID 25863582).

How We Validate It — and Where Validation Succeeds

There are sites of success just as much as sites of failure. A properly validated surrogate marker actually works.

An example where predictions later panned out comes from trial data on adjuvant therapy for resected melanoma. Recurrence-free survival is the length of time until recurrence or death. When researchers modeled the relationship between recurrence-free survival and overall survival using individual patient data from 13 randomized trials (involving 5,826 or more people), 12 of the 13 trials showed the two markers moving consistently together. In the remaining trial, recurrence-free survival improved while survival did not, and including this trial made it impossible to reliably estimate the coefficient of determination. The failure pattern we saw earlier appeared right in this single trial.

Based on the 12 trials, the individual-level correlation was 0.89, the trial-level coefficient of determination was 0.91, and the surrogate threshold effect was estimated at a hazard ratio of 0.77. This means that in future similar adjuvant therapy trials, if the hazard ratio for recurrence-free survival is 0.77 or lower, we can predict that there will also be an effect on survival.

And the authors applied that prediction to an actual trial. In one immune checkpoint inhibitor trial, the recurrence-free survival hazard ratio was 0.75, and the model predicted a survival benefit. That prediction was subsequently confirmed. The overall survival hazard ratio was 0.72 (95.1% CI 0.58–0.88) (PMID 28922786). A validated surrogate marker does this: before a trial ends, with fewer events, it puts out a falsifiable prediction—and that prediction turns out to be right.

An example of calculating and reproducing a threshold is found in osteoporosis. Researchers calculated a threshold value by running a meta-regression between 24-month changes in total hip bone mineral density and reductions in fractures, using individual patient data from 16 randomized trials involving 61,415 people. For a trial to predict a reduction in fractures, treatment had to increase total hip bone mineral density compared to the control group by at least 1.83% for all fractures, 1.42% for vertebral fractures, and 3.18% for hip fractures. This is a trial-level benchmark rather than an individual patient's target, and the threshold differed depending on the type of fracture.

Validation was performed in two batches. When checking whether bone mineral density changes crossed the threshold and fractures were also significantly reduced, 27 out of the 28 trials used to build the threshold met this condition. So did 10 out of 11 separate trials that were not used to build the threshold (PMID 34490915). This means the threshold held true even in trials it had never seen before.

A similar site exists in nephrology. In chronic kidney disease, proteinuria (albuminuria) has long been a leading candidate. In a 2019 meta-analysis pooling individual patient data from 41 randomized trials with 29,979 people, every 30% reduction in albuminuria from treatment lowered the risk of clinical endpoints by an average of 27% (median coefficient of determination 0.47). This relationship was stronger in patients with high baseline albuminuria and less certain in those with low baseline albuminuria (PMID 30635226).

A 2026 follow-up analysis scaled up the size to 48 trials and 85,681 people. Here, every 30% reduction lowered the risk by an average of 19%, and the median coefficient of determination was 0.66 (PMID 41198855). The numbers from the two analyses differ. The strength of a surrogate relationship is an estimate that shifts depending on which trials are pooled, not a fixed constant.

Progression vs. Survival — The Most Widely Used Surrogate Endpoints in Oncology

Using progression-free survival instead of overall survival as an endpoint in cancer trials is now close to standard. A methodological review examining how well its validity has been tested found 91 studies evaluating the surrogate validity of progression-free survival across 24 cancer types. Among the 47 studies concluding that "progression-free survival is a good surrogate endpoint," 15 (32%) presented no quantitative evidence to support that conclusion (PMID 32214230).

The actual predictive power can be seen across 260 phase 3 trials in metastatic solid tumors (172,133 participants). Trials with progression-free survival as the primary endpoint met their endpoints in 66.9% of cases, succeeding much more often than trials with overall survival as the primary endpoint (33.3%). However, the proportion of trials that were positive for progression-free survival and translated into a survival benefit was only 38% (31 out of 82) (PMID 32702645).

Collecting 2,109 randomized oncology trials over the past 50 years reveals a bigger picture. The correlation between the treatment effect on progression-free survival and the treatment effect on overall survival was moderate (r = 0.60, 95% CI 0.56–0.64). Across 300 randomly selected trials, median progression-free survival increased by an average of 1.6 months, and median overall survival increased by an average of 1.4 months. One in five trials met the criteria for a clinically meaningful improvement in survival (PMID 35777712).

The variation is wide depending on the cancer type. Pooling 20 studies in follicular lymphoma (10,724 participants) showed a coefficient of determination between progression-free survival and overall survival of only 0.15 (95% CI 0.002–0.35). This means that only 15% of the variation in survival is explained (PMID 38571449). Conversely, in 8 adjuvant therapy trials for HER2-positive early breast cancer (21,480 participants), the individual-level correlation between disease-free survival—the time lived without recurrence or death—and overall survival was 0.90, and the trial-level coefficient of determination was 0.75 (95% CI 0.50–1.00). The authors concluded that disease-free survival can continue to be used as a surrogate endpoint in this setting, while also noting the limitation that the results depend on the presence of a single trial that stood apart from the others (PMID 30709633). The same metric passes for some cancer types and fails for others.

Moreover, there are instances where a metric that tracks survival well does not track the patient experience well. In a post hoc analysis of a phase 3 trial in metastatic prostate cancer (831 participants), radiographic progression-free survival correlated strongly with overall survival (ρ ≥ 0.7). In the same data, the correlation with time to symptomatic skeletal events was weak or modest, and the correlation with time to quality-of-life deterioration was modest to moderate (PMID 39031642). "Progression was delayed" and "fewer bone-related symptomatic events were experienced" are not the same thing. The authors note that these results warrant further investigation.

Pathology Findings and Functional Indicators — Good News for an Individual Is Not the Answer of a Trial

In neoadjuvant trials for early breast cancer, pathological complete response has long been used as a primary endpoint. Pathological complete response is a state in which no cancer remains in the tissue removed during surgery. Intuitively, this is good news. The results of pooling 54 randomized trials (32,611 patients) to calculate trial-level surrogacy showed a coefficient of determination of 0.14 (95% CI 0.00–0.29) between the treatment effect on pathological complete response and the treatment effect on disease-free survival, and 0.08 (0.00–0.22) for overall survival. The authors conclude that this indicator should not be used as a primary endpoint in regulatory neoadjuvant trials (PMID 34933868).

The distinction between the individual level and the trial level remains alive here as well. The proposition that the prognosis of a person who achieves a complete response is good, and the proposition that a treatment creating more complete responses increases survival even more, are separate things.

In rehabilitation, a particularly important item is exercise capacity. Exercise capacity is a meaningful function for patients in its own right, but it is also used as a surrogate indicator replacing death, hospitalization, and quality of life. In a meta-analysis pooling 31 randomized trials of exercise-based cardiac rehabilitation in heart failure patients (4,784 patients) to evaluate the validity of exercise capacity as a surrogate indicator for death, hospitalization, and quality of life, cardiac rehabilitation actually improved exercise capacity more than the control group. Peak oxygen uptake was +3.10 mL/kg/min, and the 6-minute walk distance was +41.15 m.

The problem is what comes next. The trial-level association between exercise capacity improvement and death or hospitalization was low, and the association with health-related quality of life was moderate. The surrogate thresholds predicting quality of life improvement were estimated to be 5 mL/kg/min for peak oxygen uptake and 80 m for the 6-minute walk distance (PMID 29957192).

Reading these results as "rehabilitation is useless" is reading them in exact reverse. The same study confirmed rehabilitation's effect of improving exercise capacity itself. This study says only one thing. You must not convert and read the magnitude of improved exercise capacity as the magnitude of reduced death and hospitalization.

Test Numbers — Reading Clinical Outcomes and Surrogacy Separately

Here, the order matters. We first look at what each paper reported about the clinical outcome itself, and then look at surrogacy.

We look at lipids first. In an umbrella review gathering 20 randomized trials of statins versus placebo or usual care (194,686 participants), statin treatment significantly improved all clinical outcomes. Building on that, a separate question the authors asked was this. Among the trials, did trials with greater reductions in LDL cholesterol also show greater improvements in clinical outcomes? This trial-level association was weak (coefficient of determination for LDL 0–0.1). The authors conclude that while these numbers are useful biomarkers for tracking patients, they should not be used as core evidence in drug trials (PMID 39999862). This paper does not deny the efficacy of that class of drugs. What this paper denies is the conversion formula that says, "Because the number dropped by this much, the outcome must have improved by that much."

Blood glucose looks like this. When gathering 18 cardiovascular outcome trials (161,156 participants), treatment with a new class of diabetes drugs was associated with a lower risk of major adverse cardiovascular events. Major adverse cardiovascular events are usually a composite endpoint grouping cardiovascular death, non-fatal myocardial infarction, and non-fatal stroke. In meta-regression, there was also an association between the magnitude of glycated hemoglobin reduction in the treatment group and the reduction in the risk ratio of major adverse cardiovascular events (β = −0.298, P = 0.007). However, this association was entirely explained by the reduction in the risk of non-fatal stroke (100% of the variance between trials), and the magnitude of glycated hemoglobin reduction was not associated with the risk ratio of heart failure hospitalization or all-cause mortality (PMID 34663316). This does not mean the drug had no effect on those outcomes, but rather that it was not explained by the drop in blood sugar. The authors write that they hope this result, showing that glucose lowering contributes to some cardiovascular benefit, will not change the indications for these drugs, but will help broaden their use. This also does not mean that blood sugar control is meaningless. When talking about a surrogate marker, we must state together which outcome it is a surrogate for.

Blood pressure is often cited as a model example of a valid surrogate marker. When 18 antihypertensive treatment trials were gathered to calculate the surrogate threshold effect for stroke, it was a 7.1 mmHg systolic and 2.4 mmHg diastolic. In the evaluation grading scheme, systolic blood pressure received a B+ and diastolic blood pressure received an A. However, with the same data, a surrogate threshold effect could not be estimated for cardiovascular death and all-cause mortality (PMID 22409774). This means a threshold could not be calculated from this data, not that treatments lowering blood pressure are unrelated to mortality. The same surrogate marker was grade A for stroke and could not even be graded for mortality. Surrogacy must be graded outcome by outcome.

Common Misunderstandings — Three of Them

The first misunderstanding is "The numbers improved, so the treatment works." This is the most basic and widespread misunderstanding, and the examples above are direct counterexamples. The magnitude of the effect of treatment on pathological complete response hardly predicted the effect on progression-free survival and overall survival (coefficients of determination 0.14 and 0.08) (PMID 34933868), and only 38% of trials that were positive in progression-free survival led to a survival benefit (PMID 32702645). And there was actually a case where death increased while the indicator of arrhythmia improved clearly (PMID 8780325).

A change in a numerical value is a candidate for an effect, not proof of an effect.

The second misunderstanding is "If it is valid in one place, it is valid in another." This misunderstanding can be more dangerous in practice. Bone mineral density predicted fracture reduction in trials of antiresorptive agents, but did not do so in trials of calcium and vitamin D supplementation (PMID 21060990). Progression-free survival passed in HER2-positive early breast cancer (coefficient of determination 0.75) (PMID 30709633), while progression-free survival was only 0.15 in follicular lymphoma (PMID 38571449). The conclusions of a colorectal cancer analysis state this principle as is (PMID 25863582), and a review examining the surrogate validity of progression-free survival states the same principle at the starting point. Validation must be performed for each indication and each intervention (PMID 32214230).

Surrogacy is a property attached to a combination of indicator, disease, treatment, and outcome.

The third misunderstanding is "The composite endpoint decreased, so the serious events within it also decreased." For composite endpoints, one must look together at the clinical importance of the components, their frequency of occurrence, and the effect of the intervention on each (PMID 32739473).

Blood glucose data is a practical example of this check. When dissecting the association with glycated hemoglobin reduction in the composite endpoint of major adverse cardiovascular events, non-fatal stroke explained all the variance between trials, and the magnitude of glycated hemoglobin reduction was not associated with heart failure hospitalization or all-cause mortality (PMID 34663316). If you only read the sentence "The composite endpoint decreased," this structure is not visible. Composite endpoints must be read separately to see which component moved.

An Honest Unresolved Dilemma — Some Diseases Are Too Difficult to Test Without Surrogate Markers

This is the biggest reason why you should not read this article as a call to abandon surrogate markers.

In slowly progressing ultra-rare neurological diseases, it is often difficult to measure definitive clinical endpoints like death or sustained functional decline within a feasible trial period. In such diseases, fluid or imaging biomarkers become the path to accelerating drug development. Recent regulatory reviews frame the problem this way and propose a four-stage validation model. The cases show that the validity of a surrogate marker is tied to its context of use, and differing evidentiary standards across countries and the absence of international guidelines specifically for rare disease surrogate markers are cited as structural gaps (PMID 42525369).

Do Confirmatory Trials Actually Finish?

The logic of accelerated approval is clear. You approve first on a surrogate marker, and after approval, you confirm actual benefit with a confirmatory trial. The problem is whether that second half is actually carried out.

When U.S. oncology indications that received accelerated approval between 1992 and 2021 were tracked until December 2024, out of 77 confirmatory trials, 25 (32%) showed a survival benefit at the time of conversion and 52 (68%) relied on surrogate markers. When tracked for an additional median of 5 years after conversion, a survival benefit newly appeared in 7 trials (9%). The authors conclude that new evidence of benefit emerging after conversion is rare (PMID 40690366).

Timelines have sped up. An analysis of oncology accelerated approvals from 1992 to 2024 showed that the median duration to regular approval dropped from 4.3 years to 2.3 years, and the duration to withdrawal dropped from 9.5 years to 3.2 years. The proportion of confirmatory studies already underway at the time of approval increased from 63% to 85% (PMID 40472969). The authors suggest this trend may reflect stronger oversight, and that advancements in trial design and regulatory coordination may also have played a role.

A remaining problem persists, however. In a study pairing and comparing decisions by U.S. and European regulators for 21 oncology indications authorized via the accelerated pathway between 2009 and 2013, 40% of post-approval requirements imposed through the accelerated pathway were delayed in the U.S. and 61% in Europe (median follow-up 7.25 years). 41% of confirmatory trials following U.S. accelerated approval were being conducted in populations different from the approved indication (PMID 33021339).

A widely known individual case came from obstetrics. An injection used to prevent recurrent preterm birth received accelerated approval. In a confirmatory trial completed over the course of 9 years, delivery before 35 weeks was 11% versus 11.5% (relative risk 0.95, 95% CI 0.71–1.26), failing to confirm efficacy, and an advisory committee recommended withdrawal of approval by a vote of 9 to 7. The authors note that this confirmatory trial also had flaws (PMID 33035472).

A conceptual paper looking at this trend on a larger scale was also recently published. The authors divide the forms of regulatory evidence into three: a system requiring two independent randomized trials, a system where one is sufficient, and the current "narrative system." The narrative system relies on surrogate markers, single-arm designs, external controls, real-world evidence, and extrapolation. When assumptions are set realistically, the authors estimate the posterior probability that a positive conclusion is actually false to be about 0.3% in the two-trial system, about 4–9% in the one-trial system, and potentially exceeding 15–20% in the third system. Indeed, the proportion of U.S. approvals backed by two or more pivotal randomized trials dropped from 80.6% in 1995–1997 to 52.8% in 2015–2017 (PMID 42580367). The authors' conclusion is that while flexible evidence frameworks must be maintained, uncertainty must be explicitly managed.

The Question Without an Answer — How Long to Wait

Everything discussed so far is about the strength of the evidence. The really difficult questions cannot be answered by evidence alone.

In a disease with no treatment, should we approve a drug now based on a surrogate marker and wait for confirmation, or make patients wait until it is confirmed? The former choice leads to using treatments that may be ineffective or harmful, while the latter delays a potentially effective treatment by years. Which side is right depends on how urgent the disease is, whether alternatives exist, and what kind of uncertainty that person is willing to tolerate. There are notches on this scale that statistics do not answer.

The answers for each field have not yet gathered either. An international nephrology society's conference focused on expanding patient-reported outcome measures, moving beyond single biomarker targets, and introducing hierarchical composite endpoints. One of the core recommendations is verbatim to "avoid over-reliance on isolated biomarkers" (PMID 41033457). The direction is set, but we have not arrived yet.

To Go Deeper

• Read outcome measures by splitting them apart first. When looking at the results of a treatment, there is only one question to ask first. Is this number something the patient experiences, or is it a proxy for it? Death, fractures, hospitalization, symptoms, and function are on the front end, while test numbers, imaging findings, pathology findings, and time to progression are on the back end.
• If it is a surrogate marker, check four things. ① What outcome is it a surrogate for? (The same marker might be grade A for stroke and have no grade for death — PMID 22409774) ② In what disease and what treatment class has it been validated? (If the class changes, the relationship can collapse — PMID 21060990) ③ Was the validation done at the trial level, or is there only the fact that it is a prognostic factor? (The two are different — PMID 38458207, PMID 15615079) ④ Is there a threshold for how much it needs to move? (PMID 17080751, PMID 34490915).
• When you encounter a composite endpoint, pick it apart. See whether the importance, frequency, and effects of the components are similar to one another (PMID 32739473). Any single one of them may have carried the rest (PMID 34663316).
• Distinguish between "approved" and "benefit confirmed". Treatments approved via an expedited pathway may still be awaiting confirmation of benefit, and the rate at which that wait actually reaches a conclusion is lower than you might think (PMID 40690366, PMID 33021339).
• At the same time, do not belittle what surrogate markers do. The predictions made by recurrence-free survival in melanoma adjuvant therapy turned out to be actually correct (PMID 28922786), bone mineral density thresholds were reproduced in separate trials (PMID 34490915), and albuminuria showed a consistent relationship across data from over 80,000 people (PMID 41198855). In ultra-rare diseases, without such markers, there are often no suitable endpoints that can be measured within a feasible timeframe (PMID 42525369).
• If you leave behind only one thing — a surrogate marker is a borrowed promise. In exchange for a short trial telling us about a long-term outcome in advance, we borrow the promise that "this marker can stand in for that outcome." Knowing how far that promise has been validated when you borrow it, versus borrowing it without knowing, are two entirely different things.

Source Papers

Study designs and cohorts are noted together. Numerical values present in the main text are not repeated.

How Widely Used It Is — PMID 31504034 (2019) Complete survey of 51 EMA accelerated pathway products and literature grading of each surrogate marker · PMID 39695642 (2024) 156 FDA accelerated approval drug-indication pairs · PMID 25034049 (2015) Overview of glaucoma biomarkers and surrogate endpoints.

Where It Betrayed Us — PMID 8780325 (1996) Increased mortality with antiarrhythmic drugs, placebo-controlled trials and meta-analysis of small trials · PMID 7654403 (1995) Contemporary overview summarizing changes in cardiology practice following that event.

Individual Level and Trial Level — PMID 38458207 (2024) Regulatory pooled analysis of 13 lung cancer immunotherapy trials (9,285 patients) · PMID 15615079 (2004) Overview where conclusions diverged by level of analysis in the bone mineral density-fracture relationship.

Thresholds and Composite Markers — PMID 17080751 (2006) Introduction of the surrogate threshold effect (STE) concept · PMID 32739473 (2020) Three checks for reading composite markers.

When You Capture Only Part of the Pathway — PMID 32707115 (2020) Individual patient data from 23 osteoporosis trials (91,779 patients) · PMID 22696079 (2012) Patient-level pooling of 10 pulmonary arterial hypertension regulatory submission trials (2,404 patients).

When It Operates Through a Different Pathway or Harms — PMID 32665641 (2020) Overview of benefit mechanisms of SGLT2 inhibitors.

The Conditionality of It — PMID 21060990 (2011) Meta-regression of 15 calcium and vitamin D supplementation trials (47,365 patients) · PMID 25863582 (2015) 101 randomized trials of advanced colorectal cancer.

Where Validation Succeeded — PMID 28922786 (2018) Individual patient data for resected melanoma adjuvant therapy, STE of 0.77 and post-hoc confirmation of its prediction · PMID 34490915 (2022) Bone mineral density threshold derivation and reproduction in a separate trial · PMID 30635226 (2019) 41 albuminuria trials (29,979 patients) · PMID 41198855 (2026) Expansion of the same question to 48 trials (85,681 patients).

Reality in Oncology — PMID 32214230 (2020) Methodology review of 91 progression-free survival surrogate validation studies · PMID 32702645 (2020) 260 phase 3 metastatic solid tumor trials (172,133 patients) · PMID 35777712 (2022) 2,109 decades of oncology randomized trials · PMID 38571449 (2024) 20 follicular lymphoma trials (10,724 patients) · PMID 30709633 (2019) 8 HER2-positive early breast cancer adjuvant trials (21,480 patients) · PMID 39031642 (2024) Post-hoc analysis of a metastatic prostate cancer phase 3 trial (831 patients), correlation with quality of life and skeletal events.

Pathology, Function, and Lab Values — PMID 34933868 (2021) Trial-level surrogacy of 54 early breast cancer neoadjuvant treatment trials (32,611 patients) · PMID 29957192 (2018) 31 heart failure exercise-based cardiac rehabilitation trials (4,784 patients) · PMID 39999862 (2025) Umbrella review of 20 statin placebo- or usual-care-controlled trials (194,686 patients) · PMID 34663316 (2021) Meta-regression of 18 cardiovascular outcome trials (161,156 patients) · PMID 22409774 (2012) 18 antihypertensive trials, STE and grading by outcome.

Regulatory Side — PMID 42525369 (2026) Ultra-rare neurological disease regulatory review and 4-stage validation model · PMID 40690366 (2025) Tracking of 77 confirmatory trials for converted oncology indications · PMID 40472969 (2025) Changes in conversion and withdrawal durations for accelerated approvals from 1992 to 2024 · PMID 33021339 (2020) Matched comparison of post-approval mandate fulfillment in the US and Europe · PMID 33035472 (2021) Preterm birth prevention injection confirmatory trial failure and advisory committee vote · PMID 42580367 (2026) Conceptual paper addressing the shift in regulatory evidence frameworks · PMID 41033457 (2025) International nephrology society research roadmap.