Surrogate Endpoint

Surrogate Endpoint in Trials and Health Technology Assessment

A surrogate endpoint lets a trial report a result long before the outcome that matters to patients has happened, for example a fall in blood pressure in place of fewer strokes, or delayed tumour growth in place of longer life. For health economists it leaves a gap, because a cost-utility model needs survival and quality of life, so the surrogate effect has to be converted through an estimated relationship. How far a reimbursement decision can rest on that conversion depends on evidence that treatment effects on the surrogate predict treatment effects on the final outcome, a stronger demand than showing that patients with better surrogate values do better. Neighbouring pages treat the clinical endpoint, the biomarker endpoint and accelerated approval.

What a surrogate has to capture

Prentice, writing in 1989, proposed a deliberately restrictive meaning, under which the true endpoint rate at any follow-up time is independent of treatment given the earlier history of the surrogate. The surrogate must carry the whole effect of treatment, which links the idea to mediation analysis and the causal pathway. NICE Decision Support Unit Technical Support Document 20 (TSD 20) notes that meeting the Prentice criteria does not guarantee a causal relationship between the treatment effects, and repeats the warning of Fleming and DeMets that "a correlate does not a surrogate make".

The ICH E9 guideline on statistical principles for clinical trials names two concerns. A surrogate may not truly predict the clinical outcome, because it may reflect one pharmacological mechanism and miss the treatment's other effects, positive or negative. It may also fail to give a measure of benefit that can be weighed against adverse effects. The Cardiac Arrhythmia Suppression Trial shows the first concern. Encainide and flecainide suppressed ventricular ectopy after myocardial infarction yet raised arrhythmic deaths, 43 on drug against 16 on placebo.

Levels of evidence for a surrogate relationship

Regulators and HTA bodies grade surrogates by the evidence behind them. ICH E9 ties the strength of evidence to biological plausibility, epidemiological evidence of prognostic value, and trials showing that treatment effects on the surrogate correspond to effects on the clinical outcome, and warns that a relationship found for one product need not hold for one with a different mode of action. In the United States, 21 CFR 314.510 allows approval on an effect on a surrogate that is reasonably likely to predict clinical benefit, with a requirement to study the drug further to verify that benefit where uncertainty remains.

The NICE manual (PMG36, section 4.6.6) uses a three-level hierarchy citing Ciani and colleagues. Level 3 is biological plausibility, level 2 is a consistent association between surrogate and final outcome, usually from observational studies, and level 1 is evidence from randomised controlled trials that the technology's effect on the surrogate corresponds to a commensurate effect on the final outcome.

Trial-level validation and the surrogate threshold effect

Level 1 validation needs many trials reporting effects on both outcomes. Buyse and colleagues separated individual-level agreement between the two endpoints from trial-level agreement between the treatment effects on them, and the trial-level model predicts the effect on the true endpoint from an observed effect on the surrogate. TSD 20 gives the standard surrogacy model of Daniels and Hughes, shown here in equivalent notation (TSD 20 writes the effects as $\delta$, the intercept and slope as $\lambda_0$ and $\lambda_1$ and the conditional variance as $\psi_2^2$).

$$\theta_{Ti} = \beta_0 + \beta_1 \theta_{Si} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \psi^2)$$

where $\theta_{Si}$ and $\theta_{Ti}$ are the true treatment effects on the surrogate and the final outcome in trial $i$, often log hazard ratios, $\beta_0$ is the intercept, $\beta_1$ is the slope and $\psi^2$ is the conditional variance left unexplained. TSD 20's surrogacy criteria are an intercept of zero, so that no effect on the surrogate implies no effect on the final outcome, a slope different from zero, and a conditional variance of zero for a perfect surrogate.

Burzykowski and Buyse noted that the trial-level coefficient of determination, $R^2_{trial}$, is hard to interpret and introduced the surrogate threshold effect, the smallest treatment effect on the surrogate needed to predict a non-zero effect on the true endpoint. TSD 20 prefers bivariate meta-analysis, a form of multivariate meta-analysis, to meta-regression, which treats the surrogate effect as a fixed covariate and ignores its uncertainty. It checks a candidate by take-one-out cross-validation, and for a valid surrogate the 95% prediction intervals should contain the observed estimate in at least 95% of trials. In its multiple sclerosis example, relapse rate predicted disability progression well in all but one of 25 studies.

How NICE expects surrogate evidence to enter a cost-utility model

PMG36 states that, for cost-utility analyses, endpoints reflecting how a patient feels, functions or how long a patient lives are more informative than surrogate outcomes (section 4.6.5). Where other outcomes are used to infer effects on mortality and quality of life, evidence for the relationship and an explanation of how it is quantified must be provided. A surrogate counts as validated only when the relative effect of a technology on it predicts the relative effect on the final outcome, preferably shown by a meta-analysis of randomised trials using the TSD 20 methods (section 4.6.7), although the level of evidence may be lower for some technologies for rare conditions, diagnostics and devices.

Validation is specific to the population and technology type (section 4.6.8), and carrying a relationship to a different class or mechanism of action must be thoroughly justified (section 4.6.9). Section 4.6.11 says the surrogate is most useful for estimating QALYs when it strongly predicts quality of life or survival, and that the uncertainty in the relationship should be quantified, included in probabilistic sensitivity analysis and can be explored further in scenario analysis. Taylor and Elston found 4 of 35 UK HTA Programme reports from 2005 and 2006 with a cost-effectiveness model based mainly on a surrogate, only one of them with level 1 validation evidence.

Worked example with a progression-free survival hazard ratio and an ICER

All numbers in this example are illustrative. A new cancer drug reports a hazard ratio of 0.70 for progression-free survival, with a standard error of 0.08 on the log scale, and no mature data on overall survival. A trial-level validation in the same class gives an intercept of 0.02, a slope of 0.60 and a residual standard deviation $\psi$ of 0.06, treated as known for simplicity.

1. Predict the effect on overall survival. The log hazard ratio for progression-free survival is ln(0.70) = minus 0.3567.

$$\hat\theta_T = 0.02 + 0.60 \times (-0.3567) = -0.1940$$

where $\hat\theta_T$ is the predicted log hazard ratio for overall survival. The predicted hazard ratio is exp(minus 0.1940) = 0.824, a smaller benefit than on the surrogate.

2. Attach a prediction interval. The prediction variance adds the residual variance to the slope squared times the variance of the new estimate.

$$SD_{pred} = \sqrt{0.06^2 + (0.60 \times 0.08)^2} = \sqrt{0.0059} = 0.0768$$

where $SD_{pred}$ is the standard deviation of the prediction. The 95% interval is minus 0.1940 plus or minus 1.96 × 0.0768, or minus 0.3446 to minus 0.0434 on the log scale, a hazard ratio of 0.708 to 0.958.

3. Find the surrogate threshold effect. For a very large trial only $\psi$ matters, and the upper limit falls below zero when, for a surrogate log hazard ratio x, 0.02 + 0.60x + 1.96 × 0.06 < 0, that is when x is below minus 0.1376 / 0.60 = minus 0.2293. The threshold hazard ratio is exp(minus 0.2293) = 0.795, so a surrogate hazard ratio above about 0.80 could not predict a survival benefit.

4. Carry the prediction into a cost-utility model. Survival on the comparator is exponential with a mean of 2 years, utility is 0.70, the incremental cost is GBP 15,000 and nothing is discounted. If the surrogate hazard ratio were wrongly applied to survival, mean survival would be 2 / 0.70 = 2.8571 years, a gain of 0.8571 life years and 0.6000 QALYs.

$$ICER_{naive} = \frac{15{,}000}{0.6000} = 25{,}000$$

where $ICER_{naive}$ is the incremental cost-effectiveness ratio in GBP per QALY gained. With the predicted hazard ratio, mean survival is 2 / 0.8237 = 2.4281 years, a gain of 0.4281 life years and 0.4281 × 0.70 = 0.2997 QALYs.

$$ICER_{pred} = \frac{15{,}000}{0.2997} = 50{,}050$$

where $ICER_{pred}$ is the ratio using the predicted survival effect.

5. Interpretation. Treating the surrogate effect as the survival effect halves the ICER. The naive figure sits at the lower edge of the GBP 25,000 to GBP 35,000 per QALY range over which, under PMG36 section 6.3.7, NICE committees refer explicitly to further factors, and the predicted figure sits well above it. At the upper prediction limit, a hazard ratio near 0.96, the QALY gain falls to about 0.06 and the ICER passes GBP 240,000.

Evidence that surrogate-based trials can mislead

Two meta-research studies help explain why HTA bodies treat surrogate results with caution. Ciani and colleagues compared 84 trials with surrogate primary outcomes and 101 with patient-relevant primary outcomes. The surrogate trials reported larger effects, with a ratio of odds ratios of 1.47 (95% CI 1.07 to 2.01), not explained by risk of bias.

Prasad and colleagues found that 34 of 65 trial-level correlations between a surrogate and overall survival in oncology (52%) were of low strength, at 0.7 or less. The CONSORT-Surrogate reporting extension recommends that trials with a surrogate primary outcome also collect the final outcome.

Surrogate, intermediate and final outcomes compared

A final outcome, such as death, stroke or a patient-reported outcome of symptoms, measures benefit directly. An intermediate outcome lies earlier on the disease pathway and becomes a surrogate only when it stands in for a final outcome. A biomarker endpoint is defined by what is measured and surrogacy by the use made of the result, so some surrogates, such as progression-free survival, are clinical events rather than biomarkers. A mediator explains how an effect arises within a study, whereas a surrogate is judged by how well it predicts effects across trials.

Misreadings that change a decision

Most errors come from treating one step in the argument as the whole. Each one changes the predicted benefit that reaches the economic model.

1. Reading a significant surrogate effect as proven benefit. The effect on the final outcome is a prediction with its own interval, as step 2 of the example shows.
2. Entering only the mean prediction. PMG36 asks for the uncertainty in the relationship to be included in the probabilistic analysis.
3. Equating regulatory acceptance with HTA acceptance. Approval on a reasonably likely surrogate carries a duty to verify benefit, and the remaining uncertainty is what managed entry agreements and value of information analysis address.

Sources

• Bujkiewicz S, Achana F, Papanikos T, et al. NICE DSU Technical Support Document 20: Multivariate meta-analysis of summary data for combining treatment effects on correlated outcomes and evaluating surrogate endpoints. Version 2. Decision Support Unit; 2019 (updated December 2022). https://www.sheffield.ac.uk/nice-dsu/tsds/multivariate-meta-analysis
• Burzykowski T, Buyse M. Surrogate threshold effect: an alternative measure for meta-analytic surrogate endpoint validation. Pharmaceutical Statistics. 2006;5(3):173-186. https://doi.org/10.1002/pst.207
• Buyse M, Molenberghs G, Burzykowski T, et al. The validation of surrogate endpoints in meta-analyses of randomized experiments. Biostatistics. 2000;1(1):49-67. https://doi.org/10.1093/biostatistics/1.1.49
• Ciani O, Buyse M, Drummond M, et al. Time to review the role of surrogate end points in health policy: state of the art and the way forward. Value in Health. 2017;20(3):487-495. https://doi.org/10.1016/j.jval.2016.10.011
• Ciani O, Buyse M, Garside R, et al. Comparison of treatment effect sizes associated with surrogate and final patient relevant outcomes in randomised controlled trials: meta-epidemiological study. BMJ. 2013;346:f457. https://doi.org/10.1136/bmj.f457
• Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo. The Cardiac Arrhythmia Suppression Trial. New England Journal of Medicine. 1991;324(12):781-788. https://doi.org/10.1056/NEJM199103213241201
• International Council for Harmonisation. ICH E9 Statistical principles for clinical trials. Step 4 version, 5 February 1998. Section 2.2.6. https://database.ich.org/sites/default/files/E9_Guideline.pdf
• Manyara AM, Davies P, Stewart D, et al. Reporting of surrogate endpoints in randomised controlled trial reports (CONSORT-Surrogate): extension checklist with explanation and elaboration. BMJ. 2024;386:e078524. https://doi.org/10.1136/bmj-2023-078524
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). Published 31 January 2022, last updated 13 October 2026. Sections 4.6.5 to 4.6.11 and 6.3.7. https://www.nice.org.uk/process/pmg36
• Prasad V, Kim C, Burotto M, et al. The strength of association between surrogate end points and survival in oncology: a systematic review of trial-level meta-analyses. JAMA Internal Medicine. 2015;175(8):1389-1398. https://doi.org/10.1001/jamainternmed.2015.2829
• Prentice RL. Surrogate endpoints in clinical trials: definition and operational criteria. Statistics in Medicine. 1989;8(4):431-440. https://doi.org/10.1002/sim.4780080407
• Taylor RS, Elston J. The use of surrogate outcomes in model-based cost-effectiveness analyses: a survey of UK Health Technology Assessment reports. Health Technology Assessment. 2009;13(8). https://doi.org/10.3310/hta13080
• US Code of Federal Regulations. 21 CFR 314.510, Approval based on a surrogate endpoint. eCFR, accessed 1 November 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-D/part-314/subpart-H/section-314.510