Evaluating a Biomarker as a Surrogate in a Treatment Case

What Does It Mean for a Biomarker to Be a Surrogate?

A biomarker can change with treatment and correlate with clinical outcomes without being a reliable substitute for measuring those outcomes. A useful evaluation therefore asks a more demanding question: across randomized treatment conditions, do treatments that produce larger biomarker effects also produce larger clinical benefits? The distinction matters because patients with favorable biomarker values may have better outcomes for reasons unrelated to the treatment effect on the biomarker.

The analysis below uses paired treatment-effect estimates from multiple doses, regimens, or independent studies. Exposure data can help characterize how treatment produces biomarker changes, but observed exposure is not automatically randomized: it can reflect adherence, disease severity, or other patient characteristics. The primary comparison is between treatment effects, not between patients selected for high versus low exposure.

Pairing is essential: \(x_j\) and \(y_j\) must refer to the same treatment contrast, population, follow-up period, and relevant estimand. For example, a biomarker effect at week 4 should not be paired casually with a clinical effect accumulated over a substantially different period. When exposure-response models are used to estimate treatment effects at particular exposures, the exposure range and the assumptions used to obtain those estimates should also be reported.

Building Comparable Treatment-Effect Pairs

Use a consistent benefit scale before comparing effects. Suppose a biomarker falls when treatment is beneficial and a clinical score rises when patients improve. The biomarker effect can be defined as the control mean minus the treated mean, while the clinical effect is treated minus control. Both are then positive for benefit. Reversing one sign without noticing can make a genuinely concordant pattern look negative.

This example also shows why patient-level exposure analyses and treatment-effect analyses answer different questions. A model relating individual exposure to individual outcome may be informative about exposure-response, but high-exposure patients are not necessarily exchangeable with low-exposure patients. For surrogate evaluation, retain the randomized treatment contrasts and make the source of each estimated effect explicit.

Regression Across Treatment Conditions

A simple exploratory model relates the clinical treatment effect to the biomarker treatment effect across \(J\) independent treatment conditions:

Here \(\alpha\) is the expected clinical treatment effect when the biomarker treatment effect is zero, and \(\beta\) is the expected change in clinical benefit per one-unit increase in biomarker benefit. A positive \(\beta\) indicates that conditions with larger biomarker effects tend to have larger clinical effects. Ordinary least squares estimates \(\alpha\) and \(\beta\) by minimizing the sum of squared residuals.

A conventional test of association uses \(H_0:\beta=0\). For a prespecified directional question, one may use \(H_a:\beta>0\); otherwise, a two-sided alternative is appropriate. Under the stated model, the slope standard error is \(SE(\hat{\beta})=s/\sqrt{S_{xx}}\), where \(s=\sqrt{\mathrm{SSE}/(J-2)}\), \(\mathrm{SSE}=\sum_j(y_j-\hat y_j)^2\), and \(S_{xx}=\sum_j(x_j-\bar{x})^2\). The test statistic is \(t=\hat{\beta}/SE(\hat{\beta})\), with \(df=J-2\) under \(H_0\).

The standard error follows from the least-squares slope formula. If the \(x_j\) values are treated as fixed and the errors are independent with variance \(\sigma^2\), then

Replacing \(\sigma\) with the residual standard deviation \(s\) gives the estimated standard error above. This derivation also shows a limitation: little spread in biomarker effects makes \(S_{xx}\) small and the slope imprecise, even if the fitted line appears visually steep.

Prediction and Its Uncertainty

The fitted mean clinical effect at a specified biomarker effect is not the same as the clinical effect expected for a new treatment condition. The latter includes both uncertainty in the estimated regression line and the residual variation of a new condition around that line. For a new condition with biomarker effect \(x_0\), a \(100(1-\alpha)\%\) prediction interval under the regression assumptions is

The leading 1 inside the square root is the new condition's residual variation. Omitting it produces an interval for the mean response at \(x_0\), which is narrower and answers a different question. A prediction interval applies only to a new condition comparable to those used to fit the model and within a range supported by the observed biomarker effects.

Diagnostics, Alternatives, and Limits

A positive slope and small \(p\)-value are not sufficient grounds for surrogate use. Inspect a scatterplot of the paired effects and residuals. Curvature can indicate that a linear relationship is inappropriate; a single influential condition can dominate a slope when there are few pairs; and changing residual spread can invalidate the usual standard errors and intervals. Compare the estimated slope with and without influential conditions as a sensitivity analysis, while reporting the full set of conditions rather than deleting points solely to improve the result.

Uncertainty in both axes is an especially important diagnostic concern. Each \(x_j\) and \(y_j\) is usually an estimate, not a known value. Ordinary least squares ignores that uncertainty. Substantial error in \(x_j\) can attenuate the estimated slope toward zero, while differing precision across conditions can make an unweighted fit inefficient or misleading. When available, use the estimated variances and covariance of paired effects in a suitable errors-in-variables or hierarchical analysis. With only a few conditions, however, more elaborate models can be weakly identified; transparent sensitivity analyses may be more informative than a falsely precise result.

The main alternatives answer related but distinct questions. Patient-level mediation analyses examine whether individual biomarker changes relate to individual clinical outcomes, often conditional on treatment and covariates; such associations alone do not establish treatment-level surrogacy. A meta-regression across randomized trials or dose arms targets the treatment-effect relationship directly, but needs enough independent conditions and careful handling of effect-estimate uncertainty. Mechanistic PK/PD models can describe exposure, biomarker response, and clinical response jointly, but model-based predictions still require validation against clinical outcomes and do not turn an association into proof of mediation.

A common mistake is to conclude, “Patients with larger biomarker changes did better, so the biomarker is a valid surrogate.” This confuses an individual-level association with the question of whether treatment effects on the biomarker predict treatment effects on the clinical outcome. Another mistake is to treat a significant cross-condition slope as proof that the treatment acts only through the biomarker. A treatment may affect clinical outcomes through pathways the biomarker does not capture, or a biomarker may track treatment effects in the studied range and fail for a different mechanism or population. A single case can support a provisional prediction claim; stronger surrogate claims require evidence across relevant treatments, settings, and clinical outcomes.