Biomarker Endpoint Planning Under Surrogate Uncertainty

Why Biomarker Precision Is Not the Same as Clinical Evidence

A biomarker can provide a timely, sensitive measure of biological activity and may therefore be useful for planning or evaluating a trial. But two questions must remain separate: How precisely can the trial estimate the treatment effect on the biomarker? And how strongly does that biomarker effect support a conclusion about a clinical outcome? Measurement error affects the first question. Uncertainty in the biomarker-to-clinical-outcome relationship affects the second.

This distinction matters in model-informed development. A small standard error for a biomarker contrast can show that the biomarker effect is estimated precisely; it does not by itself show that the treatment improves a clinical outcome. Such a clinical interpretation depends on evidence connecting treatment effects on the biomarker to treatment effects on the clinical outcome, including the uncertainty in that connection.

Definition: A biomarker endpoint is a trial endpoint based on a biological characteristic measured to assess a process, response, or effect of treatment. A biomarker is a surrogate endpoint for a clinical outcome only when evidence supports its use to predict the treatment effect on that clinical outcome in the relevant context. A biomarker may be useful without being an established surrogate.

Measurement Error and Biomarker Endpoint Precision

For a simple continuous-endpoint planning calculation, let \(B\) denote the underlying biomarker value and let the measured value be \(B^*=B+e\), where \(e\) is measurement error. Assume the error has mean zero, is independent of the underlying biomarker and treatment assignment, and has the same variance in both randomized groups. Under these assumptions, measurement error does not systematically shift the expected treatment contrast, but it increases the variance of individual measurements and therefore generally increases the standard error of the contrast.

The reliability of a measurement is the proportion of its total variance attributable to variation in the underlying biomarker rather than measurement error. In this simple model, it is the ratio below. The formula assumes the error variance is independent of the underlying biomarker.

\lambda = \frac{\operatorname{Var}(B)} {\operatorname{Var}(B)+\operatorname{Var}(e)}

For a two-arm trial with equal allocation, \(n\) participants per arm, and observed endpoint variance \(\sigma_{B^*}^2\) in each arm, the variance of the difference in sample means is \(2\sigma_{B^*}^2/n\). Thus its standard error is \(\sqrt{2\sigma_{B^*}^2/n}\). This planning expression assumes independent participants, a common variance in the arms, and no precision gain from additional covariate adjustment. If the planned estimand is a change score, the variance used must be the variance of that change score, including any covariance between its repeated measurements.

Worked Example: Planning for a Noisy Biomarker

Suppose the underlying biomarker has a standard deviation of 10 units, and the measurement error has a standard deviation of 6 units. Assume the error is mean-zero, independent, and nondifferential between the treatment groups. The observed variance is the sum of the underlying and error variances:

\sigma_{B^*}^2 = 10^2+6^2=136\ \text{units}^2, \qquad \lambda=\frac{10^2}{10^2+6^2} =\frac{100}{136}=0.7353

The reliability is about 73.5%. To target a standard error of 0.80 biomarker units for the treatment contrast, use the equal-allocation formula and solve for the number per arm:

0.80=\sqrt{\frac{2(136)}{n}} \quad\Longrightarrow\quad n=\frac{2(136)}{0.80^2} =\frac{272}{0.64}=425

The plan therefore requires 425 participants per arm under these assumptions. If measurement error were absent while the underlying variance remained \(100\ \text{units}^2\), the corresponding calculation would give \(n=2(100)/0.64=312.5\), or 313 per arm after rounding up. This comparison isolates the precision cost of the stated measurement error; it does not imply that the biomarker is or is not a valid surrogate.

Reliability is not a universal property of a biomarker divorced from its measurement context: it depends on the population variability and the measurement process. A low reliability can arise from substantial assay noise, restricted biological variation, or both. In trial planning, the quantity directly needed for the standard error is the observed endpoint variance in the intended population; reliability helps explain how much of that variance is measurement error and can guide assay improvements.

For example, if \(k\) independent replicate measurements are averaged and each has error variance \(\tau^2\), the average has error variance \(\tau^2/k\). This reduction requires independent replicate errors; correlated errors do not shrink at that rate. Replicates may improve endpoint precision, but the benefit must be weighed against collection burden, cost, and any change in the endpoint definition.

Worked Example: Replicate Measurements and Required Sample Size

Continue with an underlying biomarker variance of \(100\ \text{units}^2\) and single-measurement error variance of \(36\ \text{units}^2\). Suppose each participant has four independent replicate assays, averaged to form the endpoint. The error variance of the average is \(36/4=9\ \text{units}^2\), so the endpoint variance becomes:

\sigma_{\text{average}}^2=100+9=109\ \text{units}^2

The reliability of the averaged measurement is \(100/109=0.9174\), or about 91.7%. At 200 participants per arm, the standard error of the contrast is:

SE=\sqrt{\frac{2(109)}{200}} =\sqrt{1.09}=1.044\ \text{units}

For comparison, with one assay per participant at the same sample size, \(SE=\sqrt{2(136)/200}=\sqrt{1.36}=1.166\ \text{units}\). If the target remains an \(SE\) of 0.80 units, four replicates give \(n=2(109)/0.80^2=340.625\), so at least 341 participants per arm are needed. Replication improves precision, but it does not remove all biological variation or establish a link to a clinical outcome.

Propagating Uncertainty Into a Clinical-Effect Projection

Suppose a development team uses the treatment effect on a biomarker to project a treatment effect on a clinical outcome. Define \(\theta_B\) as the true treatment contrast on the biomarker, in biomarker units, and \(\theta_Y\) as the corresponding contrast on the clinical outcome, in clinical units. A simple projection model is:

\theta_Y=\gamma\theta_B+\delta

Here \(\gamma\) is the relationship between biomarker and clinical treatment effects, and \(\delta\) represents a possible clinical-effect component not captured by that relationship. The scale and direction of the contrasts must be defined consistently before applying this model. For instance, if higher values mean greater benefit for one endpoint but greater harm for the other, the direction must be harmonized or explicitly reflected in \(\gamma\). Both \(\gamma\) and \(\delta\) require evidence; treating them as known without justification would conceal surrogate uncertainty.

For a transparent planning calculation, suppose the biomarker estimate \(\hat{\theta}_B\) is unbiased, has variance \(V_B\), and is independent of \(\gamma\) and \(\delta\). Suppose also that \(\gamma\) and \(\delta\) are independent, with means \(g\) and 0 and variances \(v_\gamma\) and \(v_\delta\), respectively. The following derivation gives the variance of the projected clinical effect \(Z=\gamma\hat{\theta}_B+\delta\) over biomarker sampling and relationship uncertainty.

Find the projection mean.
Independence and unbiasedness give \(E(Z)=E(\gamma)E(\hat{\theta}_B)+E(\delta)=g\theta_B\).

Use the second moment of the product.
\(E(\gamma^2) = g^2+v_\gamma\), and \(E(\hat{\theta}_B^2)=\theta_B^2+V_B\). Independence therefore gives \(E[(\gamma\hat{\theta}_B)^2]=(g^2+v_\gamma)(\theta_B^2+V_B)\).

Subtract the squared mean and add residual uncertainty.
Since \(\delta\) is independent, \(\operatorname{Var}(Z)=E[(\gamma\hat{\theta}_B)^2]-[E(\gamma\hat{\theta}_B)]^2+v_\delta=(g^2+v_\gamma)V_B+v_\gamma\theta_B^2+v_\delta\).

The result separates three sources of uncertainty: biomarker sampling error, uncertainty in the relationship \(\gamma\), and the residual surrogate gap represented by \(\delta\). It assumes the specified independence structure and a mean-zero residual gap. In real applications, dependence between \(\gamma\) and \(\delta\), or other model features, may require additional covariance terms. A projection interval based on this variance is conditional on the model and its inputs; it is not a confidence interval for an observed clinical-endpoint contrast.

Worked Example: Projecting a Clinical Effect Under Surrogate Uncertainty

Suppose the estimated biomarker treatment effect is \(2.00\) biomarker units, with \(SE=0.50\) units, so \(V_B=0.25\ \text{units}^2\). For illustration, assume \(g=1.00\) clinical unit per biomarker unit, \(v_\gamma=0.080\) in squared slope units, and \(v_\delta=0.155\ \text{clinical units}^2\). Treat these inputs as independent and take the residual gap to have mean zero. The projected mean is:

E(Z)=g\hat{\theta}_B=1.00(2.00)=2.00\ \text{clinical units}

Using the variance expression, with the observed estimate substituted for the planning value of \(\theta_B\):

\operatorname{Var}(Z) =(1.00^2+0.080)(0.25)+(0.080)(2.00^2)+0.155 =0.270+0.320+0.155 =0.745\ \text{clinical units}^2

Thus \(SE(Z)=\sqrt{0.745}=0.8631\) clinical units. Under a normal approximation for the projection distribution, an approximate 95% projection interval is:

2.00\pm1.96(0.8631) \quad\Longrightarrow\quad [0.3083,\ 3.692]\ \text{clinical units}

This interval describes substantial uncertainty in the projected clinical effect under the stated model, even though the biomarker estimate itself has an \(SE\) of only 0.50 biomarker units. The interval does not say there is a 95% probability that a fixed clinical effect lies within those limits. It is a model-based projection interval whose interpretation depends on the assumptions and uncertainty inputs.

Precision Limits, Diagnostics, and Alternatives

Increasing the biomarker trial sample size reduces \(V_B\), but it cannot eliminate uncertainty in the biomarker-to-clinical-outcome relationship. In the example, as \(V_B\) approaches zero, the projection variance approaches:

v_\gamma\theta_B^2+v_\delta =(0.080)(2.00^2)+0.155 =0.475\ \text{clinical units}^2

The remaining standard deviation is \(\sqrt{0.475}=0.6892\) clinical units. This is an uncertainty floor under the assumed model and inputs, not a guarantee that the projection will be accurate. It shows why a very large biomarker trial may still be unable to support a narrow clinical-effect projection.

Worked Example: Checking Whether More Biomarker Precision Can Meet a Target

Suppose the planning goal is a normal-approximation 95% projection interval with a half-width no greater than 1.00 clinical unit. This requires a projection standard deviation no greater than \(1.00/1.96=0.5102\) clinical units, or variance no greater than:

\left(\frac{1.00}{1.96}\right)^2=0.2603\ \text{clinical units}^2

But the relationship-uncertainty floor in the preceding example is \(0.475\ \text{clinical units}^2\), which already exceeds \(0.2603\ \text{clinical units}^2\). Therefore, no increase in the biomarker trial sample size alone can attain the target under this model. The plan would need better evidence about the biomarker-to-clinical relationship, a smaller residual gap, a different endpoint strategy, or a less stringent precision goal. Any revised assumptions need empirical or scientific justification rather than being chosen merely to produce a narrower interval.

Diagnostics should assess both the measurement process and the projection model. For measurement, inspect assay repeatability, replicate disagreement, missing or out-of-range results, and whether error varies by treatment group, time, or biomarker level. Differential measurement error can bias a contrast rather than merely inflate its variance. Check that the variance used for planning matches the intended population, endpoint definition, and time point.

For surrogate projection, examine how the relationship was estimated: the number and range of relevant treatment effects, the clinical settings represented, and uncertainty in both biomarker and clinical estimates. A strong association among individual patients is not, by itself, proof that treatment effects on a biomarker predict treatment effects on the clinical outcome across interventions. Look for influential studies, poor calibration, residual patterns, and disagreement between predicted and observed clinical effects. If the estimated relationship changes substantially under reasonable modeling choices or exclusion of a study, the projection should be reported as fragile.

The main alternatives answer different questions. A biomarker endpoint can efficiently quantify biological response, but its clinical interpretation depends on surrogate evidence. A direct clinical endpoint measures the outcome of interest and avoids projecting through an uncertain relationship, although it may require more time, participants, or follow-up. Using both biomarker and clinical outcomes can help assess whether the biomarker predicts the clinical effect, but the trial must be planned to estimate each endpoint adequately. These choices concern endpoint evidence; a biomarker result alone should not be described as proof of clinical benefit.

Common Pitfalls and Practical Interpretation

Ignoring measurement error. Planning with the underlying variance when the observed endpoint contains assay error understates the standard error. Use the variance of the endpoint that will actually be analyzed.
Assuming high reliability validates a surrogate. Reliability concerns measurement quality; it does not establish that treatment effects on the biomarker predict clinical effects.
Treating the estimated slope as known. Plugging in a single \(\gamma\) while omitting its uncertainty can make a clinical projection look more precise than the evidence supports.
Equating individual-level association with treatment-effect prediction. Prediction of treatment effects requires evidence about the relationship between treatment effects, not merely correlation between biomarker and clinical values among individuals.
Reporting a projection as an observed clinical result. A biomarker-based projection is conditional on a model and its inputs. State that distinction, the clinical units, and the uncertainty assumptions.

Key takeaway: Plan biomarker precision using the observed endpoint variance, including measurement error. Then separately propagate uncertainty in the biomarker-to-clinical-outcome relationship. A precise biomarker contrast may still yield a wide clinical projection, and increasing the trial sample size cannot remove uncertainty in the surrogate relationship itself.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A biomarker has underlying variance \(64\ \text{units}^2\) and independent measurement-error variance \(16\ \text{units}^2\). What are its reliability and observed variance?
For a two-arm equal-allocation trial with observed biomarker variance \(100\ \text{units}^2\), what is the treatment-contrast standard error with 50 participants per arm?
Why does uncertainty in \(\gamma\) matter more when the biomarker treatment effect is larger?
In the projection model, what uncertainty remains as \(V_B\) tends to zero?
Does a highly reliable biomarker measurement establish that the biomarker is a surrogate endpoint? Explain.

Solutions. (1) Reliability is \(64/(64+16)=0.800\), and observed variance is \(80\ \text{units}^2\). (2) \(SE=\sqrt{2(100)/50}=\sqrt{4}=2.00\) biomarker units. (3) The relationship-uncertainty contribution is \(v_\gamma\theta_B^2\), so it increases with the squared size of the biomarker effect when \(v_\gamma\) is fixed. (4) Under the stated model and independence assumptions, the remaining variance is \(v_\gamma\theta_B^2+v_\delta\). (5) No. Reliability describes measurement precision relative to measurement error; surrogate status requires evidence that treatment effects on the biomarker predict treatment effects on the clinical outcome in the relevant context.