Measurement Error in Surrogate Validation

Why Measurement Error Matters for Surrogate Validation

A surrogate may appear weak because it is a poor substitute for the clinical outcome, but its observed relationship with that outcome can also be weakened by noisy measurement. Biomarker assays have analytical and biological variability, while treatment effects estimated in individual trials have sampling error. These are different sources of uncertainty: one concerns measurements on individuals, and the other concerns estimates used to compare trials.

The distinction matters for both levels of surrogacy. Individual-level assessments relate a participant’s biomarker value or change to that participant’s clinical outcome, usually accounting for treatment assignment and other specified covariates. Trial-level assessments relate treatment effects on the biomarker to treatment effects on the clinical outcome across trials. In either case, the quantity of interest is not necessarily the same as the association calculated from noisy observed values.

For an individual-level analysis, let \(B\) denote a latent biomarker value and \(W=B+U\) its measured value, where \(U\) is measurement error. Let \(Y\) be a clinical outcome. For a trial-level analysis, let \(\alpha_j\) and \(\delta_j\) denote the true treatment effects in trial \(j\) on the biomarker and clinical outcome, and let \(\hat{\alpha}_j\) and \(\hat{\delta}_j\) be their estimates. The hats indicate estimates, which vary because of sampling error even when the underlying effects are fixed.

Individual-Level Attenuation: Conditions and Derivation

Consider a linear outcome model \(Y=\beta B+\varepsilon\) within a defined treatment and covariate stratum, or after adjustment for those factors. Here \(\beta\) is the association of interest between the latent biomarker and the outcome. For the familiar attenuation result, assume \(E[U]=0\), \(U\) is independent of both \(B\) and \(\varepsilon\), and \(\operatorname{Cov}(B,\varepsilon)=0\). The last condition is essential: it holds, for example, if \(E[\varepsilon\mid B]=0\). The outcome model alone does not imply it.

Regressing \(Y\) on \(W\), rather than on \(B\), gives a population slope equal to \(\operatorname{Cov}(W,Y)/\operatorname{Var}(W)\). Under the stated conditions, the numerator and denominator are:

Thus the observed slope is:

The quantity \(\lambda\), between 0 and 1 under positive variances, is the reliability ratio for this measurement model. The observed slope is closer to zero than \(\beta\) when \(\beta\ne 0\). This is not a universal rule for every form of error: correlated error, error depending on the biomarker or outcome, and nonlinear models can produce bias in either direction. If error is instead added to the outcome while the predictor is measured without error, independent mean-zero outcome error does not bias the population slope, although it increases residual variability and reduces precision.

Trial-Level Error in Estimated Treatment Effects

Across trials, let \(\hat{\alpha}_j=\alpha_j+e_j\) and \(\hat{\delta}_j=\delta_j+u_j\), where \(e_j\) and \(u_j\) are estimation errors. Suppose these errors have mean zero, are independent of the true effects, and are independent across trials. If the two estimation errors are also uncorrelated within each trial, then adding them does not change the across-trial covariance:

However, the observed variances increase by the respective error variances. Consequently, the observed correlation and the ordinary regression \(R^2\) can be smaller than their latent counterparts. The effect is not guaranteed if estimation errors are correlated. For example, shared participants or related outcome definitions can induce covariance between \(e_j\) and \(u_j\); the observed covariance then includes that error covariance. Further, when trial estimates have different standard errors, a single common error variance is generally inappropriate.

Correction and Sensitivity-Analysis Strategies

Correction requires information about the error process. A reliability ratio can be estimated from replicate assays, validation data, or a justified variance-components model. If a classical measurement model is credible, a regression slope can be divided by the reliability ratio. For a trial-level \(R^2\), one can subtract the estimated sampling-error variances from the observed variances while retaining the observed covariance only when the estimation errors are uncorrelated with each other and with the true effects. Uncertainty in the estimated error variances must also be reflected in uncertainty for the corrected result.

Several approaches address these limitations, but they answer the problem in different ways:

For trial-level validation, a hierarchical model can represent the true effects as varying across trials and the estimated effects as noisy observations with known or estimated sampling variances. This is preferable to treating each point estimate as exact when precision differs substantially. It also makes clear that more trials, more precise treatment-effect estimates, or both may be needed to distinguish a weak surrogate relationship from a strong relationship obscured by noise.

Diagnostics, Failure Modes, and Interpretation

Before applying a correction, inspect assay replicates, standard errors of trial-specific estimates, and the covariance structure of the treatment-effect estimates. Plot estimated surrogate effects against estimated clinical effects with uncertainty information, not just point estimates. Look for a few imprecise trials dominating an unweighted regression, substantial variation in precision, or evidence that assay variability changes with concentration. These patterns challenge a constant-variance classical-error correction.

• Correlated errors: If \(U\) is correlated with \(B\) or \(\varepsilon\), the individual-level covariance formula changes. At trial level, correlated \(e_j\) and \(u_j\) alter the observed covariance itself. A simple reliability adjustment may then be biased in either direction.
• Systematic or differential error: A measurement that is shifted by treatment arm, outcome status, or trial can create spurious associations or conceal real ones. Mean-zero random-error formulas do not address this.
• Restricted range: If true biomarker values or treatment effects vary little in the analyzed sample, reliability and correlation can be low even with a well-performing assay. The resulting assessment may not transport to a setting with a different range.
• Uncertain correction: Corrected estimates can be highly unstable when reliability is low or estimated from few replicates. Report their uncertainty and assumptions, not only a corrected point estimate.

A common mistake is to observe a modest trial-level \(R^2\), assume that the surrogate is intrinsically poor, and ignore uncertainty in the trial-effect estimates. The correction is not to assume measurement error explains the result, either. Instead, quantify the likely error using standard errors or validation data, fit an appropriate errors-in-variables model when feasible, and show sensitivity to plausible alternatives. The introductory material on trial-level \(R\)-squared and the surrogate threshold effect concerns how those quantities are used; measurement-error analysis asks how reliably they have been estimated.

All such corrections remain conditional on the defined endpoint, population, treatment contrast, and analysis model. Measurement-error correction cannot establish that a biomarker captures all relevant treatment pathways, nor can it repair bias from a misspecified surrogate-outcome relationship. It addresses distortion due to a specified measurement process, not every source of uncertainty in surrogate validation.