Misclassification Correction For Biomarker-Defined Effects

Why Misclassification Changes a Subgroup Effect

Enrichment decisions and subgroup claims depend on knowing which participants truly have the biomarker of interest. An assay that sometimes labels a biomarker-positive participant as negative, or a negative participant as positive, mixes the true subgroups. Consequently, treatment effects calculated within the assay-reported groups generally do not equal the effects in the true biomarker groups. The direction and size of the distortion depend on prevalence as well as assay accuracy.

This tutorial considers a binary baseline biomarker and a treatment comparison on an additive scale: a difference in mean outcome, or a risk difference for a binary outcome. The correction is not generally a direct correction of odds ratios or hazard ratios; those measures do not combine as simple weighted averages.

Write \(A=1\) for assignment to the experimental treatment and \(A=0\) for control. For an outcome \(Y\), define the true subgroup treatment effect as \(\tau_b=E[Y(1)-Y(0)\mid B=b]\), where \(Y(a)\) is the potential outcome under assignment \(a\). For a continuous outcome, \(\tau_b\) is in outcome units; for a binary outcome, it is a risk difference.

Assumptions for the Basic Correction

The formulas below require more than a sensitivity and specificity number. First, the biomarker is measured at baseline, before treatment assignment can affect either the biomarker or its measurement. Second, treatment is randomized, so treatment assignment is independent of true biomarker status and potential outcomes. Third, the assay properties apply to the trial population and are the same across treatment assignments.

The central misclassification assumption is nondifferential measurement conditional on true biomarker status: given \(B\), the assay result \(W\) is independent of treatment assignment and of the potential outcomes. In particular, the assay error must not depend on prognosis or treatment response within a true biomarker group. Under this assumption, people classified into a given assay stratum are mixtures of the true groups with the same mixture proportions in both randomized arms. If assay error depends on outcome or treatment, the simple mixture correction need not hold.

The basic correction also assumes \(Se\) and \(Sp\) are known or have been estimated from suitable validation data, and that \(J>0\). When \(J=0\), the assay result contains no information about true status, so prevalence and subgroup effects cannot be recovered from it using these equations. A value of \(J\) near zero makes the correction highly unstable.

Correcting Biomarker Prevalence

The observed assay-positive proportion combines true positives correctly detected and true negatives incorrectly called positive. The law of total probability gives

Collecting the terms in \(p\) yields \(q=1-Sp+Jp\). Thus, when \(J\ne 0\), the prevalence implied by the observed assay-positive proportion is

The corrected value must be between 0 and 1. Equivalently, under the assumed assay properties, \(q\) must lie between the false-positive rate \(1-Sp\) and sensitivity \(Se\) when \(J>0\). A value outside that range is a warning of sampling variation, mismatched assay properties, or a violated model; it is not a reason to silently truncate the result and proceed as if the assumptions held.

Correcting Treatment Effects

To correct an additive treatment effect, first determine the true-status composition of each assay stratum. By Bayes' rule, the proportion truly positive among assay-positive participants is \(a=P(B=1\mid W=1)=Se\,p/q\). The proportion truly positive among assay-negative participants is \(r=P(B=1\mid W=0)=(1-Se)p/(1-q)\). The remaining proportions in those strata are truly negative.

Let \(\delta_+\) be the randomized treatment contrast among participants with \(W=1\), and let \(\delta_-\) be the corresponding contrast among participants with \(W=0\). Under the assumptions above, the two observed contrasts are mixtures of the true subgroup effects:

Subtracting the second equation from the first eliminates the common component of the effects and gives \(\delta_+-\delta_-=(a-r)(\tau_1-\tau_0)\). Therefore, if \(a\ne r\), the contrast between true subgroup effects is

Substitution into either mixture equation recovers the individual effects. For example, \(\tau_0=\delta_- - r(\tau_1-\tau_0)\), and \(\tau_1=\tau_0+(\tau_1-\tau_0)\). These equations correct both the estimated effect difference and the subgroup-specific effect levels; simply dividing the assay-positive effect by sensitivity would not do so.

Why the Effect Correction Works

The preceding subtraction is the key justification. In matrix form, the mixture equations are

The determinant of the mixing matrix is \(a(1-r)-r(1-a)=a-r\). If \(a-r\ne0\), the matrix is invertible, and the two true effects are identifiable from the two observed contrasts, provided the mixture weights are known. If \(a=r\), assay-positive and assay-negative strata have the same true-status composition. They then provide no distinct information with which to separate \(\tau_1\) from \(\tau_0\).

The connection between this determinant and assay accuracy can be made explicit. Using the Bayes expressions for \(a\) and \(r\), and \(q=Se\,p+(1-Sp)(1-p)\), gives

Thus, if both true groups are present and \(J>0\), the assay strata have distinct compositions. But when prevalence is near 0 or 1, or \(J\) is small, \(a-r\) can be small. Dividing by a small value amplifies sampling error and any error in the assumed assay properties.

Estimation, Diagnostics, and Alternatives

In practice, replace \(q\) with the observed assay-positive proportion and use estimates of sensitivity and specificity from validation data. Calculate \(\hat p\), then the estimated composition weights, and finally solve the mixture equations using the randomized treatment contrasts within assay strata. Keep the analysis on a prespecified additive estimand, with its units stated. The earlier material on population-specific estimands is relevant to choosing which population and treatment contrast the corrected effects are intended to describe.

A plug-in correction does not make uncertainty disappear. The uncertainty in the observed treatment contrasts, prevalence, sensitivity, and specificity all contribute to uncertainty in the corrected effects. A standard error that treats estimated assay properties as fixed can be too small. Suitable approaches include joint likelihood methods that incorporate validation data, or a bootstrap that resamples trial and validation participants according to the sampling design and recalculates the complete correction in each replicate. Report the source and precision of assay-performance estimates, not only their point values.

Diagnostics should include checking that the corrected prevalence lies in \([0,1]\), inspecting whether the assay validation population and laboratory procedures match the trial, and examining whether sensitivity or specificity plausibly varies by treatment arm, outcome-related factors, or participant characteristics. Large corrected effects, wide intervals, or estimates that change sharply under small changes in \(Se\) or \(Sp\) indicate instability. A negative corrected prevalence or an implausible effect is a model warning, not evidence that the formula should be adjusted informally.

The direct analysis by observed assay status is simpler and avoids relying on externally supplied assay accuracy, but estimates effects in assay-defined groups rather than true biomarker groups. Restricting analysis to assay-positive participants does not remove false positives or recover missed positives. A validation substudy can support correction or a likelihood model that propagates assay uncertainty; if the assay properties are unknown, the observed positive rate alone cannot identify prevalence and both subgroup effects. These approaches answer different questions and rely on different information.

The main limitation is the nondifferential-error assumption. If false negatives have different outcomes within the true-positive group, for example, then the assay-positive and assay-negative mixtures need not combine the same \(\tau_1\) and \(\tau_0\) in both treatment arms. The simple linear equations then fail. Further, the correction is specific to additive effects: a risk difference can be mixed by subgroup proportions, but an odds ratio or hazard ratio generally cannot be corrected by applying these same formulas. Finally, correcting a point estimate does not address multiplicity or selection from searching across many subgroups.