Assay Error and Exposure-Metric Stability

Why Assay Error Affects Exposure Metrics Differently

A concentration assay does not report the true concentration without error. Even when sampling times are fixed and every scheduled sample is available, measurement error can change the calculated exposure metrics. Its effect depends on how the metric uses the observations: trapezoidal AUC combines concentrations across time, whereas observed Cmax selects the largest value and Tmax selects the time attached to that value.

This distinction matters when nearby concentrations are similar. A modest error at one sample may have a limited effect on the area, because that concentration contributes only part of the weighted profile. The same error can change which sample is largest, shifting Cmax to a different value and Tmax to a different time. The analysis below isolates assay error from sampling-grid limitations: the times are treated as fixed, and no samples are missing.

Definition: Let \(C_i\) be the true concentration at fixed sampling time \(t_i\), and let \(Y_i=C_i+\varepsilon_i\) be the measured concentration. Here \(\varepsilon_i\) is assay error. For a profile with \(n+1\) samples, observed Cmax is \(\max_i Y_i\), and observed Tmax is the sampling time \(t_i\) at which that maximum occurs. Trapezoidal AUC is calculated from the measured concentrations \(Y_i\), rather than the unobserved \(C_i\).

A useful starting model assumes that assay errors have mean zero, so \(E(\varepsilon_i)=0\), and that their variances and dependence are specified. Independence is a simplifying assumption, not a universal property of assays. Some assays have concentration-dependent variability, and shared calibration or batch effects can make errors correlated. The calculations must reflect the error model that is plausible for the measurements.

How Assay Error Propagates into Trapezoidal AUC

For sampling times \(t_0<t_1<\cdots<t_n\), the linear trapezoidal rule calculates AUC by adding the area of each trapezoid. It can also be written as a weighted sum of the sampled concentrations:

\widehat{\mathrm{AUC}}=\sum_{i=0}^{n} w_iY_i, \qquad w_0=\frac{t_1-t_0}{2},\quad w_i=\frac{t_{i+1}-t_{i-1}}{2}\ (1\le i<n),\quad w_n=\frac{t_n-t_{n-1}}{2}.

The weights have units of time; multiplying a concentration by a weight gives concentration-times-time, the units of AUC. Substituting \(Y_i=C_i+\varepsilon_i\) separates the calculated area into the trapezoidal area from the true sampled concentrations and a weighted assay-error contribution:

\widehat{\mathrm{AUC}} = \sum_{i=0}^{n}w_iC_i+\sum_{i=0}^{n}w_i\varepsilon_i.

If the assay errors have mean zero, the expected calculated AUC equals the trapezoidal area based on the true sampled concentrations. This does not imply that the trapezoidal area exactly equals the continuous-time area under the unknown concentration curve: that is a separate approximation issue. Assay error affects the precision of the calculated area, while the trapezoidal rule also has approximation error between sample times.

Result: Write \(\Sigma\) for the covariance matrix of the assay errors and \(w\) for the vector of trapezoidal weights. Under the stated error model, the assay-error variance of calculated AUC is \(w^\mathsf{T}\Sigma w\). If errors are independent, this reduces to \(\sum_{i=0}^{n}w_i^2\sigma_i^2\), where \(\sigma_i^2=\operatorname{Var}(\varepsilon_i)\). Thus a sample with a larger time weight can contribute more to AUC uncertainty, even if its assay-error variance is the same as that of other samples.

The independent-error result follows directly from the variance of a weighted sum. Variance of the weighted sum includes each weighted variance and, when errors are correlated, covariance terms:

\operatorname{Var}\left(\sum_i w_i\varepsilon_i\right) = \sum_i w_i^2\operatorname{Var}(\varepsilon_i) + 2\sum_{i<j}w_iw_j\operatorname{Cov}(\varepsilon_i,\varepsilon_j).

When errors are independent, each covariance is zero and only the sum of weighted variances remains. Positive correlation can increase AUC variance; negative covariance can reduce it. Treating correlated errors as independent can therefore misstate the precision of an exposure estimate.

Worked Example: AUC Error With Unequally Spaced Samples

Suppose hypothetical samples are collected at \(t=0,1,3\) hours, with measured concentrations \(Y=(0,4,2)\) mg/L. Assume independent, mean-zero assay errors with standard deviation 0.20 mg/L at each time. Calculate the trapezoidal AUC and its assay-error standard deviation.

The trapezoidal area is \(1(0+4)/2+2(4+2)/2=2+6=8.00\) mg·h/L. The corresponding weights are \(w_0=0.5\) h, \(w_1=(3-0)/2=1.5\) h, and \(w_2=1.0\) h.

Because all three error standard deviations are 0.20 mg/L, each error variance is \(0.20^2=0.0400\) (mg/L)\(^2\). The AUC error variance is \(0.5^2(0.0400)+1.5^2(0.0400)+1.0^2(0.0400)=0.140\) (mg·h/L)\(^2\).

The standard deviation is \(\sqrt{0.140}=0.3742\) mg·h/L. This describes the standard deviation of the assay-error contribution across repeated measurements under the assumed model; it is not a confidence interval for the true continuous-time AUC.

The example illustrates why the time pattern matters. The middle sample has weight 1.5 h, so its error is multiplied more strongly in AUC than the error at either endpoint. A larger weight does not mean the middle measurement is intrinsically less accurate; it means its measurement error has greater leverage on this particular weighted sum.

Peak Selection: Cmax and Tmax

Unlike AUC, observed Cmax is not a fixed weighted sum. It is selected after measurement as the largest of the observed concentrations. As a result, errors can affect both the size of the selected value and its rank among samples. The chance of a change in peak selection is particularly relevant when two true concentrations are close relative to their measurement error.

Worked Example: Probability of Reversing Two Nearby Concentrations

At two hypothetical sampling times, suppose the true concentrations are 5.0 and 4.8 mg/L. Assume independent, mean-zero normal assay errors with standard deviation 0.15 mg/L at each time. What is the probability that the second measured concentration exceeds the first?

Let the measured values be \(Y_1=5.0+\varepsilon_1\) and \(Y_2=4.8+\varepsilon_2\). The second exceeds the first when \(\varepsilon_2-\varepsilon_1>0.2\) mg/L. Independence gives the difference an error standard deviation of \(\sqrt{0.15^2+0.15^2}=\sqrt{0.045}=0.2121\) mg/L.

Therefore, the reversal probability is \(P(\varepsilon_2-\varepsilon_1>0.2)=1-\Phi(0.2/\sqrt{0.045})\). Using the unrounded standard deviation, \(0.2/\sqrt{0.045}=0.9428\), and the probability is approximately 0.1729. Under these assumptions, there is about a 17.29% chance that assay error alone makes the lower true concentration appear higher.

If each concentration is instead the average of two independent replicate assays, the error standard deviation of each average is \(0.15/\sqrt{2}=0.1061\) mg/L. The standard deviation of the difference is then \(\sqrt{0.1061^2+0.1061^2}=0.1500\) mg/L, giving a reversal probability \(1-\Phi(0.2/0.15)\approx0.09121\). Replication reduces this ranking uncertainty under the assumed independence and unchanged assay process.

This pairwise calculation is useful for understanding instability, but a profile can have more than two candidate peaks. With multiple measurements, each observed value competes with all the others, and the probability that a particular time is selected depends on the full set of true concentrations and the joint error distribution.

Result: If each measured concentration is unbiased, selecting the largest measured value can nevertheless bias observed Cmax upward. For each possible error realization, \(\max_i(C_i+\varepsilon_i)\ge C_j+\varepsilon_j\) for every \(j\). Taking expectations gives \(E[\max_i(C_i+\varepsilon_i)]\ge C_j\) when \(E(\varepsilon_j)=0\). Choosing \(j\) as the index of the largest true concentration shows that the expected observed maximum is at least the true maximum among the sampled times. The amount of this selection effect depends on the number and spacing of competing concentrations and on the error distribution.

Tmax is an index-selection statistic: it is the time of the selected maximum, not a direct measurement of the underlying continuous-time peak. If two neighboring concentrations are close, a small error can change the selected time. If the maximum is well separated from all other sampled concentrations relative to the error, the selection is more stable. This assay-error issue is distinct from the limitation that a sampling schedule may not include the actual peak time.

Worked Example: A Peak-Time Shift With a Small AUC Change

Consider a hypothetical profile sampled at 0, 1, 2, and 3 hours, with true concentrations \(1.0, 5.0, 4.8,\) and \(2.0\) mg/L. Suppose one assay realization has errors \(0, -0.4, +0.5,\) and \(0\) mg/L. Compare the measured peak time and AUC with those based on the true sampled concentrations.

The measured concentrations are \(1.0, 4.6, 5.3,\) and \(2.0\) mg/L. The true sampled maximum is 5.0 mg/L at 1 hour, whereas the observed maximum is 5.3 mg/L at 2 hours. Thus this realization shifts observed Tmax by 1 hour and raises observed Cmax by 0.3 mg/L.

For equal one-hour spacing, the trapezoidal weights are \(0.5,1,1,0.5\) h. The AUC change is the weighted sum of the errors: \(0.5(0)+1(-0.4)+1(0.5)+0.5(0)=0.10\) mg·h/L. The true-sampled trapezoidal AUC is \(0.5(1+5)+0.5(5+4.8)+0.5(4.8+2)=11.30\) mg·h/L; the measured AUC is therefore 11.40 mg·h/L.

Here peak selection changes substantially in time, while the AUC changes by only 0.10 mg·h/L. This single realization does not show that AUC is always unaffected or that Tmax is always unstable. It demonstrates how an error can reorder nearby concentrations while entering the integrated metric as a weighted contribution.

Assumptions, Diagnostics, and Alternatives

Before applying these calculations, check whether the error model is credible. The zero-mean assumption means measurements are not systematically high or low at the relevant concentrations. The constant-variance model assumes the same error variance across concentrations; this may be inappropriate when assay variability is proportional to concentration. Independence assumes no shared error component across samples. The fixed-time setup assumes that the recorded times themselves are sufficiently accurate for the intended analysis.

Diagnostics should focus on the measurement process as well as the derived metrics. Replicate measurements can reveal within-sample variability and concentration-dependent spread. Assay quality-control results and calibration information can reveal systematic bias or batch effects. For the profile, examine the gap between the highest and next-highest measured concentrations, the time interval between those samples, and whether small changes within plausible assay-error ranges alter Cmax or Tmax. A narrow peak margin is a warning about ranking stability, not proof that the assay failed.

When assay variability changes with concentration, use concentration-specific variances rather than a single common standard deviation. When errors share a batch or calibration component, include covariance terms or analyze an appropriate error model. Replicate assays can reduce independent random error in an average, but do not necessarily remove systematic bias or shared error. A sensitivity analysis can recalculate the metrics after perturbing measurements within plausible error ranges; its results depend on the chosen error ranges and should be reported as such.

Approach	What it addresses	Main limitation
Direct trapezoidal calculation	Produces AUC from the observed profile	Does not quantify assay uncertainty by itself
Variance propagation	Quantifies AUC assay-error variance under a specified covariance model	Depends on correct error variances and covariances
Replicate assay or sensitivity analysis	Evaluates repeatability or metric changes under plausible perturbations	Does not automatically eliminate systematic error or resolve sparse sampling

These approaches answer different questions. The direct calculation gives the metric for the measured profile. Variance propagation describes AUC uncertainty under explicit assumptions. Replication and sensitivity analysis help assess robustness to measurement variation. None of them can establish an unobserved continuous-time peak from the sampled concentrations alone, and none substitutes for checking whether the assay is biased or the sampling grid is adequate.

Common Mistake

A common mistake is to assume that because zero-mean assay error does not bias a linear trapezoidal sum, it cannot bias any exposure metric. The conclusion does not extend to a selected maximum. AUC is a weighted sum, so zero-mean errors preserve its expectation relative to the true-sampled trapezoidal area. Cmax is the maximum selected after errors are added; selection favors whichever observation happens to have a positive error. Therefore unbiased concentration measurements do not guarantee an unbiased observed Cmax.

A second error is to interpret a changed Tmax as evidence that the underlying absorption peak moved. A change in the largest observed sample can arise from assay noise, and even error-free sampled concentrations identify only the largest value on the sampling grid. Interpret the selected time in the context of both measurement precision and sampling resolution.

Check Your Understanding

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

At times 0, 2, and 5 hours, what are the trapezoidal weights? If independent assay errors all have standard deviation 0.10 mg/L, calculate the assay-error standard deviation of AUC.
Two true concentrations differ by 0.30 mg/L. Independent normal assay errors at both times have standard deviation 0.20 mg/L. What is the probability that the lower true concentration is measured as higher?
Why can expected observed Cmax exceed the true maximum among sampled concentrations when every concentration assay is unbiased?
A small perturbation changes the selected maximum from 4 hours to 5 hours, while AUC barely changes. Is this internally inconsistent? Explain.

Solutions. (1) The weights are \(1\), \(2.5\), and \(1.5\) h. The variance is \(0.10^2(1^2+2.5^2+1.5^2)=0.0950\) (mg·h/L)\(^2\), so the standard deviation is \(\sqrt{0.0950}=0.3082\) mg·h/L. (2) The error-difference standard deviation is \(\sqrt{0.20^2+0.20^2}=0.2828\) mg/L. The reversal probability is \(1-\Phi(0.30/0.2828)=1-\Phi(1.0607)\approx0.1444\). (3) The maximum is selected after errors are added, favoring observations with positive error; the maximum of the noisy values has expectation at least as large as the true maximum at the sampled times under mean-zero errors. (4) No. AUC is a weighted sum of all sampled concentrations, while Tmax is the time attached to the largest one. A small perturbation can reorder two nearby concentrations and change Tmax while producing only a small weighted change in AUC.