Analytical Uncertainty in Drug Concentration Results

Why Analytical Uncertainty Matters

A reported drug concentration is a measurement, not an exact reading of the patient’s underlying concentration. The assay can vary from one measurement to another, and it can consistently read above or below the value it is intended to measure. Those features matter especially when a result is close to a target boundary: a small analytical difference may change whether the reported value falls above or below that boundary.

The focus here is uncertainty introduced by the analytical measurement itself. It is distinct from changes in a patient’s concentration over time, uncertainty from sample timing, or uncertainty in interpreting pharmacokinetics. A precise description of assay performance can show whether a result clearly lies on one side of a boundary or whether analytical error makes that classification uncertain. It does not, by itself, determine the appropriate clinical action.

Definition: Analytical imprecision is random variation among repeated measurements of the same stable sample under specified assay conditions. Analytical bias is the systematic difference between the assay’s expected result and an appropriate reference value. The limit of quantification (LOQ) is the lowest concentration at which an assay has been validated to quantify with specified performance criteria for precision and bias. These are distinct properties: an assay can be precise but biased, or unbiased on average but imprecise.

A Measurement Model for Assay Error

Let \(C\) denote the concentration in a stable sample and \(Y\) a reported assay result. A simple model is

Y = C + b(C) + \varepsilon,

where \(b(C)\) is the assay bias at concentration \(C\), and \(\varepsilon\) is a random error with mean zero and standard deviation \(\sigma_A(C)\). This model assumes the sample is homogeneous, the assay conditions are specified, and the bias and imprecision are characterized at the concentration of interest. A normal approximation, \(\varepsilon \sim N(0,\sigma_A^2(C))\), can be useful when assay validation supports it. Assay variability may change with concentration, so a precision estimate from a very different concentration may not be suitable.

For repeated measurements of the same sample, imprecision is commonly summarized by the sample standard deviation \(s\). When the mean concentration is positive, the coefficient of variation (CV) expresses the standard deviation relative to the mean:

\mathrm{CV} = 100\% \times \frac{s}{\bar{x}}.

The CV is dimensionless; \(s\) and \(\bar{x}\) have the concentration units. Bias is assessed by comparing the mean result with an appropriate assigned reference value \(C_{\mathrm{ref}}\):

\text{estimated bias} = \bar{x} - C_{\mathrm{ref}}, \qquad \text{relative bias} = 100\% \times \frac{\bar{x}-C_{\mathrm{ref}}}{C_{\mathrm{ref}}}.

An estimate of bias is not automatically the true bias. It depends on the reference value being suitable, and its uncertainty may matter. In practice, assay performance is characterized using validation data across relevant concentrations, rather than inferred from one patient result.

Worked Example: Separating Imprecision From Bias

Suppose four replicate measurements of the same hypothetical sample are 9.6, 10.2, 9.8, and 10.4 mg/L. For this illustration, the assigned reference concentration is 10.5 mg/L and is treated as exact.

The mean is \(\bar{x}=(9.6+10.2+9.8+10.4)/4=40.0/4=10.0\) mg/L. The squared deviations from the mean sum to \(0.16+0.04+0.04+0.16=0.40\) \((\mathrm{mg/L})^2\). Thus the sample standard deviation is \(s=\sqrt{0.40/(4-1)}=\sqrt{0.1333}=0.365\) mg/L, and the CV is \(100\%\times 0.365/10.0=3.65\%\).

The estimated bias is \(10.0-10.5=-0.5\) mg/L, or \(100\%\times(-0.5)/10.5=-4.76\%\). The replicates show variation around their mean, while that mean is below the assigned reference. The standard deviation describes imprecision; the difference from the reference estimates bias. Neither statistic alone establishes assay performance across all concentrations.

Interpreting a Result Near a Target Boundary

Suppose an assay is adequately characterized as unbiased near concentration \(C\), with analytical error approximately normal and standard deviation \(\sigma_A\). For a fixed true concentration, a repeat result then follows \(Y \sim N(C,\sigma_A^2)\). If a target boundary is \(T\), the probability that a repeat result falls at or above the boundary, conditional on a specified concentration \(C\), is

P(Y \geq T \mid C) = 1-\Phi\left(\frac{T-C}{\sigma_A}\right),

where \(\Phi\) is the standard normal cumulative distribution function. This is a statement about the distribution of possible measurements if the concentration is fixed. It is not the probability that the patient’s concentration is above the boundary given one observed result.

When a result \(y\) is observed and the assay is unbiased with known analytical standard deviation \(\sigma_A\), the interval \(y \pm 1.96\sigma_A\) is a 95% measurement-error interval under the normal model. Equivalently, for a fixed true concentration, intervals formed this way from repeated measurements cover that concentration about 95% of the time. This interval is not a posterior probability statement about the particular patient’s concentration; it describes the coverage of a procedure under its assumptions.

Worked Example: A Result Close to a Lower Boundary

Consider a hypothetical target with a lower boundary of 10.0 mg/L. A measured concentration is 9.7 mg/L. Suppose assay validation supports negligible bias and a local analytical CV of 5.0% near this concentration. For a local approximation, take \(\sigma_A=0.050\times9.7=0.485\) mg/L.

If the fixed concentration were 9.7 mg/L, the chance that a repeat measurement would be at least 10.0 mg/L is \(1-\Phi((10.0-9.7)/0.485)=1-\Phi(0.619)\approx1-0.732=0.268\). Thus, under the stated model, about 27% of repeat results would be at or above the boundary. This is a repeat-measurement probability conditional on a specified concentration, not the probability that the patient is above target.

For the observed result, the 95% measurement-error interval is \(9.7\pm1.96(0.485)=9.7\pm0.951\), or approximately 8.75 to 10.65 mg/L. The interval includes the 10.0 mg/L boundary. Under these assumptions, the assay result alone does not sharply distinguish values below the boundary from values above it. The calculation does not establish which value is clinically most likely.

The distance from a result to a boundary, measured in analytical standard deviations, provides a useful diagnostic:

z_T = \frac{y-T}{\sigma_A}.

A large positive \(z_T\) means the result is many analytical standard deviations above the boundary; a large negative value means it is many standard deviations below. A value near zero signals that small analytical fluctuations could readily change the side on which a repeat result falls. This is a description of measurement robustness, not a universal dosing rule. The relevant tolerance depends on assay validation and the clinical context.

Quantification Limits and Results Below the LOQ

The LOQ is a performance limit, not a concentration at which the drug suddenly becomes absent. Below the LOQ, the assay may still produce a signal, but the method has not been shown to quantify it with the specified accuracy and precision. A result reported as “less than 2 mg/L” therefore means the concentration is below the assay’s validated quantification range, not that it equals zero or exactly 2 mg/L.

The limit of detection (LOD) is a different concept: it concerns whether a signal can be distinguished from background under specified conditions. Detection does not guarantee reliable quantification. LOQ and LOD depend on the method and validation criteria; neither is a universal biological threshold.

Worked Example: A Boundary Below the Quantification Limit

Suppose a hypothetical assay has an LOQ of 2.0 mg/L and reports a sample as “less than 2.0 mg/L.” The monitoring question is whether the concentration is below or above a boundary of 1.5 mg/L.

The report does not resolve that question. Concentrations of 1.0 mg/L and 1.8 mg/L are both below the LOQ, but they lie on different sides of the 1.5 mg/L boundary. Replacing the censored result with 2.0 mg/L, zero, or half the LOQ would invent a numerical value not established by the assay. The appropriate conclusion is that this method’s reported result cannot distinguish the two sides of this boundary. An assay with a sufficiently lower LOQ may be needed if that distinction is important.

Checking Assumptions and Comparing Approaches

Before using an uncertainty calculation, check that it matches the result being interpreted. The imprecision estimate should apply to the assay, concentration range, and relevant operating conditions. The sample should be stable and sufficiently homogeneous for repeated measurements to represent analytical variation. A normal approximation should be supported by validation or be recognized as an approximation. If bias is appreciable, an interval centered on the uncorrected result may be misleading; a justified bias correction and its uncertainty would need consideration.

Replicate measurements can estimate random imprecision when the sample remains stable and the errors are independent. Under those assumptions, averaging \(m\) replicates reduces the standard error of their mean to \(\sigma_A/\sqrt{m}\). This reduction applies to random error, not to a shared systematic bias. By contrast, validation against reference materials helps characterize bias and precision over a range of concentrations. A result below the LOQ is not made quantitatively reliable by simply applying the same numerical uncertainty formula used within the validated range.

These approaches answer different questions. Replicates characterize variability; reference comparisons characterize agreement with an assigned value; a validated LOQ marks the low end of quantitative performance. For a result near a boundary, an uncertainty interval or a repeat-measurement calculation can communicate how readily analytical error might change the classification. Neither replaces clinical interpretation, and a repeat measurement cannot correct a persistent assay bias. A method with better validated performance may be more informative when the boundary lies near or below the current method’s LOQ.

Common pitfall: “The assay CV is 5%, so a result just above the target boundary is definitely above target.” A CV describes relative random variation under specified conditions; it does not make an individual result exact, account automatically for bias, or establish the patient’s concentration. Compare the result’s distance from the boundary with the relevant analytical uncertainty, and state when the assay alone does not resolve the distinction.

Analytical uncertainty is only one part of concentration interpretation. A result can be analytically precise yet difficult to interpret for other reasons, and a result with modest analytical uncertainty can still be inconclusive when it lies very close to a boundary. Keeping imprecision, bias, and quantification limits separate makes that uncertainty visible without overstating what the assay can establish.

Check Your Understanding

Work through each question, then compare your reasoning with the solutions.

A hypothetical assay gives replicate results with a low CV, but their mean is consistently below an assigned reference value. Which assay property is low, and which is indicated by the reference comparison?
A result is 12 mg/L, a boundary is 10 mg/L, and the local analytical standard deviation is 1 mg/L. Calculate \(z_T\). What does its value say about the result’s distance from the boundary in analytical standard deviations?
An assay reports a sample as less than 3 mg/L. A decision boundary is 2 mg/L. Can the report establish which side of the boundary the concentration lies on? Why?
Under independent, unbiased analytical errors with standard deviation 0.6 mg/L, what is the standard error of the mean of four replicate measurements? Does averaging remove a shared bias?

Solutions. (1) The low CV indicates low imprecision; a mean consistently below the assigned reference indicates negative bias. These properties are distinct. (2) \(z_T=(12-10)/1=2\). The result is two analytical standard deviations above the boundary, but this calculation alone is not a clinical decision rule. (3) No. Values below 2 mg/L and values between 2 and 3 mg/L are both consistent with the report, so the concentration could lie on either side of the boundary. (4) The standard error is \(0.6/\sqrt{4}=0.30\) mg/L. Averaging reduces independent random error but does not remove a shared systematic bias.