Why a Checklist Is Needed
A concentration-time data set records measurements, not the underlying concentration path itself. Variation in an observed value can therefore have different explanations: the subject’s concentration may have changed unpredictably, or the assay may have measured a concentration imperfectly. Those explanations lead to different predictions and different model components. In practice, a plot of scattered observations rarely settles the question by itself.
The decision is best made by combining evidence. Ask whether variation persists across time, whether same-time replicates disagree, whether its size changes with concentration, and what quantity the analysis is meant to predict. Each answer is informative only under stated assumptions. This checklist supports model choice; it does not turn an observational pattern into proof of a mechanism.
The equation is a working decomposition, not a claim that every real error is additive or independent. For the diagnostic results below, suppose initially that observation errors have mean zero, are independent of the latent path, and are independent across distinct assays, with variance \(\sigma_\epsilon^2\). Replicates must be taken sufficiently close together that the latent concentration is effectively unchanged between them. If these assumptions are doubtful, the checklist still helps frame the question, but the simple calculations need not apply.
The Four Questions
After accounting for known dosing history and subject-level parameter differences, do deviations at one time tend to be associated with deviations later? Persistence is compatible with process noise, but can also arise from unmodeled subject effects, time-varying covariates, or correlated assay errors.
Replicate assays of the same specimen, or samples taken close enough to share the same latent concentration, can expose observation error. Replicates do not reveal how the concentration evolved before or after that time.
Compare residual spread across the concentration range. Roughly constant standard deviation is consistent with an additive observation-error model; roughly constant coefficient of variation is consistent with a proportional model. Either pattern can also arise from concentration-dependent process variability or model misspecification.
For a latent concentration prediction, observation error is not part of the target. For a future reported assay result, both uncertainty in the concentration and uncertainty in its measurement matter.
Two Useful Results Behind the Checklist
Replicates and cross-time associations are useful because their expected variability differs under the decomposition. The following results state exactly what the simple assumptions allow us to conclude.
To prove the result, subtract two measurements: \(Y_r-Y_s=(X+\epsilon_r)-(X+\epsilon_s)=\epsilon_r-\epsilon_s\). The shared latent concentration cancels, so the difference has expectation zero. Independence gives \(\operatorname{Var}(\epsilon_r-\epsilon_s)=\operatorname{Var}(\epsilon_r)+\operatorname{Var}(\epsilon_s)=2\sigma_\epsilon^2\). For the mean, \(\bar Y=X+m^{-1}\sum_r\epsilon_r\); conditional on \(X\), its variance is \(m^{-2}\sum_r\operatorname{Var}(\epsilon_r)=m^{-2}m\sigma_\epsilon^2=\sigma_\epsilon^2/m\). Thus, replicate averaging reduces the measurement contribution to variance, but it does not remove uncertainty in the latent concentration across subjects or times. \(\square\)
Expand the covariance using bilinearity: \(\operatorname{Cov}(Y(t),Y(s))=\operatorname{Cov}(X(t),X(s))+\operatorname{Cov}(X(t),\epsilon(s))+\operatorname{Cov}(\epsilon(t),X(s))+\operatorname{Cov}(\epsilon(t),\epsilon(s))\). Independence of each error from the latent path makes the two mixed covariances zero, and independence of the distinct-time errors makes the last covariance zero. Only the latent-path covariance remains. The conclusion fails if, for example, errors are serially correlated or persistent subject effects remain in the residuals. \(\square\)
Worked Examples
Worked Example: Same-Time Replicate Assays
Four assays of one specimen give concentrations \(9.6,\ 10.0,\ 10.2,\ 10.2\) mg/L. Assume the specimen represents one shared concentration and that replicate errors are independent with a common variance.
The replicate mean is \((9.6+10.0+10.2+10.2)/4=40.0/4=10.0\) mg/L. The deviations from the mean are \(-0.4,\ 0,\ 0.2,\ 0.2\) mg/L. Their squared sum is \(0.16+0+0.04+0.04=0.24\). The usual sample variance is therefore \(0.24/(4-1)=0.08\) \((\mathrm{mg/L})^2\), and the sample standard deviation is \(\sqrt{0.08}\approx0.283\) mg/L.
The disagreement among these replicates is evidence about assay-level variability for this specimen; it is not evidence that the subject’s concentration changed four times. The estimate is imprecise with only four replicates, and it describes the assay conditions represented by this specimen. It does not establish that the error variance is constant at other concentrations or in other batches.
Worked Example: Persistent Deviations Across Time
After a reference concentration model and measured subject characteristics have been accounted for, four subjects have paired residuals at an early and a later time: \((-0.6,-0.4)\), \((0.6,0.4)\), \((-0.3,-0.2)\), and \((0.3,0.2)\) mg/L. The residual means at both times are zero. Calculate the average cross-product.
The four cross-products are \(0.24,\ 0.24,\ 0.06,\ 0.06\) \((\mathrm{mg/L})^2\). Their average, using divisor \(4\), is \((0.24+0.24+0.06+0.06)/4=0.60/4=0.15\) \((\mathrm{mg/L})^2\), which is positive. The pattern is consistent with deviations that persist across time. Under the assumptions of Result 2, independent observation error cannot by itself create this expected cross-time covariance.
This is a clue, not a classification by itself. A persistent subject-specific clearance difference, a systematic omitted covariate, or correlated errors could also produce related residuals. The interpretation as within-subject process noise is more credible only after plausible persistent subject effects and data-processing sources have been considered.
Worked Example: Concentration-Dependent Spread
Suppose residual standard deviations from replicate or assay-validation data are approximately \(0.10,\ 0.50,\ 1.00\) mg/L at concentrations \(2,\ 10,\ 20\) mg/L, respectively. Compare the additive and proportional descriptions.
The coefficients of variation are \(0.10/2=0.05\), \(0.50/10=0.05\), and \(1.00/20=0.05\), or 5% at each level. The absolute standard deviation is not constant, but the relative spread is. A proportional observation-error model, with standard deviation about \(0.05X\), is therefore a reasonable candidate. For example, at \(10\) mg/L it predicts \(0.05(10)=0.50\) mg/L, matching the stated spread.
This pattern describes how measurement variability changes with concentration in these data; it does not prove that all concentration-dependent scatter is assay error. If the variability is in the latent concentration process instead, a state-dependent process model may be needed. Replicate design and the prediction target help determine which interpretation is appropriate.
Using the Answers to Choose a Model
The checklist is a way to organize model comparison, not a scoring rule. A simple observation model may be adequate when independent replicates show appreciable measurement error, residual spread has a credible assay-based form, and there is little unexplained persistence after subject effects are handled. A process-noise component becomes more plausible when unexplained deviations persist across time in a way consistent with the proposed concentration dynamics, and that persistence cannot be accounted for by stable subject differences or correlated measurement errors. Both components may be needed when the evidence supports both.
The prediction target makes the distinction operational. If the target is the latent concentration \(X(t)\), its uncertainty concerns the possible concentration paths and parameter uncertainty; the assay error is not itself part of that latent quantity. If the target is a future assay \(Y(t)\), then under independent additive error the predictive variance includes both contributions:
This follows by expanding the variance of \(X(t)+\epsilon(t)\): independence makes the covariance term zero. A prediction interval for measured concentrations should not silently omit the observation component, while a statement about latent concentration should not treat assay noise as biological change.
Diagnostics, Alternatives, and Limits
Useful diagnostics should be matched to the question being asked. Inspect replicate differences by concentration, batch, and specimen type; a changing spread or batch pattern warns against one common error variance. Examine residuals at separated times after accounting for subject-level effects; unexpected temporal association suggests that the current decomposition is incomplete. Compare observed and predicted variability on both the concentration scale and, where suitable, a relative or transformed scale. A pattern that disappears only after a transformation may indicate that an additive error model is poorly matched to the data, but it does not identify a unique mechanism.
The main alternative to assigning persistent variation to process noise is to represent stable between-subject heterogeneity, such as differences in individual PK parameters. Such heterogeneity can create correlations over time even if each subject follows a deterministic path conditional on those parameters. Likewise, correlated assay errors violate the independence assumption used in the covariance result. Sparse sampling, few replicates, and a narrow concentration range may leave these explanations difficult to distinguish. Do not force a source label when the design does not identify it; report the ambiguity and check whether key predictions change across plausible models.
- Warning sign: Replicate differences vary by batch or concentration, but the model assumes one common error variance.
- Warning sign: Residual persistence is called process noise before stable subject differences or time trends are considered.
- Warning sign: Replicate measurements are treated as independent even though they share specimen preparation or assay conditions.
- Warning sign: A model is selected from residual appearance alone, without checking the scientific prediction target.
No single diagnostic has universal priority. Replicate assays are strongest for same-time measurement variability when their errors are independent; time-separated measurements are informative about persistence when stable effects and correlated errors are addressed; concentration-stratified residuals assess variance structure; and the target defines which components must enter a prediction. These sources of evidence are complementary rather than interchangeable.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Three same-specimen replicates are \(4.8,\ 5.0,\ 5.2\) mg/L. Find their mean and sample variance. What source of variability do their differences directly inform?
- Under the assumptions of Result 2, why does independent measurement error not create expected covariance between distinct-time observations?
- At concentrations \(5\) and \(15\) mg/L, residual standard deviations are \(0.25\) and \(0.75\) mg/L. Is the spread more consistent with constant absolute SD or constant CV? Show the calculation.
- Residuals at two times are positively associated. Give two explanations other than within-subject process noise.
- For a future assay result, what variance components should be included if latent concentration uncertainty and independent additive assay error are both present?
Solutions. (1) The mean is \((4.8+5.0+5.2)/3=15.0/3=5.0\) mg/L. Squared deviations sum to \(0.04+0+0.04=0.08\), so the sample variance is \(0.08/(3-1)=0.04\) \((\mathrm{mg/L})^2\). Replicate differences directly inform observation error under the shared-concentration and independent-error assumptions. (2) Expanding the covariance leaves the latent-path covariance plus two mixed covariances and the covariance of the errors. Independence makes all three error-involving terms zero. (3) The absolute SDs differ, \(0.25\) versus \(0.75\) mg/L, while the CVs are \(0.25/5=0.05\) and \(0.75/15=0.05\). The pattern is consistent with constant 5% CV, not constant absolute SD. (4) Stable subject-specific PK differences and serially correlated assay errors are two alternatives; an omitted time-varying covariate is another. (5) Include both latent concentration variance and observation-error variance: with independent additive error, \(\operatorname{Var}(Y(t))=\operatorname{Var}(X(t))+\sigma_\epsilon^2\).