Why “Known” PK Inputs Can Give a False Sense of Precision
A sequential PK/PD workflow first estimates pharmacokinetic quantities and then uses their fitted values in a pharmacodynamic analysis. For example, an individual PK fit may produce an estimated concentration at each PD observation time. Those concentrations are then supplied to a PD model as if they were known exactly. This is convenient, and it may be a deliberate approximation, but it changes what the reported PD uncertainty means.
The PD standard error from such a fit generally describes uncertainty conditional on the supplied PK estimates. It does not automatically include uncertainty because the PK parameters and concentrations were estimated from data. Thus a narrow conditional confidence interval need not imply that the complete PK/PD analysis has equally narrow uncertainty.
This distinction is about uncertainty, not necessarily about whether the point estimate is biased. Plugging in an estimate can affect point estimates in some settings, but the central issue here is that treating an estimated quantity as fixed suppresses variation that would occur if the PK data and PK fit were repeated. The size and direction of the resulting uncertainty error depend on the estimand, the data, and the dependence between stages.
What the Conditional Standard Error Includes—and Omits
Suppose \(Y\) denotes the PD data, \(X\) the PK data, \(\hat{\eta}\) the fitted PK quantities, and \(\hat{\theta}\) a PD parameter estimate. A common sequential analysis estimates \(\hat{\theta}\) using \(Y\) and the fixed input \(\hat{\eta}\). Its standard error describes the variability of the PD estimate under the fitted PD model with those inputs held fixed. It may reflect PD residual variability and uncertainty in estimating PD parameters, as specified by that PD fit.
It does not, merely by being a standard error, include the change in the PD result that would arise if a new PK data set produced a different \(\hat{\eta}\). Possible omitted components include uncertainty in population PK parameters, uncertainty in an individual’s PK parameters or random effects, uncertainty in concentrations calculated from those estimates, and uncertainty in summaries such as exposure. If the components are statistically dependent, their covariance can matter too.
Conversely, the plug-in analysis does not necessarily omit every source of uncertainty. It may include PD residual error and conditional uncertainty in PD parameters. It may also include some PK-related uncertainty if that uncertainty was explicitly represented in the PD model or passed forward by another procedure. The analyst must inspect what was actually held fixed rather than infer coverage from the label “two-stage.”
A precise way to express the distinction is with the law of total variance. For any estimate or target \(\hat{\theta}\) and fitted PK information \(\hat{\eta}\),
The first term is average variability conditional on the PK fit. The second is variability in the conditional mean of the PD result as the PK fit changes. A plug-in standard error commonly represents the first term at the one observed \(\hat{\eta}\), rather than the full unconditional variance. The identity does not mean that the second term is always positive and large enough to make a particular reported interval too narrow: conditional variances can vary, estimates can be dependent, and the target may not change appreciably across plausible PK fits. It does show why a conditional standard error alone cannot establish that all relevant uncertainty has been included.
Worked Examples
Worked Example: Conditional Versus Total Uncertainty in an Exposure Effect
Hypothetical data. A linear PD effect is represented by \(\psi=\beta C\), where \(\beta\) is a PD slope in effect units per mg/L and \(C\) is an individual concentration in mg/L. Suppose the PD fit gives \(\hat{\beta}=0.80\) with \(SE=0.10\), while the PK fit gives \(\hat{C}=5.0\) mg/L with \(SE=1.0\) mg/L. For this illustration only, treat the two estimates as independent and use their reported means and variances to quantify uncertainty in their product.
The plug-in effect estimate is \(0.80\times5.0=4.0\) effect units. If \(C\) is held fixed at 5.0 mg/L, the conditional standard error for this product is \(5.0\times0.10=0.50\) effect units.
Under the stated independence assumption, the variance of the product of two independent random quantities is \(\operatorname{Var}(\hat{\beta}\hat{C})=\hat{\mu}_{\beta}^{\,2}\operatorname{Var}(\hat{C})+\hat{\mu}_{C}^{\,2}\operatorname{Var}(\hat{\beta})+\operatorname{Var}(\hat{\beta})\operatorname{Var}(\hat{C})\), where the means are approximated by the reported estimates. Substitution gives \(0.80^2(1.0^2)+5.0^2(0.10^2)+(0.10^2)(1.0^2)=0.64+0.25+0.01=0.90\) effect-units squared. The corresponding standard deviation is \(\sqrt{0.90}=0.949\) effect units.
The conditional calculation gives 0.50 effect units; the illustrative calculation including concentration uncertainty gives 0.949. This difference follows from the assumptions and estimand in this example, not from a universal adjustment factor. If the estimates were dependent, the product variance would also depend on their joint distribution; the independence calculation could not simply be reused.
Worked Example: Decomposing Variation Across Repeated Analyses
Hypothetical analysis summary. Imagine repeating a two-stage analysis across many comparable data sets. In each repetition, the PK stage produces a fitted input \(\hat{\eta}\); conditional on that input, the PD estimator has variance \(0.18^2\) effect-units squared. Across PK fits, the conditional mean of the PD estimator has standard deviation \(0.24\) effect units. Assume for this illustration that the conditional variance is constant across fits.
By the law of total variance, the unconditional variance is the average conditional variance plus the variance of the conditional mean: \(0.18^2+0.24^2=0.0324+0.0576=0.0900\) effect-units squared. Therefore the total standard deviation is \(\sqrt{0.0900}=0.300\) effect units.
A single plug-in PD fit would report the conditional standard error, 0.180 effect units. Under the assumptions of this hypothetical repeated-analysis summary, the full repeated-sampling standard deviation is 0.300 effect units. The difference is not a correction inferred from one fit: it is what follows when the between-PK-fit variation is known from the stated repeated-analysis information. If the conditional variance changed across PK fits, its average—not one selected conditional variance—would enter the identity.
Worked Example: Why Covariance Cannot Always Be Ignored
Hypothetical estimand. Suppose an effect contrast is \(\psi=\beta+0.40q\), where \(\beta\) is a PD contribution in effect units and \(q\) is an exposure summary in units chosen so that \(0.40q\) is also in effect units. Let \(SE(\hat{\beta})=0.10\), \(SE(\hat{q})=0.50\), and suppose their estimation-error correlation is \(0.60\). These values are hypothetical and are supplied to illustrate the variance calculation.
The covariance is \(0.60(0.10)(0.50)=0.0300\). For a linear combination, the variance is \(\operatorname{Var}(\hat{\psi})=\operatorname{Var}(\hat{\beta})+0.40^2\operatorname{Var}(\hat{q})+2(0.40)\operatorname{Cov}(\hat{\beta},\hat{q})\). Substituting gives \(0.10^2+0.16(0.50^2)+0.80(0.0300)=0.0100+0.0400+0.0240=0.0740\) effect-units squared. The standard error is \(\sqrt{0.0740}=0.272\) effect units.
If the exposure estimate were held fixed, only the PD contribution would remain, giving a conditional standard error of 0.100 effect units for this contrast. Even including the exposure variance but incorrectly setting covariance to zero would give \(\sqrt{0.0100+0.0400}=0.224\), not 0.272. The assumed positive covariance increases uncertainty here; negative covariance could reduce it. Independence between stages may justify a zero covariance in some designs, but it must be justified rather than presumed.
How to Evaluate a Plug-In Result
State whether the target is a PD parameter conditional on exposure, a predicted response at a specified exposure, or an effect that itself combines PK and PD quantities. Different targets require different uncertainty accounting.
Identify which population PK estimates, individual parameters, random-effect estimates, and calculated concentrations were supplied to the PD fit without uncertainty.
Determine which PD variability and parameter-estimation uncertainty the PD fitting procedure includes, and whether any PK uncertainty has already been propagated.
Ask whether plausible changes in PK fits would materially change the PD target, and whether the relevant PK and PD estimates are dependent. A conditional interval cannot answer these questions by itself.
Practical warning signs include PD confidence intervals that are reported without saying that exposures were estimated, individual concentration profiles treated as exact despite sparse PK sampling, and uncertainty calculations that use only the final PD fit’s covariance matrix. A useful report states explicitly that inference is conditional on the fitted PK inputs and describes which PK uncertainty was not propagated.
A common mistake is to say, “The PD model reports a standard error, so uncertainty in the entire PK/PD analysis is included.” That reasoning confuses conditional uncertainty with total uncertainty. The correction is to identify the quantities held fixed and to determine whether repeated PK estimation would add variation to the target. Another mistake is to add every variance component as though independent. Variances add in that way only under appropriate assumptions; covariance terms may be required.
Alternatives, Uses, and Limitations
| Approach | What it does | Key uncertainty implication |
|---|---|---|
| Plug-in sequential fit | Fits the PD model using estimated PK inputs as fixed values | Typically reports uncertainty conditional on those inputs |
| Uncertainty propagation | Passes PK estimation uncertainty into the PD target or its uncertainty calculation | Can represent variation from estimated PK quantities; results depend on the propagation method and assumptions |
| Simultaneous PK/PD estimation | Estimates PK and PD quantities within a combined model | Can represent their dependence within the specified model, but does not guarantee accurate or precise inference |
The plug-in approach remains useful when the PK inputs are estimated precisely relative to the PD question, when the goal is explicitly conditional on those inputs, or when computational and operational constraints favor staging. It can also be a clear first analysis. Its limitation is not that it is always wrong, but that its standard errors answer a narrower question unless additional uncertainty is incorporated.
No simple variance decomposition by itself supplies a generally valid correction for nonlinear PD models, estimated random effects, shrinkage, or shared-data dependence. The exact impact depends on the target and model. A conditional result can be reported honestly, but it should not be presented as though it automatically covered uncertainty from the PK stage. In particular, uncertainty propagation cannot repair a misspecified PK or PD model; it addresses uncertainty under the assumptions used, not whether those assumptions are adequate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A PD model reports a standard error after using fitted individual concentrations as fixed covariates. What uncertainty does that standard error usually describe?
- In the law of total variance, what do the conditional-variance term and the variance-of-conditional-mean term represent in a staged PK/PD analysis?
- A target is \(\psi=\beta+2q\). If \(\operatorname{Var}(\hat{\beta})=0.04\), \(\operatorname{Var}(\hat{q})=0.01\), and \(\operatorname{Cov}(\hat{\beta},\hat{q})=-0.005\), calculate its variance.
- Does a plug-in analysis necessarily omit PD residual variability? Explain.
Solutions. (1) It usually describes PD estimation uncertainty conditional on the supplied fitted concentrations, including the variability represented by that PD fit; it does not automatically include variation from repeating the PK estimation. (2) The first is average PD-estimator variability with the PK fit fixed; the second is variation in the conditional PD result as the PK fit changes. (3) \(\operatorname{Var}(\hat{\psi})=0.04+2^2(0.01)+2(2)(-0.005)=0.04+0.04-0.02=0.060\), in squared units of \(\psi\). (4) No. A PD fit can include residual variability and conditional parameter uncertainty while still treating estimated PK inputs as fixed.