Common Errors in Sequential PK/PD Estimation

Why Small Shortcuts Can Change a Staged Analysis

In a sequential analysis, PK parameters are estimated first, and the resulting concentration predictions are then used in a PD analysis. This workflow can be practical, but it creates several opportunities to mistake an intermediate estimate for data known without error. A PD fit may look numerically precise while ignoring uncertainty inherited from the PK stage, or it may no longer correspond to the model that the analyst intends to report.

This tutorial focuses on four recurring errors: treating predicted concentrations as observed without uncertainty, omitting covariance among PK parameter estimates, reusing fitted values after changing the model or inputs without refitting, and reporting conditional uncertainty as though it were marginal. None of these errors means sequential estimation is inherently invalid. The correction is to state what is being conditioned on, keep the likelihood and model consistent, and carry forward uncertainty appropriate to the inferential target.

Set Up the Staged Model Before Diagnosing It

Let \(C_i^{\mathrm{obs}}\) denote an observed concentration for individual \(i\), and let \(C_i(\theta)\) be the concentration predicted by a PK model with parameter vector \(\theta\). The PK stage estimates \(\theta\) from PK data. The PD model then relates a response \(Y_j\) to an exposure such as \(C_j(\theta)\), with PD parameter vector \(\psi\). For example, a response model may be written \(Y_j=f(C_j(\theta),\psi)+\epsilon_j\).

A basic staged PD fit substitutes the PK estimate \(\hat\theta\) and treats the resulting concentrations \(C_j(\hat\theta)\) as fixed. The PD estimate is therefore conditional on the selected PK model and its fitted parameter values. It is not automatically equivalent to a joint analysis of the PK and PD observations, and the usual PD standard error from that fit does not automatically include uncertainty in \(\hat\theta\).

The calculations below assume the PK and PD models are specified, the relevant likelihood or error model is appropriate, and the estimates are in a region where local uncertainty approximations are meaningful. When using a first-order delta method, the prediction function must be differentiable near \(\hat\theta\), and the estimated covariance matrix must reasonably describe uncertainty in \(\hat\theta\). These assumptions should be examined rather than inferred from a successful optimizer run.

Error 1: Treating a Predicted Concentration as Observed

Predicted concentrations often appear in a data table beside measured responses, which can make them look like ordinary covariates. But \(C_j(\hat\theta)\) depends on estimated PK parameters. Treating it as error-free can make PD uncertainty conditional on one PK estimate appear to be the full uncertainty. It can also change PD estimates: the direction and size of any change depend on the model and error structure, so it is not correct to claim that plug-in use always biases a PD parameter in one particular direction.

A useful first-order calculation shows how PK parameter uncertainty affects a predicted concentration. If \(\hat\theta\) has estimated covariance matrix \(\widehat{\operatorname{Cov}}(\hat\theta)\), then for a differentiable concentration function,

The gradient describes how the prediction changes with each PK parameter. The covariance matrix describes both the variances of the parameter estimates and their covariances. Dropping either information can give a misleading uncertainty estimate.

Error 2: Ignoring Covariance Among PK Estimates

PK parameter estimates are often correlated. For instance, different combinations of clearance and volume may describe similar concentration profiles over a limited sampling window. A PD exposure prediction that depends on both parameters can therefore inherit uncertainty from their joint movement. Computing a separate contribution from each parameter variance and adding those contributions is correct only when the relevant covariances are zero, or when the derivatives make their covariance contributions zero.

More generally, for a scalar prediction \(g(\theta)\), the first-order variance approximation is

This formula follows by approximating \(g(\hat\theta)\) near the estimate: \(g(\hat\theta)\approx g(\theta)+\nabla g^\mathsf{T}(\hat\theta-\theta)\). Taking the variance of the linear expression gives \(\nabla g^\mathsf{T}\operatorname{Cov}(\hat\theta)\nabla g\), which expands into the variance and covariance terms above. The approximation is local; it may be inadequate for strongly nonlinear predictions, parameters near boundaries, or highly asymmetric uncertainty.

A practical diagnostic is to inspect the full PK covariance or correlation matrix and identify which parameters influence the PD exposure or prediction. If the matrix is unavailable, a standard error computed from parameter variances alone is not a complete propagation of uncertainty. A parameter bootstrap or simulation from the fitted joint parameter distribution can retain the dependence structure, subject to the validity of the fitted model and resampling procedure.

Error 3: Reusing Fitted Values Without Refitting

A fit is tied to its data, model, and fixed inputs. If the PK model changes, the predicted concentrations generally need to be recalculated. If those concentrations change, PD parameters estimated using them generally need to be refitted. Likewise, changing the PD model, response data, or weighting scheme requires fitting the revised PD model rather than carrying over its previous estimates as if they still optimized the new objective.

A useful check is to record the exact PK model version, parameter estimates, and prediction settings used for each PD fit. If any of these change, regenerate the concentrations and refit the PD model. When comparing analyses, evaluate each fit using a clearly defined objective for the same underlying observations; a PD-only objective conditional on fixed concentrations is not automatically comparable to a joint likelihood.

Error 4: Reporting Conditional Uncertainty as Marginal

A PD standard error computed with \(C_j(\hat\theta)\) fixed describes uncertainty conditional on those fitted PK quantities, under the PD model and its assumptions. It does not include uncertainty from estimating \(\theta\). A marginal prediction or interval that is intended to reflect both PK and PD uncertainty must include both sources. This distinction is especially important when the PK data are sparse or the PD response is sensitive to exposure.

The law of total variance makes the distinction precise. For a response \(Y\) and uncertain concentration \(C\),

The first term is average variability conditional on concentration; the second is variability in the conditional mean as concentration varies. Additional uncertainty in PD parameters can also contribute. The terms cannot in general be combined by simply adding standard errors; variances and relevant covariances must be handled on a common scale.

In a real PK/PD analysis, uncertainty may also involve estimated PD parameters, correlations between PK and PD estimates, residual error structure, and the distinction between uncertainty about a mean response and prediction of a new individual response. State which quantity an interval describes and which sources of uncertainty it includes. Bootstrap uncertainty for staged and joint fits, likelihood profiles, or Bayesian inference may be appropriate when a local normal approximation is not adequate, but each method still depends on its assumptions and implementation.

Diagnostics, Alternatives, and Limitations

The errors above can be difficult to spot from a successful fit summary alone. Check the analysis path, not only the final parameter table. In particular, compare measured concentrations and model-predicted concentrations in the records supplied to the PD stage; confirm whether the reported standard errors condition on PK estimates; inspect the complete covariance matrix; and verify that the PD fit was rerun after any change to PK predictions or model specification. Residual diagnostics from staged PK/PD fits can reveal response patterns not captured by the assumed PD model, but good residual behavior does not by itself show that PK uncertainty was propagated.

No single approach eliminates all limitations. A delta-method calculation can be efficient but unreliable for highly nonlinear predictions or weakly identified parameters. Resampling can preserve complex dependence, but only if the resampling scheme reflects the data-generating and estimation process. Joint estimation can propagate uncertainty coherently under its model, but may be computationally demanding and can be sensitive to model misspecification. The appropriate interpretation depends on the stated estimand and the information in the PK and PD data.

Common Mistakes

• Wrong: “The concentration is in the PD data set, so it is observed without error.” Correction: Identify whether each value was measured or predicted; a plug-in prediction carries uncertainty from the PK fit.
• Wrong: “Adding the separate parameter-variance contributions is enough.” Correction: Include covariance terms unless they are known to be zero or irrelevant to the prediction.
• Wrong: “The old PD estimate remains valid after I update the PK model.” Correction: Recalculate the concentrations and refit the PD model using the revised inputs.
• Wrong: “The PD standard error is the overall uncertainty.” Correction: State whether it is conditional on fitted PK values or includes PK uncertainty.