Choosing an Estimation Strategy for a PK/PD Analysis

Why the Estimation Strategy Is a Modeling Decision

A PK/PD analysis connects a model for drug concentrations to a model for response. The estimation strategy determines how information and uncertainty move between those parts of the analysis. In a staged analysis, a PK model is estimated first and its concentration predictions are then used in the PD analysis. In a joint analysis, PK and PD information contribute to estimation together. Neither choice is universally preferable: the appropriate strategy depends on the data, the dependence structure, the model, computational behavior, and the quantity the analysis is intended to estimate.

The practical question is not simply whether PK samples are “dense” or “sparse.” It is whether the PK data identify the exposure features that matter for the PD model, whether PD observations contain useful information about those features, and whether the model can combine that information without relying on unsupported assumptions. The earlier material on sequential-versus-joint estimation, likelihood profiles, simulation comparisons, and bootstrap uncertainty provides tools for assessing these issues; here they are assembled into a decision framework.

What the Likelihood Must Represent

Let \(D_{PK}\) and \(D_{PD}\) denote the PK and PD observations, \(\theta_{PK}\) and \(\theta_{PD}\) their respective parameter vectors, and \(C(\theta_{PK})\) the concentration predictions used by the PD model. If there are no shared latent effects and the PK and PD observations are conditionally independent given the parameters, a joint likelihood can be written as a product:

The product is not justified merely because two conditional distributions have been specified. It requires conditional independence of the PK and PD observations given the variables being conditioned on. If a latent effect \(b\) is shared, and the observations are conditionally independent given \(b\) and the model parameters, the conditional likelihood includes that effect in both factors and includes its distribution:

This is the marginal likelihood when the shared effect is integrated out. It is generally not the product of the separately marginalized PK and PD likelihoods: integrating each factor separately and multiplying would discard the dependence induced by the shared \(b\). If the effect is instead conditioned on during estimation, the conditional factors must still include it, and inference must account for the effect and its distribution as appropriate to the model.

How to Choose: Evidence, Dependence, and Target

A useful decision begins by stating the target estimand: the precise parameter, prediction, or population summary the analysis is meant to report. For example, the target might be an individual PD parameter conditional on that individual’s PK data, a typical population parameter, or a population-average response under a dosing regimen. The best computational fit is not automatically the best analysis if it targets a different quantity.

Next ask whether PK data identify the exposure features used by the PD model. Many PK samples, or samples near informative portions of the concentration-time curve, can make the relevant concentrations precise. A large sample count alone is not enough: samples may be poorly timed, concentrations may be below quantification limits, or key parameters may be weakly identified. Conversely, sparse PK data can sometimes be adequate if the design and model identify the exposure summaries relevant to the PD response.

Then consider how strongly the PD likelihood can inform PK parameters. If PD depends sensitively on an exposure parameter that is weakly determined by PK data, joint estimation may use PD information to improve precision. That can be helpful when the PK and PD models are correctly specified and the dependence structure is appropriate. It can be misleading when PD misspecification, an unmodeled time trend, or an incorrect exposure-response form pulls PK estimates away from values supported by PK data. Joint estimation does not make a misspecified PK or PD model correct.

Finally, inspect the random-effect and uncertainty structure. Shared latent effects, correlation, and uncertainty in predicted concentrations can make simple plug-in staging inadequate. If there is no meaningful shared structure and PK estimates are precise for the target, plug-in staging may be a transparent and stable choice. When uncertainty matters but joint fitting is difficult, an uncertainty-propagating staged approach can be a useful middle ground. It does not reproduce every feature of a joint likelihood, especially feedback from PD data to PK parameters, so its inferential target and limitations should be stated.

Why Plug-In Staging Can Understate Uncertainty

A plug-in analysis replaces uncertain PK parameters with a point estimate, then estimates PD conditional on that value. The law of total variance shows one consequence. For a target quantity \(\phi\) and PK parameters \(\theta_{PK}\), assuming the relevant variances exist:

The first term is average uncertainty remaining when the PK parameters are fixed; the second is variation in the conditional mean of the target across plausible PK parameter values. Plug-in staging at one estimate generally omits the second term. It can therefore yield intervals that are too narrow when the PD target is sensitive to PK uncertainty. This is a reason to propagate uncertainty or use a joint model, not a guarantee that either alternative will be reliable: both still depend on model adequacy and correct implementation.

The approximation can be especially consequential when the concentration prediction is nonlinear in PK parameters, the exposure-response relationship is steep, or the data provide little information about an influential PK parameter. In contrast, when the PK estimates are precise and modest perturbations in them barely change the PD target, the omitted component may be practically small.

Estimand and Computational Stability

The strategy must preserve the target definition. For nonlinear response models, applying the PD function to a mean exposure generally differs from averaging responses over an exposure distribution. A staged analysis that uses a single typical concentration may therefore answer a conditional question, not a population-average one. A joint fit also requires a clearly specified estimand and the appropriate prediction or averaging step; joint estimation alone does not turn a conditional prediction into a marginal population quantity.

Joint estimation can be computationally demanding because PK and PD parameters may be strongly correlated, weakly identified, or coupled through latent effects. A successful optimizer termination is not enough to establish a trustworthy result. Check that multiple reasonable initial values lead to compatible estimates, gradients and convergence diagnostics are acceptable, estimates are not held at implausible boundaries, and profiles or uncertainty summaries are informative. Compare predictions and residual diagnostics with the model’s assumptions. If staged and joint results differ materially, investigate which data or assumptions drive the difference rather than selecting the preferred answer by convenience.

Simulation-based recovery and bootstrap comparisons can reveal bias, poor coverage, or frequent fitting failure under the actual design. Such checks should reflect the intended model and realistic alternatives; simulation under a single assumed model cannot rule out misspecification. When both strategies are stable, sensitivity analyses across plausible PK and PD structures help show whether the substantive conclusion depends on the estimation choice.

A Practical Decision Sequence

Common Pitfalls and Limitations

• “Joint is always more rigorous.” A joint model can transfer information efficiently, but it can also let a misspecified PD component distort PK estimates. Better integration is useful only when the model and dependence assumptions are defensible.
• “The PK fit converged, so plug-in exposure is known.” Convergence is not precision. Examine uncertainty in the concentrations or exposure summaries that the PD model uses.
• “Separate likelihoods can always be multiplied.” A product requires the stated conditional independence. With shared effects, use the conditional product including those effects and integrate over their joint distribution when forming the marginal likelihood.
• “A stable staged fit solves a weak-identification problem.” Staging may improve numerical behavior, but it cannot create information absent from the data. A stable estimate can still be weakly identified or conditional on an arbitrary fixed PK value.
• “Similar point estimates mean the strategies are interchangeable.” Their uncertainty estimates, estimands, and sensitivity to model assumptions may still differ. Compare all of these before drawing a conclusion.

The framework has limits: there is no universal numerical threshold for “enough” PK information, and comparisons are only as credible as the model structures considered. A sound report explains why the chosen strategy matches the design and estimand, how uncertainty was handled, which diagnostics were checked, and what alternative explanations remain.