Worked Example: Effect-Compartment Estimation With Sparse Concentrations

Why Sparse Concentrations Complicate Effect-Compartment Estimation

When a pharmacodynamic response rises or falls later than plasma concentration, an effect-compartment model represents the delay with an equilibration rate. Estimating that rate is challenging when plasma samples are sparse: the unobserved parts of the plasma curve are uncertain, and those parts determine the input to the effect compartment. The estimated equilibration rate and the estimated Emax can therefore depend on whether the PK fit is frozen before PD estimation or whether both data sources contribute to one joint fit.

The calculations below use a deliberately simple simulated model to make that dependence visible. They are not a claim that joint estimation must always give better estimates. The result depends on the model, sampling design, residual assumptions, and information in both data streams.

This example assumes a known bolus time, a single linear effect compartment, an inhibitory-free increasing Emax response, and conditionally independent Gaussian errors in measured plasma concentrations and responses. It also assumes the same structural PK model is appropriate for the concentration and response observations. For the worked calculations, \(E_0\) and \(EC_{50}\) are fixed; in an applied analysis, whether to estimate them must be considered in light of the design and identifiability.

Deriving the Effect-Compartment Concentration

The effect-compartment equation says that \(C_e(t)\) moves toward the current plasma concentration at a rate controlled by \(k_{e0}\). Multiplying the differential equation by the integrating factor \(e^{k_{e0}t}\) and integrating from zero to \(t\) gives the convolution expression below. It makes explicit that the effect compartment retains a weighted history of plasma concentrations, rather than responding only to the current plasma value.

Substitute \(C_p(s)=C_0e^{-ks}\). When \(k_{e0}\ne k\), integration gives the following closed-form expression. If the two rates are equal, the limiting expression applies instead.

The equal-rate case follows by taking the limit as \(k_{e0}\) approaches \(k\). For example, the difference \(e^{-kt}-e^{-k_{e0}t}\), divided by \(k_{e0}-k\), approaches \(t e^{-kt}\); multiplying by \(k_{e0}C_0\) then gives \(C_0kt e^{-kt}\). This limit matters numerically: direct use of the unequal-rate formula near equal rates can cause loss of precision.

Sequential and Joint Likelihoods

In a sequential fit, the concentration data are used first to estimate the PK parameters. The resulting PK estimates are then held fixed while the response data are used to estimate PD parameters such as \(k_{e0}\) and \(E_{\max}\). This is computationally convenient, but the second stage treats the estimated plasma profile as though it were known. Uncertainty in the sparse-data PK fit is not automatically carried into the PD likelihood.

In a joint fit, the concentration and response observations contribute to one likelihood. For example, suppose the concentration residuals have variance \(\sigma_C^2\), and response residuals have variance \(\sigma_R^2\). Under the stated independent Gaussian-error assumption, the negative log likelihood is:

The indices \(i\) refer to individuals, \(j\) to concentration observations, and \(l\) to response observations. If the variances are fixed, the log-variance terms are constant with respect to the model parameters and may be omitted when comparing parameter estimates. If either variance is estimated, its log-variance term must be included: otherwise the objective is not the Gaussian negative log likelihood and can favour inappropriate variance estimates.

A Simulated Sequential-versus-Joint Comparison

The next example illustrates how the parameter estimates can move when the response data help determine the latent PK profile. The reported fits are hypothetical optimizer outputs from one simulated data set with sparse concentration observations and delayed responses. Assume \(E_0=5.00\) response units and \(EC_{50}=2.00\) mg/L were fixed for both fits, so that the comparison focuses on \(k_{e0}\) and \(E_{\max}\), together with PK parameters.

Assumptions, Diagnostics, and Alternatives

An effect-compartment fit is meaningful only if the structural and error assumptions are reasonably compatible with the data. In particular, a single first-order equilibration rate assumes the effect compartment approaches plasma concentration with a constant rate. A delayed response can also arise from an indirect-response process, active metabolites, tolerance, or another mechanism. The effect-compartment model may reproduce a delay without identifying its biological cause.

Useful checks include plotting observed and predicted concentrations and responses against time, inspecting residuals for systematic trends, and checking whether the predicted response shows a plausible delay relative to the plasma curve. A fitted \(k_{e0}\) near a parameter boundary, a very broad profile likelihood, or large changes in \(k_{e0}\) and \(E_{\max}\) under modest changes in the PK fit signal weak identification. With sparse PK samples, intermediate plasma concentrations are model-based predictions, not direct observations. If those predictions are uncertain, a narrow PD confidence interval conditional on a frozen PK fit can overstate the information available.

The central comparison is not “sequential is wrong, joint is right.” Sequential estimation can be reasonable when the PK profile is well characterized and the loss of PK uncertainty has little practical effect. Joint estimation carries the PK and PD residual contributions into one objective, but it can also let a misspecified PD model pull the PK estimates away from what the concentration data support. The table summarizes the distinction.

When rates are close, use a numerically stable implementation of the equal-rate limit rather than subtracting nearly equal exponentials. When concentrations or responses have different units or measurement variability, their likelihood contributions should be scaled by their appropriate variances; raw sums of squared residuals from the two assays are not directly comparable. If the response model includes estimated residual variances, retain the corresponding log-variance terms in the likelihood.

Common Pitfall

• Wrong reasoning: “The sequential PD standard error is small, so \(k_{e0}\) and Emax are precisely estimated.”
• Correction: That standard error may be conditional on a frozen PK curve. With sparse concentrations, uncertainty in that curve can affect the effect-compartment trajectory and the PD estimates. Check sensitivity to PK uncertainty or fit a joint model, then assess whether the assumptions and fit support the result.

A second common error is to minimize only the weighted squared residuals while estimating residual variances and call the result a Gaussian likelihood fit. The variance-dependent log terms are part of that likelihood; omit them only when the variances are fixed, and then only because they are constant with respect to the parameters being optimized.