Why Sparse PK Sampling Matters
A PD model uses exposure to explain response. If an individual’s concentrations are measured at only a few times, their exposure may be uncertain even when the population PK model appears well estimated. That uncertainty matters because the PD model is fitted against exposure: an uncertain concentration trajectory can weaken the evidence for a PD parameter, such as a slope or EC50, and can make different parameter values fit almost equally well.
The effect depends on both the sampling times and the question being asked. A sample near one time point may describe concentration there but give little information about the trajectory between or beyond observations. Likewise, even accurately estimated concentrations may not identify PD parameters if they cover too narrow a range.
The examples use a one-compartment model after an intravenous bolus, with dose \(D\), volume \(V\), clearance \(CL\), and elimination rate constant \(k=CL/V\). This is a deliberately simple model: it assumes linear elimination, a known dose and sampling time, and no additional distribution phase. The examples are not claims about a real study. In other models the precise parameter combinations affected by sparse samples will differ, but the central issue remains: the data must distinguish the parameters that determine exposure.
One Concentration Does Not Determine an Exposure Profile
For this model, the concentration at time \(t\) is \(C(t)=(D/V)\exp(-kt)\). The area under the concentration-time curve after the dose is \(AUC=D/CL\). Thus \(V\) controls the initial concentration, while \(k\) controls the rate of decline; together they determine the profile and its AUC.
Worked Example: Many PK Profiles Fit One Sparse Sample
Suppose a hypothetical individual receives \(D=100\) mg and has one concentration measured at \(t=2\) h: \(C(2)=3.033\) mg/L. Under the one-compartment model, this single measurement supplies one equation for the two unknowns \(V\) and \(k\):
For \(k=0.25\) h\(^{-1}\), the equation gives \(V=20.00\) L and \(CL=kV=5.000\) L/h. The AUC is \(100/5.000=20.00\) mg·h/L. But \(k=0.20\) h\(^{-1}\) also fits the same observation, with \(V=22.10\) L, \(CL=4.420\) L/h, and \(AUC=22.62\) mg·h/L. Taking \(k=0.30\) h\(^{-1}\) gives \(V=18.10\) L, \(CL=5.429\) L/h, and \(AUC=18.42\) mg·h/L. Rounding accounts for small differences in the displayed values.
The three parameter pairs give the same concentration at 2 h but different clearances and AUCs. In this design, the PK parameters are structurally non-identifiable: a single concentration cannot uniquely determine both \(V\) and \(k\). A second concentration at a distinct time could distinguish these profiles in an ideal error-free setting. For example, at 6 h the three predicted concentrations are approximately 1.116, 1.363, and 0.913 mg/L for \(k=0.25\), 0.20, and 0.30 h\(^{-1}\), respectively. With noisy observations, whether that added sample is informative enough also depends on its measurement error and timing.
In an individual staged analysis, PK parameters are estimated first and then used to supply exposure for the PD fit. A point estimate of exposure hides uncertainty along alternatives like those above. A joint analysis can retain PK uncertainty by estimating PK and PD parameters together, but it cannot make an uninformative PK design informative by itself. Any improvement depends on the full model and on whether the PD observations contribute valid information about the uncertain exposure.
Exposure Range and PD Parameter Identifiability
Consider an Emax response model with baseline fixed at zero for simplicity: \(E(C)=E_{\max}C/(EC_{50}+C)\). Here \(E_{\max}\) is the maximum effect and \(EC_{50}\) is the concentration producing half of that maximum. Estimating both requires observations that distinguish their effects on the response. If the observed concentrations occupy only a narrow range, changes in \(E_{\max}\) and \(EC_{50}\) can compensate for one another.
A local way to examine this issue is to consider how predicted response changes with each parameter. The sensitivity derivatives are:
If observations occur at identical concentrations, the two columns of these sensitivities are proportional across observations. The data then cannot separate the two parameters locally. Concentrations that vary meaningfully can make the sensitivity patterns less alike and improve practical identifiability, provided the PD observations are sufficiently precise and the model is appropriate.
Worked Example: A Narrow Concentration Range Weakens Emax Estimation
Suppose the hypothetical true values are \(E_{\max}=100\) effect units and \(EC_{50}=10\) mg/L. Compare measurements at 1 and 1.2 mg/L with measurements at 1 and 20 mg/L. The predicted effects at 1 and 1.2 mg/L are \(100(1)/(10+1)=9.091\) and \(100(1.2)/(10+1.2)=10.71\) units. Both concentrations are far below \(EC_{50}\).
At the first concentration, the sensitivity pair \((\partial E/\partial E_{\max},\partial E/\partial EC_{50})\) is \((0.09091,-0.8264)\). At 1.2 mg/L it is \((0.1071,-0.9566)\). The two rows are nearly proportional: their determinant is approximately \(0.00158\). This small value indicates that the two parameters have very similar effects on these predictions, so separating them will be difficult with noisy data.
At 20 mg/L, the predicted effect is \(100(20)/(10+20)=66.67\) units, and the sensitivity pair is \((0.6667,-2.222)\). The determinant for the sensitivity rows at 1 and 20 mg/L is approximately \(0.3489\), much larger than for the two nearby low concentrations. The broader range provides more distinct information about the two parameters in this local comparison. It does not guarantee precise estimates: residual variability, the number of observations, and other model parameters also matter.
This reasoning also applies to other PD models. If the observed exposures barely vary, a slope, potency parameter, maximum effect, or baseline can be weakly identified depending on the model. In particular, at concentrations far below \(EC_{50}\), the Emax model is approximately \(E(C)\approx (E_{\max}/EC_{50})C\). Those observations primarily inform the ratio \(E_{\max}/EC_{50}\), rather than each parameter separately.
How PK Uncertainty Reaches a PD Estimate
A staged fit that treats estimated concentrations as exact can understate uncertainty in PD parameters. A first-order calculation shows the mechanism: uncertainty in concentration contributes through the sensitivity of the PD estimate to concentration. This calculation assumes small errors, a locally smooth relationship, and independence between the PK concentration estimate and the PD residual in the example below. It is an approximation, not a substitute for a full joint model.
Worked Example: Propagating Concentration Uncertainty to a PD Slope
Suppose a hypothetical linear PD model has no intercept, \(Y=\beta C+\epsilon\), and one individual has observed effect \(Y=12\) units. The estimated concentration is \(C=2.0\) mg/L with standard error \(0.40\) mg/L. Assume the PD residual standard deviation is \(1.0\) unit and the PK and PD errors are independent.
Plugging in the concentration gives \(\hat{\beta}=Y/C=12/2.0=6.00\) units per mg/L. If \(C\) were known exactly, the conditional standard error from PD residual error alone would be \(1.0/2.0=0.500\) units per mg/L. But \(\beta=Y/C\) is sensitive to \(C\), with derivative \(\partial\beta/\partial C=-Y/C^2=-12/4=-3.00\) units per mg/L for each mg/L change in concentration. The approximate variance contribution from concentration uncertainty is therefore \((-3.00)^2(0.40)^2=1.44\) in squared slope units.
The PD residual contribution to variance is \(0.500^2=0.250\). Under the stated independence and first-order assumptions, the approximate combined variance is \(1.44+0.250=1.69\), giving standard error \(\sqrt{1.69}=1.300\) units per mg/L. The staged conditional standard error, 0.500, represents only PD residual uncertainty; it omits the uncertainty contributed by the estimated concentration.
With only this one individual and both \(C\) and \(\beta\) unknown, the equation \(Y=\beta C\) alone identifies only their product, not each quantity separately. PK information is what constrains \(C\). With multiple individuals sharing a PD parameter, PD observations may also inform individual exposures in a joint fit, but that relies on the shared PD model being suitable. Joint estimation does not guarantee a more precise or more reliable estimate in every design.
Assumptions, Diagnostics, and Analysis Choices
Identifiability conclusions depend on the assumed PK and PD models and on the observation design. The one-compartment example assumes known dose, known sampling times, linear elimination, and correctly measured concentrations. The Emax example assumes that the stated response function is appropriate. If these assumptions fail, apparent parameter uncertainty may reflect model misspecification as well as sparse data. A precise estimate conditional on a wrong model is not evidence that the parameter is well supported by the data.
Useful warning signs include large standard errors or correlations among parameter estimates, estimates that change substantially when one sample is removed, and fitted individual concentration curves that vary widely while passing through the measured samples. For PD parameters, inspect the distribution of individual predicted concentrations, not only the number of PD observations. A large number of responses at nearly the same exposure may still provide little information for separating potency from maximum effect. These signs flag possible weak identifiability; none alone proves its cause.
| Approach | What it does with PK uncertainty | Main limitation in sparse sampling |
|---|---|---|
| Staged estimation with plug-in exposure | Uses an estimated PK profile as if it were known in the PD fit, unless uncertainty is explicitly propagated. | Conditional PD precision can be too optimistic when individual exposure is uncertain. |
| Staged estimation with uncertainty propagation | Carries estimated PK uncertainty into PD uncertainty using an approximation or repeated sampling. | Approximate methods may be inaccurate when uncertainty is large or the PD relationship is strongly nonlinear. |
| Joint estimation | Estimates PK and PD parameters together while retaining their dependence in the specified model. | Cannot recover information absent from the design; results depend on model assumptions and may still be weakly identified. |
The comparison is not a rule that one approach is always preferable. Staged estimates can be useful when PK is well characterized and uncertainty is adequately carried forward. Joint estimation can represent shared uncertainty directly, but it does not remove confounding or compensate automatically for sparse sampling. The central diagnostic question is whether the observed data distinguish the combinations of PK and PD parameters being interpreted.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Under the one-compartment bolus model, what does one concentration at a known time tell you about \(V\) and \(k\)?
- Why can observations only at concentrations far below \(EC_{50}\) make \(E_{\max}\) and \(EC_{50}\) difficult to estimate separately?
- In the slope example, what standard error would be reported if concentration uncertainty were ignored, and what approximate combined standard error follows from the stated assumptions?
- Does joint estimation guarantee improved precision when PK samples are sparse? Explain.
Solutions. (1) It supplies one equation, \(C(t)=(D/V)\exp(-kt)\), for two unknowns, so multiple \(V,k\) pairs can fit it; AUC need not be uniquely determined. (2) When \(C\ll EC_{50}\), \(E(C)\approx(E_{\max}/EC_{50})C\), so the data mainly inform the ratio rather than each parameter. (3) Ignoring PK uncertainty gives \(SE=1.0/2.0=0.500\) units per mg/L. Including the approximate PK contribution gives \(SE=\sqrt{0.250+1.44}=1.300\) units per mg/L. (4) No. Joint estimation retains dependence under the specified model, but cannot create information absent from the sampling design, and its precision depends on model assumptions and the information in both data types.