Why an Interim PK Update Needs More Than a Refit
At an interim analysis, concentration samples may be available for only a subset of participants, often at a small number of times per participant. The challenge is not simply to fit a concentration curve to the observed values. A population pharmacokinetic (PK) model must distinguish typical population parameters from between-participant variability and measurement-level residual variation. It must also carry uncertainty in its parameter estimates into the predictions that inform an adaptation.
A model update can change a predicted concentration even when the observed interim concentrations are close to earlier predictions. Sparse data may provide little information about some parameters, and correlations between parameters can make predictions at one time more or less certain than predictions at another. For that reason, an interim report should identify what was updated, what information was used, the uncertainty in the updated parameters, and how that uncertainty affects the decision target defined for the trial.
A Model for Sparse Concentration Data
Consider a hypothetical single-dose intravenous bolus study with a one-compartment model. For participant \(i\), concentration at time \(t_{ij}\) is represented as
Here \(D_i\) is dose, \(CL_i\) is the participant's clearance in liters per hour, \(V_i\) is volume in liters, and \(C_{ij}\) is concentration in milligrams per liter at time \(t_{ij}\) in hours. \(CL\) and \(V\) are typical population values. The random effects \((\eta_{CL,i},\eta_{V,i})\) are assumed to have a mean-zero bivariate normal distribution with covariance matrix \(\Omega\). The residual errors \(\epsilon_{ij}\) are assumed to be mean-zero normal with variance \(\sigma^2\), independent across observations and independent of the random effects. This log-additive residual model implies a lognormal observation distribution conditional on the individual PK parameters.
Those assumptions are modeling choices, not automatic truths. A fit also relies on an adequately specified structural model, accurate dose and sample-time records, and a defensible account of which measurements are available at the interim. The one-compartment bolus equation is used here for clarity; real regimens may require different structural models.
A participant's random effects are not themselves population parameter estimates. With sparse data, an individual may contribute little direct information about their own \(CL_i\) and \(V_i\); their estimates can consequently be strongly pulled toward the population distribution. The population update must account for that shrinkage rather than treating estimated individual effects as error-free observations.
How Interim Data Update the Population Model
For a given participant, the marginal likelihood contribution is obtained by integrating over the unobserved random effects. If \(y_i\) denotes that participant's observed concentrations, the contribution has the form
The likelihood for the available participants is the product of these contributions when their data are independent conditional on the model. An interim maximum-likelihood update estimates the population parameters by maximizing that likelihood, commonly using numerical integration or a validated approximation. A Bayesian update instead combines the likelihood with a specified prior distribution. Either approach should use the actual observed sampling times and the observed-data pattern; sparse records are not made equivalent to complete profiles by filling in unmeasured concentrations.
A useful justification for carrying parameter uncertainty into predictions is the first-order delta method. Let \(g(\theta)\) be a predicted quantity based on parameter vector \(\theta\), and suppose its estimate has covariance matrix \(\widehat{\Sigma}\). Near the estimated parameter vector \(\hat{\theta}\), a first-order Taylor expansion gives \(g(\theta)\approx g(\hat{\theta})+\nabla g(\hat{\theta})^\mathsf{T}(\theta-\hat{\theta})\). Therefore the approximate prediction variance due to parameter uncertainty is
This approximation accounts for parameter variances and covariances. It does not, by itself, add between-participant variability or residual error. Those are needed for a prediction interval for a future participant's measured concentration, but not for an interval describing uncertainty in the population-typical prediction alone.
Worked Example: A Local Interim Update for Typical Clearance
Suppose four hypothetical participants each contribute one concentration at 4 hours after a 1,000 mg bolus. For illustration, take \(V=20\) L as fixed, and use a local linear approximation around typical clearance \(CL_0=5\) L/h. Assume the log-clearance random effect has variance \(0.04\), and the log-scale residual variance is \(0.01\). The four observed concentrations are \(20.233\), \(17.474\), \(19.314\), and \(16.555\) mg/L.
At \(CL_0=5\) L/h, the model's typical concentration at 4 hours is
Write \(\theta=\log CL\). The derivative of log concentration with respect to \(\theta\), evaluated at \(CL_0\), is \(-CL_0t/V=-1\). Thus, locally, a one-unit increase in \(\theta\) lowers log concentration by approximately one unit. The marginal log-scale variance for one observation is approximately \(0.04+0.01=0.05\), combining the linearized random-effect and residual variances.
The displayed observations are \(C_0\) multiplied by \(1.10\), \(0.95\), \(1.05\), and \(0.90\), respectively. Their mean log ratio to \(C_0\) is
Because the local log-concentration slope is \(-1\), the estimated shift in \(\theta\) is the negative of that mean. The approximate interim estimate is therefore
Under this local model, the standard error of \(\hat{\theta}\) is \(\sqrt{0.05/4}=0.1118\). The estimate is close to the starting value, and its standard error is substantial relative to the estimated shift. It would be misleading to report only the revised clearance point estimate. The calculation is an illustration of the marginal population update under a local approximation, not a substitute for fitting and checking the full nonlinear mixed-effects model.
Propagating Parameter Uncertainty to Concentration
For the one-compartment model, the typical log concentration at time \(t\) is a function of \(\log CL\) and \(\log V\). Let \(k=CLt/V\). Differentiating the log of the concentration equation gives the gradient
The gradient changes with time and with the fitted parameters. In particular, volume uncertainty has little first-order effect when \(k\) is near one, but it matters at times for which \(k\) differs from one. The covariance between log clearance and log volume can either increase or decrease prediction uncertainty, depending on the signs and sizes of the gradient components.
Worked Example: A Typical-Concentration Interval
Suppose an interim fit gives typical values \(CL=5\) L/h and \(V=20\) L, with estimated covariance matrix for \((\log CL,\log V)\)
Calculate the uncertainty in the typical concentration after the same 1,000 mg bolus at 2 hours. The point prediction is
Here \(k=0.5\), so the gradient for log concentration is \((-0.5,-0.5)^\mathsf{T}\). Its approximate variance from parameter uncertainty is
The standard error on the log-concentration scale is \(\sqrt{0.020625}=0.1436\). A normal-approximation 95% interval for the typical concentration, transformed back to the concentration scale, is
This interval describes uncertainty in the typical concentration due to uncertainty in the estimated population parameters, under the model and approximation. It is not a 95% interval for the concentration of an individual participant. A future individual's concentration also varies because of random effects and residual error.
Bayesian Updating as a Separate Illustration
The likelihood-based update and the Bayesian update answer related but distinct questions. A likelihood fit estimates parameters from the observed data; a Bayesian analysis combines those data with a prior distribution and summarizes the resulting posterior distribution. The prior and likelihood assumptions must both be reported, and the posterior uncertainty should not be mistaken for an uncertainty calculation that has automatically included every form of participant-level variation.
Worked Example: Combining an Interim Estimate With a Prior
Continue with the local log-clearance estimate above. For a simple Bayesian illustration, suppose the prior for \(\theta=\log CL\) is normal with mean \(\log(5)\) and variance \(0.04\). Approximate the interim likelihood by a normal distribution centered at \(\log(5)+0.003138\), with variance \(0.05/4=0.0125\). Under this normal-prior, normal-likelihood approximation, posterior precision is the sum of prior and likelihood precision:
The posterior mean is the precision-weighted average of the prior mean and interim estimate:
Thus the posterior median of \(CL\) is \(\exp\{\log(5)+0.002391\}\approx5.012\) L/h. It is closer to 5 L/h than the likelihood estimate of 5.016 L/h because the prior contributes information centered at 5 L/h. This result depends on the stated prior and the normal approximation; it is not a prior-free re-expression of the likelihood estimate.
Diagnostics, Comparisons, and Failure Modes
An interim update is only as credible as the information and model assumptions behind it. Before using updated predictions, examine whether the fitted model can reproduce the available observations and whether estimation is stable. Useful checks include observed-versus-predicted concentrations by time and dose, residual patterns on the modeled scale, convergence and sensitivity to starting values, plausible parameter estimates, and the estimated covariance structure. With very sparse data, these checks have limited power: an apparently acceptable fit does not establish that the structural model is correct.
Also inspect which participants and samples are represented. If concentrations are available sooner for some participants than others, the interim subset may differ systematically from the enrolled population. A model that assumes representative availability can then give a distorted population update. Record the interim data cut, apply prespecified data-cleaning and exclusion rules, and check that observed sampling times and dosing histories are correctly represented.
| Approach | What it uses | Important limitation |
|---|---|---|
| Marginal likelihood refit | Available concentrations integrated over participant random effects | Can be unstable when sparse data weakly identify parameters or the model is misspecified |
| Bayesian update | Likelihood plus an explicit prior distribution | Results can depend on the prior, especially when interim information is weak |
| Individual estimates followed by a population average | Participant-level estimates, often empirical Bayes estimates | Can understate uncertainty and distort the population update because sparse-data estimates are shrunk and noisy |
A common mistake is to average participant-level empirical Bayes clearance estimates and treat the result as if it were a direct, fully precise estimate of typical clearance. That reasoning ignores the fact that each estimate depends on the population model and is shrunk toward the population mean; participants with very few samples may have estimates dominated by that shrinkage. The correction is to update the population model using a method that accounts for random effects and their uncertainty, then report the resulting uncertainty in population parameters and predictions.
Another failure mode is to carry only the revised parameter point estimates into a decision calculation. The worked prediction example shows why this can be inadequate: covariance in clearance and volume creates a range of plausible typical concentrations even before individual variability is added. Conversely, adding random-effect and residual variance to an interval for the typical prediction without stating the prediction target mixes distinct sources of uncertainty. Name the target clearly: a typical concentration, a new participant's latent concentration, or a future measured concentration.
The update also has limits. Sparse data may not identify all elements of \(CL\), \(V\), \(\Omega\), and \(\sigma\) at once. Numerical precision cannot compensate for weak information, and model-based predictions may be sensitive to the chosen structural model. Therefore, an interim update should communicate uncertainty and limitations alongside estimates, and adaptations should use the prespecified decision target and rules rather than an informal reading of a fitted curve.
Specify the data cut, participants included, recorded concentrations, dose histories, sample times, and prespecified rules for handling incomplete or questionable records.
Use the stated structural and error models, account for participant random effects, and document the estimation method and any prior information.
Review residual and prediction diagnostics, convergence, parameter plausibility, and whether sparse observations leave important parameters weakly informed.
Use parameter covariance or the posterior distribution for the defined prediction target; add participant-level and residual variation only when that target requires it.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the stated model, what are the population parameters, and what do the participant random effects represent?
- Why is the marginal likelihood contribution integrated over random effects rather than evaluated at a participant's estimated random effect as if known?
- For \(k=1\), what is the gradient of log typical concentration with respect to \((\log CL,\log V)\)?
- Does the interval \([22.89,40.18]\) mg/L from the worked prediction example describe future measured concentrations for individual participants? Why or why not?
- Why can averaging empirical Bayes clearance estimates be misleading in a sparse interim data set?
Solutions. (1) \(CL\) and \(V\) are typical population parameters; \(\eta_{CL,i}\) and \(\eta_{V,i}\) represent participant deviations from those typical values, with their joint distribution described by \(\Omega\). (2) The random effects are unobserved and uncertain; integrating over their population distribution accounts for that uncertainty when evaluating the probability of the observed data. (3) The gradient is \((-1,0)^\mathsf{T}\), since \((-k,-1+k)=(-1,0)\) at \(k=1\). (4) No. It describes uncertainty in the typical concentration due to parameter uncertainty; a future measured concentration also includes between-participant and residual variability. (5) Such estimates are shrunk toward the fitted population distribution and have differing precision, so their average is not generally an unbiased, fully uncertainty-adjusted population estimate.