Why Parameter Uncertainty Matters
A dose simulation uses a fitted pharmacokinetic model to predict exposure under proposed regimens. A common starting point is to insert the estimated parameter values and simulate a virtual population. This gives a useful plug-in prediction, but the fitted values are estimates, not known truths. If plausible parameter values imply meaningfully different target-attainment rates, a recommendation based only on the estimates can convey more certainty than the data support.
The focus here is uncertainty in the population model’s fitted parameters. It is distinct from interindividual variability, which describes how patients differ from one another under a model, and from residual variability, which describes unexplained variation in observations. A simulation may include all three, but they enter at different levels: parameter uncertainty concerns which population parameter values are plausible, while patient and residual variability concern outcomes conditional on those values.
The earlier material on comparing regimens by simulated target attainment establishes the basic target-attainment calculation. Here, the additional task is to repeat that calculation over plausible fitted parameter values and determine whether the dose comparison changes.
That parameter distribution might be an approximate sampling distribution of the estimator, a distribution generated by a parametric bootstrap, or a posterior distribution. These approaches have different interpretations, so the chosen source should be stated. In a Bayesian analysis, averaging over the posterior gives a posterior predictive probability. With frequentist parameter draws, it is an uncertainty-averaged prediction; it should not automatically be described as a posterior probability.
Propagating Uncertainty Through the Simulation
Suppose the fitted model has estimate \(\widehat{\boldsymbol{\theta}}\) and estimated covariance matrix \(\widehat{\Sigma}_{\theta}\). When the estimator is approximately normal and parameters are away from boundaries, one possible approximation is to draw parameter vectors jointly:
The joint draw matters: off-diagonal entries in \(\widehat{\Sigma}_{\theta}\) represent estimated dependence among parameter estimates. Drawing each parameter independently can produce the wrong range of combinations, especially when estimates are strongly correlated. For positive parameters, it may be more appropriate to draw on a log scale and transform back, or to use bootstrap or posterior draws that respect the parameter constraints.
For each parameter draw, simulate patients under the proposed regimen, including the model’s interindividual variability if the target concerns a population of patients. Let \(I_i^{(b)}(d)\) equal 1 if simulated patient \(i\) meets the target and 0 otherwise. With \(M\) simulated patients per parameter draw, estimate conditional attainment as:
The average \(\widehat{A}_{\mathrm{avg}}(d)\) describes target attainment averaged over the parameter draws and simulated patient variability. The distribution of \(\widehat{A}^{(b)}(d)\) across draws shows how much the conditional prediction varies with parameter uncertainty. Its spread is not the same as a confidence interval for the true attainment rate unless the draw-generation and interval-construction procedure justify that interpretation.
State the exposure criterion, regimen, population, and whether the target concerns individual attainment or another summary.
Use a joint covariance approximation, bootstrap estimates, or posterior draws, and preserve parameter constraints and dependence.
Hold the population parameters fixed within an outer draw; simulate patient variability within that draw when needed.
Report average attainment and the distribution across parameter draws, then compare dose rankings and decisions with the plug-in result.
Worked Examples
Worked Example: Averaging Attainment Across Parameter Draws
In a hypothetical analysis, the target is a trough concentration from 1.0 to 2.0 mg/L. For a proposed regimen, the plug-in simulation at \(\widehat{\boldsymbol{\theta}}\) predicts attainment of 0.640. To examine parameter uncertainty, four equally weighted parameter vectors are used for illustration. At each vector, a large simulated cohort gives the following conditional attainment estimates: 0.720, 0.660, 0.580, and 0.500.
The uncertainty-averaged estimate is:
Thus the plug-in estimate is 64.0%, while the average across these parameter draws is 61.5%. The four conditional estimates range from 50.0% to 72.0%, indicating that the prediction depends appreciably on the fitted parameters. These four draws are only a teaching illustration, not enough to characterize uncertainty reliably in an applied analysis. The average also does not mean that every patient has a 61.5% chance of meeting the target; it is a population-level probability under the specified simulation and parameter-averaging scheme.
Worked Example: Why Parameter Covariance Affects an Exposure Prediction
Consider a hypothetical local approximation to log trough concentration, \(h\), as a function of two fitted log-scale parameters, \(\theta_1\) and \(\theta_2\). At the fitted values, predicted trough is 1.50 mg/L. Near those values, suppose the change in \(h\) is approximated by a linear function with gradient \((-0.8,-0.3)\). The estimated covariance matrix of the two parameters is:
For a differentiable function, the first-order delta method approximates the variance of its estimate by the gradient multiplied by the parameter covariance matrix and the gradient’s transpose. Here:
Therefore, the approximate standard deviation on the log-concentration scale is \(\sqrt{0.0481}=0.2193\). Since the fitted trough is 1.50 mg/L, its log is \(\ln(1.50)=0.4055\). A rough 95% interval from this normal approximation is \(0.4055 \pm 1.96(0.2193)\), or \((-0.0243,0.8353)\) on the log scale. Transforming the endpoints gives approximately 0.976 to 2.305 mg/L.
This calculation illustrates that the positive covariance contributes \(0.0144\) to the approximate variance. If the covariance were incorrectly set to zero, the variance would instead be \(0.0256+0.0081=0.0337\), understating uncertainty for this gradient. The interval is a local, parameter-only approximation: it does not include patient variability, and it may be poor if the exposure function is strongly nonlinear or the parameter distribution is far from normal.
Worked Example: A Dose Ranking That Changes After Propagation
Suppose a hypothetical model compares three regimens against the same exposure target. Four equally weighted parameter draws produce the conditional attainment estimates below. The plug-in column gives predictions at the fitted parameter estimates.
| Regimen | Plug-in | Draw 1 | Draw 2 | Draw 3 | Draw 4 | Average across draws |
|---|---|---|---|---|---|---|
| 100 mg | 0.750 | 0.700 | 0.620 | 0.540 | 0.460 | 0.580 |
| 200 mg | 0.740 | 0.780 | 0.760 | 0.720 | 0.680 | 0.735 |
| 300 mg | 0.650 | 0.740 | 0.690 | 0.630 | 0.570 | 0.6575 |
For 200 mg, the average is \((0.780+0.760+0.720+0.680)/4=2.940/4=0.735\). The corresponding averages for 100 mg and 300 mg are 0.580 and 0.6575. The plug-in predictions rank 100 mg first (75.0%), narrowly ahead of 200 mg (74.0%). The uncertainty-averaged predictions instead rank 200 mg first (73.5%).
This changes the preferred regimen if the decision rule is to select the regimen with the highest average target attainment. It does not, by itself, establish that 200 mg is clinically preferable: a real selection must also account for the prespecified decision criteria and relevant safety or practical constraints. The broad variation in the 100 mg results also signals that its plug-in advantage is sensitive to parameter uncertainty.
Diagnostics, Alternatives, and Limitations
Parameter propagation is only as credible as the parameter draws and the simulation they feed. Before interpreting results, check that the fitted model and uncertainty calculation are numerically reliable. For a covariance-based approximation, inspect whether estimates are near boundaries, whether the covariance matrix is valid and well-conditioned, and whether the approximate normal distribution generates impossible values. For any draw method, inspect marginal distributions and parameter correlations, and check that simulated exposures behave plausibly across the range of draws.
Monte Carlo error is a separate source of uncertainty. Too few parameter draws can give an unstable average or misleading impression of the spread; too few simulated patients per draw can add noise to each conditional attainment estimate. Increase both counts and check whether the reported summaries stabilize. Using the same simulated patient-level random numbers across regimens where appropriate can reduce noise in regimen comparisons, but it does not remove parameter uncertainty.
The main alternatives answer related but distinct questions:
- Plug-in simulation uses \(\widehat{\boldsymbol{\theta}}\) alone. It is straightforward and useful as a reference, but it hides sensitivity to parameter estimation.
- Covariance-based draws are computationally convenient when a normal approximation is adequate. They can be unreliable near constraints or for strongly nonlinear models.
- Parametric bootstrap repeatedly simulates and refits data under the model, then propagates the resulting estimates. It can better reflect estimator behavior, but depends on the assumed model and can be computationally demanding.
- Posterior draws propagate a Bayesian posterior distribution. Their interpretation depends on the model and prior, and they require adequate sampling and convergence checks.
- Delta-method approximations are efficient for smooth predictions near the fitted values, but describe local uncertainty and may miss skewness, nonlinear threshold behavior, or multimodality.
A common mistake is to simulate many patients at the single fitted estimate and conclude that the resulting narrow Monte Carlo error captures uncertainty in the dose recommendation. It captures simulation precision, not uncertainty about \(\boldsymbol{\theta}\). The correction is to add an outer layer of parameter draws, summarize how the prediction changes across those draws, and keep Monte Carlo error distinct from parameter uncertainty.
Finally, parameter uncertainty is conditional on the model structure, the data, and the estimation procedure. It does not automatically account for model misspecification, uncertainty about the target, differences between the simulated and intended patient population, or uncertainty in adherence and clinical practice. Report the parameter-draw method, the number of draws and simulated patients, the target definition, and both plug-in and uncertainty-propagated results. When a dose ranking changes across plausible parameter values, describe the decision as uncertainty-sensitive rather than presenting one ranking as definitive.
Check Your Understanding
Answer each question, then compare with the concise solutions.
- In a dose simulation, what is the difference between interindividual variability and parameter uncertainty?
- Three equally weighted parameter draws give conditional attainment estimates of 0.60, 0.70, and 0.80. What is their average?
- Why can drawing correlated model parameters independently distort uncertainty propagation?
- A model’s plug-in prediction is 80% attainment, but the uncertainty-averaged prediction is 72%. Does this mean 72% of parameter values predict exactly 72% attainment?
- Name one limitation of using a multivariate normal approximation to generate parameter draws.
Solutions. (1) Interindividual variability describes patient-to-patient differences conditional on model parameters; parameter uncertainty describes uncertainty about the fitted parameter values themselves. (2) \((0.60+0.70+0.80)/3=0.70\), or 70.0%. (3) Independent draws discard the estimated covariance, changing the combinations of parameter values and potentially overstating or understating prediction uncertainty. (4) No. It is an average of conditional predictions over the chosen parameter draws; individual draws can predict different attainment rates. (5) For example, it can generate impossible values for constrained parameters or poorly approximate skewed or boundary-adjacent uncertainty.