Prior Predictive Calibration for NLME Models

Why Calibrate Priors Through Predicted Profiles?

A prior can look reasonable when each parameter is considered separately and still imply implausible concentrations when those parameters act together. For example, a large clearance paired with a small volume produces rapid elimination, while a small clearance paired with a large volume produces a long half-life. In an NLME model, population parameters, random effects, their dependence, and the observation model jointly determine the concentration-time profiles the model considers plausible before any study data are used.

Prior predictive calibration makes those implications visible. The goal is not to make simulated profiles resemble a particular observed data set; that would use the data to tune the prior. Instead, specify clinically defensible ranges or patterns in advance, simulate from the model and priors, and ask whether the resulting profiles are plausible for the intended drug, dose, population, and sampling times.

Definition: The prior predictive distribution for a future observation $Y^{\mathrm{rep}}$ is the distribution obtained by first drawing population parameters from their prior, then drawing individual random effects from their model, and finally drawing observations from the likelihood. In a hierarchical model, $$ p(y^{\mathrm{rep}})=\int p(y^{\mathrm{rep}}\mid \theta,\eta)\,p(\eta\mid\theta)\,p(\theta)\,d\eta\,d\theta. $$ Here $\theta$ denotes population-level parameters and $\eta$ denotes individual random effects. The distribution describes what the model predicts before conditioning on the observed study outcomes.

For a simple one-compartment intravenous bolus model, a subject's latent concentration can be written $C_i(t)=D/V_i\exp(-CL_i t/V_i)$, where $D$ is dose, $CL_i$ is individual clearance, and $V_i$ is individual volume. A common hierarchical specification is $\log(CL_i)=\mu_{CL}+\eta_{CL,i}$ and $\log(V_i)=\mu_V+\eta_{V,i}$, with the vector of random effects drawn from a mean-zero multivariate normal distribution with covariance matrix $\Omega$. The priors on $\mu_{CL}$, $\mu_V$, and the elements of $\Omega$ complete the parameter-generating part of the model. An observation model then maps latent concentrations to measured concentrations.

Prior predictive simulation requires the full generative model, not just marginal prior draws for the population means. Each replicate should draw the population parameters, draw individual random effects conditional on the covariance structure, calculate the concentration profile at the intended times, and generate observations if the observation model is part of the question. Keep units explicit throughout: a clearance in L/h and volume in L give a rate constant in $1/\mathrm{h}$, so its product with time in hours is dimensionless.

What the Simulation Reveals

A prior predictive check examines distributions and profiles, not a single “typical” parameter set. Useful summaries include concentrations at clinically relevant times, peak and trough concentrations, time above or below a threshold, area under the curve when it can be computed appropriately, and the frequency of biologically impossible or clinically implausible trajectories. For a population model, inspect both the distribution of individual profiles and the range of population-level summaries across prior draws.

The sources of predictive spread can be separated conceptually. Given population parameters and an individual's random effects, the likelihood contributes residual observation variation. Across individuals, random effects add heterogeneity. Across prior draws, uncertainty in population parameters changes the distribution of individuals and their profiles. A basic variance identity makes the distinction precise. For any predicted quantity $Y$, conditional on parameters $\theta$ and effects $\eta$, $$ \operatorname{Var}(Y)=\operatorname{E}\{\operatorname{Var}(Y\mid\theta,\eta)\} +\operatorname{Var}\{\operatorname{E}(Y\mid\theta,\eta)\}. $$ To justify this, write $Y-\operatorname{E}(Y)=[Y-\operatorname{E}(Y\mid\theta,\eta)]+[\operatorname{E}(Y\mid\theta,\eta)-\operatorname{E}(Y)]$, square and take expectations. The cross-product has expectation zero because the first bracket has conditional mean zero. The first term is average conditional variation; the second is variation in conditional means induced by the distributions of parameters and effects.

This decomposition helps interpret a broad predictive distribution but does not, by itself, say which component is too broad. Plot conditional latent profiles without observation noise as well as generated observations when useful. If latent profiles are plausible but observations are excessively dispersed, examine the observation model; if latent profiles already span implausible values, the issue lies in population priors, random effects, or their combination.

Worked Example: Checking One Simulated Concentration Profile

Consider a hypothetical prior predictive draw for a 100 mg intravenous bolus. The draw produces $V_i=18.10$ L and $CL_i=1.71022$ L/h. These are illustrative simulated individual parameters, not estimates from a real study. Calculate the latent concentration at 8 h.

For the one-compartment bolus model, $$ C_i(8)=\frac{D}{V_i}\exp\left(-\frac{CL_i}{V_i}\,8\right). $$ Substituting the draw and units gives $$ C_i(8)=\frac{100\ \mathrm{mg}}{18.10\ \mathrm{L}} \exp\left(-\frac{1.71022\ \mathrm{L/h}}{18.10\ \mathrm{L}}\times 8\ \mathrm{h}\right). $$ The exponent is $-0.755898$, so $\exp(-0.755898)\approx0.4696$, and $100/18.10\approx5.5249$ mg/L. Using the unrounded parameter values in the calculation gives $C_i(8)\approx2.595$ mg/L.

This arithmetic checks that a single simulated profile can be evaluated correctly; it does not establish that the priors are calibrated. A calibration decision requires examining many such profiles, including combinations in the tails of the population-parameter and random-effect distributions, and comparing them with criteria set for the intended application.

Worked Calibration With Prespecified Criteria

A useful check begins by defining what “plausible” means in context. Criteria may be broad rather than exact: for example, a concentration should be nonnegative, a profile should decline after an intravenous bolus under the model, or only a small fraction of subjects should exceed a specified concentration threshold at a specified time. Criteria should be scientifically justified and chosen before inspecting the simulation output. They are not universal cutoffs; the same value may be acceptable for one compound and unacceptable for another.

Worked Example: Estimating the Prior Probability of Extreme Concentrations

Suppose a hypothetical model is simulated for 1,000 prior predictive subjects receiving a specified dose. Before simulation, the team sets two illustrative screening criteria: concentrations above 15 mg/L at 1 h are uncommon, and concentrations above 8 mg/L at 24 h are uncommon. The intended judgment is that each event should occur in fewer than 5% of prior predictive subjects. These thresholds are hypothetical and would need application-specific justification.

Suppose the simulated profiles show 84 subjects above 15 mg/L at 1 h and 27 subjects above 8 mg/L at 24 h. The estimated prior predictive probabilities are $$ \widehat{P}(C(1)>15\ \mathrm{mg/L})=\frac{84}{1000}=0.084, \qquad \widehat{P}(C(24)>8\ \mathrm{mg/L})=\frac{27}{1000}=0.027. $$ Thus, 8.4% of the simulated subjects exceed the first threshold, which is above the prespecified 5% screening limit. The second event occurs in 2.7%, below that limit. On these criteria, the prior requires review because of the early high-concentration tail; the result does not imply that every parameter prior is wrong or identify the responsible component.

A next diagnostic is to inspect the simulated parameters and profiles among the 84 high-concentration subjects. If they tend to have small volumes, the volume prior may be influential; if the high concentrations persist or rise unexpectedly, check the model implementation and units as well. The threshold proportions are Monte Carlo estimates: with a finite number of simulated subjects, they have simulation error. Increasing the number of draws can clarify whether a proportion near the criterion is stably above or below it, but more draws cannot make an implausible model plausible.

Calibrating Prior Scales From Implied Ranges

Prior scales should be judged on the natural parameter and profile scales, not only on the scale used to write the prior. For example, if $\log(CL)\sim N(\log(5),0.6^2)$, then clearance is lognormal with median 5 L/h. Its central 95% prior interval is approximately $5\exp(\pm1.96\times0.6)$, or 1.54 to 16.2 L/h. That interval is a marginal interval for clearance alone; it is not a joint 95% region for clearance, volume, and random effects.

The interaction between parameters can be even more revealing. Under the same example, let $V$ have median 50 L and log-scale standard deviation 0.5. Its central 95% prior interval is $50\exp(\pm1.96\times0.5)$, or approximately 18.8 to 133 L. Combining the marginal endpoint values gives a broad range of possible half-lives, since $t_{1/2}=(\log 2)V/CL$. Such combinations do not all have equal probability, especially if random effects are correlated, but the calculation signals why profile simulation is needed rather than checking marginal intervals alone.

Worked Example: Comparing Two Log-Scale Prior Choices

For illustration, suppose the clearance prior is $\log(CL)\sim N(\log(5),0.6^2)$, with clearance in L/h, and the volume prior is $\log(V)\sim N(\log(50),0.5^2)$, with volume in L. The implied central 95% marginal intervals are $$ CL:\ 5\exp(\pm1.96\times0.6)\approx[1.54,16.2]\ \mathrm{L/h}, \qquad V:\ 50\exp(\pm1.96\times0.5)\approx[18.8,133]\ \mathrm{L}. $$ For the combination of low clearance and high volume at these interval endpoints, the half-life is approximately $0.693\times133/1.54=59.8$ h. This is an endpoint illustration, not a claim that the combination has 2.5% probability or belongs to a joint 95% interval.

Now consider narrower log-scale standard deviations of 0.25 for clearance and 0.20 for volume, keeping the same medians. The marginal central 95% intervals become $$ CL:\ 5\exp(\pm1.96\times0.25)\approx[3.06,8.15]\ \mathrm{L/h}, \qquad V:\ 50\exp(\pm1.96\times0.20)\approx[33.8,74.0]\ \mathrm{L}. $$ These priors imply less dispersion in the marginal parameters and generally narrower profile variation, but they are not automatically better. The scale change is defensible only if the narrowed parameter ranges and the simulated profiles remain compatible with substantive knowledge. Simulate the full profiles under each candidate prior and compare them with the same prespecified criteria; do not choose a scale merely because it produces a visually neat plot.

Practical Workflow, Diagnostics, and Alternatives

A disciplined calibration workflow makes each modeling choice auditable. Preserve the prior version, simulation settings, random seed, and summaries used for the decision. For every prior draw, retain enough information to trace an unusual profile back to its population parameters and individual effects. This is particularly helpful when an extreme prediction results from a joint parameter combination rather than an obviously extreme marginal draw.

Set the predictive target.
Specify dose, population, sampling times, and the concentration or exposure features that matter. Write down the plausibility criteria before generating profiles.

Simulate from the complete prior model.
Draw population parameters, random effects with their specified covariance, latent profiles, and observations if required by the target. Apply identical units and equations in every replicate.

Inspect summaries and trajectories.
Check distributions at key times, profile shapes, tail frequencies, and parameter combinations associated with extremes. Separate latent variability from observation noise when diagnosing spread.

Revise and repeat transparently.
If a criterion fails, identify the influential prior component or model assumption, revise with a scientific rationale, and rerun the same checks. Record both the original and revised implications.

Warning signs include negative or otherwise impossible generated values when the chosen likelihood permits them, profiles with biologically impossible shapes, large fractions beyond clinically defensible ranges, and extreme tails driven by implausible parameter combinations. Also check numerical failures: exponent overflow, concentrations reported in the wrong units, dose or time entered on an inconsistent scale, and random effects applied to the wrong transformed parameter. A strange profile is not always evidence that the prior is too diffuse; it can expose a coding error or a mismatch between the structural model and the intended use.

Prior predictive checks differ from posterior predictive checks. A prior predictive check integrates over the prior before using the observed outcomes and asks what the model permits in advance. A posterior predictive check uses the posterior distribution after observing data and asks whether the fitted model can reproduce relevant features of those data. Neither replaces the other. Prior predictive calibration is also distinct from checking whether parameter values are individually “weakly informative”: it tests the implications of their joint distribution through the model.

The method has limitations. Plausibility criteria can be uncertain, sparse prior knowledge can leave multiple defensible choices, and a finite simulation may miss rare but consequential tails. A profile that looks reasonable at one dose may be implausible at another dose or in another population. Prior predictive agreement with broad expectations does not prove that the model is correct, nor does it guarantee adequate posterior computation or fit. It is a structured way to expose implications early, not a substitute for scientific judgment or later data-based model assessment.

Common Mistake

Common pitfall: “The prior medians for clearance and volume are plausible, so the prior predictive profiles must be plausible.” This reasoning ignores the spreads, dependence, and nonlinear interaction of parameters. A plausible median pair can coexist with implausible tails or half-lives. Instead, simulate joint draws using the specified covariance and likelihood, inspect the resulting concentration-time profiles, and compare them with criteria set for the intended application.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In what order are population parameters, individual random effects, and observations generated in a prior predictive simulation?
A prior predictive simulation has 40 concentrations above a prespecified threshold among 800 subjects. What is the estimated prior predictive probability of exceeding the threshold?
Why is it not sufficient to inspect the marginal prior intervals for clearance and volume separately?
If latent concentration profiles are plausible but simulated measured concentrations are too dispersed, which model component should be investigated first?

Solutions. (1) Draw population parameters from their priors, draw individual random effects from the hierarchical distribution conditional on those parameters, calculate latent profiles, and then generate observations from the likelihood when observations are part of the predictive target. (2) The estimate is $40/800=0.05$, or 5%. (3) The concentration model is nonlinear in clearance and volume, and their joint values—including dependence—determine the profile shape and exposure; plausible marginal intervals do not ensure plausible combinations. (4) Investigate the observation model, because the latent profiles are plausible while the additional spread appears when generating measured concentrations.