Tutorials › Pharmacometrics › Model Averaging in Pharmacometric Modeling
Pharmacokinetics · Pharmacometrics

Model Averaging in Pharmacometric Modeling

Learn how model averaging accounts for structural model uncertainty by combining predictions or parameter estimates from multiple plausible pharmacometric models rather than treating one selected model as unquestionably correct.

Intermediate Pharmacometrics Model Selection Model Uncertainty
01 · The big picture

1. What Is Model Averaging?

Model averaging is a framework for incorporating uncertainty about which candidate model best represents the data. Instead of selecting one structural model and treating that selection as fixed, model averaging combines information from several plausible models.

This distinction matters in pharmacometric modeling because different structural models can sometimes provide similar descriptions of the observed data while producing different predictions, parameter estimates, or extrapolations.

Observed data Model 1 one-compartment Model 2 two-compartment Model 3 alternative absorption Model averaging weighted combination averaged prediction or inference

Model averaging retains information from multiple plausible models and combines them according to prespecified or estimated weights.

Core idea: model averaging does not ask only, “Which model wins?” It asks, “How should uncertainty across plausible models be incorporated into the quantity we want to estimate or predict?”
02 · Why it matters

2. Why Is Model Uncertainty Important?

Pharmacometric analyses often involve structural decisions. Examples include choosing the number of compartments, the form of absorption, the representation of nonlinear elimination, or the structure of a covariate relationship.

These decisions are often informed by diagnostics, prior knowledge, plausibility, and statistical criteria. But when several models remain plausible, choosing one model can understate the uncertainty associated with that choice.

Modeling decisionCandidate alternativesPotential consequence
Number of compartmentsOne vs. two vs. three compartmentsDifferent distribution and terminal-phase predictions
AbsorptionFirst-order vs. transit vs. zero-order inputDifferent early concentration profiles and Tmax
Covariate formLinear vs. power vs. nonlinear relationshipDifferent predictions for individuals outside the central covariate range
Residual errorAdditive vs. proportional vs. combinedDifferent treatment of observation variability
PD relationshipEmax vs. sigmoid EmaxDifferent exposure-response predictions

A model may fit the observed data well while alternative models produce similarly acceptable fits. The differences may become more important when the model is used for simulation, extrapolation, dose selection, or prediction in populations that differ from the development dataset.

03 · Selection versus averaging

3. Model Selection and Model Averaging Are Different

Model selection chooses one candidate model as the basis for subsequent inference. The selected model may be chosen using scientific considerations, diagnostics, likelihood-based criteria, information criteria, or a combination of evidence.

Model averaging retains several models and combines their results. The contribution of each model is determined by a model weight.

ApproachBasic questionOutput
Model selectionWhich model should be used?One selected model
Model averagingHow should information from plausible models be combined?Weighted combination of models

Model averaging is therefore not simply another model-selection criterion. It is a way of propagating structural model uncertainty into the quantity of interest.

Important distinction: a model with the lowest information criterion is not automatically the model that should receive 100% of the inferential weight. Model averaging explicitly allows multiple models to contribute.
04 · Candidate models

4. Building a Set of Candidate Models

Model averaging begins with a candidate model set. The set should contain scientifically plausible alternatives rather than arbitrary models generated solely to increase the number of candidates.

For example, suppose a concentration-time dataset can plausibly be described by either a one-compartment or two-compartment disposition model. The candidate set might be:

\[ \mathcal{M}=\{M_1,M_2\} \]

where \(M_1\) is the one-compartment model and \(M_2\) is the two-compartment model.

For a more complex analysis, the candidate set could vary several structural features:

\[ \mathcal{M}=\{M_1,M_2,\ldots,M_K\} \]

Each model is estimated using the same underlying dataset, although the models can contain different structural assumptions.

What belongs in the candidate set?

  • Models that are scientifically plausible for the drug and study.
  • Models that can be estimated with the available data.
  • Models whose structural differences are relevant to the scientific question.
  • Models that have been subjected to appropriate diagnostics.

The candidate set should be defined carefully because model averaging cannot compensate for a candidate set that excludes important plausible alternatives.

05 · Model weights

5. How Are Models Weighted?

Let \(w_k\) denote the weight assigned to candidate model \(M_k\). The weights typically satisfy:

\[ w_k\geq 0,\qquad \sum_{k=1}^{K}w_k=1 \]

A common approach is to derive weights from an information criterion. For example, if \(\mathrm{IC}_k\) is an information criterion for model \(k\), define:

\[ \Delta_k=\mathrm{IC}_k-\min_j(\mathrm{IC}_j) \]

A commonly used information-criterion weighting scheme has the form:

\[ w_k= \frac{\exp(-\Delta_k/2)} {\sum_{j=1}^{K}\exp(-\Delta_j/2)} \]

This expression converts relative information-criterion differences into normalized weights.

The exact interpretation of the weights depends on the criterion, candidate model set, and modeling framework. They should not automatically be interpreted as literal probabilities that a model is the “true” biological model.

Interpretation: a larger weight means that the model receives a larger contribution under the selected weighting procedure. It does not establish that the model is biologically true.
06 · Information criteria

6. AIC-Based Model Averaging

The Akaike information criterion (AIC) balances model fit against model complexity. In its common likelihood-based form:

\[ AIC=-2\log L+2p \]

where \(L\) is the maximized likelihood and \(p\) is the number of estimated parameters.

The first term rewards models that fit the data well, while the second term penalizes additional parameters.

After calculating AIC values for all candidate models, the differences are:

\[ \Delta_k=AIC_k-AIC_{\min} \]

and the corresponding Akaike weights can be calculated as:

\[ w_k= \frac{\exp(-\Delta_k/2)} {\sum_{j=1}^{K}\exp(-\Delta_j/2)} \]

AIC weights are therefore relative weights within the specified candidate set. Adding or removing candidate models can change the normalized weights.

07 · Alternative criteria

7. Other Ways to Construct Model Weights

AIC is not the only possible basis for model weighting. Other criteria may be useful depending on the inferential objective and modeling framework.

Criterion or approachGeneral ideaImportant consideration
AICBalances likelihood and parameter count using a relatively mild complexity penaltyOften used for predictive-oriented model comparison
AICcSmall-sample correction to AICCan be relevant when the effective sample size is not large relative to model complexity
BICUses a stronger complexity penalty related to sample sizeHas a different asymptotic motivation from AIC
Cross-validationEvaluates predictive performance on held-out dataRequires an appropriate validation strategy
Bayesian model averagingWeights models using posterior model probabilitiesDepends on prior distributions and model evidence

These approaches are not interchangeable. The weighting method should match the scientific objective, data structure, estimation framework, and assumptions of the analysis.

08 · Prediction averaging

8. Averaging Predictions

One of the most direct applications of model averaging is to combine model predictions.

Suppose model \(M_k\) produces a prediction \(\hat{y}_k\) and receives weight \(w_k\). The model-averaged prediction is:

\[ \hat{y}_{MA}=\sum_{k=1}^{K}w_k\hat{y}_k \]

For a concentration prediction at time \(t\), this becomes:

\[ \hat{C}_{MA}(t)= \sum_{k=1}^{K}w_k\hat{C}_k(t) \]

This approach is especially natural when the scientific target is a prediction rather than a single parameter.

Model-averaged prediction Time Concentration

The averaged trajectory combines predictions from multiple candidate models. The resulting curve need not correspond to the trajectory generated by any single candidate model.

Key point: a model-averaged prediction can be scientifically useful even when no individual candidate model is treated as definitively correct.
09 · Parameter averaging

9. Can PK Parameters Be Averaged?

Parameters can sometimes be model-averaged, but this requires considerably more care than averaging predictions.

Suppose a parameter \(\theta\) has a directly comparable interpretation across all candidate models. A weighted model-averaged estimate can be written:

\[ \bar{\theta}=\sum_{k=1}^{K}w_k\hat{\theta}_k \]

However, this calculation is meaningful only when \(\theta\) has a sufficiently common interpretation across the models.

For example, clearance may be directly comparable across several candidate structural models. In contrast, a peripheral volume in a two-compartment model has no direct counterpart in a one-compartment model.

QuantityPotentially comparable?Reason
Systemic clearanceOften yesMay represent the same conceptual elimination quantity across models
BioavailabilityOften yesCan have a common interpretation across candidate absorption models
One-compartment volumeNot necessarilyIts interpretation may differ from central or peripheral volumes in multi-compartment models
Peripheral volumeOften noMay exist only in models containing a peripheral compartment
Intercompartmental clearanceOften noOnly defined for models with the relevant distribution structure

For this reason, prediction averaging is often easier to interpret than blindly averaging every parameter in the model.

10 · Uncertainty

10. Model Averaging and Uncertainty

One of the important motivations for model averaging is to recognize that uncertainty can arise from two different sources:

  • Within-model uncertainty: uncertainty in parameter estimates conditional on a particular model.
  • Between-model uncertainty: uncertainty associated with the choice among competing candidate models.

A conventional confidence interval or prediction interval based on one selected model primarily represents uncertainty conditional on that model.

Model averaging attempts to incorporate the additional uncertainty associated with the candidate model set.

Conceptually, if \(Y\) is a prediction and \(M\) denotes the model, the total predictive variance can be decomposed as:

\[ \operatorname{Var}(Y) = E_M[\operatorname{Var}(Y\mid M)] + \operatorname{Var}_M[E(Y\mid M)] \]

The first term represents average within-model uncertainty, while the second represents variation among model-specific predictions.

Why this matters: if competing models make materially different predictions, the between-model component can contribute substantially to overall predictive uncertainty.
11 · Worked example

11. Worked Example: Averaging a PK Prediction

Consider three hypothetical candidate PK models used to predict concentration at 12 hours after a dose. Suppose the models have the following information-criterion values and predictions.

ModelInformation criterion12-h prediction
Model 12104.0 mg/L
Model 22125.0 mg/L
Model 32147.0 mg/L

Step 1: Calculate the minimum criterion

\[ IC_{\min}=210 \]

Step 2: Calculate the differences

\[ \Delta_1=0,\qquad \Delta_2=2,\qquad \Delta_3=4 \]

Step 3: Calculate the relative likelihood quantities

\[ e^{-\Delta_1/2}=1 \] $$ e^{-\Delta_2/2}=e^{-1}\approx0.3679 $$ $$ e^{-\Delta_3/2}=e^{-2}\approx0.1353 $$

Step 4: Normalize the weights

The denominator is:

\[ 1+0.3679+0.1353=1.5032 \]

Therefore:

\[ w_1\approx0.665,\qquad w_2\approx0.245,\qquad w_3\approx0.090 \]

Step 5: Calculate the model-averaged prediction

\[ \hat{C}_{MA} = (0.665)(4.0) + (0.245)(5.0) + (0.090)(7.0) \] $$ \hat{C}_{MA} \approx4.58\text{ mg/L} $$

The model-averaged prediction is therefore approximately 4.58 mg/L.

Notice that the averaged prediction is not identical to the prediction from any one of the three models. It incorporates the predictions of all three according to their relative weights.

Interpretation: the result should be described as a weighted prediction under the specified candidate models and weighting procedure. It is not evidence that a concentration of exactly 4.58 mg/L is generated by a fourth biological model.
12 · Conditional versus marginal inference

12. Conditional and Model-Averaged Inference

Suppose a parameter or prediction is estimated conditional on model \(M_k\). The resulting quantity is model-specific:

\[ \hat{\theta}_k=\hat{\theta}\mid M_k \]

Model averaging instead combines information over the candidate model set:

\[ \bar{\theta}=\sum_{k=1}^{K}w_k\hat{\theta}_k \]

This distinction can be expressed conceptually as:

\[ \text{conditional inference} \quad\longrightarrow\quad \text{given one model} \] $$ \text{model-averaged inference} \quad\longrightarrow\quad \text{across a candidate model set} $$

Model averaging therefore changes the inferential target. Instead of asking only what follows if one model is assumed, it incorporates uncertainty across the specified alternatives.

13 · Population PK

13. Model Averaging in Population Pharmacokinetics

In population PK, model uncertainty can occur at several levels. Structural uncertainty may involve compartment number, absorption model, covariate relationships, interindividual variability, or residual error structure.

For example, consider two candidate population models:

FeatureModel AModel B
DispositionOne compartmentTwo compartments
Clearance covariatePower relationshipPower relationship
Weight effectIncludedIncluded
Residual errorProportionalCombined

If both models remain plausible after diagnostic evaluation, model averaging can be used to combine predictions rather than relying exclusively on one structural specification.

Potential applications include:

  • Population concentration prediction.
  • Individual exposure prediction.
  • Covariate-adjusted predictions.
  • Simulation under alternative dosing regimens.
  • Prediction in populations where structural uncertainty is consequential.
14 · Covariate model averaging

14. Averaging Alternative Covariate Models

Model uncertainty is particularly relevant when the goal is to predict drug exposure for patients with covariate values outside the center of the observed dataset.

Suppose clearance is modeled using body weight in two different ways:

\[ CL_i=CL_{\mathrm{typ}} \left(\frac{WT_i}{70}\right)^{\theta} \]

versus a model with an alternative covariate structure:

\[ CL_i=CL_{\mathrm{typ}} \left[1+\theta\left(\frac{WT_i-70}{70}\right)\right] \]

These relationships can produce similar predictions near the center of the observed weight range but diverge at more extreme values.

Model averaging can combine the resulting predictions when both covariate structures are considered plausible.

Practical implication: model uncertainty can matter more for extrapolation than for interpolation. Two models that appear nearly equivalent within the observed data range can produce meaningfully different predictions outside it.
15 · PK/PD

15. Model Averaging in PK/PD Modeling

Model averaging can also be applied to pharmacodynamic models. Consider alternative exposure-response relationships such as a simple \(E_{\max}\) model and a sigmoid \(E_{\max}\) model.

A simple \(E_{\max}\) model is:

\[ E(C)=E_0+\frac{E_{\max}C}{EC_{50}+C} \]

A sigmoid \(E_{\max}\) model adds a Hill coefficient:

\[ E(C)=E_0+ \frac{E_{\max}C^\gamma} {EC_{50}^\gamma+C^\gamma} \]

If both models are plausible, model-specific predictions can be weighted and combined:

\[ \hat{E}_{MA}(C) = w_1\hat{E}_1(C)+ w_2\hat{E}_2(C) \]

The same principle can be applied to indirect-response models, turnover models, disease-progression models, or other pharmacometric structures when multiple candidate models are scientifically plausible.

16 · Simulation

16. Model Averaging for Simulation

Simulation is often where structural uncertainty becomes especially important. A single selected model can generate a precise-looking simulated distribution even when several plausible models would produce different results.

There are several ways to incorporate model uncertainty into simulation. One conceptual approach is to:

  1. Define the candidate model set.
  2. Estimate each candidate model.
  3. Calculate model weights.
  4. Select a model according to its weight for each simulated replicate.
  5. Simulate from the selected model, including parameter and residual uncertainty as appropriate.
  6. Combine the resulting simulated distribution across replicates.

This approach differs from simply averaging the simulated concentration curves. Sampling a model according to its weight can preserve the fact that each model represents a different data-generating structure.

Two distinct ideas: averaging model predictions produces a weighted expected prediction, whereas sampling from the model set can produce a predictive distribution that reflects between-model variation.
17 · Sources of uncertainty

17. What Uncertainty Should Be Propagated?

Model averaging is only one component of uncertainty analysis. A pharmacometric prediction may contain several sources of uncertainty.

SourceExamplePotential treatment
Parameter uncertaintyUncertainty in CL or VCovariance-based or simulation-based propagation
Interindividual variabilityBetween-subject variability in CLPopulation simulation
Residual variabilityObservation errorResidual-error simulation
Covariate uncertaintyUnknown future patient characteristicsAppropriate covariate distributions or scenarios
Structural model uncertaintyOne vs. two compartmentsModel averaging or model-selection sensitivity analysis

A robust simulation strategy should identify which uncertainty sources are relevant to the decision rather than treating model averaging as a universal substitute for uncertainty propagation.

18 · Limitations

18. Limitations and Practical Considerations

Model averaging can be useful, but it does not eliminate the need for careful pharmacometric reasoning.

  • The candidate model set matters. Important plausible models omitted from the set cannot contribute to the averaged result.
  • Weights depend on the weighting method. AIC, AICc, BIC, cross-validation, and Bayesian approaches can produce different weights.
  • Model weights depend on the candidate set. Adding or removing models can change normalized weights.
  • Parameter averaging requires common interpretation. Parameters that exist only in some models should not automatically be averaged.
  • Averaged predictions need interpretation. The averaged curve may not correspond to the output of any single mechanistic model.
  • Model averaging does not correct structural misspecification automatically. If all candidate models are poor, averaging them does not create a valid model.
  • Complexity can increase. Maintaining, validating, and communicating multiple models requires additional work.
  • Decision relevance matters. Model uncertainty should receive particular attention when competing models produce materially different conclusions for the scientific question.
Modeling principle: model averaging is most useful when the candidate models are scientifically defensible, individually evaluated, and meaningfully different for the quantity of interest.
19 · Practical workflow

19. A Practical Model-Averaging Workflow

  1. Define the scientific question. Decide whether the target is a parameter, concentration prediction, exposure measure, simulation outcome, or another quantity.
  2. Identify structural uncertainties. Determine which modeling choices are genuinely uncertain and scientifically relevant.
  3. Construct the candidate model set. Include plausible alternatives rather than arbitrary variations.
  4. Estimate every candidate model. Use a consistent dataset and appropriate estimation procedures.
  5. Evaluate model adequacy. Examine diagnostics, parameter plausibility, identifiability, and predictive performance.
  6. Select a weighting framework. For example, information-criterion weights or a Bayesian model-averaging framework.
  7. Calculate model weights. Confirm that the weights are nonnegative and sum to one.
  8. Generate model-specific quantities of interest. These may include concentrations, AUC, exposure-response predictions, or simulations.
  9. Combine the results. Average predictions or otherwise integrate across models according to the chosen framework.
  10. Propagate uncertainty. Include parameter uncertainty and other relevant sources in addition to model uncertainty.
  11. Perform sensitivity analyses. Assess how conclusions change under alternative candidate sets or weighting procedures when appropriate.
  12. Communicate the model set and weights. Report enough information for the reader to understand how the averaged result was constructed.
20 · Reporting

20. How Should Model Averaging Be Reported?

A model-averaging analysis should make the source of the averaged result transparent.

At minimum, reporting should describe:

  • The candidate models considered.
  • The scientific rationale for including those models.
  • The estimation method used for each model.
  • The model-comparison criterion or weighting framework.
  • The resulting model weights.
  • The quantity that was averaged.
  • How parameter and predictive uncertainty were propagated.
  • Any sensitivity analyses involving the candidate model set.

For example, a concise report might state that three prespecified structural PK models were evaluated, AIC-based weights were calculated, and concentration predictions were averaged using those weights.

Transparency matters: reporting only the final model-averaged prediction without describing the candidate models and weights makes it difficult to understand what uncertainty has actually been incorporated.
21 · Interpreting the result

21. What Does a Model-Averaged Result Mean?

A model-averaged result should be interpreted as an inference or prediction that integrates across a defined set of candidate models according to a defined weighting scheme.

It does not mean that the drug simultaneously follows several compartment structures in a literal biological sense.

For example, if a model-averaged concentration is:

\[ \hat{C}_{MA}(t)=5.2\text{ mg/L} \]

the appropriate interpretation is that 5.2 mg/L is the weighted combination of the model-specific predictions under the specified candidate set and weights.

The interpretation therefore depends on three things:

  1. Which models were included?
  2. How were they weighted?
  3. What quantity was averaged?
22 · A practical comparison

22. Model Selection Versus Model Averaging

FeatureModel selectionModel averaging
Number of models used for final inferenceUsually oneMultiple
Structural uncertaintyConditioned on selected modelExplicitly incorporated across candidate models
OutputModel-specific estimate or predictionWeighted combination
InterpretabilityUsually straightforward for model parametersRequires explanation of candidate models and weights
Parameter averagingDirect within selected modelRequires parameters to have comparable meanings
Prediction averagingPrediction from selected modelWeighted prediction across models
Computational burdenLower after model selectionHigher because multiple models must be estimated and evaluated

Neither framework removes the need for scientific judgment. The important distinction is whether uncertainty about the structural model is conditioned away through selection or retained as part of the analysis.

23 · Implementation

23. Implementing Model Averaging in Pharmacometric Software

Model averaging is not limited to a particular software package. The general workflow can be implemented whenever the analyst can estimate multiple candidate models, calculate appropriate weights, generate model-specific predictions, and combine the results.

A practical implementation can be organized into four computational layers:

LayerTask
Model fittingEstimate each candidate pharmacometric model
Model comparisonCalculate likelihood-based or predictive criteria
Weight calculationConvert relative model performance into normalized weights
Prediction combinationGenerate and combine model-specific predictions or simulations

For reproducibility, it is useful to maintain a structured record containing each model's specification, parameter estimates, objective function or likelihood, information criterion, weight, and predictions.

This is particularly important when model averaging is performed across a large set of candidate population PK or PK/PD models.

24 · Decision relevance

24. When Does Model Averaging Matter Most?

Model uncertainty becomes particularly important when candidate models make similar predictions within the observed data but diverge for the intended application.

Examples include:

  • Predicting concentrations at sampling times not represented in the original dataset.
  • Simulating alternative dosing regimens.
  • Extrapolating to different patient populations.
  • Predicting exposure at extreme covariate values.
  • Evaluating long-term disease or treatment trajectories.
  • Using a PK/PD model to predict effects at concentrations beyond those directly observed.

If all plausible models lead to essentially the same decision-relevant prediction, structural model uncertainty may have limited practical impact. If predictions diverge substantially, explicitly accounting for model uncertainty can become much more consequential.

25. Key Takeaways

  • Model averaging incorporates uncertainty across multiple plausible pharmacometric models rather than conditioning all inference on one selected model.
  • A candidate model set should contain scientifically plausible and adequately evaluated alternatives.
  • Model weights quantify each candidate model's contribution under the chosen weighting framework.
  • A common information-criterion approach uses differences such as \(\Delta_k=IC_k-IC_{\min}\) to construct normalized weights.
  • AIC-based weights are relative to the specified candidate model set and should not automatically be interpreted as probabilities that models are biologically true.
  • Model-averaged predictions can be calculated as a weighted combination of model-specific predictions.
  • Prediction averaging is generally easier to interpret than parameter averaging when candidate models have different parameterizations.
  • Parameters should only be averaged when they have sufficiently comparable meanings across candidate models.
  • Model averaging can incorporate between-model uncertainty in addition to within-model parameter uncertainty.
  • Simulation can incorporate model uncertainty either by averaging predictions or by sampling from candidate models according to their weights.
  • Model averaging does not compensate for a poor candidate model set or structural misspecification shared by all candidate models.
  • The weighting method, candidate model set, and quantity being averaged should be clearly reported.
  • Model uncertainty is especially relevant when plausible models produce materially different predictions for the intended scientific application.
Next step

Where to Go Next

A natural progression is to study population PK model selection and validation, followed by covariate model building, bootstrap and visual predictive checks, Bayesian model averaging, and simulation-based evaluation of structural model uncertainty.

The next tutorial can build directly on this topic by examining population PK model selection in detail, including structural model development, covariate selection, diagnostic evaluation, and the distinction between statistical criteria and pharmacometric plausibility.