1. What Is Bayesian Model Averaging?
Bayesian model averaging (BMA) is a framework for making inference while explicitly accounting for uncertainty about which candidate model generated the data.
In PK/PD modeling, the uncertainty may concern questions such as whether a drug follows one- or two-compartment disposition, whether absorption is best represented by a first-order or more complex input model, whether an effect compartment is necessary, or which covariates belong in a population model.
A conventional workflow often chooses one model and then treats that model as if it were known with certainty. BMA instead allows several plausible models to contribute to the final inference, with each model receiving a weight based on its posterior probability.
BMA propagates uncertainty about the structural model into the final inference rather than conditioning all inference on one selected model.
2. Why Is Model Uncertainty Important in PK/PD?
PK/PD models are often nonlinear and mechanistically structured. Several models can fit the available observations reasonably well while making different assumptions about unobserved processes.
For example, consider concentration and effect observations collected after repeated dosing. Plausible models might include:
- a direct-effect model in which effect responds immediately to plasma concentration;
- an effect-compartment model with a delay between plasma and effect-site concentrations;
- a turnover model in which drug inhibits or stimulates the production or loss of an endogenous response;
- a mechanistic receptor or target-mediated model;
- alternative covariate structures for clearance, volume, or PD sensitivity.
If several models remain scientifically plausible, selecting one model and ignoring the others can understate uncertainty. BMA provides a formal way to propagate that structural uncertainty into parameter estimates and predictions.
3. Bayes' Theorem for Competing PK/PD Models
Let \(M_k\) denote candidate model \(k\), and let \(y\) denote the observed PK/PD data. Bayesian model comparison begins with:
Here:
- \(P(M_k)\) is the prior probability assigned to model \(k\).
- \(P(y\mid M_k)\) is the marginal likelihood of the observed data under model \(k\).
- \(P(M_k\mid y)\) is the posterior model probability.
- \(K\) is the number of candidate models being considered.
The denominator ensures that the posterior model probabilities sum to one:
The key quantity is the marginal likelihood. Unlike an ordinary likelihood evaluated at one parameter estimate, the marginal likelihood integrates over the model's parameter uncertainty:
This integration means that a model is evaluated not only on how well it can fit the data, but also on how much prior-supported parameter space produces that fit.
4. How Does Bayesian Model Averaging Work?
Suppose the quantity of scientific interest is \(\Delta\), such as clearance, an exposure metric, an effect at a future time point, or a predicted concentration.
Within each model, Bayesian inference gives a posterior distribution:
BMA combines these model-specific posterior distributions using their posterior model probabilities:
This is a mixture distribution. It contains both parameter uncertainty within each model and model uncertainty across the candidate models.
For a posterior mean, the corresponding model-averaged estimate is:
Thus, a model with posterior probability 0.70 contributes 70% of its model-specific posterior expectation to the model-averaged expectation, while a model with posterior probability 0.20 contributes 20%.
5. Posterior Model Probabilities
The posterior model probability is the natural BMA weight. For two models, the posterior odds can be written as:
The second ratio is the Bayes factor in favor of \(M_1\) over \(M_2\):
Consequently, model evidence and prior model probabilities both matter.
| Quantity | Interpretation | Role in BMA |
|---|---|---|
| Prior model probability | How plausible a model is before seeing the current data | Provides the starting model weight |
| Marginal likelihood | How well the model predicts the observed data after integrating over its parameters | Updates the model weight |
| Posterior model probability | Updated support for a candidate model after observing the data | Used as the averaging weight |
| Bayes factor | Relative evidence supplied by the data for one model versus another | Quantifies relative model evidence |
6. BMA for Competing PK Models
Consider a concentration-time dataset for an orally administered drug. Suppose the following structural models are scientifically plausible:
| Model | Structural assumptions |
|---|---|
| M₁: One compartment | Central disposition with first-order elimination |
| M₂: Two compartments | Central and peripheral disposition with intercompartmental exchange |
| M₃: Two compartments + flexible absorption | Two-compartment disposition with an absorption model that allows more complex input behavior |
Each model has its own parameter vector, for example:
and a more complex parameter vector for \(M_3\).
The important point is that these models do not necessarily define the same parameter space. A parameter such as peripheral volume exists naturally in a two-compartment model but has no direct counterpart in a one-compartment model.
7. BMA for Competing PD Models
Model uncertainty is often even more consequential on the PD side because different mechanistic hypotheses can produce similar observations over the sampled time range.
Suppose concentration is linked to effect using three candidate models.
Direct Emax model
Effect-compartment model
Indirect-response model
For an inhibitory drug effect on response production, one possible structure is:
These models make different mechanistic assumptions about the relationship between exposure and response. BMA can retain uncertainty across these hypotheses instead of forcing the analysis to condition on one selected PD mechanism.
8. Model-Averaged Predictions
Prediction is one of the most intuitive applications of BMA in PK/PD.
Suppose a future concentration at time \(t^\*\) is predicted under each model:
The model-averaged predictive distribution is:
The corresponding posterior predictive mean is:
This distinction is important: BMA does not simply average point estimates. In a full Bayesian analysis, it averages entire predictive distributions.
A model-averaged prediction can fall between competing model predictions, while its predictive uncertainty can reflect disagreement among the models as well as parameter uncertainty within each model.
9. What About PK/PD Parameter Estimates?
Parameter averaging requires more care than prediction averaging.
Suppose every candidate model contains clearance \(CL\). Then a model-averaged posterior for clearance can be constructed as:
However, some structural parameters are model-specific. For example, \(Q\), the intercompartmental clearance, exists in a two-compartment model but not in a one-compartment model.
In such cases, averaging the numerical values directly can be scientifically meaningless. A better strategy is often to average a common derived quantity.
| Quantity | Usually suitable for cross-model averaging? | Reason |
|---|---|---|
| Predicted concentration | Yes | All models can generate a prediction on the same concentration scale |
| AUC | Often | Can be defined consistently across many structural models |
| Cmax | Often | Common prediction target |
| Clinical response | Often | Can be expressed on a common response scale |
| Clearance | Often | Usually has a common scientific interpretation across candidate PK structures |
| Peripheral volume | Not always | May not exist in simpler models |
| Intercompartmental clearance | Not always | Only defined in models containing the corresponding compartmental structure |
10. Worked Example: Averaging Two PK Models
Consider two candidate PK models for a hypothetical drug:
- Model 1: one-compartment model.
- Model 2: two-compartment model.
Assume equal prior model probabilities:
Suppose the marginal likelihoods obtained from the Bayesian analysis are:
Step 1: Calculate the model evidence
Step 2: Calculate posterior model probabilities
The data therefore update the model probabilities from 50/50 before observing the data to 40/60 afterward.
Step 3: Predict a concentration at 12 hours
Suppose the posterior predictive means at 12 hours are:
The model-averaged prediction is:
Step 4: Interpret the result
The model-averaged estimate of 6.8 mg/L is not the concentration predicted by either model individually. It is the posterior expectation after accounting for uncertainty about which structural model is appropriate.
11. Why Model Averaging Can Increase Predictive Uncertainty
Suppose each model has a posterior predictive mean and variance:
The variance of the model-averaged distribution can be decomposed using the law of total variance:
The first term represents the average within-model uncertainty. The second term represents between-model disagreement.
This is one of the most important conceptual advantages of BMA. If two models make substantially different predictions, that disagreement contributes to the final uncertainty distribution rather than disappearing after model selection.
12. Bayesian Model Averaging for Covariate Models
Model uncertainty in population PK/PD frequently involves covariate selection.
Suppose clearance is modeled as:
Alternative models might add renal function, age, sex, or other clinically relevant covariates.
For example:
| Model | Clearance covariates |
|---|---|
| M1 | Weight |
| M2 | Weight + renal function |
| M3 | Weight + age |
| M4 | Weight + renal function + age |
BMA can distribute posterior probability across these candidate covariate structures rather than treating one selected covariate model as certain.
A particularly useful summary is the posterior inclusion probability for a covariate:
This answers a different question from identifying the single highest-probability model. It measures the total posterior support for including the covariate across the model space.
13. How Is BMA Computed?
The computational strategy depends on the size and structure of the candidate model space.
Enumerating a small model set
If only a few candidate models are scientifically plausible, each model can be fit separately. Posterior model probabilities can then be calculated from model evidence or Bayes factors.
Stochastic model search
If there are many possible models, enumerating every combination can become computationally infeasible. Bayesian model determination can instead use stochastic search methods.
Reversible-jump Markov chain Monte Carlo (RJMCMC) is one approach that allows an MCMC algorithm to move between models with different parameter dimensions.
Conceptually, the chain samples:
Posterior model probabilities can then be estimated from the proportion of posterior samples associated with each model.
Prediction-based averaging
In some applied workflows, model-specific predictions are combined using weights derived from model support. This can be useful for practical prediction, but the interpretation of the weights depends on how they were constructed.
14. The Role of Prior Model Probabilities
BMA requires prior probabilities for the candidate models. A common starting point is equal prior probability:
Equal prior probabilities are convenient, but they are not mandatory.
Scientific knowledge can motivate unequal prior model probabilities. For example, a mechanistically implausible model might receive less prior probability than a model strongly supported by previous experimental evidence.
Parameter priors also matter because marginal likelihoods integrate over the parameter space. Consequently, Bayesian model comparison can be sensitive to prior choices, particularly when the data are not highly informative.
| Prior choice | Potential consequence |
|---|---|
| Equal model probabilities | Starts all candidate structures with the same model weight |
| Informative model probabilities | Allows prior scientific knowledge to influence model comparison |
| Very diffuse parameter priors | Can substantially affect marginal likelihoods because the prior parameter volume enters the evidence calculation |
| Mechanistically informed priors | Can incorporate prior biological or clinical knowledge into the model comparison |
Prior sensitivity analysis is therefore an important part of a serious Bayesian model-averaging analysis.
15. Bayesian Model Averaging Versus Choosing One Model
Model selection and model averaging answer related but different questions.
| Approach | Primary question | What happens to model uncertainty? |
|---|---|---|
| Single-model selection | Which candidate model will be used for inference? | Usually treated as resolved after selection |
| Bayesian model averaging | What should we infer after accounting for uncertainty across candidate models? | Propagated into the final posterior or predictive distribution |
| Model comparison | How does the evidence differ across candidate models? | Summarized through quantities such as Bayes factors or posterior model probabilities |
These approaches are not mutually exclusive. Model comparison may be used to understand the evidence, model selection may be appropriate when a single operational model is required, and BMA may be especially useful when structural uncertainty itself is scientifically important.
16. A Practical BMA Workflow for PK/PD
- Define the scientific question. Decide whether the main target is a PK parameter, exposure metric, PD parameter, biomarker trajectory, clinical response, or future prediction.
- Define the model space. Specify scientifically plausible structural and covariate models before inspecting results from the competing models.
- Specify priors. Define prior distributions for model probabilities and model-specific parameters.
- Fit the candidate models. Estimate the posterior distribution within each candidate model or jointly sample the model space.
- Assess computational convergence. Check MCMC diagnostics and, when model-space sampling is used, verify that relevant models are adequately explored.
- Calculate model support. Obtain posterior model probabilities or another explicitly defined model-weighting quantity.
- Generate model-specific predictions. Calculate concentrations, AUC, Cmax, responses, or other quantities of interest under each model.
- Average across models. Combine posterior distributions or predictions using the model weights.
- Perform sensitivity analyses. Examine how conclusions change with alternative priors, model spaces, or plausible structural assumptions.
- Validate predictions. Use posterior predictive checks, external validation, or prospective prediction when appropriate.
The model space is particularly important. BMA can account for uncertainty among the models that are included, but it cannot automatically protect against a model space that excludes an important scientifically plausible structure.
17. How Should a BMA Analysis Be Evaluated?
Model averaging does not eliminate the need for model diagnostics. Several checks remain important.
- Within-model diagnostics: examine residuals, posterior predictive checks, parameter plausibility, and convergence for each important model.
- Model-space diagnostics: determine whether the candidate models adequately represent the scientific hypotheses of interest.
- Posterior model probability diagnostics: verify that the estimated model weights are stable under the computational method.
- Prior sensitivity: evaluate whether conclusions are strongly dependent on model or parameter priors.
- Prediction diagnostics: compare model-averaged predictions with observed data or independent validation data.
- Model disagreement: identify situations where competing models make materially different predictions.
A model-averaged result can look stable simply because one model dominates the posterior model probability. Conversely, substantial model disagreement may be scientifically important even when the averaged prediction itself appears reasonable.
18. Limitations and Common Pitfalls
1. Treating BMA as automatic protection against misspecification
BMA averages over the specified model space. If every candidate model is inappropriate, averaging them does not solve the underlying problem.
2. Averaging parameters that do not have a common meaning
Model-specific parameters should not be averaged mechanically merely because they have similar names or numerical units.
3. Ignoring prior sensitivity
Posterior model probabilities depend on model priors and parameter priors through the marginal likelihood. Sensitivity analysis can therefore be important.
4. Confusing posterior probability with predictive accuracy
A model can have high posterior probability under the assumed Bayesian specification while still performing poorly for a particular prediction task. Predictive validation remains important.
5. Using an overly large model space without adequate computation
Large numbers of candidate covariates and structural alternatives can create a massive model space. Computational strategies must be appropriate for the number and complexity of candidate models.
6. Averaging across scientifically incompatible models
Candidate models should represent plausible alternative explanations of the same scientific process. Including arbitrary or implausible models can make the resulting model probabilities difficult to interpret.
19. Where Can BMA Be Used in Pharmacometric Modeling?
| Application | Example model uncertainty | Potential BMA target |
|---|---|---|
| Structural PK | One versus two compartments | Concentration, AUC, Cmax, exposure |
| Absorption | First-order, zero-order, transit, or mixed input | Exposure and concentration predictions |
| PD delay | Direct effect versus effect compartment | Future effect predictions |
| Mechanism-based PD | Direct versus turnover versus mechanistic model | Effect trajectory or dose-response prediction |
| Population PK | Alternative covariate structures | Individual or population exposure |
| Population PD | Alternative covariate-effect relationships | Predicted response |
| MIDD | Alternative mechanistic assumptions | Scenario-based prediction |
| Model-informed dosing | Multiple plausible published PK models | Individual exposure or target concentration prediction |
The common theme is that the scientific target is often a prediction or derived quantity that can be defined consistently across models.
20. BMA and Simulation
Simulation is particularly useful because BMA naturally produces a distribution over possible future trajectories.
For each posterior draw:
- sample a model according to its posterior model probability;
- sample model parameters from that model's posterior distribution;
- simulate the PK/PD system under the desired dose or covariate scenario;
- repeat many times to generate a posterior predictive ensemble.
Conceptually, the process is:
The resulting simulated trajectories incorporate both parameter uncertainty and structural model uncertainty.
This can be valuable for dose selection, exposure-response simulation, clinical trial simulation, and model-informed precision dosing when the structural model is not known with certainty.
21. Worked Example: Model-Averaged PK/PD Response
Suppose three PD models have been fit to a hypothetical drug:
| Model | Description | Posterior probability | Predicted effect at 24 h |
|---|---|---|---|
| M1 | Direct Emax | 0.20 | 62 units |
| M2 | Effect compartment + Emax | 0.50 | 70 units |
| M3 | Indirect response | 0.30 | 76 units |
The model-averaged posterior mean effect is:
The result is close to the prediction from the second model because that model has the greatest posterior probability, but all three models contribute.
More importantly, the full model-averaged posterior would retain the uncertainty associated with the differences among these models. The final interval therefore need not be identical to the interval from the highest-probability model alone.
22. How Should Model-Averaged Results Be Reported?
A transparent BMA analysis should make it possible for a reader to understand both the model uncertainty and the final inference.
At minimum, report:
- the candidate model space;
- the prior model probabilities;
- the parameter priors or important prior assumptions;
- the method used to calculate or approximate marginal likelihoods;
- posterior model probabilities;
- the target quantity being averaged;
- model-specific estimates or predictions when informative;
- the final model-averaged estimate and uncertainty interval;
- sensitivity analyses for important prior or model-space choices.
For example, a concise result might state that three prespecified PK/PD structures were considered, that their posterior model probabilities were 0.20, 0.50, and 0.30, and that the model-averaged 24-hour effect was 70.2 units with its corresponding posterior credible interval.
23. BMA Within the Broader Bayesian PK/PD Framework
Bayesian PK/PD analysis can be viewed as containing several nested layers of uncertainty:
Traditional Bayesian inference usually conditions on one structural model and propagates parameter uncertainty. BMA adds another layer by treating the structural model itself as uncertain.
This distinction becomes particularly important when multiple mechanistic explanations are compatible with the available data.
BMA extends Bayesian uncertainty propagation by allowing the structural model itself to be uncertain.
24. Key Takeaways
- Bayesian model averaging accounts for uncertainty about which candidate model is appropriate.
- In PK/PD, model uncertainty can involve compartments, absorption models, effect models, mechanisms, covariates, or other structural assumptions.
- Posterior model probabilities combine prior model probabilities with the marginal likelihood of the observed data.
- BMA averages posterior distributions or predictions across models rather than simply selecting one model.
- The model-averaged posterior is a mixture of the model-specific posterior distributions.
- Model-averaged prediction incorporates both parameter uncertainty within models and disagreement among models.
- Common derived quantities such as concentration, AUC, Cmax, or clinical response are often more appropriate averaging targets than model-specific structural parameters.
- Posterior inclusion probabilities can summarize uncertainty about population PK/PD covariates.
- Model averaging does not protect against a poor model space: important scientific alternatives must still be included.
- Prior model probabilities and parameter priors can influence Bayesian model comparison and should be considered in sensitivity analyses.
- Model diagnostics, convergence assessment, predictive validation, and prior sensitivity remain essential in BMA analyses.
- BMA is especially useful when several scientifically plausible PK/PD models make materially different predictions and the structural uncertainty should be propagated into the final inference.
Where to Go Next
A natural progression is to study Bayesian population PK/PD modeling, followed by Bayesian estimation of individual PK parameters, prior specification, MCMC diagnostics, posterior predictive checks, Bayesian covariate selection, and model-informed precision dosing.
For more advanced work, BMA can be extended to hierarchical population models, mechanistic PK/PD systems, nonlinear mixed-effects models, and simulation-based decision making. The central idea remains the same: uncertainty about the model should be propagated whenever that uncertainty is scientifically relevant.
25. References
- Hoeting JA, Madigan D, Raftery AE, Volinsky CT. Bayesian Model Averaging: A Tutorial. Statistical Science. 1999;14(4):382–401. DOI: 10.1214/ss/1009212519.
- Lunn DJ, Best N, Thomas A, Wakefield J, Spiegelhalter D. Bayesian analysis of population PK/PD models: general concepts and software. Journal of Pharmacokinetics and Pharmacodynamics. 2002;29(3):271–307. DOI: 10.1023/A:1020206907668.
- Lunn DJ. Automated covariate selection and Bayesian model averaging in population PK/PD models. Journal of Pharmacokinetics and Pharmacodynamics. 2008;35(1):85–100. DOI: 10.1007/s10928-007-9077-x.
- Csajka C, Verotta D. Pharmacokinetic-pharmacodynamic modelling: history and perspectives. Journal of Pharmacokinetics and Pharmacodynamics. 2006;33(3):227–279. DOI: 10.1007/s10928-005-9002-0.
- Gould AL. BMA-Mod: A Bayesian model averaging strategy for determining dose-response relationships in the presence of model uncertainty. Biometrical Journal. 2019;61:1141–1159. DOI: 10.1002/bimj.201700211.
- Yates JWT. Integrated pharmacokinetic-pharmacodynamic and agent-based modelling in drug development: Current status and future perspectives. Advanced Drug Delivery Reviews. 2025;216:115474. DOI: 10.1016/j.addr.2024.115474.
These references provide background on Bayesian model averaging, Bayesian population PK/PD analysis, covariate model averaging, and the broader use of PK/PD models in quantitative drug development.