Empirical Bayes Estimates in Population PK

1. What Is an Empirical Bayes Estimate?

An empirical Bayes estimate (EBE) is a subject-specific estimate of a random effect or individual pharmacokinetic parameter obtained by combining information from the population PK model with the concentration data observed for that individual.

In nonlinear mixed-effects modeling, the population model describes typical pharmacokinetic behavior and between-subject variability. The EBE step then asks a more individual question: given the fitted population model and this subject's observations, what individual parameter values are most consistent with both sources of information?

EBEs are also commonly called post hoc estimates. In the classical NONMEM workflow, they are obtained after population parameters such as the typical values and between-subject variability have been estimated.

2. Where Do EBEs Come From?

Consider a typical population PK model in which the individual parameter for subject \(i\) is represented as:

Here, \(\theta_{\mathrm{pop}}\) represents the population parameter values and \(\eta_i\) represents the subject-specific random effects. For a simple exponential model:

The random effect \(\eta_{CL,i}\) describes how subject \(i\)'s clearance differs from the typical population value. A common assumption is:

The matrix \(\Omega\) describes between-subject variability and, when appropriate, relationships among the random effects.

The population model therefore supplies a prior distribution for an individual's parameters. The individual's concentration data then provide additional information about where that person lies within the population distribution.

3. The Bayesian Interpretation

The Bayesian interpretation is particularly useful for understanding EBEs. Conceptually, the posterior distribution of the individual random effect is proportional to the product of the individual's likelihood and the population prior:

where \(y_i\) represents the concentration observations for subject \(i\), \(\psi\) contains the estimated population parameters, and \(\Omega\) describes the estimated between-subject variability.

The EBE is commonly the mode of this posterior distribution. In other words, it is the value of the individual random effect that maximizes the posterior density, conditional on the fitted population model and the individual's observations.

This is why the term maximum a posteriori (MAP) estimate is often used when describing the EBE concept.

4. The Balance Between Population Information and Individual Data

An EBE can be understood as a compromise between two sources of information:

• Population information: what the fitted model says is plausible for a typical individual and for the distribution of individual variability.
• Individual information: what the subject's observed concentration data say about that particular subject.

When the individual has little informative data, the population model has greater influence. When the individual has highly informative data, the observations exert greater influence.

This relationship can be illustrated conceptually:

This behavior is one reason EBEs are particularly useful in sparse-data population PK: the population model provides information that the individual's observations alone may not contain.

5. A One-Parameter Example

Suppose the population model estimates typical clearance as:

and estimates substantial between-subject variability. Now suppose an individual has concentration observations suggesting that their clearance may be higher than typical. The EBE combines that evidence with the population distribution.

If the observations are sparse, the resulting estimate may be closer to 5 L/h than a subject-specific estimate obtained from the observations alone.

If the observations are rich and strongly identify clearance, the EBE may move considerably away from 5 L/h.

The exact amount of movement depends on the model, the residual error, the amount and timing of individual data, and the estimated between-subject variability.

6. From EBEs to Individual PK Parameters

In many population PK models, the EBE is reported as an estimate of the random effect \(\eta\), while the clinically interpretable quantity is an individual parameter such as clearance or volume.

For example, under an exponential random-effects model:

If:

then:

The EBE of \(\eta\) therefore determines the individual's model-based parameter estimate through the structural relationship used in the population model.

7. Why Are EBEs Called “Post Hoc” Estimates?

Population estimation first produces estimates of the fixed effects and variability parameters. Once those population parameters are available, an individual-level conditional estimation step can be performed.

In simplified form, the workflow is:

The term “post hoc” therefore refers to the ordering of the estimation process: the individual estimates are obtained conditional on population parameter estimates rather than being estimated as independent parameters for every subject.

This distinction is important because a population PK analysis is not simply a collection of separate PK analyses performed subject by subject.

8. What Is EBE Shrinkage?

Because EBEs combine individual information with the population distribution, they can become systematically closer to the population mean when individual data contain limited information.

This phenomenon is called shrinkage.

For a random effect whose population standard deviation is \(\omega\), a commonly used summary of eta-shrinkage is:

When the empirical standard deviation of the EBEs is substantially smaller than the estimated population variability, the EBE distribution has contracted toward zero.

For example, if:

then:

or approximately 30%.

9. Why Do EBEs Shrink?

Several features of a study can reduce the amount of information available for estimating individual parameters.

• Few observations per subject. Sparse PK sampling may provide limited information about individual clearance, volume, or absorption.
• Uninformative sampling times. Samples taken at times that do not distinguish competing parameter values may contribute relatively little information.
• High residual variability. Large measurement or unexplained variability can make the individual's observations less informative.
• Small between-subject variability. When the population distribution is narrow, individual estimates are naturally constrained near the population value.
• Parameter correlation. Strongly correlated parameters may be difficult to identify separately from sparse data.
• Model structure. Some parameters are inherently less identifiable than others under a particular sampling design.

Shrinkage can therefore be viewed as an interaction between the information in the individual data and the information supplied by the population model.

10. Population Shrinkage and Individual Information

A population-level shrinkage statistic summarizes behavior across subjects, but it does not imply that every individual has the same degree of shrinkage.

The information available for one subject can depend on their actual sampling schedule, concentration values, dose history, residual error, and where their true parameters lie relative to the population distribution.

This distinction becomes especially important in therapeutic drug monitoring and other applications in which individual parameter estimates are used for subsequent decisions.

Consequently, a single shrinkage statistic should not be interpreted as a complete description of the uncertainty of every individual estimate.

11. Why Are EBEs Useful for Model Diagnostics?

EBEs are widely used to generate individual-level diagnostic quantities. For example, an EBE can be used to calculate an individual's predicted concentration:

where \(IPRED\) denotes an individual prediction and \(t_{ij}\) is an observation time.

Comparisons between observed concentrations and individual predictions can help reveal patterns that are not obvious from population predictions alone.

Examples include:

• Observed versus individual-predicted concentrations.
• Individual random effects versus covariates.
• Distributions of EBEs.
• Relationships among individual random effects.
• Individual predictions and residual-based diagnostics.

However, the usefulness of these diagnostics depends on the amount of shrinkage. When shrinkage is substantial, EBE-based relationships can become distorted or obscured.

12. EBEs and Covariate Relationships

A common use of individual parameter estimates is exploratory assessment of relationships between PK parameters and patient characteristics such as body size, renal function, age, or other prespecified covariates.

For example, suppose:

A plot of EBE-derived individual clearance against body weight may provide a useful visualization of the relationship.

But shrinkage can make the relationship appear weaker than it truly is, alter its apparent shape, or under some circumstances create misleading patterns. Consequently, formal covariate model evaluation should generally rely on the population model and appropriate likelihood-based or other model-based methods rather than treating EBEs as if they were directly observed clearance values.

13. Is Shrinkage “Bad”?

Shrinkage is not automatically evidence that a population PK model is wrong. It is often a consequence of limited information at the individual level.

For example, sparse sampling can be scientifically appropriate and may still support reliable estimation of population parameters even though some individual EBEs exhibit substantial shrinkage.

The important question is therefore not simply whether shrinkage exists, but whether its magnitude affects the particular use being made of the EBEs.

Published simulation work has shown that substantial eta-shrinkage can compromise the interpretation of EBE-based diagnostics. A commonly cited practical warning point is roughly 20–30%, although this should not be treated as a universal pass/fail threshold. The appropriate interpretation depends on the model, parameter, data richness, and intended application.

14. Worked Example: Estimating Individual Clearance

Consider a hypothetical one-compartment IV population PK model. Suppose the population analysis estimates:

• Typical clearance: \(CL_{\mathrm{pop}}=5\) L/h
• Between-subject SD on log clearance: \(\omega_{CL}=0.40\)
• Residual variability: represented by an appropriate proportional error model

For one individual, the concentration observations suggest a clearance somewhat above the population typical value.

Step 1: Population model

Step 2: Individual observations

Suppose the individual's data produce an EBE:

Step 3: Convert the EBE to individual clearance

Step 4: Interpret the result

The model therefore estimates this individual's clearance as approximately 6.11 L/h, compared with the population typical value of 5 L/h.

The important point is that 6.11 L/h should not be interpreted as though it were a directly measured clearance. It is the value supported by the individual's observations after accounting for the population distribution and residual variability.

Step 5: What if the individual's data were sparse?

If the concentration data contained little information about clearance, the EBE might be closer to zero on the \(\eta\) scale. For example, an EBE of 0.05 would give:

This illustrates the central idea of shrinkage: when the individual's data are weakly informative, the population model has more influence.

15. EBEs and Individual Predictions

Once individual parameters have been estimated, the model can generate an individual prediction for each observed or hypothetical time point.

For a one-compartment IV bolus model:

If \(CL_i\) and \(V_i\) are replaced by their EBE-derived individual values, the resulting curve is an individual prediction.

A useful distinction is:

• PRED: prediction based primarily on population-typical parameters and covariates.
• IPRED: prediction using individual parameter estimates, typically incorporating EBEs.
• OBS: the observed concentration.

The difference between these quantities is often central to population PK diagnostic plots.

16. EBEs, IPRED, and Residuals

Individual predictions can be used to construct residual diagnostics. For example, an individual weighted residual can compare an observed concentration with its individual prediction after accounting for the assumed residual error model.

Conceptually:

The exact formula depends on the observation model.

One important consequence of shrinkage is that residual diagnostics based on EBEs can appear unusually well behaved when the individual estimates have insufficient information. This occurs because the same data used to estimate the individual parameters are then used to evaluate the resulting predictions.

Therefore, diagnostic interpretation should consider both eta-shrinkage and epsilon-shrinkage rather than relying on an apparently attractive residual plot alone.

17. What Does Shrinkage Look Like?

Suppose a population model estimates a broad distribution of true individual clearances. If the individual EBEs are highly informative, the EBE distribution may resemble the underlying population variability.

With substantial shrinkage, the EBE distribution becomes narrower.

18. Why Can Some Parameters Shrink More Than Others?

Shrinkage is often parameter-specific. The study design may contain substantial information about clearance but little information about an absorption parameter.

For example, in a sparse oral PK study:

• Clearance may be reasonably informed by concentrations collected during the elimination phase.
• Volume may be less precisely identified if the early distribution phase is poorly sampled.
• Absorption rate may be weakly informed if there are few samples around \(T_{\max}\).

Consequently, one should avoid summarizing a model with a single generic statement such as “the EBEs have 30% shrinkage.” Shrinkage should generally be examined separately for the important random effects.

19. What Are EBEs Used For?

EBEs have several important applications in pharmacometrics.

19.1 Individual prediction

EBEs allow the fitted population model to generate subject-specific concentration predictions and model-based PK profiles.

19.2 Exposure estimation

Individual clearance estimates can be used to derive model-based exposure measures. For a linear IV dose:

When \(CL_i\) is EBE-derived, the resulting AUC is a model-based individual estimate, not a directly observed AUC.

19.3 Therapeutic drug monitoring

Bayesian individualization can use population PK information together with a patient's concentration measurements to estimate individual PK and support dose adjustment. The quality of those individual estimates depends on the model and the information in the patient's data.

19.4 Model diagnostics

EBEs can support visual exploration of individual variability, parameter-covariate relationships, and individual predictions.

19.5 Sequential PK/PD modeling

Individual PK estimates can sometimes be used as inputs to subsequent analyses, although the impact of shrinkage and the propagation of uncertainty must be considered carefully.

20. Are EBEs the Same as Full Bayesian Estimates?

EBEs and full Bayesian inference are related but should not be treated as identical.

An EBE is generally a subject-specific conditional estimate obtained after fitting the population model. The population parameters are treated as fixed at their estimated values during the post hoc step.

A full Bayesian analysis instead specifies prior distributions for unknown parameters and propagates uncertainty in those parameters through the posterior distribution. Rather than producing only a single individual estimate, a Bayesian analysis can provide a posterior distribution for the individual parameter.

The two approaches can therefore produce related individual parameter estimates while answering somewhat different inferential questions.

21. What EBEs Do Not Tell You Automatically

An EBE is useful, but several limitations should remain explicit.

• An EBE is not an observed PK parameter. It is estimated from a model.
• An EBE is conditional on the population model. A misspecified structural or stochastic model can affect the individual estimates.
• High shrinkage limits interpretation. The population model may dominate the individual's data.
• Different parameters can have different information content. Clearance may be well informed while absorption parameters are not.
• Point estimates do not fully represent uncertainty. An EBE alone does not provide the complete uncertainty distribution of the individual's parameter.
• EBE-based correlations can be misleading. Shrinkage can obscure, distort, or sometimes induce apparent relationships.
• Exposure derived from EBEs inherits model limitations. Individual AUC or other exposure measures can be sensitive to shrinkage in the underlying parameters.

22. A Practical EBE Workflow

1. Develop an adequate population PK model. Establish a reasonable structural model, variability model, residual error model, and covariate structure.
2. Estimate population parameters. Obtain typical values and between-subject variability using an appropriate nonlinear mixed-effects estimation method.
3. Evaluate the population model. Use population-level diagnostics, parameter plausibility, model comparisons, and appropriate predictive checks.
4. Obtain EBEs. Perform the individual conditional/post hoc estimation step using the fitted population model.
5. Calculate individual parameters. Transform EBEs into clearance, volume, absorption parameters, or other quantities according to the model.
6. Assess shrinkage. Calculate and report shrinkage for important random effects.
7. Inspect EBE-based diagnostics carefully. Look at EBE distributions, EBE-covariate plots, IPRED plots, and residual diagnostics while considering shrinkage.
8. Use individual estimates for their intended purpose. Distinguish exploratory diagnostics from individual prediction, exposure estimation, TDM, or downstream PK/PD analysis.

23. Common Mistakes When Interpreting EBEs

Mistake 1: Treating an EBE as the “true” individual parameter

An EBE is a model-based estimate. It is not a direct measurement of the individual's true clearance, volume, or absorption rate.

Mistake 2: Ignoring shrinkage

A narrow EBE distribution can look reassuring, but it may indicate that individual data contain insufficient information to distinguish subjects from one another.

Mistake 3: Using EBEs as though they were independent observations

EBEs are estimated from the same population model and concentration data that generated the analysis. They therefore have model-dependent properties and uncertainty.

Mistake 4: Assuming more samples always eliminate shrinkage

The information content of samples depends on when they are collected and which parameters they inform. Additional observations are not automatically informative for every parameter.

Mistake 5: Treating 20% or 30% as a universal cutoff

Published literature commonly discusses shrinkage in this range as a practical warning for EBE-based diagnostics, but the appropriate interpretation depends on the model, parameter, sampling design, and intended analysis.

Mistake 6: Confusing EBE estimation with full Bayesian inference

EBEs are conditional individual point estimates. Full Bayesian inference explicitly propagates uncertainty through posterior distributions.

24. Worked Diagnostic Example: Interpreting 35% Shrinkage

Suppose a population PK model estimates:

and the standard deviation of the estimated individual \(\eta_{CL}\) values is:

The approximate eta-shrinkage is:

This indicates substantial contraction of the EBE distribution relative to the estimated population variability.

The correct interpretation is not simply that the model is “bad.” Instead, it raises a specific question:

For example, if the EBEs are being used primarily for exploratory visualization, the plots may still be informative but should be interpreted cautiously. If they are being used to estimate highly individualized exposure or to infer a subtle covariate relationship, the shrinkage may be more consequential.

25. How Study Design Influences EBE Quality

The quality of individual parameter estimates begins with study design.

This illustrates a broader principle in population PK: model-based individualization cannot recover information that was never contained in the observations.

26. The EBE Concept in One Equation

The central Bayesian idea can be summarized as:

The EBE is a point estimate obtained from this posterior relationship. If the individual's data are highly informative, the likelihood contributes strongly. If the data are sparse or noisy, the population distribution contributes relatively more.

This is the fundamental reason EBEs are useful in population PK: they allow individual information to be interpreted in the context of the population rather than forcing each subject to be analyzed independently.

Where to Go Next

A natural next topic is shrinkage in population PK models. That tutorial can build on the EBE framework introduced here and examine eta-shrinkage, epsilon-shrinkage, individual versus population shrinkage, how shrinkage is calculated, and how substantial shrinkage affects diagnostic plots and covariate analysis.

From there, the natural progression is to Bayesian therapeutic drug monitoring, individualized dosing, model-informed precision dosing, and the use of population PK models for simulation and exposure-response analysis.

References

1. Savic RM, Karlsson MO. Importance of Shrinkage in Empirical Bayes Estimates for Diagnostics: Problems and Solutions. AAPS Journal. 2009;11(3):558–569. doi:10.1208/s12248-009-9133-0. Open article.
2. Sheiner LB, Ludden TM. Population pharmacokinetics/dynamics. Annual Review of Pharmacology and Toxicology. 1992;32:185–209.
3. Mould DR, Upton RN. Basic Concepts in Population Modeling, Simulation, and Model-Based Drug Development—Part 2: Introduction to Pharmacokinetic Modeling Methods. CPT: Pharmacometrics & Systems Pharmacology. 2013;2:e38. Open article.
4. Xu S, Yuan M, Yang H, et al. Further Evaluation of Covariate Analysis using Empirical Bayes Estimates in Population Pharmacokinetics: the Perception of Shrinkage and Likelihood Ratio Test. AAPS Journal. 2017;19(1):264–273. doi:10.1208/s12248-016-0001-4. PubMed.
5. Yano Y, Beal SL, Sheiner LB. Evaluating pharmacokinetic/pharmacodynamic models using posterior predictive checks.
6. NONMEM Users Guide, Part VII: Conditional Estimation Methods. ICON plc / NONMEM. The documentation describes the post hoc estimate as the mode of the empirical Bayes posterior density conditional on the fitted population parameters. NONMEM Users Guide.
7. Lavielle M, Mentré F. Estimation of population pharmacokinetic parameters of saquinavir in HIV patients using the SAEM algorithm.
8. Panhard X, Mentré F. Evaluation of shrinkage in nonlinear mixed-effects population models.