Tutorials › Pharmacometrics › Shrinkage in Population PK Models
Pharmacokinetics · Population PK

Shrinkage in Population PK Models

Understand why individual population-PK estimates can move toward the typical population value, how eta and epsilon shrinkage arise, how shrinkage affects diagnostics and covariate analysis, and why shrinkage is an important property of the model rather than simply a modeling defect.

Intermediate Population PK NONMEM Concepts Model Diagnostics
01 · The big picture

1. What Is Shrinkage?

Shrinkage in population pharmacokinetic (PK) modeling describes the tendency of individual parameter estimates to move toward the typical population value when the individual data contain limited information about that parameter.

This is most commonly discussed for empirical Bayes estimates (EBEs), also called individual estimates or post hoc estimates. In a population model, an individual's parameter is informed by two sources:

  • the individual's observed concentration data, and
  • the population distribution of the parameter.

When the individual's data are highly informative, the individual estimate can be substantially different from the population typical value. When the individual's data provide little information, the population distribution contributes more strongly and the estimate tends to remain closer to the population value.

Core idea: shrinkage is a consequence of limited individual information. An individual's estimate is pulled toward the population distribution because the model combines individual observations with prior population information.
02 · Population PK

2. Why Does Shrinkage Occur in Population PK?

A basic population PK model separates typical population parameters from between-subject variability.

For example, suppose clearance for individual \(i\) is modeled using an exponential random-effects model:

\[ CL_i = CL_{\mathrm{pop}}e^{\eta_{CL,i}} \]

Here:

  • \(CL_{\mathrm{pop}}\) is the typical population clearance.
  • \(\eta_{CL,i}\) describes how individual \(i\)'s clearance differs from the typical value.
  • The distribution of \(\eta_{CL}\) represents between-subject variability in clearance.

Typically, the random effect is assumed to have mean zero:

\[ \eta_{CL,i}\sim N(0,\omega^2_{CL}) \]

The observed concentration data provide information about \(\eta_{CL,i}\). If the data are sparse, noisy, or poorly positioned in time to estimate clearance, the model has limited evidence that the individual's \(\eta\) differs substantially from zero.

Consequently, the individual's EBE can be pulled toward zero.

Population typical value η = 0 Individual estimate Individual estimate

Conceptually, less informative individual data can cause EBEs to move toward the population typical value. The amount of shrinkage depends on the information available in the individual data relative to the population variability and residual error.

03 · Eta shrinkage

3. What Is Eta Shrinkage?

Eta shrinkage describes the degree to which individual estimates of between-subject random effects are compressed toward their population mean, conventionally zero.

If an individual's true deviation from the population mean is represented by \(\eta_i\), the model produces an estimate such as \(\hat{\eta}_i\). With substantial information, \(\hat{\eta}_i\) can reflect the individual's underlying deviation. With weak information, \(\hat{\eta}_i\) tends to be closer to zero.

A commonly used descriptive measure of eta shrinkage is:

\[ \text{Eta shrinkage} = 1- \frac{\operatorname{SD}(\hat{\eta})}{\omega} \]

where \(\omega\) is the estimated population standard deviation of the corresponding eta distribution.

Equivalent expressions are sometimes written using the variance rather than the standard deviation, so the exact convention should always be checked when comparing software outputs or publications.

Situation Individual information Typical effect on EBE
Rich sampling with informative PK phases High EBEs can vary substantially across individuals
Moderate sampling Intermediate Some movement toward the population mean
Very sparse sampling Low EBEs tend to cluster more closely around zero

High eta shrinkage therefore means that the observed spread of individual EBEs can be much smaller than the estimated population variability.

04 · Residual variability

4. What Is Epsilon Shrinkage?

Epsilon shrinkage concerns the individual residuals rather than the between-subject random effects.

A simple observation model might be written as:

\[ C_{ij}=f(\theta,\eta_i,t_{ij})+\epsilon_{ij} \]

where \(C_{ij}\) is an observed concentration, \(f(\cdot)\) is the structural PK model, \(\eta_i\) represents between-subject variability, and \(\epsilon_{ij}\) represents residual unexplained variability.

Individual residuals are often estimated conditional on the model and can also be pulled toward zero when the data contain limited information.

A commonly used descriptive expression is:

\[ \text{Epsilon shrinkage} = 1- \frac{\operatorname{SD}(\hat{\epsilon})}{\sigma} \]

where \(\sigma\) represents the estimated residual-error standard deviation under the relevant error model.

Eta versus epsilon: eta shrinkage concerns individual between-subject random effects, whereas epsilon shrinkage concerns individual residual errors. They describe different parts of the hierarchical population-PK model.
05 · Bayesian interpretation

5. Shrinkage From a Bayesian Perspective

One of the clearest ways to understand shrinkage is through the Bayesian interpretation of an individual parameter estimate.

Suppose the individual parameter is represented by \(\eta_i\), with a population distribution:

\[ \eta_i\sim N(0,\omega^2) \]

The individual's concentration observations provide a likelihood for \(\eta_i\). Combining that likelihood with the population distribution produces a conditional distribution for the individual's random effect:

\[ p(\eta_i\mid y_i) \propto p(y_i\mid\eta_i)p(\eta_i) \]

The EBE can be viewed as a mode or conditional estimate of this individual-specific distribution, depending on the estimation framework.

When the individual data are highly informative, the likelihood contributes substantial information and the conditional estimate can move away from zero.

When the individual data are weakly informative, the population distribution exerts greater influence and the estimate remains closer to zero.

Important interpretation: shrinkage does not mean that the model has arbitrarily forced every individual to have the same PK. It reflects the relative strength of the individual's information and the population-level information in the estimation problem.
06 · What controls shrinkage?

6. What Causes More or Less Shrinkage?

Shrinkage is influenced by the amount of information available for estimating an individual parameter and by the relative magnitude of the model's variability components.

6.1 Sparse sampling

If only a small number of concentrations are available for an individual, there may be limited information about parameters such as clearance, volume, or absorption rate.

6.2 Sampling times

The number of samples is not the only issue. Their timing matters. For example, observations concentrated around a single PK phase may provide little information about another parameter.

6.3 Residual unexplained variability

Large residual error makes it harder to distinguish an individual's underlying PK behavior from measurement and model noise. This can increase shrinkage.

6.4 Between-subject variability

The estimated magnitude of the population variability also matters. The balance between \(\omega^2\), residual error, and individual information affects how strongly estimates are pulled toward the population distribution.

6.5 Parameter identifiability

If two parameters produce similar effects on the observed concentrations over the available sampling schedule, individual estimates may be weakly identified even when the structural model fits the population data adequately.

6.6 Dose and concentration range

The dose and resulting concentrations can influence how much information the observations provide about particular parameters, especially in nonlinear models.

Factor Why it can affect shrinkage
Few observations Less individual information is available
Poorly timed observations Important PK processes may not be adequately observed
Large residual error Individual PK differences become harder to distinguish from noise
Weak parameter identifiability Individual parameters cannot be estimated precisely from the available data
Extensive parameter correlation Information about one parameter may be difficult to separate from another
07 · A visual example

7. How Shrinkage Changes the Distribution of EBEs

Consider a population model in which the true between-subject random effects for clearance have a standard deviation of \(\omega=0.5\).

If the individual data are highly informative, the estimated EBEs might have a distribution that resembles the population random-effect distribution.

With substantial shrinkage, the EBE distribution becomes narrower.

Underlying η distribution EBE distribution 0 Estimated random effect

Conceptual illustration: shrinkage reduces the apparent spread of individual EBEs relative to the underlying population random-effect distribution.

This narrowing is important because an EBE distribution should not automatically be interpreted as a direct empirical estimate of the true between-subject variability.

08 · Diagnostics

8. Why Does Shrinkage Matter for PK Diagnostics?

Individual EBEs are often used to create diagnostic plots. For example, a modeler may plot an EBE for clearance against body weight, age, renal function, or another covariate.

This can be useful when shrinkage is modest. However, substantial shrinkage changes how much information the EBE contains about the individual's underlying parameter.

When EBEs are strongly compressed toward zero, relationships between EBEs and covariates can become attenuated or difficult to detect.

For example, suppose the true relationship is:

\[ \log(CL_i) = \log(CL_{\mathrm{pop}}) + \beta\log\left(\frac{WT_i}{WT_{\mathrm{ref}}}\right) + \eta_{CL,i} \]

If the estimated \(\hat{\eta}_{CL,i}\) values are strongly shrunk toward zero, an EBE-versus-weight plot may show a weaker apparent relationship than would be obtained from more informative individual data.

Diagnostic warning: a weak relationship between an EBE and a covariate does not necessarily prove that the covariate is unimportant. High shrinkage can reduce the information contained in the individual estimates and therefore weaken EBE-based exploratory plots.
09 · Covariate analysis

9. Shrinkage and Covariate Relationships

Covariate analysis is one of the most important settings in which shrinkage needs to be considered.

Suppose a population PK model contains a covariate relationship for clearance:

\[ CL_i = CL_{\mathrm{pop}} \left(\frac{WT_i}{70}\right)^{\theta_{WT}} e^{\eta_{CL,i}} \]

A modeler might initially examine the relationship between estimated individual clearance and body weight. But if the individual estimates have substantial shrinkage, the observed spread of the estimates may be substantially smaller than the true underlying spread.

This creates an important distinction:

  • EBE-based exploration uses individual estimates and can be affected by shrinkage.
  • Model-based covariate evaluation estimates the covariate relationship directly within the population model and does not rely solely on the visual spread of EBEs.

Consequently, high shrinkage should not be interpreted as evidence that covariates have no relationship with PK parameters.

10 · PK/PD implications

10. Why Shrinkage Matters for PK/PD Modeling

Shrinkage becomes particularly important when individual PK estimates are subsequently used as inputs to a pharmacodynamic analysis.

Suppose a modeler first estimates individual clearance values and then uses those EBEs as predictors in a separate exposure-response analysis.

If the clearance EBEs are highly shrunk, they may contain limited information about true individual differences in clearance. This can affect downstream analyses.

For example, an analysis might attempt to relate an estimated individual exposure metric to an observed response:

\[ E_i = E_0+ \frac{E_{\max}AUC_i}{EC_{50}+AUC_i} \]

If \(AUC_i\) is calculated using highly shrunk individual parameter estimates, the resulting exposure measure may not preserve all of the relevant between-subject variation.

This is one reason integrated population PK/PD modeling can be preferable to a simple two-stage procedure when the scientific objective requires accurate propagation of parameter uncertainty and individual variability.

Two-stage caution: using highly shrunk individual PK estimates as if they were directly observed individual parameters can discard information and underrepresent uncertainty in a downstream analysis.
11 · What shrinkage can distort

11. Common Problems Caused by High Shrinkage

High shrinkage is not automatically a reason to reject a model. It becomes important when the modeler's intended use of individual estimates requires information that the data do not adequately provide.

11.1 Compressed EBE distributions

EBEs can appear much less variable than the estimated population distribution.

11.2 Weak EBE-covariate relationships

Relationships can be visually attenuated when individual estimates contain little information about individual deviations.

11.3 Diagnostic interpretation

Plots based on individual random effects may be less informative when shrinkage is substantial.

11.4 Downstream exposure-response analyses

Highly shrunk individual parameters may provide poor substitutes for directly estimated individual exposure or PK quantities in a two-stage analysis.

11.5 Misleading residual diagnostics

When individual predictions depend heavily on population predictions because of shrinkage, diagnostic behavior involving conditional predictions and residuals can require careful interpretation.

12 · Residual diagnostics

12. Shrinkage and Residual-Based Diagnostics

Population PK diagnostics often include residual-based measures such as individual weighted residuals (IWRES), conditional weighted residuals (CWRES), and related quantities.

The purpose of these diagnostics is to assess whether observed concentrations are consistent with model predictions after accounting for the specified variability structure.

High eta shrinkage can affect diagnostics that depend strongly on individual parameter estimates because the individual predictions may become increasingly similar to population predictions.

For this reason, diagnostic interpretation should consider:

  • the amount of eta shrinkage;
  • the amount of epsilon shrinkage;
  • the sampling design;
  • the residual error model;
  • the number and timing of observations per individual; and
  • whether the diagnostic requires reliable individual parameter estimates.

Shrinkage is therefore best considered alongside visual predictive checks, prediction-corrected visual predictive checks, residual diagnostics, parameter estimates, and other model-evaluation tools rather than in isolation.

13 · Worked example

13. Worked Example: Calculating Eta Shrinkage

Suppose a population PK model estimates the between-subject variability in clearance as:

\[ \omega_{CL}=0.50 \]

After obtaining individual EBEs, suppose their empirical standard deviation is:

\[ SD(\hat{\eta}_{CL})=0.30 \]

Step 1: Apply the shrinkage formula

\[ \text{Eta shrinkage} = 1- \frac{SD(\hat{\eta}_{CL})}{\omega_{CL}} \]

Step 2: Substitute the values

\[ \text{Eta shrinkage} = 1-\frac{0.30}{0.50} \]

Step 3: Calculate

\[ \text{Eta shrinkage} = 1-0.60 = 0.40 \]

Thus, the estimated eta shrinkage is:

\[ \boxed{40\%} \]

Conceptually, the observed spread of the EBEs is about 60% of the estimated population random-effect standard deviation under this descriptive calculation.

Interpretation: 40% shrinkage indicates meaningful compression of the individual clearance EBEs toward zero. It does not mean that 40% of the true between-subject variability has disappeared biologically.
14 · Individual estimates

14. How Shrinkage Affects an Individual Parameter Estimate

Suppose the typical population clearance is:

\[ CL_{\mathrm{pop}}=5\text{ L/h} \]

Assume an individual's underlying deviation would correspond to an individual clearance of approximately 8 L/h. In the exponential model:

\[ CL_i=CL_{\mathrm{pop}}e^{\eta_i} \]

the corresponding random effect is:

\[ \eta_i=\log\left(\frac{8}{5}\right)\approx0.470 \]

Now suppose the individual's data are sparse and the EBE is estimated as only:

\[ \hat{\eta}_i=0.20 \]

The EBE-based individual clearance becomes:

\[ \widehat{CL}_i = 5e^{0.20} \approx6.11\text{ L/h} \]

The estimated individual clearance is therefore closer to the population value of 5 L/h than the hypothetical underlying value of 8 L/h.

This illustrates the central idea of shrinkage: the EBE reflects the combination of individual information and population information, rather than being a direct measurement of the individual's true parameter.

15 · Interpretation

15. Shrinkage Is Parameter-Specific

It is important not to describe a population PK model simply as having "high shrinkage" without specifying which random effect is shrinking.

For example, a model could have:

Random effect Example shrinkage Possible interpretation
\(\eta_{CL}\) 15% Individual clearance estimates retain relatively substantial information
\(\eta_V\) 35% Individual volume estimates are more compressed
\(\eta_{ka}\) 70% Individual absorption-rate estimates are strongly informed by the population distribution

Different parameters can have very different levels of individual identifiability.

For example, a sparse sampling design may contain enough information about overall exposure to estimate clearance reasonably well while providing little information about an absorption rate constant.

Practical point: always identify the parameter associated with the shrinkage estimate. High shrinkage for one parameter does not imply equally high shrinkage for every parameter in the model.
16 · Study design

16. Shrinkage and Sampling Design

Shrinkage is strongly connected to the information content of the sampling design.

Consider a one-compartment oral PK model with first-order absorption:

\[ C(t) = \frac{FDk_a}{V(k_a-k)} \left(e^{-kt}-e^{-k_at}\right) \]

To estimate \(k_a\), the study needs observations that provide information about the absorption portion of the concentration-time profile.

If all samples are collected long after the absorption phase, the data may provide little information about \(k_a\). The model can still estimate a population-level absorption rate, but individual \(k_a\) EBEs may exhibit substantial shrinkage.

Similarly, if there are few samples during the terminal elimination phase, individual clearance or terminal-rate estimates may be weakly informed.

This leads to an important design principle:

Sampling principle: reducing shrinkage is not primarily a matter of changing the estimation software. It often requires collecting observations at times that contain information about the parameters of interest.
17 · What can be done?

17. How Can Shrinkage Be Reduced?

If high shrinkage limits the intended use of individual estimates, several approaches may improve the information available to the model.

17.1 Improve sampling

Collect samples at times that are informative for the parameters of interest. This can be guided by optimal design or simulation-based design evaluation.

17.2 Increase the number of informative observations

Additional samples can provide more information, although simply increasing the number of samples is not sufficient if the new observations are redundant.

17.3 Improve the residual error model

An inappropriate or overly large residual-error model can reduce the apparent information available from individual observations.

17.4 Reconsider the structural model

An unnecessarily complex model can create parameter-identification problems. Conversely, an oversimplified model can leave systematic patterns in the residuals that obscure the intended interpretation.

17.5 Incorporate important covariates

Explaining systematic between-subject differences through measured covariates can reduce unexplained heterogeneity, although adding covariates should be based on scientific and statistical considerations rather than used solely to manipulate shrinkage.

17.6 Use an integrated model when appropriate

For exposure-response analyses, directly modeling PK and PD together can avoid treating highly shrunk individual PK estimates as though they were observed quantities.

18 · High shrinkage

18. Is High Shrinkage Always Bad?

No. High shrinkage is not automatically evidence that a population PK model is invalid.

A model can adequately describe population concentration-time data while producing highly shrunk individual estimates for one or more parameters.

This may be entirely expected when the study design contains limited information at the individual level.

The important question is therefore:

What are the individual estimates being used for?

If the objective is population-level prediction, high shrinkage for an individual parameter may be less problematic than if the objective is to use that individual's EBE as a reliable measure of their actual PK parameter.

Use of model Potential relevance of shrinkage
Population-level concentration prediction May be relatively limited if the population model adequately describes the data
EBE versus covariate plots Can be important because shrinkage compresses individual variability
Individual dosing based on EBEs Important when individual parameter estimates are expected to be reliable
Two-stage PK/PD analysis Potentially important because shrunk estimates may lose individual information
Simulation of population behavior Depends primarily on the population model and its variability structure, not simply on the empirical spread of EBEs
19 · Common mistakes

19. Common Misinterpretations of Shrinkage

Mistake 1: "High shrinkage means the model is wrong."

Not necessarily. High shrinkage can be a consequence of sparse or poorly informative data.

Mistake 2: "Shrinkage means individuals are actually becoming more similar."

No. Shrinkage concerns the estimated individual parameters. It does not imply that biological between-subject variability has disappeared.

Mistake 3: "A narrow EBE distribution means low population variability."

Not necessarily. High shrinkage can make the EBE distribution substantially narrower than the estimated underlying random-effect distribution.

Mistake 4: "No EBE-covariate relationship means the covariate does not matter."

High shrinkage can attenuate EBE-covariate relationships.

Mistake 5: "Every parameter has the same shrinkage."

Different parameters can have very different levels of individual information and therefore different shrinkage.

Mistake 6: "Shrinkage is a property of the drug."

Shrinkage is a property of the model-data combination. Sampling design, residual variability, model structure, parameter variability, and the population all contribute.

20 · Reporting

20. How Should Shrinkage Be Reported?

When shrinkage is important to the interpretation of a population PK analysis, it is useful to report it explicitly and identify the associated parameter.

For example, a report might state that eta shrinkage was approximately 20% for clearance, 35% for volume, and 65% for absorption rate.

Good reporting should make clear:

  • which random effect was evaluated;
  • the definition or formula used for shrinkage;
  • whether the reported quantity is eta or epsilon shrinkage;
  • which estimation method and software generated the individual estimates; and
  • how shrinkage affected the interpretation of individual diagnostics or downstream analyses.

Because different shrinkage definitions and estimation procedures can be used, numerical values should be interpreted in the context of the calculation method.

21 · Worked interpretation

21. Worked Example: Interpreting Different Shrinkage Values

Suppose a population PK model produces the following results:

Parameter Population variability Eta shrinkage
Clearance \(\omega_{CL}=0.40\) 12%
Volume \(\omega_V=0.35\) 28%
Absorption rate \(\omega_{ka}=0.70\) 72%

The results suggest that individual clearance estimates contain substantially more information than individual absorption-rate estimates in this hypothetical dataset.

The high shrinkage for \(k_a\) might be expected if the study has few observations during the absorption phase.

It would therefore be inappropriate to conclude that absorption-rate variability is necessarily small simply because the \(k_a\) EBEs show a narrow distribution.

Key distinction: the estimated \(\omega_{ka}\) describes the population variability in the model, whereas the empirical spread of the \(\hat{\eta}_{ka}\) values reflects that variability after individual-data information and shrinkage have been taken into account.
22 · A useful mental model

22. Shrinkage as a Balance of Information

A useful conceptual framework is to think of an individual estimate as a balance between two sources of information:

Individual data concentrations sampling times Population model typical parameters between-subject variability combined information Individual EBE

An EBE combines information from the individual's observations with information supplied by the population model. Shrinkage becomes more pronounced when individual information is weak relative to the population information.

This perspective is useful because it explains why shrinkage can change when the same drug is studied using a different sampling design, population, residual-error model, or dosing regimen.

23 · Practical workflow

23. A Practical Workflow for Evaluating Shrinkage

  1. Identify the individual estimates. Determine which EBEs or post hoc estimates are being evaluated.
  2. Calculate or obtain shrinkage. Identify whether eta, epsilon, or another shrinkage metric is being reported.
  3. Associate shrinkage with the parameter. Examine clearance, volume, absorption parameters, and other random effects separately.
  4. Inspect the sampling design. Determine whether the data contain adequate information for the parameter.
  5. Review residual variability. Large unexplained residual variability can contribute to weak individual information.
  6. Examine EBE distributions. Compare the empirical spread of EBEs with the estimated population random-effect distribution.
  7. Interpret EBE-covariate plots cautiously. High shrinkage can weaken apparent relationships.
  8. Consider the intended use. Ask whether individual estimates are being used for diagnostics, covariate exploration, dosing, or downstream PK/PD analysis.
  9. Use integrated modeling when appropriate. Avoid treating highly shrunk EBEs as directly observed individual parameters in analyses where individual information is critical.

24. Key Takeaways

  • Shrinkage describes the tendency of individual population-PK estimates to move toward the population typical value when individual data contain limited information.
  • Eta shrinkage concerns individual estimates of between-subject random effects, which conventionally have a population mean of zero.
  • Epsilon shrinkage concerns individual residual estimates and reflects limited information about individual residual deviations.
  • Shrinkage arises naturally from the hierarchical structure of population PK models, where individual observations are combined with population-level information.
  • Sparse sampling, poorly timed observations, high residual variability, and weak parameter identifiability can increase shrinkage.
  • Different parameters in the same population PK model can have very different levels of shrinkage.
  • A narrow EBE distribution does not necessarily imply that the underlying population variability is small.
  • High shrinkage can attenuate apparent relationships between EBEs and covariates.
  • High shrinkage is not automatically evidence that a population PK model is inadequate.
  • The importance of shrinkage depends strongly on how individual estimates are being used.
  • Highly shrunk individual PK estimates should be interpreted cautiously in two-stage PK/PD or exposure-response analyses.
  • Improving sampling design and individual information is often more effective than simply changing estimation procedures when the objective is to reduce shrinkage.
  • Shrinkage should be interpreted together with model diagnostics, parameter estimates, sampling design, residual variability, and the scientific purpose of the analysis.
Next step

Where to Go Next

A natural next step is to study Empirical Bayes Estimates in Population PK in more detail, including MAP estimation, post hoc parameter estimation, conditional distributions, and the distinction between population parameters and individual parameters.

From there, the concepts connect naturally to Covariate Modeling in Population PK, Visual Predictive Checks, Model Diagnostics, Optimal Sampling Design, and PK/PD Modeling.

References

References

  1. Sheiner LB, Ludden TM. Population pharmacokinetics/dynamics. Annual Review of Pharmacology and Toxicology. 1992;32:185–209.
  2. Sheiner LB, Ludden TM, Beal SL. Hierarchical statistical models for the analysis of population pharmacokinetic data. Journal of Pharmacokinetics and Biopharmaceutics. 1979;7:305–318.
  3. Ette EI, Williams PJ. Population pharmacokinetics I: background, concepts, and models. Annals of Pharmacotherapy. 2004;38:1702–1706.
  4. Ette EI, Williams PJ. Population pharmacokinetics II: estimation methods. Annals of Pharmacotherapy. 2004;38:1907–1915.
  5. Food and Drug Administration. Population Pharmacokinetics: Guidance for Industry. U.S. Department of Health and Human Services.
  6. European Medicines Agency. Guideline on Reporting the Results of Population Pharmacokinetic Analyses.

Definitions and shrinkage calculations can vary across software and publications. When reporting a numerical shrinkage estimate, specify the calculation used and the parameter to which it applies.

← Back to Pharmacokinetics Tutorials