1. What Is Residual Error?
In a population pharmacokinetic model, the structural model describes the expected concentration-time profile, while the residual error model describes discrepancies between observed concentrations and the concentrations predicted by the model.
Residual error is often denoted by \(\varepsilon\). It represents sources of variability that remain after accounting for the structural model, between-subject variability, covariates, and other modeled components.
A useful conceptual decomposition is:
The precise mathematical form depends on the chosen residual error model.
2. What Does \(\varepsilon\) Represent?
In population PK notation, \(\varepsilon\) commonly represents the random component associated with the observation model. For example, an additive error model can be written as:
where \(Y_{ij}\) is the observed concentration for individual \(i\) at observation \(j\), \(F_{ij}\) is the model-predicted concentration, and \(\varepsilon_{ij}\) is the residual error.
A proportional error model can instead be expressed as:
A combined additive and proportional model may take the form:
These formulations make an important distinction: \(\varepsilon\) belongs to the observation model, whereas interindividual random effects such as \(\eta\) belong to the model for individual-specific parameters.
3. Epsilon Is Not Eta
One of the most important distinctions in population PK is the difference between interindividual variability and residual unexplained variability.
| Component | Common notation | What it represents |
|---|---|---|
| Interindividual variability | \(\eta\) | Differences between individuals in PK parameters |
| Residual unexplained variability | \(\varepsilon\) | Differences between observations and model predictions after modeled effects are accounted for |
| Typical population parameter | \(\theta\) | Population-level structural parameter |
For example, clearance might be modeled as:
while the resulting concentration observation could then be modeled using a residual error term:
Thus, \(\eta\) describes how individual PK parameters differ, whereas \(\varepsilon\) describes how individual observations differ from their model-predicted values.
4. What Are Individual Epsilon Estimates?
For a fitted population PK model, software can estimate individual-level residual errors using the information available for each individual's observations. These estimates are often called individual epsilon estimates or EBEs of epsilon.
Conceptually, the model has a population distribution for residual error and uses the observed data to estimate how much an individual's observations appear to deviate from their model predictions.
For a simple additive model, if:
then an intuitive residual quantity is:
However, an estimated individual epsilon is not simply an ordinary residual calculated independently at every observation. In a population model, empirical Bayes estimation accounts for the assumed residual-error distribution and the information available in the individual's observations.
5. What Is Epsilon Shrinkage?
Epsilon shrinkage describes the tendency of individual empirical Bayes estimates of residual error to be closer to the population mean of the residual distribution than would be expected if the individual residual effects were estimated with very strong information.
Because residual errors are conventionally centered around zero, strong epsilon shrinkage causes individual epsilon estimates to cluster toward zero.
Conceptually, shrinkage pulls empirical Bayes estimates toward the population mean of the random-effect distribution, which is zero for a conventional epsilon distribution.
The phenomenon is therefore about the informativeness of the individual data for estimating the individual residual effect. When the data provide limited information, the population distribution contributes more strongly to the individual estimate.
6. Why Does Epsilon Shrinkage Occur?
Epsilon shrinkage can occur when individual observations contain limited information about the individual's residual-error effects.
Several characteristics can contribute:
- Sparse sampling: few observations provide limited information about within-subject residual behavior.
- Large residual variability: a broad residual distribution can make individual deviations difficult to estimate precisely.
- Strong information from the population distribution: when the individual's data are weakly informative, empirical Bayes estimates are pulled toward the population mean.
- Repeated observations with correlated information: the amount and structure of available data influence how individual residual effects can be estimated.
- Model structure: the residual-error model itself affects the resulting empirical Bayes estimates.
Importantly, shrinkage is not automatically evidence that the model is wrong. It can arise naturally from the estimation process when individual-level information is limited.
7. How Is Epsilon Shrinkage Quantified?
A commonly used conceptual measure of shrinkage compares the observed variability of individual empirical Bayes estimates with the estimated population variability.
For a residual effect with population variance \(\omega_\varepsilon^2\), a commonly used approximate expression is:
When the variance of the individual estimates is close to the estimated population variance, shrinkage is small. When the variance of the individual estimates is substantially smaller, shrinkage is larger.
| Approximate shrinkage | Interpretation |
|---|---|
| Near 0% | Individual epsilon estimates retain variability similar to the modeled population distribution. |
| Moderate | Individual epsilon estimates show some contraction toward zero. |
| High | Individual epsilon estimates are strongly concentrated near zero and contain limited individual-level information. |
The exact numerical definition and implementation can vary across software and diagnostic workflows, so shrinkage values should always be interpreted together with the estimation method and diagnostic context.
8. Why Does Epsilon Shrinkage Matter for Residual Diagnostics?
Residual diagnostics are commonly used to assess whether systematic patterns remain between observations and model predictions. These diagnostics can include residuals plotted against predictions, time, covariates, or other quantities.
When epsilon shrinkage is substantial, empirical Bayes estimates of residual error may have artificially reduced variability. Consequently, plots based directly on individual epsilon estimates can become less informative about the true residual distribution.
This distinction is particularly important because residual variability is a population-level observation-model parameter, whereas an individual epsilon EBE is an estimated subject-specific quantity.
9. Why Are IWRES and CWRES Often Used?
Because ordinary individual residual EBEs can be affected by shrinkage, population PK diagnostics often use residuals designed to have more appropriate statistical properties under the fitted model.
Two commonly encountered quantities are:
- IWRES: individual weighted residuals.
- CWRES: conditional weighted residuals.
These approaches attempt to account for aspects of the model and variance structure when evaluating residual behavior.
CWRES, in particular, are constructed using conditional information from the model and are commonly used to examine whether residuals are approximately centered around zero and whether systematic patterns remain.
| Diagnostic quantity | General purpose | Potential shrinkage concern |
|---|---|---|
| Individual epsilon EBE | Estimate individual residual effects | Can show strong contraction toward zero when epsilon shrinkage is high |
| IWRES | Scale individual residuals using model-based variance information | Interpretation depends on model and implementation |
| CWRES | Assess conditional residual behavior | Designed to provide more suitable diagnostic behavior under the fitted model |
10. Epsilon Shrinkage Versus Eta Shrinkage
Epsilon shrinkage and eta shrinkage arise from related empirical Bayes principles, but they concern different random effects.
| Eta shrinkage | Epsilon shrinkage | |
|---|---|---|
| Random effect | \(\eta\) | \(\varepsilon\) |
| Represents | Between-subject variability in model parameters | Residual unexplained variability |
| Individual estimate | Individual PK parameter deviation | Individual residual deviation |
| Population mean | Usually zero | Usually zero |
| Common consequence of high shrinkage | Individual PK EBEs cluster toward zero | Individual residual EBEs cluster toward zero |
High eta shrinkage can complicate the interpretation of individual PK parameters and covariate diagnostics. High epsilon shrinkage primarily affects interpretation of individual residual diagnostics.
11. Epsilon Shrinkage in Sparse Sampling
Population PK studies frequently use sparse sampling designs. A participant may contribute only a small number of concentration observations, sometimes at different sampling times from other participants.
With few observations per participant, there may be limited information for estimating individual residual effects. The resulting empirical Bayes epsilon estimates can therefore exhibit substantial shrinkage toward zero.
This is one reason that population PK diagnostics should be interpreted in light of the sampling design.
| Sampling situation | Potential implication |
|---|---|
| Many observations per individual | More information may be available for estimating individual residual behavior. |
| Few observations per individual | Individual epsilon estimates may be strongly influenced by the population residual distribution. |
| Highly informative sampling times | May provide better information about structural and residual components. |
| Limited or poorly distributed sampling | Can make model components harder to distinguish. |
12. Worked Example: Interpreting Epsilon Shrinkage
Suppose a population PK model estimates the standard deviation of an additive residual-error distribution as:
The corresponding population variance is therefore:
Now suppose the variance of the individual empirical Bayes epsilon estimates is approximately:
Step 1: Calculate approximate shrinkage
Thus the approximate epsilon shrinkage is:
Step 2: Interpret the result
The individual epsilon estimates have substantially less variability than the modeled population residual distribution. This indicates that the empirical Bayes estimates have been strongly pulled toward zero.
Step 3: What does it mean for diagnostics?
The result does not mean that the true residual variability is only one-quarter of the modeled residual variance. Rather, it indicates that the individual epsilon EBEs contain limited information about the underlying residual effects.
Consequently, diagnostics based directly on the individual epsilon EBEs should be interpreted cautiously, and other model-based residual diagnostics may be more informative for assessing systematic residual patterns.
13. What Epsilon Shrinkage Does Not Mean
Several interpretations are easy to make incorrectly.
- It does not mean that the residual error is zero. Individual EBEs can be near zero even when the population residual variance is substantial.
- It does not automatically mean the model is wrong. Shrinkage can result from limited individual information.
- It does not mean observed concentrations have little variability. The observed data can still show substantial deviations from predictions.
- It does not mean the residual-error parameter should simply be reduced. The population residual variance should be estimated and evaluated using the full model and data.
- It does not automatically invalidate the population model. The consequences depend on the purpose of the analysis and the diagnostics being used.
14. What Should You Do When Epsilon Shrinkage Is High?
High epsilon shrinkage should prompt an evaluation of the model and data rather than an automatic model modification.
A practical assessment can include:
- Check the sampling design. Determine how many observations are available per individual and whether the sampling schedule is informative for the processes being modeled.
- Review the residual-error model. Consider whether additive, proportional, or combined error adequately represents the observation process.
- Inspect observed-versus-predicted concentrations. Look for systematic departures that could indicate structural model misspecification.
- Evaluate residual diagnostics. Examine suitable residual diagnostics rather than relying exclusively on individual epsilon EBEs.
- Assess the assay and data quality. Consider assay precision, below-quantification-limit observations, dose timing, sample timing, and other data characteristics.
- Consider the scientific purpose. The implications of epsilon shrinkage differ depending on whether the model is being used for description, simulation, individual prediction, covariate evaluation, or another purpose.
The appropriate response therefore depends on the underlying cause and the intended use of the model.
15. Using Simulation to Understand Epsilon Shrinkage
Simulation can be useful for understanding how sampling design and residual variability influence epsilon shrinkage.
A basic simulation workflow is:
- Specify a structural PK model.
- Specify a population residual-error distribution.
- Simulate concentration observations at the planned sampling times.
- Fit the same model to the simulated data.
- Calculate individual epsilon estimates.
- Compare their empirical distribution with the known simulated residual distribution.
- Repeat under alternative sampling designs or residual-error assumptions.
Because the true simulated residual effects are known, simulation can separate limitations caused by the data structure from problems caused by the fitted model.
16. A Practical Workflow for Evaluating Epsilon Shrinkage
- Identify the residual-error model. Determine whether the model uses additive, proportional, combined, or another observation-error structure.
- Estimate the population residual variability. Review the estimated residual-error parameters and their uncertainty.
- Obtain individual epsilon estimates. Examine their distribution and compare it with the assumed population distribution.
- Quantify shrinkage. Use an appropriate shrinkage metric and document how it was calculated.
- Assess the sampling design. Determine whether limited observations could explain strong shrinkage.
- Review residual diagnostics. Use appropriate residual-based diagnostics to investigate systematic model deficiencies.
- Check structural model adequacy. A residual pattern can sometimes reflect an inadequate structural model rather than an inappropriate residual-error model.
- Interpret shrinkage in context. Consider whether the shrinkage affects the specific scientific conclusions or diagnostics for which the model is being used.
17. Key Takeaways
- Residual error describes unexplained variability between observed concentrations and model-predicted concentrations.
- \(\varepsilon\) is the random component of the observation model, whereas \(\eta\) represents between-subject random effects on model parameters.
- Epsilon shrinkage describes the contraction of individual empirical Bayes epsilon estimates toward their population mean, conventionally zero.
- High epsilon shrinkage commonly occurs when individual observations provide limited information about individual residual effects.
- Sparse sampling can contribute to substantial epsilon shrinkage.
- A high epsilon-shrinkage value does not mean that the population residual variability is small or that observed concentrations have little variability.
- Individual epsilon EBEs should be interpreted cautiously when shrinkage is substantial.
- Residual diagnostics such as IWRES and CWRES can provide useful complementary information when evaluating residual behavior.
- Epsilon shrinkage and eta shrinkage concern different random effects and have different implications for model interpretation.
- High epsilon shrinkage should be investigated in the context of sampling design, residual-error specification, structural model adequacy, and the intended use of the model.
Where to Go Next
A natural progression is to study Interindividual Variability in Pharmacokinetic Models and Eta Shrinkage and Its Interpretation, followed by residual diagnostics such as IWRES and CWRES.
Understanding the distinction between \(\eta\), \(\varepsilon\), their empirical Bayes estimates, and their respective shrinkage measures provides an important foundation for interpreting population PK model diagnostics.