1. What Is Residual Unexplained Variability?
In a population pharmacokinetic model, the structural model describes the expected concentration-time behavior of a drug. Between-subject variability can then describe how individual PK parameters differ from the population-typical values. Even after those sources of variation have been modeled, observed concentrations will generally not fall exactly on the model-predicted curve.
The remaining difference between an observation and the model prediction is represented through an observation error model, often called the residual unexplained variability (RUV) model.
The structural PK model generates the expected concentration, while the residual error model describes the remaining observation-level discrepancy.
2. Where Does Residual Variability Fit in a Population PK Model?
A population PK model is commonly understood as having several layers. The structural model describes the typical concentration-time behavior. Between-subject variability (BSV) describes how individual parameters vary around population-typical values. The residual unexplained variability describes discrepancies between observations and the corresponding individual model predictions.
These components represent different sources of variation and should not be interpreted as interchangeable.
| Model component | What it describes | Typical level |
|---|---|---|
| Structural model | Expected concentration-time behavior given PK parameters and dosing | Population or individual prediction |
| Between-subject variability | Differences in PK parameters between individuals | Subject level |
| Residual unexplained variability | Difference between observed concentrations and the model's corresponding prediction | Observation level |
This distinction is particularly important when interpreting model diagnostics. A systematic pattern in residuals may indicate that the structural model is inadequate, whereas random scatter around the prediction may be consistent with an appropriate residual error model.
3. What Does Residual Variability Represent?
The word unexplained should not be interpreted to mean that every residual is caused by random laboratory noise. Residual variability is a practical modeling category for variation that has not been captured elsewhere in the model.
Depending on the study, residual variability can reflect several sources, including:
- Analytical assay measurement error.
- Sample handling and processing variability.
- Timing discrepancies between nominal and actual sampling times.
- Short-term fluctuations not represented by the structural model.
- Model simplification or structural misspecification.
- Unmodeled within-subject variability.
- Other observation-level sources of variation.
For this reason, a very large residual variance can sometimes indicate more than imprecise measurements. It may also signal that the structural model, covariate model, sampling design, or other assumptions do not adequately describe the data.
4. Observations, Predictions, and Residuals
Let \(y_{ij}\) denote the observed concentration for subject \(i\) at observation \(j\). Let \(f(\theta_i,t_{ij})\) denote the model-predicted concentration at the same time.
The prediction can be written as:
where \(\theta_i\) represents the individual's PK parameters.
A simple additive observation model is:
Here, \(\epsilon_{ij}\) represents the residual error. Under a conventional additive model, one common assumption is:
Thus, the observed concentration is viewed as the model prediction plus an observation-level random deviation.
5. The Additive Residual Error Model
The additive error model assumes that the absolute size of residual variability is approximately constant across the concentration range.
with:
Under this model, the standard deviation of the observation error is \(\sigma\), regardless of the predicted concentration.
When can an additive model be useful?
- When assay error is approximately constant in absolute concentration units.
- When residual scatter appears similar at low and high concentrations.
- When concentrations can reasonably be modeled on their original scale.
- When the observation process supports approximately constant absolute variability.
For example, if the residual standard deviation is approximately 0.5 mg/L, a concentration of 2 mg/L and a concentration of 20 mg/L might both have residual standard deviations near 0.5 mg/L under an additive model.
| Predicted concentration | Approximate SD under additive model | Relative variability |
|---|---|---|
| 2 mg/L | 0.5 mg/L | 25% |
| 5 mg/L | 0.5 mg/L | 10% |
| 10 mg/L | 0.5 mg/L | 5% |
| 20 mg/L | 0.5 mg/L | 2.5% |
This illustrates a limitation: although the absolute variability remains constant, the relative variability becomes much larger at low concentrations.
6. The Proportional Residual Error Model
The proportional error model assumes that the magnitude of residual variability increases with the predicted concentration.
where commonly:
Equivalently, this can be written as:
The residual standard deviation is therefore proportional to the prediction:
For example, if the proportional error parameter is 20%, a prediction of 2 mg/L corresponds to an approximate residual standard deviation of 0.4 mg/L, while a prediction of 20 mg/L corresponds to approximately 4 mg/L.
7. The Combined Additive + Proportional Model
Many PK datasets show a combination of approximately constant absolute variability and concentration-dependent variability. A common observation model is therefore the combined additive and proportional error model.
where \(\epsilon_1\) and \(\epsilon_2\) represent additive and proportional components, respectively.
Under independent zero-mean components with variances \(\sigma_1^2\) and \(\sigma_2^2\), the conditional variance is:
and the corresponding standard deviation is:
The combined model therefore allows the residual spread to be approximately constant near the low end while increasing with concentration at higher values.
| Model | Residual scale | Typical interpretation |
|---|---|---|
| Additive | \(\sigma\) | Constant absolute error |
| Proportional | \(\widehat{y}\sigma\) | Constant relative error |
| Combined | \(\sqrt{\sigma_1^2+\widehat{y}^2\sigma_2^2}\) | Both absolute and relative components |
8. The Exponential Residual Error Model
Another commonly used formulation is an exponential error model:
where \(\epsilon\) is often assumed to follow a normal distribution:
This formulation implies that the observation is positive when the model prediction is positive, because \(e^\epsilon\) is always positive.
Taking logarithms gives:
Thus, the exponential error model on the original concentration scale corresponds to an additive error model on the log-concentration scale.
9. Modeling Residual Variability on the Log Scale
Instead of modeling concentration directly, one can model the logarithm of concentration. The observation model can be written as:
with:
On the original scale this implies:
This is closely related to the exponential residual error model. The interpretation is naturally multiplicative: residual deviations are expressed as ratios between observed and predicted concentrations rather than as simple arithmetic differences.
The log-scale formulation can be especially attractive when concentration variability increases approximately in proportion to concentration and when the distribution of log concentrations is closer to symmetric than the distribution of raw concentrations.
Observed-to-predicted ratios
Under a log-scale model, the residual can be interpreted as:
A residual of zero corresponds to \(y=\widehat{y}\). Positive residuals indicate that the observation exceeds the prediction, while negative residuals indicate that the observation is below the prediction.
10. Additive, Proportional, Combined, and Exponential Models
The major residual error formulations differ primarily in how they allow the variability of observations to change with the magnitude of the prediction.
| Error model | Basic formulation | How variability changes | Useful interpretation |
|---|---|---|---|
| Additive | \(y=\widehat{y}+\epsilon\) | Constant absolute scale | Error measured in concentration units |
| Proportional | \(y=\widehat{y}(1+\epsilon)\) | Increases with prediction | Error expressed approximately as a percentage |
| Combined | \(y=\widehat{y}+\epsilon_1+\widehat{y}\epsilon_2\) | Absolute + concentration-dependent | Useful when neither pure additive nor pure proportional behavior is adequate |
| Exponential | \(y=\widehat{y}e^\epsilon\) | Multiplicative | Equivalent to additive error on the log scale |
No single formulation is universally correct. The appropriate model depends on the observation process, concentration range, assay characteristics, data distribution, and scientific purpose.
11. Worked Example: Comparing Residual Error Models
Suppose a population PK model predicts a concentration of 10 mg/L. Consider three hypothetical residual-error specifications:
- Additive error SD = 0.5 mg/L.
- Proportional error = 20%.
- Combined error with additive SD = 0.3 mg/L and proportional SD = 15%.
Step 1: Additive error
The residual spread is 0.5 mg/L regardless of the predicted concentration.
Step 2: Proportional error
At a prediction of 10 mg/L, the proportional model therefore implies a residual standard deviation of 2 mg/L.
Step 3: Combined error
Step 4: Compare the implied variability
| Error model | Implied residual SD at 10 mg/L |
|---|---|
| Additive | 0.50 mg/L |
| Proportional | 2.00 mg/L |
| Combined | 1.53 mg/L |
The example demonstrates why residual error structure matters. The same structural PK prediction can imply very different observation distributions depending on the chosen residual model.
12. How Are Residual Error Models Evaluated?
Residual error models should be evaluated using diagnostic plots and statistical behavior rather than selected solely because a particular formulation is conventional.
Common diagnostic quantities include individual weighted residuals, conditional weighted residuals, and related normalized residual measures, depending on the modeling framework.
A basic residual is:
But raw residuals are difficult to compare when the expected observation variance changes with concentration. A weighted residual accounts for the model-implied residual variance.
Conceptually:
The exact residual definition and weighting depend on the estimation method and model implementation.
Useful diagnostic plots
- Residuals versus time.
- Residuals versus population predictions.
- Residuals versus individual predictions.
- Residuals versus covariates.
- Observed versus predicted concentrations.
- Observed versus individual predicted concentrations.
- Distribution or quantile plots of normalized residuals.
13. What Do Common Residual Patterns Mean?
Residual diagnostics are most useful when interpreted as evidence about model behavior rather than as automatic pass/fail tests.
| Pattern | Possible interpretation |
|---|---|
| Increasing residual spread with prediction | Residual variance may be concentration-dependent; proportional or combined error may be worth considering. |
| Large spread only near low concentrations | An additive component may be important, or assay behavior may change near the lower quantification range. |
| Curved residual trend versus prediction | The structural model or error model may not adequately describe the relationship. |
| Residual trend versus time | Time-dependent misspecification, delayed processes, or other structural issues may be present. |
| Residual relationship with a covariate | A systematic source of variability may not be adequately represented by the current model. |
| Strong outliers | Possible data, dosing, sampling, assay, or model issues should be investigated. |
These interpretations are not deterministic. The same residual pattern can have several possible explanations, and diagnostic evidence should be considered alongside study design, data quality, biological plausibility, and other model diagnostics.
14. Why Are Low Concentrations Especially Important?
Residual error modeling can become particularly important near the lower end of the observed concentration range.
Suppose an additive residual standard deviation is 0.5 mg/L. At 20 mg/L, this represents only 2.5% of the prediction. At 1 mg/L, it represents 50%.
Therefore, an additive model that appears reasonable at high concentrations may imply very large relative variability near low concentrations.
Conversely, a pure proportional model can imply extremely small absolute error as predicted concentration approaches zero.
This is one reason combined error models can sometimes be useful: they allow an absolute error component to remain important at low concentrations while retaining proportional behavior at higher concentrations.
15. What About Concentrations Below the Limit of Quantification?
Observations reported as below the limit of quantification (BLQ) are not ordinary numeric concentration measurements. They indicate that the concentration is below a quantification threshold.
This distinction matters because simply replacing BLQ observations with zero, half the quantification limit, or another fixed value changes the information represented by the data.
Depending on the analysis objective and modeling framework, BLQ observations can be handled using approaches that explicitly recognize their censored nature.
16. Are Residual Errors Independent?
Many standard population PK implementations assume that residual errors are conditionally independent between observations after accounting for the structural model and other modeled variability.
In a simple formulation:
conditional on the relevant model components.
This assumption is convenient, but it is not automatically guaranteed by the data. Repeated observations from the same individual can sometimes display temporal or serial correlation that is not captured by a conventional independent residual error model.
If meaningful correlation remains after accounting for the structural and variability models, more specialized observation models may be considered.
17. Between-Subject Variability Is Not Residual Variability
One of the most important distinctions in population PK is the difference between between-subject variability and residual unexplained variability.
Suppose clearance for subject \(i\) is modeled using an exponential random effect:
The random effect \(\eta_{CL,i}\) describes how that subject's clearance differs from the population-typical clearance.
By contrast, residual error describes why an observed concentration may differ from the concentration predicted using that individual's parameters:
Thus:
| Feature | Between-subject variability | Residual variability |
|---|---|---|
| Notation commonly used | \(\eta\) | \(\epsilon\) |
| Primary level | Subject | Observation |
| Often affects | PK parameters | Observed concentration around prediction |
| Example | Subject-specific clearance differs from typical clearance | Measured concentration differs from model-predicted concentration |
Confusing these two sources of variability can lead to incorrect interpretation of population PK model results.
18. Why Does the Residual Model Affect Parameter Estimates?
The residual error model is not merely a descriptive add-on applied after parameter estimation. It forms part of the likelihood or objective function used to estimate model parameters.
Suppose two observations have the same absolute difference from their predictions:
If the model assumes that the variance depends on the prediction, those two discrepancies may not have the same standardized magnitude.
For example, under a proportional model, a 1 mg/L discrepancy is relatively much larger at 2 mg/L than at 20 mg/L.
Consequently, changing the residual error model can change how different observations contribute to the estimation procedure. This can, in turn, affect estimates of structural PK parameters and between-subject variability.
19. How Should a Residual Error Model Be Selected?
Residual error model selection should be guided by the scientific and statistical characteristics of the data rather than by a universal rule.
A practical sequence is:
- Understand the assay. Determine whether measurement variability is approximately constant, proportional, or more complex across the concentration range.
- Inspect the concentration distribution. Examine the range and scale of observations.
- Fit a reasonable initial model. Additive and proportional formulations are common starting points.
- Examine residual diagnostics. Look for systematic trends and changes in variance.
- Consider a combined model when justified. This may be useful when both absolute and proportional variability are evident.
- Compare model behavior. Consider objective-function changes where appropriate, parameter plausibility, diagnostics, numerical stability, and predictive performance.
- Check whether the model is scientifically coherent. Statistical improvement alone does not establish that a more complex observation model is appropriate.
Model selection should consider the entire model rather than relying on a single criterion.
20. More Complex Residual Models Are Not Always Better
A combined error model has more flexibility than a pure additive or pure proportional model. That flexibility can be useful, but it also introduces additional parameters and can make the model harder to estimate or interpret.
For example, if the additive component is essentially zero over the concentration range represented in the study, estimating a separate additive variance may provide little practical benefit.
Similarly, if the data contain very few low-concentration observations, the study may contain limited information about the low-concentration behavior of the residual error model.
21. Residual Error and Simulation
The residual error model becomes especially important when a population PK model is used for simulation.
A structural model alone can generate a smooth concentration-time curve. A simulation intended to reproduce realistic observations generally requires variability to be added around that curve.
For example, under a proportional error model:
Under an exponential model:
Thus, the residual model determines how simulated observations scatter around the underlying individual concentration-time trajectory.
22. A Simple Diagnostic Thought Experiment
Imagine a dataset containing concentrations from 0.5 to 50 mg/L. Suppose the residual plot shows very small absolute scatter at low concentrations but increasingly large scatter as predictions rise.
Three candidate models might be considered:
- Additive: assumes approximately constant absolute spread.
- Proportional: assumes spread increases approximately in direct proportion to prediction.
- Combined: allows both behaviors.
If the residual standard deviation is approximately 0.2 mg/L at very low concentrations but approximately 10% to 20% of the prediction at higher concentrations, a combined or multiplicative formulation may represent the observed behavior more naturally than a pure additive model.
However, the residual pattern should also be examined for signs of structural misspecification. For example, systematic positive residuals at one part of the concentration-time curve and negative residuals at another part may indicate that the structural model is inadequate rather than that the residual error variance is incorrectly specified.
23. Common Mistakes in Residual Error Modeling
Mistake 1: Treating residual variability as biological variability
Residual error represents variation not explained elsewhere in the model. It should not automatically be interpreted as biological variability.
Mistake 2: Confusing residual error with between-subject variability
Between-subject variability concerns differences between individuals, whereas residual variability concerns observation-level deviations from the model prediction.
Mistake 3: Choosing proportional error simply because concentrations span a wide range
A wide concentration range alone does not establish that proportional error is appropriate. The observed residual behavior and assay characteristics should be considered.
Mistake 4: Ignoring low-concentration behavior
An error model that looks reasonable at high concentrations may behave poorly near the lower end of the concentration range.
Mistake 5: Treating the lowest objective function as the only criterion
Statistical criteria should be considered together with diagnostics, parameter plausibility, model stability, scientific interpretation, and predictive performance.
Mistake 6: Assuming a good residual plot proves the structural model is correct
Residual diagnostics are only one component of model evaluation. A model can display acceptable residual behavior while still being an imperfect representation of the underlying biology.
24. A Practical Residual Error Modeling Workflow
- Understand the observation process. Review assay characteristics, sampling procedures, concentration range, and data quality.
- Specify the structural PK model. Establish a reasonable representation of absorption and disposition before focusing on residual error.
- Account for between-subject variability where appropriate. Separate subject-level parameter variability from observation-level error.
- Start with a simple residual model. Additive or proportional error can provide useful initial formulations.
- Inspect residual diagnostics. Evaluate residuals against predictions, time, and relevant covariates.
- Assess variance behavior. Determine whether absolute or relative variability changes with concentration.
- Consider combined or log-scale formulations when supported.
- Compare candidate models. Evaluate statistical, diagnostic, numerical, and scientific considerations together.
- Check low-concentration and BLQ behavior. Ensure that the model does not produce implausible observation distributions.
- Validate the final model. Consider predictive checks and simulation-based diagnostics where appropriate.
The final residual error model should be understandable, estimable from the available data, and appropriate for the intended use of the population PK model.
25. Worked Interpretation: What Does the Error Model Say?
Suppose a fitted population PK model uses an exponential residual error model with an estimated residual standard deviation of 0.20.
The observation model is:
This means the model describes observation variability on the log scale. A typical log-scale deviation has a standard deviation of 0.20.
For a prediction of 10 mg/L, a simulated observation can be generated by drawing \(\epsilon\) from the specified normal distribution and calculating:
For example, if a particular simulated residual were \(\epsilon=0.20\):
If instead \(\epsilon=-0.20\):
The two deviations are symmetric on the log scale but not on the arithmetic concentration scale. This illustrates the multiplicative nature of the model.
26. What Residual Error Models Do Not Tell Us Automatically
Residual error estimates should be interpreted carefully.
- A residual variance is not a direct estimate of assay imprecision. It can contain multiple sources of unexplained variation.
- A smaller residual variance does not automatically indicate a better model. Changes in structural model or variability assumptions can affect the residual estimate.
- A proportional error estimate is not automatically a biological coefficient of variation. Its interpretation depends on the exact observation model.
- Residual error does not replace structural model evaluation. Systematic residual patterns can indicate structural misspecification.
- Residual diagnostics do not prove model truth. They provide evidence about model adequacy relative to the observed data.
- BLQ observations require separate consideration. Censored observations should not automatically be treated as ordinary quantified concentrations.
- Residual error may depend on the study and assay. An error model that works well for one dataset may not be appropriate for another.
27. Key Takeaways
- Residual unexplained variability describes the difference between observed concentrations and the concentrations predicted by the population PK model.
- Residual variability is distinct from between-subject variability: \(\eta\) commonly represents subject-level parameter variation, whereas \(\epsilon\) commonly represents observation-level error.
- An additive error model assumes approximately constant absolute variability.
- A proportional error model assumes that residual variability increases with predicted concentration.
- A combined additive-plus-proportional model allows both constant absolute and concentration-dependent variability.
- An exponential error model is equivalent to an additive error model on the log-concentration scale.
- The residual error model can influence parameter estimation because it determines how observations are statistically distributed around model predictions.
- Residual diagnostics should be examined against predictions, time, and relevant covariates to identify systematic patterns.
- Large residual variability does not necessarily mean poor assay performance; it may reflect several sources of unexplained observation-level variation.
- Low concentrations require particular attention because additive error can represent a large relative error near the lower end of the concentration range.
- BLQ observations are censored observations and should be considered separately from ordinary quantified concentrations.
- Residual error assumptions also affect simulation because they determine how observations are generated around individual model predictions.
- The most appropriate residual error model is the one that adequately represents the observation process and available data without unnecessary complexity.
Where to Go Next
After understanding residual unexplained variability, a natural next step is to study between-subject variability and random-effects models. That topic explains how population PK models represent differences in clearance, volume of distribution, absorption, and other parameters between individuals.
From there, the concepts can be combined into a complete population PK model containing a structural model, interindividual variability, residual unexplained variability, and covariate relationships.
Understanding these layers provides the foundation for more advanced topics such as NONMEM modeling, nonlinear mixed-effects estimation, covariate model development, visual predictive checks, bootstrap validation, model qualification, and PK/PD population modeling.