1. What Are Individual and Population Predictions?
Population pharmacokinetic models describe both the typical PK behavior of a population and the variability between individuals. Once such a model has been fitted, it can generate different kinds of predictions depending on whether individual-specific information is used.
A population prediction describes what the model expects for a typical individual with specified covariates, without using that individual's observed drug concentrations to estimate their personal random effects.
An individual prediction incorporates information about a particular subject—typically including their estimated individual random effects—to produce a prediction tailored to that subject.
Population predictions use typical population parameters and covariates. Individual predictions additionally incorporate subject-specific information, commonly through empirical Bayes estimates of individual random effects.
2. The Population PK Model
A basic population PK model separates parameters into a typical population component and a between-subject variability component.
For an individual parameter such as clearance, a common exponential model is:
Here, \(CL_{pop}\) is the typical population clearance and \(\eta_{CL,i}\) represents the deviation of subject \(i\)'s clearance from the typical value.
When covariates are included, the typical parameter may depend on subject characteristics:
The covariate portion describes systematic differences associated with measured characteristics, while \(\eta_i\) represents unexplained between-subject variability remaining after the modeled covariates are accounted for.
| Component | Role | Used in population prediction? |
|---|---|---|
| Typical parameter \(\theta\) | Describes the typical value in the modeled population | Yes |
| Covariates | Describe systematic relationships between subject characteristics and PK parameters | Yes, when available |
| Random effect \(\eta_i\) | Describes subject-specific deviation from the typical prediction | No |
| Estimated random effect \(\hat{\eta}_i\) | Subject-specific estimate obtained using individual observations | No |
3. What Is a Population Prediction?
A population prediction, commonly denoted PRED, is generated from the fixed-effects portion of the population model. It uses the typical population parameters and applicable covariates but sets the individual random effects to their population-average value, usually zero.
For the exponential clearance model:
Notice that the subject's weight can make the prediction different from the overall typical value even though the subject-specific random effect is not used.
This distinction is important. A population prediction is not necessarily the same for every subject. If the model includes covariates, two subjects with different covariate values can have different PRED values.
4. What Is an Individual Prediction?
An individual prediction, commonly denoted IPRED, incorporates the estimated individual random effects for a subject.
For the same clearance model:
The estimated random effect \(\hat{\eta}_{CL,i}\) is obtained by combining the population model with the individual's observed concentration data, together with the residual-error model and other relevant information.
Thus, IPRED is informed by the subject's own data. If a subject's observations consistently suggest higher clearance than predicted from their covariates alone, their estimated individual effect may shift the IPRED trajectory upward or downward as appropriate.
5. Where Does the Individual Prediction Come From?
Population PK models typically assume that individual parameters vary around population-typical values. In a simple model:
The random effect \(\eta_i\) is unobserved directly. Instead, it is estimated from the individual's data after the population model has been fitted.
These estimates are often called empirical Bayes estimates (EBEs) or individual parameter estimates. They represent estimates of the individual's random effects conditional on the available observations and the fitted population model.
6. PRED vs. IPRED
| Feature | PRED | IPRED |
|---|---|---|
| Population parameters | Uses typical population parameters | Uses typical population parameters |
| Covariates | Uses modeled covariates | Uses modeled covariates |
| Individual random effects | Not included | Estimated individual effects are included |
| Uses individual observations | No | Yes, indirectly through individual-effect estimation |
| Represents | Population-model expectation for the subject's covariates | Subject-specific model prediction |
| Common diagnostic use | Population-level goodness of fit | Individual-level goodness of fit |
The distinction is especially useful in diagnostic plots. Comparing observations with PRED can reveal systematic population-level model deficiencies, while comparing observations with IPRED can reveal how well the model can describe the observations after accounting for estimated individual variability.
7. PRED, IPRED, and Residuals
Residual diagnostics often distinguish between population-level and individual-level prediction errors.
A commonly used population residual is:
where \(DV_i\) is the observed dependent variable and \(PRED_i\) is the corresponding population prediction.
An individual prediction error can similarly be written as:
The terminology and exact residual definitions vary across modeling software and diagnostic workflows, but the conceptual distinction is the same: one prediction excludes the individual's estimated random effects, while the other includes them.
8. Worked Example: Clearance for Two Subjects
Suppose a population PK model estimates a typical clearance of 5 L/h for a 70-kg subject and uses the following weight relationship:
Consider two subjects, one weighing 70 kg and another weighing 100 kg.
Step 1: 70-kg subject
Step 2: 100-kg subject
Now suppose the 100-kg subject has an estimated individual random effect of:
The individual prediction becomes:
The population prediction is therefore approximately 6.52 L/h, while the individual prediction is approximately 7.97 L/h.
The difference is produced by the subject-specific random effect. The population model says that a 100-kg subject would be expected to have a clearance around 6.52 L/h based on weight alone. The individual data provide additional information suggesting that this particular subject has higher clearance than that population-based expectation.
9. How Shrinkage Affects Individual Predictions
Individual predictions depend on estimates of individual random effects. When individual data are sparse or weakly informative, those estimates may be pulled toward the population mean.
This phenomenon is known as eta shrinkage. With substantial shrinkage, individual random-effect estimates may be closer to zero than the underlying individual differences would be if they could be observed directly.
Consequently, IPRED may be relatively close to PRED when the data provide limited information about an individual's random effects.
For this reason, shrinkage should be considered when interpreting individual-level diagnostics based on EBEs or IPRED.
10. Why Compare PRED and IPRED?
PRED and IPRED answer different diagnostic questions.
| Diagnostic question | Prediction of interest | What it can reveal |
|---|---|---|
| Does the model capture systematic population trends? | PRED | Bias associated with time, dose, covariates, or concentration |
| How well can the model describe individual observations? | IPRED | Within-subject agreement after accounting for individual effects |
| Are observations systematically above or below population expectations? | PRED | Potential structural or covariate misspecification |
| Are individual random effects being used to explain observations? | IPRED | Subject-level fit and information captured by individual effects |
A common diagnostic strategy is therefore to examine both population-level and individual-level predictions rather than relying on only one.
11. Common PRED and IPRED Plots
Several standard diagnostic plots use population and individual predictions.
Observed vs. PRED
A plot of observed concentrations against PRED evaluates how well the population component of the model predicts the data. Systematic deviations from the line of identity can indicate model deficiencies or unmodeled structure.
Observed vs. IPRED
A plot of observed concentrations against IPRED evaluates the model after subject-specific random effects have been incorporated. It generally represents a closer fit because the individual predictions have access to information from each subject's observations.
Residuals vs. PRED
Population residuals plotted against PRED can help identify heteroscedasticity, systematic bias, or inappropriate residual-error assumptions.
Residuals vs. IPRED
Individual residual diagnostics can help evaluate the residual-error model after accounting for estimated individual PK parameters.
12. Covariates Can Change Population Predictions
Population predictions are often individualized through measured covariates even before individual concentration data are considered.
For example, if clearance depends on body weight:
then two subjects with different weights can have different population predictions.
Other covariates may include age, renal function, sex, disease status, formulation, or concomitant medications, depending on the scientific context and the model.
This creates an important hierarchy:
- Population prediction: typical population behavior adjusted for measured covariates.
- Individual prediction: population prediction further adjusted using estimated individual random effects.
13. Population Predictions Are Especially Important for New Subjects
For a subject with no concentration observations yet, an individual random-effect estimate cannot be obtained from that subject's PK data. The model therefore starts with the population prediction, adjusted for available covariates.
This is particularly important in prospective dosing applications. A population PK model can generate an initial prediction using known patient characteristics. As concentration data become available, individual parameter estimates can potentially be updated through Bayesian estimation or another appropriate individualization approach.
A population prediction can be available before individual PK observations exist. Individual prediction becomes possible after subject-specific information is incorporated into the model.
14. From Population Prediction to Individual Prediction
The transition from population to individual prediction can be understood as a form of model-based updating.
Before observing a subject's PK concentrations, the population model supplies a distribution of plausible individual parameters. Covariates may shift the expected value of that distribution.
After observations are collected, the individual's data provide additional information. The estimated individual parameters then reflect both the population model and the subject's observed concentrations.
In Bayesian terms, the population model supplies prior information about individual parameters, while the observed data contribute likelihood information. The resulting posterior distribution provides the basis for individual parameter estimates and predictions.
In practical population PK workflows, the resulting empirical Bayes estimates are commonly used to obtain subject-specific predictions.
15. PRED, IPRED, and Simulation-Based Predictions
Population PK software and diagnostic workflows may use several related prediction concepts. It is useful to distinguish deterministic model predictions from simulation-based prediction distributions.
| Prediction type | Main information used | Typical purpose |
|---|---|---|
| PRED | Population parameters + covariates | Assess population-level model behavior |
| IPRED | Population parameters + covariates + individual random effects | Assess individual-level fit |
| Population simulation | Population parameters + variability + residual error | Describe expected distributions of future observations |
| Individual simulation | Subject-specific parameters + residual error | Explore predicted trajectories for an individual |
These concepts should not be treated as interchangeable. A fitted prediction for an observed subject and a simulated future observation answer different questions.
16. What a Difference Between PRED and IPRED Means
The difference between PRED and IPRED provides information about how much the subject-specific random effects contribute to the fitted trajectory.
If PRED and IPRED are very similar, the estimated individual parameters may be close to the population prediction. This can occur because the subject is genuinely close to the population expectation, because covariates explain much of the variation, or because the subject's data provide limited information and the individual estimates are strongly influenced by the population distribution.
If PRED and IPRED differ substantially, the individual observations have provided evidence supporting a subject-specific deviation from the population prediction.
Neither situation is automatically desirable or undesirable. The interpretation depends on the model, the amount of information per subject, shrinkage, sampling design, and the scientific question.
17. A Practical Workflow for Using PRED and IPRED
- Fit the population PK model. Estimate typical parameters, between-subject variability, covariate effects, and residual variability.
- Generate population predictions. Calculate PRED using the fixed-effects model and relevant covariates.
- Estimate individual effects. Use each subject's available observations to estimate individual random effects when appropriate.
- Generate individual predictions. Calculate IPRED using the estimated individual effects.
- Compare observations with both predictions. Examine observed vs. PRED and observed vs. IPRED plots.
- Inspect residual behavior. Evaluate whether residual patterns suggest bias, heteroscedasticity, or model misspecification.
- Assess shrinkage. Determine whether individual-effect estimates are sufficiently informed by the data for the intended diagnostic or interpretation.
- Interpret the predictions in context. Distinguish population-level model performance from individual-level fit.
This workflow helps prevent a common interpretive error: treating an improved individual fit as evidence that the population structural model itself is necessarily adequate.
18. Key Takeaways
- Population predictions (PRED) use typical population parameters and applicable covariates without incorporating subject-specific random-effect estimates.
- Individual predictions (IPRED) incorporate estimated individual random effects and therefore use information from the subject's observed PK data.
- A population prediction can differ between subjects when covariates such as body weight or renal function are included in the model.
- Individual predictions are commonly based on empirical Bayes estimates of individual random effects.
- PRED is particularly useful for evaluating population-level model behavior, whereas IPRED is useful for evaluating individual-level fit.
- A difference between PRED and IPRED is expected when individuals deviate from the population prediction.
- Large PRED–IPRED differences are not automatically evidence of model failure; they can reflect legitimate between-subject variability.
- Shrinkage can pull individual random-effect estimates toward the population mean, particularly when individual data are sparse or weakly informative.
- PRED and IPRED should be interpreted together with residual diagnostics, parameter estimates, variability estimates, and other model-evaluation tools.
- For a new subject without PK observations, population predictions provide the starting point; subject-specific data can subsequently support individualization.
- Population and individual predictions answer different scientific questions and should not be treated as interchangeable.
Where to Go Next
A natural progression is to study eta shrinkage and its interpretation, followed by epsilon shrinkage and residual error, visual predictive checks, prediction-corrected visual predictive checks, and normalized prediction distribution errors.
These diagnostics build directly on the distinction between population and individual predictions and show how model-based predictions can be evaluated across individuals, time, and the observed concentration distribution.