Why Individual Fit Can Be Misleading
A population pharmacokinetic model is fitted to observations from many individuals, but its interpretation often includes individual-specific quantities such as clearance or volume. Those quantities are informed by each individual's concentration-time data and by the model's assumptions about the population distribution. When an individual's samples are sparse, the data may not distinguish that person's parameters from several plausible alternatives. The model then relies more heavily on the population distribution, and the estimated individual random effects are pulled toward their typical value, usually zero.
This pull is called shrinkage. It matters for model evaluation because a shrunken individual estimate is not a direct summary of what that person's data alone imply. Sparse sampling and high shrinkage can make a correct model appear to have individual-level problems, or conceal systematic model failure when individual diagnostics are interpreted without population-level evidence. The central discipline is to distinguish a weakly informed individual estimate from a repeatable pattern across the population.
This measure is not the only shrinkage definition. For example, a variance-based measure is \(1 - Var(\widehat{\eta})/\omega^2\). These measures are not interchangeable: one compares standard deviations, while the other compares variances. A report should say which definition it uses. Neither measure is an individual-level probability or a direct estimate of how accurately each person's parameter is known.
Why Sparse Data Pull Estimates Toward the Population
The mechanism is easiest to see in a simplified normal model. Suppose an individual's data provide a noisy estimate \(y^*\) of the person's random effect \(\eta\), with \(y^*=\eta+\epsilon\), where \(\eta\sim N(0,\omega^2)\), \(\epsilon\sim N(0,s_e^2)\), and the two are independent. Here \(s_e^2\) represents uncertainty in the individual's data-based estimate. Under these assumptions, the conditional mean of \(\eta\) given \(y^*\) is
The multiplier is between zero and one. If the data are precise, \(s_e^2\) is small and the multiplier is close to one: the conditional estimate can follow the individual's signal. If the data are weak, \(s_e^2\) is large and the multiplier is closer to zero: the estimate is pulled toward the population mean of zero. In a nonlinear PK model this equation is an explanatory approximation, not a universal formula for the fitted EBE; nevertheless, it captures the key role of information in the individual likelihood.
Sparse sampling can increase uncertainty because there may be too few observations, too little coverage of the concentration-time profile, or inadequate information about the parameter of interest. For example, a single concentration can often be compatible with different combinations of clearance, volume, absorption, and residual error. A fitted EBE may therefore be close to zero even when the individual's true parameter differs from the population typical value. Conversely, an isolated unusual observation can pull an EBE away from zero without providing reliable evidence of a persistent individual difference.
Worked Example: Calculating Eta Shrinkage
Suppose a hypothetical analysis estimates \(\omega=0.50\) on a log-clearance random-effect scale. The five EBEs for that random effect are \(-0.20,-0.10,0,0.10,0.20\). Calculate standard-deviation-based eta shrinkage.
The EBE mean is zero. The sum of squared deviations is \(0.04+0.01+0+0.01+0.04=0.10\), so the sample variance is \(0.10/(5-1)=0.0250\), and the sample standard deviation is \(\sqrt{0.0250}=0.1581\). Therefore,
The standard-deviation-based eta shrinkage is about 68.4%. This indicates that the spread of these EBEs is substantially smaller than the fitted population standard deviation. It does not mean that each individual's estimate is 68.4% wrong, nor does it establish that the estimated \(\omega\) is incorrect.
For comparison, the variance-based measure is \(1-0.0250/(0.50^2)=1-0.0250/0.25=0.900\), or 90.0%. The difference is a consequence of the definitions, not an arithmetic inconsistency. The metric must be identified before its value is interpreted.
How Shrinkage Can Mimic or Conceal Misfit
Individual diagnostic plots are useful only when interpreted in light of how much information each individual contributes. With high eta shrinkage, EBEs cluster near zero and individual predictions based on those EBEs tend to resemble the population-typical prediction. That can conceal genuine between-person differences: the fitted individual profiles may look similar even when the data are too sparse to estimate each person's deviation. It can also make a population-typical prediction look inadequate for a particular person's observations, inviting an unsupported change to that person's clearance or absorption parameter.
Residual diagnostics can also be affected. Conditional predictions use individual parameter estimates informed by the observations, so they are not independent predictions of those same observations. When individual estimates are weakly informed or strongly pulled toward typical values, conditional residuals and individual plots can be difficult to interpret. A small-looking conditional residual is not, by itself, proof that the model describes that individual well; an unusual residual is not, by itself, proof that the individual's PK parameters are unusual. The residual definition, estimation method, sampling design, and degree of shrinkage all matter.
The more dependable question is whether errors recur in a structured way across individuals. For example, residuals repeatedly positive in one part of the profile and negative in another may indicate a time-dependent population pattern, even when each person's data are too sparse to support a stable individual diagnosis. Conversely, scattered deviations without a consistent population pattern may simply reflect residual variability or weak information. Population-level residual summaries and predictive checks should therefore be considered alongside individual plots, not replaced by them.
Worked Example: A Sparse Individual Does Not Establish High Clearance
Consider a hypothetical one-compartment, intravenous bolus setting with a known dose of 100 mg. One individual has a single concentration of 5.0 mg/L at 2 hours. Suppose the population-typical values are \(CL=5.0\) L/h and \(V=20\) L, and the fitted individual clearance EBE corresponds to \(CL=5.0\) L/h. Under the model, \(C(t)=(Dose/V)e^{-CLt/V}\). The population-typical prediction at 2 hours is
The observed value is \(5.0-3.03=1.97\) mg/L above that prediction. It is tempting to conclude that this individual has lower clearance. But a single concentration does not separately identify clearance and volume, and its interpretation also depends on residual error and the sampling time. The data-based estimate may be weak, so the EBE can remain close to the population typical value. Without additional information, the correct conclusion is that the observation is above the stated prediction; the individual's clearance is not established to be lower.
Use Population Patterns to Test the Model
A practical evaluation separates two questions: does the model describe the population's systematic concentration-time behavior, and is there enough information to estimate this individual's deviation? Population-level checks address the first question. Individual EBEs and conditional predictions address the second, with reliability limited by each person's sampling. These questions should not be collapsed into a single judgment based on an individual plot.
Worked Example: A Recurring Time Pattern Despite Sparse Profiles
In a hypothetical evaluation, 12 individuals each contribute a small number of observations. Define a raw residual as observed concentration minus the corresponding population prediction. For illustration, suppose the residuals in three time windows are as follows, in mg/L.
| Time window | Residuals (mg/L) | Mean residual (mg/L) |
|---|---|---|
| Early | 1.0, 0.8, 1.2, 0.6 | 0.900 |
| Middle | -0.5, -0.8, -0.6, -0.7 | -0.650 |
| Late | 0.7, 0.9, 0.5, 0.7 | 0.700 |
For the early window, the mean is \((1.0+0.8+1.2+0.6)/4=0.900\) mg/L. For the middle window it is \((-0.5-0.8-0.6-0.7)/4=-0.650\) mg/L, and for the late window it is \((0.7+0.9+0.5+0.7)/4=0.700\) mg/L. The repeated positive, negative, positive sequence is a population-level time pattern worth investigating. It is not evidence that every individual has the same parameter error, and this small illustration does not establish statistical significance. It does show why pooled patterns can be more informative than trying to assign a different clearance to each sparsely sampled person.
Diagnostics, Alternatives, and Limits
When shrinkage is substantial, treat individual-level diagnostics as qualified evidence. Examine the number and timing of observations per person, whether the data cover informative phases of the profile, and whether the model's population predictions show structured residual patterns. Check whether EBEs are compressed relative to the fitted random-effect distribution, and report the shrinkage definition used. If an apparent pattern depends on conditional predictions, compare it with a population-level diagnostic that does not rely on the same individual's EBE in the same way. Predictive checks can also assess whether the fitted model reproduces relevant population summaries, provided the checks reflect the design and the feature being assessed.
There are several alternatives to an immediate individual parameter revision. Additional samples at informative times can improve individual identifiability. A population-level model revision may be more appropriate when residual patterns recur across individuals. If the evidence is only an isolated deviation, retain uncertainty rather than fitting a bespoke explanation. These approaches answer different questions: more sampling improves individual information; population-level diagnostics assess shared model behavior; and individual estimates describe deviations only to the extent the data support them.
Shrinkage values are not universal pass-fail thresholds. They depend on the model, the data design, the random-effect magnitude, residual variability, and the shrinkage definition. A high value can result from limited information without model misspecification, while low shrinkage does not guarantee that the model is correct. The simple normal derivation assumes a correctly specified normal prior and a noisy estimate with independent normal error; real PK likelihoods can be nonlinear, correlated, and non-normal. Treat the derivation as an explanation of the direction of the effect, not as a substitute for model-specific diagnostics.
Common Pitfall
Wrong reasoning: “This individual's EBE is close to zero, so the individual must be typical; the observed discrepancy must be a structural failure for that individual.” The EBE may be close to zero precisely because sparse observations provide too little information to move it away from the population mean. The reverse mistake is to interpret one unusual concentration as proof of an unusual individual parameter.
Correction: Separate the observed discrepancy from its explanation. Check whether a pattern recurs across individuals and times, assess the individual's sampling information, and avoid an individual-level parameter claim unless the data support it. Population-level evidence can motivate model evaluation; an EBE alone cannot resolve weak identifiability.
Check Your Understanding
Answer each question, then compare with the solutions.
- In the standard-deviation-based definition, what does high eta shrinkage indicate, and what does it not prove?
- In the simplified normal derivation, what happens to the conditional mean multiplier when \(s_e^2\) increases?
- Why can one concentration above a population prediction fail to establish that an individual's clearance is low?
- Suppose mean residuals are positive early and late but negative in the middle across several individuals. Why is that more informative than an isolated individual EBE?
Solutions. (1) It indicates that the across-individual spread of EBEs is small relative to \(\omega\); it does not prove that the model or \(\omega\) is wrong, or quantify the error in any one EBE. (2) The multiplier \(\omega^2/(\omega^2+s_e^2)\) decreases toward zero, so the conditional mean is pulled more strongly toward zero. (3) Clearance and other parameters such as volume can produce similar concentrations at one time, and residual variability also affects the observation; the individual data may not identify clearance. (4) A repeated time pattern is evidence of a potentially systematic population-level discrepancy, whereas one EBE is an individual estimate whose reliability depends on that person's information and shrinkage.