Separating Shrinkage From Model Misspecification

Why Shrinkage and Misspecification Must Be Separated

A population pharmacokinetic model can yield empirical Bayes estimates (EBEs) close to zero for many individuals and individual predictions that appear to fit their observations well. One explanation is weak individual information: the data do not distinguish each person’s random effect well from the population distribution, so the EBEs are pulled toward the population mean. Another possibility is that the model is systematically wrong—for example, it may inadequately describe a time-dependent change in clearance. These explanations are not interchangeable, and either can occur without the other.

The practical distinction is important. High eta shrinkage warns that individual estimates and diagnostics based on them may be uninformative. A residual pattern, especially one repeated across individuals or aligned with time, dose, or a covariate, can indicate that the model is systematically missing structure. Neither observation proves the other: high shrinkage does not establish model misspecification, and low shrinkage does not establish that the model is correct.

In this tutorial, \(PRED\) denotes a prediction based on population parameters and the individual’s recorded design and covariates, without that individual’s EBEs. \(IPRED\) denotes the corresponding individual prediction that uses the EBEs. A population-level residual compares an observation with \(PRED\); a conditional or individual-level residual compares it with \(IPRED\). Standardized residuals also account for a model-based measure of residual variability. Their exact scaling depends on the residual definition, so compare like with like and state which diagnostic is being examined.

What Each Diagnostic Can—and Cannot—Tell You

Population-level residuals are useful for asking whether the model’s typical prediction has a systematic discrepancy. For example, residuals that are predominantly positive at late times and negative at early times may suggest a time-related pattern the model has not captured. Repetition across many individuals is more concerning than a similar-looking pattern from one person. Still, sampling variation, an incorrect residual error model, influential observations, or data issues can produce patterns that resemble structural bias.

Individual predictions incorporate EBEs and therefore adapt to each person’s observed data. That adaptation can reduce individual residuals even when individual information is weak. With high eta shrinkage, the EBEs carry limited information about between-person differences, and a plot of EBEs against a covariate may look flat even when the relationship exists. Conditional residuals and individual fits must therefore be interpreted alongside shrinkage, not as independent confirmation of good individual fit.

Information is parameter-specific. A sampling schedule may provide useful information about clearance while providing little information about absorption rate, or the reverse. A single overall shrinkage value, or a judgment that “the model has enough samples,” can conceal this difference. The relevant questions include which parameter is implicated by the residual pattern, whether the sampling times are informative about that parameter, and whether the pattern persists across subjects and relevant design or covariate strata.

A Short Derivation: Why Weak Information Produces Shrinkage

A simple normal model shows the mechanism. Suppose an individual random effect \(\eta\) has prior distribution \(\eta\sim N(0,\omega^2)\). Suppose the individual’s data can be summarized locally by \(z=\eta+e\), where \(e\sim N(0,s^2)\), independent of \(\eta\). Here \(s^2\) represents uncertainty in the individual-level information about \(\eta\). The posterior distribution is normal, and completing the square in its density gives:

The posterior mean is a fraction \(\lambda=\omega^2/(\omega^2+s^2)\) of the individual’s data summary. If \(s^2\) is large, \(\lambda\) is small and the estimate is pulled toward zero. Across individuals under this model, \(\operatorname{Var}\{E(\eta\mid z)\}=\lambda\omega^2\), so the variance-based shrinkage is \(1-\lambda=s^2/(\omega^2+s^2)\). Thus shrinkage reflects the balance between individual information and population variability. This derivation is an explanatory normal approximation; in a nonlinear mixed-effects model, the actual information and EBE behavior can be more complicated.

Use Residual Patterns as Separate Evidence

When residuals appear patterned, first describe the pattern precisely: which residual type, which direction, and which design or subject feature? A residual defined as observation minus prediction is positive when the observation exceeds the prediction. For standardized residuals, compare their location and spread across meaningful groups and look for systematic ordering, not isolated large values alone. Then ask whether the pattern occurs in population-level residuals, conditional residuals, or both. This comparison helps separate a discrepancy in typical predictions from apparent improvement produced by individual adaptation.

A group-level pattern is more suggestive of systematic misspecification when it recurs across subjects under comparable conditions. For instance, a rising residual trend over time in population predictions could be consistent with a missing time-dependent process. But it is not a unique signature of that mechanism: a mismatched residual error model, an omitted covariate, an incorrect dose history, or another model component could also contribute. The diagnostic identifies a question to investigate, not a unique answer.

Conversely, high conditional fit does not erase a population-level pattern. If EBEs allow predictions to adjust toward each individual’s observations, conditional residuals may look less patterned than population residuals. Under high shrinkage, even that apparent improvement should be interpreted cautiously because the EBEs may not represent well-estimated individual differences. This is why shrinkage and residual patterns should be reported and interpreted as complementary evidence.

Parameter-Specific Information Changes the Interpretation

A residual pattern should be connected to the parameters and processes that could plausibly affect it, while recognizing that several model features can produce similar patterns. Then assess whether the design contains information about the implicated parameter. A parameter with weak individual information may have strongly shrunk EBEs even if another parameter is well informed. In turn, population-level residuals can reveal a shared discrepancy even when individual estimates for that parameter are unreliable.

For example, observations concentrated in a narrow portion of the concentration-time profile may offer limited separation of absorption and disposition features. A parameter-specific shrinkage calculation can flag that limited individual resolution. It cannot by itself identify a missing process, and the absence of a visible EBE-covariate relationship under substantial shrinkage is weak evidence against such a relationship. Examine the actual sampling times, dose history, and pattern in the relevant residuals before drawing a conclusion.

Assumptions, Checks, and Common Failure Modes

The variance-based calculation is most interpretable when the EBEs and \(\omega^2\) refer to the same fitted random effect and the same model parameterization. It summarizes contraction across individuals; it is not an individual-level uncertainty interval. Its usual interpretation also relies on the fitted model providing a meaningful estimate of between-subject variability. Model-estimation problems, few subjects, unusual observations, or misspecification can affect both the estimated \(\omega\) and the EBE distribution.

Residual interpretation also depends on the chosen residual definition and its variance scaling. A raw residual’s spread can change with the predicted concentration or measurement error, so a trend in raw residuals may reflect heteroscedasticity rather than a trend in the mean prediction. Standardization addresses modeled variability but cannot repair an incorrect variance model. Check whether the apparent pattern is consistent across subjects and whether it remains after considering the relevant design, dose, and covariate structure.

• Do not equate high shrinkage with a bad structural model. It signals limited separation of individual effects from the population distribution; a systematic residual pattern is separate evidence.
• Do not treat good individual fit as proof of good population fit. IPRED uses EBEs to adapt to observations and can conceal discrepancies in PRED.
• Do not generalize one parameter’s shrinkage to every parameter. Information and shrinkage can differ substantially by random effect.
• Do not interpret one residual excursion as a systematic pattern. Look for coherent replication and consider influential observations and data or error-model issues.
• Do not treat absence of a pattern as proof of adequacy. Sparse or uninformative data may have little ability to reveal a discrepancy.

The main alternative to shrinkage-based qualification is to rely only on residual fit, especially conditional residuals. That approach can overlook weak individual information and overstate how much the EBEs explain. The reverse alternative—relying only on shrinkage—can overlook a repeated population-level discrepancy. The more defensible interpretation keeps both evidence streams distinct: shrinkage qualifies what can be learned about individuals, while residual patterns indicate where predictions may be systematically inadequate. Neither diagnostic identifies a unique cause, and both are conditional on the model, data, and diagnostic construction.