Reporting Shrinkage and EBE Reliability by Parameter

Why Report Shrinkage by Parameter?

Shrinkage is a useful warning about how much individual estimates are pulled toward the population prediction, but a single percentage cannot describe every parameter or every individual. A dataset may support individual clearance estimates well while providing little information about absorption rate. Reporting only one overall shrinkage value hides that difference. Reporting only an EBE plot can hide it too: a narrow distribution of EBEs may reflect genuinely similar individuals, substantial shrinkage, or both.

A useful report therefore pairs population-level summaries of shrinkage with parameter-specific EBE distributions, individual uncertainty, and a concise description of the observations supporting those estimates. These quantities answer related but distinct questions. Shrinkage summarizes the spread of individual estimates across the sample relative to population variability; a reliability measure summarizes how uncertain an individual estimate remains relative to its prior variability.

The notation \(\hat{\omega}_k\) denotes the fitted standard deviation, so the comparison is between two standard deviations on the same random-effect scale. Shrinkage is often reported as a percentage after multiplying the fraction by 100. The formula is a descriptive estimator, not a universal diagnostic test. Finite-sample variation, model misspecification, covariate conditioning, and the way EBEs are computed can all affect it.

What the Numbers Do—and Do Not—Say

Eta shrinkage near zero means that the across-subject EBE spread is close to the fitted population standard deviation. Larger positive shrinkage means the EBE spread is more contracted relative to that fitted variability. It does not establish that every EBE is poor, nor does a low value prove that every individual estimate is precise. Conversely, high shrinkage is a reason to examine individual information and the model, not an automatic reason to discard all EBEs.

An EBE distribution should be reported in the coordinates in which the random effects were modeled. For example, if clearance is modeled as \(CL_i=CL_{\mathrm{typ}}\exp(\eta_{CL,i})\), the shrinkage calculation applies to \(\eta_{CL,i}\), not directly to clearance in volume-per-time units. Describe the corresponding clinical parameter separately when useful, making clear that the transformed parameter distribution may be asymmetric. For correlated random effects, report the effect-specific marginal variances and the fitted correlation structure when relevant; componentwise summaries do not make the effects independent.

A parameter-specific EBE summary can include the number of subjects, mean and standard deviation of the EBEs, selected quantiles, and a plot that reveals skewness, outliers, or multimodality. State whether EBEs are conditional modes or means and whether the reported spread is sample SD. In a model with covariates, specify whether the reported random effects are deviations around the covariate-conditioned typical values. Do not treat a visually narrow EBE distribution as evidence of low population variability without comparing it with the fitted \(\hat{\omega}_k\).

Individual Reliability and a Short Justification

For an individual, an uncertainty-based reliability measure can be defined on the random-effect scale as the fraction of prior variance removed by conditioning on that individual's data. Let \(\widehat{\mathrm{Var}}(\eta_{ik}\mid y_i)\) denote the estimated conditional variance for subject \(i\)'s random effect, using the fitted model and that subject's observations. A practical measure is:

This definition must be named and its uncertainty method stated, for example, a conditional posterior covariance estimated from a local Hessian approximation. It is not interchangeable with the sample variance of EBEs. In a correlated-effects model, the conditional variance in the numerator is the corresponding diagonal element of the full conditional covariance matrix, while \(\hat{\omega}_k^2\) is the marginal prior variance for that effect. An alternative conditional-prior comparison is possible, but it answers a different question and must be explicitly defined.

The relationship between shrinkage and this reliability measure can be seen in a simple Gaussian illustration. Suppose \(\eta_i\sim N(0,\omega^2)\), and a subject contributes a noisy measurement \(z_i=\eta_i+e_i\), with \(e_i\sim N(0,\tau^2)\) independent of \(\eta_i\). The conditional mean and variance are:

The conditional variance formula follows by combining the prior precision \(1/\omega^2\) and measurement precision \(1/\tau^2\): the posterior precision is their sum. Across repeated subjects under this model, the variance of the conditional means is \(\omega^4/(\omega^2+\tau^2)\). Thus the SD ratio of conditional means to prior SD is \(\sqrt{\omega^2/(\omega^2+\tau^2)}\), while the reliability defined above is \(\omega^2/(\omega^2+\tau^2)\). The two summaries are related, but are not numerically the same. Nonlinear models, estimated population parameters, sparse sampling, and non-Gaussian conditional distributions weaken this simple relationship; do not infer an individual's reliability from an aggregate shrinkage percentage.

Report Data Support, Not Just a Shrinkage Percentage

Reliability varies across individuals. Useful summaries include the median and range or selected quantiles of \(\widehat{R}_{ik}\), plus a plot against an interpretable measure of data support. Support is parameter-specific: the number of observations alone may not indicate how informative a profile is. Relevant context can include the timing of samples, whether observations cover the rise and decline of concentrations, the spread of doses, and whether the observations are precise enough to distinguish competing parameter effects. Describe the actual design features rather than declaring a profile informative solely because it has many rows.

Conditional variance estimates also have limitations. A Hessian-based approximation describes local curvature around an individual conditional estimate; it may be unreliable when the likelihood is asymmetric, flat, multimodal, or near a parameter boundary. For a derived parameter such as clearance, uncertainty on the eta scale does not directly state uncertainty in physical units. If reliability is computed after a transformation, explain the transformed-scale definition and method rather than silently comparing unlike variances.

Keep uncertainty about the fitted population model separate from uncertainty for a particular subject. The conditional variance above is typically calculated while treating fitted population parameters as fixed. It therefore does not include all uncertainty in \(\hat{\omega}_k\), fixed effects, or model choice. If inferential precision for the reported shrinkage itself matters, assess its sampling variability—for example, by resampling subjects and refitting the model—and report that procedure and its limitations. A narrow conditional interval for one individual is not a confidence interval for the population shrinkage statistic.

Epsilon Shrinkage: Use the Intended Residuals

For conventional epsilon shrinkage, use the standard deviation of individual weighted residuals (iWRES), calculated according to the residual definition used for the analysis. Under the conventional summary, \(1-s_{\mathrm{iWRES}}\) describes contraction of their spread. State how residuals were standardized and what observations entered the calculation. A value above one for \(s_{\mathrm{iWRES}}\) produces negative shrinkage; report and investigate it rather than truncating it to zero.

Conditional weighted residuals (CWRES) are a different model-based residual diagnostic. Their SD is not generally an epsilon-shrinkage measure: substituting it for \(s_{\mathrm{iWRES}}\) changes the quantity being summarized and its interpretation. CWRES can be reported for their own diagnostic purpose, but label them separately. Also explain any residual handling that affects the iWRES SD, such as exclusions or special weighting, so that the number is reproducible.

How to Assemble a Defensible Report

A concise reporting table can make parameter-level differences visible, but the table should not replace the explanation of scale, data, and uncertainty. The following sequence keeps the summaries interpretable.

An alternative to relying on shrinkage alone is to report conditional uncertainty for each subject; this is more directly individualized but depends on the conditional-distribution approximation and can be cumbersome to summarize. EBE plots reveal distributional shape but do not by themselves quantify reliability. Residual diagnostics such as CWRES assess different aspects of model fit and should remain distinct from the iWRES-based epsilon-shrinkage calculation. These approaches complement rather than replace one another.

Common Pitfalls and Limitations

• Using one cutoff as a verdict. The same shrinkage percentage can have different implications for different parameters and designs. Treat it as a prompt for parameter-specific checks, not a pass/fail rule.
• Equating EBE spread with individual precision. Across-subject spread and an individual's conditional variance are different summaries. Report both when individual reliability is important.
• Substituting CWRES for iWRES. CWRES SD is not generally epsilon shrinkage. Use the iWRES-based definition for the conventional calculation and label other residual summaries accurately.
• Comparing mismatched scales. The EBE SD and population SD must refer to the same random effect and scale. Do not compare log-effect variability with variability in the exponentiated parameter.
• Ignoring model and approximation limits. A conditional Hessian can be misleading for irregular likelihoods, and treating fitted population parameters as fixed omits their uncertainty.

Shrinkage summaries are most useful when they lead to a transparent account of what the model learned from each parameter's data. Their role is not to certify EBEs, but to make contraction, remaining uncertainty, and uneven information visible to the reader.