Eta Shrinkage Bias in Covariate Screening

Why Eta Shrinkage Matters for Covariate Screening

In a population pharmacokinetic model, an individual random effect represents an individual's deviation from the population model. A covariate can explain some of that deviation. Yet the individual effect is not observed directly: it is inferred from that individual's concentration data, with the population distribution contributing information as well. When individual data are limited or noisy, the resulting empirical Bayes estimate (EBE) is pulled toward its population mean, often zero. This is eta shrinkage.

A common exploratory practice is to plot EBEs against a candidate covariate or regress the EBEs on that covariate. Such plots can be useful for describing model output, but a weak relationship is not reliable evidence that the covariate has little effect. Shrinkage can attenuate the apparent relationship, and differences in shrinkage across individuals can alter its shape. Covariate effects should therefore be evaluated by fitting and testing the covariate in the population model, not by treating EBEs as directly observed individual effects.

For a simple random effect with population standard deviation \(\omega\), one commonly reported summary is \(1-\operatorname{SD}(\hat{\eta})/\omega\), where \(\operatorname{SD}(\hat{\eta})\) is the standard deviation of the EBEs across individuals. Software and reporting conventions can differ, so the exact definition should be checked. This summary describes the spread of the EBEs relative to the estimated population spread; it does not correct individual EBEs or quantify the bias in a particular covariate slope.

How Shrinkage Attenuates a Covariate Relationship

Consider a deliberately simple approximation that makes the mechanism explicit. Let \(z_i\) be a centered covariate, and suppose the latent effect satisfies \(\eta_i=\beta z_i+u_i\), where \(u_i\) has mean zero and variance \(\omega^2\). Imagine that an individual-level data summary is \(y_i=\eta_i+\epsilon_i\), where \(\epsilon_i\) has mean zero and variance \(\tau^2\), independently of \(u_i\) and \(z_i\). The covariate is omitted when calculating the EBE, so the model treats \(\eta_i\) as having mean zero. With these normal assumptions, the posterior mean, which is also the posterior mode, is a weighted average of \(y_i\) and zero.

The factor \(\lambda\) is between zero and one. It is smaller when individual-level uncertainty \(\tau^2\) is large relative to the population variance \(\omega^2\). Since \(E(y_i\mid z_i)=\beta z_i\), the expected EBE conditional on the covariate is

Thus, with the same \(\lambda\) for everyone, the expected slope from regressing EBEs on \(z_i\) is \(\lambda\beta\), not \(\beta\). The slope is attenuated toward zero. This derivation relies on the stated normal approximation and on using an EBE from a model that omits the covariate effect. It illustrates the mechanism; it is not a universal correction formula for fitted population PK models.

Unequal Information Can Distort More Than the Slope

In a population PK analysis, the amount of information about an individual's random effect can vary with sampling times, number of observations, residual variability, and the individual's covariate values. The degree of shrinkage can therefore differ across individuals. If the shrinkage factor is \(\lambda_i\), the simple approximation gives \(E(\hat{\eta}_i\mid z_i)=\lambda_i\beta z_i\), not a single constant times \(\beta z_i\). A regression of EBEs on the covariate may then reflect both the covariate effect and the way information varies across covariate values. The apparent relationship can be curved or otherwise distorted, rather than merely reduced by one common multiplier.

Prefer Tests of the Population Covariate Effect

A model-based test asks whether including a covariate improves the population model, while accounting for the random-effects distribution and the concentration data. For example, a model might specify \(\log(CL_i)=\log(CL_{\mathrm{pop}})+\beta z_i+\eta_i\), where \(CL_i\) is individual clearance, \(CL_{\mathrm{pop}}\) is typical clearance at \(z_i=0\), and \(\eta_i\) is a mean-zero random effect. The null hypothesis of no covariate effect is \(H_0:\beta=0\); a two-sided alternative is \(H_a:\beta\ne 0\).

One standard comparison uses the likelihood-ratio statistic for nested models fitted to the same data. Under \(H_0\), with an identifiable added parameter, an adequate model and the regularity conditions for the usual large-sample approximation, the statistic is approximately chi-square distributed with degrees of freedom equal to the number of added parameters. For one added coefficient, \(df=1\). Boundary parameters, substantial non-identifiability, or other violations of regularity can make this reference distribution unreliable.

Diagnostics, Alternatives, and Limitations

An EBE-versus-covariate plot is still useful as an exploratory display, but it should be read with shrinkage in mind. Check the reported eta-shrinkage summaries and the spread of the EBEs relative to the estimated random-effect distribution. Look for uneven individual data support across covariate values and for patterns in model diagnostics that suggest a poor structural or residual model. These checks can warn that EBE screening is unreliable; none turns EBEs into unbiased observations of the latent effects.

The main alternatives have different purposes. An EBE plot or post hoc regression is quick and may help identify patterns to investigate, but its apparent slope can be attenuated and its usual regression uncertainty does not account for the fact that EBEs are estimated. A population-model likelihood-ratio test directly compares nested models under their stated assumptions. Other model-based inferential procedures may be appropriate when the likelihood-ratio approximation is unsuitable, but their assumptions also need to be checked. A covariate-search simulation can assess whether a proposed screening and testing procedure behaves adequately in a design resembling the analysis; it cannot compensate for an implausible model or poorly chosen simulation assumptions.

Model-based testing is not automatically trustworthy. A significant test can reflect structural misspecification, an unmodeled covariate relationship, or data features not represented by the model. A nonsignificant result does not prove that the effect is exactly zero. The model's parameterization, data support, uncertainty, and scientific relevance remain essential to interpretation. Eta shrinkage is therefore a reason to avoid treating individual EBEs as direct screening measurements, not a reason to accept every model-based result without scrutiny.

Common Pitfall

Incorrect reasoning: “The EBEs show little association with body weight, so body weight has little effect on clearance.” This conclusion treats shrunken estimates as if they were the latent individual effects and ignores how much individual information supported each estimate.

Correction: Use the EBE plot as a descriptive diagnostic, then evaluate the prespecified population covariate effect by fitting the appropriate model and testing or estimating its coefficient. Check shrinkage, data support, model fit, and uncertainty before interpreting either a weak or strong EBE pattern.