Why Eta Shrinkage Flattens Covariate Plots

Why a Covariate Trend Can Disappear

A plot of empirical Bayes estimates (EBEs) against a covariate is often used to look for a systematic relationship that the model has not explained. For example, if an EBE for a clearance random effect tends to increase with body size, that pattern may suggest that the population model has not captured a body-size relationship. But an EBE is not the subject’s random effect observed without error. It is an estimate informed both by that subject’s data and by the model’s population distribution. When the subject’s data provide limited information, the estimate is pulled toward the population value, often zero. This pull can flatten the plotted relationship.

The key implication is about interpretation: a weak or flat EBE–covariate trend does not, by itself, establish that the underlying covariate relationship is absent. The plot may reflect a genuinely weak relationship, substantial uncertainty in the individual estimates, or both. The simple normal model below isolates the mechanism and shows exactly when the slope is attenuated.

A Normal-Model Derivation of Slope Attenuation

Consider a single random effect and a centered covariate \(x_i\). Suppose the true random effect has a linear covariate relationship:

Here \(\gamma\) is the true change in the mean random effect per one-unit increase in \(x_i\), and \(u_i\) is the remaining between-subject deviation. Now suppose the fitted model does not include this covariate relationship and instead uses a zero-centered prior, \(\eta_i \sim N(0,\omega^2)\). To represent the information in subject \(i\)’s concentration data, consider a simple estimate \(z_i\) satisfying

The quantity \(z_i\) is an idealized data-based estimate of the random effect, with uncertainty \(s_i^2\). Under these assumptions, the posterior distribution of \(\eta_i\) is normal. Its mean, which is also its mode, is a weighted average of the data-based estimate and the prior mean:

The weight \(\lambda_i\) lies between zero and one. A large \(s_i^2\) makes \(z_i\) less informative, reducing its weight and pulling the EBE more strongly toward zero. In this simplified setting, \(E(z_i\mid x_i)=\gamma x_i\), so the expected EBE at a given covariate value is

when \(\lambda_i\) is constant across subjects. Thus, the expected EBE–covariate slope is \(\lambda\gamma\), not \(\gamma\). The reduction is not evidence that the true covariate effect has changed: it arises because the plotted estimates are pulled toward the prior center.

The derivation follows directly from normal updating. Multiplying the prior density by the likelihood gives a posterior kernel whose terms involving \(\eta_i\) are

Collecting the quadratic and linear terms in \(\eta_i\), the posterior precision is \(1/\omega^2+1/s_i^2\), and the posterior mean is \((z_i/s_i^2)/(1/\omega^2+1/s_i^2)\). Multiplying numerator and denominator by \(\omega^2s_i^2\) yields \(\omega^2 z_i/(\omega^2+s_i^2)\), as above. This makes the source of the pull explicit: when the data-based estimate is uncertain, the prior contributes more to the posterior estimate.

From the Expected Trend to a Finite Plot

The derivation describes an expected relationship, not a promise that every finite sample will have a slope equal to \(\lambda\gamma\). The plotted EBEs also vary because of the residual deviations \(u_i\), the uncertainty in \(z_i\), the covariate values observed in the sample, and sampling variation in the fitted model. Consequently, one EBE plot can look nearly flat even when the expected slope is not zero. Conversely, sampling variation can produce a visible trend when the underlying relationship is weak or absent.

When Attenuation Is Not a Simple Multiplication

The factor \(\lambda\) is the same for all subjects only under the simplified assumption that their information is comparable. In a population PK analysis, the amount and timing of concentration data, residual variability, and the model’s sensitivity to an individual parameter can differ between subjects. The amount of shrinkage can therefore vary across the covariate range. In general, if the effective weight is \(\lambda_i\), then the conditional mean of the EBE is approximately \(\lambda_i\gamma x_i\) in this simple data-based-estimate representation. When \(\lambda_i\) changes with \(x_i\), the expected pattern need not be a straight line with one uniformly reduced slope. It may be curved, uneven, or especially compressed in one part of the covariate range.

Assumptions, Diagnostics, and Alternatives

The exact slope result requires a normal prior, a normal data-based estimate with the stated conditional variance, a zero-centered prior in the fitted model, and a common reliability weight that does not vary with the covariate. The simple observation model also treats the data-based estimate as centered on the individual random effect. These assumptions make the mechanism transparent, but real nonlinear mixed-effects models need not satisfy them exactly. In particular, an EBE may be a posterior mode rather than a posterior mean, and the relationship between the concentration data and a random effect may be nonlinear. In those cases, the formula is a useful explanation of the direction of the effect, not an exact correction to apply to a plotted slope.

When reviewing an EBE–covariate plot, consider the pattern alongside the information supporting the estimates and the model’s population predictions. A trend can be difficult to see if EBEs are concentrated near zero; uneven compression may appear as different scatter or spread across the covariate range. Also check whether the covariate range is broad enough to reveal a relationship and whether a few subjects dominate the apparent pattern. These checks help describe what the plot can show, but they do not turn the EBE plot into a formal test of the covariate effect.

A practical alternative is to fit and compare population models that include the covariate relationship and assess the change in the model’s fit and parameter estimates using appropriate inferential procedures. A model-based display can also compare population predictions across covariate values. These approaches address the relationship within the fitted likelihood rather than treating noisy, shrunken individual estimates as if they were directly observed random effects. They still depend on model assumptions and adequate data; they are not automatic proof that a covariate has a causal effect.

A common mistake is to reason, “The EBE plot is flat, so the covariate has no effect.” That conclusion treats each EBE as an unshrunk measurement of the true random effect. Under the normal model, however, the expected plotted slope is \(\lambda\gamma\), so it can be near zero because \(\lambda\) is small even when \(\gamma\) is not. The correction is to interpret the plot as a diagnostic of the fitted individual estimates, consider how shrinkage may affect its appearance, and evaluate a proposed covariate relationship with an appropriate population model.