Empirical Bayes Estimate Precision Versus Shrinkage

One Aggregate Diagnostic, One Individual Uncertainty

An empirical Bayes estimate (EBE) is an individual-specific estimate of a random effect, such as a subject’s \(\eta_i\). Eta shrinkage summarizes how concentrated EBEs are across a group relative to the population variation in that random effect. The posterior variance, by contrast, describes uncertainty about one subject’s random effect given that subject’s data and the fitted model. Both quantities reflect how much information the data provide, but they do not describe the same target.

This distinction matters when an EBE is used to characterize an individual. A low estimated \(\eta_i\) might reflect a genuinely near-typical subject, or it might reflect weak individual information that pulls the estimate toward the population mean. The EBE alone does not distinguish those possibilities. Eta shrinkage is a warning about aggregate compression; it is not an uncertainty interval for any one subject.

The variance-based definition is used here so that the ratio and its interpretation are explicit. Some reports use the standard-deviation form \(1-\operatorname{SD}(\hat\eta_i)/\omega\), which is a different transformation and generally gives a different numerical value. Always identify the convention when reporting shrinkage. Neither version is an individual posterior variance.

A Normal Model Makes the Relationship Visible

A simple normal-normal model gives a transparent calculation. Suppose \(\eta_i\sim N(0,\omega^2)\). For one subject, suppose the individual-level information can be summarized by \(z_i=\eta_i+e_i\), where \(e_i\sim N(0,\tau_i^2)\), independently of \(\eta_i\). Here \(z_i\) is a hypothetical noisy measurement on the random-effect scale; \(\tau_i^2\) represents its measurement variance. Assume for now that \(\omega^2\) and \(\tau_i^2\) are known.

In this model, the posterior distribution is normal. Its mean is the EBE, and its variance is the uncertainty remaining after observing \(z_i\). Completing the square in the product of the normal prior density and the normal likelihood gives the result below.

Write \(\lambda_i=\omega^2/(\omega^2+\tau_i^2)\). Then \(\hat\eta_i=\lambda_i z_i\), while the posterior variance is \(V_i=\omega^2\tau_i^2/(\omega^2+\tau_i^2)=\omega^2(1-\lambda_i)\). When the subject’s information is weak, \(\tau_i^2\) is large, \(\lambda_i\) is small, and the EBE is pulled toward zero. The posterior variance is then large. This is the basic reason shrinkage and uncertainty are related.

But \(V_i\) is an individual quantity, whereas \(S_\eta\) summarizes the spread of estimates across subjects. To see the precise relationship under the idealized model, apply the law of total variance:

If the EBE is the posterior mean, \(\operatorname{E}(\eta_i\mid Z_i)=\hat\eta_i\), and the population variance is \(\omega^2\). Rearranging gives

Thus, in a correctly specified model with known parameters and posterior-mean estimates, the variance-based shrinkage corresponds to the expected posterior variance as a fraction of the population variance. It is an average relationship, not an identity between aggregate shrinkage and the uncertainty for every individual. With different information levels across subjects, their posterior variances differ even though one aggregate shrinkage value can be reported.

Why Equal EBEs Need Not Have Equal Precision

The EBE is a point estimate; precision concerns how much uncertainty surrounds that point. A subject with little information can have an EBE near zero because the estimate is strongly pulled toward the population mean. Another subject can also have an EBE near zero because the data support a near-zero random effect. Their point estimates match, but the evidence behind them need not.

Aggregate Shrinkage Does Not Identify Each Subject’s Uncertainty

The law-of-total-variance relationship can hold even when subjects have different information. It connects a population-level average of posterior variances to the population-level dispersion of posterior means. It does not allow a single shrinkage statistic to recover each subject’s posterior variance. The following example shows both the useful aggregate connection and its limit.

Interpretation, Diagnostics, and Limitations

Use eta shrinkage to qualify how much the group’s EBEs are compressed relative to the fitted random-effect distribution. Use an individual posterior variance or standard deviation when the question concerns the uncertainty in one subject’s random effect. If individual uncertainty is not available, do not convert an aggregate shrinkage value into a subject-specific confidence or posterior interval.

The normal-normal derivation relies on a normal random-effect distribution, a normal measurement model on the stated scale, independence of measurement error and random effect, and known variance parameters. In the simple model, the posterior mean and mode coincide. In a nonlinear population PK model, the individual likelihood may not be normal, posterior distributions can be asymmetric, and a mode-based EBE need not equal a posterior mean. Moreover, fitted population parameters are estimated rather than known. Posterior variances calculated while treating those fitted parameters as fixed do not capture all uncertainty from estimating them.

Several checks help prevent overinterpretation:

• Check the scale. A posterior variance has squared random-effect units; its square root has random-effect units. Eta shrinkage is a dimensionless proportion or percentage.
• Check the target. Ask whether the number summarizes dispersion across subjects or uncertainty conditional on one subject’s data.
• Check how uncertainty was obtained. A curvature-based approximation or a plug-in posterior calculation relies on model and estimation assumptions; it is not automatically a complete measure of uncertainty.
• Check whether the aggregate hides variation. Subjects can have different sampling information and different posterior variances even when reported shrinkage is one number.

A common mistake is to say, “Eta shrinkage is 50%, so each subject’s EBE has a posterior standard deviation equal to half the population standard deviation.” This is not generally correct. Under the ideal normal model, variance-based shrinkage may equal the expected posterior variance divided by \(\omega^2\); taking square roots changes the scale, and an average does not specify any one subject’s variance. If the question concerns an individual, use that individual’s uncertainty calculation and state its assumptions.

The main alternative is not a competing definition of the same number, but a different summary for a different purpose: report aggregate shrinkage to describe EBE dispersion, and report subject-level posterior uncertainty to describe an individual estimate. In practice, both can be informative. Neither establishes that the fitted model is correct, and neither should be interpreted without the model and information structure that produced it.