Computing Eta Shrinkage and Its Finite-Sample Caveats

What Eta Shrinkage Measures

Empirical Bayes estimates (EBEs) are subject-specific estimates informed both by that subject’s data and by the fitted population distribution. When individual data are limited, the estimates tend to be pulled toward the population value. Eta shrinkage summarizes this effect by comparing the observed spread of the estimated random effects across subjects with the fitted population spread.

The calculation is component-wise: for each random-effect component, compare the across-subject spread of its EBEs with the fitted standard deviation of that component. It is a descriptive summary of a set of estimates, not a probability that an individual estimate is correct and not a direct measure of the information in any one subject’s data.

Definition: For random-effect component $k$, let $\hat{\eta}_{ik}$ be the EBE for subject $i$, and let $\hat{\omega}_k^2$ be the fitted variance of that component. For $n\ge 2$ subjects and $\hat{\omega}_k^2>0$, define $$ \bar{\hat{\eta}}_k=\frac{1}{n}\sum_{i=1}^{n}\hat{\eta}_{ik}, \qquad s_{\hat{\eta}_k}= \sqrt{\frac{\sum_{i=1}^{n}(\hat{\eta}_{ik}-\bar{\hat{\eta}}_k)^2}{n-1}}, \qquad \widehat{S}_{\eta,k}=1-\frac{s_{\hat{\eta}_k}}{\sqrt{\hat{\omega}_k^2}}. $$ Here $s_{\hat{\eta}_k}$ is the sample standard deviation across subjects, and $\sqrt{\hat{\omega}_k^2}$ is the fitted random-effect standard deviation. The calculation is undefined if there is only one subject or if the fitted variance is zero.

Both quantities in the ratio are standard deviations on the same random-effect scale, so their ratio is unitless. For random effects on a log scale, for example, both spreads are on that log scale. The sample standard deviation centers the EBEs on their observed mean; it does not require their average to be exactly zero. A finite set of EBEs need not have a mean of zero even when the model’s random effects have a population mean of zero.

Calculation and a Short Justification

The formula expresses shrinkage as one minus a relative spread. If the EBE standard deviation equals the fitted random-effect standard deviation, the calculated shrinkage is zero. If the EBEs have less spread, shrinkage is positive; if they have more spread, the result is negative. This description follows directly from the formula: the ratio is less than, equal to, or greater than one, respectively.

The denominator $n-1$ in the sample variance is the usual sample-standard-deviation convention. It centers each value on the sample mean and uses $n-1$ rather than $n$ squared-deviation units. Replacing it with $n$ produces a different, population-style spread for this particular set of EBEs. That convention gives a smaller spread for any nonconstant set, and therefore a larger calculated shrinkage. For comparisons or reporting, state which convention was used.

Worked Example: Computing Shrinkage for Five Subjects

Suppose five hypothetical subjects have EBEs for a log-scale clearance random effect of $-0.30$, $-0.10$, $0.00$, $0.10$, and $0.30$. The fitted variance is $\hat{\omega}_k^2=0.16$ on the squared log scale. There are at least two subjects, and the fitted variance is positive, so the definition applies.

The EBE mean is $({-0.30-0.10+0.00+0.10+0.30})/5=0$. The sum of squared deviations is $0.09+0.01+0+0.01+0.09=0.20$. Thus,

s_{\hat{\eta}_k}=\sqrt{\frac{0.20}{5-1}} =\sqrt{0.05}=0.2236.

The fitted random-effect standard deviation is $\sqrt{0.16}=0.40$. Therefore,

\widehat{S}_{\eta,k}=1-\frac{0.2236}{0.40} =1-0.5590=0.4410.

The calculated eta shrinkage is about $44.1\%$. In this five-subject set, the EBE standard deviation is about $55.9\%$ of the fitted random-effect standard deviation. This comparison describes the relative spreads; it does not say that every subject’s EBE is shrunk by exactly $44.1\%$.

If the same squared deviations were divided by $n=5$, the spread would be $\sqrt{0.20/5}=0.2000$, giving $1-0.2000/0.40=0.5000$. That is a different convention and yields a different value, illustrating why the denominator should be reported.

Finite-Sample Cautions

With few subjects, the across-subject standard deviation can vary substantially from one sample of subjects to another. Using $n-1$ is the standard sample-SD calculation, but it does not make the shrinkage ratio an unbiased estimate of some fixed population quantity. In particular, the ratio is nonlinear, and the EBEs and fitted variance are typically obtained from the same dataset. Uncertainty in both parts of the ratio matters.

The estimated random-effect variance is especially consequential because it appears under a square root in the denominator. Holding the EBE spread fixed, a larger fitted variance produces a larger shrinkage value; a smaller fitted variance produces a smaller value. When the variance estimate is imprecise, a single shrinkage number can therefore give a misleading impression of precision. A value calculated from a small cohort should be read as a cohort-specific diagnostic, not as a stable property guaranteed to hold in another sample.

Worked Example: A Negative Estimate Is Possible

For four hypothetical subjects, suppose the EBEs for a log-scale volume random effect are $-0.12$, $-0.04$, $0.04$, and $0.12$. The fitted variance is $\hat{\omega}_k^2=0.0100$, so its standard deviation is $0.1000$. The EBE mean is zero, and the squared deviations sum to $0.0320$. The sample SD is

s_{\hat{\eta}_k}=\sqrt{\frac{0.0320}{4-1}} =\sqrt{0.010667}=0.1033.

Consequently,

\widehat{S}_{\eta,k}=1-\frac{\sqrt{0.0320/(4-1)}}{0.1000} =-0.0328.

This negative value is the arithmetic result: the observed EBE spread is slightly larger than the fitted random-effect standard deviation. It is not evidence of “negative information,” nor should it be silently reset to zero. With only four subjects, sampling variation and uncertainty in the fitted variance are plausible reasons for this modest discrepancy. The result can also prompt a check of the model fit, the variance estimate, and the calculation convention.

Worked Example: Sensitivity to the Fitted Variance

Suppose a hypothetical cohort has a sample SD of $0.2000$ for the EBEs of a log-scale absorption-rate random effect. Compare two plausible fitted variances for that component, $0.0900$ and $0.1600$. In both calculations, assume at least two subjects and use the same sample-SD convention.

With variance $0.0900$, the fitted standard deviation is $0.3000$, giving

\widehat{S}_{\eta,k}=1-\frac{0.2000}{0.3000} =0.3333.

With variance $0.1600$, the fitted standard deviation is $0.4000$, giving

\widehat{S}_{\eta,k}=1-\frac{0.2000}{0.4000} =0.5000.

The same EBE spread gives shrinkage estimates of $33.3\%$ and $50.0\%$ because the fitted variance differs. This calculation does not determine which variance estimate is more credible; it shows why shrinkage should be considered alongside the fitted variance and its estimation context.

Checks, Comparisons, and Limitations

Before interpreting a reported value, verify that it is for the intended random-effect component, that the denominator is the square root of a variance rather than the variance itself, and that the across-subject SD and fitted SD use the same scale. Also check that there are at least two subjects and that the fitted variance is strictly positive. If every EBE is identical and the fitted variance is positive, the formula gives $1-0/\sqrt{\hat{\omega}_k^2}=1$; this describes zero observed EBE spread in that set, not proof that every subject has no individual information.

Eta shrinkage differs from simply reporting the EBE SD: the latter gives an absolute spread, while shrinkage compares that spread with the fitted population spread. It also differs from an individual-level uncertainty measure: a cohort-level ratio does not reveal whether one particular subject’s estimate is more or less precise than another’s. The empirical Bayes modes and means discussed in the earlier material can differ, especially in nonlinear models; whichever EBE definition is used should be stated and applied consistently. This formula alone does not resolve that difference.

A high positive value indicates that the EBE distribution is narrow relative to the fitted random-effect distribution, but it is not a universal pass/fail threshold. A negative value is possible, and a small positive value is not a guarantee of accurate individual estimates. The summary can also be affected by misspecification, unusual subjects, covariate structure, and uncertainty in the variance estimate. It should be read as one diagnostic alongside model and data checks, not as a substitute for them.

Confirm the calculation is defined.
Use at least two subjects and a strictly positive fitted variance for the component.

Compute the across-subject spread.
Find the EBE mean and sample SD using denominator $n-1$.

Put spreads on the same scale.
Take the square root of the fitted variance, then divide the EBE SD by it.

Interpret with the estimation context.
Report the component, convention, and result; consider subject count and variance-estimate uncertainty.

Key takeaway: Eta shrinkage is $1-$ the across-subject EBE sample SD divided by the fitted random-effect SD. It requires $n\ge 2$ and positive fitted variance, can be negative, and can be sensitive to finite-sample variation and variance estimation.

Check Your Understanding

Answer each question, then compare with the solutions below.

For $n=1$, can the sample-SD eta shrinkage formula be calculated? Why?
Three hypothetical EBEs are $-0.20$, $0.00$, and $0.20$, and the fitted random-effect variance is $0.0400$. Calculate eta shrinkage using the sample SD.
If the EBE sample SD is $0.15$ and the fitted variance is $0.0100$, what is the calculated shrinkage? How should its sign be interpreted?
Keeping the EBE sample SD fixed, what happens to calculated shrinkage when the fitted variance increases?
Why does a positive eta shrinkage value not mean that every subject’s EBE is shrunk by that percentage?

Solutions. (1) No. The sample-SD denominator is $n-1=0$, so the calculation is undefined. (2) The mean is zero and the squared deviations sum to $0.0800$, so $s=\sqrt{0.0800/(3-1)}=0.2000$. The fitted SD is $\sqrt{0.0400}=0.2000$, giving $1-0.2000/0.2000=0$. (3) The fitted SD is $0.1000$, so $1-0.15/0.10=-0.5000$. The negative result means the observed EBE spread exceeds the fitted SD; it is not negative information. (4) The fitted SD rises, the ratio of EBE SD to fitted SD falls, and shrinkage rises. (5) It compares cohort-level spreads. Individual estimates can differ in their data support and their amount of shrinkage, so the cohort summary is not an individual percentage.

Computing Eta Shrinkage and Its Finite-Sample Caveats

What You'll Learn