Eta Distributions And Shrinkage

Why Eta Plots Need a Shrinkage Check

Random effects, commonly denoted by \(\eta_i\), describe how subject \(i\)'s model parameters deviate from the population parameters. A common NLME assumption is that these effects follow a multivariate normal population distribution with mean zero and covariance matrix \(\Omega\). An eta histogram or density plot can help identify unusual features in the estimated subject effects, but the plotted values are usually not the unobserved \(\eta_i\) themselves. They are estimates calculated using each subject's data and the fitted model.

Those subject-specific estimates are often called empirical Bayes estimates (EBEs). When a subject's data provide little information about that subject's random effects, the estimate is pulled toward the population mean, usually zero. This is eta shrinkage. Consequently, an EBE distribution can be narrower than the population random-effect distribution, and its shape can also be distorted. A plot of EBEs is therefore a diagnostic of estimates and model fit, not a direct sample from the population distribution.

Definition: In an NLME model, \(\eta_i\) is subject \(i\)'s latent random-effect vector, commonly assumed to have population distribution \(\eta_i\sim N(0,\Omega)\). An empirical Bayes estimate, written \(\hat{\eta}_i\), is a subject-specific estimate conditional on the fitted model and that subject's observations. Eta shrinkage describes the tendency for these conditional estimates to be less dispersed than the corresponding population random effects, especially when individual data are weakly informative.

Quantifying Shrinkage for One Eta Component

For eta component \(j\), let \(\omega_{jj}\) be its estimated population variance, so its estimated population standard deviation is \(\sqrt{\omega_{jj}}\). Let \(s_{\hat{\eta}_j}\) be the sample standard deviation of the subject-specific estimates \(\hat{\eta}_{ij}\) across \(n\) subjects. A commonly reported descriptive measure is:

\text{eta shrinkage}_j = 1-\frac{s_{\hat{\eta}_j}}{\sqrt{\omega_{jj}}}

The result is often multiplied by 100 and reported as a percentage. For example, a value of 0.30 corresponds to 30% shrinkage: the EBE standard deviation is 70% of the estimated population standard deviation. This is a comparison of spreads, not a measure of how accurately any one subject's eta has been estimated. It is also a descriptive summary, not a test of the normality assumption.

This formula uses the marginal standard deviation for the component being examined. If \(\Omega\) has off-diagonal elements, those describe population covariance between components; they do not replace \(\sqrt{\omega_{jj}}\) in the component-wise calculation. The sample standard deviation of the EBEs centers them at their sample mean. Inspect that mean separately: a nonzero average estimate may be a finite-sample feature or a clue to model structure, but it is not itself the shrinkage measure.

Worked Example: Calculating Shrinkage From Seven Eta Estimates

Suppose a hypothetical fitted model has an estimated population standard deviation of 0.300 for one eta component. The EBEs for seven subjects, in the component's dimensionless units, are \(-0.30,-0.20,-0.10,0,0.10,0.20,0.30\). Calculate the descriptive eta shrinkage.

The estimates have mean zero. Their squared deviations sum to \(0.28\), so their sample variance and sample standard deviation are:

s_{\hat{\eta}}^2=\frac{0.28}{7-1}=0.04667, \qquad s_{\hat{\eta}}=\sqrt{0.04667}=0.2160

The population standard deviation is \(\sqrt{\omega_{jj}}=0.300\). Therefore:

\text{eta shrinkage} = 1-\frac{0.2160}{0.300} = 0.280 = 28.0\%

The EBE spread is 72.0% of the estimated population spread. This calculation does not show that any particular subject's EBE is 28% too small, nor does it establish whether the latent random effects are normally distributed. With only seven subjects, the histogram is also too small to support a strong shape judgment.

Why Conditional Estimates Contract Toward Zero

A simple normal-normal model shows why low-information subjects tend to have EBEs close to zero. Suppose a latent random effect \(\eta\) has prior distribution \(N(0,\omega^2)\), and a subject's data are summarized by \(z=\eta+e\), where \(e\sim N(0,s^2)\) is independent measurement error. Here \(s^2\) represents uncertainty in the subject-level summary, not the population variance. The conditional distribution of \(\eta\) given \(z\) is normal. Completing the square in its density gives the conditional mean and variance below.

\begin{aligned} -\frac{1}{2}\left(\frac{\eta^2}{\omega^2} +\frac{(z-\eta)^2}{s^2}\right) &= -\frac{1}{2}\left(\frac{1}{\omega^2}+\frac{1}{s^2}\right) \left(\eta-\frac{\omega^2}{\omega^2+s^2}z\right)^2 +\text{constant in }\eta,\\ E(\eta\mid z)&=\frac{\omega^2}{\omega^2+s^2}z,\\ \operatorname{Var}(\eta\mid z)&=\frac{\omega^2s^2}{\omega^2+s^2}. \end{aligned}

The multiplier on \(z\) is between zero and one. When \(s^2\) is large, the data are uncertain about \(\eta\), the multiplier is small, and the conditional estimate is pulled toward zero. When \(s^2\) approaches zero, the data are highly informative and the estimate approaches \(z\). In realistic nonlinear models, the calculation is more complex and EBEs are often conditional modes rather than conditional means, but the same information-dependent tendency is important.

In this simple model, the contraction also appears in the spread across subjects. The marginal variance of \(z\) is \(\omega^2+s^2\). Since the conditional mean is a fixed multiple of \(z\), its variance across subjects is:

\operatorname{Var}\{E(\eta\mid z)\} = \left(\frac{\omega^2}{\omega^2+s^2}\right)^2 (\omega^2+s^2) = \frac{\omega^4}{\omega^2+s^2} < \omega^2 \quad\text{when }s^2>0.

Thus even if the assumed population distribution is exactly normal, the conditional estimates are less dispersed than the latent effects. This derivation assumes a scalar normal prior, a linear noisy summary, known variances, and independent normal error. It illustrates the mechanism; it is not an exact shrinkage formula for a general NLME model.

Worked Example: The Information in a Subject's Data Changes the EBE

For a hypothetical subject, suppose the scalar model above has \(\omega=0.500\), \(s=0.600\), and observed summary \(z=0.800\). The conditional mean, which is also the conditional mode in this normal model, is:

E(\eta\mid z) = \frac{0.500^2}{0.500^2+0.600^2}(0.800) = \frac{0.250}{0.610}(0.800) = 0.3279

The estimate is \(0.4098\) times \(z\), rather than \(z\) itself, because the data summary is uncertain relative to the population distribution. The conditional variance is:

\operatorname{Var}(\eta\mid z) = \frac{0.250(0.360)}{0.610} = 0.1475, \qquad \operatorname{SD}(\eta\mid z)=\sqrt{0.1475}=0.3841

This subject's conditional estimate is not the known true eta; the conditional standard deviation indicates substantial remaining uncertainty. Across subjects with this same \(s\), the conditional-mean variance is \(0.250^2/0.610=0.1025\), giving a standard deviation of \(0.3201\). Relative to the population standard deviation of \(0.500\), that spread corresponds to \(1-0.3201/0.500=0.3598\), or about 36.0% shrinkage. The calculation depends on the simplified model and does not imply that every subject in a real NLME analysis has the same information.

Reading Eta Histograms and Density Plots

An eta plot is useful when read alongside the model assumptions and the degree of shrinkage. A normal curve with mean zero and variance \(\omega_{jj}\) can provide a visual reference for the assumed marginal distribution of component \(j\). However, an EBE histogram compared with that curve is not a direct goodness-of-fit test for the latent distribution: the histogram contains conditional estimates, while the curve describes the assumed population random effects. A mismatch can reflect shrinkage, unequal subject-level information, limited sample size, model misspecification, or features of the random-effect distribution.

Similarly, an EBE histogram that looks roughly normal does not confirm the normality assumption. Shrinkage can compress estimates toward zero and mask tail behavior or asymmetry. Different subjects can have different amounts of information, so their EBEs need not be contracted equally. In nonlinear models, the mapping from random effects to observations can further affect the conditional estimates. A visually narrow distribution may therefore coexist with a substantially wider estimated population distribution.

In practice, inspect the eta plots together with the estimated \(\Omega\), component-wise shrinkage, the number of subjects, and evidence about how informative their observations are. A marked mean shift, asymmetry, extreme estimates, or a discrepancy from the model-based reference can flag a question for further investigation; none identifies its cause on its own. Check whether conclusions change across plausible model specifications. If distributional shape is consequential, simulation under the fitted model followed by re-estimation can help show how the same design and estimation process would shape EBE plots when the assumed model is true. It does not turn a plot into proof of the assumption.

Worked Example: A Compressed, Asymmetric EBE Density

Suppose a hypothetical analysis estimates an eta population standard deviation of \(0.250\), while the sample standard deviation of its EBEs is \(0.180\). The density plot is concentrated near zero and has a modest right tail. Calculate shrinkage and state what can and cannot be concluded.

The descriptive shrinkage is:

1-\frac{0.180}{0.250}=0.280=28.0\%

The plot's EBE spread is therefore 72.0% of the estimated population spread. Its right tail is a feature of the conditional estimates, but it does not by itself demonstrate that the population eta distribution is skewed. The apparent shape could be influenced by which subjects have informative data, by the finite number of subjects, or by model fit as well as by the latent distribution.

A cautious interpretation is: “The EBE density is compressed relative to the fitted population scale and has a modest right tail; with 28.0% estimated shrinkage, the plot alone does not establish the population distribution's shape.” The next checks are to review individual information and model fit, and to evaluate whether the distributional conclusion is stable under an appropriate model-based assessment.

Comparison, Diagnostics, and Common Pitfalls

Approach	What it can show	Main limitation
Histogram or density of EBEs	Unexpected features in the subject-specific conditional estimates	Estimates may be compressed and unequally informative; their shape is not the latent population shape itself
EBE spread compared with \(\sqrt{\omega_{jj}}\)	A component-wise summary of estimated shrinkage	Does not measure an individual's uncertainty or test distributional normality
Model-based likelihood assessment	Assesses the population distribution through the fitted model and observed data	Depends on model assumptions, adequate estimation, and the information in the data
Simulation followed by re-estimation	Shows how the assumed model and study design can produce EBE plots and shrinkage	Results depend on the simulated model; agreement does not prove that model is true

Treat eta plots as prompts for investigation, not verdicts. Warning signs include a very narrow EBE distribution relative to \(\Omega\), conclusions about tails based on few subjects, or an interpretation that treats conditional estimates as observed random effects. A high shrinkage value means the EBE spread is substantially reduced by this summary measure; it does not establish that the model is wrong. Conversely, low shrinkage does not guarantee that the assumed distribution is correct.

A common mistake is to say, “The EBEs look nearly normal, so the population random effects must be normal,” or, “The EBEs are compressed, so the estimated \(\Omega\) must be too small.” Both statements confuse different objects. EBEs are data-informed conditional estimates; \(\Omega\) describes the model's population distribution. The correction is to report the EBE plot as a conditional-estimate diagnostic, quantify shrinkage against the estimated population scale, and assess distributional assumptions using the fitted model and suitable model-based checks.

Key takeaway: An eta histogram or density plot displays empirical Bayes estimates, not a direct sample of latent population random effects. Quantify shrinkage, examine the available subject-level information, and avoid using the apparent EBE shape alone to confirm or reject the population distribution assumption.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A component has estimated population standard deviation \(0.400\) and EBE sample standard deviation \(0.300\). What is its descriptive eta shrinkage?
Why can a normal population random-effect distribution produce a narrower distribution of EBEs?
If an EBE density is asymmetric when shrinkage is substantial, what conclusion about the latent population distribution is justified?
In the scalar normal example, what happens to \(E(\eta\mid z)\) as the data-summary variance \(s^2\) becomes very large?
Does low estimated eta shrinkage prove that the population random-effect distribution is normal? Explain.

Solutions. 1. \(1-0.300/0.400=0.250\), or 25.0% shrinkage. 2. Each EBE is conditional on subject data and is pulled toward the population mean when those data are uncertain; conditional estimates can therefore have less spread than the latent effects. 3. The plot alone does not establish the latent distribution's shape; the asymmetry may reflect conditional estimation, unequal information, finite sample size, or model fit. 4. Since \(E(\eta\mid z)=\{\omega^2/(\omega^2+s^2)\}z\), the multiplier approaches zero, so the estimate approaches the population mean of zero for fixed \(z\). 5. No. Low shrinkage indicates less compression by the spread-based measure, but it does not confirm the distributional assumption.