Reading Eta Distributions and EBE Plots Under Shrinkage

Why an EBE Plot Is Not a Direct Picture of the Eta Distribution

In an NLME model, the random effects describe population variability in individual parameters. An empirical Bayes estimate (EBE) is an individual-specific estimate of a random effect, conditional on that individual's data and the fitted population model. These are related quantities, but they are not interchangeable. In particular, an EBE plot is not simply a histogram of the latent random effects.

The distinction matters when interpreting apparent spread, skewness, tails, or clusters. With limited individual information, EBEs are pulled toward the population mean. A narrow plot can therefore reflect substantial shrinkage rather than genuinely small between-individual variability. The introductory material on conditional distributions and eta shrinkage establishes the conditional-estimation framework; here the focus is how that framework changes what EBE distributions look like.

For a scalar random effect, a common descriptive eta-shrinkage summary is \(1-s_{\hat{\eta}}/\sqrt{\hat{\omega}^2}\), where \(s_{\hat{\eta}}\) is the sample standard deviation of the EBEs and \(\hat{\omega}^2\) is the fitted population variance. This summary compares EBE spread with fitted latent spread. It is not a direct estimate of the fraction of each individual's effect that is known, and in finite samples it can be unstable. Its interpretation also depends on using the same random-effect scale and parameterization in both quantities.

How Shrinkage Compresses a Distribution

A Gaussian random-intercept model gives a transparent derivation. Suppose the data for individual \(i\) are \(y_{ij}=\mu+\eta_i+\epsilon_{ij}\), for \(j=1,\ldots,m\), with \(\eta_i\sim N(0,\omega^2)\), independent \(\epsilon_{ij}\sim N(0,\sigma^2)\), and independence across individuals. Here \(\mu,\omega^2,\sigma^2\) are treated as known. These assumptions make the conditional distribution exactly Gaussian. Define the individual mean residual \(r_i=\bar y_i-\mu\). Then \(r_i\mid\eta_i\sim N(\eta_i,\sigma^2/m)\).

The conditional density of \(\eta_i\) is proportional to its prior density times the likelihood of \(r_i\). Collecting the terms involving \(\eta_i\) in the exponent gives:

Thus the conditional mean and mode coincide, and the EBE is \(\hat{\eta}_i=\lambda r_i\), where \(\lambda=\omega^2/(\omega^2+\sigma^2/m)\). Since \(0<\lambda<1\), each observed residual mean is contracted toward zero. The posterior variance is \(\omega^2(1-\lambda)\). Across hypothetical repeated data sets generated under this model, the variance of the conditional means is \(\omega^2-\omega^2(1-\lambda)=\lambda\omega^2\), which is smaller than the latent variance. This makes explicit why an EBE distribution can be narrower even when the random-effects distribution is correctly specified.

In nonlinear models, conditional means or modes are generally not a single common linear contraction of the observations. The amount and direction of shrinkage can vary by individual, parameter, and region of the likelihood. Consequently, a distorted EBE shape is possible: tails may be suppressed, asymmetry may be muted or altered, and apparent clusters may become less distinct. The Gaussian derivation explains a mechanism, not a universal transformation that can be reversed by rescaling every EBE.

Reading Shape, Center, and Scale

The center of an EBE distribution is often near zero when the model has a zero-centered random-effect distribution and the data provide balanced information. But this is not a test of distributional fit. EBEs are constructed using the fitted population model, including its centering and variance assumptions. A plot centered at zero may simply reflect that conditional estimates are pulled toward the assumed population center. In finite samples, the EBE mean need not be exactly zero, and an offset can also arise from sampling variability, parameter estimation, model misspecification, or the way the random effects are parameterized.

The same caution applies to apparent normality. Shrinkage can reduce extreme EBEs, making a histogram look more compact and sometimes more normal than the latent effects. Conversely, uneven information, nonlinear likelihoods, or a small sample can make the EBE distribution look irregular even when a Gaussian latent distribution is plausible. A histogram or Q–Q plot of EBEs describes those estimates; it does not isolate the latent distribution from the conditional-estimation process.

Plots remain useful as descriptive diagnostics, provided their scope is stated correctly. Inspect the EBE histogram or density for gross irregularities, and use a Q–Q plot to display departures from a normal reference among the estimates. Read these alongside the fitted \(\Omega\), the amount of individual information, and shrinkage for the specific random effect. When shrinkage is substantial, tail behavior and apparent spread in the EBE plot deserve especially little weight as evidence about the latent distribution.

Diagnostics, Alternatives, and Limits

A responsible reading combines plots with information about how the estimates were obtained. First, confirm which quantity is plotted: conditional mode or mean, raw eta or an individual parameter transformed from eta, and which model parameterization and scale are used. Then compare EBE spread with the fitted latent SD or covariance structure, and consult eta-shrinkage summaries for that parameter. High shrinkage warns that the plotted shape may be a poor representation of the latent effects; low shrinkage makes the estimates more data-informed but does not make an EBE plot a formal goodness-of-fit test.

When distributional adequacy is important, a main alternative is a model-based check that simulates data under the fitted model and applies the same estimation and plotting procedure to those simulated data. Comparing observed EBE patterns with the simulated EBE patterns accounts, at least under the simulation assumptions, for shrinkage and finite-sample behavior. A complementary alternative is to examine the model-implied latent distribution directly, while remembering that it is imposed or estimated as part of the model rather than observed person by person. Neither approach removes the need to assess the assumptions and the adequacy of the overall model.

Typical warning signs include a mismatch between the EBE scale and the reported \(\Omega\) scale, extreme sensitivity of the histogram to bin width, a few influential EBEs dominating the apparent tails, and strong shrinkage paired with claims about latent skewness based only on the EBE plot. These are reasons to investigate, not automatic evidence of a particular alternative distribution. EBE plots also have limits in small samples: their shape is noisy, and estimated population parameters add uncertainty not reflected by a simple histogram or Q–Q plot.

The central practical rule is to separate two questions: “What do the conditional estimates look like?” and “What latent random-effect distribution is compatible with the data and model?” The first is answered descriptively by an EBE plot. The second requires interpretation in light of shrinkage and, where warranted, model-based checks. A visually narrow or centered EBE distribution alone answers neither the question of whether the random-effect variance is appropriate nor whether the assumed distribution fits well.