Why an EBE Regression Can Mislead
Empirical Bayes estimates (EBEs) are estimates of individual random effects, not direct measurements of those effects. In particular, an EBE is informed both by an individual's observations and by the population distribution. When the individual data are weak, the estimate is pulled toward the population mean. A regression of EBEs on a covariate therefore relates that covariate to shrunk estimates, not to the unobserved individual effects themselves.
This distinction matters in covariate screening. A small EBE regression slope does not establish that a covariate effect is small: shrinkage can reduce the slope. Conversely, an EBE trend does not establish that the covariate has an independent effect on the structural parameter. It may reflect an omitted covariate correlated with the one being examined, or other model misspecification. The Gaussian calculation below makes the attenuation mechanism explicit; the confounding example shows how an apparently clear signal can arise for the wrong covariate.
A Gaussian Derivation of Attenuation
Consider a deliberately simple model with \(m\) observations for individual \(i\). Suppose the fitted model has a random intercept \(\eta_i\), with \(\eta_i\sim N(0,\omega^2)\), and observation errors \(\epsilon_{ij}\sim N(0,\sigma^2)\) that are mutually independent and independent of \(\eta_i\). The model is \(y_{ij}=\eta_i+\epsilon_{ij}\). Assume for this derivation that \(\omega^2\) and \(\sigma^2\) are known and that every individual has the same \(m\). These assumptions give a closed-form result; nonlinear NLME EBEs need not have this exact form.
The conditional mean is also the posterior mode in this Gaussian model. The factor \(\lambda\) lies between 0 and 1: as observation error increases, or as the number of observations decreases, \(\lambda\) gets smaller and the EBE is pulled more strongly toward zero. The key point for covariate regressions is that this pull applies to systematic patterns in the observations as well as to individual departures.
Suppose the true data-generating process has a covariate effect that the fitted model omits: \(y_{ij}=\beta x_i+u_i+\epsilon_{ij}\), where \(u_i\sim N(0,\omega^2)\) is independent of \(x_i\). Under the fitted random-intercept model, the EBE is \(\hat\eta_i=\lambda\bar y_i\). Conditional on \(x_i\), its expectation is \(\lambda\beta x_i\), since \(E(u_i\mid x_i)=0\) and \(E(\epsilon_{ij}\mid x_i)=0\). Thus, in this balanced Gaussian illustration, the expected slope from regressing the EBE on \(x_i\) is \(\lambda\beta\), not \(\beta\). This derivation depends on the assumptions just stated; it is not a general correction formula for nonlinear models or unequal information.
Worked Example: Calculating an Attenuated EBE Slope
Suppose a hypothetical PK analysis uses a standardized covariate \(x_i\), and the true log-clearance signal is \(\beta x_i\), with \(\beta=0.60\) log-clearance units per one-unit increase in \(x_i\). The baseline fitted model omits \(x_i\) and estimates an individual random intercept. For this illustration, take \(\omega^2=0.25\), \(\sigma^2=1.00\), and \(m=4\) observations per individual. Assume the Gaussian model and independence conditions in the derivation hold, and treat the variance parameters as known.
The variance of an individual's sample mean due to observation error is \(\sigma^2/m=1.00/4=0.25\). Therefore the shrinkage factor is \(\lambda=0.25/(0.25+0.25)=0.50\).
For an individual with \(x_i=1\), the expected sample mean is \(0.60\), so the expected EBE is \(0.50(0.60)=0.30\). For \(x_i=-1\), the expected sample mean is \(-0.60\), so the expected EBE is \(0.50(-0.60)=-0.30\). The expected EBE slope is \((0.30-(-0.30))/(1-(-1))=0.30\) log-clearance units per covariate unit.
The underlying signal has slope \(0.60\), but the EBE regression has expected slope \(0.30\) in this example. The EBE regression is attenuated by a factor of \(0.50\). Dividing a real EBE slope by an estimated shrinkage factor is not a generally valid remedy: the derivation relies on a simple, balanced Gaussian model, and actual NLME EBEs may have different information levels and nonlinear dependence on the data.
How Confounding Creates a Misleading Signal
Attenuation is only one concern. An EBE trend against a covariate can also arise because that covariate is associated with an omitted cause of the individual data pattern. In that case, the EBE may capture the omitted signal. The regression can look orderly and still fail to identify an independent effect of the examined covariate.
Worked Example: A Correlated Omitted Covariate
Consider a hypothetical analysis of log clearance. Let \(X_i\) be a standardized covariate under investigation and \(Z_i\) a second standardized covariate. Suppose the true individual signal is \(0.50Z_i+u_i\), with no direct \(X_i\) effect; \(u_i\) has mean zero and is independent of \(X_i\) and \(Z_i\). Suppose further that \(E(Z_i\mid X_i)=0.80X_i\). The baseline model omits both covariates and estimates individual random effects. For a simple illustration, assume its EBE behaves like a Gaussian shrinkage estimate with a common factor \(\lambda=0.60\), and assume its observation errors have conditional mean zero.
Because the fitted model omits \(Z_i\), its individual data contain a systematic signal \(0.50Z_i\) that can be absorbed into the random-effect estimate. Conditional on \(X_i\), the expected EBE is \(0.60\{0.50E(Z_i\mid X_i)+E(u_i\mid X_i)\}\). Independence and the zero mean of \(u_i\) give \(E(u_i\mid X_i)=0\), so \(E(\hat\eta_i\mid X_i)=0.60(0.50)(0.80X_i)=0.24X_i\).
The expected EBE regression slope is therefore \(0.24\) log-clearance units per one-unit increase in standardized \(X_i\), even though the true direct \(X_i\) effect is zero. The pattern is induced by the association between \(X_i\) and the omitted \(Z_i\), together with the model's assignment of unexplained individual signal to the random effect. Shrinkage reduces the size of this particular signal, but does not make it an independent \(X_i\) effect. A fitted joint model that represents relevant covariates and their relationship to the parameter is needed to assess such an effect.
This example is a confounding argument, not a claim that shrinkage alone creates a nonzero population slope whenever a covariate is unrelated to the latent effects. Under a correctly specified symmetric model, with random effects conditionally mean-zero and no systematic covariate-related error, varying precision alone need not create a nonzero expected slope. It does, however, change EBE reliability across individuals and can make a realized slope unstable. In finite samples, chance patterns, misspecification, or covariate-related data quality can further complicate the interpretation.
Unequal Information Changes What the Regression Represents
In practice, individuals may have different numbers or timing of observations, different residual variability, or different sensitivity of their profiles to a parameter. Their EBEs then need not share one shrinkage factor. In the Gaussian illustration, individual \(i\) has \(\lambda_i=\omega^2/(\omega^2+\sigma^2/m_i)\) when the number of observations is \(m_i\). In a nonlinear model, information depends more broadly on the design, parameter values, residual model, and observed profile.
When \(\lambda_i\) differs, an EBE regression compares estimates with different degrees of pull toward the population mean. If information also varies with the covariate, the apparent pattern can be distorted. The following calculation shows that unequal shrinkage can change both the slope and the intercept, even when a covariate-related signal is present in the individual data.
Worked Example: A Covariate-Related Difference With Unequal Shrinkage
Suppose a hypothetical, centered covariate takes values \(x=-1\) and \(x=1\), and the underlying omitted individual signal is \(0.50x\) log-clearance units. Individuals with \(x=-1\) have relatively informative profiles with \(\lambda=0.80\); those with \(x=1\) have less informative profiles with \(\lambda=0.20\). Assume the simple shrinkage approximation applies within each group and that random deviations have conditional mean zero.
The expected EBE for \(x=-1\) is \(0.80(0.50)(-1)=-0.40\). For \(x=1\), it is \(0.20(0.50)(1)=0.10\). The slope through these two expected group means is \((0.10-(-0.40))/(1-(-1))=0.25\) log-clearance units per covariate unit. The intercept at \(x=0\) is their midpoint, \((0.10+(-0.40))/2=-0.15\) log-clearance units, despite the underlying signal being centered at zero across these two covariate values.
The slope is not simply the underlying \(0.50\) signal times one common shrinkage factor, and the intercept reflects the unequal attenuation. This calculation does not prove that unequal information always biases an EBE slope in a particular direction. It demonstrates why a single shrinkage percentage is not enough to interpret a regression when individual information varies systematically.
Diagnostics, Alternatives, and Limitations
An EBE plot can be useful for generating hypotheses, but it should be read alongside information and model diagnostics. Before assigning meaning to a trend, examine whether the individuals contributing to different covariate ranges have comparable profile information and whether the fitted structural and residual models describe their data adequately. A trend concentrated among individuals with sparse or uninformative profiles deserves particular caution. So does a pattern that changes materially when influential individuals or plausible model assumptions are examined.
The central diagnostic question is not just whether a regression line is visible, but what data and model features produced the estimated EBEs. Check covariate relationships with design and observation count, examine whether relevant covariates were omitted from the population model, and compare EBE patterns with the fitted model's predictions and residual behavior. These checks can identify warning signs, but they do not turn a second-stage EBE regression into a valid substitute for fitting the covariate effect in the NLME model.
| Approach | What it uses | Main interpretation | Key limitation |
|---|---|---|---|
| Regression of EBEs on covariates | Estimated individual random effects from a fitted model | Exploratory association between the covariate and the shrunk estimates | Attenuation, unequal reliability, and omitted-signal confounding can mislead |
| Joint NLME covariate model | Individual observations and a covariate function in the population model | Estimates the covariate effect while accounting for the random-effects model | Still depends on model specification, identifiability, and adequate data |
| Plots and summaries of profiles or residuals | Observed data and model-based fit diagnostics | Helps locate patterns, poor fit, and data limitations | Does not by itself estimate or validate a covariate effect |
The joint NLME approach is the main alternative for estimating a covariate effect: specify the covariate's role in the population parameter model and fit it using the individual observations and random-effects distribution. This lets the model distinguish, to the extent supported by the data, a systematic population relationship from individual deviations. It does not guarantee an unbiased or precise result if the model is wrong or the effect is weakly identified, but it targets the covariate effect directly rather than treating EBEs as observed outcomes.
A common mistake is to read a near-zero EBE slope as evidence that the covariate has little effect, or to interpret a nonzero slope as proof of an independent effect. The correction is to recognize that EBEs are shrunk estimates whose precision may differ, and that omitted correlated predictors can leave covariate-related signal in them. Use EBE regressions as exploratory diagnostics, not as a final covariate-effect estimate or a stand-alone inferential test. In particular, ordinary regression standard errors and p-values do not automatically account for the estimation, shrinkage, and shared population-model uncertainty in the EBEs.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the Gaussian random-intercept example, what happens to the shrinkage factor when the number of observations per individual increases, with the variance parameters fixed?
- A fitted baseline model gives an EBE regression slope of \(0.18\) per covariate unit. Does that result alone establish that the covariate has a direct effect? Give two reasons.
- In the confounding example, identify the true direct effect of \(X_i\) and the expected EBE regression slope against \(X_i\).
- Why can individuals with different profile information make an EBE regression harder to interpret?
- What is the primary alternative when the goal is to estimate a covariate effect on a population parameter?
Solutions. (1) Since \(\lambda=\omega^2/(\omega^2+\sigma^2/m)\), increasing \(m\) decreases \(\sigma^2/m\) and increases \(\lambda\) toward 1. (2) No. Shrinkage may attenuate a real effect, and an omitted covariate correlated with the tested covariate may create an EBE trend; differing EBE reliability and model misspecification are additional concerns. (3) The true direct \(X_i\) effect is zero, while the expected EBE slope is \(0.60(0.50)(0.80)=0.24\) log-clearance units per standardized covariate unit. (4) Their EBEs are pulled toward the population mean by different amounts, so the regression combines estimates with unequal reliability and may be distorted if information relates to the covariate. (5) Fit the covariate in the population parameter model using the joint NLME model and the individual observations.