What an Eta-Zero Individual Represents

What Does “Eta-Zero” Mean?

In a nonlinear mixed-effects model, a subject’s parameter can differ from the population parameter because of a subject-level random effect. Setting that random effect to zero gives a useful reference value. But the phrase “eta-zero subject” does not, by itself, mean “the average subject.” Its interpretation depends on how the random effect enters the parameter equation and on the distribution assumed for the random effect.

Here, \(\theta_{\mathrm{pop}}\) denotes a population parameter, \(\theta_i\) the corresponding parameter for subject \(i\), and \(\eta_i\) a subject-level random effect. A random effect equal to zero means that subject \(i\) has no deviation from the model’s reference value on the scale where \(\eta_i\) is defined.

This definition is about a value in the model, not necessarily a particular person in the data. A fitted model can use the eta-zero value to describe its reference prediction even if no observed subject has exactly \(\eta_i=0\). Whether the value corresponds to a central subject in the population depends on the random-effect distribution.

Eta-Zero in an Additive Parameter Model

In an additive parameter model, a subject’s parameter is the population parameter plus a subject-specific deviation:

If \(\eta_i=0\), then \(\theta_i=\theta_{\mathrm{pop}}\). Thus, the eta-zero subject has the population parameter value on the natural parameter scale. For example, if the parameter is clearance in L/h, the additive deviation is also expressed in L/h.

Suppose the random effects have mean zero, so \(E(\eta_i)=0\), and the parameter values have a finite mean. Taking expectations in the equation gives:

Under these conditions, the eta-zero value equals the population arithmetic mean in the additive model. This result depends on the additive form and the zero-mean assumption. It does not imply that every individual’s parameter equals \(\theta_{\mathrm{pop}}\), nor that a finite sample’s average must equal it exactly.

An additive model also has a practical constraint: if \(\theta_i\) must be positive, the model and random-effect distribution must make negative values impossible or sufficiently implausible for the intended use. The eta-zero interpretation itself does not remove this modeling concern.

Eta-Zero in an Exponential Parameter Model

For a positive parameter, a common alternative is to let the random effect act multiplicatively:

With \(\eta_i=0\), \(\exp(0)=1\), so the eta-zero parameter is \(\theta_{\mathrm{pop}}\). A positive eta multiplies the population value by a factor greater than 1; a negative eta multiplies it by a factor between 0 and 1. This parameterization keeps \(\theta_i\) positive whenever \(\theta_{\mathrm{pop}}>0\).

The important distinction is the scale of the deviation. In the additive model, \(\eta_i\) is added on the parameter scale. In the exponential model, \(\eta_i\) is exponentiated and acts as a multiplicative factor. Even if \(E(\eta_i)=0\), it does not follow that \(E(\exp(\eta_i))=1\). Consequently, in general:

The eta-zero value in the exponential model is therefore not generally the population arithmetic mean on the natural parameter scale. Under a symmetric distribution for \(\eta_i\), however, zero is its center, and the increasing transformation \(\exp(\eta_i)\) makes \(\theta_{\mathrm{pop}}\) the corresponding central value on the parameter scale. A central value and an arithmetic mean are different summaries.

Covariates and the Eta-Zero Reference Value

A covariate can shift the reference value before the random effect is applied. For example, a log-linear parameter model may be written:

Here, \(x_i\) is the subject’s covariate value, \(x_{\mathrm{ref}}\) is the reference covariate value, and \(\beta\) is the coefficient for the covariate. Setting \(\eta_i=0\) gives \(\theta_{\mathrm{pop}}\exp\{\beta(x_i-x_{\mathrm{ref}})\}\), not necessarily \(\theta_{\mathrm{pop}}\). The eta-zero value is conditional on the subject’s covariate value. It is the covariate-adjusted reference parameter for that value of \(x_i\).

Why the Distinction Matters

The model’s eta-zero value is useful because it provides a clear reference for interpreting individual deviations and generating predictions. But calling it “the average subject” without qualification can lead to incorrect summaries, comparisons, and explanations of model output.

The difference follows from the scale transformation. In an additive model, a zero-mean random effect passes directly through the expectation because addition preserves means. In an exponential model, the random effect is transformed first. An average of transformed values need not equal the transformation evaluated at the average random effect.

If \(E(\eta_i)=0\), the right-hand side is \(1\), but the left-hand side is not generally \(1\). In the hypothetical three-subject example, it is \((\exp(-0.4)+1+\exp(0.4))/3\), approximately \(1.054\). This is enough to show why a zero-mean random effect does not imply that the natural-scale arithmetic mean equals the eta-zero value. The precise population mean requires specifying the random-effect distribution and calculating the expectation under that distribution.

Assumptions, Diagnostics, and Common Mistakes

Interpretations of zero require attention to the model’s assumptions. A mean-zero random effect is a statement about its distribution, not a guarantee that the observed estimates for a finite set of subjects sum to zero. A symmetric distribution provides a central zero on the eta scale; if the distribution is not symmetric, zero need not be its median or other central point. The parameterization also matters: an additive effect and an exponential effect have different implications on the natural scale.

When reviewing a fitted model, check that the equation, scale, and covariate reference values support the intended interpretation. If a positive parameter is modeled additively, inspect whether the implied individual values can be negative. For an exponential model, verify that the reference parameter is interpreted at the stated covariate values, and do not infer the natural-scale arithmetic mean from the eta-zero value alone. Distributions of estimated individual random effects can also be examined for substantial asymmetry or patterns, while recognizing that estimated effects are not the same as directly observed population random effects.

• Mistake: “The eta-zero subject is the average subject.” This is not generally true. State the parameterization and the scale on which “average” is meant.
• Mistake: “The random effects average to zero, so exponentiating them still gives factors that average to one.” Exponentiation is nonlinear. For example, the three factors \(\exp(-0.4)\), \(1\), and \(\exp(0.4)\) average to about \(1.054\), not \(1\).
• Mistake: “The population parameter is always the eta-zero value, even with covariates.” In a covariate model, the eta-zero value depends on the covariate values unless they are at the reference values.
• Mistake: “An eta-zero person must appear in the data.” Eta-zero is a model-based reference value. An observed subject need not have exactly zero random effect.

The main alternative to describing an eta-zero individual is to report a population summary on the scale relevant to the question, such as an arithmetic mean or another clearly defined central summary. These alternatives answer different questions. The eta-zero value describes the model prediction at a specified random effect and covariate value; a population summary describes a feature of the distribution of individual parameters. Choosing one does not make the other incorrect, but they should not be treated as interchangeable.