Selecting Covariates for Between-Subject Variability

Two Different Questions About Covariates

A covariate can help explain why the typical value of a pharmacokinetic or pharmacodynamic parameter differs across the population. It can also help explain why subjects differ from one another around that typical value. These are distinct hypotheses: a covariate that shifts the population typical value need not explain between-subject variability, and a covariate associated with variability need not shift the typical value.

This distinction matters for model building and interpretation. A covariate effect on a typical value describes a systematic change in the parameter’s population location. An effect on random-effect variance describes a change in the spread of subject-specific deviations. Variance effects are often harder to estimate because they are inferred indirectly from the distribution of individual deviations, which themselves are estimated with uncertainty.

For clarity, consider one positive parameter \(\theta\), such as clearance. Let \(x_i\) be a centered and scaled covariate for subject \(i\), with \(x_i=0\) at a chosen reference value. A typical-value model is

Here \(\theta_0\) is the typical value at the reference covariate value, and \(\beta\) describes how the typical value changes with \(x_i\). In contrast, a variance model can leave the typical value unchanged and allow the random-effect variance to depend on the covariate:

The exponential form ensures a positive variance. At \(x_i=0\), the variance is \(\omega^2\); for a one-unit increase in \(x_i\), it is multiplied by \(\exp(\gamma)\). The corresponding standard deviation is multiplied by \(\exp(\gamma/2)\), not by \(\exp(\gamma)\).

Modeling and Interpreting Variance Effects

The random effect remains centered at zero in this variance model. Thus, the covariate changes how widely individual effects are dispersed around the typical parameter, not the expected value of \(\eta_i\) on the log scale. Subjects at the same covariate value still vary: the model assigns them a distribution, not a deterministic parameter value.

A useful derivation makes the scale clear. At covariate value \(x\), the variance is \(\omega^2(x)=\omega^2\exp(\gamma x)\). Comparing it with the reference variance gives

Taking square roots gives the standard-deviation ratio, \(\omega(x)/\omega(0)=\exp(\gamma x/2)\). For example, \(\gamma<0\) means that the random effects are less dispersed at larger \(x\), while \(\gamma>0\) means greater dispersion. Neither sign alone says that the typical parameter rises or falls.

A variance covariate model should be considered only when the data and scientific question support it. The covariate should have adequate representation across the study population, and its relationship with design features and other candidate covariates should be considered. The structural model, residual-error model, and random-effects distribution must also be plausible. Otherwise, unexplained model misspecification can appear as a covariate-related difference in random-effect spread.

Testing a Covariate Effect on Variability

For a single continuous covariate in the variance model, the null hypothesis of no variance effect is \(H_0:\gamma=0\), against the alternative \(H_a:\gamma\ne 0\). A likelihood-ratio test compares the model with \(\gamma\) estimated to the nested model with \(\gamma=0\). Under regular conditions, the statistic \(2(\ell_{\mathrm{full}}-\ell_{\mathrm{reduced}})\) is approximately chi-square with one degree of freedom. Those conditions include a correctly specified likelihood, comparable fits to the same data, and estimable parameters away from problematic boundaries. The baseline random-effect variance should be estimable and positive; if it is effectively zero, the variance covariate may be unidentified and the usual reference distribution may not apply.

A Wald test instead divides the estimate of \(\gamma\) by its standard error and compares the resulting statistic with a standard normal distribution for a two-sided test. The Wald approach is convenient, but can be unreliable when the likelihood is asymmetric, the variance effect is weakly identified, or the estimate is near a boundary-related problem. Likelihood-based comparisons are often preferable when the model fits stably, but they are not automatically reliable if the underlying estimation or assumptions fail.

Estimation Challenges, Diagnostics, and Alternatives

A typical-value covariate effect is often informed by systematic differences in the observations across covariate values. A variance effect is supported by differences in the amount of subject-to-subject spread across those values. This second signal is harder to distinguish from individual-level estimation error, residual variability, and misspecification. Many observations from a small number of subjects do not necessarily provide strong information about a population variance relationship: the number and distribution of subjects are central.

Before interpreting a fitted variance effect, check whether the model converged, whether estimates and standard errors are finite and stable, and whether the estimated variance remains positive throughout the observed covariate range. Examine uncertainty and the fitted distribution of individual deviations across covariate values. Compare simulated or predicted variability with observed patterns where appropriate. A steep variance trend driven by a few subjects or by covariate values with little support is a warning, not persuasive evidence by itself.

Check also whether the proposed variance trend is compensating for another problem. An incorrect structural model, residual-error model, or distributional assumption can create systematic residual patterns that appear to be differences in between-subject variability. Study design can confound a covariate with dose, sampling schedule, or another factor that affects the information available per subject. Collinearity can likewise make it difficult to assign a variability pattern to one covariate. These concerns are especially important when some covariate regions are represented by few subjects.

A common mistake is to say, “The negative variance coefficient means older subjects have lower typical clearance.” That conclusion confuses a spread parameter with a location parameter. The correction is to interpret \(\gamma\) through the variance ratio \(\exp(\gamma x)\) and, separately, interpret \(\beta\) through the typical-value ratio \(\exp(\beta x)\). A second pitfall is selecting a variance covariate simply because an empirical Bayes estimate plot looks more or less dispersed across groups. Such estimates are uncertain and may be shrunk toward zero; use them as exploratory evidence, then assess the covariate in the population model.

The main alternative is to model a covariate effect on the typical parameter value, when the scientific question concerns systematic differences in that parameter. A model can also include both location and variance effects if there is a reason to consider both and the data support estimating them. These models answer different questions, so one should not be substituted for another merely because one yields a smaller p-value. Variance models add flexibility but also demand stronger information about the spread of random effects and can be sensitive to sparse support or misspecification. Statistical evidence should therefore be considered alongside plausibility, precision, and the covariate’s observed support.