Separating Animal Variability From Species Effects

Why the Two Sources of Variation Matter

Suppose a preclinical PK parameter differs across animals given the same compound. Some differences may reflect variation among animals of one species; others may arise because species have different typical values. Those sources have different implications for translation. A wide spread among animals in one species is not, by itself, evidence that species differ. Conversely, a consistent shift between species should not be hidden by pooling all animals into one undifferentiated variance.

A mixed-effects model represents observations as a combination of systematic predictors and random deviations at specified grouping levels. Here, animals are nested within species: each animal belongs to one species. The model can therefore represent a species-level deviation shared by animals of that species and an animal-level deviation that differs among animals within that species. The resulting variance components describe distinct levels of variation, provided the design and assumptions support their separation.

A Nested Model and Its Assumptions

Let \(y_{sj}\) be the outcome for animal \(j\) in species \(s\). It could be a PK parameter or a transformed parameter such as log clearance; the chosen scale must be stated because variance components depend on scale. A useful starting model is

Here \(x_s\) is a species-level predictor when one is used; without it, set \(\beta_1=0\). The fixed part, \(\beta_0+\beta_1x_s\), describes the expected trend across species. The random species effect \(u_s\) represents departure from that trend, and the animal effect \(a_{sj}\) represents departure from its species-specific expected value. The usual model assumes the random effects are mutually independent across species and animals, independent of the predictors, and normally distributed with mean zero and constant variances at their respective levels.

This model treats species as sampled from a population of species, an exchangeability assumption that may be scientifically strong when only a few named species are available. It also assumes the fixed trend is adequate and that animals are independent conditional on the model. If the data are repeated measurements rather than one animal-level summary per animal, their within-animal correlation needs an appropriate structure. With only one recorded value per animal, unexplained animal-level variation and measurement error cannot generally be separated; \(\sigma_A^2\) then describes their combined contribution to the observed animal-level outcome.

When the aim is to compare a prespecified set of species, species can instead be represented by fixed effects, with animal deviations nested within species. That estimates contrasts among those species, but does not by itself estimate a distribution of effects for a new species. A random species effect is useful for prediction to an unobserved species only when the exchangeability assumption is defensible.

What the Variance Components Mean

Conditional on the predictors, the variance of a single animal outcome is the sum of the species and animal variances:

Two animals from the same species share \(u_s\), so their model-implied covariance is \(\tau_S^2\), and their correlation is

This correlation is often called the intraclass correlation. It is the fraction of the total marginal variance in individual outcomes attributable to differences between species under this model. It is not the fraction of uncertainty in every possible human prediction: the answer depends on whether the target is a species mean or an individual, and on uncertainty in the fitted parameters.

Why the Balanced-Design Calculation Works

The preceding calculation follows from the expected mean squares under the model, with no covariate and the same number \(m\) of animals in every species. Within each species, subtracting its sample mean removes the shared species effect. The remaining animal deviations have variance \(\sigma_A^2\), so the expected within-species mean square is \(\sigma_A^2\). Between species, the variance of a species mean includes the species variance plus the variance of an average of \(m\) independent animal deviations, which is \(\sigma_A^2/m\). Multiplying by \(m\) gives the between-species expected mean square:

Subtracting the first expectation from the second and dividing by \(m\) yields \(\tau_S^2\). Replacing expectations by observed mean squares gives the method-of-moments estimates used above. This justification depends on a balanced design and the stated random-intercept model. In unbalanced studies or models with predictors, variance components are ordinarily estimated by mixed-model methods rather than by this simple formula. A moment estimate can even be negative when \(MS_{\text{between}}<MS_{\text{within}}\); a negative variance is impossible, and constrained estimation handles the boundary at zero.

From Variance Components to Human Predictions

Prediction uncertainty must be defined for a target. Suppose a fitted cross-species trend predicts a human species mean on the modelled scale as \(\hat{\mu}_H=\hat{\beta}_0+\hat{\beta}_1x_H\). If human is treated as a new species with an unobserved random effect, the prediction for its population mean includes the between-species variance \(\tau_S^2\). A prediction for one human individual additionally includes the within-species animal variance \(\sigma_A^2\).

For a simple illustration, suppose the fixed-effect estimates have covariance matrix with entries \(\operatorname{Var}(\hat{\beta}_0)\), \(\operatorname{Var}(\hat{\beta}_1)\), and \(\operatorname{Cov}(\hat{\beta}_0,\hat{\beta}_1)\). The estimated variance contribution from fixed-effect estimation at \(x_H\) is

Ignoring uncertainty in the estimated variance components themselves, the approximate prediction variances are \(V_{\text{fixed}}(x_H)+\tau_S^2\) for a new human species mean and \(V_{\text{fixed}}(x_H)+\tau_S^2+\sigma_A^2\) for an individual human. These expressions assume the new random effects are independent of the fitted fixed effects. If the target is a species mean conditional on a well-supported human-specific effect estimated from human data, the prediction problem differs; the new-species term is not added in the same way. An interval for the mean trend alone reflects fixed-effect estimation, not the variability of a new species or a new individual.

Diagnostics, Alternatives, and Limitations

The model is useful only if its grouping structure and assumptions are credible. Examine fitted species means and animal-level residuals, and look for a species with a systematic pattern that the fixed trend fails to capture. Compare the estimated components with the observed spread at each level. When species counts are small, the between-species variance may be weakly identified even if each species has many animals: adding animals mainly improves estimates of within-species variation, while adding independent species is what informs the species-level distribution.

Other warning signs include strong non-normality, unequal spreads across species, influential species, and animal observations that are dependent for unmodelled reasons. Variance estimates at zero can indicate little evidence for a component in these data, but may also reflect limited information or a misspecified structure. A large estimated intraclass correlation is not proof that species caused the differences; it is a model-based description of clustering after the included predictors.

A fixed-species model is often preferable when inference concerns contrasts among a small, specific set of species and no generalization to a new species is intended. A random-species model is more directly suited to describing a distribution across exchangeable species and predicting a new one, but relies on that exchangeability assumption. A pooled model with no species effect is simpler, yet can understate uncertainty for animals from a new species and obscure systematic species differences. None of these formulations resolves structural uncertainty in the translation relationship; the variance components quantify variation represented by the chosen model, not every source of prediction error.