Why Quantify Variability Explained by a Covariate?
A covariate can explain why subjects differ in a population PK/PD parameter. For example, body weight may be associated with a subject’s typical drug-response parameter. Adding that relationship can reduce the random-effect variance left to describe differences among subjects with the same covariate value. Quantifying this change helps describe what the covariate adds to the model.
The comparison requires care. A smaller estimated random-effect variance is a model-based description of less unexplained variability on a specified scale. It is not automatically the proportion of all observed response variability explained, evidence that the covariate causes the parameter to change, or proof that the effect matters clinically. Those conclusions require additional evidence and interpretation.
What the Variance Comparison Measures
For a positive PD parameter \(\theta_i\), a common model is
Here \(\theta_{\mathrm{typ}}\) is the typical parameter at the covariate reference value, \(x_i\) is the subject’s covariate value as coded in the model, \(\beta\) is its fixed-effect coefficient, and \(\eta_i\) is the subject-specific random effect. The coefficient describes a shift in the modeled typical parameter across covariate values. The variance \(\omega^2\) describes the residual spread of individual parameters around those covariate-specific typical values, on the log scale.
To quantify the change, fit a model without the covariate and a model with it, then compare their estimates of the same random-effect variance. If the estimates are \(\hat{\omega}^2_{\mathrm{without}}\) and \(\hat{\omega}^2_{\mathrm{with}}\), the absolute reduction is their difference, and the relative reduction is
This is a descriptive comparison of estimated variances, not a separate test of the covariate effect. It is meaningful only when both estimates refer to the same parameter, scale, data, random-effects structure, residual model, and compatible estimation procedures. A value below zero means the estimated variance increased after adding the covariate; it should not be forced to zero.
Why Covariate Adjustment Can Reduce Random-Effect Variance
The distinction can be seen by considering the log parameter across subjects. Under the model above, suppose \(x_i\) varies across the population and is independent of \(\eta_i\). The law of total variance gives
The first term is variation in the modeled mean across covariate values; the second is average variation remaining among subjects conditional on the covariate. If the covariate relationship is real and adequately modeled, including it can assign some systematic differences to the first term rather than leaving them in the random-effect term. The reduction in \(\omega^2\) is not generally equal to \(\beta^2\operatorname{Var}(x_i)\) in a fitted PK/PD analysis: estimation, nonlinear relationships, other random effects, and model misspecification can all affect the estimates.
This justification relies on a correctly specified covariate relationship, a suitable random-effect distribution, and an appropriate description of dependence between covariates and random effects. Independence is a useful simplifying assumption for the variance decomposition; if it does not hold, the covariance contributes to the marginal variance. A fitted reduction therefore needs to be interpreted alongside the model assumptions and estimation quality.
Worked Example: Calculating a Reduction in Random-Effect Variance
Suppose a hypothetical analysis estimates the random-effect variance for a log-scale response parameter as \(0.16\) without a supported covariate and \(0.09\) with it. Assume both models use the same subjects, parameter scale, random-effects structure, residual model, and estimation method, and both estimates are sufficiently stable for a descriptive comparison.
The absolute reduction in variance is
The relative reduction is
Thus, the covariate model has an estimated log-scale random-effect variance 43.75% lower than the model without the covariate. This does not mean that 43.75% of total response variability, or 43.75% of all clinical variability, is explained.
The corresponding random-effect standard deviations are \(\sqrt{0.16}=0.4000\) and \(\sqrt{0.09}=0.3000\). Their relative reduction is
Variance and standard deviation do not have the same percentage reduction. If a lognormal interpretation is appropriate, the corresponding coefficients of variation are \(\sqrt{\exp(\omega^2)-1}\): about 41.65% for \(\omega^2=0.16\) and 30.69% for \(\omega^2=0.09\), a relative decrease of about 26.33%. These are different summaries on the original parameter scale and should be labeled as such.
Variance Reduction and the Covariate Effect Are Different Quantities
The estimated coefficient \(\beta\) describes the modeled covariate-associated shift in the typical parameter. The change in \(\omega^2\) describes how the estimated residual random-effect spread changes between models. A covariate can have a notable estimated effect but reduce \(\omega^2\) only slightly, or it can reduce the estimate substantially while its effect remains too uncertain or too small to matter clinically.
Worked Example: Interpreting a Weight Effect Separately
Consider the hypothetical model
where \(\mathrm{WT}_i\) is weight in kilograms and \(\theta_{\mathrm{typ},70}\) is the typical parameter at 70 kg. For a subject at 80 kg compared with the 70 kg reference, the covariate contrast is \((80-70)/10=1\). Holding the random effect at the same value for both subjects, the modeled parameter ratio is
The model therefore predicts a parameter about \(100(1-0.8025)=19.75\%\) lower at 80 kg than at 70 kg, conditional on the same random effect. This is an estimated covariate contrast, not a statement that weight causes the change.
Suppose the estimated random-effect variance for this parameter changes from \(0.12\) in the model without weight to \(0.08\) in the model with weight, with all other comparison conditions held constant. The relative variance reduction is
The 19.75% parameter contrast and 33.33% variance reduction answer different questions. Neither number alone establishes whether the effect changes a clinically relevant PD outcome; that judgment depends on the parameter’s role in the response model and the clinical context.
Assumptions, Comparability, and Diagnostics
Before interpreting a variance reduction, check that the comparison is coherent and that the estimates are credible. The following conditions are important:
- Same target and scale. Compare the variance for the same parameter and the same random-effect scale. A log-scale variance cannot be directly compared with a variance estimated on the natural scale.
- Comparable model specification. Keep the observations, other covariates, random-effects structure, residual-error model, and estimation approach aligned. If several features change at once, the difference cannot be attributed specifically to the covariate addition.
- Supported covariate relationship. The functional form and coding must be biologically plausible and supported by the observed data. A poorly specified relationship can leave systematic variation in the random effect or produce an unstable reduction.
- Reasonable random-effect assumptions. The assumed distribution and dependence structure should be defensible. A single normal random effect may be inadequate if the population is strongly heterogeneous in a way the model cannot represent.
- Stable estimation. Check convergence, parameter boundaries, and sensitivity to starting values or reasonable modeling choices. An estimate at or near zero, or an imprecise estimate, makes a percentage reduction especially fragile.
A large reduction can be driven by a small or unstable denominator, so report the two variance estimates as well as the percentage. If the no-covariate estimate is \(0.01\) and the covariate-model estimate is \(0.005\), the relative reduction is 50%, but the absolute change is only \(0.005\). Conversely, an apparently small percentage may correspond to a meaningful absolute change when the starting variance is large. Estimates and uncertainty should be considered together.
Useful checks include examining whether the fitted covariate relationship is sensible over the observed range, whether the variance estimate is stable, and whether the comparison models differ only in the intended covariate term. Population and individual prediction checks and covariate-stratified visual predictive checks can provide complementary evidence about model adequacy. A favorable variance comparison does not replace those checks, and an unfavorable diagnostic may indicate that the model form or error assumptions need attention.
What the Reduction Does—and Does Not—Establish
A random-effect variance reduction is not automatically a measure of clinical importance. Clinical interpretation should consider the size and uncertainty of the covariate effect, the range of covariate values represented, the parameter’s influence on predicted response, and whether the implied differences could alter a clinical decision. A statistically supported covariate effect may have little clinical consequence, while a clinically relevant effect may remain uncertain if the available data are limited.
Nor is the reduction a causal effect. In an observational data set, a covariate may be associated with other features of the subjects or their treatment. The fitted relationship describes the model’s conditional association; it does not by itself identify what would happen if the covariate were changed. The earlier material on separating covariate effects from random effects and on collinearity and confounding is relevant when assessing this interpretation.
Variance reduction also depends on the model and population being studied. A covariate that explains some heterogeneity in one data set may be less informative in another population with a narrower covariate range. A reduction in \(\omega^2\) is not the same as improved precision for every individual prediction, and it does not guarantee that the fitted model captures all important sources of variability.
Comparison With Other Summaries
| Summary | What it describes | Key caution |
|---|---|---|
| Change in \(\omega^2\) | Change in estimated conditional random-effect variance on the model’s parameter scale | Not the percentage of total observed response variability explained |
| Change in \(\omega\) | Change in random-effect standard deviation | Its percentage change differs from the variance percentage change |
| Covariate coefficient \(\beta\) | Modeled shift in the typical parameter per covariate contrast | Does not quantify remaining between-subject variability or establish causality |
| Clinical effect on a PD outcome | Implication of the covariate contrast for a response or decision of interest | Requires interpretation through the PD model and clinical context |
These summaries complement rather than replace one another. A likelihood-ratio test or another model-comparison method addresses evidence for the added model term under its assumptions; the variance reduction describes a change in a fitted variance estimate. Neither should be reported as though it answered the other question.
Common Pitfall
Correction: The 40% refers only to the relative change in the estimated conditional random-effect variance, on the specified scale and between the compared models. Total variability also depends on the covariate distribution and other sources of variation, and clinical importance must be evaluated from the modeled effect, uncertainty, PD consequences, and clinical context.
Another frequent error is to report a percentage reduction in variance while calculating it from standard deviations. Use the variance estimates in the variance formula; if standard deviations or original-scale coefficients of variation are also reported, label and calculate each separately.
A Practical Reporting Sequence
Name the PD parameter, the random-effect scale, and the covariate model being compared.
Report \(\hat{\omega}^2_{\mathrm{without}}\) and \(\hat{\omega}^2_{\mathrm{with}}\), including units or scale where appropriate, before giving a percentage.
Give the absolute and relative variance changes, and state that they describe estimated conditional variability rather than total response variability.
Describe its estimated parameter contrast and uncertainty, then assess the clinical implications and model adequacy without treating variance reduction as proof of importance or causation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A model’s estimated log-scale random-effect variance decreases from \(0.25\) to \(0.15\) after adding a covariate. What are the absolute and relative reductions?
- Why is a 36% reduction in variance not the same as a 36% reduction in random-effect standard deviation?
- A covariate model has a lower estimated \(\omega^2\), but its coefficient is very uncertain. What can and cannot be concluded from the variance comparison alone?
- Name two conditions needed for a meaningful comparison of random-effect variance estimates from models with and without a covariate.
- In a log-linear model, what is the modeled parameter ratio for a covariate contrast of one unit when \(\beta=0.10\)?
Solutions. (1) The absolute reduction is \(0.25-0.15=0.10\). The relative reduction is \((0.10/0.25)\times100\%=40.00\%\). (2) Standard deviation is the square root of variance: the estimates are \(0.5000\) and \(\sqrt{0.15}=0.3873\), so the SD reduction is about \((0.5000-0.3873)/0.5000\times100\%=22.54\%\), not 40%. (3) One can report a lower estimated conditional variance in that fitted comparison, assuming it is stable and comparable. The reduction alone does not establish a precise covariate effect, clinical importance, or causality. (4) For example, use the same parameter and scale and keep the data, other model components, and estimation approach comparable. (5) The ratio is \(\exp(0.10)=1.105\), so the modeled parameter is about 10.52% higher for that contrast, conditional on the same random effect.