Interpreting Clearance–Volume Random-Effect Correlation in Population PK

What Clearance–Volume Correlation Describes

A population PK model can allow clearance and volume to vary between individuals. When both parameters have random effects, their joint distribution describes not only how much each varies, but also whether individuals with above-typical clearance tend to have above-typical or below-typical volume. This pairing can matter when interpreting the range of individual parameter combinations supported by a model.

The correlation usually refers to the random effects on the model’s parameter scale, commonly the log scale for positive PK parameters. It is not a statement that clearance causes volume to change, nor does a correlation of a given size prescribe an exact volume for an individual with a known clearance.

Definition: In a lognormal random-effects model, individual clearance and volume may be written as \(CL_i=TVCL\exp(\eta_{CL,i})\) and \(V_i=TVV\exp(\eta_{V,i})\). Here \(TVCL\) and \(TVV\) are typical population values, and \(\eta_{CL,i}\) and \(\eta_{V,i}\) are the individual random effects. The clearance–volume random-effect correlation is \(\rho_{CL,V}=\operatorname{Corr}(\eta_{CL,i},\eta_{V,i})\). It describes the linear association between the two log-scale deviations across individuals under the model.

For the usual interpretation, assume that the random-effect pair has a bivariate normal distribution with mean zero and covariance matrix

\begin{pmatrix} \eta_{CL,i}\\ \eta_{V,i} \end{pmatrix} \sim N\left[ \begin{pmatrix}0\\0\end{pmatrix}, \begin{pmatrix} \sigma_{CL}^{2} & \rho_{CL,V}\sigma_{CL}\sigma_V\\ \rho_{CL,V}\sigma_{CL}\sigma_V & \sigma_V^{2} \end{pmatrix} \right].

The standard deviations \(\sigma_{CL}\) and \(\sigma_V\) describe the between-individual spread of the log-scale effects. The off-diagonal covariance is \(\rho_{CL,V}\sigma_{CL}\sigma_V\), whereas the correlation is dimensionless and lies between \(-1\) and \(1\). A positive correlation means that the deviations tend to have the same sign; a negative correlation means they tend to have opposite signs. Neither sign means that every individual follows that pattern.

Interpreting the Sign and Magnitude

Because the exponential function is increasing, a positive clearance random effect means clearance is above its typical value, and a negative effect means it is below its typical value. The same sign interpretation applies to volume. Thus, a positive correlation makes high-clearance/high-volume and low-clearance/low-volume combinations more prevalent than would be expected under zero correlation. A negative correlation makes high-clearance/low-volume and low-clearance/high-volume combinations more prevalent.

These are tendencies in a population distribution, not rules for each person. Even a strong positive correlation permits individuals with high clearance and low volume; such combinations are simply less likely under the specified bivariate normal model than same-direction combinations. A correlation near zero indicates little linear association between the two random effects. Under the bivariate normal assumption, zero correlation implies independence of the random effects. Without joint normality, zero correlation alone does not establish independence.

Correlation must also be distinguished from covariance. If the random-effect standard deviations change, the covariance changes even when the correlation is unchanged. Conversely, covariance values cannot be compared directly across parameterizations or scales without considering their units and the associated variances.

Worked Example: Positive Correlation and Paired Deviations

Suppose a hypothetical model has \(\sigma_{CL}=0.25\), \(\sigma_V=0.30\), and \(\rho_{CL,V}=0.60\). For an individual with \(\eta_{CL}=0.20\), calculate the clearance multiplier and describe the conditional distribution of the volume effect.

Clearance. The individual clearance relative to \(TVCL\) is \(\exp(0.20)=1.221\). This individual’s clearance is therefore about 22.1% above the typical value.

Conditional volume effect. For a bivariate normal pair, the conditional distribution is

\eta_V\mid\eta_{CL}=x \sim N\left( \rho_{CL,V}\frac{\sigma_V}{\sigma_{CL}}x,\; \sigma_V^2(1-\rho_{CL,V}^2) \right).

Substituting the values gives conditional mean \(0.60(0.30/0.25)(0.20)=0.144\) and conditional variance \(0.30^2(1-0.60^2)=0.0576\), so the conditional standard deviation is \(0.240\). The conditional median volume multiplier is \(\exp(0.144)=1.155\). Thus the model associates this positive clearance deviation with a positive expected log-volume deviation, but the conditional spread is substantial.

The conditional mean volume multiplier is not \(\exp(0.144)\). Since the conditional random effect is normal, its exponential has mean \(\exp(0.144+0.0576/2)=\exp(0.1728)=1.189\). This distinction matters: exponentiating a normal mean gives the median multiplier, while the mean multiplier includes the variance term.

Why the Conditional Distribution Gives the Interpretation

The conditional normal result makes the meaning of the correlation more precise. Given a clearance effect \(x\), the conditional mean of the volume effect is proportional to \(x\), with slope \(\rho_{CL,V}\sigma_V/\sigma_{CL}\). The conditional variance is \(\sigma_V^2(1-\rho_{CL,V}^2)\). These formulas follow from the standard conditional distribution of a bivariate normal vector.

The conditional mean has the same sign as \(x\) when the correlation is positive and the opposite sign when it is negative. At zero correlation, the conditional mean is zero regardless of \(x\). The conditional variance decreases as the absolute correlation approaches one: under the exact bivariate normal model, a correlation of \(1\) or \(-1\) leaves no conditional variation in one random effect once the other is known. Such boundary cases are special and should not be treated as ordinary moderate correlations.

The slope is not the correlation itself. It also depends on the relative standard deviations. For example, a given correlation can correspond to a steeper change in the conditional mean of \(\eta_V\) when \(\sigma_V/\sigma_{CL}\) is larger. Correlation describes standardized association; the conditional slope describes the expected change in the volume effect per unit change in the clearance effect.

Worked Example: Negative Correlation

Consider a hypothetical model with \(\sigma_{CL}=0.20\), \(\sigma_V=0.40\), and \(\rho_{CL,V}=-0.75\). For an individual with \(\eta_{CL}=0.10\), find the conditional mean and standard deviation of \(\eta_V\), then interpret the associated parameter multipliers.

The clearance multiplier is \(\exp(0.10)=1.105\), so clearance is about 10.5% above \(TVCL\). The conditional mean volume effect is \( -0.75(0.40/0.20)(0.10)=-0.150\). The conditional variance is \(0.40^2(1-(-0.75)^2)=0.16(0.4375)=0.0700\), giving conditional standard deviation \(\sqrt{0.0700}=0.2646\).

The conditional median volume multiplier is \(\exp(-0.150)=0.861\), below the typical value. The conditional mean multiplier is \(\exp(-0.150+0.0700/2)=\exp(-0.115)=0.891\). In this model, above-typical clearance tends to accompany below-typical volume, but the sizeable conditional standard deviation shows that the association does not determine an individual’s volume.

Correlation on the Log Scale and on the Parameter Scale

The model correlation is defined for \(\eta_{CL}\) and \(\eta_V\), not directly for \(CL_i\) and \(V_i\). With positive typical values and lognormal parameters, however, the sign of the correlation is preserved by exponentiation. The numerical value generally changes.

This can be seen from the lognormal moment formulas. For jointly normal random effects, \(E[\exp(\eta_{CL})]=\exp(\sigma_{CL}^2/2)\), and likewise for volume. The covariance of the exponentiated effects is

\operatorname{Cov}(CL_i,V_i) =TVCL\,TVV\exp\left(\frac{\sigma_{CL}^2+\sigma_V^2}{2}\right) \left[\exp(\rho_{CL,V}\sigma_{CL}\sigma_V)-1\right].

Dividing this covariance by the two parameter-scale standard deviations gives

\operatorname{Corr}(CL_i,V_i)= \frac{\exp(\rho_{CL,V}\sigma_{CL}\sigma_V)-1} {\sqrt{\left[\exp(\sigma_{CL}^2)-1\right] \left[\exp(\sigma_V^2)-1\right]}}.

For nonzero standard deviations, the denominator is positive, and the numerator has the same sign as \(\rho_{CL,V}\). Therefore the signs agree, but the two correlation values need not be equal. The formula also shows why the scale matters: the parameter-scale correlation depends on both the random-effect correlation and the random-effect variances.

Worked Example: Near-Zero Correlation

Suppose a hypothetical model has \(\sigma_{CL}=0.30\), \(\sigma_V=0.20\), and \(\rho_{CL,V}=0.05\). The covariance is \(0.05(0.30)(0.20)=0.00300\). If an individual has \(\eta_{CL}=0.30\), the conditional mean volume effect is \(0.05(0.20/0.30)(0.30)=0.0100\).

The corresponding conditional median volume multiplier is \(\exp(0.0100)=1.010\), close to the typical value. The conditional variance is \(0.20^2(1-0.05^2)=0.0399\), with standard deviation about \(0.1997\). The individual’s clearance multiplier is \(\exp(0.30)=1.350\), but the small correlation provides little information about that individual’s volume effect: its conditional spread remains close to the marginal spread of \(0.20\) on the log scale.

A near-zero estimate does not prove that clearance and volume are unrelated in every possible sense. It indicates little estimated linear association between these random effects under this model; the bivariate normal assumption additionally gives independence when the correlation is exactly zero.

Interpreting an Estimated Correlation in Practice

A fitted correlation summarizes the model’s estimated joint random-effect distribution. It should be interpreted alongside the two estimated random-effect variances, the model structure, and the available data. A large magnitude can imply that certain combinations of individual clearances and volumes are relatively unusual under the model, but it does not establish a biological mechanism or a causal relationship.

Estimation can be difficult when the data contain limited information about both parameters for the same individuals. A correlation may be imprecise if one random-effect variance is small, if the design poorly distinguishes clearance from volume, or if the model’s structural and residual assumptions do not describe the data adequately. A point estimate alone does not show how precisely the correlation has been estimated. Check its uncertainty and the stability of the fit, and consider whether the estimated signs and magnitudes are sensitive to reasonable changes in model specification.

The main alternatives represent different assumptions, not simply different labels for the same estimate. A diagonal covariance structure fixes the correlation at zero. Estimating an off-diagonal covariance allows a nonzero correlation. A block-diagonal structure allows correlation within specified groups of random effects while setting cross-group covariances to zero. These choices should be guided by scientific plausibility and the information in the data; adding a correlation is not automatically an improvement.

Common Mistakes and Limitations

Reading correlation as a percentage change. A correlation of \(0.60\) does not mean that a 1% increase in clearance corresponds to a 0.60% increase in volume. It is a standardized association between random effects.
Treating a tendency as a rule. Positive correlation does not require every high-clearance individual to have high volume. Conditional distributions retain variation unless the correlation is exactly at a boundary under the bivariate normal model.
Confusing covariance and correlation. The covariance equals the correlation multiplied by both random-effect standard deviations. It is not dimensionless and is not interchangeable with the correlation.
Assuming zero correlation always means independence. That conclusion holds for a jointly normal pair, not for arbitrary joint distributions.
Interpreting the estimate causally. A random-effect correlation describes a modeled population association. It does not show that changing clearance would change volume.

The interpretation is also conditional on the chosen parameterization and distribution. A correlation on log clearance and log volume is not a complete description of all possible dependence patterns, and a misspecified joint distribution can make its summaries misleading. Correlation is useful for understanding paired variability, but it does not replace checks of model fit, uncertainty, and scientific plausibility.

Check Your Understanding

Answer each question using the lognormal, bivariate normal random-effects model described above. Solutions follow the questions.

If the clearance–volume random-effect correlation is negative, which paired deviations are more prevalent: high clearance with high volume, or high clearance with low volume?
Let \(\sigma_{CL}=0.20\), \(\sigma_V=0.30\), and \(\rho_{CL,V}=0.50\). For an individual with \(\eta_{CL}=0.10\), calculate the conditional mean of \(\eta_V\) and the conditional median volume multiplier.
Calculate the random-effect covariance when \(\sigma_{CL}=0.40\), \(\sigma_V=0.25\), and \(\rho_{CL,V}=-0.30\).
Under what assumption does exactly zero random-effect correlation imply independence? Does a correlation near zero require every individual’s clearance and volume deviations to be unrelated?

Solutions. (1) High clearance with low volume, and low clearance with high volume, are more prevalent under a negative correlation; neither pairing is mandatory for every individual. (2) The conditional mean is \(0.50(0.30/0.20)(0.10)=0.0750\). The conditional median volume multiplier is \(\exp(0.0750)=1.078\). (3) The covariance is \((-0.30)(0.40)(0.25)=-0.0300\). (4) Exactly zero correlation implies independence when the random-effect pair is jointly normal. A near-zero correlation indicates little linear association in the model, but does not make every individual’s paired deviations unrelated or rule out other forms of dependence.