Parameter Covariance in Translational Predictions

Why Parameter Covariance Matters

A translational PK prediction often depends on several parameters at once. For example, a concentration prediction after an intravenous dose depends on both clearance and volume of distribution. Uncertainty in each parameter matters, but so does whether their uncertain values tend to move together. Treating correlated parameters as independent can therefore widen or narrow a prediction interval incorrectly.

The mapping of PK parameters across species provides the predicted parameter values; this tutorial focuses on propagating their joint uncertainty into a prediction. The calculations below use a one-compartment model to make the role of covariance explicit. The same principle applies to other model outputs: the effect depends on how sensitive the output is to each parameter and on the joint uncertainty of those parameters.

Definition: Parameter covariance describes how two uncertain parameters vary together. For parameters \(X\) and \(Y\), their covariance is \(\operatorname{Cov}(X,Y)=E[(X-E[X])(Y-E[Y])]\). Positive covariance means values above their respective means tend to occur together; negative covariance means a value above one mean tends to accompany a value below the other. Covariance is scale-dependent, so the parameter scale must be stated.

Propagation on the Log Scale

Let \(CL\) denote clearance, \(V\) volume of distribution, \(D\) an intravenous bolus dose, and \(t\) time after dosing. Under a one-compartment model with first-order elimination, concentration is

C(t)=\frac{D}{V}\exp\left(-\frac{CL}{V}t\right).

Suppose uncertainty is represented for the log parameters \(x=\log(CL)\) and \(v=\log(V)\). This scale is convenient when parameter values are positive and uncertainty is naturally expressed as relative rather than additive changes. Write their covariance matrix as \(\Sigma\), with variances \(\sigma_x^2\) and \(\sigma_v^2\), and covariance \(\sigma_{xv}\). The correlation is \(\rho=\sigma_{xv}/(\sigma_x\sigma_v)\), provided both standard deviations are positive.

The log concentration is \(h(x,v)=\log C(t)=\log(D)-v-(CL/V)t\). Define \(a=CLt/V\) at the parameter values used for the prediction. Its sensitivity vector, the gradient of log concentration with respect to the log parameters, is

g= \begin{pmatrix} \partial h/\partial x\\ \partial h/\partial v \end{pmatrix} = \begin{pmatrix} -a\\ -1+a \end{pmatrix}.

A first-order Taylor approximation around the nominal log parameters gives the approximate variance of log concentration. This is the parameter-uncertainty calculation used in the examples.

Approximation: If the log parameters have covariance matrix \(\Sigma\), then the first-order propagated variance of log concentration is \(g^\mathsf{T}\Sigma g\). For two parameters, this is \(g_x^2\sigma_x^2+g_v^2\sigma_v^2+2g_xg_v\sigma_{xv}\). The last term is the covariance contribution; its sign depends on both the covariance and the sensitivities.

To justify the result, write the deviations from the nominal log parameters as \(\delta x\) and \(\delta v\). The first-order change in log concentration is \(g_x\delta x+g_v\delta v\). Taking its variance and expanding gives

\begin{aligned} \operatorname{Var}(g_x\delta x+g_v\delta v) &=g_x^2\operatorname{Var}(\delta x)+g_v^2\operatorname{Var}(\delta v) +2g_xg_v\operatorname{Cov}(\delta x,\delta v)\\ &=g_x^2\sigma_x^2+g_v^2\sigma_v^2+2g_xg_v\sigma_{xv} =g^\mathsf{T}\Sigma g. \end{aligned}

The approximation is most useful when uncertainty is moderate and the output is nearly linear over the plausible parameter range on the log scale. If log concentration is approximately normal with nominal value \(\log C_0\) and propagated standard error \(s_{\log C}\), an approximate central 90% interval for concentration is \(C_0\exp(\pm1.645s_{\log C})\). This interval describes uncertainty due to the parameters represented in \(\Sigma\); it is not automatically an interval for future individual concentrations.

Worked Predictions

Worked Example: Positive Covariance Increases Uncertainty

Consider a hypothetical human concentration prediction with nominal value \(C_0=3.0327\) mg/L. At the prediction time, suppose \(a=CLt/V=0.5\), so the log-concentration sensitivities to log clearance and log volume are both \(-0.5\). Suppose the hypothetical covariance matrix for these log parameters is

\Sigma= \begin{pmatrix} 0.0400&0.01800\\ 0.01800&0.0225 \end{pmatrix}.

The marginal standard deviations are \(0.2000\) and \(0.1500\), and the correlation is \(0.01800/(0.2000\times0.1500)=0.6000\). The matrix is admissible: its determinant is \(0.0400(0.0225)-(0.01800)^2=0.000576>0\), and its variances are positive.

With \(g_x=g_v=-0.5\), the propagated log variance is

$$ (-0.5000)^2(0.0400)+(-0.5000)^2(0.0225) +2(-0.5000)(-0.5000)(0.01800) =0.024625. $$

Thus \(s_{\log C}=\sqrt{0.024625}=0.1569\). Using the approximate 90% multiplier 1.645 gives \(1.645(0.1569)=0.2581\), and the concentration interval is \(3.0327\exp[\pm1.645(0.1569)]=(2.343,\ 3.926)\) mg/L.

If the same marginal variances were incorrectly treated as independent, the covariance term would be omitted. The variance would instead be \((0.5000)^2(0.0400)+(0.5000)^2(0.0225)=0.015625\), giving \(s_{\log C}=0.1250\). The resulting interval would be \(3.0327\exp[\pm1.645(0.1250)]=(2.469,\ 3.725)\) mg/L. Here the two sensitivities have the same sign, so positive covariance increases the prediction variance; independence understates it.

Worked Example: Negative Covariance Reduces Uncertainty

Keep the hypothetical nominal concentration, sensitivities, and marginal variances from the preceding calculation, but suppose the log-parameter covariance is \(-0.01800\), corresponding to correlation \(-0.6000\). The matrix remains admissible because its determinant is still \(0.000576>0\).

The covariance contribution now has the opposite sign:

$$ (-0.5000)^2(0.0400)+(-0.5000)^2(0.0225) +2(-0.5000)(-0.5000)(-0.01800) =0.006625. $$

Therefore \(s_{\log C}=\sqrt{0.006625}=0.08139\), and the approximate central 90% concentration interval is \(3.0327\exp[\pm1.645(0.08139)]=(2.653,\ 3.467)\) mg/L.

The negative covariance partly offsets the uncertainty contributions because the output responds in the same direction to changes in both log parameters, while the parameters tend to move in opposite directions. Ignoring that dependence would again give the independent interval (2.469, 3.725) mg/L, which is wider in this case. Covariance does not always inflate uncertainty; its effect is determined by \(g_xg_v\sigma_{xv}\).

Worked Example: The Prediction Target Changes the Covariance Effect

For a second hypothetical concentration prediction, let \(C_0=1.800\) mg/L and \(a=1.2\). Then \(g_x=-1.2\), whereas \(g_v=-1+1.2=0.2\). Use the positive covariance matrix from the first example. The diagonal variance contributions sum to

$$ (-1.2)^2(0.0400)+(0.2)^2(0.0225)=0.05850. $$

Now the sensitivities have opposite signs, so positive covariance contributes negatively: \(2(-1.2)(0.2)(0.01800)=-0.008640\). The total variance is \(0.05850-0.008640=0.04986\), and \(s_{\log C}=\sqrt{0.04986}=0.2233\). The approximate 90% interval is \(1.800\exp[\pm1.645(0.2233)]=(1.247,\ 2.599)\) mg/L.

Under an independence assumption, the variance would be \(0.05850\), with standard error \(\sqrt{0.05850}=0.2419\), producing \((1.209,\ 2.680)\) mg/L. Here positive covariance reduces uncertainty in this prediction. The contrast with the first example shows why the correlation alone does not determine whether an interval gets wider or narrower: the output's parameter sensitivities matter too.

Assumptions, Checks, and Alternatives

These calculations rely on a specified PK model and prediction target, a covariance matrix for the parameters on the same scale as the sensitivities, and a local first-order approximation. The approximate interval also uses a normal approximation for log concentration. Before interpreting it, check the following:

Check the scale. A covariance for \(CL\) and \(V\) on their natural scales is not interchangeable with a covariance for \(\log(CL)\) and \(\log(V)\). Transform the uncertainty representation consistently before combining it with log-scale sensitivities.
Check matrix validity. Variances must be nonnegative, and a two-parameter covariance matrix must satisfy \(\sigma_{xv}^2\leq\sigma_x^2\sigma_v^2\). A violation can produce a negative propagated variance and indicates inconsistent inputs or numerical error.
Check the covariance sign and sensitivities. The covariance contribution is \(2g_xg_v\sigma_{xv}\), not simply twice the covariance. A sign error can reverse whether dependence widens or narrows the interval.
Check local adequacy. Strong nonlinearity over the plausible parameter range, very large uncertainty, or an interval that behaves implausibly can make the first-order approximation unreliable. In such cases, assess the prediction with a more suitable uncertainty-propagation method rather than treating this approximation as exact.

Setting covariance to zero is a useful comparison only when independence is scientifically justified or when explicitly showing the consequence of that assumption. Marginal intervals alone do not specify joint uncertainty. A conservative sensitivity analysis can explore combinations of parameter values, but it does not generally have the same probability interpretation as a covariance-based interval. More elaborate propagation methods can accommodate nonlinear output transformations, but the central requirement remains: represent the joint parameter uncertainty rather than silently replacing it with independent inputs.

Finally, distinguish parameter uncertainty from variability among people. The covariance matrix here represents uncertainty in the parameters used to make a translational prediction. It does not by itself describe between-individual differences, residual unexplained variability, or future measurement error. Those sources require their own model terms and should not be folded into this parameter covariance without justification.

Common Mistake

A frequent error is to conclude that positive correlation always widens a prediction interval. That reasoning ignores the signs of the output sensitivities. The relevant contribution is \(2g_xg_v\sigma_{xv}\): positive covariance increases variance when the sensitivities have the same sign, but decreases it when they have opposite signs. Another error is to omit covariance while still describing the result as a joint uncertainty interval. That calculation is an independence scenario, not the propagation of the stated correlated inputs.

Key takeaway: Propagated uncertainty depends jointly on parameter variances, their covariance, and the prediction's sensitivities. Calculate \(g^\mathsf{T}\Sigma g\) for the stated parameter scale and target, then interpret the resulting interval as approximate parameter uncertainty under the model and assumptions used.

Check Your Understanding

Answer each question, then compare with the solutions below.

If two parameters have positive covariance and the prediction has sensitivities with opposite signs, is the covariance contribution positive or negative?
For sensitivities \(g_x=-1\), \(g_v=2\), variances \(0.09\) and \(0.04\), and covariance \(0.015\), calculate the first-order propagated variance.
What condition must a two-parameter covariance matrix satisfy to be admissible?
Does a parameter-uncertainty interval automatically describe the range of concentrations among future individuals? Explain.

Solutions. (1) Negative, because \(g_xg_v<0\) while covariance is positive. (2) \(1^2(0.09)+2^2(0.04)+2(-1)(2)(0.015)=0.09+0.16-0.06=0.190\). (3) Both variances must be nonnegative and \(\sigma_{xv}^2\leq\sigma_x^2\sigma_v^2\). (4) No. It represents uncertainty in the parameters included in the calculation; between-individual variability and other sources of variation need separate representation.