Why Joint Sampling Matters
Human predictions combine inputs such as clearance, bioavailability, body size, enzyme abundance, and pharmacodynamic potency. These inputs are not necessarily independent. For example, two physiological quantities may share a common size-related driver, or model parameters estimated from the same data may have correlated uncertainty. If each input is sampled independently despite known dependence, simulated combinations can be too extreme, too narrow, or biologically implausible. The resulting exposure and response distributions can therefore be misleading even when every input has a reasonable marginal distribution.
The earlier material on parameter covariance in translational predictions establishes why dependence belongs in uncertainty calculations. Here the focus is practical: how to generate joint samples that preserve it and how to see its consequences in human predictions. A joint sample is one coherent row of inputs; the model is evaluated on each row, not on independently rearranged columns.
Sampling From a Multivariate Normal Model
A standard construction begins with a multivariate normal distribution. If \(X\sim N(\mu,\Sigma)\), with mean vector \(\mu\) and positive-definite covariance matrix \(\Sigma\), find a matrix \(L\) such that \(\Sigma=LL^\mathsf{T}\). Draw a vector \(Z\) of independent standard normal variables and set \(X=\mu+LZ\). The resulting vector has mean \(\mu\) and covariance \(\Sigma\). A covariance matrix must be symmetric and positive semidefinite; a nonsingular Cholesky factorization requires it to be positive definite.
This construction is appropriate when a multivariate normal distribution is a defensible joint model for the quantities being sampled. Many pharmacokinetic and physiological inputs are positive, however, so a normal distribution on the original scale may generate impossible negative values. One option is to work with logarithms. If \(Y=\log X\) is multivariate normal, then \(X\) has positive lognormal marginals, with dependence induced through \(Y\). In that case, specify whether the covariance matrix is on the log scale or the original scale; these are not interchangeable.
Worked Example: Correlated Log-Scale Clearance and Volume
Suppose a hypothetical human prediction uses clearance \(CL\) and volume \(V\). Their log-scale uncertainty is represented by \(\log(CL)=\log(10)+0.25Z_1\) and \(\log(V)=\log(50)+0.20Z_2\), where \(Z_1,Z_2\) are standard normal with correlation \(0.60\). Thus the medians are \(10\) L/h and \(50\) L, and the log-scale standard deviations are \(0.25\) and \(0.20\). To illustrate one joint draw, take independent standard normal values \(e_1=0.40\) and \(e_2=-1.20\).
A Cholesky factor for the latent correlation matrix is \(\begin{pmatrix}1&0\\0.60&0.80\end{pmatrix}\). Therefore \(Z_1=e_1=0.40\) and \(Z_2=0.60e_1+0.80e_2=0.24-0.96=-0.72\). Transforming to the model inputs gives \(CL=10\exp(0.25\times0.40)=10\exp(0.10)=11.05\) L/h and \(V=50\exp(0.20\times(-0.72))=50\exp(-0.144)=43.29\) L.
For a hypothetical 100 mg intravenous dose, the corresponding exposure measure \(AUC=D/CL\) is \(100\text{ mg}/11.05\text{ L/h}=9.048\text{ mg}\cdot\text{h/L}\). This is one paired draw: the sampled \(CL\) and \(V\) arose from the same latent vector. Sampling each in isolation and pairing values arbitrarily would not preserve the specified dependence.
Copulas for Specified Marginal Distributions
A copula separates the choice of marginal distributions from the model of dependence. A Gaussian copula starts with correlated standard normal variables \(Z\), transforms each to a uniform variable \(U_i=\Phi(Z_i)\), and then applies the desired inverse marginal distribution \(X_i=F_i^{-1}(U_i)\). Here \(\Phi\) is the standard normal cumulative distribution function and \(F_i\) is the chosen marginal cumulative distribution function. The uniform variables are marginally uniform on \((0,1)\); the inverse transforms produce the requested marginals.
The matrix \(R\) describes dependence on the latent normal scale. After nonlinear transformations, the Pearson correlations among the \(X_i\) will generally differ from the entries of \(R\). The copula preserves the specified marginal distributions and a dependence structure, not a promise that transformed variables have exactly the same Pearson correlations as the latent normals. If a target correlation on the original scale is important, calibrate the latent dependence and check the resulting simulated correlation.
Worked Example: A Gaussian Copula for Clearance and Bioavailability
Consider a hypothetical model with a lognormal clearance marginal having median \(8\) L/h and log-scale standard deviation \(0.30\), and a bioavailability marginal uniform between \(0.40\) and \(0.80\). For this worked draw, the latent normal correlation is \(0.40\). Use independent standard normal inputs \(e_1=0.50\) and \(e_2\approx0.3974\), and set \(Z_{CL}=e_1\) and \(Z_F=0.40e_1+\sqrt{1-0.40^2}\,e_2\). Then \(Z_F\approx0.5642\).
Transforming the latent values gives \(U_{CL}=\Phi(0.50)\approx0.6915\) and \(U_F=\Phi(0.5642)\approx0.7137\). For the specified marginals, \(CL=8\exp(0.30\times0.50)=8\exp(0.15)\approx9.295\) L/h and \(F=0.40+0.40U_F\approx0.40+0.40(0.71369)=0.6855\).
For a hypothetical 100 mg oral dose, the model \(AUC=DF/CL\) gives \(AUC=100(0.6855)/9.295\approx7.375\text{ mg}\cdot\text{h/L}\). The same uniform values would be transformed differently if different marginal distributions were selected. The calculation also illustrates why marginal checks matter: \(F\) remains within its specified bounds, while \(CL\) remains positive.
How Dependence Changes a Prediction
Dependence matters when the prediction combines multiple inputs. For a simple ratio, its effect can be shown directly. Suppose \(\log(V)\) and \(\log(CL)\) are jointly normal, with standard deviations \(\sigma_V\) and \(\sigma_{CL}\) and correlation \(\rho\). Since the terminal half-life is \(t_{1/2}=(\log 2)V/CL\), its logarithm is a constant plus \(\log(V)-\log(CL)\). The variance of a difference includes a covariance term:
The minus sign is important. Positive correlation between \(V\) and \(CL\) can reduce uncertainty in their ratio because high values of one tend to occur with high values of the other. Conversely, for a product, the log-scale variance has a plus sign on the covariance term, so positive correlation can increase uncertainty. The direction depends on how inputs enter the prediction, not on a rule that correlation always widens or narrows prediction distributions.
Worked Example: Correlation and the Half-Life Distribution
Assume, hypothetically, that \(\log(V)\) and \(\log(CL)\) are jointly normal with medians \(50\) L and \(10\) L/h, log-scale standard deviations \(0.20\) and \(0.25\), and correlation \(0.70\). The median half-life is \((\log 2)(50/10)=3.466\) h. The log-scale variance of the half-life is \(0.20^2+0.25^2-2(0.70)(0.20)(0.25)=0.0400+0.0625-0.0700=0.0325\), so its log-scale standard deviation is \(\sqrt{0.0325}=0.1803\).
Because the log half-life is normal, a central 90% interval under this input-uncertainty model is the median multiplied and divided by \(\exp(1.6449\times0.1803)=1.345\). The interval is approximately \(3.466/1.345=2.576\) h to \(3.466(1.345)=4.662\) h.
If the same marginals were instead sampled independently, the log-scale variance would be \(0.20^2+0.25^2=0.1025\), with standard deviation \(0.3202\). The corresponding interval is approximately \(3.466/\exp(1.6449\times0.3202)=2.047\) h to \(3.466\exp(1.6449\times0.3202)=5.868\) h. In this example, ignoring positive correlation substantially widens the half-life distribution. These are intervals implied by the stated input model, not confidence intervals for an estimated parameter.
Implementation, Diagnostics, and Alternatives
Choose the marginal distributions and the dependence representation using available parameter estimates, posterior draws, physiological knowledge, or an explicit assumption. State the scale on which covariance or correlation is defined.
Use a covariance factorization for a multivariate normal model, or latent correlated normals followed by marginal inverse transforms for a Gaussian copula. Each row must retain the generated pairings.
Compare simulated marginal summaries with their targets, inspect scatterplots and rank relationships, check bounds and positivity, and verify that the covariance matrix is symmetric and positive semidefinite.
Run the human prediction model for each row, then summarize the resulting exposure or response distribution. Compare with independent sampling as a diagnostic contrast, not as a default substitute for dependence.
A common practical failure is a covariance matrix that is not positive semidefinite. This can arise from rounding pairwise correlations or combining estimates from incompatible sources. A Cholesky failure is a warning to investigate, not a reason to silently replace the matrix. Also inspect the marginal draws and the transformed correlations: the intended latent correlation can look quite different after nonlinear transformations.
A multivariate normal model is attractive when uncertainty is reasonably symmetric on the chosen scale and covariance estimates are available. A Gaussian copula is more flexible about marginals, but still represents dependence through latent normal variables. It may be a poor description when dependence is strongly nonlinear or includes important joint tail behavior. If relevant posterior or bootstrap draws are available, resampling whole parameter vectors can preserve their empirical dependence without imposing a normal copula. Independent sampling is simpler, but is defensible only when independence is supported or used deliberately as a scenario.
The approach has limits. Estimated correlations can be imprecise when data are sparse; a plausible covariance matrix does not establish that the chosen marginal distributions are correct; and a copula cannot compensate for an inadequate human prediction model. In particular, uncertainty in the joint inputs is distinct from structural uncertainty in the cross-species model. A well-specified sampling scheme makes the uncertainty conditional on its assumptions visible, but does not remove those assumptions.
Common Pitfall
The incorrect reasoning is, “Clearance has the right distribution and volume has the right distribution, so their joint samples are adequate.” That conclusion does not follow: the same marginals can be paired with positive, negative, or zero dependence, and those alternatives can give different distributions for functions such as \(V/CL\) or \(F/CL\). The correction is to specify and justify the dependence structure, propagate it through the prediction model, and report conclusions as conditional on that joint-input assumption.
Check Your Understanding
Answer each question using the stated assumptions; solutions follow the questions.
- If two log-scale inputs have standard deviations \(0.10\) and \(0.20\) and correlation \(0.50\), what is the variance of their log-scale sum?
- Why does a Gaussian copula not generally make the Pearson correlation of transformed inputs equal to the latent normal correlation?
- For a prediction proportional to \(V/CL\), what covariance term appears in the variance of its logarithm?
- What should be investigated if a proposed covariance matrix fails Cholesky factorization?
Solutions. (1) The variance is \(0.10^2+0.20^2+2(0.50)(0.10)(0.20)=0.0100+0.0400+0.0200=0.0700\). (2) Nonlinear inverse marginal transformations preserve the chosen marginals but generally alter Pearson correlation; the resulting transformed correlation should be checked or calibrated if it is the target. (3) Since \(\log(V/CL)=\log(V)-\log(CL)\), the covariance contribution is \(-2\operatorname{Cov}\{\log(V),\log(CL)\}\). (4) Check symmetry, positive semidefiniteness, rounding, and whether the entries came from compatible estimates. Do not silently alter the matrix without documenting and justifying the change.