Why Uncertainty Must Travel Through the IVIVC
An IVIVC prediction is not determined by a single error term. Dissolution measurements are uncertain; the relationship translating dissolution into in vivo input has estimated parameters; and in vivo quantities such as the absorbed fraction or disposition may themselves be estimated. Each uncertainty source can affect predicted plasma exposure, and their effects can reinforce or offset one another. A point prediction alone does not show whether the uncertainty is small enough for the intended decision.
The aim here is to propagate uncertainty from those inputs to outcomes such as AUC and Cmax. The earlier material on convolution establishes how an absorption input and a disposition function combine to produce a concentration profile. The earlier material on internal and external predictability concerns predictive performance; here the emphasis is on uncertainty in a particular prediction, including which input sources contribute most.
A useful generic representation is \(Y=f(Z,\theta)\), where \(Y\) is a predicted exposure measure, \(Z\) collects uncertain inputs such as dissolution or estimated in vivo absorption, and \(\theta\) denotes IVIVC and disposition parameters. The function \(f\) may be a direct equation or a numerical convolution followed by calculation of AUC or Cmax. An interval is meaningful only relative to the uncertainty sources included in it.
First-Order Propagation and Its Conditions
When the prediction function is smooth near the estimated inputs and their uncertainty is not too large, a first-order Taylor expansion gives a convenient approximation. Let \(W\) collect all uncertain inputs and parameters, with estimate \(\hat W\), covariance matrix \(\Sigma\), and prediction \(g(W)\). Locally, \(g(W)\) is approximately \(g(\hat W)+\nabla g(\hat W)^\mathsf{T}(W-\hat W)\). Taking the variance of this linear expression gives the delta-method approximation.
The covariance matrix matters: uncertainty sources should not be treated as independent merely for convenience. For two uncertain quantities \(W_1,W_2\), the approximation contains the cross-term \(2(\partial g/\partial W_1)(\partial g/\partial W_2)\operatorname{Cov}(W_1,W_2)\). A negative covariance can reduce the propagated variance; a positive covariance can increase it.
This approximation assumes the model is locally differentiable, the covariance estimates are adequate, and a local linear approximation is reasonable over the uncertainty range. Approximate normal intervals additionally rely on the output distribution being sufficiently close to normal on the chosen scale. Exposure is positive and often right-skewed, so propagation on the log scale is frequently more suitable. If \(L=\log Y\) is approximately normal with mean estimate \(\hat L\) and standard error \(SE_L\), an approximate 95% interval is \(\exp(\hat L\pm1.96SE_L)\). It is asymmetric on the original scale.
Worked Example: Correlated IVIVC Parameter Uncertainty
Suppose a hypothetical Level A model predicts log AUC by \(\log(\mathrm{AUC})=\alpha+\beta x\), where \(x=\log(r/1)\) and \(r\) is the dissolution fraction at the specified time. At \(r=0.90\), take \(\hat\alpha=\log(10)\), \(\hat\beta=1.20\), \(SE_\alpha=0.06\), \(SE_\beta=0.15\), and \(\operatorname{Cov}(\hat\alpha,\hat\beta)=-0.0036\). AUC is measured in mg·h/L. These are hypothetical inputs; assume the parameter estimates have the stated covariance and that the log-linear model is adequate locally.
The prediction is \(10\exp\{1.20\log(0.90)\}\). Since \(\log(0.90)=-0.10536\), this is \(10\exp(-0.12643)=8.812\) mg·h/L. For the log prediction, the gradient with respect to \((\alpha,\beta)\) is \((1,x)\). Therefore,
The approximate 95% interval for the mean model prediction, conditional on the fixed dissolution value and excluding future-outcome variability, is \(8.812\exp\{\pm1.96(0.06788)\}\), or approximately (7.71, 10.07) mg·h/L. It is not a 95% probability statement about the fixed true AUC.
Suppose a new outcome also has an independent residual standard deviation of 0.08 on the log scale. Its variance must be added for a future-outcome prediction interval: \(0.0046084+0.08^2=0.0110084\), giving \(SE=\sqrt{0.0110084}=0.1049\). The approximate interval becomes \(8.812\exp\{\pm1.96(0.1049)\}\), or (7.18, 10.82) mg·h/L. This interval is wider because it includes both uncertainty in the estimated mean prediction and residual future-outcome variability. The covariance term is positive here because both \(x\) and the covariance are negative; omitting it would understate the variance.
Carrying Dissolution and In Vivo Input Uncertainty
An uncertainty calculation should follow the actual prediction chain. For example, dissolution may be transformed to an absorbed fraction, which is then combined with dose and clearance to obtain AUC. If a log-scale exposure model is \(L=\log(A_{\mathrm{ref}})+\gamma\log(D)+\eta\), then \(D\) is a measured dissolution fraction, \(\gamma\) is a model parameter, and \(\eta\) represents uncertainty in an in vivo input or other specified log-scale component. The derivatives show how changes in each input affect the prediction. The variance contributions can be added separately only when those components are independent.
Worked Example: Dissolution Measurement and In Vivo Input
Consider the hypothetical model \(\mathrm{AUC}=A_{\mathrm{ref}}D^\gamma\), with \(D=0.80\), \(\gamma=1.10\), and \(A_{\mathrm{ref}}=12.0\) mg·h/L. Assume the dissolution measurement has \(SE_D=0.03\), the parameter has \(SE_\gamma=0.12\), the reference exposure has log-scale standard error 0.04, and an independently estimated in vivo input contributes log-scale standard deviation 0.06. Assume all four sources are independent and the local approximation is adequate.
The point prediction is \(12.0(0.80)^{1.10}=12.0\exp\{1.10\log(0.80)\}=9.388\) mg·h/L. On the log scale, \(L=\log(A_{\mathrm{ref}})+\gamma\log D+\eta\). The sensitivity to \(D\) is \(\gamma/D=1.10/0.80=1.375\), and the sensitivity to \(\gamma\) is \(\log(D)=-0.22314\). The resulting variance contributions are:
The in vivo input accounts for about \(0.003600/0.007619=47.2\%\) of this variance, dissolution measurement for 22.3%, the reference exposure for 21.0%, and the model parameter for 9.4%. These shares identify where uncertainty is largest under the stated model and independence assumptions; they do not establish that one source is biologically more important in general.
If a future outcome also has independent log-scale residual standard deviation 0.05, the total variance is \(0.007619+0.002500=0.010119\), so \(SE_L=0.1006\). The approximate 95% prediction interval is \(9.388\exp\{\pm1.96(0.1006)\}\), or about (7.71, 11.43) mg·h/L. This calculation assumes approximate normality on the log scale and includes the stated future-outcome residual component.
Convolution Predictions and Dominant Sources
A concentration prediction from a convolution depends on the timing of the in vivo input as well as its total amount. In a discretized convolution, a concentration at a specified time can be written as a weighted sum of absorption increments. The weights encode disposition and the time relationship between each increment and the observation time. Uncertainty in the increments and weights therefore contributes to concentration uncertainty. For Cmax, there is an additional complication: the time at which the maximum occurs can shift across plausible inputs.
Worked Example: Uncertainty in a Convolution-Based Cmax
For a hypothetical two-interval convolution, suppose the concentration at the predicted peak time is \(C=8q_1+5q_2\) mg/L, where \(q_1\) and \(q_2\) are absorbed fractions in the two intervals. Let their estimates be 0.40 and 0.35, with standard deviations 0.04 and 0.03 and covariance \(-0.0006\). The convolution weights have estimates 8 and 5 mg/L per unit fraction, standard deviations 0.5 and 0.4 mg/L per unit fraction, and are independent of each other and of the \(q\)'s. Assume the predicted peak time remains fixed locally, and add an independent output residual standard deviation of 0.20 mg/L for a future outcome.
The point prediction is \(8(0.40)+5(0.35)=4.95\) mg/L. The first-order variance from the input fractions is \(8^2(0.04)^2+5^2(0.03)^2+2(8)(5)(-0.0006)=0.07690\) (mg/L)\(^2\). The weight-uncertainty contribution is \(0.40^2(0.5)^2+0.35^2(0.4)^2=0.05960\) (mg/L)\(^2\). Including future-outcome residual variance gives \(0.07690+0.05960+0.20^2=0.1765\) (mg/L)\(^2\), and \(SE_C=\sqrt{0.1765}=0.4201\) mg/L.
An approximate normal 95% prediction interval is \(4.95\pm1.96(0.4201)\), or (4.13, 5.77) mg/L. The input-fraction uncertainty contributes about 43.6% of the total variance, weight uncertainty about 33.8%, and residual variability about 22.7% (rounding accounts for the total slightly exceeding 100%). The negative covariance between the absorbed fractions reduces their combined variance. If the peak time changes over plausible inputs, however, the fixed-time linear approximation may not capture uncertainty in the maximum; simulation of the full concentration-time profile is preferable.
Choosing a Propagation Method and Checking It
The delta method is quick, transparent, and useful for showing local sensitivities and variance contributions. It can perform poorly when the prediction is strongly nonlinear, the input uncertainty is broad, parameters are near boundaries, or the output depends on a changing maximum such as Cmax. In those situations, simulation generally represents the prediction process more faithfully.
For Monte Carlo propagation, draw jointly from the estimated distribution of dissolution inputs and model parameters, preserving their covariance. For each draw, apply the IVIVC, calculate the in vivo input and disposition, and compute the exposure measure. Add residual variability only when the target is a future outcome rather than the mean model prediction. The resulting simulated distribution can be summarized with suitable quantiles; its interval need not be symmetric. A parametric bootstrap or resampling approach can also be used when refitting the model for each resample is important, but resampling must reflect the study design and the source of uncertainty.
Specify whether the output is AUC, Cmax, a concentration at a fixed time, or another quantity, and whether the interval is for a mean prediction or a future outcome.
Include relevant dissolution measurements, IVIVC parameters, in vivo input estimates, disposition quantities, and—if appropriate—future-outcome residual variability.
Use the delta method with a covariance matrix for a local approximation, or simulate jointly when nonlinearities or changing peak times matter.
Check interval plausibility, sensitivity to assumptions, covariance handling, and the contribution of each uncertainty source.
Diagnostics should compare delta-method results with simulation when feasible, especially for skewed or bounded outputs. Inspect whether simulated predictions violate physical limits—for example, an absorbed fraction outside its allowable range—and whether a few extreme draws dominate the interval. Confirm that covariance matrices are valid and that uncertainty sources have not been counted twice. If the interval changes greatly when a local approximation is replaced by simulation, report that sensitivity rather than presenting the approximation as definitive.
A common mistake is to combine standard errors as if all inputs were independent, or to report only parameter uncertainty as a prediction interval. The correction is to include covariance terms or joint draws, state exactly which uncertainty sources are included, and add future-outcome variability when the target is a new outcome. A second pitfall is to interpret a narrow interval as proof that the IVIVC is valid outside the data and conditions supporting it. Propagation quantifies uncertainty conditional on a model and its assumptions; it does not repair model misspecification, unrepresentative dissolution inputs, or unsupported extrapolation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a log-scale prediction, the slope estimate and dissolution value are uncertain and correlated. What must be included beyond their separate variance contributions?
- A delta-method interval is symmetric on the log scale. What happens to its endpoints when reported on the exposure scale?
- Why might simulation be preferable to the delta method for Cmax?
- A mean-prediction interval omits residual future-outcome variability. What does that imply if it is presented as a prediction interval for a new outcome?
Solutions. (1) Include the covariance cross-term \(2(\partial g/\partial W_1)(\partial g/\partial W_2)\operatorname{Cov}(W_1,W_2)\), or preserve the dependence in joint simulation. (2) Exponentiation gives positive, generally asymmetric endpoints on the original scale. (3) Cmax is a maximum, and its peak time may shift with the inputs; the resulting output can be nonlinear and non-normal. (4) The interval will generally be too narrow because it describes uncertainty in the mean prediction, not the additional variability of a future outcome.