From Model Parameters to an Exposure Ratio
A pharmacokinetic model estimates parameters; a bioequivalence analysis often needs an exposure contrast. The conversion is direct when the log exposure is a linear function of model parameters, but can require integration when exposure is derived from a nonlinear concentration-time curve or when subject random effects affect the formulation contrast. The key is to define the target ratio first, then carry the model’s uncertainty through that same calculation.
Let \(M_F(\theta,b)\) denote a positive model-derived metric, such as AUC or Cmax, for formulation \(F\). Here \(\theta\) is the vector of fixed-effect parameters and \(b\) represents subject-level random effects. For a test formulation \(T\) and reference formulation \(R\), define the paired log contrast for a subject with effects \(b\) as \(d(\theta,b)=\log M_T(\theta,b)-\log M_R(\theta,b)\). The corresponding subject-specific ratio is \(\exp\{d(\theta,b)\}\).
This definition uses the same subject effect for the two formulations when the model defines a paired comparison. Shared subject effects may cancel in the contrast; formulation-by-subject effects generally do not. The estimand determines whether the goal is a typical-subject contrast or a population-average log contrast. The introductory material on model-based bioequivalence estimands establishes that this choice must be made before interpreting a model-derived result.
Calculating and Interpreting the GMR
If the model gives a log-metric contrast directly, exponentiation converts it to a ratio. For example, a log AUC contrast of \(0.10\) corresponds to a test/reference GMR of \(\exp(0.10)=1.105\), or about 10.5% higher model-derived AUC for test. The contrast is dimensionless, so the ratio has no units, although the underlying AUC and Cmax metrics do.
For oral administration, a simple structural relationship is \(AUC=F\cdot Dose/CL\), where \(F\) is bioavailability and \(CL\) is clearance. Taking logarithms gives \(\log AUC=\log F+\log Dose-\log CL\). For equal doses, the formulation contrast is therefore the log bioavailability contrast minus the log clearance contrast. More generally, a model may calculate AUC by integrating its predicted concentration curve, or calculate Cmax by maximizing that curve. In either case, compute the metric for each formulation under the specified model, take logs, and form the contrast.
Worked Example: A Linear Log-AUC Contrast
Suppose a hypothetical model for equal-dose treatments has \(\log AUC=\alpha+\beta I(T)\), where \(I(T)=1\) for test and \(0\) for reference. The fitted test coefficient is \(\hat\beta=0.120\), with \(SE(\hat\beta)=0.050\). Assume the coefficient estimate is approximately normal and that this is the only uncertain parameter in the contrast.
The reference log AUC is \(\alpha\), and the test log AUC is \(\alpha+0.120\). Thus \(\hat\Delta=0.120\) and \(\widehat{GMR}=\exp(0.120)=1.127\). The model estimates test AUC to be 1.127 times reference AUC, or about 12.7% higher.
The standard error of the log contrast is \(0.050\). A normal-approximation 95% interval on the log scale is \(0.120\pm1.96(0.050)=(0.0220,0.2180)\). Exponentiating both endpoints gives \((\exp(0.0220),\exp(0.2180))=(1.022,1.244)\). Under the stated model and approximation, this interval describes uncertainty in the estimated GMR; it is not a 95% probability statement about a fixed true ratio. No equivalence conclusion follows from this calculation alone.
Propagating Fixed-Effect Uncertainty
When the log contrast is a function \(f(\theta)\) of estimated parameters, its uncertainty depends on both the sensitivity of the contrast to each parameter and the covariance among parameter estimates. Let \(\hat\theta\) be the fitted parameter vector and \(V\) its estimated covariance matrix. A first-order Taylor expansion around \(\hat\theta\) gives the delta-method approximation:
The gradient contains the partial derivatives of the log contrast with respect to the parameters. The covariance matrix includes the variances of estimates on its diagonal and their covariances off the diagonal. Ignoring covariance is justified only when the relevant estimates are uncorrelated or the omitted terms are negligible. For a large-sample normal approximation, a 95% interval for the GMR is obtained by forming \(\widehat{\Delta}\pm1.96SE(\widehat{\Delta})\) and exponentiating the two limits. If inference uses a \(t\) distribution with an appropriate \(df\), use its critical value instead.
The reason for this calculation is that, near the fitted value, a small parameter change \(\theta-\hat\theta\) changes the contrast by approximately \(\nabla f(\hat\theta)^{\mathsf T}(\theta-\hat\theta)\). The variance of that linear combination is the gradient multiplied by \(V\) and its transpose. Exponentiation is applied after constructing the interval because the uncertainty calculation is on the log-ratio scale.
Worked Example: AUC Depends on Bioavailability and Clearance
Consider a hypothetical equal-dose model in which the test/reference log-AUC contrast is \(f(\theta)=\delta_F-\delta_{CL}\). The fitted log bioavailability contrast is \(\hat\delta_F=0.080\), with standard error \(0.040\); the fitted log clearance contrast is \(\hat\delta_{CL}=-0.030\), with standard error \(0.030\). Assume their estimated correlation is \(0.25\), and use a normal approximation.
The estimated log-AUC contrast is \(0.080-(-0.030)=0.110\), so the GMR is \(\exp(0.110)=1.116\). The negative clearance contrast increases AUC for test relative to reference, because clearance is subtracted in the log-AUC equation.
The covariance of the two estimates is \(0.25(0.040)(0.030)=0.000300\). Since the contrast subtracts the clearance estimate, its variance is \(0.040^2+0.030^2-2(0.000300)=0.00190\), and its standard error is \(\sqrt{0.00190}=0.04359\). The 95% log-scale limits are \(0.110\pm1.96(0.04359)=(0.0246,0.1954)\). Exponentiating gives a GMR interval of approximately \((1.025,1.216)\). The covariance reduces the variance here; with a different sign or correlation, its effect could be different.
Random Effects: Variation Versus Uncertainty
Random effects serve two distinct roles in a GMR analysis. Their distribution describes variation among subjects, including any subject-specific formulation contrasts. The estimated random-effects parameters, such as a variance or covariance in \(\Omega\), also have estimation uncertainty. These are not interchangeable: a large between-subject spread does not automatically mean the estimated population GMR has equally large standard error, and uncertainty in an estimated variance is not the same as subject-to-subject variability.
For the population GMR defined above, first average the log contrast over the random-effects distribution, then exponentiate: \(\Delta(\theta,\Omega)=E_b[d(\theta,b)]\). When this expectation has no closed form, approximate it using numerical integration or a sufficiently large set of draws from the specified random-effects distribution. If the contrast is nonlinear in \(b\), simply setting \(b=0\) can give a different answer. When the random-effects distribution is symmetric and the contrast is linear in \(b\), its mean-zero effects may cancel in the population mean, even though individual ratios vary substantially.
Uncertainty in the estimated population GMR can be propagated by treating \(\Delta\) as a function of both fixed-effect and random-effects parameters. If \(\psi\) collects those parameters and \(V_\psi\) is their joint covariance matrix, the delta-method variance is approximately \(\nabla_\psi\Delta^{\mathsf T}V_\psi\nabla_\psi\Delta\). A parametric simulation alternative draws plausible parameter vectors using their estimated joint sampling distribution, evaluates the population contrast for each, and summarizes the resulting GMR distribution. Numerical integration error should be checked separately from parameter uncertainty.
Worked Example: A Nonlinear Random-Effects Contrast
Suppose a hypothetical model-derived log-Cmax contrast for a subject is approximated by \(d(b)=\beta+b+c b^2\), where \(b\sim N(0,\tau^2)\). This quadratic form represents a nonlinear dependence of the contrast on a subject effect. Let \(\hat\beta=0.080\), \(c=0.50\) (treated as known for this example), and \(\hat\tau=0.200\). Assume \(SE(\hat\beta)=0.030\), \(SE(\hat\tau)=0.040\), and that these two estimates are independent.
Because \(E(b)=0\) and \(E(b^2)=\tau^2\), the population mean log contrast is \(\Delta=\beta+c\tau^2\). Substitution gives \(\hat\Delta=0.080+0.50(0.200^2)=0.100\), so the population GMR is \(\exp(0.100)=1.105\). Setting \(b=0\) would instead give \(\exp(0.080)=1.083\), missing the contribution of the quadratic term.
For the uncertainty calculation, the derivatives are \(\partial\Delta/\partial\beta=1\) and \(\partial\Delta/\partial\tau=2c\tau=0.200\). Under the stated independence assumption, the approximate variance is \(1^2(0.030^2)+(0.200)^2(0.040^2)=0.000964\), giving \(SE(\hat\Delta)=0.03105\). The 95% log-scale limits are \(0.100\pm1.96(0.03105)=(0.0392,0.1608)\); exponentiating yields approximately \((1.040,1.174)\). This interval reflects uncertainty in \(\beta\) and \(\tau\). It is not the distribution of individual subject ratios, which also varies because \(b\) varies.
Diagnostics, Alternatives, and Limitations
The calculation is only as reliable as the fitted model, the target definition, and the numerical transformation. Check that the contrast uses the intended formulation ordering and that the model-derived metric is computed on the correct scale. For AUC, verify the integration interval and that the model’s curve represents the intended exposure. For Cmax, verify that the calculated maximum is on the relevant time range and not merely a value at a sparse set of evaluation times.
- Check parameterization. Confirm whether model coefficients are on a log scale and whether the formulation contrast has the correct sign before exponentiating.
- Inspect the covariance structure. Large correlations, near-singular covariance matrices, or variance estimates near a boundary can make standard errors and normal approximations unstable.
- Check nonlinear calculations. Compare numerical gradients with small finite differences, and increase the number of integration points or random-effect draws until the estimated contrast is stable.
- Separate variability from precision. Report subject-level random-effect spread as variability, and report uncertainty in the estimated GMR as inferential uncertainty.
A plug-in typical-subject calculation is simpler and may be appropriate when that is the defined target or when the contrast is linear in mean-zero random effects. Population integration is preferable when the target averages over subjects and nonlinear random-effect terms change the average log contrast. The delta method is efficient when the transformation is smooth and uncertainty is moderate; parametric simulation can better reflect nonlinear transformations or asymmetric uncertainty, but depends on the assumed sampling distribution and model. Neither approach corrects structural model misspecification, weak identifiability, or poorly estimated random-effects distributions.
A common mistake is to exponentiate each model coefficient separately, divide those exponentiated values, and ignore covariance. That can be wrong because a metric may depend on several parameters jointly, and the uncertainty of their contrast includes covariance terms. The correction is to derive the log-metric contrast first, calculate its variance using the full covariance information, and exponentiate the contrast and interval only afterward. The resulting GMR is an estimate of a specified model-based exposure contrast, not by itself a bioequivalence decision.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A model estimates a log test/reference AUC contrast of \(-0.060\). What is the GMR, and how should it be interpreted?
- Why can a population GMR differ from the ratio calculated at random effects \(b=0\)?
- Two log-scale parameter estimates have standard errors \(0.05\) and \(0.04\), with correlation \(0.50\). What is the variance of their difference?
- In a model with a linear, mean-zero random effect in the log contrast, does that random effect change the population mean log contrast? Does it imply there is no subject-level variability?
Solutions. (1) \(\exp(-0.060)=0.942\); the model estimates test AUC to be about 5.8% lower than reference. (2) Nonlinear terms in the random effects need not average to their value at zero; for example, \(E(b^2)=\tau^2\), not zero. (3) The covariance is \(0.50(0.05)(0.04)=0.001\), so the variance of the difference is \(0.05^2+0.04^2-2(0.001)=0.0021\). (4) A linear mean-zero random effect contributes zero to the population mean log contrast, but its variance can still produce substantial differences among individual subject ratios.