Predicting Population Bioequivalence From Exposure Distributions

From Model Predictions to Population Coverage

A geometric mean ratio summarizes the center of a test-to-reference exposure comparison. It does not, by itself, say what fraction of people in a target population are predicted to have ratios within specified bounds. Two populations can have the same geometric mean ratio but very different spreads, and therefore different fractions inside the bounds. Model-based simulation can estimate that fraction by generating exposure predictions for representative virtual subjects under both formulations.

The target quantity must be defined before simulating. Specify the exposure endpoint, the population to which the result applies, the test-to-reference ratio, and the lower and upper bounds. The calculation then asks how often a virtual subject’s predicted ratio falls within those bounds. This is a population coverage estimand: it describes a proportion under a specified population distribution and model. It is not, by itself, the separate moment-based regulatory criterion sometimes called population bioequivalence.

The paired notation matters. The ratio compares predictions for the same simulated profile under test and reference. A model therefore needs a defensible way to represent both outcomes for that profile, such as shared covariates and shared subject-level characteristics. Two unrelated samples of test and reference predictions provide marginal distributions, but do not by themselves define which test response should be divided by which reference response. The pairing rule is part of the estimand and model assumptions, not a harmless computational detail.

This tutorial concerns ratios of model-predicted exposures such as AUC or Cmax, not observed concentration at a single time. The endpoint, limits, and target population must be specified on the relevant scale. Limits may be stated as ratios or as equivalent log-ratio bounds: if \(X(Z)=\log R(Z)\), then the event is \(\log L\le X(Z)\le\log U\). Equality at a boundary is included in the definition; for continuous distributions it has probability zero.

Assumptions Before Simulation

A numerical estimate is only meaningful relative to the model and population used to produce it. Before simulating, establish the following conditions:

• Target-population representation. The distribution used to generate profiles reflects the intended population, including relevant covariates and any modeled subject-level variation. A sample from a narrow or unrepresentative covariate range does not automatically represent the target population.
• Valid exposure predictions. The structural and formulation components of the pharmacokinetic model must generate plausible predictions for each formulation over the relevant range. Exposure responses must be positive for ratios to be defined.
• Defensible pairing. Test and reference predictions must correspond to the same simulated profile, with shared characteristics assigned consistently. Any formulation-specific differences in the model must match the scientific question. This calculation does not establish that an unmodeled source of subject-by-formulation variation is absent.
• Prespecified decision bounds and endpoint. The limits and exposure metric should be fixed before interpreting the simulated coverage. Changing them after seeing the results changes the estimand.
• Adequate simulation. Simulated profiles must be drawn according to the intended population distribution, and enough draws must be used to make Monte Carlo error small for the intended precision.

These assumptions are distinct from the mechanics of counting ratios. A very large simulation can reduce random simulation error, but it cannot repair a misspecified structural model, an incorrect target population, or an unjustified pairing rule.

Estimating the Proportion

Suppose \(Z_1,\ldots,Z_M\) are independently generated profiles from the target-population distribution. For each profile, predict the exposure under both formulations and calculate \(R_i=Y_T(Z_i)/Y_R(Z_i)\). Record an indicator \(I_i\), equal to 1 when \(L\le R_i\le U\), and 0 otherwise. The estimated coverage is the average of these indicators.

The justification is direct. Under independent draws from the stated population and model, each indicator is a Bernoulli random variable with \(P(I_i=1)=p_{\mathrm{cov}}\). Its expected value is \(p_{\mathrm{cov}}\). By linearity of expectation, the expected average of \(M\) indicators is also \(p_{\mathrm{cov}}\), so the simulated proportion is unbiased for the model-implied coverage. Independence gives variance \(p_{\mathrm{cov}}(1-p_{\mathrm{cov}})/M\); replacing the unknown probability with its estimate yields the usual Monte Carlo standard error.

This standard error describes variation from generating a finite number of simulation draws, conditional on the fitted model and its parameters. It is not a confidence interval for the true population coverage that includes uncertainty in model parameters or model structure. If the model parameters were estimated from data, that additional uncertainty requires a separate method.

Using a Distributional Model

When the log ratio has a tractable distribution, coverage can sometimes be calculated directly rather than estimated by simulation. This also provides a useful check on a simulation implementation. If \(X=\log R\) follows a normal distribution with mean \(\mu_X\) and standard deviation \(\sigma_X\), then the log ratio has distribution \(X\sim N(\mu_X,\sigma_X^2)\), and the coverage is the probability between the log-transformed limits. Let \(\Phi\) denote the cumulative distribution function of a standard normal variable. Standardizing each limit gives

The subtraction follows because the event \(X\le\log U\) includes all values below the upper bound, including those below the lower bound. Subtracting \(P(X<\log L)\) leaves only the probability between the two limits. This formula relies on the normal log-ratio assumption; it is not automatically appropriate for every exposure distribution.

Simulation Precision, Diagnostics, and Alternatives

A simulation result should be checked both for numerical precision and for scientific plausibility. First inspect the generated target profiles: compare their key covariate distributions with the intended population and check that the simulated range is plausible. Then inspect paired predictions and ratios for nonpositive values, extreme values, or patterns suggesting that the model is being used outside its supported range. A coding check should confirm that each test prediction is divided by the reference prediction for the same profile, and that the bounds are applied on the intended scale.

Monte Carlo stability can be assessed by increasing \(M\), repeating the simulation with independent random draws, or examining the Monte Carlo standard error. If the estimate changes materially between runs or its simulation error is too large for the intended interpretation, increase the number of draws. When draws are weighted, stratified, or otherwise dependent, the simple Bernoulli standard error above may not apply unchanged; the calculation must reflect the actual simulation design.

A useful alternative is the analytic calculation for a specified ratio distribution, as in the previous example. It avoids random simulation error when its distributional assumptions are appropriate. Model-simulated distributions are more flexible when the response ratio is generated from covariates or a more complex pharmacokinetic model, but they still depend on correct model specification and adequate sampling.

The geometric mean ratio comparison taught in the earlier material on estimating geometric mean ratios summarizes the center of the comparison. It is useful for that purpose, but it is not a substitute for coverage: a central estimate does not tell how much probability lies inside two limits. Conversely, a coverage estimate does not replace every other bioequivalence summary or decision criterion. The estimand answers a particular question about the proportion of the specified population whose predicted ratios meet specified bounds.

For example, with \(M=1000\), a simulated estimate of 0.922 has approximate Monte Carlo standard error \(\sqrt{0.922(0.078)/1000}=0.00848\), or 0.848 percentage points. A rough normal interval based only on simulation error is \(0.922\pm1.96(0.00848)\), approximately 0.905 to 0.939. Even if a hypothetical target coverage were 0.90, this interval would address only finite simulation noise conditional on the model. It does not establish that the model is correct or account for uncertainty in its fitted parameters.

Common Pitfalls and Limits

• Using marginal predictions as if they were paired. Dividing unrelated test and reference predictions creates a ratio distribution that depends on an arbitrary pairing. Generate both predictions for the same profile under a scientifically specified pairing rule.
• Equating a geometric mean ratio with population coverage. The center does not determine the spread or the fraction inside the bounds. Calculate the coverage probability directly from the distribution of ratios.
• Treating simulation error as total uncertainty. More simulated profiles reduce Monte Carlo error, not uncertainty caused by estimated parameters, model misspecification, or an incorrect target population.
• Changing the scale or limits silently. Ratio bounds \(L,U\) correspond to log bounds \(\log L,\log U\). Applying ratio limits directly to log ratios is incorrect.

The main limitation is that the result is model-implied and population-specific. Changing the covariate distribution, endpoint, pairing assumptions, or limits can change the estimated fraction. A precise estimate under an inappropriate model is still an inappropriate estimate. The result should therefore be reported with its endpoint, target population, limits, model assumptions, simulated proportion, and Monte Carlo precision.