Why Vary the Absorption Assumption?
Model-based exposure comparisons depend on how the model represents drug entry into systemic circulation. A first-order absorption model assumes a constant fraction of the unabsorbed dose enters per unit time. A transit-compartment model represents a distribution of delays before absorption, while a lag-time model postpones the start of absorption by a specified duration. These structures can describe different concentration-time profiles, even when they imply the same eventual amount absorbed.
Sensitivity analysis asks whether the formulation contrast of interest changes materially when the absorption structure is changed. The comparison should hold the study data, exposure definition, disposition assumptions, and other analysis choices fixed as far as possible. Each absorption model is fitted separately, and the same exposure contrast is recalculated from each fitted model. The point is not to choose whichever model gives the preferred result, but to assess how much the conclusion relies on a structural assumption.
Exposure Under Alternative Absorption Structures
For a clear comparison, consider a one-compartment disposition model with dose \(D\), apparent systemic availability \(F\), volume \(V\), and elimination rate constant \(k_e\). Clearance is \(CL=k_eV\). Assume linear disposition, a known dose, no time-varying clearance, and complete absorption of the fraction represented by \(F\). The absorption structure determines when that fraction enters systemic circulation.
With first-order absorption at rate \(k_a\), and no lag, the concentration for \(t\geq 0\) is
For a lag time \(t_{\mathrm{lag}}\), replace \(t\) by \(t-t_{\mathrm{lag}}\) when \(t>t_{\mathrm{lag}}\), and set \(C(t)=0\) before the lag. This expression assumes first-order absorption begins after the lag and that the same disposition model applies thereafter.
A transit-compartment model instead treats the time \(U\) at which drug enters systemic circulation as having an Erlang distribution. With \(m\) transit stages and rate \(k_{\mathrm{tr}}\), its density is
The mean transit time is \(m/k_{\mathrm{tr}}\). The resulting concentration is the convolution of this input-time distribution with the disposition response:
A partial AUC through time \(T\) is the area under the concentration curve from zero to \(T\). It measures exposure only within that window; it is not generally equal to total AUC. For any normalized input-time density \(g(u)\), integration of the disposition response gives
For first-order absorption, \(g(u)=k_ae^{-k_au}\). For a lag-time model, the input density is shifted: \(g(u)=0\) for \(u<t_{\mathrm{lag}}\), and \(g(u)=k_ae^{-k_a(u-t_{\mathrm{lag}})}\) thereafter. These expressions make the central issue explicit: input arriving near or after \(T\) contributes less to \(\mathrm{AUC}_{0-T}\).
Worked Calculations and a Sensitivity Comparison
Worked Example: First-Order Absorption and Partial AUC
Suppose a hypothetical formulation has \(k_a=1.00\ \mathrm{h}^{-1}\), the reference has \(k_a=0.500\ \mathrm{h}^{-1}\), and both have \(k_e=0.500\ \mathrm{h}^{-1}\). Let \(F=1\) for both and \(D/CL=100\) exposure units. Calculate the partial-AUC ratio through \(T=4.00\ \mathrm{h}\).
For first-order absorption with \(k_a\ne k_e\), the dimensionless fraction of total exposure captured by time \(T\) is
For the formulation, substituting the values gives
Thus its partial AUC is \(100(0.7476)=74.76\) exposure units. For the reference, \(k_a=k_e\), so the limiting expression is \(G=1-e^{-k_eT}(1+k_eT)\). Therefore,
The faster first-order absorption produces more exposure within the four-hour window, even though the two formulations have the same \(F\), dose, and disposition parameters. This is a finite-window contrast, not evidence that their total exposures differ.
Worked Example: Recalculating a Contrast Across Absorption Models
Consider hypothetical outputs from separate fits to the same concentration data. Let \(k_e=0.500\ \mathrm{h}^{-1}\), \(D/CL=100\) exposure units for the reference, and \(F_{\mathrm{form}}/F_{\mathrm{ref}}=0.995\). The exposure of interest is \(\mathrm{AUC}_{0-4}\). Suppose the first-order fit gives \(k_{a,\mathrm{form}}=1.00\ \mathrm{h}^{-1}\) and \(k_{a,\mathrm{ref}}=0.500\ \mathrm{h}^{-1}\). The first-order fractions from the preceding calculation are 0.7476 and 0.5940, so
and the formulation-to-reference ratio is \(74.39/59.40\approx 1.252\).
Now suppose a two-stage transit fit gives \(k_{\mathrm{tr}}=1.00\ \mathrm{h}^{-1}\) for both formulations. For \(m=2\), the input density is \(g(u)=u e^{-u}\). The fraction of total exposure within four hours is
The reference partial AUC is \(100(0.5869)=58.69\) units, and the formulation partial AUC is \(100(0.995)(0.5869)=58.40\) units. Their ratio is \(58.40/58.69=0.995\).
Finally, suppose a lag-time fit gives both formulations a lag of \(1.00\ \mathrm{h}\) and a post-lag \(k_a=1.00\ \mathrm{h}^{-1}\). The available absorption duration within the four-hour window is three hours. The first-order fraction over that duration is
The reference partial AUC is \(60.35\) units; the formulation partial AUC is \(100(0.995)(0.6035)=60.05\) units. The ratio is again \(0.995\). In this hypothetical comparison, the point estimate changes from about 1.252 under first-order absorption to 0.995 under the other fits. If 1.25 were used as an illustrative point-estimate boundary, only the first-order estimate would exceed it. This is not a formal bioequivalence decision: that requires the relevant prespecified inferential procedure and uncertainty, not a point estimate alone.
Why Total AUC Can Behave Differently
The preceding example concerns a finite window. Under linear disposition and complete eventual absorption of fraction \(F\), total AUC through infinite time is invariant to the timing distribution of absorption. This follows by integrating the convolution:
The last equality uses two assumptions: \(g\) is a normalized input-time density, so its integral is one, and \(CL=Vk_e\). Thus changing \(k_a\), adding a lag, or using a transit distribution changes the shape and timing of concentrations but not total AUC under these assumptions. A different estimated \(F\), incomplete absorption, nonlinear disposition, or a finite observation horizon can change the exposure comparison.
Worked Example: Checking Total-AUC Invariance
Use the hypothetical availability ratio \(F_{\mathrm{form}}/F_{\mathrm{ref}}=0.995\), equal doses, and common linear clearance. Under the assumptions above, each formulation’s total AUC is \(FD/CL\), regardless of its absorption rate, transit time, or lag. Therefore,
This total-AUC ratio is not the first-order partial-AUC ratio of 1.252 calculated above. The difference is not a contradiction: early exposure depends on the timing of input, while total exposure integrates the full concentration curve. If the target is a partial AUC, replacing it with the total-AUC invariance argument would answer a different question.
Conducting and Interpreting the Sensitivity Analysis
A defensible comparison starts by defining the estimand: the exposure measure, time window, formulation contrast, and scale (for example, a formulation-to-reference ratio). Then fit each plausible absorption structure using the same observations and, where appropriate, the same disposition, period, sequence, subject, and residual-error specifications. These controls help isolate sensitivity to absorption structure rather than to several simultaneous analysis changes.
Identify scientifically plausible structures, such as first-order absorption, a prespecified number of transit stages, and a lag-time model. State which parameters are formulation-specific and which are shared.
Review convergence, parameter estimates, uncertainty, residual patterns, and whether the fitted curves represent the observed sampling times. A failed or poorly identified fit is not a meaningful sensitivity result.
Compute the same exposure measure from each fitted curve, then report its value, units, and uncertainty under each structure. Keep the estimand and contrast direction unchanged.
Describe whether the magnitude, interval, or decision changes across plausible structures. If bootstrap uncertainty is used, refit the relevant models and recalculate the contrasts within each resample.
Useful diagnostics include systematic residual patterns around the rising phase or peak, implausible absorption parameters, estimates near parameter boundaries, large standard errors, and strong dependence of the contrast on a small number of early samples. Lag time and slow absorption can be difficult to distinguish when early sampling is sparse. Transit-stage count and transit rate can likewise trade off against one another. These issues can make different absorption structures fit similarly while producing different extrapolations or partial AUCs.
A fit statistic alone does not establish that the absorption structure is correct. Inspect fitted curves and residuals in relation to time, examine parameter precision and identifiability, and consider whether the sampling schedule can distinguish the candidate structures. If a model does not converge reliably or its key parameters are weakly identified, report that limitation rather than treating its contrast as equally informative.
Comparison, Limitations, and a Common Pitfall
| Absorption structure | Timing assumption | What to examine in sensitivity analysis |
|---|---|---|
| First-order | A constant fraction of the remaining dose is absorbed per unit time. | Whether the estimated absorption rate captures the observed rise and whether the partial-window contrast changes with that rate. |
| Transit-compartment | Input times follow a staged delay distribution; the mean is \(m/k_{\mathrm{tr}}\). | Whether stage count and transit rate are identifiable and whether the implied delay is supported by the sampling schedule. |
| Lag-time | Absorption starts only after a specified delay, then proceeds according to its post-lag absorption process. | Whether the data support a distinct lag rather than gradual slow absorption, and whether the lag is estimated precisely. |
The main alternative to a multi-model sensitivity analysis is to choose one absorption structure as the primary model and justify it using pharmacological knowledge, design, and model diagnostics. A prespecified primary model gives a direct analysis, but it does not show how dependent the answer is on the absorption assumption. Sensitivity analysis adds that perspective; it does not replace the primary analysis, model assessment, or uncertainty quantification.
A common mistake is to conclude that a formulation has greater total exposure because its fitted curve rises faster or its early partial AUC is larger. The correction is to identify the exposure estimand first. Under linear disposition and complete eventual absorption, faster input can increase early exposure without changing total AUC. Conversely, conclusions about finite-window exposure can genuinely depend on the input structure.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Under linear disposition and complete eventual absorption, what determines whether changing \(k_a\) changes total AUC?
- Why can two formulations with equal \(F\) have different \(\mathrm{AUC}_{0-T}\)?
- A lag-time model and a slow first-order model fit similarly, but give different partial-AUC ratios. Name one diagnostic concern and one appropriate interpretation.
- What must be held fixed, as far as possible, when comparing absorption assumptions?
- Why is a point-estimate ratio alone insufficient for a formal bioequivalence conclusion?
Solutions. (1) Under the stated assumptions, total AUC is \(FD/CL\), so changing only the timing distribution or \(k_a\) does not change it. (2) The formulations may deliver different fractions of their dose to systemic circulation before \(T\); later input contributes less to the partial AUC. (3) Lag and slow absorption may be weakly distinguishable, particularly with sparse early sampling. Check residuals, parameter precision, and identifiability; report that the partial-AUC contrast is sensitive to the absorption structure rather than claiming either fit is established as true. (4) Keep the data, estimand, exposure window, contrast direction, and other model components consistent, changing the absorption structure as the intended sensitivity factor. (5) Formal inference also requires uncertainty and the prespecified decision criterion; a point estimate does not show whether the relevant interval meets that criterion.