Interindividual Variability On Absorption Parameters

Why the Scale of Absorption Variability Matters

Population models can allow subjects to differ in how quickly drug enters the systemic circulation, whether absorption is delayed, and what fraction of an administered dose becomes available. These are not interchangeable kinds of variation. An absorption rate must be positive, a lag time is usually nonnegative, and absolute bioavailability is bounded between zero and one. A random-effect formulation should represent those constraints in the individual values it generates, not merely in the population typical value.

The choice of scale also determines what a variance means. A variance on the log scale describes multiplicative variation, while a variance on the original scale describes additive variation in the parameter's units. For bioavailability, a random effect on the logit scale changes log-odds rather than adding a fixed amount to the fraction. The earlier material on transforming random effects and lognormal parameters provides the general background; here the emphasis is on the consequences for absorption parameters.

Positive Absorption Rates: Multiplicative Effects

Let \(k_{a,i}\) denote subject \(i\)'s first-order absorption rate, measured in inverse time, and let \(k_{a,\mathrm{pop}}>0\) be its typical population value. A standard model is

Because the exponential is always positive, every individual rate is positive. Taking logarithms gives \(\log(k_{a,i})=\log(k_{a,\mathrm{pop}})+\eta_{ka,i}\), so the rate is lognormally distributed under this assumption. The typical value is the median of the individual distribution, not generally its arithmetic mean. The distinction follows from the symmetry of \(\eta_{ka,i}\) around zero and the asymmetry of the exponential transformation.

The model's relative spread can be summarized by its coefficient of variation (CV). For a lognormal parameter with log-scale variance \(\omega_{ka}^2\), the CV is \(\sqrt{\exp(\omega_{ka}^2)-1}\). A central 95% model-based range can be obtained by transforming the approximate normal range \(-1.96\omega_{ka}\) to \(1.96\omega_{ka}\). This is a range for individual rates implied by the specified population model, not a confidence interval for the typical rate.

An additive alternative, \(k_{a,i}=k_{a,\mathrm{pop}}+\eta_{ka,i}\), puts variation in \(\mathrm{h}^{-1}\) rather than in a multiplicative factor. If \(\eta_{ka,i}\) is normally distributed, this model assigns some probability to negative rates, which are not physically meaningful. It can be defensible only if the parameterization or constraints prevent impossible values and the additive interpretation is appropriate; otherwise, a log-scale effect is a natural choice for a positive rate.

Lag Time: Nonnegative Values and the Zero Boundary

Let \(T_{\mathrm{lag},i}\) be subject \(i\)'s absorption lag time, measured in hours. If lag times are strictly positive and vary proportionally, one possible formulation is

This yields positive individual lag times and a lognormal distribution. However, it cannot represent a population typical lag of exactly zero: multiplying zero by an exponential still gives zero for every subject. If a zero lag is scientifically plausible, an additive formulation around zero, a different parameterization, or a model with a separate probability of having a nonzero lag may be needed. Each choice changes the implied distribution and should be motivated by the model and data rather than selected only for computational convenience.

A log-scale model is not automatically superior for every lag-time problem. It excludes zero and can create a long right tail. An additive model allows values on both sides of the typical value and has a variance in squared time units; it may be reasonable when the estimated spread is small relative to a safely positive typical lag. The choice should be checked against plausible individual values and the possibility of a true zero lag.

Bioavailability: Respecting a Bounded Fraction

Absolute bioavailability \(F_i\) is a fraction, with \(0<F_i<1\) under a continuous model. A normal additive effect, \(F_i=F_{\mathrm{pop}}+\eta_{F,i}\), can produce values below zero or above one. A common alternative places the random effect on the logit scale:

Equivalently, \(F_i=\{1+\exp[-\operatorname{logit}(F_{\mathrm{pop}})-\eta_{F,i}]\}^{-1}\). The inverse-logit maps every finite real number to a value strictly between zero and one. Consequently, the model enforces the bounds for every subject. Because the transformation is nonlinear, \(F_{\mathrm{pop}}\) is the median of \(F_i\) under a zero-median symmetric random effect, but it is not generally the arithmetic mean. Also, \(\omega_F\) describes variability in log-odds, not a standard deviation in fraction units.

How to Select and Check the Formulation

The transformation should reflect both the parameter's support and the scientific meaning of its variability. The following comparison summarizes common choices.

Several random effects can be modeled jointly. For example, effects on \(k_a\), lag time, and \(F\) may have a covariance structure, as described by the omega matrix. A correlation on the random-effect scales is not necessarily the same as a simple correlation between the natural-scale parameter values. In particular, a logit-scale correlation concerns log-odds deviations, and a log-scale correlation concerns log deviations. Such a structure should be supported by the data and interpreted on the scales where it is defined.

Diagnostics should ask whether the fitted distribution and its consequences are plausible, not only whether the optimizer converged. Examine the estimated variability and its uncertainty, the implied individual ranges, and model-based predictions of concentration-time profiles across subjects. A large estimated variance can imply very long absorption delays or rates far outside credible values; a near-zero estimate may indicate little information or little supported variability. Check whether the model reproduces the observed timing and shape of absorption-related profiles, especially early concentrations when absorption rate and lag have their clearest influence.

Absorption parameters can be difficult to distinguish from one another and from other structural parameters. Sparse early sampling may provide weak information about \(k_a\) or lag time. Bioavailability affects exposure and may be confounded with clearance or other dose-related factors in a given design. These are interpretive limitations: a plausible transformed distribution does not guarantee that the data identify its variance. Simulated individual values, uncertainty estimates, and predictive checks help reveal problems, but cannot replace adequate design or a suitable structural model.

Deriving the Lognormal Summaries

The mean and CV formulas used for the rate example follow directly from the normal random effect. If \(P_i=P_{\mathrm{pop}}\exp(\eta_i)\), with \(\eta_i\sim N(0,\omega^2)\), then the normal moment-generating function gives \(E[\exp(\eta_i)]=\exp(\omega^2/2)\) and \(E[\exp(2\eta_i)]=\exp(2\omega^2)\). Therefore,

Dividing the variance by the squared mean and taking the square root yields \(\mathrm{CV}=\sqrt{\exp(\omega^2)-1}\). The median is \(P_{\mathrm{pop}}\) because \(P_i\leq P_{\mathrm{pop}}\) exactly when \(\eta_i\leq0\), which has probability one half. These results require a normally distributed random effect and the stated exponential transformation. If the random-effect distribution changes, these summaries need not hold.

Common Mistakes and Practical Interpretation

• Calling the typical value the mean. Under a lognormal model it is the median; the mean is higher when \(\omega^2>0\).
• Interpreting \(\omega\) as a natural-scale standard deviation. For a log-rate, it is a standard deviation of log deviations; for a logit bioavailability model, it is a standard deviation of log-odds deviations.
• Adding a normal effect to a bounded or positive parameter without checking support. Such a model can imply negative rates or bioavailabilities outside zero to one.
• Assuming a transformation solves an identification problem. A log or logit link enforces support, but does not make weakly informed variability estimable.
• Treating a transformed percentile range as an estimation interval. It describes individual variation conditional on model parameters, not uncertainty about those parameters.

The practical goal is not to use the most elaborate transformation. It is to specify an individual distribution that respects the parameter's meaning, to interpret its variability on the correct scale, and to verify that the resulting values and concentration profiles are credible for the intended population.