Sharing Parameters Across Metabolizer Classes

Why Decide What to Share?

A two-class metabolizer model can represent fast and slow clearance while keeping other aspects of the pharmacokinetic (PK) model common. This is useful when the scientific hypothesis concerns clearance specifically: estimating separate values for every PK parameter may consume information and make the model harder to interpret, without improving its description of the data. The opposite mistake is to force parameters to be equal merely to keep the model small, even when the class-specific concentration profiles suggest otherwise.

The decision is about the model's constraints. A shared population parameter has one value used in both classes. A class-specific population parameter has a separate value for each class. Sharing a population value does not, by itself, remove individual variability: for example, individuals may still have random deviations in clearance around their class's typical clearance.

A Clearance-Focused Model

Let \(c_i\) denote individual \(i\)'s latent class, with \(c_i \in \{\mathrm{fast},\mathrm{slow}\}\). Write the probability of the fast class as \(\pi_{\mathrm{fast}}\); the slow-class probability is \(1-\pi_{\mathrm{fast}}\). For a simple one-compartment model after an intravenous dose, let \(CL_{\mathrm{fast}}\) and \(CL_{\mathrm{slow}}\) be the class-specific typical clearances, and let \(V\) be a shared volume of distribution. A shared absorption parameter \(k_a\) would be relevant for an oral model, not for this intravenous example.

For illustration, suppose clearance has a log-normal individual deviation:

Here the typical clearance differs by class, while the random-effect variance \(\omega_{CL}^2\) is shared. This is one possible constraint, not an automatic requirement. A model could instead estimate a separate clearance random-effect variance in each class if there is sufficient information and a scientific reason to expect different within-class variability.

Let \(f(t;CL_i,V)\) denote the concentration predicted by the structural model at time \(t\), and let \(p(y_i\mid c_i,\eta_{CL,i})\) denote the likelihood contribution from individual \(i\)'s observations given class and random effect. If the class probabilities are common to individuals, that individual's likelihood contribution is a mixture:

The shared-parameter constraint means that the same \(V\), for example, enters both class-specific conditional likelihoods. It does not mean that the two integrals are identical: their typical clearances differ. The population likelihood is the product of the individual contributions, or equivalently its log likelihood is the sum of their logarithms.

Choosing Constraints Parameter by Parameter

A useful decision is not simply “share everything” versus “share nothing.” Consider each parameter according to what it represents and what the data can inform.

• Typical clearance: estimate separate \(CL_{\mathrm{fast}}\) and \(CL_{\mathrm{slow}}\) when the class distinction is intended to represent different clearance levels.
• Volume or absorption: share \(V\) or \(k_a\) if the working scientific model attributes the class difference to clearance alone. Allow separate values if there is a justified mechanism or a clear, reproducible improvement in fit that the available design can support.
• Random-effect variances and correlations: sharing \(\omega_{CL}^2\), or a covariance structure, is a separate decision from sharing typical clearances. A common variance asserts similar within-class heterogeneity; separate variances allow it to differ.
• Residual-error parameters: a common residual model is often reasonable when the observation process is common across classes. Separate residual parameters need a specific rationale, since they can absorb differences that might otherwise be attributed to the PK model.

Each added class-specific parameter increases the number of quantities the data must estimate. That can be a poor trade when class-specific observations are limited or the classes produce similar predictions. Conversely, a shared value can systematically misrepresent one class if its true parameter differs. The aim is a scientifically interpretable model with enough flexibility for the observed data, not the smallest possible parameter count.

Comparing Nested Sharing Models

When one model is obtained from another by setting parameters equal, the models are nested. For example, a model with separate \(V_{\mathrm{fast}}\) and \(V_{\mathrm{slow}}\) reduces to a shared-volume model under the restriction \(V_{\mathrm{fast}}=V_{\mathrm{slow}}\). The maximized log likelihood of the less constrained model cannot be lower, because it includes the constrained model as a special case. A likelihood-ratio statistic quantifies the improvement:

Under the null hypothesis that the equality constraints hold, the usual large-sample reference is a chi-square distribution with degrees of freedom equal to the number of independent constraints. This approximation requires regular likelihood conditions, including an appropriately identified model and null values that are not on a problematic boundary. In mixture models, weak class separation or sparse class information can make the approximation unreliable. A comparison of sharing constraints within a fixed two-class model is also different from testing whether a second class is needed.

Derivation and Practical Checks

The likelihood-ratio calculation follows directly from the constrained parameter space. Let \(\Theta\) denote the parameter space of the less constrained model, and let \(\Theta_0\) be the subset satisfying the sharing restrictions. Since \(\Theta_0\subseteq\Theta\), maximizing the log likelihood over the larger set cannot produce a smaller maximum:

Twice the difference between these maxima is the likelihood-ratio statistic. Under the regular null conditions, its large-sample distribution is approximately chi-square, with one degree of freedom for each independent equality imposed. For example, imposing \(V_{\mathrm{fast}}=V_{\mathrm{slow}}\) and \(k_{a,\mathrm{fast}}=k_{a,\mathrm{slow}}\) imposes two constraints. This justification explains both why the statistic cannot be negative for correctly fitted nested models and why its reference distribution depends on the conditions, not merely on the parameter count.

Before trusting a sharing decision, check that both fits converged to credible maxima and that the more flexible fit did not simply reach extreme or poorly determined values. Compare estimates and uncertainty, not just the objective function. Inspect whether the shared model leaves a consistent class-specific pattern in concentration predictions or residuals, and whether relaxing the constraint resolves that pattern without creating implausible estimates elsewhere. If results change substantially with reasonable starting values or the estimated class-specific difference is highly uncertain, the data may not support a firm conclusion about sharing.

The main alternatives have different interpretations:

A common mistake is to reason, “The fast and slow classes are different, so every PK parameter must be class-specific.” That does not follow: a clearance-defined class distinction does not imply different volumes, absorption rates, or residual errors. The correction is to state the scientific hypothesis for each parameter, fit a defensible shared model, and relax individual constraints only when the assumptions and data support doing so. The reverse mistake—sharing every parameter because the model is simpler—can also conceal systematic class differences.

No single comparison settles the choice. Likelihood-based tests depend on their regularity assumptions; information criteria balance fit and complexity but do not establish biological truth; and inspection of fitted profiles is informative but should not replace a clearly stated model and uncertainty assessment. The best-supported sharing structure is conditional on the structural model, observation design, and available class information.