Covariates on Metabolizer-Class Membership

Why Model Class Membership With Covariates?

A two-class NLME model allows individuals to arise from different clearance subpopulations, such as fast and slow metabolizers. Class membership is latent: concentration data inform which class is more plausible for an individual, but the model does not treat that individual’s class as directly observed. A measured covariate such as genotype can provide additional information by explaining why the prevalence of the fast class may differ among individuals.

The key modeling distinction is which part of the model the covariate changes. A genotype effect on class membership changes the probability of drawing an individual from the fast rather than the slow class. A genotype effect on clearance within a class changes the clearance distribution among individuals who belong to that class. These are different mechanisms and have different interpretations, even if both can produce differences in average clearance between genotype groups.

Definition: Let \(Z_i=1\) indicate that individual \(i\) belongs to the fast-clearance class, and let \(Z_i=0\) indicate the slow-clearance class. The covariate-dependent fast-class membership probability is \(\pi_i=P(Z_i=1\mid x_i)\), where \(x_i\) is a measured covariate. A logistic membership model specifies \(\operatorname{logit}(\pi_i)=\log\{\pi_i/(1-\pi_i)\}=\beta_0+\beta_1x_i\). Thus \(1-\pi_i\) is the probability of slow-class membership.

Logistic Membership Model and Interpretation

For a binary genotype coded \(x_i=0\) for one genotype group and \(x_i=1\) for the other, the model gives

\pi_i=\frac{\exp(\beta_0+\beta_1x_i)} {1+\exp(\beta_0+\beta_1x_i)}.

Here \(\beta_0\) is the fast-class log odds for the reference group, and \(\beta_1\) is the difference in fast-class log odds between the \(x_i=1\) group and the reference group. Exponentiating the coefficient gives an odds ratio: \(\exp(\beta_1)\) is the ratio of fast-to-slow membership odds for \(x_i=1\) compared with \(x_i=0\). It is not a ratio of probabilities.

For a continuous covariate, the same equation can be used with \(x_i\) measured in its stated units. A one-unit increase multiplies the fast-to-slow odds by \(\exp(\beta_1)\), provided the linear log-odds relationship is an adequate description over the covariate range. Centering a continuous covariate changes the interpretation of the intercept: it becomes the log odds at the centering value rather than at zero. It does not change the fitted probabilities if the intercept is adjusted consistently.

Worked Example: Genotype and Fast-Class Probability

Suppose a hypothetical model uses \(x=0\) for genotype group A and \(x=1\) for genotype group B, with \(\beta_0=-1.20\) and \(\beta_1=1.10\). Calculate the fast-class probabilities and interpret the coefficient.

For group A, the linear predictor is \(-1.20+1.10(0)=-1.20\). Therefore,

\pi(0)=\frac{\exp(-1.20)}{1+\exp(-1.20)} =\frac{0.3011942}{1+0.3011942} \approx 0.2315.

The model-implied fast-class probability for group A is about 23.15%. For group B, the linear predictor is \(-1.20+1.10(1)=-0.10\), so

\pi(1)=\frac{\exp(-0.10)}{1+\exp(-0.10)} =\frac{0.9048374}{1+0.9048374} \approx 0.4750.

The estimated odds ratio is \(\exp(1.10)\approx 3.004\). Under this model, group B has about 3.004 times the fast-to-slow membership odds of group A. This does not mean its fast-class probability is three times as large: the probabilities are approximately 0.4750 and 0.2315, respectively. These are model-based probabilities, not proof that genotype causes class membership.

How Membership Probabilities Enter the NLME Likelihood

A covariate-dependent membership model changes the weights in the two-class likelihood. Let \(y_i\) denote all concentration observations for individual \(i\), and let \(L_{i,F}\) and \(L_{i,S}\) be the likelihoods of those data under the fast- and slow-class models, respectively. These likelihoods include the relevant within-class parameter and random-effects structure. Conditional on \(x_i\), the individual’s contribution is the weighted sum of the class-specific likelihoods:

L_i=\pi_iL_{i,F}+(1-\pi_i)L_{i,S}.

This equation sums over the unobserved class rather than choosing a class first and treating that choice as known. The covariate informs the prior class weights \(\pi_i\) and \(1-\pi_i\); the concentration data inform how well each class-specific model explains \(y_i\). Both sources of information contribute to the likelihood.

Worked Example: Combining Membership and Concentration Evidence

Consider a hypothetical individual in genotype group A, with \(\pi_i=0.2315\). Suppose the likelihood of that individual’s concentration data is \(L_{i,F}=0.0800\) under the fast-class model and \(L_{i,S}=0.2000\) under the slow-class model. Calculate the mixture likelihood contribution and the fast-class probability after accounting for the observed data.

The individual likelihood contribution is

$$ L_i=(0.2315)(0.0800)+(1-0.2315)(0.2000) =0.01852+0.15370 =0.17222. $$

The prior membership probability is 0.2315, but the concentration data favor the slow-class likelihood, which is larger. The probability of fast-class membership conditional on the covariate and these data is

P(Z_i=1\mid x_i,y_i) =\frac{\pi_iL_{i,F}}{\pi_iL_{i,F}+(1-\pi_i)L_{i,S}} =\frac{0.01852}{0.17222} \approx 0.1075.

Thus the data-informed fast-class probability is about 10.75%, lower than the genotype-based probability of 23.15%. The likelihood contribution itself is not a class probability; it is the weighted likelihood used to fit the model.

Membership Effects Are Not Within-Class Clearance Effects

A useful way to keep the two roles separate is to write both parts of the model. For example, suppose \(\pi_i\) follows a logistic regression, while clearance conditional on class \(k\) follows

CL_{i,k}=\theta_{CL,k}\exp(\eta_{CL,i,k}), \qquad k\in\{F,S\}.

In this form, the covariate changes class prevalence if it appears in the equation for \(\pi_i\). It changes clearance within class if it also appears in the equation for \(CL_{i,k}\), for example as a multiplicative covariate term. A model may include either effect or both, but each additional effect needs adequate data support and a clear scientific interpretation.

Worked Example: Different Prevalence, Same Class-Specific Clearances

Use the probabilities from the first example. Suppose, hypothetically, the typical clearance is 12 L/h in the fast class and 5 L/h in the slow class for both genotype groups. There is no genotype effect on clearance conditional on class. Calculate the mixture-weighted typical clearance for each genotype group, and identify what differs.

For group A, the mixture-weighted value is

(0.231475)(12\ \mathrm{L/h})+(1-0.231475)(5\ \mathrm{L/h}) =2.7777+3.8426 \approx 6.620\ \mathrm{L/h}.

For group B, it is

(0.475021)(12\ \mathrm{L/h})+(1-0.475021)(5\ \mathrm{L/h}) =5.7003+2.6249 \approx 8.325\ \mathrm{L/h}.

These weighted values differ because the groups have different probabilities of belonging to the fast class. The class-specific typical clearances, 12 L/h and 5 L/h, are unchanged between groups. The weighted calculation illustrates the prevalence mechanism; it is not a substitute for the full NLME prediction, which may also account for random effects and the observation model.

Assumptions, Checks, and Alternatives

The logistic membership specification is useful under identifiable conditions. The class labels and class-specific likelihoods must be defined consistently; each individual’s covariate must be measured and coded correctly; and the log odds should be adequately represented by the chosen covariate terms. The full model also relies on its stated within-class structural, random-effects, and residual assumptions. A membership equation cannot compensate for a misspecified concentration model.

Before interpreting an estimated covariate effect, check the coding and the range of the covariate, calculate fitted probabilities across that range, and verify that they remain between 0 and 1 as required. For a categorical genotype, inspect how many individuals are represented in each category. Very sparse categories may provide little information about a membership effect, especially if their concentration data also weakly distinguish the classes. Compare the predicted prevalence pattern with the observed covariate distribution and examine uncertainty in the coefficient and predicted probabilities. Extreme fitted probabilities or a large coefficient with high uncertainty can signal limited information or poor support; they are not, by themselves, proof of a particular failure.

Modeling choice	What the covariate changes	Interpretation
Membership model	\(\pi_i\), the probability of fast-class membership	Changes fast-to-slow class odds across covariate values
Within-class clearance model	\(CL_{i,k}\), conditional on class \(k\)	Changes clearance among individuals in the same class
Both components	Membership probabilities and class-conditional clearance	Allows both mechanisms, if the data support their separate estimation
No covariate effect	Neither component	Uses a common class prevalence and no modeled clearance association with that covariate

The alternative to a logistic membership model may be a constant class prevalence, which sets \(\pi_i=\pi\) for all individuals. This is simpler and appropriate when no covariate effect on prevalence is being modeled. Alternatively, a covariate can be placed only in the within-class clearance model if the scientific question concerns conditional clearance differences rather than class prevalence. These choices are not interchangeable: fitting a prevalence effect does not estimate a within-class clearance effect, and vice versa.

A common mistake is to see higher average clearance in one genotype group and conclude that genotype raises clearance within each class. As the worked calculation shows, average clearance can differ solely because the class proportions differ. The correction is to state explicitly whether the proposed effect operates on \(\pi_i\), on class-conditional clearance, or on both, and to interpret the fitted component accordingly. Another pitfall is to call \(\exp(\beta_1)\) a probability ratio. It is an odds ratio; predicted probabilities must be calculated from the logistic equation.

Key takeaway: A covariate in the logistic membership equation changes who is more likely to belong to the fast class. A covariate in a class-specific clearance equation changes clearance conditional on class. Keep these model components distinct when fitting, checking, and reporting results.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In a logistic membership model, what does \(\exp(\beta_1)\) represent for a binary covariate coded 0 and 1?
If \(\beta_0=0\) and \(\beta_1=0\), what is the fast-class probability for either covariate value?
Can two groups have different mixture-weighted typical clearances even if their class-specific typical clearances are identical? State why.
For an individual, why is \(\pi_iL_{i,F}+(1-\pi_i)L_{i,S}\) used instead of assigning the individual to the class with the larger likelihood?
What does it mean if a covariate is included in both the membership and the within-class clearance equations?

Solutions. (1) It is the ratio of fast-to-slow membership odds for covariate value 1 relative to value 0, holding the other modeled terms fixed. (2) The linear predictor is zero, so \(\pi=1/(1+\exp(0))=0.5\) for either value. (3) Yes. Different class probabilities produce different weighted averages even when fast- and slow-class clearances themselves are the same across groups. (4) Class membership is latent, so the likelihood sums the two class-specific likelihoods weighted by their membership probabilities rather than treating one class as observed. (5) The model allows the covariate to affect both the probability of class membership and clearance conditional on class; those are separate effects and require separate interpretation.