Categorical Covariates and Subgroup Exposure Comparisons

Why Categorical Covariates Need Careful Comparisons

Clinical development decisions often ask whether a dose or exposure is appropriate across categories such as age bands, renal-function groups, or formulation groups. A population PK model can estimate how a category relates to a parameter such as clearance, but the interpretation depends on the model scale and on which group is chosen as the reference. A parameter difference is not automatically an exposure difference of the same size, and neither statistical significance nor a subgroup average alone establishes clinical importance.

This tutorial focuses on estimating categorical covariate effects and translating them into exposure comparisons under explicit assumptions. It does not treat a covariate association as causal: the model describes differences conditional on its specification and data. The earlier material on functional forms for continuous covariates covers how continuous predictors can affect parameters; a categorical predictor instead represents a finite set of levels.

Reference-Group Coding and Parameter Interpretation

For a two-category covariate \(G_i\), let \(G_i=0\) denote the reference category and \(G_i=1\) the comparison category. A common model for clearance is

Here \(CL_{\mathrm{ref}}\) is the typical clearance in the reference category, \(\beta\) is the fixed categorical effect on log clearance, and \(\eta_i\) represents individual variation on the log scale. The typical clearance ratio for comparison versus reference is \(\exp(\beta)\). Thus \(\beta=0\) means equal typical clearances in this model, while \(\beta>0\) means the comparison category has higher typical clearance.

With \(K\) categories, choose one as reference and use \(K-1\) indicator variables. For example, with categories A, B, and C, the typical clearance model can be written \(CL_i=CL_A\exp(\beta_B I_{Bi}+\beta_C I_{Ci})\exp(\eta_i)\). The contrasts are \(\exp(\beta_B)=CL_B/CL_A\) and \(\exp(\beta_C)=CL_C/CL_A\). The contrast between B and C is not either coefficient by itself: it is \(\exp(\beta_C-\beta_B)\) for C versus B.

A model's reference group is a coding choice, not a biological claim that the group is normal or intrinsically preferable. Choosing a different reference changes coefficient values and their interpretation, but it does not change fitted group predictions or pairwise contrasts when the model is re-expressed correctly.

From a Parameter Contrast to an Exposure Contrast

For linear PK after a single dose, the AUC relationship depends on route and bioavailability. For an extravascular dose, \(AUC=F\,\mathrm{Dose}/CL\), where \(F\) is bioavailability; for an IV dose, \(F=1\). Therefore, comparing groups on the same dose and route gives

If bioavailability is the same in both categories, it cancels in this ratio. If bioavailability differs, the clearance contrast alone does not determine the AUC contrast. This relationship assumes linear PK over the dose range, the same dose and dosing conditions, and comparable definitions of the exposure metric. For an IV dose, the ratio simplifies to the inverse clearance ratio.

Predicted Exposure Distributions, Not Just Typical Values

A population model can represent both a typical group value and variation among individuals. Under the lognormal clearance model above, if \(\eta_i\sim N(0,\omega^2)\), individual AUC after an IV dose \(D\) is \(D/(CL_{\mathrm{typ}}\exp(\eta_i))\). Its log is normally distributed with standard deviation \(\omega\). The typical AUC, obtained at \(\eta_i=0\), is the median of this individual distribution, not its arithmetic mean.

If the same \(\omega\) applies to both categories, their within-group distributions can have the same relative spread but different locations. A central interval can be calculated from the normal quantiles for \(\eta_i\), then transformed to the AUC scale. Such a prediction interval describes modeled individual variability conditional on the parameters; it is distinct from a confidence interval for a group contrast, which reflects uncertainty in estimated parameters. A full predictive assessment can also account for parameter uncertainty, but the calculations below isolate individual variability to make the distinction clear.

Contrasts Across More Than Two Categories

With three or more levels, all comparisons should be stated explicitly. A reference-coded model provides contrasts with the selected reference, while contrasts between nonreference groups require subtracting their log-scale coefficients. Their uncertainty depends on the covariance between estimates, so it cannot generally be obtained by treating the coefficients as independent. Reparameterizing the same fitted model with a different reference is often useful when a clinically relevant comparison is not against the original reference.

Assumptions, Diagnostics, and Common Pitfalls

A categorical effect is interpretable only to the extent that its coding, model scale, and data support are clear. Before interpreting a contrast, confirm the category definitions, the reference level, the parameter affected, and whether the reported coefficient is on a log scale or another scale. For exposure translation, state dose, route, bioavailability assumptions, and the PK relationship being used. A simple inverse-clearance relationship is not appropriate if the relevant PK is nonlinear or if bioavailability differs across categories.

Useful checks for this analysis include verifying category counts and coding against the analysis dataset, inspecting whether each group has adequate observations and concentration-time information, and comparing model predictions with observations within each category. Examine whether residual patterns or systematic prediction errors are concentrated in one group. Check that uncertainty in the contrast is reported, especially for small or imbalanced categories. These checks can reveal an unstable estimate, coding error, or model mismatch; they do not by themselves establish that an observed association is causal.

• Wrong: “The coefficient is 0.20, so exposure is 20% higher.” Correction: If the coefficient is on log clearance, the clearance ratio is \(\exp(0.20)\), and under equal-dose IV conditions the AUC ratio is \(\exp(-0.20)\), not \(\exp(0.20)\).
• Wrong: “The 90% individual prediction intervals overlap, so the groups have the same typical exposure.” Correction: Prediction intervals describe individual variability; estimate and quantify the uncertainty of the group contrast separately.
• Wrong: “A statistically significant covariate effect proves that a clinically important dose change is needed.” Correction: Statistical evidence concerns compatibility with a null contrast under the model. Clinical importance requires contextual interpretation against relevant efficacy, safety, and decision criteria.
• Wrong: “The reference group is the untreated or normal group.” Correction: A reference is a coding choice. Name it and define it from the actual categories rather than implying a clinical status it may not have.

The main alternatives to reference-group reporting are to present all pairwise contrasts or to re-express the model using another reference. These are different presentations of the same fitted categorical effects when the model is unchanged. A descriptive, unadjusted subgroup comparison may be easier to understand, but it does not generally answer the same question as a model-based contrast adjusted for included covariates. Conversely, a model-based result relies on its specification and data support. Avoid interpreting either presentation as a causal effect without the assumptions and design needed for that claim.