Binary Covariates as Multiplicative Parameter Effects

Why express a binary covariate as a ratio?

A binary covariate divides individuals into two groups, such as treatment versus control or impaired versus unimpaired renal function. In a nonlinear mixed-effects model, the covariate can alter a positive structural parameter by multiplying its reference value. This is useful when a proportional change is more plausible than an equal absolute change: an increase of 1 L/h in clearance, for example, has a different relative meaning at 2 L/h than at 20 L/h.

The key distinction is between a change in a model parameter and a change in the response predicted by the model. A parameter multiplier is an exact ratio for that parameter. The corresponding response ratio depends on how the parameter enters the structural model and can vary with time, dose, and other parameter values.

The parameter must be positive for this exponential form, which is natural for quantities such as clearance, volume, and rate constants. With \(\eta_i=0\), the typical parameter is \(P_{\mathrm{ref}}\) when \(z_i=0\), and \(P_{\mathrm{ref}}\exp(\beta)\) when \(z_i=1\). Thus, \(\beta\) itself is a log ratio, not a difference in the original parameter units.

Deriving the log-scale coefficient

First consider the typical parameter, setting the individual random effect to zero. Denote the typical parameters in the two groups by \(P_0\) and \(P_1\). Substitution of the two binary values gives

Dividing the comparison-group parameter by the reference-group parameter cancels the reference value. Taking logarithms then isolates the coefficient:

This is the central interpretation: exponentiating the coefficient gives a ratio, while the coefficient is the natural logarithm of that ratio. A ratio above 1 means the typical parameter is larger in the comparison group; a ratio below 1 means it is smaller. The percentage change relative to the reference group is \(100\{\exp(\beta)-1\}\%\). A negative coefficient therefore corresponds to a decrease, not a negative parameter.

When \(\eta_i\sim N(0,\omega^2)\), the parameter is lognormally distributed. The typical parameter \(P_{\mathrm{ref}}\exp(\beta z_i)\) is its median, not its arithmetic mean. The arithmetic mean is \(P_{\mathrm{ref}}\exp(\beta z_i+\omega^2/2)\); if the same \(\omega^2\) applies to both groups, its comparison-to-reference ratio is still \(\exp(\beta)\). These statements concern the parameter distribution. They do not imply that the mean of a nonlinear response is obtained by simply substituting the mean parameter into the structural model.

Worked interpretations and prediction consequences

A fitted coefficient is often reported with uncertainty. Because the coefficient is on the log-ratio scale, an interval for \(\beta\) should be exponentiated at both endpoints to obtain an interval for the parameter ratio. The resulting ratio interval is generally asymmetric, even when the coefficient interval is symmetric.

A parameter ratio is not automatically the ratio of predicted concentrations, effects, or other responses. The response function may depend nonlinearly on the parameter. The following calculation checks this distinction using a simple one-compartment intravenous bolus model with elimination.

Model checks, alternatives, and limitations

A multiplicative binary effect is a modeling choice, not a consequence of having two groups. It is appropriate when a constant parameter ratio is scientifically plausible and the parameter is positive. The fitted model should still be checked against the data and the model assumptions. Useful checks include:

• Compare observed and model-predicted profiles separately by covariate group. Systematic overprediction in one group and underprediction in the other can indicate a misspecified group effect.
• Inspect residual patterns across time, dose, and the binary groups. A group-specific pattern may point to an effect on another parameter or to a structural-model problem rather than the assumed multiplicative effect.
• Check whether the parameter and coefficient are estimated with reasonable precision. Sparse data in one group, little overlap in dosing or covariates, or strong parameter correlations can make the ratio unstable.
• Assess whether the random-effect distribution is plausibly comparable between groups. A fixed-effect ratio alone does not describe differing variability or group-specific random-effect distributions.
• Check the coding and units. The reference group must actually be coded 0, and both group parameter values must refer to the same parameter and units.

There are alternatives. An additive effect \(P_i=P_{\mathrm{ref}}+\delta z_i\) represents a constant difference \(\delta\) in parameter units; it may be reasonable when absolute differences are the scientific target, but it can predict nonpositive values for a positive parameter. Separate group-specific parameter estimates allow the group values to be unrelated, but use more parameters and may be imprecise with limited data. A multiplicative model is parsimonious because it uses one coefficient to express a relative shift. The earlier material on exponential effects of continuous covariates uses the same log-scale logic, but here the binary indicator gives exactly two group values rather than a continuously varying trend.

The binary model also has limits. It assumes the same ratio for every individual, conditional on the modeled random effects, and does not by itself explain why groups differ. An estimated covariate effect is an adjusted model comparison, not proof that group membership causes a parameter change. If the binary covariate was created by dividing a continuous measurement into categories, the resulting two-group effect also discards information within each category. Finally, a good parameter-level fit does not guarantee that every prediction or clinical consequence has the same proportional change; calculate predictions from the complete structural model for the conditions of interest.