Covariate Effects On Typical PD Parameters

Why Put a Covariate on a PD Parameter?

A population PD model describes both a shared response structure and differences among individuals. A measured characteristic may help explain why the typical value of a PD parameter differs across subjects. For example, a biomarker could be associated with the typical maximum drug effect, while subjects with the same biomarker value could still differ from one another. A useful model must represent these two sources of variation separately.

The distinction is important for interpretation and prediction. A covariate effect is a population-level relationship estimated from the data; it is shared by subjects with the same covariate values. A random effect describes an individual deviation from the covariate-conditioned typical value. It is not another name for the covariate effect, nor is it a known subject characteristic.

A general parameter-level model can be written as

Here \(h\) specifies how the random effect enters the individual parameter. For a positive parameter, one common specification is

This lognormal random-effects specification ensures \(\theta_i>0\). The exact covariate function and random-effects structure should be chosen for the parameter and scientific question; the equations here define the roles of the terms, not a universal functional form. The covariate function is a population model, while \(\eta_i\) varies by subject. In a model with several random effects, their joint distribution may include covariance, but each effect’s role remains an individual deviation.

Defining and Interpreting the Population Relationship

To put a covariate on a PD parameter, first name the parameter and define its covariate-conditioned typical value. Then specify the relationship using a population parameter such as a slope or contrast. For example, a model might write \(\theta_{\mathrm{typ}}(x_i;\beta)=\theta_{\mathrm{ref}}\exp(\beta x_i)\). In this expression, \(\theta_{\mathrm{ref}}\) sets the typical parameter at the model’s reference covariate value, and \(\beta\) controls how the typical value changes with \(x_i\). The function’s precise form, covariate coding, and interpretation depend on how \(x_i\) is defined.

The covariate coefficient is a fixed effect in the population model: it is one estimated value shared across subjects, not a new coefficient for every person. “Fixed” does not mean known without uncertainty. Its estimate has sampling uncertainty, and a fitted coefficient may be imprecise. By contrast, \(\eta_i\) is a subject-specific latent deviation described by a population distribution. An individual’s random effect is not generally observed directly.

A covariate relationship is conditional on the model and the observed data. It does not by itself establish that changing the covariate would cause a change in the PD parameter. The interpretation should state the modeled parameter, the covariate contrast or change, and the relevant units. For example, a relationship with typical \(E_{\max}\) concerns the model’s maximum-effect parameter; it is not automatically a statement that the covariate changes every observed response by the same amount.

Separating the Fixed Effect from Individual Deviations

For the lognormal random-effects model, the covariate relationship determines \(\theta_{\mathrm{typ}}(x_i;\beta)\). The random effect then scales that value for subject \(i\). When \(\eta_i=0\), the individual parameter equals the conditional typical value. When \(\eta_i\ne0\), the individual value differs, even though the subject has the same covariate value and the same population coefficient as everyone else.

A covariate on a parameter also differs from a covariate added directly to the response equation. The parameter-level model changes an input to the PD response function; the response change then depends on the response model and the other parameter values. This distinction matters when interpreting predictions.

Typical Value Is Not Always the Population Mean

The term “typical value” should not automatically be read as the arithmetic mean of individual parameters. With a lognormal random effect, \(\eta_i\) has mean zero on the log scale, but \(\exp(\eta_i)\) has an arithmetic mean greater than one when \(\omega^2>0\). Consequently, the conditional typical value at \(\eta_i=0\) is the median individual parameter under this model, not its conditional arithmetic mean.

For \(\eta_i\sim N(0,\omega^2)\), the normal moment-generating result gives \(E[\exp(\eta_i)]=\exp(\omega^2/2)\). Therefore,

This result assumes the stated lognormal model, with the random-effect variance \(\omega^2\) not varying with \(x_i\). If the variance depends on covariates, the conditional mean also depends on that variance relationship. More generally, the expected individual parameter depends on both the typical-value model and the random-effects distribution.

Assumptions, Diagnostics, and Alternatives

A covariate-on-typical-value model is interpretable only relative to its assumptions. The covariate must be defined for the subjects to whom the model is applied, and the specified function must make sense over the covariate range used for estimation and prediction. The population relationship is assumed to apply across the modeled subjects, while the random-effects distribution describes residual subject-to-subject variation after conditioning on covariates. The PD response model, observation model, and random-effects assumptions must also be adequate for the intended inference.

Diagnostics should check whether the fitted model leaves systematic structure unexplained. For example, plots of individual parameter estimates or suitable individual-level diagnostics against a covariate can reveal a residual trend after the covariate relationship is included. Such plots are exploratory: estimated individual effects are uncertain, and strong shrinkage can obscure patterns. Also check whether predictions and residuals show problems across covariate ranges or subgroups. A flat-looking diagnostic does not prove the relationship is correct, particularly when data are sparse or the covariate has little variation.

Warning signs include unstable covariate coefficients, implausible predicted typical values, a covariate effect that is supported by only a narrow data range, and systematic residual patterns that persist after fitting. A covariate may also appear to explain variation when the parameter itself is weakly identified by the PD observations. In that case, a numerical estimate alone is not evidence that the covariate relationship is reliably estimated.

The main alternative is to leave the typical parameter unrelated to the covariate and allow between-subject random effects to describe heterogeneity. That model does not estimate a systematic covariate relationship. Another distinct choice is to put the covariate directly into the response equation rather than on a PD parameter; this encodes a different scientific model and can imply different predictions. These alternatives are not interchangeable merely because they fit the observed responses similarly.

A common mistake is to interpret a subject’s fitted random effect as the covariate effect, or to treat a fixed covariate coefficient as if it described each subject’s deviation. The correction is to write the two levels separately: the covariates and \(\beta\) determine the conditional typical parameter, and \(\eta_i\) describes an individual deviation around it. A second pitfall is to call the typical value the population mean without checking the random-effects distribution; for a lognormal model, the arithmetic mean includes the factor \(\exp(\omega^2/2)\).