Fixed Effects And Random Effects In Virtual Populations

Why Separate Typical Behavior From Individual Variability?

A population pharmacometric model can describe both a typical response and differences among individuals. These are related, but not interchangeable, parts of the model. Fixed effects specify the shared structure and typical parameter values for a defined scenario. Random effects represent modeled differences among individuals around that typical behavior. A virtual population that includes random effects therefore generates a distribution of individual predictions, not merely repeated copies of the typical prediction.

This distinction matters when a simulation is used to compare regimens, describe variability in exposure, or predict the fraction of virtual patients above a threshold. A fixed-effects-only run can be useful for a typical-subject result, but it cannot by itself represent between-subject variability. Conversely, drawing random effects does not mean that the fixed effects have also been varied to represent uncertainty in their estimates.

A Standard Population-Model Representation

For a positive pharmacokinetic parameter such as clearance, a common specification is

Here \(TVCL\) is the typical clearance for the reference covariate values, and \(\omega_{CL}^2\) is the variance of the clearance random effect on the log scale. The random effect is centered at zero; consequently, when \(\eta_{CL,i}=0\), the individual's clearance equals \(TVCL\). Because exponentiation keeps clearance positive, this specification produces a lognormal distribution of individual clearances. \(TVCL\) is the median, also the geometric mean, of that distribution when covariates are held at their reference values; it is not generally its arithmetic mean.

Fixed effects can also determine how typical behavior changes with covariates. For example, a model might specify \(TVCL_i=TVCL(WT_i/70)^{\theta_{WT}}\), where \(WT_i\) is body weight in kilograms and \(\theta_{WT}\) is a shared fixed-effect coefficient. The individual parameter then becomes \(CL_i=TVCL_i\exp(\eta_{CL,i})\). The covariate term shifts the typical value for a patient with those covariates; the random effect describes the patient's deviation from that covariate-adjusted typical value.

These statements rely on a specified model and a specified random-effects distribution. In the standard example, random effects have mean zero and variance \(\omega_{CL}^2\); they are sampled across subjects according to that distribution, and covariates are assigned according to the simulation target. If several random effects are included, their joint distribution and any modeled correlations must also be respected. A particular simulation can use other distributional assumptions, but the resulting virtual population represents those assumptions, not a distribution-free set of patients.

What Random Effects Add to a Virtual Population

A virtual population is produced by applying the model to a set of simulated individuals. Each individual receives covariates and, when the target includes between-subject variability, random-effect draws. Holding the fixed effects constant while drawing different random effects gives individuals different parameters under the same population model. If random effects are omitted or set to zero, the resulting subjects share the same random-effect component, even if their covariates differ.

The distinction between a typical individual and a random individual is especially clear for the exponential model. The typical individual has \(\eta_{CL,i}=0\) and clearance \(TVCL\). A simulated individual instead receives a draw from the modeled distribution. Most draws may be near zero, but some are farther away; the frequency and magnitude of those deviations depend on \(\omega_{CL}^2\).

Why the Mean of Individual Predictions Can Differ

The individual random effect enters the parameter through a transformation, and that parameter in turn enters a prediction. In general, the average of predictions from simulated individuals need not equal the prediction calculated at the typical parameter. A short derivation shows this even for clearance alone. If \(\eta\sim N(0,\omega^2)\), the normal moment-generating function gives \(E[\exp(\eta)]=\exp(\omega^2/2)\). Therefore,

which exceeds \(TVCL\) whenever \(\omega^2>0\). The derivation assumes the stated normal distribution and the exponential parameterization. It does not say that \(TVCL\) is incorrectly specified: \(TVCL\) is the typical value on the model's scale, while the arithmetic population mean is a different summary.

Implementation Checks and Common Failure Modes

Before interpreting a virtual-population result, check that the implemented model generated the kind of individuals the simulation target requires. The earlier material on conditional and unconditional simulation targets distinguishes targets that condition on individual effects from those that average over them. In this tutorial's terms, the key practical question is whether the random effects were included, and whether they were sampled or fixed at specified values for the target at hand.

• Check the scale of the random effect. In \(CL_i=TVCL\exp(\eta_i)\), \(\eta_i\) is on the log scale. Adding \(\eta_i\) directly to clearance, or exponentiating a variance rather than a random-effect draw, changes the model.
• Check the distribution and spread. Compare simulated random effects with their intended mean, variance, and distribution. For a finite sample, the empirical mean and variance will not be exactly their theoretical values, but large discrepancies can indicate a coding, scaling, or sampling error.
• Check the parameter transformation. Positive parameters modeled exponentially should remain positive. Extremely large or small values may be possible under a broad distribution; inspect their frequency and consider whether the assumed distribution is scientifically plausible.
• Check the covariate reference and units. A fixed covariate effect is applied at the covariate values specified by the model. Incorrect units or an unintended reference value can shift all typical parameters before random effects are applied.
• Check whether repeated observations share the subject effect. In a model with subject-level random effects, repeated records for one subject generally use that subject's same draw, rather than resampling a new subject effect at every observation.

A frequent mistake is to call \(TVCL\) the arithmetic mean clearance and then treat a simulated population's higher arithmetic mean as evidence that the simulation is wrong. That conclusion does not follow: for the exponential-normal model, the arithmetic mean is \(TVCL\exp(\omega^2/2)\), whereas \(TVCL\) is the median. The correction is to name the summary precisely and compare like with like.

Choosing the Representation for the Question

A fixed-effects-only prediction is appropriate when the intended result is the model's typical behavior at specified covariates. A virtual population with random effects is appropriate when the target requires between-subject variability, such as a distribution of exposures or the proportion of individuals meeting a criterion. These approaches are not competing ways to estimate the same quantity: they answer different simulation questions.

A more flexible alternative to a parametric random-effects distribution is to use a set of empirical individual effects, when such effects are available and the analysis design justifies their use. This can preserve features that a chosen normal distribution does not represent, but it does not remove the need to define the target population and assess representativeness. A further limitation applies to either approach: random effects describe variability under the specified model and distribution. They do not, by themselves, represent uncertainty in the estimated fixed effects; that is a separate simulation component.

Thus, simulation results should be reported with enough detail to identify the fixed-effect values, covariate scenario, random-effects specification, and whether the reported quantity describes a typical individual or a summary over simulated individuals. A precise description makes the result interpretable and helps distinguish model behavior from choices made in constructing the virtual population.