Fixed Effects and Random Effects in Parameter Models

Why Separate Population Values From Individual Differences?

A pharmacokinetic model may describe concentration over time using parameters such as clearance and volume. Even when subjects receive the same regimen, their concentration profiles need not be identical. A nonlinear mixed-effects model represents this structure by separating population-level parameter values from subject-level departures. This distinction is central: the population values are treated as fixed unknown quantities, while the departures are represented as random variables.

The earlier material on the three levels of the NLME hierarchy established the roles of the structural model, parameter model, and residual model. Here the focus is the parameter model: what its fixed effects mean, what its random effects mean, and what assumptions make the distinction useful. The exact function that maps population values and random effects to individual parameters is a separate modeling choice; it is not determined merely by calling one part fixed and the other random.

The notation \(\theta_i=g(\theta,\eta_i)\) is deliberately general. It says that individual parameters depend on population quantities and on subject-level random effects, without specifying how they are combined. Different models use different mappings, and those choices can change the scale and interpretation of \(\eta_i\). Therefore, the random effect is not automatically a difference measured in the units of the pharmacokinetic parameter.

A common distributional assumption is that the random-effects vector for each subject follows the same distribution across the population, often a multivariate normal distribution with mean zero and covariance matrix \(\Omega\):

This assumption says that the model assigns a population distribution to the latent deviations. The zero mean centers that distribution; it does not say that every subject has \(\eta_i=0\). The diagonal elements of \(\Omega\) are random-effect variances, and its off-diagonal elements are covariances. The fixed effects \(\theta\) and distributional parameters such as \(\Omega\) are unknown population quantities estimated from data. They are not subject-specific random effects.

How the Two Parts Enter the Hierarchy

Let \(y_{ij}\) denote an observation for subject \(i\) at observation occasion \(j\). A structural model predicts that observation from time and the subject-specific parameter vector. A schematic form is

Here \(f\) is the structural model, \(t_{ij}\) is the observation time, and \(\epsilon_{ij}\) represents residual variation not explained by the structural prediction. The fixed effects determine the population-level parameter structure; \(\eta_i\) allows the parameters used for subject \(i\) to differ. Repeated observations from one subject share the same \(\eta_i\) in this basic formulation, while their residual errors may differ across observations.

For the usual hierarchical interpretation, subjects are treated as draws from the target population, with random effects that follow the specified distribution and are independent across subjects. The residual model must also be specified; commonly, residual errors are assumed conditionally independent across observations given the subject parameters, though other residual structures can be used. These are modeling assumptions, not consequences of the labels “fixed” and “random.” They should be checked for scientific and statistical plausibility.

The population distribution links the subject-level and population-level parts. By the law of total probability, the probability of one subject's observed data is obtained by averaging its conditional probability over possible random effects:

The first factor inside the integral describes the data conditional on the subject's random effects and the fixed effects. The second describes how likely those random effects are in the population. Integrating combines the two: the model does not treat each subject's latent deviation as an unrelated fixed parameter, but accounts for the population distribution from which deviations are assumed to arise. This marginalization is one reason nonlinear mixed-effects estimation can be computationally demanding.

Variance and Covariance Describe the Population Spread

When a model includes more than one random effect, \(\Omega\) describes both the spread of each component and how components vary together. For example, let \(\eta_{i1}\) and \(\eta_{i2}\) be two random effects associated with two parameters. A covariance matrix must be symmetric and positive semidefinite; in ordinary cases with nonzero variance, it can be written as follows.

The variances are \(\omega_{11}\) and \(\omega_{22}\). The covariance \(\omega_{12}\) describes the direction and magnitude of joint variation on the random-effects scales. Because covariance depends on scale, the correlation is often easier to compare across components:

Why Zero Mean Does Not Mean Zero Variability

A zero-mean assumption is about the center of a population distribution, not about each subject. A small hypothetical set of four random-effect values makes the distinction concrete. Suppose the values are \(-0.30\), \(-0.10\), \(0.10\), and \(0.30\), all on the same random-effect scale. Their mean is

Yet the values are not all zero. Their sample variance is

This calculation is descriptive for these four values; it is not generally a substitute for estimating \(\Omega\) in an NLME model. In fitted models, subject-level random-effect estimates are informed by each subject's data and by the population distribution, so their empirical spread need not equal the estimated population variance. The example nevertheless shows the basic point: deviations can average to zero while varying substantially across subjects.

Assumptions, Diagnostics, and Alternatives

Interpreting fixed and random effects requires more than fitting an equation. The population structure should be appropriate for the target population, the random-effects distribution should be a reasonable approximation, and the data should contain enough information to distinguish population parameters from individual variation and residual error. In particular, a random-effect variance is difficult to estimate when the study design provides little information about between-subject differences.

Useful warning signs include variance estimates close to zero, unstable estimates across reasonable model fits, correlations near \(1\) or \(-1\), and uncertainty so large that very different values remain plausible. These patterns may indicate limited information, redundant random effects, or an unsuitable model structure. They are not automatic proof that a random effect is absent or that the model is wrong. Check whether the model's predictions and residual patterns are adequate, whether the parameterization is scientifically sensible, and whether the data support the complexity being estimated.

A model with no random effects treats the parameter values as common across subjects, apart from any explicitly modeled differences and residual error. It is simpler, but it cannot represent subject-level parameter heterogeneity through \(\eta_i\). At the other extreme, estimating a separate unrestricted parameter for every subject does not impose a shared population distribution and can use many parameters. A mixed-effects model lies between these approaches: subjects have latent departures, while those departures are modeled as draws from a common distribution. This pooling can stabilize estimates, but depends on the distributional and structural assumptions being adequate.

The fixed/random distinction also has limits. It does not guarantee that the chosen population distribution is correct, that all relevant sources of variability have been represented, or that population estimates will be precise. Nor does a fitted random-effects covariance establish a biological mechanism. It summarizes variability under the specified model. A sound interpretation therefore names the fixed-effect quantity, the random-effect scale, the distributional assumption, and the parameter mapping rather than treating “fixed” and “random” as self-explanatory labels.