Weakly Identified Random-Effects Components

Why a Random-Effects Component Can Be Difficult to Estimate

A random-effects term describes variation among individuals, but its presence in a model does not guarantee that the observed data can distinguish it from other sources of variation. A variance component may be weakly identified when the likelihood changes little as that component changes, or when changes in it can be offset by changes in other parameters. In practice, the key question is not just whether the estimation algorithm converged. It is whether the individual-level observations contain a distinct signal for the component.

For individual \(i\), write the observations as \(y_i\), with a model \(y_i=f_i(\theta,\eta_i)+\epsilon_i\). Here \(\theta\) denotes fixed effects, \(\eta_i\) is a vector of individual random effects, and \(\epsilon_i\) is residual error. A common assumption is \(\eta_i\sim N(0,\Omega)\), with the random effects independent across individuals and residual errors conditionally independent given the model and random effects. The covariance matrix \(\Omega\) contains the random-effects variances and, when included, covariances. The model must also describe the mean response and residual variability adequately for likelihood-based evidence about \(\Omega\) to be trustworthy.

Weak identification is related to practical nonidentifiability, but this diagnosis focuses on how a particular random-effects component is supported by individual-level data and by the model’s structure. A small Hessian eigenvalue, high parameter correlation, or unstable covariance estimate may be clues; none alone explains which observations should have informed the component. That explanation comes from connecting the model’s sensitivity to the actual sampling patterns.

From Individual Sensitivity to Component Information

A useful local approximation linearizes the model’s response to a random effect around its population value. For component \(k\), let \(j_{ik}\) be the vector of derivatives of individual \(i\)’s predicted observations with respect to \(\eta_{ik}\). If observations are locally Gaussian, the marginal covariance of \(y_i\) can be approximated by

where \(J_i\) collects the individual sensitivities and \(R_i\) is the residual covariance matrix. When considering a single variance component \(\omega_{kk}\), its local covariance signature is approximately \(D_{ik}=j_{ik}j_{ik}^\mathsf{T}\): the way that component changes covariance among that individual’s observations.

For a Gaussian model with a parameter-dependent covariance \(V_i\) and a mean that does not depend on that parameter, the expected Fisher information for parameters \(a\) and \(b\) is

This expression is a local diagnostic, not an exact description of every nonlinear mixed-effects estimation method. It shows two important features. First, information depends on the sensitivity of each individual’s observations to the random effect. If the sensitivity is small, the component changes the data distribution very little. Second, the component must produce a covariance pattern that can be distinguished from residual error and from other random-effects components. If two components induce nearly the same pattern, their estimates can trade off.

The expression describes information before accounting for nuisance parameters. When fixed effects, residual parameters, and multiple covariance components are estimated together, their overlap can reduce the information available for a particular component. Thus, a raw sensitivity calculation is a useful screen, not proof of precise estimability. It should be considered alongside the fitted model, parameter correlations, curvature diagnostics, and the actual individual observation schedules.

Overlapping Covariance Patterns

Sampling can also make two components difficult to distinguish, even when each has some effect on the observations. Consider a random intercept and a random slope. Their covariance signatures for an individual with time vector \(t_i\) are proportional to \(\mathbf{1}\mathbf{1}^\mathsf{T}\) and \(t_i t_i^\mathsf{T}\), respectively. If observations occur in a narrow time window, \(t_i\) may be close to a constant multiple of \(\mathbf{1}\); the two signatures then become similar. More varied times can make their effects on the within-individual covariance more distinct.

The comparison concerns the timing pattern, not merely the number of observations. More observations at nearly identical times can add less distinct information than observations spanning a range of times. Likewise, repeated observations within individuals can help distinguish a persistent individual deviation from residual noise, but only if the model predicts different covariance patterns for those sources.

Check Individual-Level Support Directly

For a pharmacokinetic random effect, the relevant support depends on where that effect enters the prediction. As a simple illustration, suppose log concentration has a local dependence on a clearance random effect \(\eta_i\) of the form \(\log C_i(t)=m_i(t)-k t\eta_i+e_i(t)\), where \(k=0.2\) per hour, and independent log-scale residual errors have standard deviation \(0.1\). The derivative with respect to \(\eta_i\) is \(-kt\). With independent Gaussian errors, the conditional likelihood information about an individual’s realized \(\eta_i\) is the sum of squared derivatives divided by residual variance:

Diagnosis and Practical Responses

A component-level diagnosis should connect computational symptoms to the data and model structure. A nearly flat objective function, unstable estimates across starting values, large uncertainty, or strong parameter correlations may be consistent with weak identification. Check whether those symptoms persist after sensible estimation settings and convergence checks; a failed algorithm and a weakly informed component are not the same problem. Then ask which individuals and observations could have supplied information for that component.

Changing estimation settings can help locate a stable optimum, but it cannot create information absent from the observations. Likewise, deleting a component solely because its uncertainty is large may hide a design limitation or a misspecified competing term. Compare alternatives using the same data and a clear scientific rationale. A useful alternative may be to remove a poorly supported covariance while retaining a supported variance, or to use a simpler random-effects structure; whether that is defensible depends on the scientific question and model assumptions.

The local Gaussian information calculation has limitations. It relies on an approximation to a nonlinear model, and the relevant information metric changes when residual covariance, fixed effects, or other random-effects parameters are uncertain. Sparse data, influential individuals, model misspecification, and weak separation of residual and interindividual variation can all make a simple sensitivity calculation optimistic. Treat it as an explanation of possible information pathways, then check the fitted model’s uncertainty and stability.

Common Pitfall

Wrong reasoning: “The optimizer converged, so every random-effects variance is well identified.” Convergence means the algorithm met its numerical stopping conditions; it does not show that the likelihood contains a sharp, distinct signal for every component.

Correction: Examine the component’s individual-level sensitivities and whether its covariance signature is distinguishable from other terms. A stable optimizer can still return a weakly supported estimate, while a failed optimizer can result from numerical or implementation problems even when the data are informative. Use convergence diagnostics and information-based reasoning together.