Random-Effects Correlations Near Plus or Minus One

Why a Near-Limit Correlation Needs Interpretation

A fitted random-intercept–random-slope model may report a correlation close to \(+1\) or \(-1\). That number can be a warning: it may indicate that the estimated random-effects covariance matrix is close to singular. But the correlation alone does not tell us why the estimate is near its limit, whether it is numerically stable, or how much information the data contain about it. In particular, a small eigenvalue and an unstable correlation are related clues, not interchangeable diagnoses.

The distinction matters in a generalized linear mixed model because the correlation describes the fitted joint distribution of group-level effects, not the correlation between observed outcomes. For a logistic model, for example, the random effects act on the linear predictor scale. A near-limit estimate should therefore be interpreted alongside the covariance matrix, the model’s parameterization, and fit diagnostics.

For the standard interpretation, the groups are independent draws from this bivariate normal random-effects distribution. Conditional on their random effects and covariates, responses are independent and follow the chosen GLMM response distribution with the specified link. These assumptions define what the fitted covariance and correlation refer to; a diagnostic cannot establish that the model assumptions are true.

The question here is narrower than whether the model is appropriate overall: does a near-limit estimate correspond to a covariance matrix close to singular, and is the data-supported estimate stable enough to interpret? A correlation reported to several decimal places can look precise while the underlying covariance direction is poorly determined.

Correlation, Determinant, and Eigenvalues

A covariance matrix must be positive semidefinite. For a two-dimensional random-effects covariance matrix with positive variances, its determinant is \(\tau_{00}\tau_{11}-\tau_{01}^2\). Substituting \(\tau_{01}=\rho\sqrt{\tau_{00}\tau_{11}}\) gives a useful identity:

Thus, when both variances are positive, \(|\rho|=1\) makes the determinant zero. The matrix then has rank one: one linear combination of the two random effects has zero variance. If \(|\rho|\) is close to one, the determinant is small relative to the product of the two marginal variances. This is the precise link between an extreme correlation and a nearly singular covariance matrix.

Eigenvalues give a scale-aware companion diagnostic. For a positive semidefinite \(2\times2\) matrix, the eigenvalues are nonnegative, their sum is the trace, and their product is the determinant. A very small eigenvalue relative to the larger one means that the estimated random effects vary very little in one particular linear combination. It is the eigenvalue—not the number of decimal places displayed for \(\rho\)—that directly describes that covariance direction.

The identity does not make every near-limit correlation equally informative. If a variance is extremely small, the covariance and correlation can be sensitive to modest changes in the estimate. And if a variance is exactly zero, the correlation formula has a zero denominator and is undefined. In that situation, a printed correlation may be absent or may reflect software-specific numerical handling; it should not be treated as an ordinary estimate of \(\rho\).

A Short Derivation: Why the Correlation Depends on the Predictor Origin

The random-intercept–random-slope correlation is not invariant to changing the origin of the predictor. This is important because two models with the same fitted linear predictor can report different intercept–slope correlations if one uses a centered predictor and the other does not.

Write the random contribution to the linear predictor as \(b_{0j}+b_{1j}x\). If the predictor is recentered as \(x_c=x-c\), then \(x=x_c+c\), so \[ b_{0j}+b_{1j}x=(b_{0j}+c b_{1j})+b_{1j}x_c. \] Consequently the new random intercept is \(b_{0j}^*=b_{0j}+c b_{1j}\), while the random slope remains \(b_{1j}\). In matrix form, this transformation and its effect on the covariance are

For \(G\) as defined above, direct multiplication gives \(\operatorname{Var}(b_{0j}^*)=\tau_{00}+2c\tau_{01}+c^2\tau_{11}\) and \(\operatorname{Cov}(b_{0j}^*,b_{1j})=\tau_{01}+c\tau_{11}\). The new correlation is therefore calculated from these transformed entries, not copied from the original correlation. Because \(\det(A)=1\), \(\det(G^*)=\det(G)\). Recentring can change the correlation substantially without changing whether the covariance matrix is singular.

Distinguishing a Boundary-Like Estimate from Weak Identification

A near-limit correlation can accompany a nearly singular covariance estimate, as the determinant identity shows. That is a statement about the fitted matrix. Weak identification is a different statement: the data and model provide little information to distinguish plausible values of one or more covariance parameters. A weakly identified correlation can move considerably across repeated starts, optimizers, or small reasonable changes in representation while the fit objective changes very little. Conversely, a stable near-singular estimate may be consistently supported by the likelihood, even though the estimated covariance has a small eigenvalue.

No single numerical threshold separates these cases. A value such as \(|\widehat\rho|>0.99\) is a useful prompt to inspect the fit, not a universal decision rule. The correlation scale, the eigenvalues relative to the covariance scale, optimizer and Hessian checks, and sensitivity to predictor centering all contribute evidence. A warning or convergence message should be interpreted with these checks rather than substituted for them.

Practical Diagnostics, Alternatives, and Limitations

When a random-intercept–random-slope correlation is close to either limit, use a sequence of checks rather than treating the correlation as a verdict. First, confirm that both estimated standard deviations are meaningfully above zero so the correlation is defined in the ordinary way. Next, inspect the covariance entries and eigenvalues, comparing the smallest eigenvalue with the largest and with the overall scale of \(G\). Then review optimizer convergence, scaled gradients, and Hessian behavior. These checks answer related but distinct questions: whether the solution appears numerically stationary, whether local curvature is informative, and whether the fitted covariance is nearly singular.

A further check is to refit the same model after a scientifically sensible recentering of the predictor. The fixed-effects interpretation changes with the reference point, and the random-intercept–slope correlation can change as well. The transformed covariance should be considered alongside the original one; the two reported correlations are not competing estimates of a single origin-free quantity. If the near-limit result or other diagnostics change markedly under reasonable representations, that sensitivity is useful evidence to report.

The main alternative to reading \(\widehat\rho\) alone is to examine the full covariance structure: its standard deviations, covariance, determinant, and eigenvalues. Optimizer and Hessian diagnostics add evidence about numerical and local curvature problems, while sensitivity checks address stability. None replaces the others. In particular, a well-behaved optimizer does not guarantee that the data identify the correlation precisely, and a very small eigenvalue does not by itself identify the cause.

These diagnostics have limits. Eigenvalues depend on the units and scaling of the random-effect coefficients, so comparisons should be made with scale in mind; centering changes the intercept coordinate as well as its correlation with the slope. Fit-to-fit comparisons are informative only when the models and data are comparable. Finally, a covariance estimate near its boundary is not automatically proof of software failure. It may be the estimated solution under the specified model, while still requiring cautious interpretation.