Why Look Beyond the Convergence Message?
A binary GLMM optimizer can report convergence even when a fixed-effect estimate is moving toward infinity. Under separation, the likelihood can approach a finite supremum while the coefficient grows without bound; consequently, the optimizer may take increasingly small steps and report a small gradient. A small score alone is therefore not reassurance. A useful screen examines the score, the curvature of the likelihood, and the coefficient path together.
These diagnostics apply to a fitted binary GLMM with a finite, identifiable parameter point and a correctly specified likelihood calculation. They are numerical warning signs, not formal tests for separation. Interpret them alongside the outcome patterns and separation concepts established in the earlier material on binary GLMM separation. In particular, do not confuse a separation screen with a convergence diagnosis: optimizer failure, weak identification, and boundary estimates can produce some of the same numerical symptoms.
Score and Information in a Binary GLMM
Let \(Y_{ij}\) be a binary outcome for observation \(i\) in cluster \(j\). Conditional on the random effects \(b_j\), suppose \(Y_{ij}\) is Bernoulli with probability \(p_{ij}\), where
Here \(x_{ij}\) and \(z_{ij}\) are the fixed- and random-effect design vectors, \(\beta\) is the fixed-effect coefficient vector, and \(D\) is the random-effect covariance matrix. The marginal likelihood integrates over the random effects. Write its log likelihood as \(\ell(\beta,\psi)\), where \(\psi\) denotes the random-effect and any other nuisance parameters.
For a finite, regular maximum-likelihood estimate, the score is zero at an interior optimum. In a GLMM, both the score and information refer to the marginal likelihood, not merely to the conditional likelihood with estimated random effects plugged in. A convenient identity clarifies the distinction. Given the observed outcomes, average the conditional fixed-effect score over the posterior distribution of \(b\):
This is Fisher's identity. It shows why conditional residual contributions are informative, but not by themselves the marginal score: the latter averages over uncertainty in the random effects. Likewise, a conditional working-weight matrix can help locate weak-curvature directions, but it is not generally the observed information of the integrated GLMM likelihood.
The coordinate values of a score depend on coefficient units. For screening, inspect scaled components such as \(U_k/\sqrt{J_{kk}}\) when \(J_{kk}>0\), or score components in a suitably scaled eigenvector direction. If nuisance parameters are estimated jointly, use the joint observed information or the appropriate profiled curvature; a fixed-effect block alone can miss dependence on nuisance parameters.
Why Separation Weakens Curvature
The key mechanism is visible in the conditional Bernoulli likelihood. For a logistic model with linear predictor \(\eta_i\), the contribution to the score for a coefficient is \(x_i(y_i-p_i)\), and the contribution to the information is \(p_i(1-p_i)x_ix_i^{\mathsf T}\). As the fitted probability approaches 0 or 1, \(p_i(1-p_i)\) approaches zero. If increasing a coefficient pushes correctly classified observations farther toward their observed outcomes, the likelihood can keep increasing even as the information along that direction collapses.
A short derivation makes the pattern explicit. Consider two hypothetical observations with predictor values \(x=-1,1\) and outcomes \(y=0,1\), respectively, in a logistic model with slope \(t\) and no intercept. Both observations agree with the sign of the predictor. The log likelihood along this slope direction is
For any finite \(t\), its derivative is positive:
Thus, no finite \(t\) makes the score zero; the likelihood increases toward its supremum of zero as \(t\) tends to infinity. The information in this direction is
which also tends to zero. This is the characteristic combination: the coefficient grows, the likelihood gains become small, and the score and curvature can both become numerically small. Integration over random effects changes the exact likelihood calculations, but the same combination is a valuable warning signal in a fitted GLMM.
Worked Example: A Score That Shrinks While the Slope Grows
For the two hypothetical observations above, evaluate the score and information at \(t=8\). Since \(e^8\approx2980.958\),
The information is
The log likelihood is \(-2\log(1+e^{-8})\approx-0.0006709\). Increasing \(t\) can therefore improve the log likelihood only slightly at this point, while the positive score shows that the likelihood is still increasing. Both score and information are small, but neither fact establishes a finite maximum. This is a simple separation illustration, not a GLMM marginal-likelihood calculation.
Three Complementary Screens
Use these diagnostics together, preferably over several optimizer iterations or starting values. A single final estimate rarely provides enough context.
Check whether fixed-effect score components are small relative to their information scale, and whether they shrink while a particular coefficient continues to move. A large scaled component at a purported solution is first a warning that the fit may not be stationary; it is not specific to separation.
Examine eigenvalues of a numerically scaled information matrix and their eigenvectors. A very small eigenvalue points to a weak-curvature direction. Determine whether that direction is mainly a fixed-effect contrast, a nuisance-parameter direction, or a mixture.
Compare estimates across iterations, starting values, and reasonable optimizer settings. A coefficient that keeps increasing in magnitude while log-likelihood gains shrink and information in the same direction collapses is a stronger separation warning than any one of those signs alone.
For a finite regular maximum, score components should be close to zero relative to numerical precision and the relevant information should be nonsingular. But separation can make the score small simply because the likelihood is flattening near a nonattained supremum. Conversely, a large score at the reported solution more directly suggests inadequate optimization or inaccurate likelihood calculations. Always distinguish “not stationary” from “possibly moving along a nearly flat direction.”
Worked Example: Reading a Hypothetical GLMM Fit Path
Suppose a hypothetical binary GLMM has 12 clusters and a binary exposure. Across iterations, the optimizer reports the following fixed-effect exposure coefficient, scaled score component in its main direction, smallest eigenvalue of the scaled information matrix, and log-likelihood increase from the preceding checkpoint.
| Checkpoint | Exposure coefficient | Scaled score component | Smallest information eigenvalue | Log-likelihood increase |
|---|---|---|---|---|
| A | 3.1 | 0.42 | 0.038 | 0.160 |
| B | 6.0 | 0.12 | 0.0065 | 0.012 |
| C | 9.1 | 0.018 | 0.00019 | 0.00018 |
From B to C, the coefficient increases by \(9.1-6.0=3.1\), while the scaled score falls from 0.12 to 0.018, the smallest eigenvalue falls by a factor of \(0.0065/0.00019\approx34.2\), and the likelihood gain falls from 0.012 to 0.00018. Taken together, these are a strong screen for a weakly curved, separation-like direction. They do not prove separation: the next step is to examine the data pattern and assess whether the weak eigenvector aligns with the exposure effect, rather than with a nuisance or random-effect parameter. A convergence message at C would not resolve that question.
Separate Separation Signals from Other Problems
A small information eigenvalue is not unique to separation. Highly correlated fixed-effect predictors can make a contrast difficult to estimate while leaving fitted linear predictors stable. A random-effect variance near its boundary can also make the joint information nearly singular. In such cases, coefficient instability may reflect collinearity or weak identification rather than a direction that steadily improves classification as its magnitude grows. Review the eigenvector and the parameter block involved, and compare the coefficient path with fitted probabilities and outcome patterns.
Worked Example: A Small Eigenvalue Without a Separation Pattern
Consider a hypothetical GLMM with two continuous fixed-effect predictors, \(x_1\) and \(x_2\), that are nearly duplicates. The fitted coefficients are 0.43 and 0.41, while the smallest eigenvalue of the scaled fixed-effect information is 0.0008. Suppose the corresponding eigenvector is approximately \((1,-1)^{\mathsf T}/\sqrt{2}\). This points to the contrast between the two coefficients: changing them in opposite directions has little effect on the fitted linear predictors because the predictors are nearly the same.
The small eigenvalue warns that the individual coefficients may be imprecise. If estimates change across starts but their sum and the fitted probabilities remain stable, and there is no growing coefficient with shrinking likelihood increments in a classification direction, this pattern is more consistent with collinearity than with the particular separation signature discussed here. It still does not rule out separation elsewhere in the model. Check the complete fixed-effect information and data patterns rather than treating one eigenvalue as a verdict.
Interpretation, Alternatives, and Limits
Use score and information diagnostics as numerical screens. If they suggest a diverging direction, inspect the corresponding predictors and outcome patterns and use the geometric and linear-programming approaches from the earlier material on separation. Those approaches address whether the data admit a separating direction; score and information diagnostics describe how the fitted likelihood behaves numerically. They answer related but distinct questions. Profile-likelihood signs of separation require a separate analysis.
These numerical screens have several limitations:
- Scaled scores depend on the likelihood approximation and on how nuisance parameters are handled. A poorly computed marginal likelihood can produce misleading derivatives.
- Eigenvalues depend on parameter scaling. Scale the information before comparing directions, and interpret eigenvectors in the original parameterization.
- Small curvature may reflect limited data, collinearity, or a boundary variance estimate. It is not, by itself, evidence that a coefficient is infinite.
- A finite but extreme estimate can be weakly identified without true separation. Conversely, an optimizer may stop before a clear coefficient-growth pattern appears.
- Random-effects integration means conditional working weights are clues, not substitutes for marginal observed information.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why can a small score at a fitted binary GLMM fail to establish that a finite maximum has been reached?
- In the two-observation example, calculate the score at \(t=10\) using \(e^{10}\approx22026.47\). What does its sign imply?
- A scaled information matrix has a very small eigenvalue whose eigenvector mainly involves a random-effect variance parameter. Does this alone indicate fixed-effect separation?
- Name two numerical patterns that, when they occur together, raise a separation concern more than either one alone.
Solutions. (1) Near a nonattained supremum, the likelihood can flatten so that its score is numerically small even while it is still increasing toward larger coefficient values. (2) \(U_t=2/(1+22026.47)\approx0.0000908\), which is positive; the likelihood is still increasing at \(t=10\). (3) No. It flags weak curvature in a nuisance direction, which may arise from a boundary or weak identification; inspect the fixed-effect directions and outcome patterns as well. (4) For example, a fixed-effect coefficient that keeps growing while likelihood increments shrink, together with collapsing information in the same coefficient direction. These are warning signs, not a proof.