Explaining Cluster Variation in a Multicenter Case

What a Random-Intercept Summary Says

In a multicenter logistic GLMM, the random intercept describes how centers vary in their conditional baseline log odds after accounting for the covariates in the model. Explaining this variation is not the same as claiming that a particular center characteristic has an effect. The model summarizes a distribution of center-specific deviations; it does not, by itself, identify why centers differ.

Suppose \(Y_{ij}\) is a binary outcome for patient \(i\) in center \(j\), and \(x_{ij}\) is a vector of patient-level covariates. A random-intercept model can be written as

Here \(b_j\) is the center-specific random intercept, measured in log-odds units, and \(\tau^2\) is its variance across centers. Conditional on the random intercepts and included covariates, outcomes are assumed independent, and the center effects are assumed independent draws from a normal distribution with mean zero and common variance. The model also assumes that the included fixed effects have the same coefficients across centers; this random-intercept model does not allow those effects to vary by center.

The conditional probability scale is nonlinear: the same shift \(b_j\) in log odds changes probabilities by different amounts at different covariate values. Therefore, the variance is most directly interpreted on the log-odds scale. It is not a percentage-point difference in outcome probability, and it does not say that every pair of centers differs by the same amount.

From Variance to a More Interpretable Contrast

A common supplementary summary is the median odds ratio (MOR). Imagine selecting two centers at random from the modeled center distribution and comparing their conditional odds for otherwise identical patients. Label the center with the higher odds as the numerator. The MOR is the median of this larger-to-smaller odds ratio across such random pairs. It is at least 1; a value closer to 1 indicates less modeled center variation.

For two independent random intercepts \(b_1,b_2\sim N(0,\tau^2)\), their difference is normal with mean zero and variance \(2\tau^2\). The absolute difference has median \(0.67449\sqrt{2\tau^2}\), because \(0.67449\) is the median of the absolute value of a standard normal variable. Conditional odds ratios are exponentiated log-odds differences. Hence:

This is a model-based contrast between two randomly selected centers, not an estimated odds ratio for a named center or a measured center-level predictor. It is also conditional on the covariates in the model: the comparison holds those covariates fixed.

Reporting Uncertainty and the Model-Based Meaning

A variance estimate is uncertain, and its uncertainty matters especially when there are few centers or little information about between-center differences. If software supplies an interval for \(\tau^2\), the MOR is a monotone increasing function of that variance. Applying the MOR transformation to both variance endpoints therefore gives corresponding endpoints on the MOR scale. The result remains an interval for a model-derived quantity, not a probability statement about the parameter.

When explaining the variance itself, report its scale and connect it to the model. For example, \(\hat{\tau}^2=0.25\) means an estimated between-center variance of 0.25 conditional log-odds-squared units; \(\hat{\tau}=0.500\) gives the more directly readable standard deviation in log-odds units. The MOR can provide an odds-ratio-scale description, but does not replace the variance estimate or its uncertainty.

Assumptions, Diagnostics, and Alternatives

A random-intercept summary is only as useful as the model supporting it. The interpretation above relies on the logistic mean structure, the specified covariates, the normal random-intercept distribution, and conditional independence of outcomes given those effects and covariates. It also assumes that the centers are meaningfully treated as draws from a common distribution. If centers were selected in a way that makes this assumption implausible, or if important predictors differ systematically across centers and are omitted, the estimated random-intercept variation can absorb that misspecification.

Useful checks focus on whether the model gives a credible account of the data, rather than on treating a single variance estimate as a diagnostic test. Examine observed outcome patterns and fitted predictions across centers, including whether a few centers have unusually large residual discrepancies. Compare the observed distribution of outcomes with model-based predictions at both the patient and center levels. Consider whether the estimated variance is sensitive to influential centers or reasonable changes to covariate specification. If the estimate is near zero, remember that variance is constrained to be nonnegative; a boundary estimate is not proof that centers are identical.

The random-intercept model is not the only way to represent center structure. Treating center as a fixed categorical effect estimates separate conditional contrasts for the included centers, whereas the random-intercept model summarizes a distribution and can support partial pooling. A population-averaged model targets a different, marginal interpretation. These approaches answer different questions; a random-intercept variance should not be relabeled as a fixed center effect or as a marginal effect.

The random-intercept summary also has limits. The MOR is dependent on the assumed distribution and model specification, and does not attribute heterogeneity to staffing, policy, or any other center feature unless those features are explicitly included and appropriately modeled. Nor does a variance estimate show whether a particular center is above or below the model’s average. Center-specific random-effect predictions are shrunken estimates and have their own uncertainty; they are not the same object as the variance summary.

Common Mistake