What the Random-Intercept Variance Measures
A random intercept allows clusters to have different baselines while retaining a common distribution for those baselines. The variance of that distribution describes how much cluster baselines vary around the fixed intercept. Its meaning is on the model's link scale: it is not automatically a variance in the original response units.
For example, in a model with a log link, the random intercept varies on the log-mean scale. In a model with a logit link, it varies on the log-odds scale. In the Gaussian identity-link model, the link scale is the response scale itself. These distinctions matter because a variance is always expressed in squared units of the scale on which the random quantity is defined.
The interpretation relies on model assumptions. Conditional on the random effects and covariates, observations follow the chosen response-family distribution and are independent; clusters are independent of one another; and the random intercepts follow the specified normal distribution, with mean zero and variance \(\tau_{00}\). The random effects are also assumed independent of the covariates in the usual formulation. A model can be useful even if these assumptions are only approximations, but then the variance parameter describes heterogeneity under that fitted model, not a model-free property of the data.
The zero mean assumption makes \(\beta_0+\mathbf{x}_{ij}^{T}\boldsymbol{\beta}\) the population-average linear predictor across the random-intercept distribution. For a nonlinear link, that does not generally make the corresponding conditional mean the average response across clusters. Averaging after applying a nonlinear inverse link is not the same as applying that inverse link to an average linear predictor.
Variance, Standard Deviation, and Typical Cluster Deviations
A variance can be hard to interpret directly. Taking its square root returns the scale of the random intercept. If \(\tau_{00}=0.36\), for example, then the random-intercept standard deviation is \(0.60\) link-scale units. Under the normality assumption, about 68% of the cluster effects lie within one standard deviation of zero, and about 95% lie within approximately 1.96 standard deviations. These are statements about the assumed distribution of cluster effects, not confidence intervals for \(\tau_{00}\).
The model also implies a distribution for the difference between two independently sampled cluster intercepts. Because \(b_j\) and \(b_k\) are independent with variance \(\tau_{00}\),
Thus the standard deviation of the difference is \(\sqrt{2\tau_{00}}\), not \(\sqrt{\tau_{00}}\). This describes how far apart two randomly selected cluster deviations may be on the link scale; it does not say that every pair of clusters differs by that amount.
Worked Example: Gaussian Random Intercepts
Suppose hypothetical continuous measurements are modeled by a Gaussian identity-link random-intercept model: \(Y_{ij}=\beta_0+\beta_1x_{ij}+b_j+\varepsilon_{ij}\), with \(b_j\sim N(0,0.64)\) and \(\varepsilon_{ij}\sim N(0,2.25)\), independently. The variances are model estimates from hypothetical data.
The random-intercept standard deviation is \(\sqrt{0.64}=0.80\) measurement units. Under the normal random-effects assumption, about 95% of cluster deviations lie within \(1.96(0.80)=1.568\), or approximately \(-1.57\) to \(+1.57\) measurement units from zero. This is the between-cluster baseline spread after accounting for the included predictor.
The conditional residual standard deviation is separately \(\sqrt{2.25}=1.50\) measurement units. It describes observation-to-observation variation around each cluster's conditional mean, given its random intercept and predictors. The random-intercept variance \(0.64\) and residual variance \(2.25\) describe different sources of variation, even though both are in squared response units in this identity-link model.
For two independently selected clusters, the standard deviation of their intercept difference is \(\sqrt{2(0.64)}=\sqrt{1.28}\approx1.131\) measurement units. The calculation uses independence of their random effects; it is not the within-cluster residual standard deviation.
Non-Gaussian Models: Link-Scale Heterogeneity
For a non-Gaussian GLMM, the random intercept still adds to the linear predictor, but it does not generally add to the observed response. Conditional response variation is specified by the response family. For example, a Poisson model has conditional variance \(\operatorname{Var}(Y_{ij}\mid b_j)=\mu_{ij}\); a Bernoulli model has conditional variance \(\mu_{ij}(1-\mu_{ij})\). These are not the random-intercept variance. The latter is the variance of \(b_j\) across clusters.
Worked Example: Poisson Random Intercepts
Consider hypothetical event counts per period with a Poisson log-link model. At the covariate values under consideration, let the fixed part of the linear predictor be \(\log(3)\), and suppose \(b_j\sim N(0,0.25)\). The conditional model is \(Y_{ij}\mid b_j\sim\operatorname{Poisson}(\mu_{ij})\) with \(\log(\mu_{ij})=\log(3)+b_j\).
The random-intercept standard deviation is \(\sqrt{0.25}=0.50\) log-rate units. A cluster one standard deviation above zero has \(b_j=0.50\), giving conditional mean \(e^{\log(3)+0.50}=3e^{0.50}\approx4.946\) events per period. At one standard deviation below zero, \(b_j=-0.50\), and the conditional mean is \(e^{\log(3)-0.50}=3e^{-0.50}\approx1.820\) events per period. Thus a half-unit deviation on the log scale has a multiplicative effect on the conditional mean; it is not a difference of 0.50 events.
For these two cluster values, the conditional Poisson response variances equal their conditional means: approximately 4.946 and 1.820 events-squared per period, respectively. These conditional variances change with the means, whereas the random-intercept variance remains \(0.25\) log-rate-units-squared across clusters in this model.
The approximate central 95% range of random intercepts is \(-1.96(0.50)\) to \(+1.96(0.50)\), or \(-0.98\) to \(+0.98\). At the stated covariate values, the corresponding conditional means range from approximately \(3e^{-0.98}=1.126\) to \(3e^{0.98}=7.993\) events per period. This is a range induced by the assumed distribution of \(b_j\), not an interval estimating the fixed mean or the variance parameter.
Worked Example: Logistic Random Intercepts
Suppose hypothetical binary outcomes follow a Bernoulli logit random-intercept model, and at a particular covariate value the fixed linear predictor is \(\operatorname{logit}(0.20)=\log(0.20/0.80)\approx-1.386\). Let \(b_j\sim N(0,0.49)\). The random-intercept standard deviation is \(\sqrt{0.49}=0.70\) log-odds units.
At \(b_j=0\), the conditional probability is 0.20 by construction. At one standard deviation above zero, the linear predictor is \(-1.386+0.70=-0.686\), and the conditional probability is \(1/(1+e^{0.686})\approx0.335\). At one standard deviation below zero, the predictor is \(-1.386-0.70=-2.086\), giving \(1/(1+e^{2.086})\approx0.110\). The variance \(0.49\) quantifies spread in cluster log-odds baselines; it is not a variance of probabilities. The Bernoulli conditional response variance at each cluster probability is instead \(\mu_{ij}(1-\mu_{ij})\).
The calculation also shows why a link-scale standard deviation cannot usually be read as a constant response-scale shift: the inverse logit is nonlinear, so the probability change depends on the starting linear predictor.
Distinguishing Sources of Variation
It is useful to ask what is random in the model and at what level. The random intercept varies across clusters. Conditional response variation occurs among observations after the cluster effect and predictors have been specified. In the Gaussian identity-link case, an additive residual term makes the separation particularly visible. In a Poisson or Bernoulli GLMM, conditional variation is instead determined by the response distribution and its mean; it is not represented by a single extra additive residual on the link scale.
A common mistake is to see a random-intercept variance of \(0.25\) and say, “the response has variance \(0.25\) within every cluster.” That is wrong: the parameter is the variance of the cluster-specific link-scale deviations. The correction is to identify the link and interpret \(\tau_{00}\) in squared units of that link-scale quantity, then separately state the conditional response variance implied by the outcome family.
Another error is to call the random-intercept standard deviation a standard error. It is not the standard error of a fixed-effect estimate. It describes the assumed population distribution of cluster deviations; uncertainty in the estimated variance parameter is a separate inferential issue.
Diagnostics, Alternatives, and Limits
A variance estimate should be interpreted alongside checks of the model. Examine whether the chosen response family is plausible, whether important predictors or cluster structure have been omitted, and whether residual behavior suggests extra variation or dependence not captured by the model. For the random-effects distribution, plots of estimated cluster effects can reveal strong asymmetry or unusual clusters, but these estimates are shrunken and are not direct observations of the \(b_j\). A normality assessment is therefore informative but not decisive.
A fitted estimate at or near zero is a boundary case because \(\tau_{00}\) cannot be negative. It indicates that the fitted model finds little estimated between-cluster intercept spread after accounting for its other terms; it does not prove that clusters are identical or that the design has enough information to detect heterogeneity. In particular, estimation of a variance component can be weak when the number of clusters is small. Report the estimate and its uncertainty using an appropriate method rather than treating the point estimate alone as definitive.
| Approach | What it represents | Important distinction |
|---|---|---|
| Random-intercept model | A distribution of cluster baselines with variance \(\tau_{00}\) | Shares information across clusters and assumes a specified random-effects distribution |
| Fixed cluster indicators | A separate baseline parameter for each observed cluster | Does not estimate a population variance of cluster effects in the same way |
| Model without cluster effects | No modeled cluster-to-cluster baseline variation | May leave dependence or heterogeneity unexplained if clusters truly differ |
These approaches answer different questions. Fixed cluster indicators can be useful when interest is limited to the particular observed clusters, while a random intercept models those clusters as arising from a broader distribution. Neither the random-intercept variance nor its normality assumption is automatically appropriate: conclusions depend on the cluster sampling context, the number and size of clusters, and the adequacy of the conditional response model. The parameter also does not by itself describe all possible cluster differences; a model with predictor-dependent cluster effects would require additional terms.
Justifying the Standard-Deviation Interpretation
The conversion from variance to standard deviation is a direct consequence of the model distribution. If \(b_j\sim N(0,\tau_{00})\), standardizing gives \(Z_j=b_j/\sqrt{\tau_{00}}\sim N(0,1)\), when \(\tau_{00}>0\). Therefore \(P(|b_j|\leq1.96\sqrt{\tau_{00}})=P(|Z_j|\leq1.96)\approx0.95\). This justifies the approximate central range used above. It is a probability statement about a random cluster effect under the assumed model, not a 95% confidence statement about an unknown parameter.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A Gaussian identity-link model estimates \(\tau_{00}=0.81\). What is the random-intercept standard deviation, and what are its units if the response is measured in millimetres?
- A Poisson log-link model has \(\tau_{00}=0.25\) and a fixed conditional mean of 2 when \(b_j=0\). What is the mean at one standard deviation above and below zero?
- In a Gaussian random-intercept model, \(\tau_{00}=0.49\) and the residual variance is 4. State what each variance describes and give their standard deviations.
- What can and cannot be concluded if the estimated random-intercept variance is zero?
Solutions. (1) The standard deviation is \(\sqrt{0.81}=0.90\) millimetres, describing the spread of cluster intercept deviations on the response scale. (2) The standard deviation is \(\sqrt{0.25}=0.50\) log-rate units. At \(b_j=+0.50\), the mean is \(2e^{0.50}\approx3.297\); at \(b_j=-0.50\), it is \(2e^{-0.50}\approx1.213\). (3) The random-intercept variance describes between-cluster baseline variation, with standard deviation \(\sqrt{0.49}=0.70\) response units. The residual variance describes conditional observation-level variation, with standard deviation \(\sqrt{4}=2\) response units. (4) Zero is a boundary estimate indicating that the fitted model estimates no remaining random-intercept spread. It does not establish that true cluster differences are absent, nor does it rule out limited information, model misspecification, or other forms of clustering.