Fixed Intercepts and Cluster Deviations

Two Components, One Conditional Intercept

In a random-intercept GLMM, clusters share a fixed intercept but can differ through cluster-specific random deviations. Keeping these pieces separate matters: the fixed intercept describes the center of the modeled cluster-intercept distribution on the link scale, while a cluster’s deviation describes its modeled departure from that center. Together they determine that cluster’s conditional intercept.

The earlier material on conditional linear predictors with random intercepts introduced the role of a cluster-level shift in the linear predictor. Here the focus is on what the fixed intercept and that shift each mean, and why adding them gives the conditional intercept when the covariates are at their reference values.

Definition: For observation \(i\) in cluster \(j\), write a random-intercept GLMM as \(g\{E(Y_{ij}\mid b_j,\mathbf{x}_{ij})\}=\beta_0+\mathbf{x}_{ij}^{T}\boldsymbol{\beta}+b_j\). Here \(g\) is the link function, \(\beta_0\) is the fixed intercept, \(\mathbf{x}_{ij}^{T}\boldsymbol{\beta}\) is the fixed covariate contribution, and \(b_j\) is the random-intercept deviation for cluster \(j\). A standard model assumes \(b_j\sim N(0,\sigma_b^2)\). At the covariate reference point \(\mathbf{x}_{ij}=\mathbf{0}\), the cluster’s conditional intercept on the link scale is \(\beta_0+b_j\).

The fixed intercept is a model parameter shared by the clusters. Under the centered random-effects assumption, it is the mean of the modeled cluster-intercept distribution on the link scale. It is not automatically an intercept for an “average” response on the response scale. The deviation \(b_j\) is a cluster-level latent quantity: positive values put that cluster above the fixed intercept on the link scale, and negative values put it below. Because \(b_j\) has mean zero in the assumed distribution, deviations are centered around zero across the modeled population of clusters.

The conditional intercept \(\beta_0+b_j\) is the linear predictor at the covariate reference point, conditional on the cluster’s random effect. It is not the linear predictor for every observation in that cluster when covariates vary: the fixed covariate contribution must still be added for each observation.

Why the Components Add

The addition follows directly from the model’s conditional linear predictor. At the reference point, the covariate contribution is zero. The remaining terms are the fixed intercept and the cluster deviation.

\begin{aligned} \eta_{ij} &= \beta_0+\mathbf{x}_{ij}^{T}\boldsymbol{\beta}+b_j,\\ \mathbf{x}_{ij}=\mathbf{0} \quad\Longrightarrow\quad \eta_{ij} &= \beta_0+\mathbf{0}^{T}\boldsymbol{\beta}+b_j\\ &= \beta_0+b_j. \end{aligned}

This derivation also shows why the terms should not be confused. The fixed intercept is not already a cluster-specific intercept, because it does not include \(b_j\). And \(b_j\) alone is not the cluster’s intercept: it is only the amount added to the fixed intercept.

Key result: If all covariates are zero at the reference point, cluster \(j\)’s conditional intercept on the link scale is \(\beta_0+b_j\). If covariates are not at that reference point, the conditional linear predictor is \(\beta_0+\mathbf{x}_{ij}^{T}\boldsymbol{\beta}+b_j\), not just \(\beta_0+b_j\).

Worked Interpretations

Worked Example: Separating the Fixed Intercept and Deviation

Suppose a hypothetical fitted model for a continuous outcome uses the identity link and has fixed intercept \(\hat{\beta}_0=12\) outcome units. For a particular cluster, suppose the fitted random-intercept deviation is \(\hat{b}_j=-3\) outcome units. Find and interpret the conditional intercept.

The model’s reference covariate values are zero, so the cluster-specific conditional intercept is the sum of the two components:

\hat{\beta}_0+\hat{b}_j =12+(-3) =9

The fixed intercept is 12 outcome units, the center of the modeled cluster-intercept distribution on the identity-link scale. This cluster’s deviation is 3 units below that center. Therefore, its conditional intercept is 9 outcome units. This is the model’s conditional mean at the reference covariate values for this cluster; at other covariate values, the corresponding fixed covariate contribution must also be included.

Worked Example: A Positive Deviation on the Logit Scale

Consider a hypothetical random-intercept logistic GLMM with fixed intercept \(\hat{\beta}_0=-1.10\) and a fitted deviation \(\hat{b}_j=0.60\) for one cluster. Calculate the cluster’s conditional intercept on the link scale.

The link is the logit, and the covariate reference point is \(\mathbf{x}=\mathbf{0}\). Add the two link-scale components:

\hat{\beta}_0+\hat{b}_j =-1.10+0.60 =-0.50

The fixed intercept is \(-1.10\) logit units, and the cluster deviation is \(0.60\) logit units. The conditional intercept is \(-0.50\) logit units. The positive deviation means the cluster is above the fixed intercept on the logit scale; it does not mean the cluster’s outcome probability has increased by 0.60. The deviation is measured on the link scale, not in probability units.

Worked Example: Including a Covariate Away From Its Reference Value

In a hypothetical model with a log link, suppose \(\hat{\beta}_0=0.40\), the fixed slope for a predictor \(x\) is \(\hat{\beta}_1=0.25\) per unit, and cluster \(j\)’s fitted random-intercept deviation is \(\hat{b}_j=-0.10\). For an observation in that cluster with \(x=2\), calculate the conditional linear predictor and identify the conditional intercept separately.

At the reference point \(x=0\), the cluster’s conditional intercept is:

\hat{\beta}_0+\hat{b}_j =0.40+(-0.10) =0.30

For the observation with \(x=2\), the covariate contribution is \(\hat{\beta}_1x=0.25(2)=0.50\). Thus the conditional linear predictor is:

\hat{\eta}_{ij} =\hat{\beta}_0+\hat{\beta}_1x+\hat{b}_j =0.40+0.50-0.10 =0.80

The conditional intercept remains 0.30 on the log-link scale; the observation’s conditional linear predictor is 0.80 because its covariate contribution is 0.50. Calling 0.80 the intercept would mix the intercept with the predictor’s effect at \(x=2\).

Assumptions and Interpretation Checks

The decomposition is simple, but its interpretation depends on the model being specified correctly. In the usual random-intercept formulation, the conditional response distribution and link are specified for \(Y_{ij}\) given the covariates and random effect. Responses are assumed conditionally independent given the random effects and covariates, unless the model explicitly includes additional dependence. The cluster deviations are modeled as draws from a mean-zero distribution, commonly \(N(0,\sigma_b^2)\), and are typically assumed independent across clusters. A common formulation also assumes the random effects are independent of the covariates.

These assumptions distinguish the fixed intercept from the cluster deviations: centering the random-effects distribution at zero makes \(\beta_0\) the distribution’s center on the link scale. If deviations were instead assigned a nonzero mean, that mean and the fixed intercept would not have the same interpretation without a corresponding reparameterization.

The reference point also matters. If a continuous predictor is left in its original units, \(x=0\) may be outside the observed range or substantively unhelpful. Centering a predictor changes the value represented by the fixed intercept: after centering at a chosen value, zero corresponds to that value. In a random-intercept-only model, the cluster deviation remains a shared shift, while the fixed intercept now refers to the new covariate reference point. State the coding and reference values before interpreting an intercept.

Estimation, Diagnostics, and Alternatives

In practice, the fixed intercept and random deviations are estimated rather than known. A reported fitted deviation \(\hat{b}_j\) is an estimate of a latent cluster quantity, not a directly observed difference computed from the cluster’s raw outcomes. Estimates are informed both by data from that cluster and by the assumed distribution of deviations. In particular, they should not be treated as error-free measurements of cluster-specific differences.

Before interpreting a fitted conditional intercept, check the details that define it:

Confirm the reference values. Verify that the covariates are zero at the intended reference point. Otherwise, \(\beta_0+b_j\) is not the complete predictor for the observation being discussed.
Check the component signs and arithmetic. A negative deviation is added as a negative number; it is not subtracted as an unsigned magnitude.
Check the random-intercept distribution and variance estimate. A variance estimate near zero suggests little estimated between-cluster variation under this model. It does not justify interpreting noisy cluster deviations as substantial differences.
Look for clusters with unusual data support. Very small clusters can provide limited information about their deviation. A striking fitted value should be interpreted with that uncertainty in mind.
Assess whether a shared shift is plausible. A random intercept moves the conditional linear predictor by the same amount for observations in a cluster. If the relationship with a predictor appears to vary across clusters, a random-intercept-only structure may be inadequate.

A model with no random intercept instead imposes a common intercept, so it cannot represent cluster-to-cluster deviations in the conditional linear predictor. At the other extreme, a model with a separate fixed intercept for each cluster estimates a distinct intercept for every included cluster rather than modeling deviations as draws from a shared distribution. Those are different assumptions and serve different inferential aims; neither changes the arithmetic definition of \(\beta_0+b_j\) in the random-intercept model.

The random-intercept formulation is limited when clusters do not differ by a common shift, when the random-effects distribution is a poor description of the cluster variation, or when the covariate reference point makes the fixed intercept meaningless in context. The link-scale intercept also should not be casually described as a response-scale average. This tutorial’s sum is specifically a conditional intercept on the link scale.

Common Mistakes

Calling \(b_j\) the cluster intercept. It is the deviation from the fixed intercept. The conditional intercept is \(\beta_0+b_j\).
Calling \(\beta_0\) the response-scale mean for a typical cluster. It is the fixed intercept on the link scale, at the covariate reference point; the response-scale interpretation depends on the link and model.
Forgetting covariate contributions. The sum \(\beta_0+b_j\) is the conditional intercept at the reference point, not necessarily the full linear predictor for an observation with nonzero covariates.
Reading the deviation in response units. Its units are those of the link scale, such as logit units for a logit link, not probability units.

Key takeaway: The fixed intercept locates the center of the cluster-intercept distribution on the link scale under a mean-zero random-effects model. A cluster’s random intercept is its deviation from that center. Add them to obtain its conditional intercept at the covariate reference point.

Check Your Understanding

Answer each question, then compare with the brief solutions.

A fitted model has \(\hat{\beta}_0=2.4\) and \(\hat{b}_j=-0.7\). What is the cluster’s conditional intercept on the link scale?
In a model with \(\hat{\beta}_0=-0.8\), \(\hat{b}_j=0.3\), and a fixed covariate contribution of \(0.5\), what is the conditional intercept and what is the full conditional linear predictor?
Why is a deviation of \(0.4\) in a logit-link model not an increase of 0.4 in probability?
What does a zero-mean random-intercept distribution imply about the fixed intercept on the link scale?

Solutions. (1) \(2.4+(-0.7)=1.7\) link-scale units. (2) The conditional intercept is \(-0.8+0.3=-0.5\); the full predictor is \(-0.5+0.5=0\). (3) The deviation is measured in logit units, while probability is on the response scale. (4) The fixed intercept is the center, or mean, of the modeled cluster-intercept distribution at the covariate reference point on the link scale.