Why the Reference Point Matters
In a random-slope model, the cluster-specific intercept is not an abstract starting value: it describes the cluster's fitted level at a particular value of the predictor. If the predictor is measured from zero, that intercept refers to \(x=0\), whether or not zero is scientifically useful or represented in the data. Choosing another reference point can make the baseline more meaningful, while preserving the same fitted model when the random-effects structure is transformed consistently.
The central question is therefore not merely whether to center a predictor, but which predictor value should define the baseline? A useful choice is often a clinically or scientifically interpretable value within the observed range, such as a standard dose, a typical age, or the beginning of a follow-up period. The reference point affects the meanings of the intercept and its covariance with the slope; it does not, by itself, change the underlying cluster-specific trajectories.
Model and Meaning of the Reference Point
Let \(i\) index observations and \(j\) index clusters. For a predictor \(x_{ij}\), choose a reference value \(c\) and write the conditional GLMM as
Here \(g\) is the link function, \(\beta_0\) and \(\beta_1\) are fixed effects, and \(b_{0j}\) and \(b_{1j}\) are cluster-specific random effects. The random intercept and slope are conventionally modeled jointly as mean-zero normal variables:
This parameterization assumes a common fixed linear trend on the link scale, cluster-specific deviations in level and slope, and a valid positive-semidefinite random-effects covariance matrix. The chosen \(c\) must be the same reference value for the observations in all clusters if the goal is a common interpretation of the fixed and random intercepts. As in other GLMMs, the conditional response distribution and link must also be appropriate for the outcome, and observations are modeled as conditionally independent given the random effects and covariates.
The intercept describes a value on the link scale. With the identity link, it is a conditional mean on the response scale. With a logit or log link, it is respectively a conditional log-odds or log-mean; it is not itself a probability or expected count. Applying the inverse link gives the corresponding cluster-specific conditional response value at \(c\). That value is not generally the marginal response mean averaged over the random-effects distribution.
Reparameterizing at a New Reference Point
Start with a model written in terms of the uncentered predictor \(x\), with random intercept \(b_{0j}\) at \(x=0\) and random slope \(b_{1j}\):
To express the same trajectory around \(c\), substitute \(x_{ij}=(x_{ij}-c)+c\) and collect the constant and slope terms:
Thus the fixed intercept becomes \(\beta_0^*=\beta_0+c\beta_1\), the new random intercept is \(b_{0j}^*=b_{0j}+cb_{1j}\), and the slope remains unchanged. At \(x=c\), the cluster-specific linear predictor is \(\beta_0^*+b_{0j}^*\), as required. This is a change of coordinates, not a change to the trajectories, provided the covariance structure is transformed as well.
The random-intercept variance and its covariance with the slope consequently depend on the reference point:
For example, the variance formula follows by expanding \(\operatorname{Var}(b_{0j}+cb_{1j})\): the two variance terms are \(\tau_{00}\) and \(c^2\tau_{11}\), and the cross terms sum to \(2c\tau_{01}\). Therefore, a change in the estimated random-intercept variance after recentering need not mean that the underlying cluster trajectories have become more variable; the intercept now measures variation at a different predictor value.
Worked Example: A Meaningful Baseline for Follow-Up Time
Suppose a hypothetical identity-link model describes a clinical measurement over follow-up time \(x\), in months, as
The reference point is \(c=10\) months. The fixed intercept of 70 measurement units is the fitted mean at 10 months for a cluster with random effects equal to zero. The fixed slope is 2 measurement units per month. For a hypothetical cluster with \(b_{0j}=-3\) measurement units and \(b_{1j}=0.5\) measurement units per month, its intercept at 10 months is \(70-3=67\) measurement units, and its slope is \(2+0.5=2.5\) measurement units per month.
At 12 months, the same cluster's fitted conditional mean is \(70+2(12-10)-3+0.5(12-10)=70+4-3+1=72\) measurement units. The intercept of 67 refers to 10 months, not to the start of follow-up. If the intended baseline were the start, the reference value would need to be \(c=0\).
Worked Example: Interpreting a Logit-Scale Baseline
Consider a hypothetical binary-outcome model using age \(x\), in years, centered at \(c=20\):
For a cluster with \(b_{0j}=0.4\) and \(b_{1j}=-0.02\) per year, the conditional log-odds at age 20 are \(-1.2+0.4=-0.8\). The corresponding odds are \(\exp(-0.8)=0.449\), and the conditional probability is \(1/(1+\exp(0.8))=0.310\). Thus the random intercept \(0.4\) is a deviation of 0.4 log-odds units at age 20; it is not a 0.4 increase in probability.
The cluster-specific log-odds slope is \(0.08-0.02=0.06\) per year. At age 25, the conditional linear predictor is \(-0.8+0.06(25-20)=-0.5\), giving odds \(\exp(-0.5)=0.607\) and probability \(1/(1+\exp(0.5))=0.378\). These are conditional predictions for the stated random effects, not averages over the random-effects distribution.
Worked Example: How the Reference Point Changes Covariance
Suppose, hypothetically, that a model written with raw predictor \(x\) has random-intercept variance \(\tau_{00}=2\), intercept-slope covariance \(\tau_{01}=0.10\), and random-slope variance \(\tau_{11}=0.04\). The predictor is measured in units such that its reference point is changed from zero to \(c=2\). The new random intercept is \(b_{0j}^*=b_{0j}+2b_{1j}\).
Its variance is \(2+2(2)(0.10)+2^2(0.04)=2+0.40+0.16=2.56\), in squared linear-predictor units. The new intercept-slope covariance is \(0.10+2(0.04)=0.18\). The slope variance remains \(0.04\). The new intercept standard deviation is \(\sqrt{2.56}=1.60\) linear-predictor units, while the slope standard deviation is \(\sqrt{0.04}=0.20\) linear-predictor units per predictor unit.
These calculations describe a new intercept at \(x=2\); they do not indicate that the clusters' trajectories changed. The original covariance matrix is valid: its determinant is \(2(0.04)-(0.10)^2=0.07>0\). The transformed matrix is also valid, with determinant \(2.56(0.04)-(0.18)^2=0.0700>0\).
Choosing and Checking a Reference Point
A defensible reference point should make the intercept answer a useful question and should not force that question to depend on unsupported extrapolation. Possible choices include a prespecified clinical threshold, a meaningful treatment level, the start of follow-up, or a typical predictor value. A sample mean can be convenient, but it is not automatically the most scientifically relevant baseline. Inspect the predictor's units, observed range, and distribution across clusters before deciding.
The reference point can also affect numerical behavior. If \(c\) is far from the observed data, the fitted intercept refers to a poorly supported value and may be strongly associated with the slope. A value near the center of the observed predictor range often makes the intercept more stable and easier to interpret. This is a practical aid, not a guarantee of good estimation or a substitute for checking model fit.
For an unstructured intercept-slope covariance matrix, there is a particular value \(c=-\tau_{01}/\tau_{11}\), when \(\tau_{11}>0\), at which the transformed intercept and slope have zero covariance. This follows from setting \(\tau_{01}+c\tau_{11}=0\). That value may be outside the data range or lack scientific meaning, so zero covariance is not by itself a reason to choose it. The parameter values are also estimated, making this choice data-dependent.
Changing the reference point preserves the model only when the random-effects covariance is allowed to transform. A model that imposes zero intercept-slope covariance at one reference point generally does not impose zero covariance at another. For example, the transformed covariance is \(\tau_{01}+c\tau_{11}\), which need not be zero even if \(\tau_{01}=0\). Reimposing a diagonal covariance matrix after recentering can therefore change the model's assumptions rather than simply relabel its parameters.
Group-mean centering, covered in the earlier material on Group-Mean Centering and Conditional Effects, uses a cluster-specific reference rather than one common reference point. It answers a different interpretive question; it is not interchangeable with choosing a common value \(c\) when the goal is a shared baseline across clusters.
Diagnostics, Alternatives, and Limitations
Check whether the chosen reference value is observed or reasonably close to observations, and whether clusters provide information near it. If few clusters have data near \(c\), cluster baselines there can rely heavily on the assumed linear trajectory. Also check units and scale: changing years to months changes the numerical slope and its variance, although a correctly transformed model represents the same fitted relationships.
Convergence warnings, an estimated random-slope variance near zero, or an estimated intercept-slope correlation near \(1\) or \(-1\) can signal a difficult or weakly supported random-effects structure. These symptoms do not establish that centering is the cause. Verify the coding and reference value, inspect the data coverage, and compare the fitted results under sensible parameterizations. A reference point cannot create information absent from the data.
| Choice | What the intercept represents | Main consideration |
|---|---|---|
| Raw predictor, \(c=0\) | Conditional level at zero | Directly useful if zero is meaningful and supported; otherwise it may be an extrapolation. |
| Meaningful common value, \(c\) | Conditional level at the chosen value | Often clearest for a scientific or clinical baseline. |
| Typical value, such as the overall mean | Conditional level near a typical predictor value | Can improve interpretability, but the mean need not answer the scientific question. |
| Cluster-specific mean | Level relative to each cluster's own predictor mean | Targets a different decomposition and interpretation from a common reference. |
The main limitation is that choosing a clear reference point does not ensure that a linear random-slope model is appropriate. Curvature, limited within-cluster predictor variation, or an unsuitable conditional response model can still undermine interpretation. Nor does an interpretable conditional baseline establish a causal effect: that requires an appropriate study design and assumptions beyond the random-slope parameterization.
A Practical Choice
Decide which predictor value should define the cluster-specific level, using the scientific context and units.
Compare the proposed value with the observed range and with the predictor values represented across clusters.
Describe the fixed and random intercepts on the link scale, then apply the inverse link if a conditional response-scale value is needed.
When changing the reference point, transform the intercept and its covariance with the slope rather than assuming the old covariance values still apply.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a log-link random-slope model centered at \(c=5\), what predictor value does the random intercept describe, and what scale is it on?
- If \(\tau_{00}=1.5\), \(\tau_{01}=-0.10\), and \(\tau_{11}=0.05\), calculate the random-intercept variance after centering at \(c=2\).
- A researcher changes the reference point but fixes the intercept-slope covariance at zero in both fitted models. Why might the resulting models not be equivalent?
- A proposed reference point lies well beyond the observed predictor range. Give one interpretive risk and one appropriate check.
Solutions. (1) It describes the cluster-specific conditional linear predictor at \(x=5\); with a log link, that is a log-mean, not the mean itself. (2) The variance is \(1.5+2(2)(-0.10)+2^2(0.05)=1.5-0.40+0.20=1.30\) squared linear-predictor units. (3) Reparameterization gives new covariance \(\tau_{01}+c\tau_{11}\), which is not generally zero when the original covariance is zero. Reimposing zero covariance can change the assumed distribution of cluster trajectories. (4) The baseline may be an extrapolation and poorly estimated. Check the observed range and the predictor values available across clusters near the proposed point.