Reporting Fixed and Random Effects Together

Why Report Fixed and Random Effects Together?

A fixed slope summarizes the model’s common link-scale association, while a random slope allows that association to differ across clusters. Reporting only the fixed slope can hide meaningful cluster-to-cluster variation. Reporting only a cluster-specific slope, in turn, can obscure how that cluster’s association relates to the common pattern. A useful report makes both parts visible and keeps the model’s scale clear.

The central reporting task is to translate a fitted model into three connected statements: the fixed association, the estimated deviations for a named cluster, and the resulting cluster-specific intercept and slope. For a nonlinear link, it is also important to distinguish a link-scale statement from a statement about probabilities or expected counts.

Model and Interpretation

For observation \(i\) in cluster \(j\), let \(\mu_{ij}=E(Y_{ij}\mid b_{0j},b_{1j},x_{ij})\). A random-intercept-and-slope model with one predictor is

g(\mu_{ij})=\beta_0+\beta_1x_{ij}+b_{0j}+b_{1j}x_{ij}, \qquad \begin{pmatrix}b_{0j}\\b_{1j}\end{pmatrix} \sim N\left( \begin{pmatrix}0\\0\end{pmatrix}, \begin{pmatrix} \sigma_0^2 & \sigma_{01}\\ \sigma_{01} & \sigma_1^2 \end{pmatrix} \right).

Here \(g\) is the link function, \(\beta_0\) and \(\beta_1\) are fixed effects, and \(b_{0j}\) and \(b_{1j}\) are cluster-specific deviations. The covariance matrix allows the intercept and slope deviations to be associated. The random effects are modeled as jointly mean-zero normal across clusters; this describes their distribution in the model, not a claim that each observed set of clusters has deviations summing exactly to zero.

Definition: The cluster-specific intercept is \(\beta_0+b_{0j}\), the linear predictor at \(x_{ij}=0\). The cluster-specific slope is \(\beta_1+b_{1j}\), the change in the linear predictor for a one-unit increase in \(x_{ij}\) within cluster \(j\). The fixed slope \(\beta_1\) is the common slope when the random slope deviation is zero.

The intercept is interpretable at the covariate value coded as zero. If zero is not a meaningful or observed value, the predictor’s centering and units should be stated; the intercept then refers to the chosen reference value. The slope is per one unit of the predictor, so the unit matters. With a logit link, exponentiating a slope gives an odds ratio per unit. With a log link, it gives a multiplicative ratio of conditional means per unit. On either nonlinear scale, these are conditional interpretations given the random effects, not generally population-average effects.

The model presumes an appropriate conditional response distribution and link, a correctly specified linear predictor, and conditionally independent responses given the modeled covariates and cluster random effects. Clusters are treated as independent for the usual random-effects formulation. Interpretation also depends on adequate estimation: the model must have converged, and the data must contain enough information to estimate slope variation. A report should identify the predictor’s scale and the link, rather than presenting a coefficient without its context.

A Short Derivation for Reporting Slopes

For two predictor values in the same cluster, \(x\) and \(x+\Delta\), subtract their linear predictors. The intercept and intercept deviation cancel:

\begin{aligned} \eta_j(x+\Delta)-\eta_j(x) &=\left[\beta_0+\beta_1(x+\Delta)+b_{0j}+b_{1j}(x+\Delta)\right]\\ &\quad-\left[\beta_0+\beta_1x+b_{0j}+b_{1j}x\right]\\ &=(\beta_1+b_{1j})\Delta. \end{aligned}

Thus, the cluster-specific slope is the fixed slope plus the cluster’s random slope deviation; it is not the random deviation alone. For a logit link, exponentiating the difference gives the conditional odds ratio for that change in \(x\), \( \exp\{(\beta_1+b_{1j})\Delta\}\). For a log link, the corresponding ratio of conditional means is the same exponential expression. The intercept deviation changes the cluster’s starting level but cancels from this within-cluster slope contrast.

Worked Example: Reporting Logistic Effects for a Cluster

Suppose a hypothetical logistic random-slope model uses \(x\), measured in tens of hours of a program, to predict whether a participant meets a binary outcome criterion. Its fitted fixed effects are \(\hat\beta_0=-1.20\) and \(\hat\beta_1=0.50\). For cluster A, the fitted random-effect deviations are \(\hat b_{0A}=0.40\) and \(\hat b_{1A}=-0.15\). The logit link is \(g(\mu)=\log\{\mu/(1-\mu)\}\).

Common association. Holding the cluster random effects fixed, a one-unit increase in \(x\), or ten additional program hours, changes the linear predictor by \(0.50\). The corresponding conditional odds ratio is

\exp(\hat\beta_1)=\exp(0.50)=1.649.

The model therefore estimates 1.649 times the conditional odds per additional ten hours for a cluster whose random slope deviation is zero. This is an odds comparison, not a 64.9-percentage-point increase in probability.

Cluster A’s association. Its fitted slope is \(0.50+(-0.15)=0.35\), so its conditional odds ratio per ten hours is \(\exp(0.35)=1.419\). The cluster’s slope is less positive than the common slope, but remains positive.

Cluster A’s intercept and fitted values. Its intercept is \(-1.20+0.40=-0.80\), referring to \(x=0\). At \(x=1\), its linear predictor is \(-0.80+0.35=-0.45\). The corresponding conditional probabilities are \(\operatorname{logit}^{-1}(-0.80)=0.310\) at \(x=0\) and \(\operatorname{logit}^{-1}(-0.45)=0.389\) at \(x=1\). These are fitted probabilities conditional on cluster A’s estimated random effects.

Concise report. “For a cluster with random effects at zero, each additional ten program hours was associated with an estimated 1.65-fold increase in conditional odds. Cluster A’s fitted intercept deviation was 0.40 log-odds units and its slope deviation was \(-0.15\), giving it a fitted slope of 0.35 log-odds units per ten hours, or a conditional odds ratio of 1.42.” The wording describes an association, not a causal effect.

Worked Example: Reporting a Count-Model Slope

Consider a hypothetical Poisson model with a log link for a count outcome per observation. Let \(x\) be a predictor measured in days. Suppose \(\hat\beta_0=0.70\), \(\hat\beta_1=-0.12\), and the estimated random effects for cluster C are \(\hat b_{0C}=-0.20\), \(\hat b_{1C}=0.05\).

The common link-scale association is a decrease of \(0.12\) in the log conditional mean per day. Exponentiating gives a conditional mean ratio of \(\exp(-0.12)=0.8869\): for a cluster with random slope deviation zero, the expected count is multiplied by about 0.887 for each one-day increase in \(x\), conditional on the model.

For cluster C, the fitted slope is \(-0.12+0.05=-0.07\), so its mean ratio per day is \(\exp(-0.07)=0.9324\). Its fitted intercept is \(0.70-0.20=0.50\). At \(x=3\), its linear predictor is \(0.50+(-0.07)(3)=0.29\), and its fitted conditional mean is \(\exp(0.29)=1.336\) counts per observation.

A concise report would state that the common association corresponds to a 0.887 conditional mean ratio per day, while cluster C’s estimated slope deviation of \(0.05\) changes its slope to \(-0.07\), corresponding to a 0.932 ratio per day. The fitted mean at \(x=3\) is about 1.34 counts per observation for this cluster. The mean at that covariate value should not be confused with a ratio, and neither statement is a population-average prediction.

Presenting Estimates Without Overclaiming

Random-effect estimates are commonly reported as cluster-specific predictions or conditional modes, often called empirical Bayes estimates. They are estimates informed by both the cluster’s data and the fitted distribution of random effects; they are not known cluster parameters. In particular, these estimates are typically shrunk toward zero when a cluster provides limited information. A report should label them as estimated deviations and avoid implying that clusters with larger deviations necessarily differ reliably.

When available, report uncertainty for the fixed effects and for the cluster-specific quantities using an appropriate inferential procedure. A point estimate alone does not establish that a slope differs from zero or that two clusters differ. Uncertainty for a cluster’s slope must account for uncertainty in both the fixed slope and its random deviation, including their dependence. Do not construct a confidence interval by simply adding the endpoints of separate intervals unless the method specifically justifies that calculation.

A compact report can make the structure clear without listing every model component in prose. A table can show the fixed slope, each selected cluster’s random slope deviation, the resulting cluster-specific slope, and the transformed ratio when useful. If intercepts are also reported, include the reference value \(x=0\) and the predictor’s centering. Keep link-scale coefficients distinct from transformed effects.

Diagnostics, Alternatives, and Limitations

Before interpreting a fitted summary, check whether the model supports it. Convergence warnings or unstable estimates may make the fixed and random components unreliable. A random-slope variance estimated near zero, a singular covariance estimate, or a correlation estimate near its boundary can indicate that the data do not support the full random-effects structure well. Sparse information within clusters can make individual cluster slopes especially uncertain. Residual patterns inconsistent with the specified mean structure, or clusters that strongly influence the fit, also deserve attention. These findings call for investigation; they do not by themselves identify a single replacement model.

The main structural alternative is a random-intercept-only model, which allows cluster baselines to differ but imposes a common slope. It may be more stable when the data provide little information about slope variation, but it cannot represent genuine cluster-specific associations. A fixed-effects-only generalized linear model (GLM) has no such random cluster deviations and therefore does not express this modeled heterogeneity. Marginal approaches target population-averaged associations rather than the conditional cluster-specific associations described here; for nonlinear links, those targets need not coincide.

No one-sentence summary resolves every limitation. A random-slope model assumes a particular distribution for cluster effects and a linear trend on the link scale. Its coefficient is tied to the selected predictor scale and reference point. Cluster-specific estimates can be imprecise, and transformed effects such as odds ratios should not be mistaken for changes in probability. State what the model estimates, at what level, and on which scale.

Common pitfall: “The fixed slope is 0.50 and cluster A’s random slope is \(-0.15\), so cluster A’s slope is \(-0.15\).” This confuses a deviation with the cluster-specific slope. The correct calculation is \(0.50+(-0.15)=0.35\). For a logit link, the associated conditional odds ratio per unit is then \(\exp(0.35)\), not \(\exp(-0.15)\).

Worked Example: Turning a Partial Output Into a Report

A hypothetical log-link model has fixed slope \(\hat\beta_1=0.20\). Cluster D has estimated random slope deviation \(\hat b_{1D}=-0.08\) and random intercept deviation \(\hat b_{0D}=0.30\). The predictor \(x\) is measured in weeks, with zero defined as the chosen reference week.

The fixed slope implies a conditional mean ratio of \(\exp(0.20)=1.221\) per week when the random slope deviation is zero. Cluster D’s slope is \(0.20-0.08=0.12\), with a conditional mean ratio of \(\exp(0.12)=1.127\) per week. Its intercept deviation raises its linear predictor at \(x=0\) by 0.30 relative to the fixed intercept. The intercept deviation does not alter the slope.

One suitable report is: “At the reference cluster effect, the fitted conditional mean was multiplied by 1.22 per additional week. Cluster D’s estimated slope deviation was \(-0.08\), giving a cluster-specific slope of 0.12 and a conditional mean ratio of 1.13 per week; its fitted intercept deviation was 0.30 link-scale units.” This report includes the fixed association and cluster-specific departures without claiming that the point estimates are certain or population-average.

Check Your Understanding

Answer each question, then compare with the solutions below.

A logit model has \(\hat\beta_1=0.30\) and a cluster’s estimated slope deviation is \(0.10\). What is its cluster-specific slope and conditional odds ratio per unit?
For a log-link model, \(\hat\beta_0=0.40\), \(\hat\beta_1=-0.20\), and a cluster’s deviations are \(\hat b_0=0.10\), \(\hat b_1=0.05\). Find its intercept, slope, and mean ratio per one-unit increase.
Why is it inaccurate to say that a random slope deviation of \(-0.12\) means the cluster’s slope is \(-0.12\)?
What does the cluster-specific intercept refer to if \(x\) has been centered at its sample mean?

Solutions. (1) The slope is \(0.30+0.10=0.40\); the conditional odds ratio is \(\exp(0.40)=1.492\) per unit. (2) The intercept is \(0.40+0.10=0.50\); the slope is \(-0.20+0.05=-0.15\); the conditional mean ratio is \(\exp(-0.15)=0.8607\) per unit. (3) A random slope is a deviation from the fixed slope, so the cluster-specific slope is the fixed slope plus that deviation. (4) It is the linear predictor at the sample-mean value of the original predictor, for that cluster’s random effects; it is not the intercept at the original \(x=0\).