Why Reporting Needs More Than a Group Estimate
A reported group prediction can be misunderstood if it is given without its target, scale, and uncertainty definition. For example, a number such as 0.7 might be a predicted log-odds deviation, a predicted odds ratio after transformation, or a contribution to a linear predictor at a specified covariate value. These quantities are related, but they are not interchangeable. A clear report names which quantity is being predicted and what comparison it represents.
In a generalized linear mixed model, an empirical Bayes (EB) group prediction uses the observed data for a group together with the fitted random-effects distribution. It is therefore typically shrunk toward the model’s population-level center. It is not the group’s raw response average, nor is it known to equal the group’s true latent effect. Reporting should also make clear which model parameters were treated as fixed when the prediction and its uncertainty were calculated.
This description assumes that the stated GLMM is an adequate model for the response distribution, link, fixed-effects structure, random-effects distribution, and dependence within groups. It also assumes that the model has been fit successfully enough to provide usable estimates \(\hat\theta\). An uncertainty summary based on \(p(b_j\mid Y_j;\hat\theta)\) holds those fitted parameters fixed; it does not by itself account for uncertainty in estimating them.
A concise report can be built from five questions: What is the target? On what scale? Which fitted quantities are plugged in? What uncertainty measure is reported? What does the result not establish? Answering these in a sentence or short table makes the result interpretable without suggesting more certainty than the model supports.
Choose and Name the Prediction Target
A random-effect prediction and a response prediction answer different questions. The first concerns a group’s latent deviation, such as a random intercept \(b_j\), from the model’s population-level center. The second concerns a response-scale quantity for a specified covariate pattern, such as a conditional event probability or rate. When reporting a response prediction, state the covariate values and any exposure or offset used.
For a GLMM, the plug-in conditional response mean at covariate values \(x\) and \(z\) is \(g^{-1}(x^\mathsf{T}\hat\beta+z^\mathsf{T}\hat b_j)\). This is a prediction conditional on the fitted group effect. In general it is not the same as averaging \(g^{-1}(x^\mathsf{T}\hat\beta+z^\mathsf{T}b_j)\) over the conditional distribution of \(b_j\). Because \(g^{-1}\) is nonlinear for common GLMM links, the order of applying the inverse link and averaging matters.
Worked Example: Reporting a Binary-Outcome Random Intercept
Suppose a hypothetical logistic random-intercept model has conditional log odds \(-1.20+b_j\) for a specified reference covariate pattern, with the population random-intercept center set at zero. For one observed group, the fitted conditional mode is \(\hat b_j=0.70\). A plug-in conditional 95% interval for \(b_j\), calculated with \(\hat\theta\) held fixed, is \([-0.02,1.27]\). The task is to report the group effect and its interpretation.
Target and scale. The target is the group’s latent random-intercept deviation on the log-odds scale, not its probability of an event. The point prediction is \(0.70\) log-odds units relative to a group with random intercept zero.
Optional transformed summary. Exponentiating gives the odds ratio relative to that zero-random-intercept reference: \(e^{0.70}=2.014\), approximately 2.01. Transforming both interval endpoints gives \([e^{-0.02},e^{1.27}]=[0.980,3.56]\). Thus the report could say: “The fitted group random-intercept mode was 0.70 log-odds units (conditional plug-in 95% interval −0.02 to 1.27), corresponding to an odds ratio of 2.01 (interval 0.98 to 3.56) relative to a group at the model’s random-intercept center.”
The interval includes zero on the log-odds scale and one on the odds-ratio scale. This interval alone therefore does not clearly distinguish the group’s effect from the reference value. It is a conditional plug-in interval, not an interval that includes all uncertainty in the fitted model parameters.
A transformed point estimate and interval should remain tied to the same target. In the example, the odds ratio is a transformation of the random-intercept deviation. It is not the group’s event probability. To report a conditional probability, one would specify the covariates and calculate the inverse-logit of the full linear predictor, including the fixed effects and the group effect.
State the Plug-In Assumptions and Uncertainty
An EB prediction is empirical because the group-level distribution is estimated from the data. In a typical plug-in calculation, fitted fixed effects and random-effects parameters are inserted into the conditional distribution for \(b_j\), and that conditional distribution is summarized. A conditional standard deviation or interval describes uncertainty about \(b_j\) under that fitted model, given the observed group data. It does not automatically include the extra variability that would arise from refitting the model in new samples.
If uncertainty in fitted parameters is important, a procedure that refits the model across bootstrap samples can incorporate it. The report should identify the actual procedure—for example, “conditional plug-in interval” or “parametric-bootstrap interval from refitted models”—rather than simply calling either one a “95% confidence interval.” These labels describe different uncertainty calculations, and their results need not agree.
Worked Example: Reporting a Random-Slope Prediction at a Covariate Value
Consider a hypothetical Poisson GLMM with log link and exposure \(t\): \(\log\{E(Y_{ij}\mid b_j)\}=\log(t_{ij})+\beta_0+\beta_1x_{ij}+b_{0j}+b_{1j}x_{ij}\). Suppose the fitted values are \(\hat\beta_0=0.40\), \(\hat\beta_1=0.25\), \(\hat b_{0j}=0.15\), and \(\hat b_{1j}=-0.10\). At \(x=2\) and exposure \(t=1\), calculate and report the conditional rate.
The fixed-effects contribution is \(0.40+0.25(2)=0.90\). The predicted group-specific random-effect contribution at \(x=2\) is \(0.15-0.10(2)=-0.05\). Thus the conditional log rate is \(0.90-0.05=0.85\), and the rate is \(e^{0.85}=2.340\) expected events per exposure unit.
Suppose, for illustration, a parametric bootstrap that refits the model gives a 95% interval \([0.65,1.08]\) for the log rate at this same covariate value and exposure. Exponentiating the endpoints gives \([e^{0.65},e^{1.08}]=[1.916,2.945]\) expected events per exposure unit. A suitable report is: “At \(x=2\), the plug-in group-specific rate was 2.34 events per exposure unit; the 95% parametric-bootstrap interval, based on refitted models, was 1.92 to 2.95.”
This report specifies the covariate value, exposure scale, and uncertainty method. It describes a conditional group-specific rate, not a population-average rate. Nor does the rate interval imply that every future count will fall between its endpoints: it concerns uncertainty in the rate prediction, not the variation of a future count around that rate.
Why the Uncertainty Label Matters
A short derivation shows why conditioning on fitted parameters and accounting for parameter uncertainty are distinct. If uncertainty about \(\theta\) is included in a joint predictive distribution, the variance of a latent effect can be decomposed as
The first term is the average conditional variance of the group effect when \(\theta\) is specified. The second is the variation in its conditional mean as \(\theta\) varies. A plug-in calculation at \(\hat\theta\) summarizes the first kind of uncertainty at one fitted parameter value; it does not include the second term. In frequentist analyses, a refitting bootstrap is one way to assess the effect of parameter estimation, but it is not the same calculation as this joint-distribution decomposition.
Worked Example: Reporting a Gaussian Random-Intercept Prediction
Suppose a hypothetical Gaussian random-intercept model is fitted to measurements in a group, and the conditional distribution of that group’s intercept, with fitted model parameters held fixed, is approximately \(N(0.80,0.25^2)\), in outcome units. Report the conditional mean prediction with an approximate 95% conditional interval.
The point prediction is \(0.80\) outcome units. Using the normal approximation, the interval is \(0.80\pm1.96(0.25)=0.80\pm0.49=[0.31,1.29]\) outcome units. A clear report is: “The group’s conditional mean random-intercept prediction was 0.80 outcome units, with an approximate 95% conditional interval of 0.31 to 1.29, treating the fitted model parameters as fixed.”
This is an interval for the latent group effect under the fitted model. It is not an interval for an individual future observation, which would also reflect observation-level variation. It is also not a full parameter-uncertainty interval. Naming the target and conditioning assumption prevents both misreadings.
Diagnostics, Alternatives, and Limits
Before reporting a prediction, check that the fitted model supplies a credible basis for it. Relevant checks include convergence warnings, singular or boundary variance estimates, implausible parameter estimates, and whether the group has enough observations or covariate variation to inform its prediction. For binary outcomes, complete or near separation can make estimates unstable. If the conditional distribution is skewed, a symmetric estimate-plus-or-minus-standard-error interval may be misleading; quantile-based intervals can better reflect its shape. These checks do not prove that the model is correct, but they can reveal when a precise-looking prediction is poorly supported.
The main alternatives answer different questions. A raw group deviation uses the observed group data without the model’s shrinkage and can be highly variable, especially for small groups. A conditional EB mode or mean uses the fitted group distribution and group data to predict a latent effect. A response-scale conditional prediction applies the inverse link at a specified covariate pattern and predicted effect. A marginal response prediction instead averages over random-effect variation. Reports should not switch among these targets without saying so.
A prediction describes a model-based group deviation or response, not a causal effect of group membership. Shrinkage is part of the prediction procedure, not evidence that groups are truly identical. Rankings can be especially uncertain when group predictions are close together or individually imprecise. Predictions for groups with sparse data can be driven largely by the assumed random-effects distribution. Finally, conclusions apply only to the stated model and prediction target; they do not automatically generalize to a different population, covariate pattern, or model specification.
Say whether the result is a latent effect, a linear predictor, a conditional response mean, or another specified quantity.
State, for example, log-odds relative to a zero random intercept, or events per exposure unit at a stated covariate value.
Report whether the group prediction is a conditional mean or mode and which fitted quantities are treated as fixed.
Give the uncertainty measure and procedure, state whether parameter estimation is included, and avoid causal or broader claims.
Common Reporting Pitfall
Wrong reasoning: “The group’s estimated random intercept is 0.70, so its event probability is 0.70.” The estimate is on the model’s random-effect scale, which is not generally the response scale. A random intercept also contributes to a linear predictor alongside fixed effects and other terms.
Correction: State that 0.70 is a predicted random-intercept deviation on the log-odds scale. If a probability is needed, specify the covariates and calculate the inverse-logit of the full conditional linear predictor. Then identify whether the result is a plug-in conditional probability or an average over random-effect uncertainty. In each case, report uncertainty for that stated target rather than transferring an interval from a different scale or quantity.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A group random-intercept prediction is 0.40 on the log-odds scale. What is the corresponding odds ratio relative to a group with random intercept zero?
- Why should an interval based on \(p(b_j\mid Y_j;\hat\theta)\) not automatically be described as including parameter-estimation uncertainty?
- A report gives a conditional rate at \(x=1\). What additional information should accompany it so the target is clear?
- Why is a predicted random intercept not itself a predicted event probability?
Solutions. (1) The odds ratio is \(e^{0.40}=1.492\), approximately 1.49. (2) The calculation conditions on the fitted parameter estimate \(\hat\theta\); it does not account for variation from estimating \(\theta\). (3) State the covariate value, exposure or offset, response scale and units, and the uncertainty procedure. (4) The random intercept is a latent contribution to the linear predictor. A probability requires combining it with the fixed effects and other predictor terms, then applying the inverse link.