Writing GLMM Results with Clear Units and Contrasts

Why Units and Contrasts Belong in the Result

A coefficient is not a complete finding by itself. The numerical value \(-0.5108\), for example, does not tell a reader whether the outcome is a binary event, a count, or a continuous measurement; whether a predictor changes by one day or ten milligrams; or which groups are being compared. In a generalized linear mixed model (GLMM), a clear result also identifies the grouping level relevant to the effect. These details let a reader understand what the estimate describes without mistaking a model-scale coefficient for a direct difference in outcome probabilities or counts.

This tutorial focuses on turning fitted fixed-effect results into concise prose. It does not replace model assessment: an interpretation is meaningful only relative to an adequately specified model and its assumptions. The earlier material on conditional effects supplies the broader distinction between effects conditional on random effects and population-averaged effects; here the emphasis is on making that distinction explicit in written results.

Definition: A contrast is a specified comparison, such as a one-unit increase in a continuous predictor or treatment versus control. A fixed-effect coefficient represents a contrast on the model’s linear-predictor scale, conditional on the other model terms. With a nonlinear link, a coefficient is not generally a direct change in the outcome’s probability or expected value. A reported result should identify the contrast, the outcome scale, the predictor unit, any reference category, and the grouping level relevant to the effect.

What Must Be in Place Before Interpreting a Coefficient

For a typical GLMM, observations \(Y_{ij}\) are indexed by grouping unit \(j\) and observation \(i\). Conditional on random effects \(b_j\) and the predictors, each outcome has a distribution from a specified family, and its conditional mean is related to a linear predictor by a link function:

g\{E(Y_{ij}\mid b_j,\mathbf{x}_{ij})\} =\eta_{ij} =\mathbf{x}_{ij}^{\mathsf T}\boldsymbol{\beta} +\mathbf{z}_{ij}^{\mathsf T}b_j.

Here \(g\) is the link, \(\boldsymbol{\beta}\) contains fixed-effect coefficients, and \(b_j\) contains random effects for grouping unit \(j\). The interpretation below assumes the chosen outcome family and link are appropriate, the fixed- and random-effect structure represents the intended comparisons, and the fitted coefficient and its uncertainty are reliable enough for the stated inference. Conditional independence of observations given the modeled predictors and random effects is also typically assumed. These are modeling assumptions, not guarantees supplied by polished prose.

A result should distinguish the outcome scale from the linear-predictor scale. For a logit link, fixed-effect coefficients are changes in log odds, and exponentiating a coefficient gives an odds ratio. For a log link, coefficients are changes in log conditional means, and exponentiating gives a ratio of conditional means. Neither transformation alone yields an absolute probability or count difference. For a random-intercept model, the resulting contrast is conditional on the same grouping-unit random effect; say so when it matters to the interpretation.

Reporting rule: Name the modeled outcome and its scale; give the predictor contrast with units and reference category; report the estimate on an interpretable scale with uncertainty; and identify whether the effect is conditional on the modeled grouping effects. Use “associated with” unless the study design and assumptions justify a causal statement.

From Coefficient to Contextual Sentence

The following translation explains why exponentiated coefficients and contrasts should be described carefully. Suppose the conditional mean is \(\mu\), the link is the log link, and a continuous predictor \(x\) has coefficient \(\beta_x\). Increasing \(x\) by \(\Delta\) while holding other predictors and the random effect fixed changes the linear predictor by \(\beta_x\Delta\). Thus:

\log(\mu_{\text{new}})-\log(\mu_{\text{old}}) =\beta_x\Delta \quad\Longrightarrow\quad \frac{\mu_{\text{new}}}{\mu_{\text{old}}} =\exp(\beta_x\Delta).

The ratio depends on the chosen increment \(\Delta\). If the coefficient is per milligram, a ten-milligram contrast uses \(\Delta=10\); alternatively, the predictor can be rescaled so its coefficient directly represents that increment. For a logit link, the same argument applies to odds: the log odds difference is \(\beta_x\Delta\), and the odds ratio is \(\exp(\beta_x\Delta)\). This derivation justifies reporting a ratio for those links, but it does not justify calling that ratio a change in probability or an absolute difference.

Worked Example: Treatment Versus Placebo on the Odds Scale

Hypothetical model. Suppose a binary outcome indicates whether a patient experienced an adverse event during follow-up. A logit-link GLMM includes treatment as a fixed effect and a random intercept for clinic. Treatment is coded 1 for the new treatment and 0 for placebo, the reference category. Assume the model and its clustering structure are appropriate for the data.

The fitted treatment coefficient is \(\hat{\beta}=-0.5108\), with \(SE=0.2000\). The coefficient is the treatment-versus-placebo contrast in log odds, conditional on clinic random intercept and the other modeled covariates. Exponentiating gives the conditional odds ratio:

\exp(-0.5108)=0.6000.

A 95% Wald confidence interval for the log-odds coefficient is \(-0.5108\pm1.96(0.2000)\). Its endpoints are \(-0.9028\) and \(-0.1188\); exponentiating both gives an odds-ratio interval of approximately \((0.4054,\ 0.8880)\).

For the two-sided Wald test of \(H_0:\beta=0\), the statistic is \(z=-0.5108/0.2000=-2.554\), giving \(p\approx0.01065\). The estimate is statistically inconsistent with a zero log-odds contrast at the 0.05 level under the model and Wald approximation. It does not give the probability that the null is true.

A concise report is: “In the clinic-conditional logit model, treatment was associated with lower odds of an adverse event than placebo (odds ratio 0.6000, 95% Wald CI 0.4054 to 0.8880; \(p=0.01065\)), conditional on clinic and the adjusted covariates.” The odds are estimated to be 40% lower, but this is not a claim that the event probability is 40% lower. An absolute probability contrast would require predicted probabilities at specified covariate and random-effect values or a population-averaged calculation.

Worked Example: A Count Ratio Per Ten Milligrams

Hypothetical model. Suppose a log-link Poisson GLMM models the number of unscheduled visits per patient during a common 12-month follow-up, with a random intercept for clinic. A continuous medication-dose predictor is recorded in milligrams, and its fitted coefficient is \(0.01823\) per milligram, with \(SE=0.00600\). Assume the conditional count distribution, link, and clinic random-intercept structure are reasonable for this illustration.

A one-milligram coefficient is not the most useful contrast here. For a ten-milligram increase, the linear-predictor contrast is \(10(0.01823)=0.1823\). The corresponding ratio of conditional expected counts is:

\exp\{10(0.01823)\}=\exp(0.1823)=1.2000.

The standard error for this ten-milligram contrast is \(10(0.00600)=0.06000\), because the contrast is ten times the coefficient. The 95% Wald interval on the log scale is \(0.1823\pm1.96(0.06000)=(0.0647,\ 0.2999)\). Exponentiating yields a count-ratio interval of approximately \((1.0668,\ 1.3497)\).

A suitable sentence is: “Within the same clinic random-intercept level and at the same values of the other covariates, a 10-mg higher dose was associated with an estimated 1.20-fold conditional expected number of unscheduled visits over 12 months (95% Wald CI 1.067 to 1.350).” The common 12-month observation period is stated so the outcome’s time scale is clear. Because no exposure offset is specified, this is a ratio of expected counts over that period, not a rate ratio per unit of exposure.

Contrasts Involving Interactions and Reference Categories

With an interaction, one coefficient may not represent the comparison a reader wants. For example, in a model with treatment, subgroup, and their interaction, the treatment coefficient is the treatment contrast only in the reference subgroup. To report the treatment contrast in another subgroup, add the treatment coefficient and the interaction coefficient. Its uncertainty must use their covariance; simply adding standard errors is incorrect.

Worked Example: Treatment Effect in a Nonreference Subgroup

Hypothetical model. A binary outcome is modeled with a logit link, fixed effects for treatment and subgroup, and their interaction, plus a clinic random intercept. Control and subgroup A are the reference categories. Let \(\hat{\beta}_T=-0.30\) be the treatment coefficient and \(\hat{\beta}_{T\times B}=0.50\) the treatment-by-subgroup-B interaction. Assume the model assumptions are adequate for this illustration.

The treatment contrast in subgroup B is \(-0.30+0.50=0.20\) log odds, not \(-0.30\). Suppose the estimated variances are \(0.04\) and \(0.09\), and the covariance is \(-0.01\). The variance of the sum is \(0.04+0.09+2(-0.01)=0.11\), so its standard error is \(\sqrt{0.11}=0.3317\).

The Wald statistic is \(0.20/0.3317=0.603\), with a two-sided \(p\)-value of approximately \(0.546\). The 95% Wald interval on the log-odds scale is \(0.20\pm1.96(0.3317)=(-0.4501,\ 0.8501)\). The conditional odds ratio is \(\exp(0.20)=1.221\), with interval \((\exp(-0.4501),\exp(0.8501))\), approximately \((0.638,\ 2.340)\).

A precise statement is: “Among patients in subgroup B, the estimated treatment-versus-control conditional odds ratio was 1.221 (95% Wald CI 0.638 to 2.340; \(p=0.546\)), conditional on clinic and the other modeled terms.” The interval includes 1, and the test does not provide strong evidence of a nonzero treatment contrast in subgroup B. This does not establish that treatment has no effect. The reference categories and subgroup are essential: quoting the treatment coefficient alone would instead describe subgroup A.

Writing, Checking, and Choosing the Right Scale

A reliable reporting process makes the contrast explicit before drafting the sentence. It also checks that the reported scale matches the model, that the units survive any rescaling, and that uncertainty belongs to the exact contrast being described.

State the modelled outcome.
Name the event, count, or measurement and the relevant follow-up or measurement period. Identify the link-based scale, such as odds or conditional expected counts.

Write the contrast in words.
Specify the predictor increment and its unit, or name both categories and the reference category. For an interaction, calculate the desired linear combination of coefficients.

Match estimate and uncertainty.
Report the coefficient or transformed contrast and its confidence interval on the same interpretable scale. Use the covariance matrix for a combined contrast.

Name the conditioning level and qualify the claim.
State the relevant grouping level for a conditional effect, then describe association rather than causation unless the design and assumptions support a causal interpretation.

Several diagnostics can reveal that a draft is not communicating the fitted result faithfully. Check the model formula and coding to confirm which category is the reference, and verify the predictor’s recorded and reported units. Recompute exponentiated estimates and interval endpoints from the coefficient-scale values; a ratio interval must be positive. For an interaction, compare the stated contrast with the actual linear combination and its covariance-based standard error. Check whether the model converged and whether the estimate and standard error are plausible; extreme estimates, very large standard errors, sparse categories, or unstable fits should prompt caution rather than confident prose. Also inspect whether the assumed outcome distribution and random-effects structure are credible. A sentence cannot fix separation, overdispersion, influential clusters, a misspecified link, or an inadequate random-effects structure.

The main alternatives answer different questions. A GLM without random effects may not represent within-cluster dependence when clustering matters. A generalized estimating equation targets a population-averaged relationship under its working correlation approach, whereas a GLMM fixed effect is commonly interpreted conditional on random effects; these are not automatically interchangeable estimands. Predicted probabilities or expected counts can make an effect more tangible, but they depend on the specified covariate values and on whether predictions condition on or average over random effects. If presenting them, give that calculation’s basis rather than replacing an odds ratio or count ratio with an unqualified probability difference.

A further limitation is that confidence intervals and p-values often rely on approximate large-sample inference, especially for Wald procedures. When data are sparse or the number of grouping units is small, those approximations may be unreliable. A report should not imply greater precision than the analysis supports. Finally, a statistically detectable conditional association does not by itself show clinical importance, establish a causal effect, or describe how much an individual clinic’s outcome changes.

Common Mistakes

Reporting a coefficient without its scale. “The coefficient was 0.18” is incomplete. State whether it is log odds, log conditional mean, or another linear-predictor quantity, and translate it when useful.
Calling an odds ratio a probability ratio. Exponentiating a logit coefficient gives an odds ratio. It does not give a probability difference, and the probability contrast depends on baseline risk.
Leaving units implicit. “Dose was associated with more visits” omits whether the contrast is per milligram, per ten milligrams, or another increment.
Using the wrong treatment contrast. With an interaction, the main treatment coefficient may apply only to the reference subgroup. Calculate and report the intended subgroup contrast.
Omitting the grouping level. Do not present a clinic-conditional effect as though it were automatically a population-averaged effect.
Overstating nonsignificance. A large p-value is not proof of no association. Report the estimate and interval, and describe the uncertainty.

Key takeaway: A clear GLMM result is a contextual contrast, not a detached coefficient. State the outcome scale, predictor unit, comparison and reference category, uncertainty, and relevant grouping level; then make only the interpretation that the fitted model and study design support.

Check Your Understanding

Answer each item by identifying the contrast and scale, then compare with the brief solutions.

A logit-link GLMM gives a coefficient of \(-0.40\) for treatment coded 1 versus control coded 0. What is the odds ratio, and what key qualification should accompany it?
A log-link model gives a coefficient of \(0.03\) per one-unit increase in a predictor. What is the conditional mean ratio for a four-unit increase?
In a treatment-by-subgroup model, the treatment coefficient is \(-0.20\) and the interaction coefficient for subgroup B is \(0.35\). What is the treatment contrast in subgroup B on the linear-predictor scale?
Why is “the event probability was 40% lower” not justified by an odds ratio of \(0.60\) alone?

Solutions: 1. The odds ratio is \(\exp(-0.40)=0.6703\). Describe it as treatment versus the stated control reference, on the odds scale, conditional on the modelled grouping effects and covariates where applicable. 2. The contrast is \(4(0.03)=0.12\), so the conditional mean ratio is \(\exp(0.12)=1.127\). Name the predictor’s unit and the outcome period in an applied report. 3. The subgroup-B contrast is \(-0.20+0.35=0.15\), not \(-0.20\); exponentiating gives an odds ratio of \(\exp(0.15)=1.162\). 4. An odds ratio compares odds, not probabilities. Converting it to a probability difference requires a baseline probability and a clearly specified prediction context.