Briefing a Clinical Team on GLMM Findings

From a fitted model to a clinical briefing

A clinical team needs more than a coefficient copied from software. It needs to know what event the model describes, which treatment groups are being compared, what scale the estimate is on, and which patients the analysis represents. A short briefing can answer those questions while preserving the distinction between a model-based association and a broader clinical claim.

The examples below use hypothetical studies. The focus is not on fitting a model, but on expressing its results accurately. In particular, a binary GLMM treatment coefficient is usually on the log-odds scale and, with a random intercept, describes a treatment contrast conditional on that intercept. It is not automatically a risk difference or a population-averaged effect.

State what the fitted model represents

Before interpreting an estimate, identify the modeled event and the population to which the analysis directly pertains. For example, “readmission” is not fully specified unless the briefing says whether it means readmission within 30 days, among whom, and under what treatment comparison. The analyzed sample should be described using information available from the study: for example, the number of participants and sites, the treatment groups, and any exclusions that define the analysis set. Avoid implying that the analyzed sample includes people or settings that were not observed.

For a binary outcome \(Y_{ij}\), a common random-intercept logistic GLMM is:

Y_{ij}\mid b_j \sim \operatorname{Bernoulli}(p_{ij}), \qquad \operatorname{logit}(p_{ij})=\beta_0+\beta_1 T_{ij}+b_j,\qquad b_j\sim N(0,\tau^2).

Here, \(i\) indexes participants and \(j\) indexes sites, \(T_{ij}=1\) denotes the specified treatment and \(T_{ij}=0\) its reference group, and \(p_{ij}=P(Y_{ij}=1\mid b_j,T_{ij})\). The logit is \(\operatorname{logit}(p)=\log\{p/(1-p)\}\). The model assumes that, conditional on the site random intercept and modeled predictors, outcomes follow the stated Bernoulli distribution and are independent; it also assumes the random intercepts follow the specified normal distribution. A briefing does not need to reproduce every model detail, but its interpretation must be compatible with the model that was actually fit.

Definition: In this random-intercept logistic GLMM, \(\beta_1\) is the treatment contrast in conditional log odds, holding the site random intercept fixed. Its exponent, \(\exp(\beta_1)\), is the corresponding conditional odds ratio. It compares treatment-group odds at the same value of the modeled random intercept; it is not, by itself, a ratio of marginal risks.

Why exponentiating the coefficient gives an odds ratio

The model definition supplies the interpretation. At a fixed site effect \(b_j\), the odds for \(T_{ij}=1\) are \(\exp(\beta_0+\beta_1+b_j)\), while the odds for \(T_{ij}=0\) are \(\exp(\beta_0+b_j)\). Their ratio is:

\frac{\exp(\beta_0+\beta_1+b_j)}{\exp(\beta_0+b_j)} =\exp(\beta_1).

The intercept and site effect cancel in this within-site conditional comparison. This is why a briefing may report \(\exp(\hat{\beta}_1)\) as a conditional odds ratio, provided it names the event and the direction of the contrast. It should not silently relabel that number as a risk ratio, a percentage-point change, or an effect averaged over sites.

If the estimate has standard error \(SE\), a large-sample Wald interval on the coefficient scale is \(\hat{\beta}_1\pm 1.96SE\). Exponentiating both endpoints gives an interval for the conditional odds ratio. The associated two-sided Wald test uses \(z=\hat{\beta}_1/SE\) and compares the statistic with a standard normal distribution under \(H_0:\beta_1=0\). These are approximate procedures; the model must be adequately fit, and the approximation may be unreliable with sparse data or poorly estimated parameters.

Worked Example: Briefing a Treatment Contrast for 30-Day Readmission

Suppose a hypothetical randomized study analyzed 480 adults at 12 hospitals, with 240 assigned to a new intervention and 240 to usual care. The binary outcome is whether a participant was readmitted within 30 days. A logistic random-intercept model uses \(T=1\) for the new intervention and \(T=0\) for usual care. Suppose software reports \(\hat{\beta}_1=-0.510\) and \(SE=0.200\).

Check what the estimate means. The event is 30-day readmission, and the coefficient compares new intervention with usual care, conditional on the hospital random intercept. The estimated conditional odds ratio is:

\exp(-0.510)=0.6005.

Thus, at the same modeled hospital effect, the estimated odds of readmission in the intervention group are 0.6005 times the odds in the usual-care group, or about 40.0% lower in odds. This is not a statement that the probability of readmission is 40.0% lower.

Quantify uncertainty. The Wald statistic is \(z=-0.510/0.200=-2.55\), giving a two-sided \(p\)-value of approximately 0.0108. The 95% interval on the log-odds scale is:

-0.510\pm1.96(0.200)=-0.510\pm0.392=[-0.902,\,-0.118].

Exponentiating the endpoints gives \([\exp(-0.902),\exp(-0.118)]=[0.4058,\,0.8887]\), approximately. Under the model and its large-sample approximation, the interval procedure has 95% coverage in repeated sampling. The interval excludes 1 on the odds-ratio scale, consistent with the two-sided test at the 0.05 level. This result is evidence against \(H_0:\beta_1=0\); it does not establish that the treatment will reduce readmissions by the same amount in every hospital or patient population.

A concise briefing. “Among the 480 adults analyzed at 12 hospitals, the modeled outcome was readmission within 30 days. Comparing the new intervention with usual care, the estimated hospital-conditional odds ratio was 0.6005 (95% CI 0.4058 to 0.8887; Wald \(p=0.0108\)). The estimate is consistent with lower conditional odds of readmission in the intervention group; it is not a direct estimate of the reduction in readmission probability.”

Make contrast direction explicit

A coefficient’s sign depends on which group is coded as 1 and which group is the reference. A positive coefficient does not inherently mean benefit, and a negative coefficient does not inherently mean harm: the event and contrast direction determine the interpretation. If the contrast is reversed, the coefficient changes sign and the odds ratio becomes its reciprocal. State the comparison in words whenever reporting the estimate.

Worked Example: Reversing the Treatment Comparison

In a second hypothetical fit, the modeled event is treatment failure and the reported coefficient is for usual care relative to a new treatment. Suppose \(\hat{\beta}=0.405\) and \(SE=0.180\). The estimated odds ratio for usual care compared with the new treatment is:

\exp(0.405)=1.499.

The Wald statistic is \(0.405/0.180=2.25\), giving a two-sided \(p\)-value of approximately 0.0244. The 95% interval on the coefficient scale is:

0.405\pm1.96(0.180)=0.405\pm0.3528=[0.0522,\,0.7578].

On the odds-ratio scale this is \([\exp(0.0522),\exp(0.7578)]=[1.0536,\,2.1336]\). The appropriate wording is that the estimated odds of failure under usual care are 1.499 times the estimated odds under the new treatment, conditional on the modeled random effects. If the team instead wants the new treatment relative to usual care, reverse the contrast: the log odds ratio is \(-0.405\), and the odds ratio is \(\exp(-0.405)\approx0.6670\). The reversed interval is approximately \([0.4687,\,0.9491]\), with endpoints obtained by taking reciprocals in reverse order. Do not keep the original estimate while swapping the comparison labels.

Do not turn an odds ratio into a risk difference

An odds ratio alone does not determine an absolute change in probability. The relationship between odds and probability depends on the starting probability. A team may find probabilities easier to interpret, but probabilities must be calculated for a specified covariate and random-effect setting, or obtained by an appropriate population-averaging procedure. Name which kind of probability is being presented.

Worked Example: Conditional Probabilities at a Specified Site Effect

Return to the first hypothetical model and suppose its fitted intercept is \(\hat{\beta}_0=-2.000\). For illustration, calculate probabilities at a site random intercept of \(b_j=0\). This is a conditional calculation at that specified value, not an average over hospitals.

For usual care (\(T=0\)), the linear predictor is \(-2.000+0= -2.000\), so:

\hat{p}_{0}=\frac{1}{1+\exp(2.000)}\approx0.1192.

For the intervention (\(T=1\)), the linear predictor is \(-2.000-0.510=-2.510\), giving:

\hat{p}_{1}=\frac{1}{1+\exp(2.510)}\approx0.0752.

At \(b_j=0\), the fitted conditional probabilities are therefore about 11.92% under usual care and 7.52% under the intervention, a difference of \(0.0752-0.1192=-0.0440\), or about 4.40 percentage points lower. These values illustrate the model’s conditional predictions at \(b_j=0\). They are not the population-averaged risks for all hospitals; averaging predicted probabilities over the random-effect distribution and the study population is a different calculation.

Checks before delivering the briefing

A polished sentence cannot repair a model that failed to fit or a result whose target population is unclear. Before presenting a fitted binary GLMM, check that the outcome coding and treatment reference group match the words in the briefing, that the intended analysis sample and observation period are identified, and that the fit produced sensible estimates. Review convergence messages, unusually large standard errors, sparse events within treatment groups or sites, and whether the estimated uncertainty is plausible. A non-converged fit or near-separation can make a coefficient and its Wald interval unreliable; in that situation, do not present the numbers as settled findings without further investigation.

Also check that the reported uncertainty is on the same scale as the estimate: a coefficient interval is on the log-odds scale, whereas an odds-ratio interval is exponentiated. Verify that the interval and test agree when they use the same two-sided 0.05 Wald procedure. If they do not, explain the difference in methods rather than selecting whichever result appears more favorable. A small \(p\)-value does not measure clinical importance, and a large one does not show that the treatments are equivalent.

Name the event and time window.
Say what \(Y=1\) represents and when it was assessed, such as readmission within 30 days.

Define the analyzed population.
Give the analyzed sample and relevant setting, and avoid extending the statement to unobserved groups.

State the comparison and scale.
Identify treatment and reference groups, then report the coefficient or conditional odds ratio with the scale labeled.

Report uncertainty and qualify the conclusion.
Give a confidence interval and, when appropriate, a \(p\)-value; describe model-based evidence without claiming certainty or unsupported clinical effects.

Common pitfalls and alternatives

A frequent mistake is to say, “The treatment reduced readmission risk by 40%,” after observing an odds ratio of 0.6005. That wording treats a reduction in odds as an equal reduction in probability. The correction is to say “estimated 40.0% lower conditional odds,” or to report probabilities calculated for a clearly specified setting. Another mistake is to omit the contrast direction and say only “the odds ratio was 1.499.” Without naming which treatment is in the numerator, that number is ambiguous.

The main alternatives are complementary, not interchangeable. A coefficient and its standard error preserve the model’s log-odds scale; an exponentiated coefficient is often easier to communicate but remains a conditional odds ratio. Conditional predicted probabilities can put results on an absolute scale when their covariate and random-effect settings are stated. Population-averaged risks or risk differences can be useful for a broader clinical summary, but they require averaging or standardizing predictions over a defined population. They should not be inferred by casually relabeling the conditional odds ratio. In every approach, the conclusion remains limited by the study design, model assumptions, and analyzed population.

Key takeaway: A useful clinical briefing joins four things: the modeled event, the analyzed population, the treatment contrast, and the estimate scale. State uncertainty, keep conditional odds distinct from probabilities, and describe evidence without turning a model result into a stronger claim than it supports.

Check Your Understanding

Answer each question, then check the brief solutions below.

A logistic random-intercept model uses \(T=1\) for a new treatment and \(T=0\) for usual care, with 30-day event \(Y=1\). What does \(\exp(\beta_1)\) compare?
A fitted coefficient is \(-0.300\). Give the odds ratio to three significant figures and state what additional information is required to interpret its direction clinically.
A briefing reports an odds ratio of 0.70 as “a 30% reduction in event probability.” What is wrong, and how should the statement be corrected?
A 95% interval for a conditional odds ratio is 0.80 to 1.25. What does this interval say about the treatment contrast, and what does it not establish?

Solutions. (1) It compares the conditional odds of \(Y=1\) under the new treatment with the conditional odds under usual care, holding the random intercept fixed. (2) \(\exp(-0.300)\approx0.741\). Interpretation also requires the outcome definition and the stated contrast; with the given coding, it indicates lower estimated conditional odds of the event under the new treatment. (3) The odds ratio describes odds, not probability. Say “30% lower conditional odds,” or report probabilities from a specified prediction or population-averaging calculation. (4) The interval includes 1, so it is compatible with both lower and higher conditional odds at the 95% confidence level under the procedure. It does not prove no effect, equivalence, or clinical unimportance.