Why Rate-Ratio Reporting Needs Context
A count model with an exposure offset is designed to compare event rates when observations contribute different amounts of time, distance, or another quantity at risk. A rate ratio can summarize that comparison compactly, but it is easy to leave out information essential to interpretation: the exposure unit, the exact predictor contrast, and the group or covariate context. With interactions, even naming a predictor is not enough; the rate ratio may depend on the value of another predictor.
This tutorial focuses on how to report contrasts from an exposure-adjusted count GLMM. The aim is not simply to exponentiate a coefficient. It is to identify which model contrast answers the scientific question, quantify its uncertainty, and state what the comparison represents.
Model and Interpretation
Let \(Y_{ij}\) be a count for observation \(i\) in group \(j\), and let \(E_{ij}>0\) be its exposure, such as patient-days. A standard exposure-adjusted Poisson GLMM has a log link and a log-exposure offset:
Here \(\mu_{ij}=E(Y_{ij}\mid b_j)\), \(b_j\) is a group-level random effect, and the coefficient of \(\log(E_{ij})\) is fixed at one rather than estimated. A common specification assumes independent counts conditional on the random effects and covariates, and normally distributed random effects with mean zero and a specified covariance structure. The offset formulation also assumes that expected counts are proportional to exposure: doubling exposure doubles the conditional expected count when other terms are unchanged.
The word conditional matters: the interpretation holds at fixed random effects and covariates. When the contrast is between two predictor values for the same group, a shared random intercept cancels. If a relevant random slope is present, the contrast can vary across groups. State the group context accordingly, rather than presenting every conditional contrast as a universal population-average comparison.
From a Contrast to a Reported Rate Ratio
For a contrast between predictor settings \(A\) and \(B\), define the log-rate contrast as \(\Delta=\eta_A-\eta_B\), where \(\eta\) is the linear predictor after excluding the common offset. Then the rate ratio is \(\exp(\Delta)\). Without relevant interactions, a one-unit increase in a continuous predictor has log-rate contrast equal to its coefficient; a binary predictor’s coefficient compares its coded value of one with its reference value of zero. These shortcuts do not apply unchanged when interacting predictors differ across the contrast.
A Wald confidence interval for \(\Delta\), when its estimated standard error is available, is \(\hat{\Delta}\pm z_{1-\alpha/2}SE(\hat{\Delta})\). Exponentiating both endpoints gives a confidence interval for the rate ratio. This works because the exponential function is increasing, so it preserves the endpoint order.
When a contrast is a sum of estimated coefficients, its variance must include their covariance. For two coefficients, \(\hat{\Delta}=\hat{\beta}_1+\hat{\beta}_2\), the variance is \(\operatorname{Var}(\hat{\beta}_1)+\operatorname{Var}(\hat{\beta}_2)+2\operatorname{Cov}(\hat{\beta}_1,\hat{\beta}_2)\). Omitting the covariance generally gives the wrong standard error.
Name the predictor contrast, the exposure unit, and any interacting predictors that must be fixed at stated values.
Subtract the two relevant linear predictors. Include interaction terms that change across the contrast; shared terms cancel.
Use the fitted covariance matrix to obtain the contrast standard error, then calculate a confidence interval or test for the log contrast.
Exponentiate the contrast and interval, and state what rate is compared with what, per which exposure unit, and for which group context.
Worked Examples
Worked Example: A Treatment Rate Ratio Without an Interaction
Suppose a hypothetical Poisson GLMM compares an intervention with usual care for event counts across clinics. The offset is log patient-years, the model has a clinic random intercept, and the fitted treatment coefficient is \(0.4055\) with standard error \(0.1800\). Assume there is no treatment interaction in the model. Report the treatment comparison.
The treatment coefficient is the intervention-versus-usual-care log-rate contrast, conditional on the same clinic random intercept and the same values of the other modeled predictors. The rate ratio is \(\exp(0.4055)=1.500\). Thus, for the same covariate profile and clinic effect, the fitted event rate under intervention is 1.50 times the rate under usual care.
The 95% log-scale interval is \(0.4055\pm1.96(0.1800)=0.4055\pm0.3528=[0.0527,0.7583]\). Exponentiating the endpoints gives \([\exp(0.0527),\exp(0.7583)]=[1.054,2.135]\). The Wald statistic is \(z=0.4055/0.1800=2.253\), giving a two-sided \(p\)-value of approximately \(0.0243\).
A suitable report is: “For the same covariate profile and clinic-level random intercept, the intervention group’s event rate per patient-year was estimated to be 1.50 times the usual-care rate (95% Wald CI 1.054 to 2.135; \(p=0.024\)).” This is a comparison of rates, not a claim that the intervention caused the difference. The interval describes uncertainty in the model-based rate ratio; it does not give a probability that the fixed parameter lies in this particular interval.
Worked Example: A Treatment Contrast That Depends on Shift
In a hypothetical hospital study, a count GLMM uses patient-days as exposure and includes treatment, night shift, and their interaction, as well as a hospital random intercept. Let treatment equal one for the intervention and zero for control; let night shift equal one for night and zero for day. Suppose the estimates are \(\hat{\beta}_T=0.2231\), \(\hat{\beta}_{TN}=0.4700\), with standard errors \(0.2000\) and \(0.3000\), respectively, and covariance \(-0.0100\). Find the intervention-versus-control rate ratio on night shift.
The treatment contrast on day shift is \(\hat{\beta}_T=0.2231\). On night shift, the interaction also changes between treatment groups, so the log-rate contrast is \(\hat{\Delta}=\hat{\beta}_T+\hat{\beta}_{TN}=0.2231+0.4700=0.6931\). The night-shift rate ratio is \(\exp(0.6931)=2.000\), comparing intervention with control at the same modeled covariate values and hospital random intercept.
The variance of the sum is \(0.2000^2+0.3000^2+2(-0.0100)=0.0400+0.0900-0.0200=0.1100\). Thus \(SE(\hat{\Delta})=\sqrt{0.1100}=0.3317\). The 95% log-scale interval is \(0.6931\pm1.96(0.3317)=[0.0430,1.3432]\). Exponentiating gives the rate-ratio interval \([\exp(0.0430),\exp(1.3432)]=[1.044,3.831]\).
The Wald statistic is \(0.6931/0.3317=2.089\), with a two-sided \(p\)-value of approximately \(0.0367\). A clear report is: “On night shift, the estimated intervention-to-control event-rate ratio was 2.00 per patient-day, conditional on the same hospital random intercept and modeled covariates (95% Wald CI 1.044 to 3.831; \(p=0.037\)).” The treatment coefficient alone would describe the day-shift contrast, not this night-shift contrast.
Worked Example: Naming the Exposure Unit and Translating to Counts
Suppose a hypothetical clinic GLMM models event counts with offset \(\log(E/1000)\), where \(E\) is patient-days. At a stated covariate profile and random intercept of zero, the fitted reference-group rate is 2.40 events per 1,000 patient-days. The fitted intervention-versus-reference rate ratio is 0.800. What does this imply for 2,400 patient-days?
The offset makes the expected count equal to exposure measured in thousands of patient-days multiplied by the modeled rate per 1,000 patient-days. For the reference group, the expected count is \((2400/1000)(2.40)=2.4(2.40)=5.760\) events. For the intervention group at the same profile, its rate is \(0.800(2.40)=1.920\) events per 1,000 patient-days. Its expected count is therefore \((2400/1000)(1.920)=2.4(1.920)=4.608\) events.
A report could say: “At the stated covariate profile and a zero clinic random intercept, the estimated intervention rate was 0.800 times the reference-group rate per 1,000 patient-days (1.920 versus 2.40 events per 1,000 patient-days). For 2,400 patient-days, these rates correspond to model-based expected counts of 4.608 and 5.760 events, respectively.” These are expected counts, not guaranteed observed counts. The comparison is conditional on the stated profile; it should not be called a population-average rate unless that is the estimand actually calculated.
Checks, Alternatives, and Limits
Before reporting, verify that the offset is the log of the intended exposure in the intended units and that it has coefficient one. Check that exposure is positive, observations have meaningful exposure values, and the assumed proportionality between exposure and expected count is scientifically reasonable. Examine residual or simulation-based diagnostics for systematic lack of fit, influential observations, excess zeros, or conditional overdispersion. If the conditional Poisson variability is inadequate, uncertainty and fitted rates may be unreliable; an alternative count distribution or other model specification may be needed. These checks do not replace a clearly defined contrast.
A rate ratio is not a rate difference. A ratio of 2 can represent a change from 1 to 2 events per unit exposure or from 100 to 200; the absolute increases differ. Conversely, a rate difference depends on the covariate profile and units used to express rates. An unadjusted ratio of two observed counts divided by their exposures is also not generally the adjusted GLMM contrast, because it does not account for modeled predictors or grouping.
| Quantity | What it compares | What to state |
|---|---|---|
| Rate ratio | One rate divided by another | Contrast, exposure unit, and group/covariate context |
| Rate difference | One rate minus another | Both rate units and the profile at which rates are compared |
| Expected count | Model-based count for a specified exposure | Exposure amount and the profile or group context |
| Raw rate ratio | Observed counts divided by observed exposures | That it is unadjusted, plus the exposure units |
A conditional GLMM rate ratio is often appropriate when the scientific question is a comparison within the same group context. A marginal comparison instead averages over a random-effects distribution and may require explicit marginal predictions. Do not label a conditional coefficient as marginal merely because the random effects were included in the fitted model. For models with random slopes, report whether the contrast is evaluated for a specified group effect, summarized over groups, or averaged over a distribution.
Common Mistakes
- Exponentiating the wrong coefficient. In a model with interactions, a coefficient alone may represent the contrast only at the reference value of the interacting predictor. Form the full contrast first.
- Omitting the exposure unit. “The rate was higher” does not tell readers whether the rate is per patient-day, patient-year, or another unit. Name it in the result.
- Calling an expected count a rate. A fitted count depends on the specified exposure amount. Divide by exposure, or state the exposure amount when reporting expected counts.
- Ignoring covariance in a summed contrast. For a coefficient sum, calculate the variance with the covariance term; otherwise the standard error and interval can be incorrect.
- Overstating the conclusion. A confidence interval excluding one and a small \(p\)-value do not establish causation or practical importance. Describe the estimated association and its uncertainty in the study context.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A no-interaction model estimates a treatment coefficient of \(-0.2231\). What is the treatment rate ratio, and how should its direction be described?
- A model has treatment coefficient \(0.10\) and treatment-by-night interaction coefficient \(0.30\). What is the treatment log-rate contrast on night shift?
- Why is it incorrect to describe an offset-model coefficient as a rate ratio before exponentiating it?
- A reported rate ratio is 1.4 per 100 patient-days. What information is missing if the report does not state the predictor contrast?
- For a coefficient sum, the two coefficient variances are 0.04 and 0.09, and their covariance is \(-0.01\). Find the standard error of their sum.
Solutions. (1) \(\exp(-0.2231)=0.800\). The treatment rate is estimated to be 0.800 times the comparison rate, or 20% lower, conditional on the stated model context. (2) The contrast is \(0.10+0.30=0.40\) on the log-rate scale; its rate ratio is \(\exp(0.40)=1.492\). (3) The coefficient is a log-rate contrast; exponentiation converts it to a ratio of rates. (4) The reader needs to know which predictor levels or groups were compared, such as intervention versus control. (5) The variance is \(0.04+0.09+2(-0.01)=0.11\), so the standard error is \(\sqrt{0.11}=0.3317\).