Conditional Rate Ratios in Poisson Random-Intercept Models

Why the Conditional Rate Ratio Matters

Count outcomes are often observed repeatedly within groups: events per patient, incidents per facility, or visits per person. Observations from one group may share unmeasured characteristics. A Poisson random-intercept model represents that shared group-level variation while describing how a covariate is associated with the expected count. Its log link makes the covariate effect multiplicative on the mean scale: a one-unit change corresponds to a rate ratio.

The key qualification is conditional. The rate ratio compares expected counts for two covariate values within the same cluster, so the cluster’s random intercept is held fixed. It is not, by itself, a statement that changing the covariate causes a change in the outcome.

Model and Assumptions

Definition: For cluster $j$ and observation $i$ within that cluster, a Poisson random-intercept model with a log link is $$ Y_{ij}\mid b_j \sim \operatorname{Poisson}(\mu_{ij}), \qquad \log(\mu_{ij})=\beta_0+\beta_1x_{ij}+b_j. $$ Here $Y_{ij}$ is a count, $\mu_{ij}=E(Y_{ij}\mid b_j,x_{ij})$, $x_{ij}$ is a numerical covariate, $\beta_0$ is the fixed intercept, $\beta_1$ is its fixed slope, and $b_j$ is the random intercept for cluster $j$. The usual random-intercept specification assumes $b_j\sim N(0,\sigma_b^2)$, with random intercepts independent across clusters and independent of the covariates. Conditional on the random effects and covariates, the counts are independent and have Poisson variance $\operatorname{Var}(Y_{ij}\mid b_j,x_{ij})=\mu_{ij}$.

These assumptions define the model used for the interpretation. In particular, the Poisson mean-variance equality is conditional on the random intercept; the marginal variance after accounting for variation in $b_j$ need not equal the marginal mean. The random intercept shifts all observations in a cluster by the same amount on the log-mean scale. For this tutorial, the covariate has the same fixed slope $\beta_1$ in every cluster.

If observations have different exposure durations, a log-exposure offset can be included: $\log(\mu_{ij})=\log(T_{ij})+\beta_0+\beta_1x_{ij}+b_j$, where $T_{ij}$ is a positive exposure amount. In that model the expected count divided by exposure, $\mu_{ij}/T_{ij}$, is the conditional rate. The same rate-ratio derivation applies when comparing covariate values at the same exposure, or when comparing rates with the offset accounted for.

Deriving the One-Unit Rate Ratio

Exponentiating the linear predictor gives the conditional expected count:

\mu_{ij}(x)=\exp(\beta_0+\beta_1x+b_j).

Compare two observations in the same cluster, with all other modeled covariates and exposure held fixed, but with covariate values $x+1$ and $x$. Their conditional mean ratio is

\frac{\mu_{ij}(x+1)}{\mu_{ij}(x)} = \frac{\exp\{\beta_0+\beta_1(x+1)+b_j\}} {\exp(\beta_0+\beta_1x+b_j)} = \exp(\beta_1).

The fixed intercept and the shared random intercept cancel. The result does not depend on the particular cluster’s $b_j$, or on the starting value $x$. It describes a constant multiplicative change in the conditional mean for each one-unit increase in $x$. Since the conditional expected count is the quantity being compared, this is also called a conditional rate ratio when the model includes an exposure offset and rates are being compared.

Key result: In a log-link Poisson random-intercept model with a common fixed slope, the conditional rate ratio for a one-unit increase in $x$, comparing observations in the same cluster and holding other terms fixed, is $\exp(\beta_1)$. A change of $d$ units instead has ratio $\exp(d\beta_1)$.

The coefficient $\beta_1$ itself is on the log-mean scale. It is the change in the log conditional mean for a one-unit increase in $x$; it is not a change of $\beta_1$ counts. The exponentiated coefficient is dimensionless. Its context still depends on how $x$ is measured: if $x$ is hours, the ratio is per additional hour; if $x$ is measured in tens of hours, it is per additional ten hours.

Worked Example: Translating a Slope Into a Rate Ratio

Suppose a hypothetical model relates a count of service requests to a numerical workload measure $x$, measured in tens of units. The fitted slope is $\hat{\beta}_1=0.30$. Assume the log-link random-intercept model is appropriate, the same slope applies across clusters, and the comparison holds exposure and other modeled covariates fixed. What is the estimated conditional rate ratio for a one-unit increase in $x$?

Exponentiate the fitted slope:

\widehat{RR}_{\text{conditional}} =\exp(\hat{\beta}_1) =\exp(0.30) =1.3499\approx 1.350.

Thus, within a cluster, a one-unit increase in workload (ten units on its original scale) is associated with an estimated conditional expected count 1.350 times as large, or about 35.0% higher, holding the other terms fixed. This is an association under the fitted model, not a causal effect without additional design and causal assumptions.

What Cancels, and What Does Not

The cancellation is easiest to see by comparing two covariate values for the same cluster. For different clusters, the random intercepts generally differ, so a comparison at the same covariate value includes their baseline difference. That does not alter the within-cluster rate ratio: moving $x$ by one unit multiplies both clusters’ conditional means by the same $\exp(\beta_1)$ under this model.

The random-intercept distribution matters for fitting the model and describing variation among clusters. However, the algebraic cancellation in this conditional contrast only requires a shared intercept in the two linear predictors. It does not require substituting the average random intercept $b_j=0$. Setting $b_j=0$ describes a cluster at the center of the assumed random-intercept distribution; it is not the reason the conditional rate ratio is $\exp(\beta_1)$.

Worked Example: Comparing Two Covariate Values in One Cluster

For a hypothetical outcome, suppose $\log(\mu_{ij})=1.10+0.40x_{ij}+b_j$. Consider one cluster with $b_j=-0.50$, and compare $x=2$ with $x=3$. The model’s assumptions are that these conditional means follow the stated log-linear form and that the same cluster intercept applies to both values.

At $x=2$, the linear predictor is $1.10+0.40(2)-0.50=1.40$, so the conditional expected count is $\exp(1.40)=4.055$. At $x=3$, it is $1.10+0.40(3)-0.50=1.80$, giving $\exp(1.80)=6.050$. The ratio, using the rounded means, is approximately $6.050/4.055=1.492$.

Directly from the slope, the ratio is $\exp(0.40)=1.492$. Thus the one-unit increase corresponds to an estimated conditional mean about 1.492 times as large, or 49.2% higher, for this cluster. The intercept and the cluster deviation determine the level of the means, but not their ratio for this one-unit change.

Worked Example: Interpreting a Negative Slope

Suppose, hypothetically, that a log-link Poisson random-intercept model for monthly event counts has a fitted slope $\hat{\beta}_1=-0.22$ for $x$, measured in years. Assume the conditional Poisson model and common-slope specification are reasonable, and compare observations within a cluster at the same exposure.

The one-year conditional rate ratio is

\exp(-0.22)=0.8025.

The estimated conditional expected count after a one-year increase in $x$ is therefore 0.8025 times the expected count before the increase, or about 19.8% lower, since $(1-0.8025)\times100=19.75\%$. The percentage decrease is not 22%; that number is the magnitude of the log-scale coefficient. As always, the statement describes a conditional association, not a causal effect.

Assumptions, Diagnostics, and Alternatives

A rate ratio can be calculated by exponentiating a fitted slope, but its model-based interpretation depends on the model being adequate. Check that the response is a count, that the log conditional mean is reasonably represented by the included predictors, and that observations are conditionally independent given the modeled random effects and covariates. If there is exposure time or another exposure amount, it should be represented appropriately rather than treated as if every observation had equal exposure.

Useful checks include examining residual patterns against fitted values and covariates, looking for substantial extra-Poisson variation or clusters of zeros not captured by the model, and checking whether estimated cluster patterns are inconsistent with a single shared intercept. Such patterns can indicate mean-model misspecification, dependence not represented by the random intercept, or conditional variability beyond the Poisson assumption. The assumed normal distribution for random intercepts is also relevant to model fitting and should not be treated as automatically true. A poor fit can make the reported coefficient and its uncertainty misleading even though the cancellation is algebraically correct.

The random-intercept model is not the only way to handle clustered counts. A Poisson GLM without a random intercept is simpler, but does not represent cluster-specific baseline heterogeneity in the mean structure. Cluster-specific fixed intercepts can represent separate baselines for observed clusters, but are a different model and do not use a shared random-intercept distribution. A population-averaged analysis targets a different kind of summary. These alternatives should not be treated as interchangeable merely because each can produce a coefficient or rate ratio.

In particular, $\exp(\beta_1)$ here is a conditional rate ratio: the comparison holds the cluster effect fixed. It should not be relabelled a population-averaged effect without deriving that target under the relevant model and assumptions. Likewise, the random-intercept formulation assumes the cluster effect is independent of the covariates. If that assumption is implausible, the fitted conditional interpretation may not answer the intended comparison.

Common Pitfalls

Reading the coefficient as a percentage. A slope of 0.30 does not mean the expected count rises by 30%. The corresponding rate ratio is $\exp(0.30)=1.350$, about a 35.0% increase.
Forgetting what is held fixed. The derivation compares covariate values while keeping the same cluster intercept and other model terms fixed. A comparison between different clusters can also reflect different intercepts.
Confusing conditional and population-averaged interpretations. The result is conditional on the random intercept. A population-level summary requires its own definition and calculation.
Calling an association causal. A fitted rate ratio alone does not establish that changing the covariate would change the count. Causal claims require appropriate design and assumptions.
Ignoring covariate units or exposure. State the size of a one-unit change, and ensure that exposure differences are handled if the target is a rate.

Key takeaway: For a numerical covariate with a common fixed slope in a Poisson log-link random-intercept model, the one-unit conditional rate ratio is $\exp(\beta_1)$. The random intercept changes a cluster’s conditional baseline but cancels from a within-cluster contrast.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A fitted slope is $0.12$ for a covariate measured in days. What is the one-day conditional rate ratio, and how would you interpret it?
Why does the random intercept cancel when comparing $x+1$ with $x$ in the same cluster?
A fitted slope is $-0.40$. Calculate the one-unit rate ratio and the approximate percentage change in the conditional expected count.
Does the one-unit conditional rate ratio necessarily describe a causal effect? Briefly explain.
What interpretation changes if two observations being compared have different cluster random intercepts?

Solutions. (1) $\exp(0.12)=1.128$. Per additional day, the estimated conditional expected count is about 1.128 times as large, or 12.8% higher, within a cluster and holding other terms fixed. (2) Both linear predictors contain the same $b_j$; when their exponentiated means are divided, the common factor $\exp(b_j)$ cancels. (3) $\exp(-0.40)=0.6703$, so the conditional expected count is about 0.6703 times as large, a decrease of approximately $(1-0.6703)\times100=33.0\%$. (4) No. It is a conditional association under the fitted model; causal interpretation requires additional design and assumptions. (5) Their comparison includes the difference between the cluster intercepts as well as the covariate contrast, so it is not the same as holding a shared cluster effect fixed.