Random Effects for Heterogeneous Count Rates

Why Count Rates Vary Between Subjects

Two subjects with the same measured covariates and exposure can still have different expected event counts. Some may have a persistently higher underlying rate; others may respond differently as a covariate changes. A count mixed-effects model represents these patterns through subject-specific random effects. Because the model uses a log link, those effects act multiplicatively on rates, and averaging over them requires care.

Let \(Y_{ij}\) be the count for subject \(i\) on observation \(j\), and let \(E_{ij}>0\) be the known exposure, measured in units such as patient-weeks. Let \(z_{ij}\) be a covariate, preferably centered so that zero has a useful interpretation. A Poisson model with a random intercept and random slope is

Here \(b_i=(b_{0i},b_{1i})^\mathsf{T}\) is the pair of random effects for subject \(i\). The exposure enters as an offset \(\log(E_{ij})\), so the rate per exposure unit is \(\mu_{ij}/E_{ij}\). A random-intercept-only model sets \(b_{1i}=0\). The usual assumptions for this model are that, conditional on the subject's random effects and covariates, the counts are Poisson with the stated means; observations are conditionally independent given those effects; and the random effects follow a specified distribution, commonly a bivariate normal distribution with mean zero.

The covariance matrix \(D\) must be positive semidefinite. The zero-mean convention makes \(\beta_0\) and \(\beta_1\) the fixed effects for a subject whose random effects are both zero; it does not imply that all subjects have those effects. The model also assumes the chosen normal distribution adequately describes the distribution of subject effects for the intended analysis.

Interpreting Random Intercepts and Slopes

At \(z=0\), the subject-specific rate is \(\exp(\beta_0+b_{0i})\) per exposure unit. Thus a one-unit increase in \(b_{0i}\) multiplies this rate by \(e\), while a difference of \(d\) in random intercepts corresponds to a rate ratio of \(\exp(d)\) at every covariate value when slopes are the same. Centering \(z\) determines where the intercept comparison applies.

For subject \(i\), a one-unit increase in \(z\) changes the log rate by \(\beta_1+b_{1i}\). The subject-specific rate ratio for that increase is \(\exp(\beta_1+b_{1i})\). Consequently, a positive random slope means a larger covariate-associated increase (or smaller decrease) on the log-rate scale than the fixed slope alone indicates. The interpretation depends on the covariate's units: if \(z\) is dose in milligrams, the slope is per milligram unless dose was rescaled.

Subject-specific expected counts follow directly from the rate and exposure. At a specified \(z\) and exposure \(E\), the conditional mean is \(E r_i(z)\). This is a prediction conditional on the subject's random effects. For a new subject, those effects are not known; a population prediction instead averages over their distribution.

Random Slopes and Individual Predictions

With a random slope, the rate ratio comparing two covariate values depends on their difference. For values \(z_a\) and \(z_b\), the subject-specific rate ratio is \(\exp\{(\beta_1+b_{1i})(z_b-z_a)\}\). The random intercept cancels from this contrast, while the random slope remains. This is one reason random slopes matter when subjects have meaningfully different covariate responses.

Integrating Over Random Effects

A population-average mean is not generally obtained by setting the random effects to zero. It is obtained by averaging the conditional mean over their distribution. Write \(a(z)=(1,z)^\mathsf{T}\). The random contribution to the log rate is \(a(z)^\mathsf{T}b_i\), which is normally distributed with mean zero and variance \(v(z)=a(z)^\mathsf{T}D a(z)\). For a mean-zero normal variable \(U\) with variance \(v\), its moment-generating function gives \(E[\exp(U)]=\exp(v/2)\). Applying that identity yields the population-average rate and expected count.

The result follows by integrating the conditional rate with respect to the random-effects distribution:

The final equality uses the normal exponential moment identity, with variance \(a(z)^\mathsf{T}Da(z)\). This derivation also explains why the population-average rate exceeds the rate for a subject with \(b_i=0\) at the same covariate value whenever \(v(z)>0\). The exponential function is convex, so averaging exponentiated random effects does not equal exponentiating their average.

Population-Average Comparisons and Model Checks

If the model has only a random intercept, then \(v(z)=\sigma_0^2\) is constant. Integration multiplies every conditional rate by the same factor \(\exp(\sigma_0^2/2)\); therefore, the rate ratio across two covariate values remains \(\exp\{\beta_1(z_b-z_a)\}\). With random slopes, \(v(z)\) changes with \(z\). The population-average rate ratio is then

Thus, the fixed slope alone describes the conditional log-rate change for a subject with zero random effects, not necessarily the log rate ratio between population-average rates. Inference and communication should state which scale and which population are being described.

Useful checks include comparing observed and simulated subject-level count patterns, examining whether residual patterns remain across covariate values or exposure, and checking whether estimated random-effect variances are near zero or the covariance estimate is unstable. Subject-level predictions should also be assessed for implausible rates. A near-zero slope variance may indicate little support for a random slope, but can also arise from limited observations per subject or weak covariate variation. A variance estimate at a boundary or a highly uncertain covariance is a warning about identifiability, not automatic evidence that the model is correct or incorrect.

Random-effects predictions for observed subjects are commonly estimated using their data and the fitted population distribution. Such estimates are conditional predictions and are typically shrunk toward zero when subject-specific data are sparse. They are not known true effects, and treating them as error-free can understate uncertainty. For a new subject without counts, use the population-average prediction when the target is the mean across subjects; a zero-random-effect prediction is instead a typical-subject conditional prediction.

A random-intercept-only model is simpler and captures persistent rate heterogeneity but assumes common covariate response. A random-intercept-and-slope model allows response heterogeneity, at the cost of additional parameters and stronger data requirements. A model without random effects is simpler still, but it does not represent within-subject dependence or between-subject rate variation in this way. These are different modeling assumptions, not interchangeable computational options. Random effects can also induce extra marginal variation: for one observation with conditional mean \(\lambda_i\), the law of total variance gives \(\operatorname{Var}(Y)=E(\lambda_i)+\operatorname{Var}(\lambda_i)\). This tutorial does not identify that variation with any particular alternative count distribution; the assumed random-effects structure remains a modeling choice that should be checked against the data.

Common Pitfalls and Limitations

• Adding a random effect directly to the rate. The model adds it to the log rate; the rate is multiplied by its exponential. For example, an intercept effect of \(0.2\) multiplies the rate by \(\exp(0.2)\), not by \(1.2\).
• Calling the zero-effect rate the population mean. It is the conditional rate at \(b_i=0\). The population mean includes the variance adjustment from integrating over the random-effects distribution.
• Ignoring covariance in a random-slope model. The term \(2z\sigma_{01}\) affects the variance of the combined random contribution and hence the marginal mean.
• Interpreting a random slope without its covariate units. A slope per one milligram differs numerically from a slope per one hundred milligrams; specify the scale and centering of \(z\).

The lognormal integration formula depends on normally distributed random effects and the stated log-linear predictor. If those assumptions are poor, the formula may not describe the intended population mean. Random slopes also require enough subjects and within-subject covariate variation to estimate their variability; a complex structure can be weakly identified. Exposure must be correctly measured and enter as a proportional offset for the stated rate interpretation to hold.