Poisson Longitudinal Submodels for Counts

Why Use a Poisson Longitudinal Submodel?

Repeated outcomes are not always continuous measurements. A study may record the number of symptom episodes, hospital visits, or device failures observed for each person at several follow-up times. These responses are nonnegative integer counts, and their variability often changes with their mean. A Gaussian linear mixed model can consequently give implausible predictions, such as negative counts, or describe the mean-variance relationship poorly.

A Poisson mixed model links the expected count to covariates through a log link and represents persistent subject-to-subject differences with random effects. In a joint model, this is the longitudinal component for a count marker; the same subject-level latent effects may also enter the time-to-event component. The focus here is specifying and interpreting the count component, not the event component.

From the Model to Rates and Expected Counts

The offset is essential when observation intervals differ. Define the conditional rate per unit exposure by \(\lambda_{ij}=\mu_{ij}/E_{ij}\). Then \(\log(\lambda_{ij})=x_{ij}^{\mathsf T}\beta+z_{ij}^{\mathsf T}b_i\), while the expected count over the actual exposure is \(\mu_{ij}=E_{ij}\lambda_{ij}\). Thus a model coefficient describes a rate contrast, not automatically a contrast in total counts when exposures differ.

The Poisson distribution assumes, conditional on the covariates and random effects, that the count has variance equal to its mean: \(\operatorname{Var}(Y_{ij}\mid b_i,x_{ij},z_{ij})=\mu_{ij}\). It also assigns positive probability to zero counts. Repeated counts are commonly assumed conditionally independent across visits given the subject's random effects and covariates. The random effects induce dependence marginally: observations from one subject share the same \(b_i\).

The normal distribution for \(b_i\) is a modeling assumption about the distribution of subject-specific deviations in the population. It is not a claim that counts themselves are normally distributed. In an application, the assumed observation schedule and covariates should also be appropriate: for example, exposure duration must represent the time during which the count could actually have occurred.

The fixed-effect interpretation above requires care about the random-effects design. If treatment also has a random slope \(b_{\mathrm{trt},i}\), then the subject-specific conditional log-rate contrast is \(\beta_{\mathrm{trt}}+b_{\mathrm{trt},i}\), and the rate ratio is \(\exp(\beta_{\mathrm{trt}}+b_{\mathrm{trt},i})\). The population fixed coefficient alone does not give that subject's conditional rate ratio.

A Rate-Ratio Justification

The cancellation that gives a simple rate ratio follows directly from subtracting the two linear predictors. Let observations \(a\) and \(b\) differ only in treatment, and suppose treatment is not included in the random-effects design, so \(z_a=z_b\). If they also have equal exposure, then

Since the common exposure cancels, \(\mu_a/\mu_b=\lambda_a/\lambda_b\). Exponentiating the log contrast gives the conditional rate ratio \(\exp(\beta_{\mathrm{trt}})\). If treatment has a random slope, its contribution does not cancel: the contrast becomes \(\beta_{\mathrm{trt}}+b_{\mathrm{trt},i}\), giving the subject-specific ratio stated above. Comparisons that differ in other covariates must include their corresponding coefficient contrasts as well.

Conditional and Population-Average Predictions

A conditional prediction uses a specified value of \(b_i\), often an estimated subject-specific value for a person with observed data. A population-average mean instead averages the conditional mean over the random-effects distribution. These are different quantities because exponentiation is nonlinear. For example, with a random intercept \(b_{0i}\sim N(0,\sigma_b^2)\) and no random slope,

The last equality uses the moment-generating function of a normal random variable. It shows why simply setting the random effect to zero does not generally give the population-average mean. In a joint model, the conditional formulation is especially useful because it can preserve subject-level latent effects shared with the event process.

Fitting Checks, Alternatives, and Limitations

Before interpreting coefficients, examine whether the model's assumptions are credible. Useful checks include comparing observed and fitted count distributions, plotting residual summaries against time and fitted means, checking whether large counts are systematically underpredicted, and examining within-subject patterns not captured by the random effects. Convergence problems, very large standard errors, or random-effect estimates near a boundary can indicate sparse information or an overly complex random-effects structure.

A common failure is conditional overdispersion: after accounting for the modeled covariates and random effects, the observed variability remains greater than the Poisson mean-variance restriction allows. Excess zeros or clusters of unusually large counts may also signal a poor fit, but neither feature by itself proves which model is appropriate. Repeated-measure dependence can likewise remain if the random-effects structure does not capture the temporal pattern. These findings call for investigation of the data collection process and model specification; they should not be dismissed merely because the fitting algorithm converged.

The main alternatives answer different needs. Ordinary Poisson regression has no subject-specific random effects and does not represent residual within-person dependence in repeated data. A Poisson mixed model adds such dependence through latent effects, but relies on its conditional Poisson variance assumption. If conditional variability is persistently too large, an overdispersed count model may be more suitable. If counts arise from distinct structural processes, a different count formulation may be needed. These alternatives should be selected based on the data-generating context and diagnostics rather than by treating every zero or large count as evidence for a particular model.