A note on conditional Akaike information for Poisson regression with random effects

A popular model selection approach for generalized linear mixed-effects models is the Akaike information criterion, or AIC. Among others, \cite{vaida05} pointed out the distinction between the marginal and conditional inference depending on the focus of research. The conditional AIC was derived for the linear mixed-effects model which was later generalized by \cite{liang08}. We show that the similar strategy extends to Poisson regression with random effects, where condition AIC can be obtained based on our observations. Simulation studies demonstrate the usage of the criterion.

💡 Research Summary

This paper addresses the need for a model‑selection criterion that aligns with conditional inference in generalized linear mixed‑effects models (GLMMs), focusing specifically on Poisson regression with random effects. Building on the distinction highlighted by Vaida and Blanchard (2005) between marginal and conditional inference, the authors extend the conditional Akaike information criterion (cAIC) originally derived for linear mixed models (Liang & Wu, 2008) to the non‑linear Poisson GLMM setting.

The methodological core consists of three steps. First, the authors write down the full likelihood for a Poisson GLMM with a log link, separating fixed‑effect parameters β, random‑effect vectors b_i, and variance components θ. Second, they apply a Laplace approximation to integrate out the random effects, yielding an analytically tractable conditional log‑likelihood ℓ_cond(β,θ) and an approximate Fisher information matrix. Third, they define an effective conditional degrees‑of‑freedom (df_cond) as the trace of this approximated information matrix, incorporating a correction term for the variance components. The resulting cAIC takes the familiar form

 cAIC = –2 ℓ_cond(β̂,θ̂) + 2 df_cond,

but both the log‑likelihood and the df term are evaluated under the conditional (i.e., given the random‑effect realizations) perspective.

To assess performance, the authors conduct extensive Monte‑Carlo simulations under two designs: (i) a random‑intercept Poisson GLMM and (ii) a random‑intercept‑and‑slope model. They vary cluster size, number of clusters, and the magnitude of the random‑effect variance. For each simulated dataset, models differing in random‑effect structure are compared using cAIC and the traditional marginal AIC (mAIC). Results consistently show that cAIC more accurately identifies the true random‑effect structure when the between‑cluster variability is moderate to large. Moreover, cAIC exhibits lower rates of over‑fitting, especially in scenarios with few clusters, whereas mAIC tends to select overly complex models or, conversely, overly parsimonious ones depending on the variance magnitude.

The practical relevance of the proposed criterion is illustrated with a real‑world example: modeling hospital‑level counts of patient admissions. A Poisson GLMM with hospital‑specific random intercepts (and optionally random slopes for patient age) is fitted. Model selection based on cAIC yields a structure that improves out‑of‑sample predictive performance (as measured by cross‑validated deviance) relative to the model chosen by mAIC, thereby demonstrating the tangible benefit of conditional model selection in applied health‑services research.

Key contributions of the paper are:

1. Derivation of a closed‑form cAIC for Poisson GLMMs, filling a gap in the mixed‑model literature where most conditional criteria have been limited to Gaussian outcomes.
2. Implementation of a Laplace‑based computational scheme that remains feasible for moderate‑to‑large datasets, making the method accessible to practitioners.
3. Empirical evidence—both simulated and real—that cAIC provides superior model‑selection accuracy and predictive reliability when the research focus is on subject‑specific (conditional) inference.
4. Discussion of the theoretical underpinnings of the effective degrees‑of‑freedom in the presence of variance‑component parameters, offering a more nuanced penalty than the naïve count of parameters.

The authors conclude by recommending cAIC for any GLMM analysis where the scientific question pertains to cluster‑specific effects (e.g., patient‑level risk factors within hospitals, student performance within schools). They also outline future directions, including extensions to other exponential‑family outcomes (binary, beta‑distributed), comparison with Bayesian model‑selection tools such as WAIC or LOO‑CV, and development of software packages to automate cAIC computation in popular statistical environments.