Thall-Vail (1990) Some Covariance Models for Longitudinal Count Data with Overdispersion

Summary

Thall and Vail propose a family of covariance models for longitudinal count data that simultaneously handle overdispersion, heteroscedasticity, and within-subject dependence among repeated counts. The method is a quasi-likelihood / generalized-estimating-equation (GEE) regression in the spirit of Liang-Zeger (1986), but with a distinctive feature: two sets of estimating equations — one for the mean (covariate) parameters and a second for the variance-covariance parameters — so the correlation/overdispersion structure is itself estimated rather than treated as a nuisance. It is illustrated on the now-canonical epileptic-seizure count data from a progabide adjuvant-therapy trial.

Key Claims

• The data structure. Count response $Y_{it}$ with covariates $x_{it}$ at times $t=1,\dots,n_i$ for independent subjects $i=1,\dots,M$; when $n_i=1$ ordinary GLMs apply, but repeated biological counts routinely show variability exceeding the Poisson/exponential-family mean-variance relation.
• Overdispersion via random effects → structured covariance. Extra-Poisson variability is modeled as arising from unobserved random effects acting on the responses, which induces a parametric covariance structure on the repeated counts (overdispersion + heteroscedasticity + serial dependence), rather than assuming the working independence of a naive GLM.
• Quasi-likelihood, two GEEs. Following Wedderburn (1974) quasi-likelihood and Liang-Zeger (1986) GEE, the mean parameters solve one estimating equation; the variance-covariance parameters solve a second set of generalized estimating equations — so both the regression coefficients and the dependence/overdispersion parameters are estimated, with large-sample (asymptotic normality) theory derived for the joint estimator.
• A family of covariance models. Different specifications of the random-effect / covariance structure give a family of models, letting the analyst match the observed heteroscedasticity and correlation pattern rather than impose a single form.
• Seizure application. The methods are demonstrated on epileptic-seizure counts from a progabide clinical trial — a dataset that (partly through this paper) became the standard benchmark for longitudinal-count and GLMM methods.

Concepts Introduced or Extended

• Generalized Estimating Equations
• Count-Data Regression

Entities Mentioned

• Peter F. Thall
• Stephen C. Vail

Quotes

"A family of covariance models for longitudinal counts with predictive covariates is presented. These models account for overdispersion, heteroscedasticity, and dependence among repeated observations."

"Generalized estimating equations for both the covariate parameters and the variance-covariance parameters are presented."

My Take

The paper's contribution is best seen as making the dependence structure itself a modeled quantity for count data: Liang-Zeger GEE treats the working correlation as a nuisance to be plugged in, whereas Thall-Vail write a second estimating equation for the variance-covariance parameters, so overdispersion and serial correlation are estimated on equal footing with the mean. That is exactly the information an epidemiologist analyzing repeated seizure counts wants, and it is why the progabide seizure data it introduced became the reference example that GLMM papers (Breslow-Clayton, and the random-effects Poisson literature) return to again and again. For the wiki it extends the GEE thread from the Liang-Zeger binary/marginal setting to overdispersed counts with an explicitly estimated covariance, and sits alongside count-data regression as the longitudinal, correlation-aware member of that family — the frequentist-marginal complement to random-effects Poisson GLMMs.