Tsutakawa 1988 — Mixed Model for Analyzing Geographic Variability in Mortality Rates

Summary

Tsutakawa (1988) proposes a gamma-Poisson mixed model for estimating disease-specific mortality rates in small geographic units such as counties, where raw counts are too sparse to yield reliable direct estimates. The model has two layers of random effects: a geographic relative-risk parameter $z_i$ (shared across demographic groups within region i) and a cell-specific rate $p_{ij}$ (adding extra-Poisson variation within each region/demographic group). Demographic parameters $\theta_j$ are fixed effects. Hyperparameters are estimated by maximum likelihood (profile likelihood for the inverse-gamma shape parameter $\gamma$), and relative risks and mortality rates are estimated by their empirical Bayes posterior means. The method is illustrated on lung-cancer mortality in 115 Missouri counties, 1972–1981, for males across four age groups (45–54, 55–64, 65–74, 75+).

Key Claims

• Raw county rates are unreliable for small populations: Observed cause-specific mortality rates for small geographic units are dominated by chance variation. Aggregating small units to larger ones conceals geographic heterogeneity that is often the starting point for discovering environmental causes of disease.
• The posterior mean is a shrinkage estimator: The empirical Bayes estimate of $p_{ij}$ is a weighted average of the observed rate $y_{ij}/n_{ij}$ and the model-expected rate $\bar{p}_{ij} = \theta_j \cdot z_i$, with weights $n_{ij} \cdot \bar{p}_{ij}$ (from the data) and $\sigma_j^{-2}$ (from the prior). As $\sigma_j \to 0$, estimates shrink fully to the demographic model; as $n_{ij} \to \infty$, the observed rate dominates. For the smallest Missouri counties, empirical Bayes standard deviations can be as much as 15% smaller than under the multiplicative Poisson model.
• Gamma is the natural conjugate prior for Poisson: Unlike logistic-normal models that require multivariate integration, the gamma prior for $p_{ij}$ yields an analytically tractable posterior (also gamma). Only single integrals are required for the profile likelihood of $\gamma$, solved by Gauss-Hermite quadrature after a log transformation.
• The model is more flexible than multiplicative Poisson: Under multiplicative Poisson, geographic effects $z_i$ are treated as fixed parameters common across all age/sex groups — there is no cell-level extra-Poisson variation. The mixed model adds random $p_{ij}$, allowing geographic effects to vary across demographic subgroups. A likelihood ratio test of 51.4 (comparing full model to $\gamma = \infty$ limit with no $z_i$) confirms the presence of common geographic effects across age groups.
• Estimated relative risks $z_i$ are near 1 for most counties: In the Missouri lung-cancer application, relative risks for males are generally in the 0.9–1.1 range. Counties 3, 9, 52, 73, and 84 depart noticeably from 1. Residuals from the mean-variance structure have mean 0.060 and standard deviation (SD) 0.862 — a reasonably close fit.
• Key estimated parameters for Missouri males: $\hat{\sigma}^2 = 6.83 \times 10^{-3}$; $\hat{\theta}_1 = 7.42 \times 10^{-3}$ (ages 45–54); $\hat{\theta}_2 = 2.07 \times 10^{-2}$ (55–64); $\hat{\theta}_3 = 3.79 \times 10^{-2}$ (65–74); $\hat{\theta}_4 = 3.61 \times 10^{-2}$ (75+); $\hat{\gamma} = 58.0$. The slight decline in the 75+ group relative to 65–74 may reflect birth-cohort effects or competing risks.
• Marginal distribution of $y_{ij}$ is negative binomial: After integrating out $p_{ij}$, the observed count $y_{ij} \mid \theta_j, \sigma_j^2, z_i$ is negative binomial with mean $n_{ij} \cdot \theta_j \cdot z_i$ and variance = Poisson component + non-Poisson component ($\sigma_j^2$-scaled quadratic term), consistent with the known extra-Poisson variability in disease mortality data.

Concepts Introduced or Extended

• Small-Area Mortality Estimation — introduces gamma-Poisson hierarchical model as a solution to the sparse-data problem for county-level disease rates; empirical Bayes posterior mean as the shrinkage estimator; profile likelihood for hyperparameter estimation

Entities Mentioned

• Robert K. Tsutakawa

Quotes

"The amount of information available from any one unit is generally limited. But modeling the variability between and within units can improve estimates, as demonstrated frequently in empirical Bayes examples."

"The considerable flexibility the random effects model has in showing how geographic effects change over different age/sex groups."

My Take

The paper is an elegant solution to a practical problem: how to estimate disease-specific mortality rates for small populations without either ignoring geographic heterogeneity (by pooling) or relying on noisy raw rates (by treating each county separately). The empirical Bayes shrinkage estimator achieves the right balance — shrinking toward the demographic model for small counties and trusting the data for large ones. The gamma-Poisson structure is particularly clean because of the conjugate relationship, though the inverse-gamma prior for $z_i$ (a somewhat unusual choice) reflects a specific parametric assumption about inter-county heterogeneity. The paper stops short of a full Bayesian analysis (the hyperparameters are treated as fixed at their maximum likelihood (ML) estimates), a limitation the author acknowledges. The approach predates modern Markov chain Monte Carlo (MCMC) methods, which would allow full posterior uncertainty propagation over hyperparameters — a direction Tsutakawa (1985) and the subsequent Bayesian spatial statistics literature pursued. The profile likelihood solution for $\gamma$ is ingenious given the computational constraints of 1988.