Small-Area Mortality Estimation

Definition

Small-area mortality estimation is the statistical problem of reliably estimating disease-specific death rates for small geographic units (counties, cities, census tracts) where observed counts are too sparse to yield stable direct estimates. Raw rates for small populations are dominated by sampling variability and are unreliable for all but the largest units. The solution is to borrow strength across units using a hierarchical model: each unit's estimate is shrunk toward a demographically modeled expectation, with the degree of shrinkage determined by unit population size and estimated inter-unit heterogeneity.

Key Ideas

The Sparse-Data Problem

Observed cause-specific death counts $y_{ij}$ for small geographic units (e.g., a rural county in the age $45$–$54$ group) often equal zero or small single-digit counts, even over a decade of observation. The standard maximum likelihood estimate $(y_{ij}/n_{ij})$ is unbiased but has enormous variance. The problem is compounded when researchers wish to understand geographic variation — if small-county estimates are dominated by noise, true geographic heterogeneity (environmental, occupational, socioeconomic) cannot be separated from chance. The solution is not to aggregate units (which destroys geographic information) but to model the variability within and between units simultaneously.

The Gamma-Poisson Hierarchical Model (Tsutakawa 1988)

The model has three components:

1. Observed counts: $y_{ij} \sim \text{Poisson}(n_{ij} \cdot p_{ij})$, where $n_{ij}$ is the population at risk in region $i$, demographic group $j$, and $p_{ij}$ is the true rate.
2. Cell-level rates: $p_{ij} \mid \theta_j, \sigma_j^2, z_i \sim \text{Gamma}$ with mean $\theta_j \cdot z_i$ and coefficient of variation $\sigma_j$. This accounts for extra-Poisson variation beyond what the geographic and demographic structure explains.
3. Geographic relative risks: $z_i \sim \text{Inverse-Gamma-1}(\gamma)$, with mean 1 (normalized for identifiability). $z_i$ is common to all demographic groups within region $i$, capturing the overall environmental/geographic effect.

The marginal distribution of $y_{ij}$ (integrating out $p_{ij}$) is negative binomial, consistent with the well-known extra-Poisson variability in cancer mortality data. This differs from the multiplicative Poisson model, which treats $z_i$ as fixed parameters and has no cell-level extra-Poisson term.

Empirical Bayes Estimation

Hyperparameters $\xi = (\theta, \sigma^2, \gamma)$ are estimated by maximum likelihood (profile likelihood for $\gamma$). Given $\hat{\xi}$, the empirical Bayes estimate of $p_{ij}$ is the posterior mean:

$$\mathbb{E}(p_{ij} \mid y_{ij}, \hat{\xi}) = \frac{(y_{ij}/n_{ij}) \cdot n_{ij}\bar{p}_{ij} + \bar{p}_{ij}/\sigma_j^2}{n_{ij}\bar{p}_{ij} + 1/\sigma_j^2}$$

where $\bar{p}_{ij} = \theta_j z_i$. This is a weighted average of the observed rate $y_{ij}/n_{ij}$ and the expected rate $\bar{p}_{ij}$, with:

• More weight on $\bar{p}_{ij}$ (stronger shrinkage) when the unit is small ($n_{ij}$ small) or when inter-unit variability is low ($\sigma_j$ small)
• More weight on $y_{ij}/n_{ij}$ (weaker shrinkage) when the unit is large

For the smallest Missouri counties, this shrinkage reduces standard deviations by up to $15\%$ relative to the multiplicative Poisson model.

The Gamma as Conjugate Prior for Poisson

The choice of a gamma prior for $p_{ij}$ is natural because the gamma is the conjugate prior for the Poisson likelihood: the posterior of $p_{ij}$ given $y_{ij}$ is also gamma, with updated shape and scale parameters. This avoids the multivariate numerical integration required by logistic-normal models, reducing the computational burden to single integrals (evaluated by Gauss-Hermite quadrature after a log transformation). This computational advantage was critical for 1988 hardware.

Multiplicative Poisson vs. Random Effects Model

Under the multiplicative Poisson model, geographic effects $z_i$ are fixed parameters, common across age/sex groups. The model has no cell-level extra-Poisson variation — all unexplained variability is absorbed into $z_i$. The random effects model adds a cell-level random component $p_{ij}$, allowing geographic effects to vary across demographic subgroups. A likelihood ratio test of $51.4$ in the Missouri lung-cancer example confirms the presence of common geographic effects $z_i$ across age groups, while residual analysis reveals remaining cell-level extra-Poisson variation.

How It Works

Estimation Algorithm

1. Profile out $\gamma$: for each candidate value of $\gamma$, maximize $L(\xi)$ over $\sigma^2$ and $\theta_j$
2. Select $\hat{\gamma}$ that maximizes the profile likelihood
3. Compute posterior means $E(p_{ij} \mid y, \hat{\xi})$ by numerical integration
4. Compute empirical Bayes relative risks $\tilde{z}_i$ from the posterior distribution of $z_i$

Computational Expressions

The log-likelihood is a sum across regions: $L(\xi) = \sum_i L_i(\xi)$, where each $L_i$ involves an integral over $z_i$ of a product of negative binomial densities (for all demographic groups in region $i$) times the inverse-gamma density for $z_i$. The Gauss-Hermite quadrature after the substitution $x = \log z$ provides accurate numerical evaluation.

Residual Diagnostics

Standardized residuals $r_{ij} = (y_{ij} - m_{ij})/s_{ij}$, where $m_{ij}$ and $s_{ij}$ are posterior mean and standard deviation of $y_{ij}$ under (2.4)–(2.5), screen for subgroups with exceptionally high or low death counts. In the Missouri example, county 66 had a residual of $2.24$ for the youngest age group, consistent with its elevated relative risk.

Why It Matters

For Disease Surveillance and Epidemiology

Reliable small-area mortality estimates are the starting point for identifying environmental causes of disease. Geographic variation in lung cancer, for example, may reflect smoking rates, occupational exposures (asbestos, radon), air quality differences, or socioeconomic gradients. Without reliable county-level estimates, these patterns are obscured by noise.

For Policy and Resource Allocation

Accurate small-area rates allow health departments to target prevention and screening resources toward high-risk communities. If raw rates are used instead, the noisiest small-county estimates will dominate the apparent geographic variation, misdirecting resources.

Connection to Modern Disease Mapping

Tsutakawa (1988) is an early example of what the spatial statistics literature now calls "disease mapping" — the construction of smoothed geographic risk maps using Bayesian hierarchical models. The paper predates the Besag, York, and Mollié (1991) intrinsic conditional autoregressive (ICAR) model, which extended the approach to spatially structured priors (adjacent counties are a priori more similar than non-adjacent ones). Modern implementations use Markov chain Monte Carlo (MCMC) via Stan or integrated nested Laplace approximations (INLA) to obtain full posterior distributions over all parameters, rather than the empirical Bayes approximation used here.

Connection to Social Security and DI Research

Geographic variation in disability insurance (DI) application rates, allowance rates, and beneficiary mortality are important policy questions. The small-area estimation problem arises whenever researchers want to study county-level DI outcomes: raw rates for rural, low-population counties are unreliable. The methods introduced here — gamma-Poisson hierarchical models with empirical Bayes estimation — are directly applicable to this setting. See DI Beneficiary Mortality and Geographic Variation in Disability Insurance.

Open Questions

• How sensitive are the empirical Bayes estimates to the inverse-gamma parametric assumption for $z_i$? Alternative priors (log-normal, conditional autoregressive (CAR)) may be more appropriate when spatial autocorrelation is present.
• How should the profile likelihood approach be extended to allow different $\sigma_j^2$ for each demographic group, not just a common $\sigma^2$?
• Does the slight decline in lung-cancer risk for the $75{+}$ age group (relative to $65$–$74$) reflect a genuine cohort effect, competing risks, or model misspecification?

Related

• Period Mortality
• Lee-Carter Model
• SSA Mortality Forecasting
• DI Beneficiary Mortality
• Robert K. Tsutakawa

Sources

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