CDC WONDER (Wide-ranging ONline Data for Epidemiologic Research) is a system developed by the Centers for Disease Control and Prevention (CDC) to promote information-driven decision making by public health practitioners and researchers and provide access to detailed public health information to the general public (1). Although CDC WONDER contains a wealth of data, it has limitations. Per CDC policy (2), any counts fewer than 10 should be suppressed for confidentiality reasons, resulting in left-censored data. Because of high rates of suppression, many chronic disease researchers opt to focus their inference in a few highly populated regions (3) or state- or national-level trends (4), despite known geographic disparities in many chronic disease outcomes (5,6). This suppression may also discourage research on disparities between subsets of the population (eg, race or sex disparities) to avoid reducing already small counts below suppression thresholds. In short, suppression of small counts exacerbates many issues commonly encountered in the field of small area estimation, where the term “small area” refers to a geographic scale (eg, county, census tract) at which the observed data alone do not provide reliable inference. Thus, when CDC WONDER data are used to conduct surveillance, the ability to estimate rates for rural areas and minority populations — where the chronic disease burden is high (7) — is significantly hindered by data suppression.

To address CDC WONDER’s data suppression issue, Tiwari et al (8) proposed an algorithm for estimating age-standardized rates in which suppressed age-specific counts are replaced with estimates based on the county’s age-specific population size and the state-wide average rate for that age-group. For example, suppose y_ik denotes the number of deaths from age-bracket k in county i of a population of size n_ik and our inferential interest lies in λ_ik, the corresponding mortality rate. Tiwari et al (8) proposed replacing the suppressed y_ik <10 with yik*=λ-si0k×nik, where s_i denotes the state that county i belongs to and λ-si0k denotes the state-wide average rate for age-bracket k in state s_i such that

λ-si0k=∑j:sj = siyjk/∑j:sj = sinjk (Equation 1)

Because state-level totals are often 10 or greater, we will assume from this point forward that λ-si0k is known and publicly available; when this is not the case, rates could be smoothed toward an alternative value (eg, national estimates).

Although this approach may yield reasonable estimates, it has drawbacks. First and foremost, estimating the uncertainty in age-standardized rate estimates is not an exact science when the data are known (9,10), much less when the data are highly suppressed. Furthermore, the algorithm is not designed to account for heterogeneity in demographic information such as the racial/ethnic make-up and socioeconomic status of the counties’ populations. As a result, inference based on these substituted data may be both biased (ie, smoothing toward the wrong values) and too precise (ignoring the uncertainty due to data suppression).

When the goal is to assess geographic disparities in age-standardized rates between regions, overcoming the privacy protections to obtain trustworthy estimates of the age-specific rates and their levels of uncertainty is only half the battle. For instance, Fay (11) followed the work of Fay and Feuer (9) to construct interval estimates for ratios based on F distributions. Tiwari et al (10) modified this work to yield more efficient interval estimation for rates and ratios of rates from nonnested regions, work that was later extended by Tiwari et al (12) for when one subregion is nested within a larger region (eg, a county nested within a state); Zhu et al (13) extended these approaches to more accurately account for spatial autocorrelation. When the age-standardized rates must be estimated from suppressed data, further modifications must be made or these approaches will fail to adequately account for all sources of uncertainty, yielding interval estimates that may be too narrow (14,15).

Rather than develop the statistical theory to accurately account for substitution-based approaches to overcome CDC WONDER’s privacy restrictions in variance calculations, we consider the use of Bayesian statistical models, which rely on data augmentation to make inference on the suppressed counts. As described by Fridley and Dixon (14), data augmentation approaches estimate the suppressed counts via multiple imputation (16) while simultaneously making inference on the parameters of interest — for example, λ_ik and the effects of potential risk factors. As noted by Zhu et al (13), Bayesian methods for modeling spatial data (17) can yield improved rate estimates when data are limited while simultaneously providing a mechanism for estimating uncertainty in rate estimates — uncertainty that can be seamlessly propagated into estimates such as age-standardized rates and rate ratios. That said, a key drawback of Bayesian methods is their tendency to rely on computationally burdensome Markov chain Monte Carlo (MCMC) methods.

The objective of this analysis was to illustrate 2 Bayesian approaches for estimating county-level mortality rates, by using heart disease mortality data from 1980 obtained from CDC WONDER (18), and to compare these results with those generated by the approach of Tiwari et al (8). In particular, we used a simple, nonspatial Bayesian model, which produces estimates similar to those from Tiwari et al (8), along with a more complex Bayesian model that accounts for spatial and between-age sources of dependence.

The study population for this analysis included all residents of the contiguous United States aged 35 or older during 1980. These data have multiple advantages. Because these data were collected before CDC’s suppression guidelines (2) went into effect, the public-use data are complete and free of suppression. Furthermore, because county definitions changed in several ways during the 1980s, the choice of data from 1980 allowed use of readily available shapefiles from the US Census Bureau for the I = 3,109 counties (or county equivalents) in the contiguous United States. To replicate the analysis of Tiwari et al (8), the data were separated into K = 6 groups: those aged 35 to 44, 45 to 54, 55 to 64, 65 to 74, 75 to 84, and 85 or older. Annual counts of heart disease–related deaths per county per age-group were obtained via CDC WONDER (18) and were defined as those for which the underlying cause of death was “diseases of the heart” according to the International Classification of Diseases, Ninth Revision (codes 390–398, 402, 404–429). Of the more than 18,000 counts in this data set, nearly half were fewer than 10.

The maps of the age-standardized rates generated from the raw data (Figure 1A) and the maps generated by the Poisson-gamma model (Figure 1C) have strong similarities, while artifacts of substituting state-wide averages for suppressed counts based on the approach of Tiwari et al (8) lead to elevated estimates in many rural counties in the upper Midwest (Figure 1B). In contrast, the map of the estimates from the MCAR model (Figure 1D) preserves the overall trends in the data while producing significantly smoother rate estimates.

The correlation results (Table 1) largely support this assessment. The Poisson-gamma approach produced age-standardized rate estimates that were the most highly correlated with the true rates, although the estimates obtained by using the substitution approach of Tiwari et al (8) had nearly an identical correlation. These 2 approaches differed in age-specific rate estimates. In particular, although the Poisson-gamma approach appeared to struggle for adults aged 35 to 44 — producing estimates that were less correlated with the truth — it outperformed the substitution approach for all groups aged 55 or older. Figure 2, which displays the age-specific rate estimates for adults aged 35 to 44 and adults 85 or older, explains how this occurred. Here, although the approach of Tiwari et al (8) gave every suppressed county in each state the same rate (by design), the Poisson-gamma model tended to overestimate rate estimates for those aged 35 to 44. According to Kerman (22), this overestimation of rates when counts are very small was to be expected. Furthermore, unlike the approach of Tiwari et al (8), the Poisson-gamma model produced full posterior distributions for each age-specific rate estimate, thereby allowing quantification of the uncertainty in these estimates. (Figure B.3 in the Web Appendix illustrates how only 4.5% of estimates for those aged 35 to 44 and 42.8% of all age-specific rate estimates from the Poisson-gamma model were deemed reliable.) When estimating rates for those 85 or older, the Poisson-gamma model permitted heterogeneity within states (Figure 2E); the inability to permit such heterogeneity is a key weakness of the approach of Tiwari et al (8). Further evaluation of the low age-specific correlations is provided in the Web Appendix (Figures B.1 and B.2).

Looking at the correlation results (Table 1) and the maps in Figure 1, one may wonder why we bother fitting the complex MCAR model. The DIC results (Table 2) explain why. Here, the MCAR model offered a model fit that is similar to the fit of the Poisson-gamma model (as measured by D-) while doing so with far fewer “effective model parameters” (p_D). To understand how this can be, recall that each λ_ik in Equation 4 had its own independent prior distribution; that is, the Poisson-gamma model did not shrink the λ_ik toward each other, producing estimates of the (p_D) for older age groups that approach the full I = 3,109 number of parameters. In contrast, the MCAR model explicitly imposed dependence between its model parameters, resulting in estimates of the (p_D) that were nearly 80% less than those from the Poisson-gamma model (eg, 10,785 vs 2,307). In addition, the estimates produced by the MCAR model were more precise (Web Appendix), and the smooth geographic patterns in Figure 1D, Figure 2C, and Figure 2F may provide clearer insight into the underlying trends in heart disease mortality.

This analysis highlighted some of the benefits of using Bayesian methods to account for left-censored data like those encountered in CDC WONDER. Although the Poisson-gamma model is a relatively simple approach, models (such as the MCAR model) that explicitly account for multivariate spatial dependence structures can lead to better inference by leveraging other sources of information to produce more reliable estimates.

The strengths of the MCAR model described in this analysis extend beyond modeling censored data to the broader field of small area estimation. As alluded to in the discussion of Equation 5, many benefits are associated with using the MCAR model in conjunction with covariate information when modeling chronic disease outcomes. Combining covariate information with spatial structure can produce more reliable estimates of the rates themselves, which is beneficial for disease surveillance, while simultaneously conducting inference on the potential risk factors that are included as covariates. When the covariates in the analysis are themselves spatially structured, it can be unclear if the covariate is effecting change in the outcome or vice versa, or if an unmeasured spatial confounder is influencing both the covariate and the outcome. In these settings, including a spatial random effect can lead to a phenomenon referred to as “spatial confounding” (29) and increase the standard errors associated with these covariates. Although the notion of spatial confounding has historically been considered a drawback of spatial models (29), others have argued (30) that inference from such models can help protect against type 1 error (ie, incorrectly rejecting the null hypothesis).

Finally, although we analyzed age-specific heart disease mortality as an illustration, the MCAR model is also well suited for analyzing rarer event data via its ability to jointly model multiple outcomes. This analysis leveraged information from older age groups with higher death counts to produce more reliable estimates for those aged 35 to 44. Similarly, one could jointly model a chronic disease outcome for multiple race/ethnicities, exploiting the shared factors that may lead to increased rates for non-Hispanic white persons and racial/ethnic minorities alike. Alternatively, one could use MCAR models to simultaneously analyze multiple chronic disease outcomes with similar etiologies to improve the reliability of all estimates.

Although the suppression of data creates an obstacle to conducting chronic disease surveillance, Bayesian statistical methods such as those described in this analysis can overcome these challenges while also producing more reliable estimates with valid uncertainty measures. By illustrating the benefits of and providing code for their implementation, we hope to ease the burden of using Bayesian models and broaden their application to censored data sets available from sources like CDC WONDER, thereby improving the inference made from public-use data.