Using empirical Bayes methods to rank counties on population health measures

Cluster Analysis Methods to Support Population Health Improvement Among US Counties

CONTEXT

Population health rankings can be a catalyst for the improvement of health by drawing attention to areas in need of relative improvement and summarizing complex information in a manner understood by almost everyone. However, ranks also have unintended consequences, such as being interpreted as "hard truths," where variations may not be significant. There is a need to improve communication about uncertainty in ranks, with accurate interpretation. The most common solutions discussed in the literature have included modeling approaches to minimize statistical noise or borrow strength from covariates. However, the use of complex models can limit communication and implementation, especially for broad audiences.

OBJECTIVES

Explore data-informed grouping (cluster analysis) as an easier-to-understand, empirical technique to account for rank imprecision that can be effectively communicated both numerically and visually.

DESIGN

Cluster analysis, specifically k-means clustering with Wasserstein (earth mover's) distance, was explored as an approach to identify natural and meaningful groupings and gaps in the data distribution for the County Health Rankings' (CHR) health outcomes ranks.

SETTING

County-level health outcomes from the 2022 CHR.

PARTICIPANTS

3082 counties that were ranked in the 2022 CHR.

MAIN OUTCOME MEASURE

Data-informed health groups.

RESULTS

Cluster analysis identified 30 health groupings among counties nationwide, with cluster size ranging from 9 to 184 counties. On average, states had 16 identified clusters, ranging from 3 in Delaware and Hawaii to 27 in Virginia. Number of clusters per state was associated with number of counties per state and population of the state. The method helped address many of the issues that arise from providing rank estimates alone.

CONCLUSIONS

Public health practitioners can use this information to understand uncertainty in ranks, visualize distances between county ranks, have context around which counties are not meaningfully different from one another, and compare county performance to peer counties.

Ranking composite Cancer Burden Indices for geographic regions: point and interval estimates

PURPOSE

To develop a composite Cancer Burden Index and produce 95% confidence intervals (CIs) as measures of uncertainties for the index.

METHODS

The Kentucky Cancer Registry has developed a cancer burden Rank Sum Index (RSI) to guide statewide comprehensive cancer control activities. However, lack of interval estimates for RSI limits its applications. RSI also weights individual measures with little inherent variability equally as ones with large variability. To address these issues, a Modified Sum Index (MSI) was developed to take into account of magnitudes of observed values. A simulation approach was used to generate individual and simultaneous 95% CIs for the rank MSI. An uncertainty measure was also calculated.

RESULTS

At the Area Development Districts (ADDs) level, the ranks of the RSI and the MSI were almost identical, while larger variation was found at the county level. The widths of the CIs at the ADD level were considerably shorter than those at the county level.

CONCLUSION

The measures developed for estimating composite cancer burden indices and the simulated CIs provide valuable information to guide cancer prevention and control effort. Caution should be taken when interpreting ranks from small population geographic units where the CIs for the ranks overlap considerably.

Modeling Area-Level Health Rankings

OBJECTIVE

Rank county health using a Bayesian factor analysis model.

DATA SOURCES

Secondary county data from the National Center for Health Statistics (through 2007) and Behavioral Risk Factor Surveillance System (through 2009).

STUDY DESIGN

Our model builds on the existing county health rankings (CHRs) by using data-derived weights to compute ranks from mortality and morbidity variables, and by quantifying uncertainty based on population, spatial correlation, and missing data. We apply our model to Wisconsin, which has comprehensive data, and Texas, which has substantial missing information.

DATA COLLECTION METHODS

The data were downloaded from www.countyhealthrankings.org.

PRINCIPAL FINDINGS

Our estimated rankings are more similar to the CHRs for Wisconsin than Texas, as the data-derived factor weights are closer to the assigned weights for Wisconsin. The correlations between the CHRs and our ranks are 0.89 for Wisconsin and 0.65 for Texas. Uncertainty is especially severe for Texas given the state's substantial missing data.

CONCLUSIONS

The reliability of comprehensive CHRs varies from state to state. We advise focusing on the counties that remain among the least healthy after incorporating alternate weighting methods and accounting for uncertainty. Our results also highlight the need for broader geographic coverage in health data.

Improving the Rank Precision of Population Health Measures for Small Areas with Longitudinal and Joint Outcome Models

Objectives

The University of Wisconsin Population Health Institute has published the County Health Rankings since 2010. These rankings use population-based data to highlight health outcomes and the multiple determinants of these outcomes and to encourage in-depth health assessment for all United States counties. A significant methodological limitation, however, is the uncertainty of rank estimates, particularly for small counties. To address this challenge, we explore the use of longitudinal and pooled outcome data in hierarchical Bayesian models to generate county ranks with greater precision.

Methods

In our models we used pooled outcome data for three measure groups: (1) Poor physical and poor mental health days; (2) percent of births with low birth weight and fair or poor health prevalence; and (3) age-specific mortality rates for nine age groups. We used the fixed and random effects components of these models to generate posterior samples of rates for each measure. We also used time-series data in longitudinal random effects models for age-specific mortality. Based on the posterior samples from these models, we estimate ranks and rank quartiles for each measure, as well as the probability of a county ranking in its assigned quartile. Rank quartile probabilities for univariate, joint outcome, and/or longitudinal models were compared to assess improvements in rank precision.

Results

The joint outcome model for poor physical and poor mental health days resulted in improved rank precision, as did the longitudinal model for age-specific mortality rates. Rank precision for low birth weight births and fair/poor health prevalence based on the univariate and joint outcome models were equivalent.

Conclusion

Incorporating longitudinal or pooled outcome data may improve rank certainty, depending on characteristics of the measures selected. For measures with different determinants, joint modeling neither improved nor degraded rank precision. This approach suggests a simple way to use existing information to improve the precision of small-area measures of population health.

Big data for population-based cancer research: the integrated cancer information and surveillance system

The Integrated Cancer Information and Surveillance System (ICISS) facilitates population-based cancer research by developing extensive information technology systems that can link and manage large data sets. Taking an interdisciplinary 'team science' approach, ICISS has developed data, systems, and methods that allow researchers to better leverage the power of big data to improve population health.