Project Homepage | Key Findings and Recommendations | Documentation/Process | References, Data Sources, and Notes
The cleaned PLACES analytical table contained 3,145 county-level records before field-specific filtering.
Health-Metric Classification
2,909 counties were classified as standard.
48 counties contained at least one extreme health-metric value.
188 counties were excluded from the health-metric classification because of incomplete data.
The outlier definition used values more than three standard deviations from the national mean.
Because the formula used the absolute distance from the mean, these flags represent extreme values in either direction, not automatically high-prevalence or high-risk counties.
Across the 48 extreme-value counties, there were 72 individual health-metric outlier flags:
Stroke: 30
Mobility disability: 15
Coronary heart disease: 11
Physical inactivity: 10
Arthritis: 6
Outlier overlap showed additional concentration:
29 counties had one outlier metric.
14 counties had two outlier metrics.
5 counties had three outlier metrics.
Therefore, 19 of the 48 outlier counties contained two or more extreme health measures.
Cardiovascular Mortality Classification
21 counties were classified as cardiovascular-mortality outliers.
77 counties lacked a usable cardiovascular mortality rate.
61 mortality counts were explicitly suppressed by CDC WONDER because of low event counts.
Four counties were classified as both cardiovascular-mortality outliers and extreme outliers for at least one selected health metric.
145 counties had populations below 5,000 and were placed into the analysis-specific High Volatility population tier.
Among standard cardiovascular-mortality counties, all five analyzed health measures had statistically significant positive relationships with cardiovascular mortality:
Coronary heart disease: R² = 0.565, p < 0.0001
Trend line: Mortality per 100K = 54.956 × CHD prevalence - 65.133
Arthritis: R² = 0.507, p < 0.0001
Trend line: Mortality per 100K = 17.162 × Arthritis prevalence - 148.360
Stroke: R² = 0.485, p < 0.0001
Trend line: Mortality per 100K = 88.262 × Stroke prevalence + 0.886
Mobility disability: R² = 0.453, p < 0.0001
Trend line: Mortality per 100K = 19.931 × Mobility disability prevalence + 32.827
Physical inactivity: R² = 0.288, p < 0.0001
Trend line: Mortality per 100K = 10.498 × Physical inactivity prevalence + 62.920
Coronary heart disease had the highest individual R², while arthritis was notable for explaining slightly more variation in mortality than stroke in the bivariate model.
Among cardiovascular-mortality outlier counties included in the complete-case models, every slope reversed:
Coronary heart disease: R² = 0.341, p = 0.0086
Arthritis: R² = 0.432, p = 0.0022
Stroke: R² = 0.358, p = 0.0068
Mobility disability: R² = 0.394, p = 0.0040
Physical inactivity: R² = 0.219, p = 0.0431
The negative relationships are statistically interesting, but they do not demonstrate that lower documented disease prevalence causes greater mortality. They instead identify a subgroup that should be investigated further.
Tableau also evaluated models that included cardiovascular-outlier status as the grouping factor. The exported model structure was:
Cardiovascular Outlier Status × (Health Metric + Intercept)
The overall model R² values were:
Coronary heart disease → mortality: R² = 0.602
Arthritis → mortality: R² = 0.550
Stroke → mortality: R² = 0.530
Mobility disability → mortality: R² = 0.501
Physical inactivity → mortality: R² = 0.351
For each model, cardiovascular-outlier status was significant at p < 0.0001, supporting the decision to examine separate regression lines rather than assuming that one national relationship adequately described both groups.
Physical inactivity had some of the strongest relationships in the project.
Among standard cardiovascular-mortality counties:
Mobility disability: R² = 0.762, p < 0.0001
Stroke: R² = 0.572, p < 0.0001
Coronary heart disease: R² = 0.424, p < 0.0001
Arthritis: R² = 0.192, p < 0.0001
Cardiovascular mortality: R² = 0.288, p < 0.0001
The strongest finding in the project was the relationship between physical inactivity and mobility disability: counties with higher estimated inactivity prevalence also tended to have substantially higher estimated mobility-disability prevalence.
Within mortality-outlier counties, inactivity remained strongly associated with mobility disability (R² = 0.799, p < 0.0001) and stroke (R² = 0.565, p = 0.0002). Relationships with coronary heart disease (p = 0.0585) and arthritis (p = 0.0652) did not meet the project's p ≤ 0.05 threshold.
Among counties with populations greater than 5,000, larger population was associated with lower prevalence of each analyzed health measure, although population alone explained relatively little of the variance:
Arthritis: R² = 0.103
Coronary heart disease: R² = 0.101
Stroke: R² = 0.064
Mobility disability: R² = 0.054
Physical inactivity: R² = 0.037
All five relationships were statistically significant at p < 0.0001.
For counties below 5,000 residents, none of the corresponding population relationships were statistically significant. R² values ranged from approximately 0.005 to 0.019.
Population also had a weak negative relationship with mortality among standard mortality counties (R² = 0.065, p < 0.0001), while the relationship among mortality-outlier counties was not statistically significant (R² = 0.046, p = 0.353).
The appropriate takeaway is therefore not that population size is inherently protective, but that larger and smaller counties display different statistical behavior and that population alone is a weak explanatory variable.
Several geographic patterns were visible in the Tableau maps:
Stroke, mobility disability, and physical inactivity showed concentrations across portions of the American South.
Mobility-disability outliers were particularly visible in areas of Louisiana, Mississippi, and Alabama.
Arthritis prevalence followed a different pattern, with elevated areas in West Virginia, parts of the Northeast, northern Michigan and Wisconsin, the Pacific Northwest, and portions of the Southwest. The documented arthritis extreme outliers were concentrated in West Virginia.
Cardiovascular mortality did not form one equally distinct national cluster, although higher-rate pockets appeared throughout parts of the South and Midwest.
These geographic patterns are descriptive and should not by themselves be interpreted as evidence of regional causation.
1. Prioritize deeper review of overlapping outliers.
The four counties simultaneously classified as mortality outliers and health-metric outliers, along with the 19 counties containing two or more extreme health-metric flags, provide a reasonable starting point for more detailed investigation. Before labeling these areas "high risk," however, each metric should be checked to determine whether the outlier occurred above or below the national mean.
2. Use stroke and mobility disability as high-priority surveillance measures.
Stroke accounted for 30 of the 72 metric-outlier flags and mobility disability accounted for another 15. These measures may help identify counties where functional health burden deserves additional investigation.
3. Treat the inactivity findings as a prevention-planning signal rather than proof of causation.
The strong inactivity-mobility and inactivity-stroke relationships support further investigation of access to physical activity, transportation, rehabilitation, safe recreational space, and preventative care. Intervention effects would need to be evaluated separately.
4. Investigate the mortality-outlier reversal rather than explain it prematurely.
Future analysis should examine healthcare access, hospital and primary-care availability, insurance coverage, diagnostic access, rurality, age structure, socioeconomic conditions, and data completeness before attributing the reversed trend lines to "healthcare deserts."
5. Handle small-population counties separately.
Future versions should consider pooled multi-year mortality rates, confidence intervals, age-adjusted measures, or other rate-stabilization methods rather than relying on a single crude rate for very small populations.
6. Expand from bivariate to multivariable analysis.
Python-based regression could evaluate inactivity, age, population, income, insurance coverage, rurality, and healthcare access simultaneously, helping distinguish independent associations from confounding.
7. Add longitudinal analysis.
A multi-year dataset would allow counties to be evaluated based on whether health burden is improving, worsening, or remaining stable instead of relying on a single cross-sectional snapshot.
A more detailed breakdown of key findings can be found in the guiding case study document at this link (including documentation and process details) as well as a separate document detailing in depth descriptive statistics and trend line models.