Rural America's health is meaningfully worse than urban America's, and the gap has roughly tripled over the last two decades — but it is overwhelmingly a within-state, not a between-state, story

Summary

Rural (nonmetro) America's health is measurably worse than urban America's, and the gap has widened sharply over the last two decades. In the current County Health Rankings release, nonmetro counties average an 18% higher premature-death rate (years of potential life lost before age 75) than metro counties -- 11,051 vs 9,359 per 100,000. Nationally, USDA's Economic Research Service, analyzing CDC mortality data, found the rural disadvantage in working-age (25-54) natural-cause mortality grew from a 6% rural excess in 1999-2001 to a 43% rural excess in 2017-2019 -- and a separate peer-reviewed analysis found overall rural-urban disparities in age-adjusted mortality tripled between 2000 and 2019.

The rural penalty is overwhelmingly a within-state phenomenon, not a between-state one. Computing each state's own rural-minus-urban health gap and averaging those 47 state-level gaps (population-weighted) reproduces the entire national gap almost exactly (1,817 vs. the national raw gap of 1,692) -- meaning the disadvantage is not an artifact of rural counties happening to sit in already-worse-off states. It shows up broadly: 44 of 47 states with both metro and nonmetro counties have worse rural health, and the pattern is statistically robust (paired t=8.40, p=7.6x10^-11). By contrast, which STATE a county sits in explains only 29% of all county-to-county variation in premature death nationally, and adding rural/urban status on top only raises that to 35% -- the majority of health variation (65%) is genuinely place-specific, not explained by either geography. Note: all figures below apply the filter c.premature_death_rate IS NOT NULL, which excludes 64 privacy-suppressed counties (2.0% of the joined sample) -- see the Data, method, and limitations section for the exact count and its direction.

How much worse is rural health, and how has that changed over 20 years?

Two independent measurements, one current and one historical, both point the same direction:

• Current cross-section (CHR 2025 release, pooling recent years of NCHS mortality data): nonmetro counties (RUCC 4-9, n=1,902) average 11,051 years of potential life lost (YPLL) per 100,000 residents under 75, versus 9,359 for metro counties (RUCC 1-3, n=1,178) -- an 18.1% gap.
• 20-year trend (USDA ERS, EIB-265, analyzing CDC WONDER): for the prime working-age population (25-54), the rural excess in natural-cause mortality (disease-related deaths, excluding overdose/suicide/accidents) grew from 6% higher than urban in 1999-2001 to 43% higher in 2017-2019. Urban natural-cause mortality for this age group actually fell 37 deaths per 100,000 over that period while rural natural-cause mortality rose 14 per 100,000 -- so the gap widened from both directions, not just rural decline. The gap was driven disproportionately by women: rural working-age female natural-cause mortality rose 18% in the most rural counties versus a 3% rise for rural males in the same counties, while urban rates fell for both sexes.
• A separate 2023 peer-reviewed study (Weeks et al., PLOS Global Public Health) cites that overall age-adjusted mortality disparities between rural and large-metro areas tripled between 2000 and 2019, consistent with the ERS finding.
• An earlier long-run study (Singh & Siahpush, Am J Prev Med) found the gap in life expectancy at birth widened from 0.4 years (1969-71) to 2.0 years (2005-09) -- so the widening has been underway for over 50 years, not just the last 20; the last two decades appear to be an acceleration of a much longer trend.

Leading causes driving the recent widening: heart disease, lung disease, diabetes, liver disease, and (for working-age women specifically) pregnancy-related deaths, which rose 313% in rural areas over the period.

Is the gap a between-state or a within-state story? -- the direct decomposition

This is the sharper methodological question, and it has a clean, computable answer using County Health Rankings' current county-level premature-death data joined to USDA's Rural-Urban Continuum Codes.

Test 1 -- does the national gap survive controlling for state? For the 47 states with both metro and nonmetro counties, I computed each state's own (nonmetro mean - metro mean) gap, then took the population-weighted average across all 47 states. That within-state-average gap is 1,817 YPLL/100k -- essentially identical to (very slightly larger than) the raw, unadjusted national gap of 1,692. If the national gap were mostly a compositional artifact of rural counties clustering in worse-performing states, controlling for state would shrink the gap substantially. It does not shrink at all. A paired one-sample t-test on the 47 state-level gaps confirms this is not noise: t=8.40, p=7.6x10^-11, and 44 of the 47 states show nonmetro counties worse than metro counties within that same state.

Test 2 -- how much of ALL county health variation is state-driven vs. place-driven? A two-level ANOVA decomposition of premature-death rates across all 3,080 counties finds: State alone explains 29.0% of total variance; metro/nonmetro status alone explains only 4.6%; the two together explain 35.3% (nearly additive, confirming they are not proxies for each other -- the correlation between a state's baseline urban health and the size of its rural penalty is near zero, r=0.06). The remaining 64.7% of variation is neither a state effect nor a rural/urban effect -- it is genuinely place-specific (a specific county's own economy, hospital access, industry, demographics).

Conclusion: being rural carries a health penalty that shows up consistently within almost every state, essentially independent of which state that rural place happens to be in -- this is a within-state (place-level) phenomenon, not a between-state one. At the same time, state of residence is a considerably larger single predictor of overall county health (29%) than rural/urban status alone (4.6%) -- so if the question is 'what predicts county health in general,' state matters more; but if the question is specifically 'what produces the rural-urban gap,' state composition explains essentially none of it -- the same rural disadvantage recurs inside nearly every individual state.

There is real state-to-state variation in how LARGE the gap is (from a 7,131 YPLL/100k gap in Arizona and 5,650 in South Dakota, down to a handful of states -- Idaho, Wyoming, Massachusetts, and roughly flat Delaware -- where nonmetro counties do as well or better). That state-to-state variation in gap SIZE is real and interesting, but it rides on top of a gap that is present in the overwhelming majority of states, not caused by a few states driving a national average.

Data, method, and limitations

Tables used: health.chr_premature_death (County Health Rankings 2025 release, county-grain premature death / YPLL rate, joined via 5-digit FIPS) and geo.rural_urban_continuum (USDA ERS Rural-Urban Continuum Codes, 2023 vintage, metro_nonmetro flag). RUCC classification changes only once a decade and was applied to the current cross-section; it was not itself measured across 20 years.

Exclusion disclosed, and its direction quantified: every query in this report applies the filter c.premature_death_rate IS NOT NULL. Joining the two tables on county FIPS yields 3,144 matched counties; the filter c.premature_death_rate IS NOT NULL removes 64 of those 3,144 counties (2.0%) -- CHR's own privacy-suppressed, small-numerator counties (data_flag=2) -- leaving the 3,080 counties used in every statistic in this report (the ANOVA, the 47-state paired-gap test, and the metro/nonmetro averages of 11,051 vs. 9,359). Rerunning the same join WITHOUT the filter c.premature_death_rate IS NOT NULL (i.e. counting all 3,144 matched counties by category) shows the 64 excluded counties are not a random 2%: 56 (87.5%) are nonmetro and only 8 (12.5%) are metro. This corpus carries no imputed or model-based rate for these suppressed counties (unlike health.cdc_county_injury_mortality's rate_m column, chr_premature_death has no fallback estimate), so the headline gap cannot be recomputed with those 64 counties included using this table. Because privacy suppression triggers on a small numerator (few deaths in a small population), the 56 excluded nonmetro counties are plausibly among the smallest, most remote counties in the sample; if their true rates run even modestly above the nonmetro average (plausible on priors, not verified here), the reported 18.1% cross-sectional gap and the 1,692/1,817 gap figures would understate, not overstate, the true rural disadvantage.

A genuine data gap: this corpus's own 20-year county-level trend table for premature death, health.chr_premature_death_trends (1997-2023), returned an S3 404 error on every query attempt (including after a schema-catalog rebuild) -- its backing data file appears to be missing or misregistered. This was reported as a defect. Had it worked, it would have let us run the full state-vs-within-state variance decomposition (Test 2 above) at multiple points across the 20-year window, rather than at a single current cross-section plus externally-sourced historical trend figures. As published here, the trend magnitude (6%->43%, tripling, widening life-expectancy gap) comes from USDA ERS's and peer-reviewed researchers' own published analyses of CDC data, not from a table this corpus computed directly; the within-state/between-state decomposition is a direct computation from this corpus's current data, but only for the present-day cross-section, not repeated over the two-decade window the question asks about.

Premature Death (YPLL) vs. full mortality/life-expectancy: YPLL is a truncated measure (deaths before age 75) rather than an all-ages age-adjusted rate or life expectancy; it is standard for county comparisons but not numerically identical to the life-expectancy-gap or all-cause-mortality-rate figures cited from the literature above -- the two are directionally consistent but not the same statistic, and should not be treated as interchangeable.

Every query behind this report

11 warehouse calls ran in this session, in order. Each is reproducible against the same snapshot.

SELECT table_name FROM information_schema.tables WHERE table_schema='health' AND table_name LIKE '%premature%'

SELECT r.metro_nonmetro, count(*) n, avg(c.premature_death_rate) avg_rate, stddev_samp(c.premature_death_rate) sd
FROM health.chr_premature_death c
JOIN geo.rural_urban_continuum r ON r.county_fips = c.fips_code
WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
GROUP BY r.metro_nonmetro

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
state_grp AS (
  SELECT state_fips, metro_nonmetro, avg(rate) m, count(*) n
  FROM d GROUP BY state_fips, metro_nonmetro
),
piv AS (
  SELECT a.state_fips, a.m AS metro_mean, a.n AS metro_n, b.m AS nonmetro_mean, b.n AS nonmetro_n,
         (b.m - a.m) AS gap
  FROM state_grp a JOIN state_grp b ON a.state_fips=b.state_fips
  WHERE a.metro_nonmetro='Metro' AND b.metro_nonmetro='Nonmetro'
)
SELECT count(*) n_states,
       avg(gap) unweighted_avg_gap,
       sum(gap*(metro_n+nonmetro_n))/sum(metro_n+nonmetro_n) weighted_avg_gap,
       stddev_samp(gap) sd_gap,
       min(gap) min_gap, max(gap) max_gap,
       corr(gap, metro_mean) corr_gap_statelevel
FROM piv

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
grand AS (SELECT avg(rate) gm, count(*) NN FROM d),
state_stats AS (SELECT state_fips, avg(rate) sm, count(*) nc FROM d GROUP BY state_fips)
SELECT
 (SELECT NN FROM grand) N,
 (SELECT count(*) FROM state_stats) num_states,
 (SELECT sum(nc*(sm-gm)*(sm-gm)) FROM state_stats, grand) ss_between_state,
 (SELECT sum((d.rate-ss.sm)*(d.rate-ss.sm)) FROM d JOIN state_stats ss ON ss.state_fips=d.state_fips) ss_within_state,
 (SELECT sum((rate-gm)*(rate-gm)) FROM d, grand) ss_total

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
grp_stats AS (SELECT state_fips, metro_nonmetro, avg(rate) gsm, count(*) gc FROM d GROUP BY state_fips, metro_nonmetro),
grand AS (SELECT avg(rate) gm, count(*) NN FROM d)
SELECT
 (SELECT NN FROM grand) N,
 (SELECT count(*) FROM grp_stats) num_state_metro_groups,
 (SELECT sum(gc*(gsm-gm)*(gsm-gm)) FROM grp_stats, grand) ss_between_group,
 (SELECT sum((d.rate-g.gsm)*(d.rate-g.gsm)) FROM d JOIN grp_stats g ON g.state_fips=d.state_fips AND g.metro_nonmetro=d.metro_nonmetro) ss_within_group

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
grp_stats AS (SELECT metro_nonmetro, avg(rate) gsm, count(*) gc FROM d GROUP BY metro_nonmetro),
grand AS (SELECT avg(rate) gm, count(*) NN FROM d)
SELECT (SELECT NN FROM grand) N,
 (SELECT sum(gc*(gsm-gm)*(gsm-gm)) FROM grp_stats, grand) ss_between_metrocat,
 (SELECT sum((rate-gm)*(rate-gm)) FROM d, grand) ss_total

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
state_grp AS (SELECT state_fips, metro_nonmetro, avg(rate) m, count(*) n FROM d GROUP BY state_fips, metro_nonmetro),
piv AS (
  SELECT a.state_fips, s.state_name, s.state_abbr, a.m metro_mean, b.m nonmetro_mean, (b.m-a.m) gap
  FROM state_grp a JOIN state_grp b ON a.state_fips=b.state_fips
  JOIN geo.state_ref s ON s.state_fips=a.state_fips
  WHERE a.metro_nonmetro='Metro' AND b.metro_nonmetro='Nonmetro'
)
SELECT * FROM piv ORDER BY gap DESC

WITH d AS (
  SELECT r.state_fips, r.metro_nonmetro, c.premature_death_rate rate
  FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
  WHERE c.geo_level='county' AND c.premature_death_rate IS NOT NULL
),
state_grp AS (SELECT state_fips, metro_nonmetro, avg(rate) m FROM d GROUP BY state_fips, metro_nonmetro),
piv AS (
  SELECT a.state_fips, (b.m-a.m) gap
  FROM state_grp a JOIN state_grp b ON a.state_fips=b.state_fips
  WHERE a.metro_nonmetro='Metro' AND b.metro_nonmetro='Nonmetro'
)
SELECT gap FROM piv

SELECT DISTINCT release_year FROM health.chr_premature_death WHERE geo_level='county'

SELECT
  count(*) total_joined,
  sum(CASE WHEN c.premature_death_rate IS NULL THEN 1 ELSE 0 END) null_rate_rows,
  avg(c.premature_death_rate) FILTER (WHERE r.metro_nonmetro='Metro') metro_avg_excl_nulls,
  avg(c.premature_death_rate) FILTER (WHERE r.metro_nonmetro='Nonmetro') nonmetro_avg_excl_nulls
FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips = c.fips_code
WHERE c.geo_level='county'

SELECT r.metro_nonmetro, count(*) 
FROM health.chr_premature_death c JOIN geo.rural_urban_continuum r ON r.county_fips=c.fips_code
WHERE c.geo_level='county' AND c.premature_death_rate IS NULL
GROUP BY r.metro_nonmetro