The map
What mortality looks like here, against how many deaths were expected
Search for your locality or click it. The card shows the standardized mortality ratio (SMR): 1.00 means exactly as many deaths as expected given the locality's population, age structure, and whether it's urban or rural; above 1.00 means more deaths than expected, below means fewer. The second button shows the real distance to the nearest hospital with continuous-stay beds.
Localities with fewer than 300 residents, or fewer than 10 expected deaths over nine years, stay in the "insufficient sample" class: at small populations, a handful more or fewer deaths in one year swings the rate wildly without meaning anything real. On small screens, pinch to zoom or use the corner buttons.
The real shape of the effect
A threshold, not a slope
We split the 3,177 localities with known population and distance into seven groups by distance to the nearest hospital, and computed the adjusted mortality ratio for each group as a whole (not the average of individual rates, so a handful of small localities can't dominate the figure). The result isn't a continuous slope.
Hospital-hosting localities have a ratio of 0.99 — essentially exactly what's expected, since they're towns with a younger, more urbanized population; the urban/rural adjustment already removes the "it's a town" advantage from the calculation. From 0 to 15 kilometres, the ratio barely moves: 0.96, 0.99, 1.01. Only from 15 kilometres does it start climbing — 1.03 at 15-20 km, 1.04 at 20-30 km — and past 30 kilometres it reaches 1.11, 12% higher than at 10-15 km. The direct correlation between distance and an individual locality's ratio is weak (r=0.15 for localities without their own hospital): that means distance alone explains little of the difference between localities — but the pattern across groups, where the noise of small localities cancels out through pooling, is clear.
Only 87 localities, with 244,255 residents (2023) — 1.1% of the analyzed population — are more than 30 kilometres from a hospital. It isn't spread evenly: almost all are in the Danube Delta (the communes around Sulina and Sfântu Gheorghe, where the road often means a boat as well) or in border communes in Satu Mare, Constanța and Mehedinți.
What we already removed
The urban-rural gap is bigger than the whole gap above
Before comparing distance, we computed the age-specific mortality rate separately for urban and rural areas, so an old, structurally aging village wouldn't be confused with the effect of a distant hospital. The difference between the two is itself the bigger figure in this whole story.
| Age group | Urban (‰) | Rural (‰) | Difference |
|---|---|---|---|
| 0-4 | 1.16 | 1.84 | +59% |
| 40-44 | 1.70 | 2.71 | +59% |
| 65-69 | 22.07 | 26.39 | +20% |
| 85 and over | 168.87 | 186.71 | +11% |
The gap is larger at middle age (+59% at 40-44) than at old age (+11% at 85 and over): late in life, death approaches a biological limit that even the better healthcare access of a city can't push back as far. This difference is already removed from the calculation above through standardization — the 0.99 to 1.11 figures show what's left AFTER the "it's a town" advantage is taken out, not added to it.
The counter-example
A hospital in town is not insurance
Several towns with their own hospital have adjusted mortality 25-55% above expectations — the confidence interval clearly excludes 1.00 in every case. Five are former mining towns — Petrila and Lupeni in the Jiu Valley, Cavnic in Maramureș, Rovinari in Oltenia's lignite basin, Nucet in Bihor — where industrial decline left behind unemployment and a mass departure of the young. The other five are small agricultural or border towns, without as clear a shared explanation.
- 1Hunedoara1.542,933 deaths observed vs. 1,906 expected
- 2Maramureș1.39598 deaths observed vs. 431 expected
- 3Hunedoara1.393,004 deaths observed vs. 2,169 expected
- 4Dolj1.361,052 deaths observed vs. 776 expected
- 5Gorj1.35656 deaths observed vs. 487 expected
- 6Tulcea1.29597 deaths observed vs. 462 expected
- 7Arad1.281,052 deaths observed vs. 821 expected
- 8Bihor1.28257 deaths observed vs. 201 expected
- 9Dolj1.282,336 deaths observed vs. 1,826 expected
- 10Arad1.271,211 deaths observed vs. 949 expected
All ten have zero distance to a hospital (they host their own) and a 95% confidence interval that excludes 1.00. The industrial explanation is plausible and well known for the five former mining towns, but it isn't measured directly by this data, and the other five don't have as clear an explanation — these are correlations, not a diagnosis of each locality. Click a name and the map moves there.
The real geography of distance
Where distance actually shows
At the other end, localities more than 20 kilometres from a hospital with a statistically robust ratio above 1.00 are not spread evenly across the country — they cluster in a few areas: the Danube Delta and border communes in Satu Mare, Arad, and Constanța.
- 1Satu Mare1.5222.1 km · 738 deaths vs. 486 expected
- 2Tulcea1.4960.4 km · 171 deaths vs. 115 expected
- 3Tulcea1.4730.2 km · 1,112 deaths vs. 758 expected
- 4Satu Mare1.4626.0 km · 391 deaths vs. 267 expected
- 5Satu Mare1.4630.3 km · 389 deaths vs. 267 expected
- 6Satu Mare1.4325.6 km · 184 deaths vs. 128 expected
- 7Tulcea1.3957.2 km · 544 deaths vs. 392 expected
- 8Tulcea1.3935.4 km · 371 deaths vs. 268 expected
- 9Constanța1.3739.6 km · 271 deaths vs. 198 expected
- 10Arad1.3323.9 km · 1,070 deaths vs. 804 expected
Each has a 95% confidence interval that excludes 1.00 and at least 10 expected deaths over the nine years. Click a name and the map moves there.
What we left out
A question left open
A cluster of localities in Vâlcea and Teleorman, 20-25 kilometres from a hospital, has a very low adjusted ratio — statistically "robust" (the confidence interval excludes 1.00), but implausibly low next to the country's typical rate. We haven't published it as a list, and we draw no conclusion from it. The most likely explanation is that the population registered with a domiciliu there (the rate's denominator) includes many people who left long ago to work abroad and no longer actually live in the locality — but we couldn't verify this directly, and we found no calculation error that would explain it otherwise. It remains an open question, not a finding.
Method and sources
Where the numbers come from
Deaths, by locality and year (2014-2022), come from the INS TEMPO matrix POP206D, which allocates each death to the person's last domiciliu or usual residence — not to the locality where the death occurred, so as not to artificially inflate the mortality of towns with large hospitals that draw patients from the whole area. Population by locality, age (five-year groups) and year comes from POP107D. National age-specific rates, computed separately for urban and rural areas, come from POP206A (deaths) divided by the same POP107D population, aggregated. Hospitals — the "continuous-stay bed" category, public and private combined, year 2024 — come from SAN102C, the same data pulled for Atlas of Medical Deserts on the night of 8-9 September 2026.
The adjusted mortality ratio (SMR) is an indirect standardization: for each locality, we compute how many deaths would be "expected" if every age group died at the national rate for its residence class (urban or rural), applied to the locality's real population in that age group; the SMR is the ratio of observed to expected deaths. A ratio of 1.20 means 20% more deaths than the locality's age and residence class would suggest; 0.80 means 20% fewer. The standardization controls only for age and urban/rural status — not income, occupation, road quality, or ambulance staffing, factors that can influence real mortality just as much. For localities with fewer than 10 expected deaths over nine years, a single unusual month can move the ratio by tens of percent; these stay unclassified on the map as "insufficient sample," and the ranked lists in the text additionally require a 95% Byar confidence interval that excludes 1.00.
Distance to hospital is calculated as a straight line (the haversine formula), between each locality's geographic centroid and the centroid of the nearest locality with at least one continuous-stay hospital bed in 2024 — not real road distance, not ambulance travel time. 289 localities host their own hospital (their distance is zero); the other 2,892 are measured against the nearest one. The boundaries of the 3,181 administrative units and their centroids come from the "UAT, Romania (polygon)" dataset published by geo-spatial.org based on ANCPI data, the 26 March 2025 edition, under a CC BY-SA 4.0 licence.
The analysis window is 2014-2022, not 2014-2025: at the time of download, POP206D and POP206A had no rows yet for 2023 at this level of detail, though POP107D already did; we checked directly across all 43 county files before capping the window. An independent check of the construction of the expected rate, the distance calculation, and the distance-band tiling, done before publication, flagged this window mismatch — fixed before the final calculation — and confirmed that dropping 2023 from the calculation barely changed the results by distance band.
These figures are descriptive and correlational. They show a real, but modest, association between distance to a functioning hospital and adjusted mortality — not proof that distance alone causes the excess deaths. No figure of the "this many deaths could have been avoided" kind appears anywhere on this page: a direct comparison between localities without their own hospital and those that have one would credit distance with a difference that mostly reflects urbanization, income, and the occupational structure of hospital towns — factors this analysis does not measure.