What the index found

Underservice has two shapes

Remote areas score badly because care is far away. Dense urban areas score badly despite being surrounded by providers, because health burden and barriers are high. Both are real, and an index built only on supply or only on need would surface one and hide the other.

The rural penalty

Areas under 6,000 residents land in the worst decile 26.7% of the time, against 1.6% for larger ones, an enrichment of 16.5×.

This survives the full uncertainty treatment, which matters because small places are exactly where a rate is least stable. It is also the pattern that a social vulnerability index tends to miss: those indices weight urban poverty and demographic composition, so remote communities with heavy health burden do not register.

This number has been pushed in both directions by measuring better. Replacing straight-line distance with road drive times raised it, confirming that straight lines understate rural isolation. Weighting safety-net clinics by their staffing rather than counting their doors lowered it again, because the safety net is sited where need is highest and crediting a large health centre properly narrows the gap it exists to close.

The most underserved areas

Ranked by the median across simulations, with the span each rank could plausibly occupy. The spans are wide on purpose: they are the honest width of the estimate, and the hairline near the right edge is where the worst decile begins.

The clusters, tested

A pattern on a map means nothing until it has been tested. The eye finds clusters in random noise reliably, so every area was checked against the null hypothesis that the same scores were scattered across the state at random.

The test needs a definition of “neighbourhood”, and the answer moves with it: at a single setting of 8 neighbours the test returns 268 hot areas, roughly triple what survives every setting. So only areas that came back clustered at every definition are reported.

Inland and northern California, rather than one contiguous block. Testing at a single neighbourhood size had made the pattern look far more concentrated than it holds up as.

What a ranking cannot see

Of 87 clustered areas, 44 also rank in the worst tenth reliably on their own. The other 43 are the interesting ones: places that are unremarkable by their own numbers but sit inside a badly served region.

This is the case for running both tests. The worst-ranked areas in the state are almost all tiny, a few thousand residents at most. These are mid-ranked towns of tens of thousands, and they only look like a problem once you account for the fact that everywhere around them is short of care too. A ranking treats each area as an island. A region is not an island.

Gi* tests against complete spatial randomness, and nearly every geographic variable is spatially autocorrelated. The result establishes where clustering is and how intense, not that clustering exists.

Where the index is least sure

The median rank interval spans 302 places out of 1,389. Only 85 areas reach the worst decile in at least 80% of simulations. Everything else should be read as a region of the ranking rather than a position in it.