One Map, Many Stories: Why Most County Comparisons Are the Same Story Twice

Look across enough charts on statsmapped and a pattern repeats: whatever you plot against whatever else, richer places and poorer places tend to land on opposite ends. We tested this directly. Across 38 county-level relationships checked in our data, 23 of them, 61%, weaken sharply once you hold local affluence constant. They aren't 23 separate discoveries. They're one gradient, showing up 23 times.

The rule, and how we tested it

We used mean weekly earnings as the affluence proxy. (The electoral-area worked example below is not one of the 38; it uses median equivalised income instead.) For each of the 38 pairs, we computed the raw correlation, then a partial correlation holding that proxy constant. Our working definition: a relationship is "the affluence gradient again" if, once the proxy is held constant, its correlation falls below |r| = 0.3, or falls to less than half its original strength. Twenty-three pairs meet that bar. Four more survive the control but look, on inspection, like artefacts of it rather than independent findings, so we've kept them out of both counts below rather than calling them confirmed.

The worked example

Take the share of an electoral area's residents whose field of study was science, maths or computing, plotted against median house prices in that same electoral area. Raw correlation: r = 0.78, across 136 LEAs. That looks like a real story -- STEM graduates cluster somewhere, house prices follow. Hold local income constant and it drops to r = 0.32. Most of what looked like a STEM effect was an income effect wearing a STEM costume: places with more STEM graduates also tend to be places with higher incomes, and higher-income places have higher house prices, full stop. The residual 0.32 is not nothing, but it's a different, much smaller claim than the headline number implies. (This comparison mixes 2016 Census field-of-study data with 2024 prices, and has 30 missing LEA cells -- worth knowing before you lean on the residual figure.)

We see the same collapse elsewhere: company formations against sale price falls from r = 0.75 to 0.233 once earnings are held constant; cattle numbers against sale price go from -0.571 to -0.245; social welfare payments per person against weekly earnings from -0.693 to -0.312; disposable income against sale price from 0.694 to 0.279. Four completely different comparisons, one underlying explanation each time.

What's left when you take affluence away -- and what we're not counting yet

Eleven of the 38 relationships pass the earnings test. But three of those eleven are all versions of the same underlying comparison -- hospital discharge rates against sale price, housing affordability and second-hand prices -- and we are holding them back: "all causes" folds in dialysis day-cases, which can add roughly 150 discharges a year for a single patient. With dialysis stripped out the link weakens (against sale price in 2025, r = -0.61 becomes -0.51), and it falls below our threshold once local deprivation is taken into account (Pobal HP index: -0.28), so it looks mostly like a deprivation story rather than an independent one. They pass the statistical test but we're treating them as a candidate for further work, not a story, so we've set them aside here.

That leaves eight relationships we're confident calling independent of affluence -- though even there, "new homes completed" is one of the two variables in five of the eight, so read this as a smaller number of underlying stories checked several ways, not eight unrelated discoveries. Three are worth naming. New homes completed against sale price, at county level, starts at r = 0.697 and is still r = 0.537 after controlling for both local earnings and age structure -- construction volumes and prices are genuinely linked, not just both tracking wealth. Live Register claims against social welfare payments per person barely move, r = 0.847 to r = 0.815 -- two measures of local hardship that track each other for reasons beyond who's rich or poor locally. And housing affordability (years of income needed to buy) against homes granted planning permission, at county level, goes from r = 0.504 to r = 0.408 -- a real pipeline relationship, weakened but not erased by affluence.

One pattern we are deliberately not putting a headline on: the hospital discharge pairs above. We mention them only as a candidate for further checking, not a finding, until the data-quality questions are resolved.

A warning, now built into the site

Because this pattern is so common, statsmapped now shows a computed warning on live comparison pages when a relationship mostly follows area size or location rather than anything more specific -- our "where do young buyers buy" page is the current example of one that triggers it. The point isn't that these comparisons are wrong, only that a strong correlation between two area-level numbers is, more often than not in this dataset, a restatement of "richer places differ from poorer places" -- and that's worth knowing before treating it as a discovery about either variable specifically.

Correlation, even after these controls, is not causation. A partial correlation removes one confound; it doesn't prove the remaining number is causal, only that it isn't fully explained by the one thing we checked.