United States · methodology

What this data is, and what it is not

1. The hidden third factor is density, not wealth

In this site’s country data, every pair is adjusted for GDP per capita (income per person). That is because nearly every national measure moves with how rich or poor a country is. County data has the same shape under a different name. Almost everything moves with how urban or rural a county is. So a county link, before adjusting for anything, usually just rediscovers that cities differ from the countryside.

So every pair here is adjusted. We compare places with the same population density, median household income and share of adults with a bachelor’s degree, so those differences can’t explain the link. Density is used in percentage terms (as a logarithm), so going from 10 to 20 people per km² counts the same as going from 1,000 to 2,000. If one of the two measures being compared is itself one of these three, it is dropped from the adjustment. Then each pair is worked out again comparing counties only with other counties in the same state (“state fixed effects”). And then again giving bigger places more weight.

2. Most county data in America is estimated by models, not measured

No survey is big enough to measure diabetes rates or median income in a county of four hundred people. So federal agencies build statistical models. They use national survey data plus known facts about each county to predict a county figure. 50 of the 205 measures here are predictions of that kind. CDC PLACES lists American Community Survey data among its own inputs. SAIPE mixes the American Community Survey with tax and benefit-programme records.

That creates two hidden tangles that nothing in the numbers shows. Both are flagged on the pair:

3,756 of 20,910 pairs are tangled in one of those ways. Only 16,080 — 77% — are independent enough to rank. The rest can still be explored, with the warning attached, instead of being deleted. That way a reader looking for one of them finds the reason, not a blank.

3. A count of anything is mostly a count of people

A measure’s density loading is how much it just reflects how urban a place is: how closely it tracks population density on its own. All 54 count measures here have a density loading between +0.28 and +0.89. All 212 rates and medians fall between +0.01 and +0.74. The two ranges do not overlap, and that is no accident: a big county has more of everything. Two counts have a middle (median) link of +0.97 even after adjusting. Adding population to the adjustments changes that by only 0.003. Nothing fixes it, so count pairs are marked and kept out of the ranking.

4. Three kinds of missing data, which must not be lumped together

Each kind skews the data in a different direction, so a single “percent covered” figure would hide all three. The third is the most dangerous. If you study death rates using only the counties with complete data, you are quietly studying urban America.

5. Counting counties or counting people

Without weighting, Loving County, Texas, counts the same as Los Angeles County. Giving bigger places more weight (weighting by population) makes the link describe the average resident instead of the average county. Both are worked out for every pair, since neither is the right answer on its own. People are so packed into a few counties that a weighted county link behaves as if it rested on about 266 counties — not 3,222. This is its “effective sample size”, and its precision is reported on that basis. 3,929 pairs point in opposite directions weighted and unweighted. But in only 45 of them are both directions unlikely to be chance. Mostly a link fades away under weighting, instead of flipping.

6. Neighbours look alike

Counties in the same region resemble each other for reasons that have nothing to do with any one pair. So a link across 3,222counties acts as if it had more separate cases than it really does. To allow for this, every pair gets two p-values. A p-value is the chance of seeing a link this strong by luck alone. The plain one treats every county as separate. The second (“state-clustered”) allows for the fact that counties in the same state tend to resemble each other.

Each pair also gets a Moran’s I, which shows how much nearby places look alike. It is measured on the leftovers: the part of each county’s figures that the adjustments do not explain. It uses each county’s 8nearest neighbours, by distance over the Earth’s surface. It is tested by shuffling the data many times (a permutation test), not with the usual bell-curve shortcut, since county data is too lopsided for that. 20,794 of 20,910 pairs leave leftovers that still cluster on the map.

7. County maps change over time

Connecticut replaced its eight counties with nine planning regions in 2022, and none of the old area codes carried over. The four agencies first added to this data handled that in four different ways. Two publish each year on the map of its own time. One redoes its whole history on today’s boundaries. One keeps the old codes for past years. And one carries both maps in a single file and splits the measures between them. Every year in the data has a mismatch somewhere, so matching sources on area codes alone would lose a whole state.

So the map version each source uses (its “vintage”) is worked out from the data itself. The old and new Connecticut code sets share no codes, so the codes a source uses show exactly which map it is on. Then everything is moved onto today’s boundaries before anything is compared. Values that had to be split across the change are flagged, and can be left out.

8. What we refuse to calculate

A median cannot be combined across areas. Averaging county medians gives you the average of a set of middle values, which means nothing. The same goes for life expectancy. It is worked out from combined deaths and years lived, not by averaging other life expectancies. And where a rate is published on one map and the number it is divided by on another, it cannot be moved between them either. In each case the system refuses and records that it refused, instead of guessing. The missing row is the honest answer.

Comparing one year with later years (“lagged” correlations) is refused too, for now. American Community Survey figures one year apart share four years of the same interviews. So a link between them partly measures that shared sample, not anything about time. Survey periods with no years in common need five years between them, and this data does not have that yet.

9. The United States does not divide into counties

It divides into 3,222county-level areas (“county equivalents”), and 2,999 of them are counties.

20of them have no county government at all. Alaska’s census areas are pieces of its Unorganized Borough that the Census Bureau draws so the whole state can be counted. Connecticut abolished county government in 1960. That is why this site calls the level an area, not a jurisdiction (a body that governs).

10. Places are scored against peers, not against the nation

Ranking counties nationally on almost any measure gives the same list over and over: wealthy suburbs at one end, the rural south at the other. So every place page compares each measure with the fifty counties most like it in population density and median household income, as well as with the whole country. These fifty are its peers. Density and income are both measured in percentage terms (as logarithms) and put on the same scale (standardised), so neither one outweighs the other. The gap between the two comparisons is the part worth reading.

Counts get no peer score. A county’s population cannot be high or low for its peers when the peers were picked partly by density. Density and income get no peer score either, since they define the group. And a peer score is only worked out when at least twenty peers have data.

11. The unit changes the answer

The same pair measured across counties, metro areas and states can differ in strength, and sometimes in direction. This is called the modifiable areal unit problem. It comes with the units themselves, not from a mistake in one of them, so each pair shows its own results for each level. Where a level is missing, the measure cannot be combined up to that level, and the results say so.

Migration is where this really matters. A move from Fort Bend County to Harris County leaves a county but does not leave Houston. And 38% of all county-to-county moves in the country start and end inside the same metro area — 79% in Roanoke, 76% in St. Louis, 74% in Minneapolis–St. Paul. Add up the counties, and New York’s arrival rate reads 4.9 per 100 residents who stayed. Count only the moves that crossed the metro’s own boundary, and it is 1.7. So metro flows here are built from county-to-county moves, with the moves inside the metro taken out. States use the IRS’s own state-to-state files. That way the three levels answer the same question about three different boundaries, instead of one question three times.

The domain scores (summary scores that each combine one topic’s measures) are built three times over for the same reason, once for each level. A score is standardised: it is measured against everyone else in the comparison, so it depends on who else is being compared. Averaging a metro’s county scores tells you how those counties rank among all counties, not how the metro ranks among metros. So a topic can be scored at one size and not another. States have three of the eight scores, since the sources behind the rest publish county figures only. The scores at all three sizes.

Sources

The years each source covers differ, some by a lot. Mobility (how far children climb the income ladder) describes children born in the early 1980s. School results are averaged over a decade that ended before the pandemic. Election results are from 2016. These dates are stamped on every figure they produce, not just here.

What this cannot tell you

• Nothing here shows that one measure changes another. Everything is a link between measures, adjusted or not.
• A county link describes places, not individual people. Finding that counties with more of one thing also have more of another does not mean the same people have both. Assuming they do is called the ecological fallacy, and it is the most common misuse of this kind of data.
• Adjusting is not proof. Taking out density, income and education removes those three, and nothing else. A third factor that goes with both measures, and is not in this data, could still be behind a link.
• The record of which measures are tangled together is only as good as what publishers tell us. It holds what CDC PLACES and SAIPE say about their own inputs. A source that is less open about its model would get through unflagged.
• Nothing here is compared with another country. The country data and this data measure different things in different ways. “Poverty” means living on under $3 a day in one, and a line twenty times higher in the other. So placing a state in a world ranking would report the gap between two definitions as if it were geography. States are compared with states, counties with counties.
• County crime is missing. The FBI publishes no county figures. County totals would have to be added up from roughly 19,000 separate police agencies. And the switch to the FBI’s newer reporting system (NIBRS) left whole states partly unreported. That is why the source this site uses for health data dropped its own violent crime measure instead of publishing it with gaps.

Back to the US overview · Every measure · Explore pairs · Cross-country methodology