A common way to compare property taxes across U.S. counties is the "effective rate": the median real estate tax owners pay, divided by the median home value. Both numbers come from the Census Bureau's American Community Survey (tables B25103 and B25077), and they exist for every county, which is why county rankings are built on them. I publish one of those rankings.
The ACS is a sample survey, and every estimate ships with a margin of error. Rankings usually drop that column. I wanted to know how much of the ordering survives once it's put back in.
Short answer: for big counties, nearly all of it. For small ones, very little. The median county's plausible rank spans 760 places out of 3,111.
The data
The ACS 2020–2024 5-year summary files are pipe-delimited, one row per geography, with estimate (E) and margin of error (M) columns side by side. For B25103 the first two values are the median tax and its MOE; for B25077 the median value and its MOE.
import math
def table(path, prefix="0500000US"): # 050 = county
rows = {}
with open(path) as f:
f.readline()
for line in f:
p = line.rstrip("\n").split("|")
if p[0].startswith(prefix):
rows[p[0][len(prefix):]] = [float(x) for x in p[1:3]]
return rows
tax = table("acsdt5y2024-b25103.dat") # [median tax, MOE]
val = table("acsdt5y2024-b25077.dat") # [median value, MOE]
Two kinds of rows need handling before any statistics. Tax medians of 10001 ("$10,000 or more") and 199 ("less than $200") are caps, not measurements; 24 counties have one, and their rate is a bound, so I left them out here. That leaves 3,111 counties.
Carrying the error through a ratio
ACS margins are 90% intervals (1.645 standard errors). The Census Bureau's handbook gives an approximation for the MOE of a ratio X/Y of two estimates:
def ratio_moe(x, mx, y, my):
r = x / y
return math.sqrt(mx**2 + r**2 * my**2) / y
rate = {}
for fips in tax.keys() & val.keys():
(t, tm), (v, vm) = tax[fips], val[fips]
if t in (10001, 199) or t <= 0 or v <= 0:
continue
rate[fips] = (100 * t / v, 100 * ratio_moe(t, tm, v, vm))
It ignores the correlation between the two estimates (same sample, same homes), so it's an approximation. It's the one the Bureau documents for this case, and it's good enough to see the size of the problem.
Typical relative margins: 6.5% on the median tax, 5.9% on the median value, 9.2% on the rate. Errors add up when you divide.
It depends almost entirely on county size
| Population | Counties | Median MOE of the rate | Relative |
|---|---|---|---|
| under 10,000 | 724 | ±0.164 points | 18.9% |
| 10,000–50,000 | 1,404 | ±0.074 points | 10.0% |
| 50,000–250,000 | 710 | ±0.040 points | 4.6% |
| 250,000 and up | 273 | ±0.019 points | 2.0% |
For Los Angeles County the rate is 0.680% ± 0.004. For a county of 3,000 people a typical margin is around ±0.16 points, which at a 0.8% rate is the difference between a cheap county and an average one.
The extremes are not usable as single numbers:
- Stonewall County, Texas (pop. 1,227): median home value $48,600 ± $42,933. The rate comes out at 1.79% ± 1.59.
- De Baca County, New Mexico: home value $187,500 ± $163,183, rate 0.60% ± 0.68. The interval runs past zero.
- Eureka County, Nevada: $82,600 ± $77,687.
In 8.6% of counties the margin exceeds a quarter of the rate itself.
From intervals to rank ranges
A rank is harder to reason about than a rate, so I converted each interval into the range of ranks it is consistent with. Rank 1 is the highest rate. A county's best plausible rank counts only the counties whose lower bound is above its upper bound; its worst counts every county whose upper bound reaches its lower bound.
import bisect
n = len(rate)
lows = sorted(r - m for r, m in rate.values())
highs = sorted(r + m for r, m in rate.values())
def rank_range(r, m):
best = n - bisect.bisect_right(lows, r + m) + 1
worst = n - bisect.bisect_left(highs, r - m)
return best, worst
Overlapping intervals aren't a formal significance test, so treat these as "can't be ruled out" ranges, not confidence intervals for the rank. They're still informative:
| County | Rate | Point rank | Plausible ranks |
|---|---|---|---|
| Menominee County, WI (pop. 4,252) | 3.56% ± 0.91 | 1 | 1–12 |
| Salem County, NJ | 2.93% ± 0.12 | 2 | 1–7 |
| Camden County, NJ | 2.83% ± 0.05 | 4 | 1–8 |
| Cook County, IL | 1.908% ± 0.013 | 109 | 77–191 |
| Stonewall County, TX (pop. 1,227) | 1.79% ± 1.59 | 146 | 1–3,109 |
| Harris County, TX | 1.623% ± 0.020 | 234 | 135–396 |
| Los Angeles County, CA | 0.680% ± 0.004 | 1,940 | 1,653–2,214 |
| Maricopa County, AZ | 0.438% ± 0.003 | 2,817 | 2,571–2,940 |
The Los Angeles row is the surprising one. Its own margin is tiny, yet its rank still spans 561 places, because hundreds of small counties with wide intervals sit near its rate. A precise estimate doesn't buy a precise rank when the neighbours are noisy.
The top of the table holds up reasonably well: New Jersey and upstate New York counties lead under any reading, and Menominee's wide margin still keeps it in the top 12. In the middle of the distribution, ranks mean little. The median county's range is 760 places wide, and 502 counties have an interval that includes the median county rate (0.804%). For those, "above or below average" isn't settled by the data.
"Highest in the state" is often a tie
State pages love a "highest property tax county in X" line. I tested the top two counties in each state with the Census difference test (the difference is significant at 90% if it exceeds sqrt(moe_a**2 + moe_b**2)):
def differs(a, b):
(ra, ma), (rb, mb) = a, b
return abs(ra - rb) > math.sqrt(ma**2 + mb**2)
In only 15 of 50 states is the #1 county significantly above #2. A few cases:
- New Jersey: Salem 2.93 ± 0.12 vs Camden 2.83 ± 0.05. Not distinguishable.
- New York: Allegany 2.92 ± 0.14 vs Orleans 2.72 ± 0.16. Difference 0.20, threshold 0.21. Not quite.
- Texas: El Paso 2.03 ± 0.04 vs Menard 1.95 ± 0.65. Menard's margin covers almost every Texas county.
- Illinois: Lake 2.58 ± 0.04 vs DeKalb 2.45 ± 0.07. This one is real.
What I took from it
- Keep the MOE column through the pipeline. It costs nothing to load, and once it's dropped nobody downstream can get it back.
- Population is a good proxy for trust. Under ~10,000 people, read the rate as a range, not a number.
- Prefer statutory rates where they exist. Texas, Florida and North Carolina publish adopted rates per taxing unit; those have no sampling error, only the transcription errors of a big spreadsheet. The ACS rate answers a different question anyway: what owners actually pay relative to what they think their home is worth, after exemptions and caps.
- Say "among the highest", not "the highest", unless the difference test passes. It's 15 states out of 50.
- Rank ranges are a better headline than ranks for the middle of the table. A county at #1,400 ± 380 is "about average", and that is the honest description.
The county figures on countypropertytaxes.com are the ACS point estimates described here, labeled as estimates.
Top comments (0)