Comparing ACS demographics for site selection: keep the margin of error

Before ranking two sites by American Community Survey demographics, check that the estimates describe comparable areas and periods, then test the difference using their margins of error. A higher income estimate alone may be too uncertain to support a confident ranking.

This matters when a retail site screen puts two nearby neighborhoods a few thousand dollars apart. The spreadsheet needs a way to record that uncertainty before the difference becomes a reason to eliminate a property.

Match the period before comparing the numbers

An ACS five-year estimate summarizes 60 months of collected data. For example, the 2020–2024 release covers that whole period; it is not a snapshot taken at the end of 2024. Five-year data offer broader geographic coverage and more precision than one-year data, at the cost of currency. Census explains the tradeoff between one-year and five-year estimates.

For a comparison between locations, use the same ACS product and period. For change over time, Census recommends non-overlapping five-year periods and warns against comparing a one-year estimate with a five-year estimate. The 2019–2023 and 2020–2024 releases share four years. Subtracting them does not measure a clean one-year change. See Census comparison guidance.

Write the full period in the column heading. “Median household income, 2020–2024 ACS” is much harder to misread than “2024 income.”

Keep the estimate and its margin of error together

ACS values come from a sample. The margin of error describes sampling uncertainty around the estimate. Census recommends statistical testing when comparing estimates and provides a spreadsheet tool that accepts the estimates and their margins of error. Its default confidence level is 90%. Download the Census statistical testing tool.

Do not delete the margin-of-error column to make an export easier to present. If the source supplies an annotation instead of a usable value, retain it and check the documentation. A missing margin of error is not zero uncertainty.

Statistical significance answers a limited question about the difference between estimates. It does not establish a tenant's likely revenue, the price someone should pay for land, or whether an observed difference matters commercially.

Worked example: the apparent income advantage

The following values are hypothetical. They are not observations about actual neighborhoods or a Pillar search result.

Suppose two separate, non-overlapping census tracts from the same ACS release have these median household income estimates:

• Tract A: $92,000, with a 90% margin of error of $10,000.
• Tract B: $85,000, with a 90% margin of error of $8,000.

The point estimates differ by $7,000. Under the standard independent-estimates approximation, the 90% margin of error for that difference is the square root of the sum of the squared margins of error:

Difference = 92,000 - 85,000 = 7,000
MOE of difference = sqrt(10,000² + 8,000²) ≈ 12,806
Approximate 90% interval = 7,000 ± 12,806
                        = -5,806 to 19,806

That interval includes zero, so this comparison does not establish a statistically significant income difference at the 90% level. Record: “A has a higher point estimate; the difference is not statistically significant at 90% under this test.” Do not rewrite that as “the neighborhoods have the same income.”

The Census tool tutorial explains how to enter and compare estimates. Use the appropriate method for your actual data. The example assumes independent estimates; overlapping trade areas and other dependent estimates need different treatment. Visual overlap between two separate confidence intervals is not a substitute for testing the difference.

For the site decision, keep both candidates under review and examine evidence that addresses the intended use. For a retail proposal, that could include verified access and the tenant's requirements. The traffic-count comparison guide covers a separate part of that review.

Check whether the geography changed

A familiar tract number is not enough to establish comparable boundaries over time. Census publishes relationship files connecting 2010 and 2020 tracts. Check the geography vintage before interpreting a change as movement in the market.

Also distinguish a published tract estimate from a vendor's custom-radius estimate. Ask how the custom area was assembled and how its uncertainty was calculated. Do not apply the worked example automatically to two overlapping circles around nearby properties.

Read the release's topic notes as well. The 2024 comparison guidance discusses changes that can affect comparisons, including population controls and income adjustments. Matching column names does not settle comparability.

Copy this comparison worksheet

Use one record per comparison. This is an editorial worksheet, not a Census form or an automated site score.

Decision this comparison will inform:
Candidate sites:
Source URL and retrieval date:
ACS table and variable:
Population measured (for example, households):
Product and full period:
Geography names, IDs and boundary vintage:
Published geographies or custom trade areas:

Estimate A:
Margin of error A and confidence level:
Estimate B:
Margin of error B and confidence level:
Missing values or source annotations:

Same product, period and definition? Yes / No / Unresolved
Comparable boundaries? Yes / No / Unresolved
Overlapping areas or other dependence:
Release notes checked:
Test/tool used and assumptions:
Difference and test result:

Claim supported by this comparison:
Claim it does not establish:
Next evidence needed before advancing/eliminating a site:

Keep the worksheet beside the site's other research. If a small demographic difference is the deciding factor, resolve its uncertainty before presenting the ranking as settled.