Lesson 1 of 9 · about 16 minutes
Finding the Data Before You Need It
You'll be able to
- Match a story question to the agency most likely to hold the data
- Tell the decennial census apart from the American Community Survey and pick 1-year or 5-year ACS estimates
- Read an ACS margin of error and test whether two estimates really differ
- Write a records request for a database export with a record layout
Most data stories start the same way. Someone at a meeting says a number, and you wonder if it is true. The reporter who already knows where the numbers live can check it before the meeting ends. The reporter who doesn't spends two days searching, then quotes the official because the deadline came first.
So learn the map now, while nothing is due.
The federal shelf
A handful of federal agencies publish most of the numbers local stories need. They are free, documented, and go down to the county level for most topics. Bookmark these.
| Agency | What it has | Smallest usual geography | How often |
|---|---|---|---|
| Census Bureau, decennial census | Full population count, age, sex, race, housing units, occupancy | Census block | Every 10 years (Census Day April 1, 2020; next is 2030) |
| Census Bureau, American Community Survey | Income, poverty, commuting, rent, home values, education, internet access, veterans | Census tract and block group (5-year data) | Yearly |
| Bureau of Labor Statistics (BLS) | Unemployment (LAUS program), wages and jobs by industry, consumer prices (CPI) | County, metro area, many cities | Monthly and yearly |
| Bureau of Economic Analysis (BEA) | Personal income, per capita income, GDP by county | County | Yearly (2024 county data released February 5, 2026) |
| Energy Information Administration (EIA) | Electricity sales, prices, customers by utility (Form EIA-861), power plants, fuel | Utility, state, plant | Monthly and yearly (final 2025 EIA-861 data released October 7, 2026) |
| USDA National Agricultural Statistics Service (NASS) | Census of Agriculture, crop and livestock surveys, Quick Stats database | County | Census every 5 years (latest 2022, released February 13, 2024); surveys yearly or more often |
Census data lives at data.census.gov. BLS has a series search and a CPI inflation calculator. NASS puts everything into Quick Stats, which lets you filter by commodity, year and county and download a CSV. EIA's electricity data browser draws charts you can check against the raw tables.
Decennial census versus ACS
Beginners mix these up constantly. The decennial census tries to count every person. It asks few questions, and its numbers are counts, with no margin of error printed beside them. The ACS asks a sample of households a long list of questions every year, so its numbers are estimates that come with a margin of error. If you want how many people lived in a county on April 1, 2020, use the decennial count. If you want median household income, there is no decennial number at all. That is an ACS question.
The ACS comes in flavors, and the Census Bureau states the rules plainly:
- 1-year estimates: areas with 65,000 or more people. Most current, smallest sample, least reliable.
- 1-year supplemental estimates: areas with 20,000 or more people, a shorter list of tables.
- 5-year estimates: all areas, down to block groups. Built from 60 months of responses. Most reliable, least current.
For a rural county of 18,000 people, the 5-year estimate is the only one that exists. Say so in your copy: "the Census Bureau's 2020 to 2024 American Community Survey estimate," for example. A 5-year estimate describes a period, and writing it as if it were one year is a small, common error.
Reading an ACS margin of error
Every ACS estimate comes with a margin of error. The Census Bureau publishes these at the 90 percent confidence level. In plain words: if the survey were repeated many times, the range of estimate plus or minus MOE would catch the true value about 90 times in 100.
Example: are these two counties really different?
Invented Cedar Flats County has a median household income estimate of $58,412, MOE plus or minus $4,950. Neighboring Harlan Valley County shows $62,100, MOE plus or minus $3,880. A press release says Harlan Valley households "earn more." Do they?
The Census Bureau's own test: divide the difference by the square root of the sum of the squared MOEs. If the result is greater than 1.0, the difference is statistically significant at 90 percent confidence.
- Difference: 62,100 minus 58,412 = 3,688
- Squared MOEs: 4,950 squared = 24,502,500; 3,880 squared = 15,054,400
- Sum: 39,556,900. Square root: about 6,289
- Test value: 3,688 divided by 6,289 = 0.59
0.59 is under 1.0. The survey cannot tell these counties apart. Honest copy: "The two counties' median incomes are statistically indistinguishable in Census estimates."
Two more tools from the same Census guidance come up often.
Adding estimates. Say you combine two tracts to describe a neighborhood: 1,240 renters (MOE 310) and 860 renters (MOE 270). The total is 2,100. Its MOE is approximately the square root of 310 squared plus 270 squared, which is about 411. The MOEs do not simply add, and the combined MOE is smaller than 310 plus 270 = 580.
Changing confidence levels. If you are comparing against a poll that reports at 95 percent, convert: divide the ACS MOE by 1.645 to get the standard error, then multiply by 1.96. The $4,950 MOE becomes about $5,898 at 95 percent.
State and local shelves
Federal data tells you about people. State and local data tells you about government, which is usually the story. Places to look:
- State open data portals and transparency sites. Many states post vendor payments, salaries and contracts. Search your state name plus "open data" or "transparency."
- Agency budgets and audits. Adopted budgets, the audited annual financial report, and outside audit findings. These are PDFs more often than spreadsheets, which is why the cleaning lesson covers PDF tables.
- State agencies for licensing, inspections, crashes, school enrollment and test scores, and well permits.
- Meeting packets. Staff reports often contain the exact spreadsheet a board saw, pasted as a table.
Asking for a database
Most government records worth having sit in a database: permits, inspections, jail bookings, payments, 911 calls. You do not want a printout. You want the export.
The federal Freedom of Information Act is direct on this. 5 U.S.C. 552(a)(3)(B) requires an agency to provide a record "in any form or format requested" if it is readily reproducible in that form, and 552(a)(3)(C) requires reasonable efforts to search electronic records. State public records laws differ in wording, so read yours. For Idaho requests, Jeff's free Idaho FOIA / Public Records Letter Generator builds the letter, and the Laws and Rules module in this course covers records law in more depth.
Example: a request that gets usable data
Please send me all your building permit information.
That one invites a 4,000-page PDF. Compare:
"I request an export of the building permit database for permits issued January 1, 2023, through June 30, 2026, in CSV or Excel format, with all fields except those exempt under the law, plus the record layout or data dictionary that defines each field and code. If any field is withheld, please identify it and the exemption claimed."
Ask for the layout every time. A column named STAT_CD with values 1 through 7 means nothing until someone tells you 4 means "expired" and 6 means "withdrawn by applicant."
Reporting angle
- Ask the agency what software holds the records. A clerk who says "it's all in our permitting system" just told you an export button exists.
- Request the record layout and a count of total records, so you can check that you got everything.
- Before you build a story on ACS figures, check the MOE column. If the agency or press release dropped it, put it back.
- Ask the agency for the exact table or report a public official quoted. Officials often cite a number from a slide that came from a spreadsheet nobody published.
- Compare BLS county unemployment with the state labor department's release for the same month. They should match. If they don't, find out why.
Common mix-ups
- Claim: Census numbers are exact counts. What the evidence shows: Decennial totals are counts; ACS figures are sample estimates with published margins of error.
- Claim: A 5-year ACS estimate is out of date, so a 1-year estimate is better. What the evidence shows: For areas under 65,000 people, a 1-year estimate does not exist. Even where both exist, the 5-year number is more reliable.
- Claim: An agency can hand you a PDF when you asked for a spreadsheet. What the evidence shows: Under federal FOIA, the agency must use your requested format if the record is readily reproducible in it. State laws vary.
Try it
- Go to data.census.gov and find median household income (table B19013) for your county and one neighbor, using the latest 5-year estimates.
- Write down both estimates and both MOEs.
- Run the Census test: difference divided by the square root of the sum of the squared MOEs. Is it above 1.0?
- Write one honest sentence comparing the two counties.
- Pick one local agency and draft a records request for one database, with dates, format and record layout named.
Check yourself
1. You want median household income for a county of 18,000 people. Which ACS product will have it?
The Census Bureau publishes 1-year estimates for areas of 65,000 or more, 1-year supplemental for 20,000 or more, and 5-year estimates for all areas.
2. An ACS table shows a county's median income as $58,412 with a margin of error of plus or minus $4,950. At what confidence level is that margin published?
The Census Bureau standard for ACS margins of error is the 90 percent confidence level. Multiply by 1.96/1.645 to get a 95 percent margin.
3. County A's estimate is $58,412 (MOE $4,950) and County B's is $62,100 (MOE $3,880). Using the Census Bureau test, what do you conclude?
Divide the $3,688 difference by the square root of 4,950 squared plus 3,880 squared (about 6,289). The result, about 0.59, is under 1.0, so the estimates are not statistically different.
4. Which agency would you go to first for annual retail electricity sales and customer counts by utility?
Form EIA-861 is an annual census of U.S. electric utilities reporting sales, revenue and customers.
5. Under 5 U.S.C. 552(a)(3)(B), a federal agency must give you a record in the format you ask for when...
The federal FOIA says an agency shall provide the record in any form or format requested if it is readily reproducible in that form or format.
6. Why ask for a record layout along with a database export?
A layout (also called a data dictionary or codebook) defines each field and code. Without it, a column called STAT_CD with values 1 to 7 is a riddle.
Terms in this lesson
- American Community Survey (ACS)
- The Census Bureau's ongoing yearly survey of a sample of households, the main source for income, commuting, housing costs and similar local figures.
- Margin of error (MOE)
- The plus-or-minus range around a survey estimate that shows how far sampling alone could push it from the true value, at a stated confidence level.
- Record layout
- A document listing each field in a database, its type and what its codes mean. Also called a data dictionary or codebook.
- Open data portal
- A government website where agencies post datasets for download, often as CSV files.
- Decennial census
- The full count of the U.S. population taken every ten years, with Census Day on April 1 of years ending in zero.
Sources
- U.S. Census Bureau, When to Use 1-year or 5-year Estimates checked 2026-10
- U.S. Census Bureau, Using American Community Survey Estimates and Margins of Error (training slides, 2017) checked 2026-10
- U.S. Census Bureau, Understanding and Using ACS Data: What All Data Users Need to Know (2020) checked 2026-10
- U.S. Census Bureau, About the Decennial Census checked 2026-10
- BLS, Local Area Unemployment Statistics checked 2026-10
- BEA, Personal Income by County, Metro, and Other Areas checked 2026-10
- EIA, Annual Electric Power Industry Report, Form EIA-861 checked 2026-10
- USDA NASS, Census of Agriculture checked 2026-10
- 5 U.S.C. 552 (Freedom of Information Act), Cornell LII checked 2026-10