Skip to main content
v2026.11,986 entries · CC-BY 4.0

American Community Survey PUMS: What It Is and How to Use the Microdata

What ACS PUMS microdata contain, which weight to use, and how to compute standard errors and margins of error with replicate weights, with a worked example.

Written and maintained by CASRAI Editorial Board

Published · Last updated

On this page

The American Community Survey Public Use Microdata Sample (PUMS) is the set of anonymized, record-level files from the U.S. Census Bureau’s American Community Survey: one record per person or per housing unit, released free of charge so that analysts can build custom estimates and tables that the pretabulated ACS products do not offer. A single-year PUMS file holds roughly one percent of the U.S. population and a five-year file roughly five percent, and the smallest area it identifies is the Public Use Microdata Area (PUMA).

This guide explains what the files contain, how the person and housing files fit together, which weight to use for which statistic, how to calculate a standard error and margin of error with the replicate weights (the worked example uses illustrative numbers), how to code the calculation in Excel, R and Python, and the mistakes that most often spoil a PUMS analysis. It sits in the survey research family of CASRAI guides.

What the ACS PUMS is

PUMS files contain the individual responses behind the published ACS tables. The Census Bureau describes them as files that let users “create custom estimates and tables, free of charge, that are not available through ACS pretabulated data products” (Census Bureau, ACS microdata page). Because every record is a single person or housing unit, an analyst can cross-tabulate any variables on the file, define a population the published tables never isolate, or fit a regression model, none of which is possible from published aggregate tables.

To protect respondent confidentiality the Census Bureau modifies the files before release. According to the 2020-2024 PUMS user guide, high incomes are replaced by a top-code value, uncommon birthplace and ancestry responses are grouped, income amounts are rounded, and the only sub-state geography identified is the PUMA. Top-coded and bottom-coded values are replaced, so the extreme tails of a distribution (the very highest incomes, for example) are not recoverable from the file. Census publishes the top-code and bottom-code thresholds in the documentation rather than as variables on the file.

Person file and housing unit file

PUMS is distributed as two file types: a Person file and a Housing Unit file. Persons are organised into households, group quarters residents are included on the person side, and the Housing Unit file also includes vacant units. Group quarters records on the housing file are placeholders that exist only for the food stamp (SNAP) variable and carry zero weight (2020-2024 PUMS User Guide).

The two files are linked by the household serial number, SERIALNO. The guide’s merge instruction is to sort both files on SERIALNO and then merge them, so that each person record picks up the characteristics of the household it belongs to. Analysts typically do this in R, SPSS, Stata or SAS because the files are large.

1-year versus 5-year PUMS files

Census publishes 1-year and 5-year PUMS files. The user guide states that a PUMS file for an individual year contains data on approximately one percent of the United States population, while a file covering five years contains approximately five percent. The trade-off that follows is the usual one for survey microdata: the larger 5-year sample supports finer cross-tabulations and smaller subgroups, while the 1-year file reflects a single year’s conditions.

Feature 1-year PUMS 5-year PUMS
Approximate share of the U.S. population About 1 percent About 5 percent
Period covered One calendar year Five consecutive years (the 2020-2024 file is the one the user guide documents)
Geographic detail Nation, regions, divisions, states, PUMAs Nation, regions, divisions, states, PUMAs

Source for the table: Census Bureau, ACS microdata page and the 2020-2024 PUMS User Guide.

PUMAs: the smallest geography on the file

A PUMA is a non-overlapping area that partitions each state, with roughly 100,000 people at the time it is created; Census redraws them after each decennial census. The files currently released use PUMAs defined after the 2020 Census, in use since the 2022 PUMS, while 2021 and earlier files use the 2010 definitions. A time series that runs across 2021 and 2022 therefore crosses a geography break, and PUMAs from the two vintages should not be treated as the same areas.

The PUMA code is five digits and is not unique nationwide. The user guide’s own example is code 00100, which exists in both Connecticut and North Dakota, so a state identifier (ST) must be combined with PUMA to identify an area. PUMS does not identify towns or counties, and Puerto Rico appears only as a state-level file, whereas the District of Columbia is in the national file.

Weights: PWGTP, WGTP and the replicate weights

A PUMS record represents many people, and every statistic must be weighted. Which weight to apply depends on the unit being counted. The data table below sets out the weight variables the user guide describes, and the rules of thumb follow it.

  • Use PWGTP for any statistic about persons, such as mean age or the number of people with a given characteristic.
  • Use WGTP for any statistic about housing units or households, such as the distribution of household income.
  • Use the 80 replicate weights (PWGTP1 to PWGTP80 for persons, WGTP1 to WGTP80 for households) only to estimate standard errors and margins of error, never for the point estimate itself.

Mixing them is a common error: a household-income estimate calculated on person records with PWGTP counts each household once per resident, which overweights large households.

Formula: standard errors with successive difference replication

The 2020-2024 PUMS User Guide computes PUMS variances with successive difference replication (SDR). The variance of an estimate x is

Var(x) = (4 / 80) × Σr=1..80 (xr − x)²

where x is the estimate computed with the full-sample weight (PWGTP or WGTP), xr is the same estimate computed with replicate weight r, the sum runs over all 80 replicates, 80 is the number of replicates, and the factor 4 removes bias from the replication design. The standard error is SE = √Var(x), and the 90 percent margin of error is MOE = 1.645 × SE, as the user guide defines it. Census also describes a generalized variance function method that uses design factors, but states that the 2024 PUMS will be the final year design factors are published and strongly encourages replicate weights instead.

Worked example with the arithmetic shown

The numbers below are illustrative, chosen to show the arithmetic; they are not drawn from a real PUMS tabulation. Suppose a weighted total, such as the number of persons in a state who hold a bachelor’s degree, is x = 52,000 using PWGTP. Repeat the calculation with each of PWGTP1 to PWGTP80, subtract x from each replicate estimate, square each difference and add the 80 squares. Suppose the sum is 3,200,000.

Step Calculation Result
Sum of squared differences Σ (xr − x)² 3,200,000
Variance (4 / 80) × 3,200,000 = 0.05 × 3,200,000 160,000
Standard error √160,000 400
90% margin of error 1.645 × 400 658
Coefficient of variation 400 / 52,000 about 0.8%

The estimate would therefore be reported as 52,000 with a 90 percent margin of error of about ±658, an interval of roughly 51,342 to 52,658.

Interpretation: reading the result

Quantity What it tells you
Point estimate (full-sample weight) The best estimate of the population quantity, not a count of sampled records
Standard error How much the estimate would vary across repeated samples of the same design
Margin of error (1.645 × SE) Half-width of the 90 percent confidence interval around the estimate
Replicate estimates Re-computations of the same statistic used only to measure variability

A small margin of error relative to the estimate indicates a precise figure; a margin that is a large share of the estimate signals that the subgroup is thinly sampled, which is when moving from the 1-year to the 5-year file or collapsing categories makes sense.

How to calculate it in Excel, R or Python

Excel

With person records in rows, age in column A, PWGTP in column B and replicate weight 1 in column C, the weighted mean age is =SUMPRODUCT(A2:A1000,B2:B1000)/SUM(B2:B1000). Repeat that formula for each replicate column using that column’s weights, place the 80 replicate means in a row, then compute =SQRT(4/80*SUMPRODUCT((replicate_range-full_estimate_cell)^2)). Large PUMS files exceed practical spreadsheet limits, so Excel suits only small extracts.

R

est  <- weighted.mean(p$AGEP, p$PWGTP)
reps <- sapply(1:80, function(r) weighted.mean(p$AGEP, p[[paste0("PWGTP", r)]]))
se   <- sqrt(4/80 * sum((reps - est)^2))
moe  <- 1.645 * se

Python (pandas and NumPy)

import numpy as np

est  = np.average(p["AGEP"], weights=p["PWGTP"])
reps = [np.average(p["AGEP"], weights=p[f"PWGTP{r}"]) for r in range(1, 81)]
se   = np.sqrt(4/80 * np.sum((np.array(reps) - est) ** 2))
moe  = 1.645 * se

To attach household variables to persons, merge on SERIALNO, for example p.merge(h, on="SERIALNO") in pandas, after reading the two files with SERIALNO kept as text so identifiers are not altered.

Assumptions and common mistakes

  • Unweighted counts. Counting records without weights describes the sample, not the population.
  • Wrong weight. Use PWGTP for persons and WGTP for households; replicate weights are for variance only.
  • Treating PUMA as unique. Pair ST with PUMA; code 00100 appears in more than one state.
  • Ignoring the 2021/2022 PUMA break. Files before 2022 use 2010 PUMAs; later files use 2020 PUMAs.
  • Dollar comparisons across years without adjustment. The files carry the inflation adjustment factors ADJINC and ADJHSG, and the Census Bureau’s online microdata access tool cannot inflation adjust.
  • Reading top-coded values as real maxima. High incomes are replaced by top-code values; consult the published top-code list before interpreting the upper tail.
  • Using generalized variance functions for new work. Census encourages replicate weights, and design factors end with the 2024 PUMS.

Because PUMS is a sample, no analysis of it measures the full population the way a complete count would; the broader trade-off is covered in sampling versus census. Related designs are explained in stratified sampling and in the guide to the standard error of the mean, and sensitive microdata governance is discussed under the Five Safes framework.

How to get the files

The Census Bureau offers three routes: the microdata access tool in data.census.gov (MDAT), bulk download from the Census FTP site at www2.census.gov/programs-surveys/acs/data/pums/, and the Census Microdata API. Data dictionaries for each release are available as PDF, TXT and CSV, and the Census handbook Understanding and Using the American Community Survey Public Use Microdata Sample Files (first issued February 2021, last revised 7 October 2024) covers the basics, PUMAs, access, file preparation, data quality and linking household members. The online tool is convenient for quick tabulations, but it cannot inflation adjust, so analyses that compare dollar values across years need the downloaded files.

Sources

Data: ACS PUMS weight variables and their use

Compilation of the weight variables on the American Community Survey PUMS files, the file each belongs to, what each is used for, and how many variables there are, taken from the Census Bureau 2020-2024 PUMS User Guide.

Period covered: 2020/2024. Coverage: United States.

ACS PUMS weight variables and their use
Weight variableFileUsed forNumber of variables (count)
PWGTPPersonPoint estimates for persons (for example age)1
WGTPHousing unitPoint estimates for housing units or households (for example household income)1
PWGTP1-PWGTP80PersonStandard errors and margins of error (replicate weights)80
WGTP1-WGTP80Housing unitStandard errors and margins of error (replicate weights)80

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · full access

Ask about American Community Survey PUMS: What It Is and How to Use the Microdata

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →