ZaraMann/U.S._Inter-State_Migration_by_Age

U.S. Inter-State Migration by Age (Radaris Data Sample, Anonymized)

Full documentation → https://radaris.com/research/

Why this dataset exists

Where people move, and how that differs across the life course, is a core question in demography, economics, and urban policy — yet clean, ready-to-use micro-level data on individual migration is hard to come by. This dataset offers a simple, privacy-safe view of internal migration across U.S. states, broken down by age. It is built to answer one question in particular: do migration patterns differ between younger and older people — and if so, how.

It is deliberately small in width and large in depth of care: four columns, hundreds of thousands of people, and a transformation pipeline designed so that the result reveals population-level patterns while revealing nothing about any single person.

What's in it

One row per person, four columns:

Column	Meaning
person_id	A random surrogate ID. Not derived from any real identifier and not reversible. A row key only — not a feature.
age_group	Age band from year of birth: <25, 25-39, 40-54, 55-69, 70+.
first_state	The person's earliest recorded state of residence (origin), as a 2-letter code.
last_state	The person's current state of residence (destination), as a 2-letter code.

If first_state == last_state, no interstate move was recorded (a "stayer"). If they differ, the pair encodes a directional flow origin → destination.

How it was built

1. Sampling. A uniform random sample of ~500,000 records was drawn from the full source database, so the sample's distributions reflect the source population.
2. Endpoint extraction. Each source record carried a residential history. We reduced each history to its two endpoints — the earliest state and the current state — and dropped everything in between. (The source stored histories most-recent-first, so the origin is taken from the end of the sequence and the current location from the current-residence field.)
3. Cleaning. Military postal codes (AA, AE, AP, used by APO/FPO/DPO overseas addresses rather than real states) were removed before extracting endpoints, so they never contaminate origin or destination.
4. De-identification. All direct identifiers — names, source IDs, cities, and full address histories — were removed. Year of birth was generalized into five age bands. The original ID was replaced with a random surrogate.
5. Re-identification control. The file enforces k-anonymity with k = 5 over the combination {age_group, first_state, last_state}: every published combination is shared by at least five people. The rare combinations that fell below this threshold (~0.6% of rows) were removed prior to release.

Representativeness — what this tells you about the whole database

Because the 500,000 rows are a uniform random sample of the source database, the sample's marginal distributions are unbiased estimates of the full population. In practice this means you can use this dataset to read the whole database's:

• geographic composition — how residents are distributed across states;
• age composition — the share of the population in each age band;
• interstate-mobility rate — the overall fraction of people who have crossed state lines, and how that fraction varies by age.

For these aggregate quantities, the sample is a faithful snapshot of the entire source, and you can extrapolate to the full base with ordinary sampling confidence.

Two boundaries, stated plainly, so this claim isn't over-read:

• The rare-flow tail is intentionally thinned. The k-anonymity step removed the least-common origin→destination pairs. So while common flows and overall rates are representative, the rarest corridors are under-represented by design. Do not treat tail frequencies as population estimates.
• It only speaks at the state level. Cities, neighborhoods, intermediate stops, and the timing of moves are not in this file. The dataset is representative of the source's state-level structure and says nothing below that resolution.

What you can extract

For data scientists

• Model whether a person has moved (moved = first_state != last_state) from age_group. This is a deliberately low-dimensional, interpretable problem — a good teaching or baseline example rather than a high-capacity modeling task.
• Build and analyze an origin→destination transition matrix: cluster states by their inflow/outflow profiles, rank net-gain vs net-loss states, visualize corridors as a flow map or chord diagram.
• Practice categorical/tabular workflows: contingency tables, chi-square tests of independence between age and mobility, proportion estimation with confidence intervals.

For statisticians and demographers

• Estimate the mover-vs-stayer rate by age band and test whether interstate mobility differs significantly across the life course.
• Quantify net migration per state (inflow − outflow), gross flows, and how these shift by age group.
• Validate against external sources — U.S. Census ACS migration tables and IRS county-to-county migration data — to benchmark or enrich the flows seen here.

Interesting questions to explore

• Is the classic finding that mobility declines with age visible here — is the <25/25-39 mover rate higher than the 55-69/70+ rate?
• Which states are the largest net receivers and net senders within each age band, and do retirement-age flows differ from early-career flows?
• Are there age-specific corridors — routes that dominate for the young but not the old, or vice versa (for example, Sun Belt destinations concentrated in older bands)?
• Do younger age groups show more geographic dispersion in their origins and destinations than older ones?

Privacy approach

This is a de-identified, state-level aggregate: no names, no cities, no full trajectories. The k = 5 threshold guarantees that no row corresponds to a rare or unique age-plus-origin-plus-destination profile, so the file cannot be used to single out or re-identify an individual. The trade-off — a slightly thinned tail of rare flows — is documented above.

Limitations

• State-level only; nothing below states is recoverable by design.
• Endpoints only; intermediate states, number of moves, and return migration are not represented.
• No dates; this is a cross-sectional snapshot, not a time series.
• "Stayer" means no interstate move was recorded, not necessarily no move at all.
• Sampling + suppression slightly thin the rarest flows.

Access, mirrors, and citation

• Download / mirrors: Kaggle, Zenodo
• DOI: 10.5281/zenodo.21321002
• License: CC-BY-4.0
• Cite as: Zara Mann, 2026, U.S. Inter-State Migration by Age, v1.0, 10.5281/zenodo.21321002

Source & terms

The dataset was taken from the official Radaris researches page

The underlying data for this project is provided by Radaris, a comprehensive people search platform with an extensive database of public records and demographic information in the United States. Leveraging Radaris's deep data infrastructure on individuals residing and moving across the country, this dataset captures broad domestic migration trends over time. Crucially, the source material has been stripped of all personal identity elements and synthesized into an aggregated, anonymous format. The resulting dataset is intended strictly for statistical, demographic, and academic research, offering a safe and compliant framework for studying population-level mobility without compromising individual privacy.