Methodology

What we received, what we changed, and what we could not resolve.

The source

A copy of the Social Security Death Master File dated April 2012, distributed as three CSV files totalling about 7 GB. Together they hold 90,382,560 records, sorted by Social Security number, covering deaths reported through August 2011.

Each source row has eleven fields: Social Security number, last name, first name, middle name, a name suffix, date of birth, date of death, a verification flag, a state code, a last-residence ZIP code, and a second ZIP code used for lump-sum death payments.

What we dropped

The Social Security number. Removed entirely at ingest, for the reasons set out here. The first three digits are retained only as an integer, because they encode the issuing state and nothing else; the remaining six digits were never written to the database this site reads from.

The state code column. This one is more interesting, and we dropped it because we could not establish what it means.

The column holds two different kinds of value. In a 1.2-million-row sample drawn evenly from all three source files, 54% of populated rows carry a two-digit number and 45% carry a two-letter code (a further 290,000 rows in that sample are blank). We cross-tabulated the column against the state implied by each record's residence ZIP code:

code 30 → New Hampshire  99%      code 22 → Massachusetts  99%
code 20 → Maine          99%      code 07 → Connecticut    99%
code 10 → Florida        96%      code 33 → New York       96%
code 05 → California     98%      code 47 → Vermont        94%

code PE → no dominant state (51% / 9% / 8% ...)
code SI → no dominant state (51% / 11% / 5% ...)
code HC → no dominant state (58% / 8% / 6% ...)

The numeric codes are clearly a state-of-residence encoding and they agree with the ZIP code almost perfectly. The alphabetic codes correlate with nothing we could find — not residence, not the issuing state, not the verification flag. Rather than publish a field we cannot describe, we dropped it and derive residence from the ZIP code, which is unambiguous.

What we derived

Last known residence. From the residence ZIP code, joined against the GeoNames United States postal-code dataset (41,488 ZIPs with city, county and state). Two caveats. ZIP boundaries have shifted since these records were written, so a small number of old ZIPs will not resolve or will resolve to a neighbouring place. And this is the last address the Administration had on file, which is usually but not always where the person died.

Dates. Converted from the file's MMDDYYYY form to a sortable YYYYMMDD integer. A great many records give a month and year but a day of 00 — this is genuinely unknown in the source, and we display it as a month rather than inventing a day. Ages at death are computed only where both dates carry a real month, and are suppressed where the arithmetic gives an impossible result.

Errors we did not correct

This is a mirror, not an edit. Where the government file is wrong, this site is wrong in the same way, on purpose — silently "fixing" a public record makes it useless as evidence. Two things are worth knowing about before you trust a date.

1,756 records carry a date of death before 1936, and 349 of those are dated in the 1900s. Social Security did not exist until 1935. These dates cannot be right. They are almost certainly transcription errors for a later year.

121,321 records have no date of birth at all, and a great many more give only a month. A further 1,257 records imply an age at death of over 115, running up to a nominal 195 — the oldest reliably documented human lived to 122, so these are bad dates rather than remarkable lives. Where the arithmetic gives an age above 125 we show no age at all, but we do not alter the dates themselves.

Coverage is also extremely uneven. 89.6 million of the 90,382,560 records are deaths from 1962 onward; only 738,558 fall between 1936 and 1961. If someone died in 1950 and is not here, that is close to meaningless as evidence — the file barely covers that period.

The area-number table, which we had to rebuild

The first three digits of a Social Security number are the area number, and until 2011 they were assigned geographically. The Administration publishes a table mapping area numbers to states, and we started from it.

That table has a problem: it marks large spans — 237–246, 587–665, 667–679, 681–699, 750–772 — as "Not Issued." But this file contains 2,936,014 records whose numbers fall in exactly those spans. They were issued; the published table simply does not reflect the later reallocations.

So we reconstructed the missing rows empirically. For every area number we took the modal state of residence across all records carrying it. Migration blurs this, but not much: people overwhelmingly die in the state where they first worked.

First we validated the method against the spans SSA does document. Of 577 area numbers with at least 2,000 records and a published state, 571 matched the modal residence. All six exceptions are explanatory rather than contradictory:

580-584  SSA: Puerto Rico     modal residence: New York (28-39%)
586      SSA: Philippines     modal residence: California (50%)

Those are diaspora patterns — Puerto Ricans who moved to New York, Filipinos who moved to California — and they are precisely where a residence-based proxy should break down. The method fails where it ought to and holds everywhere else, which is about the best validation one can ask of it.

Applying it to the undocumented spans gives:

237-246  North Carolina  80%      646-647  Utah            79%
587-588  Mississippi     83%      648-649  New Mexico      76%
589-595  Florida         86%      650-653  Colorado        86%
596-599  Puerto Rico      —       654-658  South Carolina  91%
600-601  Arizona         78%      659-665  Louisiana       91%
602-626  California      91%      667-675  Georgia         89%
627-645  Texas           83%      676-679  Arkansas        96%
681-690  North Carolina  88%      750-751  Hawaii          75%
691-699  Virginia       100%      752-755  Mississippi    100%
764-765  Arizona         90%      756-763  Tennessee       80%
766-772  Florida         96%

The one judgement call is 596–599. It has no dominant state: the residences scatter across Florida, New York, New Jersey and Massachusetts in almost exactly the proportions seen in 580–584, the range SSA documents as Puerto Rico. We label it Puerto Rico on that basis, and flag it as inferred.

Two ranges are not geographic at all and we label them as such: 700–728 was issued by the Railroad Retirement Board to railroad workers nationwide until 1963, and 729–733 was Enumeration at Entry, for immigrants. Both show a flat spread across many states with no modal winner — around 10–20% — which is the signature of a non-geographic assignment, and a useful negative control on the whole method.

Everywhere this table is used on the site, empirically-derived values are marked as inferred and carry their confidence figure. You can read the whole table at /area-numbers.

Removals

Removal requests are handled with a read-layer suppression list rather than by deleting rows. A rebuild from the source data therefore cannot silently restore a record somebody asked us to take down.

Reproducing this

The derivations above are all arithmetic over a public file. If you want to check them, the JSON API is open and unmetered.