Corroborators: how much to trust a name someone else wrote

An observed name is one written by a third party: a grantmaker on Schedule I, or a related organization on Schedule R, writing a name next to an EIN. These aliases are the reason matchdb outperforms the BMF alone. Grant records use the names people actually write, and a reference that has seen those spellings matches them directly.

But one filer’s typo and a name fifty foundations wrote independently are both “observed names”. n_corroborators separates them: it counts the distinct filers who wrote that name for that EIN.

What a corroborator is

Schedule I rows are grouped by EIN and a light name key (uppercased, punctuation stripped, spaces collapsed). For each group matchdb stores:

• n_corroborators: distinct filing EINs that wrote it;
• n_obs: total times it was written, across all filings and years;
• first_year, last_year: when it was seen;
• name_raw: the most frequently written spelling.

A foundation that writes the same name in ten annual returns is one corroborator with ten observations. Ten foundations each writing it once are ten corroborators. Repetition by one filer is habit; agreement across filers is evidence.

How corroboration is distributed

Observed names by corroborator count, build of 2026-11-01:

distinct filers	names	mean times written
1	1,501,308	2.2
2	195,062	5.6
3	82,088	8.9
4	45,581	12.3
5 or more	140,770	65.0

Most observed names, 76%, were written by a single filer.

How these counts are made, and where each filer’s statement is kept, is in vignette("corroboration-ledger").

Bands, and choosing a threshold

md_spec(min_corroborators = ...) keeps observed names written by at least that many filers. It applies to observed names only; first-party names (BMF, the organization’s own filings) are always kept. The value must be a band edge, 1, 2, 5 or 20, because corpus statistics are precomputed per band and summed. A threshold between edges would cut across bands.

library(matchdb)
s <- md_spec(layers = "bmf_active",
             names = c("legal", "dba", "observed"),
             addresses = c("mailing", "observed"),
             min_corroborators = 2)

The threshold changes match results. On pfmatch’s 1k grant benchmark (dev/verify/phase2-measure.R), matched against active organizations, build of 2026-11-01:

reference	YES	MAYBE	NO
legal and DBA names only	2,945	945	853
plus observed names, 2+ filers	3,091	942	716
plus observed names, 1+ filer	2,422	1,699	648

With two or more filers, aliases add 146 YES (294 gained, 148 lost). Letting in single-filer aliases instead loses 523 YES: 389 gained, 912 lost. The likely cause is the name-frequency effect dev/PROCESS.md section 0c identified. Adding 1.5 million names, many of them noisy, makes common words look less distinctive and creates near-ties between candidates, so confident matches become MAYBEs.

Use min_corroborators = 2 unless you have measured otherwise on your own records. The published *-observed tiers use it.

What corroboration does not catch

Some errors repeat across filers because the same habit repeats:

• writing a placeholder EIN (00-0000000, 123456789);
• writing the filer’s own EIN for a grantee;
• writing a group-ruling central’s EIN for one of its congregations;
• writing a fiscal sponsor’s EIN for each project it sponsors.

Each of these can be corroborated many times and still be wrong. matchdb flags them rather than deleting them (md_guard_observed()). Each observed name carries flag_placeholder_ein, flag_placeholder_name, flag_own_ein, flag_group_central, flag_multi_name, and guard_any for any of the five. md_compile() always leaves guarded names out of a reference.

Corroboration and the guards answer different questions. Corroboration asks how many independent writers agree. The guards ask whether the EIN they wrote could be this organization at all.

Looking at the evidence for one organization

x <- md_lookup("53-0196605")
obs <- x$names[x$names$variant_class == "observed", ]
head(obs[order(-obs$n_corroborators),
         c("name_raw", "source", "n_corroborators", "n_obs", "first_year", "last_year")])

Observed addresses keep the name they were written with (observed_as in md_lookup()). A compiled reference therefore pairs each alias with its own address, rather than every alias with every address.