Affiliated organizations: operating entities and their fundraising arms

The single largest identified failure mode in the MAYBE tier is a candidate list holding both an operating organization and its own fundraising arm — EAGLE RANCH and EAGLE RANCH FOUNDATION, KANSAS FOOD BANK WAREHOUSE and KANSAS FOOD BANK FOUNDATION. Two real EINs, one body, and nothing in the name or the address separates them. Measured on 15 adjudicated cases the cascade ranks the correct member first 27% of the time — worse than chance, because the arm's name is usually the longer, more distinctive string and so scores better.

Details

This module treats the pair as one entity. It builds an EIN-to-EIN link table from three independent disclosures, groups co-located same-core candidates under a joint id, and orders the members by a stated policy rather than by a name score.

The three sources, and what each actually covers

Schedule A Part I line 12 (SA-P01-T01)

A 509(a)(3) supporting organization lists the organizations it supports, by name and EIN. This is the only source that states direction, and on 276 pairs it agrees with the name heuristic (the ... FOUNDATION row is the supporter) 271 times. It exists only for full-990 filers that are supporting organizations, and the supported-org EIN is blank on 49% of rows.

Schedule R Part II (SR-P02-T01)

Related tax-exempt organizations. Undirected in practice but far better populated — the related EIN is present on 97% of rows — and it covers pairs Schedule A never sees, because it does not require 509(a)(3) status.

BMF group exemption

Not the right tool for this relation. The affiliation_code and group_exemption_number fields describe group rulings — Lions Clubs, PTAs, BSA councils — i.e. federated siblings, not parent/arm pairs. 0000 means no group ruling and covers 1.52M of 1.92M BMF rows; treating it as a match key produces a spurious 74% link rate. Use it for the sibling case only, via pf_group_siblings().

On 927 name-detected pairs drawn from the MAYBE tier, Schedule A links 30%, Schedule R 46%, and either 49%. Restricted to pairs in the same city those rise to 39% / 60% / 64% — the monotone rise being the evidence that the geography gate removes false pairs rather than real ones. The residual is mostly non-filing: for 64% of unlinked pairs neither party files a schedule at all, which is what you would expect of public universities, school districts and 990-EZ filers.