Build a dictionary of resolved grant renderings

Turns matched grant rows into a reference the cascade can run against, keyed on the filer-written name and address rather than the BMF's. Use it with pf_recover_from_dictionary().

Usage

pf_grant_dictionary(
  grants,
  keep = "YES",
  exclude = NULL,
  key_col = NULL,
  stats = "dictionary",
  verbose = TRUE
)

Arguments

grants

Matched grant rows: the prepared recipient columns (recip_name, recip_street, recip_city, recip_state, recip_zip5) plus a tier and an EIN. A list carrying $grants is accepted, so the output of a run can be passed straight in.

keep

Tiers to build from. "YES" only, by default: a dictionary seeded from MAYBE would propagate exactly the decisions in doubt.

exclude

Optional grant keys to leave out — pass the keys of the rows you intend to match, or the dictionary answers them with themselves.

key_col

Column holding the grant key, for exclude.

stats

"dictionary" (default) computes the name and token statistics over the dictionary itself; supply a reference bundle to borrow its name_freq and token_idf instead. The IDF of ~90k filer renderings is not the IDF of 3.88M BMF names, and which one is right is an open question — the argument exists so the two are never confounded silently.

verbose

Print what was built and what was dropped.

Value

A reference bundle — reference, name_freq, token_idf, ref_index, normalizer — with attributes n_eins, n_renderings, ambiguous_dropped and built_from.

Why the address is not optional

Keying on the name alone does not work, and the margin is not close. Measured over 88,744 renderings: 727 of 63,310 bare name keys (1.15%) carry more than one EIN, and the worst is HABITAT FOR HUMANITY with 39. RONALD MCDONALD HOUSE has 31, BIG BROTHERS BIG SISTERS 29. Adding ZIP5 takes that to 0 of 66,425. Any (name_key, zip5) that still resolves to more than one EIN is dropped rather than guessed at, and the count is reported in the ambiguous_dropped attribute.

See also

pf_recover_from_dictionary()