Skip to contents

Collapse substitutable candidates before selection

Source: R/dyads.R
pf_dyad_collapse.Rd

For every grant whose tie band holds two or more candidates joined by a disclosed dyad, merges them into one and re-picks by config$policy instead of by score. Scores are not modified, so the YES / MAYBE / NO thresholds keep their meaning; what changes is which member of a merged pair is reported and how many distinct bodies remain in contention.

Usage

pf_dyad_collapse(
  x,
  dyads,
  config = pf_dyad_config(),
  np_config = NULL,
  rescore = TRUE
)

Arguments

x

A pf_match from pf_match_grants().

dyads

Output of pf_build_dyads().

config

Output of pf_dyad_config().

np_config

The npmatch config used to re-select. Defaults to the one stored on x, falling back to npmatch::np_config().

rescore

Re-run selection and tiering on the reduced pair set. See Collapse before selection.

Value

x with the selection recomputed, plus dyad_collapsed, dyad_id, dyad_class, dyad_alt_ein, dyad_from_ein, dyad_n_before and dyad_n_after recording every merge — including merges that left the pick unchanged, which are the majority and are still the point.

dyad_alt_ein and dyad_from_ein are not the same column and the difference matters. dyad_from_ein is the EIN that was reported and only fills in when the collapse changed the pick; dyad_alt_ein is the member merged away and fills in on every merge. A rollup keyed on the reported EIN needs the second one.

Details

Grouping is resolved within one grant's candidate list only. A dyad between two candidates merges them; a dyad from a candidate to some third EIN that is not a candidate here does nothing. That restriction is what keeps this from becoming the 67,000-EIN transitive closure described in pf_dyads.

Collapse before selection, not after

With rescore = TRUE (the default) the losing members of each merged group are removed from the scored pair set and the selection is re-run. That is the whole point: np_tier() demotes a YES to MAYBE whenever the margin to the runner-up is below min_margin, and when the runner-up is the same body that demotion is spurious. Rewriting the pick after the fact leaves the tier where it was and measures nothing. rescore = FALSE keeps the old post-hoc behaviour for comparison.