Assign stable cross-year person identifiers to a compensation panel

End-to-end pipeline: prepare the panel, link records across every pair of tax years within each organization, resolve the accepted links into person clusters (one record per organization-year, enforced), and stamp one deterministic EMP_ID on every record.

Usage

link_panel(
  df,
  cols = synthid_cols(),
  weights = default_weights(),
  threshold = 7,
  method = c("weighted", "em"),
  prob_threshold = 0.5,
  verbose = FALSE
)

Arguments

df

A stacked multi-year panel: the parsed output of titleclassifier and peopleparser. Must contain at least the organization id, tax year, and full name columns named by cols.

cols

Column mapping; see synthid_cols().

weights

Named base match weights; see default_weights().

threshold

Minimum match score to link two records when method = "weighted" (default 7, chosen by tuning against a labeled slice; see dev/tune_threshold.R).

method

Scoring method: "weighted" (default) uses the hand-set Fellegi-Sunter-style additive score; "em" fits an unsupervised Fellegi-Sunter latent-class model (fit_match_model()) to the candidate comparisons and links by the learned posterior match probability, also attaching a per-person link confidence.

prob_threshold

Minimum posterior match probability to link two records when method = "em" (default 0.5).

verbose

Print progress.

Value

df with added columns: EMP_ID (stable person id), EMP_N_RECORDS, EMP_N_YEARS, and – for method = "em" – EMP_LINK_CONF (the weakest posterior probability among the links holding the person's cluster together; NA for single-record persons), a ready-made confidence input for a downstream panel model. Diagnostics are attached as the "synthid" attribute (see synthid_report()).

Examples

if (FALSE) { # \dontrun{
panel <- readr::read_csv("PANEL-2019-2021.csv")
linked <- link_panel(panel)                 # hand-weighted
linked_em <- link_panel(panel, method = "em")
fs_weights(attr(linked_em, "synthid")$model) # learned weights
} # }