Fit a match model to candidate comparison vectors

Two methods, both returning an object usable by predict_match() to emit a calibrated per-pair match probability (p_match), the confidence signal the downstream panel model wants:

"em" (default, unsupervised)

Fellegi-Sunter latent-class model. Each field is binarised (similarity >= agree_cutoff is agreement) and reclin2::problink_em() learns, by EM with no labels, the probability of agreement among true matches (m) and among non-matches (u) plus the prior match rate. Weights become learned rather than hand-set; the posterior is the naive-Bayes combination of the per-field log likelihood ratios.

"logistic" (supervised)

Logistic regression of the label on the graded (not binarised) similarities, so it uses the full resolution of the comparators. Missing fields are mean-imputed with a missingness indicator. Requires labels.

Usage

fit_match_model(
  comparisons,
  method = c("em", "logistic"),
  labels = NULL,
  agree_cutoff = 0.5
)

Arguments

comparisons

Output of candidate_comparisons().

method

"em" or "logistic".

labels

Integer/logical vector of length nrow(comparisons) (1/0, NA to ignore) — required for "logistic", ignored for "em".

agree_cutoff

Similarity at or above which a field counts as agreement (EM only).

Value

An object of class synthid_model.

Examples

if (FALSE) { # \dontrun{
cmp <- candidate_comparisons(panel)
em  <- fit_match_model(cmp)          # unsupervised
fs_weights(em)                       # learned agreement/disagreement weights
} # }