An independent educational project
When a number is not a name
A real linkage decision from the ongoing joint St. Croix project: keep plot-number evidence separate from name evidence.
https://danishwestindies.org/methods/linkage/
The problem in the working pipeline
In the 1792 General Tabell linkage routine, a row can contain a plot number where an estate name would normally appear. An earlier alias expansion could turn that number into candidate estate names. Treating those derived aliases as a written name would reuse the same evidence twice and overstate the strength of a match.
The implemented decision
The existing routine treats such a row as number-based evidence: it drops the aliases derived from that plot when constructing the name candidates and records the route as number, not name. It is not included in the single-estate, name-based subset returned by the routine’s name_matched filter. This is a record-to-estate join, not linkage between named individuals.
Related safeguards already in the code
- A near-spelling candidate is not accepted when another row in the same quarter writes that matched name exactly. The comparison is between names, not simply between merged estate units.
- Two linked rows in the same quarter with the same name key and positive recorded counts are marked ambiguous rather than treated as clean name matches.
- A row recording zero people is not treated as a competing positive-count holding by that ambiguity rule. This is not a general rule that missing observations equal zero.
How this account was checked
This note paraphrases the project’s existing tabell_links.py routine (row_links and name_matched) and its six regression tests. The tests cover number-only evidence, recovered names, near versus exact names, repeated names, the selected subset and different names within one merged estate. The inspection concerns the code’s decision rules, not a new audit of the underlying historical records.
The trade-off
Restricting the name-based subset avoids treating derived aliases as independent corroboration, but it can exclude genuine matches. Remaining in that subset does not establish historical identity or eliminate sample-selection problems. Unit tests check programmed behaviour; they do not measure real-world linkage precision or identify a causal effect.
My contribution and the publication boundary
I lead transcription, linkage and cleaning in the joint thesis and am the main contributor to empirical design and planter-model calibration. This description does not claim that I alone authored this particular routine. No private records, estate-level results, source-code excerpts or new personal data are released with this note.
Prepared with AI assistance from the working code and tests inspected on 24 October 2026. The underlying working inputs are not distributed here. The separate three-cell example demonstrates transcription and provenance only; it is not a worked example of this linkage rule.
Six existing regression tests passed a bounded run on 24 October 2026. This is not a test of historical linkage accuracy. Review record with code fingerprints (JSON).