Thread by @leah_boustan

I had a viral tweet a few weeks about following the sons of slaveholders into adulthood. Did you ever wonder how we could link 350k slaveholders across Censuses?

[A thread]

I have a new methods paper comparing linking samples created by automated algorithms and by hand.

Co-authors: Ran Abramitzky, @K_A_Eriksson, @jamesfeigenbaum, @seperez84

ranabr.people.stanford.edu/sites/g/files/…

Which automated algorithm should you choose? Or should you try to link by hand?

First, links made with standard algorithms agree more than 95% of time with “gold standard” hand links made by users of the genealogical Family Tree @FamilySearch

Second, linking face inherent trade off: minimization of “false positives” vs. creation of true links.

This graph compares algorithms to hand links from @earlyindicators – some algorithms disagree with high-quality genealogical links only 5% of the time!

Third, hand linking using **standard variables** (name, age, place of birth) performs no better than automated algorithms. Compared to algorithms, humans (or at least RAs!) are less willing to “leave a match on the table” in favor of matching to the wrong guy

Bottom line: Choice of linking algorithm has little effect on coefficient estimates

All in all, good news for matching algorithms — if you choose a conservative algorithm like ABE-5 years or EM, you can get false positive rates down to 5-10% 😄🎉

We provide code and Stata command for linking algorithms here:
ranabr.people.stanford.edu/matching-codes

No doubt a methods thread will NOT go viral... but this folks is how the sausage gets made [end]