Historical Identity Triangulation
Reconstruct missing demographic data by triangulating independent historical records
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 91%
Begin with the primary dataset and define exactly which identity field is absent. Generate possible proxies, but test each against records from the relevant historical period rather than assuming modern naming patterns apply. When proxies capture too few cases or generate false positives, expand the search to independent evidence such as censuses, professional directories, government surveys, articles, obituaries, and family accounts. Match each supported identity back to the original records while deliberately including obscure as well as famous subjects. Finally, investigate known omissions and seek external validation of the reconstructed classification. The mechanism converts many individually incomplete clues into a defensible linked dataset while making proxy failure, selection bias, and uncertainty visible.
Origin
Lisa Cook developed this approach while reconstructing a dataset of African American inventors because race was not recorded in historical US patent data.
Core principles
- 01Treat missing demographic fields as a linkage problem
- 02Test convenient proxies before trusting them
- 03Use multiple independent sources to reduce selection bias
- 04Search beyond famous or self-identified subjects
- 05Externally validate any newly constructed classification
How to run it
- 1
Define the Missing Link
Specify the demographic attribute absent from the primary dataset and the unit that must be matched across records.
Pro tip Write explicit inclusion and exclusion criteria before searching.
Watch out Do not infer identity from intuition alone.
- 2
Build Period-Specific Proxies
Use sources such as historical census records to develop candidate identifiers appropriate to the period being studied.
Pro tip Separate historical naming patterns from contemporary ones.
Watch out Modern proxies may not generalize backward in time.
- 3
Test and Reject Weak Proxies
Measure whether each proxy identifies enough valid subjects without sweeping unrelated people into the sample.
Pro tip Record failed approaches because they clarify the final method's boundaries.
Watch out A culturally plausible proxy can still have poor specificity.
- 4
Triangulate Independent Archives
Search directories, articles, official surveys, obituaries, and family descriptions for corroborating evidence.
Pro tip Use sources created for different purposes so their errors are less likely to coincide.
Watch out Famous-person collections can bias the sample toward exceptional cases.
- 5
Link Back to Primary Records
Match supported identities to names and entries in the original dataset, resolving ambiguous cases individually.
Pro tip Retain the evidence trail for every accepted match.
Watch out Name collisions can create false matches.
- 6
Fill Gaps and Validate
Investigate known omissions and ask independent experts or co-authors to validate the reconstructed list.
Pro tip Report unresolved cases rather than forcing classifications.
Watch out Completeness should never be claimed solely because additional searches stopped yielding results.
In the wild
Cook first tested historically distinctive names derived from census data, but they identified only a small number of patentees. Presidential names also failed because they were common among American men generally. She then combined a 1913 Patent Office survey with directories, articles, obituaries, biographies, and family descriptions before matching people back to patent records.
→ The process produced a broader historical list of African American inventors that could support patent-rate and violence research.
A hypothetical historian studying women in an occupation with no gender field could test period-specific names, link professional directories to census households, inspect obituaries and association records, and retain only identities supported by multiple sources.
→ The historian obtains a documented cohort with explicit uncertainty instead of an opaque name-based classification.
Common mistakes
Using Modern Proxies Historically
Contemporary demographic naming patterns may not describe people in an earlier period and can miss most of the target population.
Sampling Only Famous People
Biographies of prominent subjects create a dataset unlike the underlying population and distort subsequent conclusions.
Trusting One Archival Source
A single survey or directory may contain omissions that only cross-source triangulation exposes.
Is it for you?
Best for
It is best for researchers linking archival records across incomplete historical datasets.
Not ideal for
It is not ideal when lawful, complete, and directly observed demographic data already exist.
From the transcript
“I tried everything. I tried everything.”
“I wanted it to not be biased towards famous people, because I mean, the patent data set itself is not biased towards famous people.”
“Luckily, later on my co authors Tuvon Logan and John Parman helped to externally validate that list.”
From the episode
Lisa Cook on Gender and Race in Economics (Podcast)
Lisa Cook