Matchmaking Gone Wrong: Quantifying Bias and Methods Using Non-Western Data

Alexander Persaud, University of Richmond

Despite the increased use of linked historical data, two key questions remain unresolved: do string-matching methods recover the correct underlying population? Can they be used for accurate inference? Economists have relied on hand-matching and various other techniques to attempt to recover a truth sample. I turn to an actual truth sample with proto-administative data on all 60,000-plus Indian indentured servants to Fiji from 1879-1916. I compare data on return migration to India and mortality between string-matching methods and the true, numeric links. String-matching leads to incorrect inferences about the composition of mortality and return migration. Furthermore, when compared to an economic exercise on the utility of return migration, the string-matched samples lead to nonsensical economic results. More generally, non-classical measurement error means that bias cannot be signed and all inferences using string matches may be suspect.

See extended abstract

 Presented in Session 182. Matching, Bias and Data Development: Automated Methods for Data Collection and Record Linking Assessed