Using Discrepancies to Correct for False Matches in Historical Linked Data

We propose a method to correct estimates from historical linked data for bias arising from type-I error--"false matches." We estimate the rate of false matching from the disagreement rate in characteristics that should agree across the two linked datasets. Combined with an understanding of the empirical patterns arising from false matches, knowledge of this rate enables us to correct for bias from false matches. Our method enables correction of estimates of both population moments and regression coefficients with valid inference. We illustrate the properties of our method via simulation and demonstrate them using linked US census data