Correcting Nonresponse Bias Using Panel Data on Data Requests and Responses

Abstract

When subjects who respond to requests for data, such as in surveys or post-treatment follow-up, are unrepresentative of the population of interest, inferences drawn from the data can be misleading. We show that if subjects' accumulated requests and responses over time are recorded and organized as panel data, requests can be used as instruments in standard nonresponse corrections. Randomizing total requests between subjects is insufficient for identification when responses depend on time and is unnecessary when they do not. We demonstrate our method by estimating an 18-percentage-point gender gap in entrepreneurial career intentions using a survey of undergraduates at the University of Wisconsin-Madison.