To ensure the quality of the published statistics at federal statistical agencies, it is common to conduct non-response bias analyses. While most of these analyses do the best they can to evaluate non-response bias, the methods are usually unable to capture the full effect of non-response. However, in this paper, I use administrative data to truly evaluate the effects of non-response in the Survey of Occupational Injuries and Illnesses (SOII) at the Bureau of Labor Statistics (BLS).
In 2017, the Occupational Safety and Health Administration (OSHA) began collecting some information identical to what is collected in SOII and has provided that information to the BLS. In some instances, establishments provided information to OSHA but did not respond to SOII despite being sampled. By linking the two datasets and comparing the data for SOII respondents to the non-respondents who reported to OSHA, we were able to more accurately evaluate the extent of non-response bias in SOII estimates. In this paper, I compare the results of the analysis to the results of a more traditional non-response bias analysis conducted on the same data.