Imputing Establishment Robotics Data: I’m Afraid I Can’t Do That

This technical note describes the imputation of robotics questions added to the Annual Survey of Manufactures (ASM) for the 2018 and 2019 data collections. The relative rarity of robotics use among U.S. manufacturing plants presents particular challenges in addressing non-response. We focus our efforts on capital expenditures on robotics equipment as well as a broad indicator of a plant’s exposure to robotics, which combines information on the count of active and purchased robots as well as capital expenditures on robots. We describe, estimate, and compare a large number of imputation models ranging from linear and logistic regressions to random forests and neural nets. Our analyses investigate numerous ways of assessing imputation quality and explore the value of machine learning models not typically used to impute Census data. We find that simple linear and logistic regressions models with state, industry, size, and age controls perform better than the more complex machine learning models with and without feature selection models. The results also suggest that the establishment-level accuracy of imputed values is poor. Even within detailed industry and state cells it is very difficult to identify plants exposed to robots. Despite this, we find that state and industry-level tabular estimates are robust.