Using Deep and Active Learning Classifiers to Identify Congressional Delegation to Administrative Agencies

Abstract

Congressional oversight of the federal bureaucracy remains key to understanding implementation of major laws. Essential to this are theories of how and why Congress delegates powers to administrative agencies, which posit a complex relationship between individual members of Congress and the agencies they oversee. This field of study is rich in theory but has been lacking in systematic large-scale empirics. Given advances in the use of text-as-data methods, we can begin to test the institutional and partisan conditions under which Congress delegates authority to administrative agencies. Using a convolutional neural network on the text of bills, we classify bill sections by their role in delegating authority to administrative agencies. We introduce an active learning approach to text classification, applying an iteratively improving coding scheme that enhances existing supervised learning approaches. This method allows us to study the statutory scope of administrative agencies systematically and provides a first-of-its-kind dataset to study how administrative law develops.