Predictability Hypotheses

Abstract

Abstract Perhaps owing to decades of metatheoretical work done during a time of methodological and computational limitations, it is a widespread belief that machine learning and deductive hypothesis testing are fundamentally incompatible. This chapter argues that the more flexible estimators provided by machine learning are in fact invaluable tools for testing pre-specified social theories, especially sociological ones. This chapter defines a class of hypotheses called predictability hypotheses, or general theoretical statements that posit some unspecified (though potentially causal) relationship between two or more theoretical constructs. It then points to several examples of predictability hypotheses already being used by social scientists and describes how and when they are useful. It then considers some methodological implications for testing predictability hypotheses and concludes by making predictions about their future in sociology.