Validity Concerns of Using Machine Learning in Management Research

Abstract

The adoption of machine learning (ML) techniques in business research has significantly grown in recent years, with many studies using off-the-shelf ML algorithms to replace the labor-intensive coding of large unstructured datasets. Most published studies, however, failed to examine how errors in ML outputs affect the validity of the study results. This article shows, if unchecked, even a very small machine learning error (e.g., 5% for a binary output) could be amplified in subsequent statistical analysis to produce significant changes in research outcome, hence threatening its validity. After clearly differentiating two seemingly interchangeable concepts (i.e., ML error in the ML output vs. ML-induced bias in the empirical-analysis output), we develop a theoretical framework that demonstrates the possibility of having a large ML-induced bias despite of a small ML error. We discuss three common sources of the error-to-bias amplification stemming from the design of an ML algorithm. To mitigate the amplification, we introduce a novel correction process that not only significantly outperforms baseline approaches commonly adopted in the existing literature, but also can be readily adopted for a wide variety of ML algorithms and empirical studies. These advantages are demonstrated through a study that applies a state-of-the-art ML algorithm for facial recognition using real-world data.