Using machine learning to investigate the public's emotional responses to work from home during the COVID-19 pandemic

According to event system theory (EST; Morgeson et al., Academy of Management Review, 40, 2015, 515–537), the coronavirus disease 2019 (COVID-19) pandemic and resultant stay-at-home orders are novel, critical, and disruptive events at the environmental level that substantially changed people’s work, for example, where they work and how they interact with colleagues. Although many studies have examined events’ impact on features or behaviors, few studies have examined how events impact aggregate emotions and how these effects may unfold over time. Applying a state-of-the-art deep learning technique (i.e., the fine-tuned Bidirectional Encoder Representations from Transformers [BERT] algorithm), the current study extracted the public’s daily emotion associated with working from home (WFH) at the U.S. state level over four months (March 01, 2020–July 01, 2020) from 1.56 million tweets. We then applied discontinuous growth modeling (DGM) to investigate how COVID-19 and resultant stay-at-home orders changed the trajectories of the public’s emotions associated with WFH. Our results indicated that stay-at-home orders demonstrated both immediate (i.e., intercept change) and longitudinal (i.e., slope change) effects on the public’s emotion trajectories. Daily new COVID-19 case counts did not significantly change the emotion trajectories. We discuss theoretical implications for testing EST with the global pandemic and practical implications. We also make Python and R codes for fine-tuning BERT models and DGM analyses open source so that future researchers can adapt and apply the codes in their own studies.