This paper examines spatial and social correlates of work-from-home (WFH) frequency in three major global metropolitan areas—Paris, New York, and London—using harmonized 2023 survey data and an interpretable machine learning approach (XGBoost with SHAP). Focusing on regular hybrid workers whose workplaces are located within these metropolitan areas, the analysis documents how commuting, housing, employment, and sociodemographic characteristics are associated, individually and in combination, with high-frequency WFH (3–5 days per week). Results indicate that commuting distances and proximity to the central business district (CBD) are among the most prominent correlates of WFH frequency, with substantial contextual variation: in London, greater residential distance from the CBD is associated with higher WFH frequency, whereas in the New York metropolitan area high-frequency WFH is concentrated among workers who both live and work close to the CBD. Paris differs from the other metropolitan areas, with lower overall WFH frequency and a narrower profile of workers engaging in high-frequency WFH, a pattern observed alongside metropolitan-specific institutional and organizational contexts. Income, gender, and housing conditions are also associated with WFH frequency, with heterogeneous patterns across spatial configurations and employment structures. Overall, the analysis documents distinct metropolitan patterns, indicating that WFH frequency varies across locally specific urban labor markets and spatial contexts. These findings provide descriptive insights relevant to ongoing discussions of transport systems, housing markets, and social equity in metropolitan regions. • WFH frequency is embedded in metropolitan-specific labor and spatial structures. • Paris, New York, and London exhibit distinct spatial configurations of high-frequency WFH. • In the New York metropolitan area, frequent WFH is concentrated near the CBD. • In London, frequent WFH is more common among workers living farther from the CBD. • Interpretable machine learning reveals non-linear, context-dependent WFH dynamics with planning implications.
Motte-Baumvol et al. (Sat,) studied this question.