Work From Home, Worry Less?

Abstract

Workers in occupations that cannot be performed remotely (the half of the global workforce in retail, hospitality, construction, healthcare delivery, and manufacturing) experienced sharper increases in self-reported anxiety during the first six months of the COVID-19 pandemic than workers in remote-capable jobs, even after conditioning on country-level case counts and policy stringency. The Open Science Framework COVID-19 survey, fielded across 177 countries from March 2020 onward, documents this gap at the individual level.

The behavioural-economics construct best placed to explain the residual is the **dual-process anxiety response to occupational uncertainty**: under high-uncertainty conditions, anxiety is generated less by realised income loss than by anticipated, hard-to-quantify catastrophic-outcome risk. Teleworkability, the structural property of a job that determines whether it can be performed from home, mechanically reduces the salience of one specific catastrophic-outcome category (workplace infection) and so should buffer anxiety differentially across occupations. The Dingel-Neiman occupational-teleworkability index provides a measurable exogenous job-property that varies across the respondent pool.

The existing analysis applies the classical identification stack: country fixed-effects + GMM dynamic-panel estimation, quantile regression (capturing distributional effects of the buffer), and Oaxaca-Blinder decomposition. Using 102,369 respondents (after listwise deletion from 113,083 raw) across 151 countries, merged with Johns Hopkins case counts and the Oxford Stringency Index, the estimates document a teleworkability effect on anxiety that survives the standard country-FE specification.

We propose extending the existing analysis with two ML layers: a Double-ML estimator that would partial out the 50+ country-level controls using LightGBM nuisance learners, producing standard errors valid under far weaker assumptions on the controls, and a Causal Forest CATE estimator that would recover the conditional teleworkability effect across sector × country-income-group strata. The proposed extensions would convert the population-average buffer estimate into occupational-targeting evidence for mental-health policy during future pandemic shocks.

Data & Methods