The COVID-19 pandemic highlighted the vulnerabilities of metropolitan areas (MAs), leading to increased outward migration globally. While economic factors have been studied, the empirical evidence regarding the impact of health risks and teleworking on leaving MAs (LMA migration) remains inadequate. This study addresses this gap using Japanese data, a unique case due to the absence of compulsory lockdowns. The research questions are: (1) Did health risks affect LMA migration, and were there temporal changes? (2) Did teleworking affect LMA migration, with temporal changes, occupational, and employment status variations? The study utilizes data from four waves of the Japanese government's Surveys on Changes in Attitudes and Behaviors in Daily Life under the Influence of Novel Coronavirus Infection (CABC surveys), employing entity- and time-fixed effects logit models to analyze the impact of infection rates and teleworking frequency on LMA migration decisions.

Existing research on LMA migration during the COVID-19 pandemic primarily focuses on changes in migration patterns and intensities, emphasizing economic factors like income, homeownership, and job opportunities. However, the impacts of health risks and teleworking are often discussed narratively, lacking empirical evidence. Studies inferring health-risk aversion rely on indirect indicators like population density, while conclusions about teleworking's role are often inconsistent and based on insufficient data, such as city or occupation-level data instead of individual-level. This study aims to address these limitations by providing robust empirical evidence using individual-level data and a more precise measure of health risks.

This study uses microdata from the third to sixth waves of the CABC surveys (April 2021 – March 2023), comprising an unbalanced panel dataset of over 37,000 observations. The dependent variable is LMA migration in the six months preceding each survey. The key independent variables are the logarithm of the COVID-19 infection rate in the individual's residence prefecture in the six months before the survey (COVID) and the frequency of teleworking (Telework). Control variables include socio-demographic factors (gender, age, marital status, education, employment status, occupation) and the regional unemployment rate. Entity- and time-fixed effects logit models are employed (Model 1), with separate estimations for each survey wave (Model 2) to examine temporal changes. Further models (Models 3 and 4) analyze moderating effects of employment status, and models (Models 5 and 6) analyze moderating effects of occupation on the relationship between telework and LMA migration. Odds ratios are reported, with values above one indicating a positive association and below one a negative association.

The study found a significantly negative association between COVID-19 infection rates and LMA migration (Table 3), confirming health-risk-aversion motives. This negative association persisted across the third to fifth survey waves (Table 4), but reversed in the sixth wave. This suggests a 'deferred decision' pattern, with the effect of lower infection rates persisting for several months after the end of public health emergencies, before reversing as the perceived pandemic threat diminished. Regarding teleworking (Table 3), the overall impact on LMA migration was insignificant. However, wave-specific analysis (Table 4) showed a significant positive association in the fourth wave only, suggesting a delayed effect. Analysis of moderating effects (Table 5) revealed that teleworking strengthens the negative relationship between formal employment and LMA migration, while strengthening the positive relationship between self-employment and LMA migration. This highlights the trade-off between economic opportunity costs and health risk aversion. Occupation-specific analysis (Table 6) showed that office workers and manufacturing professionals were less likely to migrate, but teleworking did not significantly moderate these relationships. Control variable analysis indicated that women, older individuals, those with higher household incomes, and the unemployed were more likely to engage in LMA migration during the pandemic; individuals with university education were also more likely to migrate, likely due to the combination of information access and safety concerns.

The findings strongly support health-risk-aversion as a primary driver of LMA migration during the COVID-19 pandemic, even in the absence of compulsory lockdowns. The 'deferred decision' pattern suggests that uncertainty surrounding the pandemic influenced migration timing, with health concerns initially outweighing other considerations. The differential impact of teleworking on formal and self-employed individuals underscores the trade-off between safety and economic factors. The results challenge the entrapment theory, demonstrating that resource constraints did not prevent migration for the unemployed. The absence of a significant overall teleworking effect on LMA migration suggests that the improvement of the teleworking environment alone is not sufficient to drive large-scale migration.

This study provides strong empirical evidence supporting health-risk-aversion as a primary driver of LMA migration during the pandemic and reveals the temporal dynamics of this behavior, especially the 'deferred decision' pattern. The results highlight the significant role of opportunity costs in shaping migration decisions, especially the difference between self-employed and employed individuals. The implications for policy suggest focusing on attracting self-employed individuals to local areas during crises and investing in long-term teleworking infrastructure improvements. Future research should incorporate living costs to provide a more comprehensive understanding of the drivers of LMA migration.

While this study provides compelling evidence, it has limitations. Although the infection rate is measured more accurately, it still reflects the situation before the decision. The study did not control for living costs due to data limitations, potentially obscuring the influence of this important factor. Future studies should address these limitations by incorporating more detailed housing market data and potentially exploring other factors such as changes in household composition and lifestyle preferences.