The impact of air pollution on career changes among Chinese workers

Baseline regression result

Table 2 shows the basic regression results for model (1). To control regional and temporal differences, all regressions include regional and temporal fixed effects, and the standard error is clustered to the county level in the basic regression (To exclude biased estimation results, we also cluster the standard error at the individual level as a robust check). To make the regression results more robust, the method of gradually adding control variables was adopted in the basic regression. Columns (1) to (4) in the table respectively show the results of no controlling variables, adding control variables at the individual level, adding regional climate variables, and further controlling regional socio-economic development variables.

The regression coefficient of PM_2.5 in Table 2 is significantly less than 0, indicating that PM_2.5 concentration has a significant negative impact on the career change of labor. The results of columns (2), (3), and (4) show that this negative effect is still significant after controlling for individual and regional control variables. According to the characteristics of the Probit model adopted in this paper, the marginal impact of PM_2.5 concentration on the probability of labor career switching can be calculated (The interpretation of the model results is different from that of OLS, for the detailed calculation method, please refer to the margins, dydx command of Stata16.0): It was found that the probability of the labor force’s career change would be reduced by about 0.8% points with the increase of PM_2.5 concentration by 1 µg/m³. Considering that the mean PM_2.5 concentration in the samples is 40 µg/m³ and the standard deviation is 11 µg/m³, the impact of air pollution on the labor career change is significant economically.

The research findings of this paper indicate that the relationship between age and the possibility of job switching presents an inverted U-shape, meaning that the impact of age on job switching varies at different age levels. Unhealthy lifestyle habits, such as smoking and drinking, are associated with a higher likelihood of job switching. Unhealthy lifestyle habits are strongly correlated with job instability, as expected. Furthermore, work experience reduces the possibility of job switching; people with more work experience have a lower likelihood of job switching. To further analyze the differences in job choices among different types of labor under the influence of air pollution, we will explore the heterogeneity from the perspectives of gender and income.

The regression results of this paper are consistent with the conclusion that air pollution hinders labor mobility in existing literature^10,12,19, that is, air pollution hinders free labor mobility.

Regression results of IV

In the benchmark regression model, due to the difficulty in controlling the impact of labor aggregation and industrial development on regional air pollution, the possible two-way causal relationship between labor aggregation and air pollution has not been observed, which may lead to biased estimation results of the basic regression model. In order to avoid biased results, we construct regression models of instrumental variables suitable for the background of this paper. Referring to the existing studies^8,44,52, we introduce instrumental variables by adopting the two-stage least square method (2SLS). To be specific, the thermal inversions are adopted as an instrumental variable of air pollution in both the linear probability model (LPM) and the binary selection model (IV Probit). The correlation between thermal inversions and PM_2.5 is shown in Fig. 2, and the estimated results are in Table 3.

Correlation between PM_2.5 and thermal inversions.

Figure 2 shows the correlation between thermal inversions and PM_2.5. Thermal inversions hinder vertical air circulation and pollutant dispersion, exacerbating air pollution. As shown, thermal inversions explain about 28.3% of the variation in PM_2.5 concentration.

Columns (1) and (2) in Table 3 (Starting from Table 3, to save space, the tables in the main text only present the estimation results of the main explanatory variables. The complete results including all control variables can be obtained by contacting the corresponding author) are regression results of the ordinary least square method (OLS), columns (3) and (4) are regression results of LPM, and columns (5) and (6) are regression results of IV Probit. Compared to columns (1), (3), and (5), columns (2), (4), and (6) have more control variables, including regional characteristics, individual characteristics, and weather characteristics variables. The results show that when PM_2.5 concentration increases by one unit, the probability of labor making a career change will decrease by about 0.8%, which is consistent with the results estimated by the Probit model in the basic regression. By comparing the regression results of OLS and IV, it can be found that the coefficient significance of PM_2.5 does not change after the thermal inversions are used as IV of PM_2.5, and the decrease in the numerical value is negligible. The F value of the first-stage regression is far greater than 10, indicating that the validity of IV meets the basic requirements. In general, the regression results of IV are consistent with the regression results of the basic model, that is, air pollution hinders the inter-occupational mobility of labor, and the probability of the career change will decrease by 0.8% for every 1 unit increase of PM_2.5 concentration.

Robustness test

For discrete choice models, most literature uses the Logit or Probit models, with only a few studies opting for the linear probability model. Both Logit and Probit models handle discrete dependent variables effectively and typically yield similar results. To test the robustness of our baseline regression, this section employs the Logit model for analysis. Furthermore, to ensure that the sample is more representative, here the age range is limited to between 16 and 60 years old, to examine whether the benchmark regression underestimates the negative impact of pollution.

Table 4 shows the results when the estimation model is changed to Logit in the first two columns, and the results when the sample is limited to those aged between 16 and 60 in the last two columns. By comparing with the benchmark regression results, it can be found that changing the estimation method and sample does not lead to significant changes in the benchmark results, indicating that the results of the benchmark model are robust.

To overcome the measuring error as much as possible, the question “Have you changed your job since 1997?” in the CHNS database questionnaire is adopted here. The answer to this survey question can be used as side evidence of career change, we use this answer as a new dependent variable (career change b), to further the analysis of the impact of pollution. Table 5 shows the robust results in the form of major variables, where columns (1) and (2) focus on the dependent variable. The results show a significant negative correlation between PM_2.5 and the new dependent variable, which is consistent with the conclusion obtained in the basic regression.

In columns (3) and (4), the calculation method of PM_2.5 is changed. Specifically, in the basic regression, this paper adopted the Bilinear Interpolation method commonly used in existing studies to calculate the mean value of PM_2.5 concentration in county areas. To test the dependence of regression results on this calculation method, the Nearest Neighbor method is adopted to calculate the average PM_2.5 concentration in the county area, namely pm2.5 b. Similarly, the regression results remain robust, and the size of the coefficient is consistent with the basic results, indicating that this result remains robust among different calculation methods.

Heterogeneity analysis

Individuals have great heterogeneity in their choice of residence and occupation, but labor groups determined by some common characteristics may have similarities in career choice. To analyze the response of labor to pollution in different groups, some pieces of literature discussed the relationship between labor mobility and pollution in different regions and types. As for individual characteristics, previous studies have found in their study that groups with higher age and education levels were more sensitive to air pollution when selecting employment sites^4,12,37. In terms of pollution tolerance, male, married, and child-bearing people are less sensitive to air pollution, and the non-agricultural population is more sensitive to pollution than the agricultural population. From the perspective of income level, some literature pointed out that there was an inverted U-shape, a non-linear relationship between haze pollution and labor mobility across regions, and wage level would weaken the impact of haze on labor mobility to a certain extent. Meanwhile, they found that the impact of smog on labor mobility across regions was particularly significant in developed cities^17,54.

Following the approach of the above-mentioned literature, we conducted heterogeneity analyses from the perspectives of gender, income, job attributes, and place of residence. Table 6 presents the results of the heterogeneity analysis. The first column indicates that there is no significant gender difference in the concept of job change in response to PM_2.5. The results in the third column further confirm this conclusion: the results of the interaction item (Work outdoors * PM_2.5) showed that the groups who often work outdoors (Drawing on the classification methods of existing literature and combining with the actual situation, primary technicians, unskilled workers, public security administrators, ordinary policemen and drivers are classified as those who often work outdoors) were less affected by pollution, that is, the reduction effect of air pollution on the change probability of outdoor workers was lower, and this result was significant. On average, outdoor workers were approximately 30% more likely to change occupations in response to pollution compared to non-outdoor workers. This could be attributed to their greater exposure to and perception of air pollution, leading to higher levels of health concerns and potential damage.

The results in the second and fourth columns of Table 6 indicate that rural residents and low-income individuals suffer more severe negative impacts from air pollution. That is, compared with urban residents and high-income individuals, rural residents and low-income individuals in areas with severe air pollution have lower mobility and more obvious welfare losses. The above results align with existing research on the differential welfare losses experienced by various groups in polluted environments⁵⁴. The likely reasons for this disparity are that low-income individuals and rural residents generally have lower income levels, poorer physical health, and fewer job opportunities compared to high-income individuals and urban residents. Consequently, these groups face more limited choices when confronted with the adverse effects of air pollution. This finding suggests that income level and physical condition may mediate the relationship between air pollution and occupational changes. The subsequent section will empirically test this hypothesis. Notably, these heterogeneous outcomes may exacerbate the existing disparities between urban and rural areas as well as among different income groups. Policymakers should therefore pay close attention to the inequality issues arising from air pollution.