Air pollution and education investment

Our study examines the impact of air pollution on household investment in children's education. We use panel data from secondary schools in Shandong Province in China and find that a one-unit increase in PM2.5 concentration leads to a decrease in the probability of parental investment in their children's education by 3.9 percentage points. Factors such as educational expectations, financial considerations and personal well-being will mediate this impact. Additionally, our results suggest that education level and living area may moderate the relationship between air pollution and educational investment. These findings imply that air pollution may undermine human capital development due to alterations in parental investment behavior prompted by environmental conditions.


Introduction
Parental investment in household education is pivotal for their children's future development.Education serves as both a consumable and an investment, potentially yielding benefits and well-being for both children and parents.Commonly, it is the primary caregivers-parents-who invest in their children's education, not the children, who are the primary beneficiaries (Alderman and King, 1998). 1  According to research on child development, the resources that families allocate to education are critical for children's academic success.The family environment, being the first learning setting for a child, can have a more significant impact on educational outcomes than the quality of schooling or teacher experience.Characteristics established during childhood play a crucial role in determining future adult achievements (Francesconi and Heckman, 2016;Huggett et al., 2011).For example, studies have indicated that around 60% of the variations in personal income are attributable to individual characteristics before adulthood (i.e., initial conditions) (Cunha et al., 2005).
Extensive studies have investigated the link between air pollution and its detrimental physical health outcomes.It has been well-documented that particulate matter in the air significantly affects human health, leading to respiratory and metabolic diseases (Chay and Greenstone, 2003;Deryugina et al., 2019).Beyond physical health, air pollution is also known to influence other personal well-being outcomes.For example, polluted air can diminish academic success, productivity and earnings (Borgschulte et al., 2022;Buoli et al., 2018;Chang et al., 2019;Chen et al., 2018a;Graff Zivin and Neidell, 2012;Hanna and Oliva, 2015).Notably, the impact of air pollution on children's learning can have prolonged consequences for their income in adulthood (Ebenstein et al., 2016).In summary, air pollution poses a profound challenge to the sustainable advancement of society and the development of human capital.
Pessimism and other negative emotions triggered by air pollution may deter individual investment in stocks, which in turn could affect stock market returns (Liu et al., 2021).Additionally, air pollution may erode personal confidence in business (Guo et al., 2022).As educational investment choices might also be affected by emotions, our study investigates whether air pollution affects education investment.Education investment is associated with human capital accumulation in individuals before entering the workforce, which can significantly impact future job prospects and lead to differences between socioeconomic and ethnic groups (Cheadle, 2008).Although previous research has focused on air pollution's direct effects on children's cognitive abilities, the effects on parental decision-making regarding education investments remain under-examined (Graff Zivin et al., 2020).
We use the case of China for two primary reasons.Firstly, the Chinese central government launched a rapid expansion program for the higher and further education sector in 1999 to stimulate domestic demand, boost economic development and alleviate employment pressure (Yao, 2019), leading to a significant rise in university enrollments.In response, Chinese parents, recognizing the value of education in a competitive environment, began to invest heavily in their children's learning.This is reflected in their commitment to educational expenses, with the goal of securing their children's success in competitive examinations and entry into top-tier schools, thereby improving their prospects in the labor, housing and marriage markets (Wang et al., 2022a).This heightened focus on education is further evidenced by the approximately 30% growth in China's total education spending from 2010 to 2018 (Wang and Cheng, 2021).Secondly, air pollution levels in China vary significantly across regions, and the air pollution problem has remained severe in recent years.While the implementation of environmental policies has led to a decrease in air pollution levels in some parts of the country, the annual average concentrations of ambient particulate matter ≤2.5 μm in diameter (PM2.5) and ambient particulate matter ≤10 μm in diameter (PM10) in many major cities still surpass China's secondary standards for ambient air quality (i.e., 35 μg/m 3 for annual mean PM2.5 and 70 μg/m 3 for annual mean PM10).For instance, Fig. 1 depicts the average PM2.5 concentrations and relevant meteorological conditions, e.g., thermal inversions, for our sample from Shandong province, with red vertical lines marking the study period.There are significant variations in air pollution and meteorological conditions in China.Notably, cities like Liaocheng and Heze exhibit high pollution levels, with average PM2.5 concentrations above 50 μg/m 3 in 2017.This concentration is twice as high as in low polluted cities, such as Weihai.Additionally, the lack of public participation in social governance and the increase in average annual concentrations of other types of pollutants, such as ozone, make pollution issues ongoing challenges (Huang et al., 2018;Li et al., 2018;Zhu et al., 2022).
We employ an instrumental variable, leveraging the predicted levels of air pollution driven by thermal inversions and wind, to assess the causal impact of air pollution on parental educational investment among secondary school students.To do so, we match the 2017 and 2020 Database of Youth Health (DYH) collected in Shandong province with three datasets: (1) high-resolution satellite air pollution data collected by the National Aeronautics and Space Administration (NASA) in the United States; (2) NASA weather data; and (3) city characteristics from the China City Statistical Yearbooks and the official documents of city government agencies.
Our research indicates that PM2.5 exposure can decrease educational investment behavior, and this finding remained robust after conducting additional tests.Moreover, air pollution impacts educational investment via education expectations, financial considerations and personal well-being.We also find that air pollution's negative effect on education investment correlates with children's education level and living area.
We acknowledge the relevance of a prior study on the association between air pollution and children's earnings in adulthood: namely, Isen et al. (2017), who compared birth cohorts in counties with varying levels of air pollution before and after the implementation of the Clean Air Act in 1970 in the United States.Their findings showed that childhood exposure to air pollution is linked to low labor force participation rates and income levels at age 30.Their study also suggests that air pollution has a significant impact on disadvantaged populations who live in high polluted areas.It is possible that air pollution affects children in the long-term through health capital, such as chronic diseases and weight, as well as cognitive and non-cognitive skills, which can be reflected in exam scores or other personality traits.However, their study does not explore the impact of air pollution on education investment and only briefly mentions the potential reinforcing or compensatory role of parental education investment in children's human capital formation.
Our study significantly contributes to the literature on environmental pollution and human capital investment in several ways.Firstly, we measure the toll of air pollution from a new perspective.Secondly, we adopt parental behaviors to provide additional explanations for the negative impact of air pollution on children's academic performance, as opposed to the previously studied mechanisms of children's physical health, mental health and school absences (Isen et al., 2017;Zhang et al., 2018).This work also contributes to understanding the role of environmental factors in the intergenerational transmission of socioeconomic status. 2 Our findings suggest that air pollution may perpetuate socioeconomic disparities by influencing parental investment in their children's education-a key factor in the accumulation of human capital.Lastly, we explore how air pollution affects parental education investment.Our study identifies income factors, beyond physiological and psychological attributes, that may contribute to air pollution-related investment behaviors (An et al., 2018;Levy and Yagil, 2011).
Our research has crucial implications for future research directions and public policy.One of the main implications is that air pollution is a critical environmental issue that requires immediate attention.Recent findings indicate that pollution was responsible for about nine million premature deaths in both 2015 and 2019, making it the largest environmental risk factor for diseases and premature deaths, with one in six deaths worldwide attributed to it.Despite some reductions in recent years due to poverty-related pollution, such as indoor household and water pollution, deaths from outdoor ambient air pollution and toxic chemical pollution (e.g., lead) have increased (WHO, 2022a). 3Our research suggests that the current measures of the social cost of pollution are insufficient and underestimate the impact of air pollution on individual behavioral choices, such as parents' investment in their children's human capital.Moreover, while low-and middle-income nations bear the brunt of air pollution's health consequences, the problem transcends national borders.The need to address air pollution's effects on both health and the development of human capital calls for a concerted global effort.Additionally, the link between air pollution and climate change is undeniable.For instance, the combustion of fuels produces short-lived climate pollutants, including methane (i.e., a precursor to ozone) and black carbon (i.e., a component of PM2.5).Therefore, targeted air quality policies and regulations are likely to yield extensive benefits, enhancing quality of life and proving cost-effective in the long term (Fuller et al., 2022;WHO, 2022b).

Literature review
The intergenerational effects of air pollution on human capital development are not yet fully understood.Studies in psychology and behavioral economics suggest that air pollution can lead to negative emotions such as anxiety, sadness, depression and risk aversion, which in turn may alter investment judgments and psychological expectations (Dong et al., 2020;Zhang et al., 2017).Such emotional responses can influence key life choices.For instance, increased caution and prudence stemming from air pollution might make individuals, like stock market investors and auditors, increase risk-averse attitude in their investment or auditing decisions.Thus, mood deterioration will affect people's decisions involving risk, such as education investment, through distortion of their probability weighting functions (i.e., increase in the recall of emotionally congruent memories, leading to high perceived probabilities of future negative or positive events in low or high mood conditions, respectively).Specifically, rises in cortisol levels due to air pollution exposure may steer individuals away from sensation-seeking behaviors (i.e., sensation-seeking is defined as pursuing and taking risks to experience new sensations, such as setting high education expectations) (Chen et al., 2020;Lepori, 2016).Moreover, as an environmental stressor, air pollution adversely affects cognitive functions, including attention, responsiveness and information processing.This can result in workplace biases and decreased productivity (Chang et al., 2019;Graff Zivin and Neidell, 2012;Grainger and Schreiber, 2019).Such negative health consequences of air pollution may also impact individuals' educational choices, as optimal decision-making requires good health and cognitive ability (Huang et al., 2020;Liu et al., 2021;Tan and Yan, 2021).
While air pollution may heighten risk aversion and thus diminish high-risk investments, the opportunity cost of air pollution-induced misperceptions and poor decision-making may be large, especially considering the importance of parental investment in their children's education.In many developing countries, educating children is often viewed as a parental responsibility, and investment in education is believed to have a crucial impact on the accumulation of children's future human capital.Previous research on migrant children in China has shown that parents' educational expectations highly correlate with their children's future educational attainment.In addition to the effects of community and school conditions, educational investment plays a significant role in this relationship (Luo et al., 2022).Our study supports earlier research on the detrimental impacts of air pollution, which have been linked to decreased psychological well-being and impaired decision-making.These impacts are evident in choices related to educational investment, highlighting the significant role of air pollution in the existence of intergenerational socioeconomic disparities.
Income factors may also influence the impact of air pollution on household investment in education.Existing research has linked air pollution to various economic impacts.Firstly, the adverse health issues related to air pollution often result in few working hours and subsequent loss of income and career opportunities (Liu et al., 2020).Exposure to air pollution and its impact on children's health (Edwards and Langpap, 2012;Jans et al., 2018) can also reduce parents' working hours and income (Currie, 2009).Secondly, reduced cognitive abilities that are essential for tasks (e.g., concentration and critical thinking) caused by air pollution could harm work productivity, reducing labor market performance and earnings linked to performance (Chang et al., 2019;Graff Zivin and Neidell, 2012).While short-term economic gains from increased pollution are possible due to the production effects (Wang et al., 2022b), the damage to health and income is substantial (Liu et al., 2020).Finally, it is worth noting that air pollution can also exacerbate poverty and inequality, with differing levels of exposure across income groups.In developed countries, air pollution levels may be particularly high in low-income communities (Finkelstein et al., 2003;Jbaily et al., 2022).In China, while there may be a positive relationship between air pollution and the socioeconomic status of residents due to the growth of urbanization and inadequate environmental regulation, individuals L. Guo et al. with high socioeconomic status tend to take protective measures to mitigate the adverse effects of air pollution.These measures may include living in areas farther away from polluting industries, utilizing their political influence to improve residential infrastructure and increase the implementation of pollution control measures, owning vehicles to minimize exposure during their daily commute, and using their financial resources to exclude low-income individuals from residing in cleaner communities.This inequality is also reflected in the different living environments between people with and without local hukou.For example, in Jiangsu Province in China, despite adjusting for the fact that rural migrants without local hukou typically work in industries with high levels of pollution, the areas where they reside still experience high levels of air and water pollution (Schoolman and Ma, 2012).The inequality between individuals with different socioeconomic statuses is further exacerbated by the fact that rural migrants lack the same rights as urban residents.Furthermore, the role of family income is critical in educational investments, as high household income tends to align with increased college enrollment (Acemoglu and Pischke, 2001).If air pollution negatively impacts parental income through reduced productivity and increased healthcare expenses, this could impose financial limitations, thereby hindering investments in children's education.
In addition, the relationship between air pollution and educational investment may be influenced by individual demographic factors.Children in high grade levels may require large educational investment due to an increased number of subjects and complex learning content.At the same time, in China, compulsory education only covers middle school education.Since the government subsidizes tuition fees for middle school education in most schools, education for high school students may require a greater financial investment than for middle school students (Xue and Zhou, 2018).Moreover, urban students, exposed to higher air pollution levels, may also be impacted differently compared to their rural counterparts.Rural educational investments may be little affected by air pollution, given the lack of educational resources and the prevalence of paternal labor migration (Shen et al., 2021;Zhao et al., 2021).Furthermore, studies have shown that in some underdeveloped regions, girls may have few opportunities to receive an education due to traditional social norms on family responsibilities and gender roles, as well as gender disparities in labor force participation and wages (Alderman and King, 1998).During the 1990s, girls in rural and low-income families frequently competed with siblings for limited educational resources, contributing to elevated dropout rates of these girls.However, recent shifts in societal attitudes toward gender equality have led to a reduction in this educational investment gap between genders (Hannum et al., 2009;Lee, 2012). 4

Data
The final combined dataset includes information on secondary school students, air pollution, thermal inversions, weather and city characteristics in 2017 and 2020 in Shandong province, China.The DYH survey was conducted by Shandong University among secondary school students (years 7-12).This dataset is the first large-scale longitudinal survey that is shared regarding the health of Chinese adolescents, which includes detailed geographical and educational background data. 5The survey randomly selected schools using a probability-proportional-tosize sampling and then randomly surveyed students in each selected school (Shandong University, 2021;Zhang et al., 2022).
The dataset allows for an analysis of education investment behaviors and related information of parents and children.Education investment behaviors are defined by whether the student attends an after-school tutorial class (yes = 1).Receiving tutoring in addition to regular school instruction (i.e., shadow education) is prevalent in many developing countries due to the existence of high-stakes examinations and low-educated parents.It supplements human capital formation outside the school (Jayachandran, 2014;Pan et al., 2022).We construct an adjusted education investment variable by aggregating the values from whether the student attends an after-school tutorial class (yes = 1) and whether the student has many books (yes = 1), the latter serving as an indicator of a family's cultural resources which supports children's reading motivation and learning stimulation (Conger and Donnellan, 2007;Heppt et al., 2022).
Data on surface-level air pollution is taken from NASA, which might provide more accurate and reliable data than data from Chinese official monitors (e.g., official air pollution data might have potential manipulation problems due to promotion incentives of government officers, Qin and Zhu, 2018).The satellite-based readings of NASA PM2.5 data with a spatial resolution of 0.1 • × 0.1 • (or roughly 11 km × 11 km) are gathered from Global/Regional Estimates (V5.GL.02) from Washington University in St. Louis. 6We utilize the average air pollution over the entire calendar year.
Information on thermal inversions and weather is also obtained and calculated from NASA in order to maintain consistency of environmental data sources.The former is calculated based on an instantaneous 3dimensional 6-hourly dataset, and the latter is obtained from postprocessed 3-hourly data, which has corrected previous grid box issues, was updated in 2020. 7The thermal inversion phenomenon is common in many regions of the world and can lead to pollutants being rapidly trapped near the ground.Thermal inversions, prevalent in Shandong province, exhibit significant temporal and spatial variations, especially in the northern and eastern regions of the Shandong Peninsula (Hao et al., 2010).Following Chen et al. (2022), we conduct a test to show the linear fit of the correlation between changes in thermal inversion occurrences and changes in air pollution.As illustrated in Fig. 2, changes in thermal inversion in Shandong Province might not be correlated to changes in air pollution during the study period.This inconsistency underscores the exogeneity of thermal inversions, suggesting their usefulness in addressing the endogeneity issues associated with air pollution (Arceo et al., 2016;Deschenes et al., 2020).Although people's health and behavior are not directly affected by inversions without pollutants; inversions might be associated with ground-level weather patterns.Therefore, we control for ground-level weather conditions in main regressions (Chen et al., 2022). 8Considering wind might influence the transmission of air pollutants, we add wind direction and wind speed in the construction of the instrumental variable.While wind direction could serve as a viable instrumental variable on its own, we choose not to depend exclusively on it due to uncertainties in air pollutant 4 According to the 2019 report on Chinese Women's Development Outline (2011− 20) from the Chinese National Bureau of Statistics, increasing females in higher education increase the number of women in high-level jobs with high salaries (China Daily, 2020). 5The geographical information (e.g., city names) of China Education Panel Survey is not available, and education information (e.g., grade of secondary school students) of China Family Panel Studies is not sufficient. 6The data can be accessed from the website: https://sites.wustl.edu/acag/datasets/surface-pm2-5/. A higher resolution will minimize measurement errors and is a better fit for city location information with a spatial resolution of 0.1 • × 0.1 • .Additionally, it helps ensure we do not inadvertently use characteristics from nearby cities when our focus is on cities with smaller land areas.
For instance, Weihai City in Shandong Province covers just 5956 km 2 , and its city center is about 66 km from the center of the adjacent Yantai City.
8 Previous studies also takes into account weather factors, like humidity and precipitation, which have a correlation with air pollution (Deschenes et al., 2020;Chen et al., 2018b).
We also collect city-level attribute data from China City Statistical Yearbook and local government websites.We merge the survey data with the city characteristics data by city and year.We assign the air pollution, thermal inversions and weather data with the closest latitude and longitude to the given city.
The combined dataset consists of 32,094 observations in ten cities in Shandong province in 2017 and 2020.There are 19,808 observations in middle school and 12,286 observations in secondary school; 48.8% are females and 51.2% are males.Panel A of Table 1 displays summary statistics for the full sample.We observe that 58.3% of observations receive parental education investment.Summary statistics and Welch's t-statistic for sub-group samples across education levels and gender are presented in Panel B and Panel C of Table 1.Results indicate significant differences between middle and high school students and between females and males: middle school students and males tend to have less education investment and experience higher air pollution levels compared to high school students and females. 9

Model and identification
We use fixed effects models and an instrumental variable strategy to overcome potential endogeneity issues for inferring the causal relationship between air pollution and education investment.Endogeneity concerns arise where education investment might be low in impoverished areas with high air pollution, and the impact of air pollution might be overstated (Bondy et al., 2020).We construct the following models to identify the link between air pollution and education investment: (1) where I i,j,t is the education investment of individual i in city j in year t and P j,t is PM2.5 concentrations in city j in year t.PM2.5 is instrumented by the Z j,t -that is, the predicted value of PM2.5, to address the endogeneity problem of air pollution.Recent research employs exogenous pollution from ventilation as an instrument for pollution (He et al.,   9 We drop observations with missing information on personal characteristics such as hukou status and whether the student belongs to key class.2019; Liu and Salvo, 2018).Following these studies, our predicted PM2.5 is based on the values of thermal inversion, wind speed, wind direction and the number of occurrences of thermal inversions. 10X i,j,t is a vector of individual-and city-level covariates.γ j and δ t represent city fixed effects and year fixed effects, respectively.ε i,j,t and u i,j,t are standard errors. 11Education investment and its related behaviors might be correlated with time-varying and time-invariant characteristics.We control these covariates in order to obtain correct estimates.We control for personal characteristics such as age, age squared, whether male (yes = 1), whether han nationality (yes = 1), hukou status (others = 1, rural = 2, urban = 3), whether live in urban area (yes = 1), whether migrant (yes = 1), education level of children (middle school = 0, high school = 1), whether key class (yes = 1), whether leader (yes = 1), whether one-child family (yes = 1), parental education level (i.e., the highest level of education attended by parents, below university = 0, university or above = 1).For city-level attributes, we include real gross domestic product per capita (thousand yuan), population density (per km 2 ), ground-level temperature ( • C), rainfall (10 − 5 kg/m 2 *s), relative humidity (%), number of industrial firms per capita, thermal power plants' effects (i.e., the distance-weighted electric energy production of thermal power plants within 200 km) and forests effects (i.e., distance-weighted areas of forests within 200 km).
We allow for city fixed effects to account for any confounding from time-invariant structural differences between cities.We include year fixed effects to control for common trends across years and absorb common year-specific shocks (Grossmann et al., 2021).Also, in our model, the error term is clustered at the city to address the concern that standard errors among students within the same city are positively correlated (i.e., unobserved within-city correlations arising) and might result in false significance.
Moreover, we conduct a three-step approach to test the mediator roles of education expectations, financial considerations and personal well-being (Baron and Kenny, 1986) and utilize interactions to measure the moderator roles of demographic characteristics.Mediators contain parental education expectation (i.e., expected education levels of parents, current or no expectation = 1, high school = 2, college education = 3, bachelor = 4, master = 5, doctor =6), child education expectation (i.e., expected education levels of children, current or no expectation = 1, high school = 2, college education = 3, bachelor = 4, master = 5, doctor =6), parental study abroad expectation (i.e., whether expect to study abroad, yes = 1) and child study abroad expectation (i.e., whether expect to study abroad, yes = 1), household income level (from 1 to 5, the higher the wealthier), health expenditure status (i.e., whether has high health expenditure, yes = 1), 12 memory (i.e., satisfaction with memory, from 1 to 4, with higher values indicating greater satisfaction), physical health (i.e., satisfaction about ability of participating physical activities, from 1 to 4, with higher values indicating greater satisfaction) and self-worth (i.e., self-perceived importance in a group, from 1 to 4, with higher values indicating greater importance).Moderators include education level of children (middle school = 0, high school = 1),  (Liu and Salvo, 2018).However, since our yearly data differs from daily data, we have included the number of inversion occurrences in our prediction model. 11We do not include individual fixed effects because our interest lies in estimating the impact of time-invariant attributes for individuals, such as gender.Moreover, including individual fixed effects could lead to unavoidable biases in parameter estimates when the number of periods is small (i.e., 2-year data in our study) (Allison, 2002). 12We classify individuals with high health expenditure as those exhibiting severe physical symptoms such as headaches, chest pain, back pain, stomach discomfort, muscle aches, breathing difficulties, chills and fever, body numbness and tingling and physical powerlessness.
whether live in urban area (yes = 1) and whether male (yes = 1).In addition, we perform several robustness checks to confirm that the results are not confounded by other variables related to education investment.

Results
We first present the coefficient estimates of education investment models.Next, we explore the mechanisms through which air pollution affects education investment and measure the moderating effects.Lastly, we show a series of robustness tests.All models include controls, city fixed effects and year fixed effects.
Table 2 displays the impacts of air pollution on two types of education investment using the OLS method.The coefficients for air pollution are negative.While the impact of air pollution on education investment is insignificant, it becomes significant at the 10% level in the adjusted education investment regression.This suggests that air pollution might negatively impact education investment.However, we argue that assessing the impact of air pollution using the OLS method can lead to endogeneity bias and inaccurate conclusions.To address this issue, the 2SLS technique is employed to correct for endogeneity in the model.
The models in Table 3 adopt first-step estimation of the Two-Stage least squares (2SLS) method.The fitted values from these first-stage regressions will be used in the second-stage regression to assess the impact of air pollution on education investment, by replacing the air pollution variable.Based on the results of the first-stage regressions, we find that the impact of exogenous air pollution on the air pollution variable is positive and significant at the 1% level.This indicates that exogenous air pollution is likely to increase air pollutants, affirming the validity of this instrumental variable.
The models in Table 4 utilize second-step estimation of the 2SLS method.The 2SLS results show that the coefficient estimates on PM2.5 are negative and significant.In education investment regression, a unit increase in PM2.5 will lead to a reduction in the possibility of receiving parental education investment by 1.4%.Adjusted education investment regression shows that one additional unit increase in PM2.5 decreases the probability of receiving adjusted parental education investment by 3.9%.The t-statistic and Kleibergen-Paap rk Wald F-statistic of the instrumental variable are >3.43 and 10, respectively, indicating no weak instrument issues.
The results in Table 5 suggest that air pollution might have effects on education investment through several channels.The results of adjusted education investment regressions indicate that education expectations (i.e., parental education expectation, child education expectation, parental study abroad expectation and child study abroad expectation), financial considerations (i.e., household income level and health expenditure status) and personal well-being (i.e., memory, physical health and self-worth) are likely to mediate the relationship between air pollution and education investment partially.These partial mediation effects were confirmed when the following criteria were met: (1) in the first regression of air pollution and education investment, the coefficient of air pollution is significant and negative if individuals with missing mediators' information are excluded; (2) the second regression shows that air pollution is significantly and negatively associated with mediators; (3) in the third regression, when including both air pollution and a mediator, the results show that the coefficient of the mediator is significant, and the coefficient of air pollution might be significant and have lower values than the coefficient of air pollution in the first regression.Moreover, considering the endogeneity issue of household income level, we utilize rainfall and household heads' unemployment status (self-employed or employed = 0, unemployment = 1) as its instrumental variable (Tanaka et al., 2010).

Table 2
The impact of air pollution on education investment: OLS estimates.We present the results after including moderators and their interactions with air pollution.In Panel B of Table 6, there is a significant indication that the impact of air pollution on adjusted education investment is moderated by education level and living area.The results show that high school students might have more education investment compared to middle school students.Moreover, the results suggest that the negative impacts of air pollution on education investment are significantly negative in urban areas, which is in line with our hypothesis that education investment behaviors in rural areas might not be sensitive to air pollution.Furthermore, the non-significant moderator effects of gender might be due to women's current high social status, which is consistent with our hypothesis.
Finally, we begin by conducting several robustness checks to address possible concerns.In Table A1, we use alternative measures of education investment as a dependent variable in the main regression.These two dependent variables measure the existence and the intensity of family relationship-adjusted education investment, respectively.The first dependent variable is measured by the sum of three dichotomous variables: whether the student attends an after-school tutorial class (yes = 1), whether the student has many books (yes = 1) and whether the student has a good relationship with parents (yes = 1).The second dependent variable is generated by principal component analysis and based on the number of after-school tutorial classes, a categorical variable representing the number of books at home (from 1 to 5, the higher the greater) and whether the student has a good relationship with parents (yes = 1). 13The results are significant and negative, which is consistent with our previous main results in Table 4.
Considering air pollution might have lagged effects on education investment, we estimate the main results using average PM2.5 values in previous years.In Table A2, we find that the coefficient estimates on PM2.5 in the previous one year are significant and negative in education investment and adjusted education investment regressions. 14Other coefficients are not significant.In alternative words, the short-term impact rather than the mid-or long-term impact of air pollution might play a more important role in the reduction in education investment behaviors when exposed to severe air pollution.
In Table A3, we apply a spatial lag model to address estimation concerns stemming from the spatial proximity of adjacent cities.It is essential that the influence of air pollution on education investment in a specific region is not affected by neighboring cities, to maintain the stable unit treatment value assumption (SUTVA).The spatial lag of education investment is determined using the inverse distance weighted average of education investment in neighboring cities based on their centroid.We generate instruments for the spatial lag of education investment using the analogous spatial lag of exogenous air pollution, i.e., inverse distance weighted average of exogenous air pollution in neighboring cities.The outcomes from this model show the influence of local air pollution on education investment remains consistent in magnitude and significance compared to our baseline.This suggests that the effects of thermal inversions and wind are well spread out across cities rather than being concentrated in few neighboring cities, and that there are not localized spillovers or spatial correlation in the error term altering the main estimates.
Table A4 displays the outcomes after accounting for surface ozone (O 3 ), which is one of the dominant air pollutants in China.This analysis investigates whether other air pollutants play important roles in the relationship between air pollution and investment in education.Specifically, we add ground-level O 3 mixing ratio from NASA data to the main regressions.We use the instrumental variable in the main analysis and the instrumental variable generated through Lewbel's method to address the endogeneity of both PM2.5 and O 3 (Lewbel, 2012).The coefficients of PM2.5 remain significant and negative, which confirms that other pollutants might not influence the negative impact of PM2.5 on education investment.
Air pollution might hinder students' academic achievements (Balakrishnan and Tsaneva, 2021;Zhang et al., 2018).Academic achievements may also be influenced by education investment associated with air pollution as increasing education investment could improve the likelihood of academic success (Becker, 1994).We explore the mediator effects of education investment in Table A5.School performance is measured by a categorical self-assessment from the survey question, i.e., "Do you consider yourself a good student?".The findings indicate that education investment might mediate the linkage between air pollution and school performance. 15 We employ the Oster test in Table A6 to further test the importance of

Table 4
The impact of air pollution on education investment: second-stage estimates.Notes: Robust standard errors are clustered by city and reported in parentheses.* p < 0.1, ** p < 0.05, *** p < 0.01. 13The education investment intensity is not incorporated into the main analysis because of potential biases from students taking different subjects across various grades in secondary schools in China.
14 When the PM2.5 value in the current year is added to the regressions, the coefficients of PM2.5 in the previous one year become insignificant. 15School performance might also play a mediator role in the relationship between air pollution and education investment.Excellent school performance might elevate parents' confidence and propensity to allocate large resources to their children's education (H.Wang et al., 2022).Yet, the impairment of cognitive abilities caused by air pollution might hinder students' academic achievements, leading parents to adopt either reinforcement or compensation strategies in children's education investment (Isen et al., 2017).We do not include the results to assess the mediator role of school performance since the school performance variable here is a subjective measure obtained during the survey, which appeared later than the education investment behavior.

Table 5
The impact of air pollution on education investment: mediation analysis.selection on unobservables (Oster, 2019).We generate a predicted PM2.5 based on the first step of the 2SLS regressions of education investment and adjusted education investment, and we calculate the delta value based on the beta value of the 2SLS estimate.Next, we run the second step of the 2SLS regressions and compute the bound of the set with the calculated delta value (Ciacci, 2021).Results show that the beta coefficients without and with controls are negative and significant.The bias-adjusted beta coefficients on education investment and adjusted education investment are also negative and are at a similar magnitude compared to the controlled effect beta (the bounds of the set do not include zero and between the +/− 2.8 standard errors of the controlled estimates).Together, these results suggest that the main findings are not sensitive to potential omitted variable bias.We utilize the wild bootstrap method in Cameron et al. (2008), which has been shown to perform better than the conventional clusterrobust standard errors method when the number of clusters is small.The bootstrap is clustered by city, with a null hypothesis imposed, Webb weights and 1000 replications.It applies the equal-tail p-value and c-test statistic, which might be reliable in instrumental variable estimation.The adjusted p-values after correcting for the small number of clusters and wild bootstrap 95% confidence interval are presented in Table A7.The findings show that the impact of air pollution on education investment is negative and statistically significant under wild cluster bootstrap, which proves the previous results' consistency.
We display the main results across survey cohorts to assess the influence of the 2019 coronavirus outbreak on the second survey (He et al., 2020).The findings of Table A8 denote that there is no significant shift in the coefficients in the second survey.This implies that the adverse impact of air pollution on education expenditure continues to endure in our sample during the virus outbreak.

Conclusions
Our study utilizes a secondary school students' dataset to identify the impact of air pollution on education investment.The dataset was combined with air pollution and weather data from NASA as well as citylevel characteristics from government reports and websites.We employ an instrumental variable method and a two-dimensional fixed effects model.The main findings show that a one unit increase in air pollution is associated with a decrease in the possibility of parental education investment by 3.9%.Additionally, our results show that education expectations, financial considerations and personal well-being mediate this causal chain.We also find the moderating effects of education level and living area.These results indicate a potential cost of air pollution that is overlooked in current environmental policy discussions.Furthermore, it is necessary to pay attention to the role of parents in the long-term effects of air pollution on children's labor market outcomes in adulthood.
We acknowledge some limitations of this study.Firstly, other factors may affect the investment in education, and we could investigate other possible explanations for our results.Secondly, we need to distinguish how air pollution affects education investment before and after implementing environmental policy.The negative impact of air pollution may differ in different periods.Lastly, future studies would be worthwhile to explore the association between air pollution and other investment behaviors.Notes: Robust standard errors are clustered by city and reported in parentheses.* p < 0.1, ** p < 0.05, *** p < 0.01.
L.Guo et al.

Table 3
The impact of air pollution on education investment: first-stage estimates.

Table 6
The impact of air pollution on education investment: moderation analysis.Robust standard errors are clustered by city and reported in parentheses.* p < 0.1, ** p < 0.05, *** p < 0.01.
L.Guo et al.