Shadows in the Air: Mental health vulnerabilities under PM10 and PM2.5, before and after COVID-19

Abstract

Air pollution is an increasing public health concern, with evidence indicating that poor air quality adversely affects health through various pathways. However, its impact on mental health remains underexplored despite indications of pollutant-induced distress. This study examines the causal relationship between air pollution and mental health disorders—such as depression, sleep disorders, ADHD, and OCD—in South Korea. Assessing this relationship is challenging due to the simultaneous occurrence of economic growth, rising stress, and worsening air quality. Additionally, mental health issues have risen despite declining pollution levels post-COVID-19, further complicating the analysis. To address these challenges, we use wind speed and direction as instrumental variables. Our results show significant adverse effects of air pollution on mental health, particularly before 2020. Although the impact diminishes post-pandemic with improved air quality, certain demographic groups remain vulnerable. This study underscores the need for policies safeguarding mental health in at-risk groups, regardless of declining average pollution levels, and contributes to the literature by highlighting effects on less-studied disorders such as ADHD and OCD.

Highlights

• •Measure air pollution's impact on mental health in South Korea pre and post COVID-19.
• •Estimate causal effects on depression, ADHD, OCD, and sleep disorder using 2SLS.
• •Air pollution strongly affects mental health, but effects weakened post-pandemic.
• •Older adults and teenagers remain vulnerable to pollution's mental health impacts.
• •Call for targeted policies to mitigate ongoing risks for high-risk populations.

1. Introduction

Air pollution has emerged as a pressing public health concern globally, with a growing body of evidence linking deteriorating air quality to various physical health issues (Dominici et al., 2006; Hong et al., 2024; Lee et al., 2014). However, the relationship between air pollution and mental health has received comparatively less attention, despite increasing indications that exposure to pollutants may contribute to psychological distress and other adverse mental health outcomes (Bakolis et al., 2020; Buoli et al., 2018; Tao et al., 2023). Therefore, further analysis is required to address a broader spectrum of mental health disorders beyond commonly studied conditions such as depressive symptoms, aggressive behaviors, and anxiety disorders.

Measuring the causal effect of air pollution on various mental health outcomes is essential, yet presents considerable challenges. A significant complicating factor is the potential endogeneity issues arising from the concurrent occurrence of economic expansion, elevated stress levels, and worsening air quality (Marques & Lima, 2011; Ruhm, 2005; Stevens et al., 2015; Wang & Tapia Granados, 2019). These interrelated factors can obscure the true extent of pollution's impact on mental health, potentially resulting in underestimations if not properly addressed. In South Korea, the complexity of this analysis is further amplified by several distinctive factors. Domestic policies and international cooperation have led to a notable reduction in pollutant levels.1 Furthermore, there has been a discontinuous decline in pollutant levels, particularly following the onset of the COVID-19 pandemic. However, despite these improvements in air pollution, the prevalence of mental health patients continues to rise, suggesting a counter-intuitive time-series relationship between declining air pollution and increasing mental health risks. This paradox—where lower pollution levels coincide with persistently high or escalating rates of mental health disorders—necessitates closer examination. Adding to this challenge, transboundary air pollution originating from various regions—particularly industrial areas along South Korea's west coast—and neighboring countries, particularly fine particulate matter such as PM10 and PM2.5, continues to pose a significant challenge to South Korea's air quality, contributing to recurrent periods of severe pollution (Bhardwaj et al., 2019; Jia & Ku, 2019; Jung et al., 2021; Kim, 2019; Lee et al., 2013; Oh et al., 2015; Park & Hwang, 2017; Park & Shin, 2017; Yoo et al., 2020). The intricate interplay between endogeneity concerns and the specific environmental and socio-economic conditions in South Korea underscores the need for a robust causal inference approach to accurately assess the impact of air pollution on mental health outcomes.

To mitigate these issues, the study employs wind speed and direction as instrumental variables, enabling a more accurate assessment of the causal link between air pollution and mental health. Marginal variations in wind speed and direction do not directly affect mental health outcomes, thereby satisfying the exclusion restriction condition. However, these changes in wind patterns contribute to the transport of air pollutants from industrial regions, both domestically and internationally, resulting in increased air pollution levels, particularly PM10 and PM2.5, thereby meeting the relevance and monotonicity conditions.

Analyses using regional and monthly variations in PM10, PM2.5, wind patterns, and mental health outcomes—measured through clinical data on depression, sleep disorders, attention deficit hyperactivity disorder (ADHD), and obsessive-compulsive disorder (OCD)— from January 2015 to December 2023, reveal that air pollution had a significant adverse impact on mental health, with this effect being particularly pronounced in the years before 2020. This adverse effect diminished as pollution levels declined post-pandemic. Nevertheless, certain demographic groups remain disproportionately affected by air pollution, with vulnerabilities varying by age and sex. Indeed, previous research suggests that social and environmental health effects may differ between males and females, as well as among adolescents and older populations (An et al., 2025; Bethmann & Cho, 2022, 2023, 2025; Cho, 2024). These works highlight the persistent influence of socio-environmental factors on mental health outcomes across different demographic groups. This study also emphasizes the need for targeted interventions and policy measures to safeguard mental health, even as overall pollution levels decline.

This study contributes by providing a comprehensive analysis of structural changes in air pollution and their effects on mental health, offering new insights into patterns observed before and after the onset of COVID-19 in a country with relatively high air pollution levels. First, by distinguishing between the impacts of PM2.5 and PM10, we identify which type of particulate matter is more closely associated with mental health conditions and highlight disproportionate effects across demographic groups, categorized by age and sex. Second, this research broadens the scope of environmental health studies by examining the effects of air pollution on mental disorders that have received less attention in previous literature, such as ADHD and OCD, thereby highlighting the broader societal costs associated with deteriorating air quality. Third, this study focuses on South Korea, where both air pollution and mental health concerns have become significant public health challenges. Given that much of the existing research has been conducted in Western countries where air pollution levels are relatively lower (see, for example, Bishop et al., 2023; Chay & Greenstone, 2003; Currie et al., 2009; Currie & Neidell, 2005; Persico & Marcotte, 2022),2 estimating the causal effects of air pollution in a country experiencing higher levels of pollution is essential as the impact of air pollution may be more pronounced in areas with severe pollution than in those with relatively cleaner air (Arceo et al., 2016; Qiu et al., 2022; Xue et al., 2023; Xu et al., 2025). In this regard, our study addresses a gap in the literature by analyzing a country with more substantial air pollution. Finally, mental health research relying on clinical data often suffers from limited representativeness, as it is typically confined to specific cases within certain hospitals (see, for example, Bernardini et al., 2019; Gao et al., 2017; Lu et al., 2020; Szyszkowicz et al., 2010; Tao et al., 2023). In contrast, this study leverages nationwide patient data, ensuring a more representative analysis of the causal relationship between air pollution and mental health outcomes.

2. Related literature

Previous research has primarily concentrated on examining the effects of air pollution on physical health, with a particular focus on cardiovascular and respiratory diseases (Dominici et al., 2006; Hong et al., 2024; Lee et al., 2014). In recent years, a growing body of studies has sought to explore the relationship between air pollution and mental health outcomes (Bakolis et al., 2020; Buoli et al., 2018; Tao et al., 2023). These investigations have largely assessed the impact of air pollution on common set of mental disorders (Bakolis et al., 2020), depressive symptoms (Buoli et al., 2018; Lim et al., 2012; Wang et al., 2014), and mental health-related emergency department visits (EDVs) and hospitalizations (Bernardini et al., 2019; Tao et al., 2023; Szyszkowicz et al., 2010).

A growing number of studies are examining the relationship between air pollution and more specific mental health conditions. Hautekiet et al. (2022) demonstrate that increased exposure to PM2.5, black carbon, and NO2 may raise the risk of generalized anxiety disorder. Similarly, Power et al. (2015) identify associations between PM2.5, PM10, and anxiety symptoms in women. Ji et al. (2021) reveal associations between exposure to PM2.5, PM10, SO2, and CO and the incidence of schizophrenia. In addition, several studies explore potential links between air pollution and conditions such as autism (Dutheil et al., 2021; Li et al., 2023) and sleep disorders (Tang et al., 2020; Yu et al., 2021). Collectively, these findings contribute to the expanding body of evidence highlighting the broader mental health implications of environmental pollutants.

Despite ongoing efforts to investigate the effects of air pollution on various mental health outcomes, significant gaps remain in the literature concerning mental disorders with strong genetic influences, such as ADHD and OCD. Numerous studies have highlighted the substantial genetic component of ADHD, with evidence suggesting that additive genetic factors may account for up to 80 % of the variance in ADHD phenotype (Albayrak et al., 2008; Biederman & Faraone, 2005; Luo et al., 2019; Thapar et al., 2013; Tarver et al., 2014). Similarly, family and twin studies have established a strong heritable component in OCD, demonstrating that individuals with the disorder exhibit significantly higher rates of OCD among their relatives compared to unaffected individuals (Mattheisen et al., 2015; den Braber et al., 2016). Despite the pronounced genetic basis of ADHD and OCD, environmental and socio-economic factors are also believed to play a role in exacerbating symptoms or triggering the onset of these disorders (Cath et al., 2008; Pineda et al., 2007; Pinto et al., 2016). If genetic predisposition alone does not fully account for the occurrence of ADHD and OCD, it is crucial whether air pollution may act as an environmental factor that worsens symptoms or contributes to the emergence of new cases.

To address this gap, the present study examines whether exposure to air pollution—particularly particulate matter (PM10 and PM2.5)—is associated with an increase in ADHD and OCD diagnoses. In addition, we conduct analyses of depressive symptoms and sleep disorders to ensure our findings align with existing literature. Through this comprehensive approach, we aim to validate prior research while extending the scope of mental health investigations by incorporating analyses of ADHD and OCD, thereby contributing a new dimension to the understanding of air pollution's impact on mental health.

3. Methods

3.1. Air pollution and mental health

Nationwide initiatives aimed at reducing air pollution, alongside international cooperation with neighboring countries, have contributed to a consistent decline in PM10 and PM2.5 levels since 2015. However, a discontinuous drop in these levels was observed in 2020, following the onset of the COVID-19 pandemic (see Fig. 1).

Fig. 1.

Annual Average PM10 and PM2.5 Levels

Notes: Using station-level hourly air pollution data for PM10 and PM2.5, we calculate the average monthly levels of PM10 and PM2.5 at the provincial level. Subsequently, we obtain the linear trends before and after the COVID-19 pandemic employing regression discontinuity plots. Bins are selected using the mimicking-variance quantile-spaced method. Researchers may obtain historical air pollution data from the Air Korea archives, accessible at https://www.airkorea.or.kr/web/(last accessed on December 31, 2024).

Simultaneously, mental health issues have emerged as a significant concern in South Korea. The number of patients diagnosed with depression, sleep disorders, ADHD, and OCD has been increasing since 2015 (see Fig. 2).

Fig. 2.

Annual Depression, Sleep Disorder, ADHD, and OCD Patients

Notes: Using provincial monthly data on the number of depressive, sleep-disorder, ADHD, and OCD patients, we calculate the yearly averages. Subsequently, we obtain the linear trends before and after the COVID-19 pandemic employing regression discontinuity plots. Bins are selected using the mimicking-variance quantile-spaced method. Researchers can access historical data on patients with mental disorders from the Health Insurance Review & Assessment Service (HIRA) archives, available at https://opendata.hira.or.kr/home.do (last accessed on December 31, 2024).

As illustrated in Fig. 1, Fig. 2, the observed relationship between particulate matter levels and the number of mental health patients may sometimes appear negative due to underlying time-series patterns. This relationship may stem from endogeneity issues arising from omitted variable bias. Specifically, while certain socioeconomic factors contribute to the deterioration of mental health, air pollution levels have continued to improve, particularly following the onset of COVID-19. As a result, ordinary least squares (OLS) regression may yield misleading findings, suggesting either that air pollution has a beneficial effect on mental health or that the decline in air pollution levels is associated with worsening mental health outcomes. Beyond these time-series patterns, additional sources of endogeneity further complicate the analysis. Air pollution levels frequently increase during periods of economic growth or boom, driven by heightened industrial production and activity. Simultaneously, mental health outcomes often deteriorate during such periods, as individuals face greater stress and pressure from intensified competition and workplace demands (Marques & Lima, 2011; Ruhm, 2005; Stevens et al., 2015; Wang & Tapia Granados, 2019). This interdependence among air pollution, mental health outcomes, and economic cycles gives rise to endogeneity issues, posing significant challenges to accurate estimation. To address endogeneity issues, previous studies have employed instrumental variables such as wind speed and direction (Cho, 2024; Deryugina et al., 2019; Hu et al., 2022; Kwak et al., 2022; Zabrocki et al., 2022), wind speeds (Gu et al., 2019, 2020; Zhang et al., 2020), and thermal inversion (Chen et al., 2024; Fu et al., 2021; Liang et al., 2023; Wang et al., 2021). In this study, to account for both time-series dynamics and endogeneity issues, wind patterns—specifically wind speed and direction, or wind speed alone—are used as instrumental variables. Although thermal inversion is a suitable alternative instrument, data on thermal inversion is unavailable. However, wind speed alone can serve as a proxy, as it operates through the same underlying mechanism. During thermal inversion, stagnant air prevents the dispersion of pollutants within a given region, coinciding with exceptionally low wind speeds. Thus, employing wind speed as an instrumental variable is conceptually equivalent to using thermal inversion, making it a viable alternative in the absence of direct thermal inversion data. Indeed, our sets of instrumental variables are particularly valid, as it effectively accounts for both the transport of air pollutants from external sources and the capture of air pollutants from both internal and external sources. Specifically, wind direction account for the transport of pollutants, while wind speed captures their accumulation and dispersion. This approach satisfies the relevance and monotonicity conditions while remaining exogenous to mental health outcomes, thereby meeting the exclusion restriction. As a result, it enables a more precise estimation of the causal relationship between air pollution and mental health.

3.2. Data and methodology

For the outcome variables, we use medical data from the Health Insurance Review & Assessment Service (HIRA).3 This dataset comprises claims data for healthcare services provided under the Korea's National Health Insurance system from 2015 to 2023. It includes detailed information on patient demographics and diagnoses (classified using Korean Standard Classification of Diseases codes). The descriptive statistics for the outcome variables are presented in Table 1.

Table 1.

Descriptive statistics for the outcome variables.

Variable	Pre- COVID19		Post- COVID19		Definition
	Obs	Mean(SD)	Obs	Mean(SD)
Total	1020	21549.6	816	32696.5	Monthly regional mental health patients diagnosed in depression, sleep disorder, ADHD, and OCD.
		(24070.6)		(39355.3)
Depression	1020	14022.0	816	20998.8	Monthly regional depression patients
		(15538.3)		(25599.1)
Sleep Disorder	1020	5207.21	816	7060.04	Montly regional sleep disorder patients
		(5306.12)		(6993.16)
ADHD	1020	1615.49	816	3623.50	Monthly regional ADHD patients
		(2419.53)		(5861.92)
OCD	1020	704.888	816	1014.11	Monthly regional OCD patients
		(931.900)		(1325.86)

Notes: Standard deviations are reported in parentheses.

In the absence of bias, including auto-correlation driven by time-series patterns between air pollution and mental health outcomes, as well as endogeneity issues, the effect of air pollution could be directly estimated using equation (1).

$$Patient{s}_{ymr}={\beta }_0+{\beta }_{1}AirPollutio{n}_{ymr}+\left({\beta }_{2}Covariate{s}_{yr}\right)+{\mu }_y+{\lambda }_m+{\gamma }_r+{ϵ}_{ymr}$$ (1)

Patients is the outcome variable, representing the number of patients diagnosed with depression, sleep disorder, ADHD, or obsessive-compulsive disorder, during year y and month m, in a specific region (province) r. Air Pollution is measured as the monthly-regional average concentration of PM10 or PM2.5.4 Covariates represent a vector of regional characteristics, which are incorporated as controls in the regressions. To assess the robustness of our findings, the first model presents the initial regression results without controlling for covariates. The second and third models, which incorporate the subset or all of these covariates, are subsequently presented. Finally, ${\mathit{\mu }}_{\mathit{y}}$, ${\mathit{\lambda }}_{\mathit{m}}$ and ${\mathit{\gamma }}_{\mathit{r}}$ are year, month and region fixed effects. Through fixed effects, we can control for year and monthly average variations as well as time-unvarying regional characteristics. The error terms $\mathit{ϵ}$ (and $\mathit{\delta }$ in equation (2)) are clustered at monthly-regional level. Additionally, we employ heteroskedasticity-robust standard errors (Huber-White standard errors) to account for potential heteroskedasticity in the regression analysis. However, as noted in the previous section, the estimated coefficient ${\beta }_1$ may be negative or smaller than the true effect. To address this, a two-stage least squares (2SLS) regression model is employed, utilizing wind patterns, as summarized in equation (2).

$$Patient{s}_{ymr}={\beta }_0+{\beta }_1\widehat{AirPollutio{n}_{ymr}}+\left({\beta }_{2}Covariate{s}_{yr}\right)+{\mu }_y+{\lambda }_m+{\gamma }_r+{ϵ}_{ymr}$$

$$AirPollutio{n}_{ymr}={\alpha }_0+{\alpha }_{1}WindPatterns+\left({\alpha }_{2}Covariate{s}_{yr}\right)+{\mu }_y+{\lambda }_m+{\gamma }_r+{\delta }_{ymr}$$ (2)

The definitions of the variables above are consistent with those in equation (1). Wind Patterns refer to either Wind Speed alone or a combination of Wind Speed and Wind Direction during year y and month m, in a specific region r.5 Wind Speed represents the monthly-regional average wind velocity. Wind Direction indicates the total number of hours per month during which the wind in a given region originates from the Mongolian Desert (Gobi Desert), a common source of air pollutants transported to South Korea (Chung et al., 2005; Ministry of Public Administration and Security, 2019; Jugder et al., 2011; Lee & Sohn, 2011).6 The use of Wind Direction as a measure of winds originating from the Gobi Desert is distinctive compared to previous studies (Deryugina et al., 2019; Hu et al., 2022; Kwak et al., 2022; Zabrocki et al., 2022) and particularly appropriate for analyzing the Korean context, where transboundary air pollution frequently occurs (Bhardwaj et al., 2019; Heo et al., 2009; Jia & Ku, 2019; Jung et al., 2021; Kim, 2019; Lee et al., 2013; Oh et al., 2015; Park & Hwang, 2017; Park & Shin, 2017; Yoo et al., 2020).7

Figure A1.

Regional Variations in Air Pollutants and Instrumental Variables

Notes: We calculate the regional averages (Wind Direction, PM10, PM2.5, and Wind Speed) from March to May, when the variations are the largest among the four seasons.

The controlled covariates include gross regional domestic product (GRDP), regional population, average regional precipitation, and average regional temperature. GRDP is a well-documented factor influencing mental health, as increased production often intensifies competition, leading to higher stress levels and greater mental health vulnerabilities (Marques & Lima, 2011; Ruhm, 2005; Stevens et al., 2015; Wang & Tapia Granados, 2019). Regional precipitation and temperature are also key determinants of mental health outcomes (Goetzmann & Zhu, 2005; Guven & Hoxha, 2015; Hua et al., 2022; Mullins & White, 2019; Nori-Sarma et al., 2022; Wu et al., 2021). Indeed, excessive precipitation, prolonged cloudy weather, and extreme temperatures may contribute to increased mental health risks. The descriptive statistics for the control and instrumental variables are presented in Table 2.

Table 2.

Descriptive statistics for the control and instrumental variables.

Variable	Pre-COVID19		Post- COVID19		Definition
	Obs	Mean(SD)	Obs	Mean(SD)
PM10	1020	43.2667	816	33.9400	Monthly regional average PM10 (microgram per m3)
		(12.3380)		(12.1731)
PM25	1012	24.3914	816	17.9183	Monthly regional average PM2.5 (microgram per m3)
		(7.0872)		(5.7617)
GRDP	1020	37389.9	816	42818.6	Yearly gross regional domestic product (1,000KRW)
		(10887.9)		(11878.0)
Population	1020	3022.8	816	3044.0	Yearly regional population (thousand persons)
		(3179.7)		(3292.4)
Precipitation	1020	100.291	816	122.149	Monthly regional precipitation (millimeter)
		(92.0524)		(136.897)
Temperature	1020	13.7070	816	13.8630	Monthly regional temperature (Celsius)
		(9.1983)		(8.9789)
WindDirection	1020	491.135	816	474.931	Monthly regional west winds (hours)
		(102.809)		(109.519)
WindSpeed	1020	2.0484	816	1.9926	Monthly regional average wind speed (meter per second)
		(0.5941)		(0.6189)

Notes: Standard deviations are reported in parentheses.

Lastly, we wish to identify which demographic groups, categorized by age and sex, are more vulnerable to air pollution. For this analysis, the monthly number of patients by age group (categorized in 10-year intervals) and sex (male and female) is obtained again from the Health Insurance Review & Assessment Service (HIRA). Since the monthly patient data by age group and sex is not available at the regional level, monthly provincial weights from our primary dataset are applied to construct a dataset at the age group, sex, and provincial levels. Subsequently, regression analyses are conducted using equation (2) for each age group and sex.

4. Results

We begin by presenting the ordinary least squares (OLS) results derived from equation (1). The results for the period preceding the COVID-19 pandemic are shown in Table 3, while those for the period following the pandemic are presented in Table 4.

Table 3.

Regression results using ordinary least squares before COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	−32.1188	−23.5925	−3.6225	−4.5385	−0.3652
	(14.3740)∗∗	(10.2092)∗∗	(2.3702)	(1.8383)∗∗	(0.4365)
P>|t|	0.026	0.021	0.127	0.014	0.403
	[18.0696]∗	[12.9028]∗	[3.0203]	[1.9576]∗∗	[0.5947]
P>|t|	0.076	0.068	0.231	0.021	0.539
Observations	1020	1020	1020	1020	1020
PM25	55.3619	40.7719	9.8557	2.1341	2.6003
	(23.7719)∗∗	(16.8835)∗∗	(3.9164)∗∗	(3.0496)	(0.7171)∗∗∗
P>|t|	0.020	0.016	0.012	0.484	0.000
	[25.5210]∗∗	[18.0077]∗∗	[4.4213]∗∗	[3.0738]	[0.8037]∗∗∗
P>|t|	0.030	0.024	0.026	0.488	0.001
Observations	1012	1012	1012	1012	1012
Model 2: GRDP & Population Controlled
PM10	−20.5698	−15.2989	−1.8058	−3.4824	0.0173
	(12.6166)	(8.9273)∗	(2.1077)	(1.7371)	(0.3545)
P>|t|	0.103	0.087	0.392	0.045	0.961
	[15.8151]	[11.2337]	[2.6582]	[1.8374]	[0.5034]
P>|t|	0.194	0.174	0.497	0.058	0.973
Observations	1020	1020	1020	1020	1020
PM25	41.5687	30.8562	7.6910	0.9145	2.1071
	(20.7642)∗∗	(14.6906)∗∗	(3.4705)∗∗	(2.8711)	(0.5783)∗∗∗
P>|t|	0.046	0.036	0.027	0.750	0.000
	[22.2426]∗	[15.7103]∗∗	[3.7738]∗∗	[2.9372]	[0.6487]∗∗∗
P>|t|	0.062	0.050	0.042	0.756	0.001
Observations	1012	1012	1012	1012	1012
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	−20.9414	−15.1036	−2.4754	−3.4341	0.0717
	(12.9209)	(9.1430)∗	(2.1538)	(1.7788)∗	(0.3629)
P>|t|	0.105	0.099	0.251	0.054	0.843
	[16.2854]	[11.5598]	[2.7163]	[1.9046]∗	[0.5166]
P>|t|	0.199	0.192	0.362	0.072	0.890
Observations	1020	1020	1020	1020	1020
PM25	44.5272	34.0201	6.7963	1.4119	2.2990
	(21.3525)∗∗	(15.1035)∗∗	(3.5651)∗	(2.9519)	(0.5940)∗∗∗
P>|t|	0.037	0.025	0.057	0.633	0.000
	[23.3168]∗	[16.5017]∗∗	[3.8580]∗	[3.1006]	[0.6762]∗∗∗
P>|t|	0.056	0.040	0.078	0.649	0.001
Observations	1012	1012	1012	1012	1012
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

Table 4.

Regression results using ordinary least squares after COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	−45.1350	−23.3365	−2.6298	−18.7984	−0.3704
	(37.9982)	(23.1101)	(3.4037)	(11.4613)	(0.9705)
P>|t|	0.235	0.313	0.440	0.101	0.703
	[37.9538]	[22.5701]	[3.4395]	[11.9509]	[0.9130]
P>|t|	0.235	0.301	0.445	0.116	0.685
Observations	816	816	816	816	816
PM25	−131.027	−71.6188	−0.5238	−56.3580	−2.5259
	(88.5454)	(53.8534)	(7.9385)	(26.6910)∗∗	(2.2610)
P>|t|	0.139	0.184	0.947	0.035	0.264
	[87.2346]	[52.2543]	[7.8843]	[27.6771]∗∗	[2.1110]
P>|t|	0.133	0.171	0.947	0.042	0.232
Observations	816	816	816	816	816
Model 2: GRDP & Population Controlled
PM10	−44.0405	−22.6678	−2.4715	−18.5716	−0.3296
	(37.7096)	(22.8285)	(3.3190)	(11.4677)	(0.9646)
P>|t|	0.243	0.321	0.457	0.106	0.733
	[37.3450]	[22.0358]	[3.3243]	[11.9336]	[0.8995]
P>|t|	0.239	0.304	0.457	0.120	0.714
Observations	816	816	816	816	816
PM25	−131.410	−72.7980	−0.5918	−55.5904	−2.4294
	(87.9690)	(53.2539)	(7.7495)	(26.7376)∗∗	(2.2501)
P>|t|	0.136	0.172	0.939	0.038	0.281
	[85.9979]	[51.1538]	[7.6445]	[27.6378]∗∗	[2.0768]
P>|t|	0.127	0.155	0.938	0.045	0.242
Observations	816	816	816	816	816
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	−41.1222	−20.4546	−2.4823	−17.9355	−0.2498
	(37.8791)	(22.9247)	(3.3354)	(11.5217)	(0.9689)
P>|t|	0.278	0.373	0.457	0.120	0.797
	[38.4718]	[22.7145]	[3.3832]	[12.3022]	[0.9286]
P>|t|	0.285	0.368	0.463	0.145	0.788
Observations	816	816	816	816	816
PM25	−124.3239	−66.7901	−0.7446	−54.5581	−2.2311
	(88.7477)	(53.7118)	(7.8212)	(26.9792)∗∗	(2.2699)
P>|t|	0.162	0.214	0.924	0.043	0.326
	[90.3312]	[53.6418]	[7.8447]	[29.1501]∗	[2.1755]
P>|t|	0.169	0.213	0.924	0.062	0.305
Observations	816	816	816	816	816
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

Given the general time-series relationship between particulate matter (PM10 and PM2.5) and the number of mental health patients, the estimated coefficients are predominantly negative. Moreover, the magnitude of these coefficients should be interpreted with caution, as they may understate the true effect.

Moving on to the 2SLS regression model, we first present the results of the first-stage regressions. The first-stage regressions utilizing wind direction and wind speed are shown in Table 5, while those using only wind speed are presented in Table 6.

Table 5.

First-stage regression results using wind speed & direction.

Variable	PM10		PM25
	Pre-Covid19	Post-Covid19	Pre-Covid19	Post-Covid19
Model 1: Covariates Not Controlled
WindDirection	0.0149	0.0036	0.0058	0.0026
	(0.0034)∗∗∗	(0.0035)	(0.0021)∗∗∗	(0.0015)∗
P>|t|	0.000	0.310	0.005	0.075
	[0.0033]∗∗∗	[0.0032]	[0.0020]∗∗∗	[0.0014]∗
P>|t|	0.000	0.258	0.004	0.061
WindSpeed	−3.6396	−2.7787	−3.6779	−3.3197
	(0.9384)∗∗∗	(0.9386)∗∗∗	(0.5665)∗∗∗	(0.3868)∗∗∗
P>|t|	0.000	0.003	0.000	0.000
	[0.9976]∗∗∗	[0.9231]∗∗∗	[0.5570]∗∗∗	[0.4750]∗∗∗
P>|t|	0.000	0.003	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	19.31	4.74	27.66	37.66
F-Stat [Hetero-Rob]	19.74	4.96	29.57	25.34
Model 2: GRDP & Population Controlled
WindDirection	0.0151	0.0038	0.0059	0.0028
	(0.0034)∗∗∗	(0.0036)	(0.0021)∗∗∗	(0.0015)∗
P>|t|	0.000	0.286	0.005	0.057
	[0.0033]∗∗∗	[0.0032]	[0.0020]∗∗∗	[0.0014]∗∗
P>|t|	0.000	0.235	0.003	0.046
WindSpeed	−3.7459	−2.7902	−3.7770	−3.3227
	(0.9449)∗∗∗	(0.9403)∗∗∗	(0.5707)∗∗∗	(0.3870)∗∗∗
P>|t|	0.000	0.003	0.000	0.000
	[0.9977]∗∗∗	[0.9280]∗∗∗	[0.5658]∗∗∗	[0.4778]∗∗∗
P>|t|	0.000	0.003	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	19.91	4.80	28.51	37.85
F-Stat [Hetero-Rob]	20.76	5.00	29.76	25.29
Model 3: GRDP, Population, Precipitation & Temperature Controlled
WindDirection	0.0136	0.0034	0.0052	0.0025
	(0.0034)∗∗∗	(0.0036)	(0.0021)∗∗	(0.0015)∗
P>|t|	0.000	0.341	0.012	0.086
	[0.0032]∗∗∗	[0.0032]	[0.0020]∗∗∗	[0.0014]∗
P>|t|	0.000	0.290	0.010	0.074
WindSpeed	−3.0624	−2.8891	−3.3058	−3.3360
	(0.9369)∗∗∗	(0.9414)∗∗∗	(0.5642)∗∗∗	(0.3857)∗∗∗
P>|t|	0.001	0.002	0.000	0.000
	[0.9753]∗∗∗	[0.9301]∗∗∗	[0.5570]∗∗∗	[0.4809]∗∗∗
P>|t|	0.002	0.002	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	14.97	5.01	22.25	38.14
F-Stat [Hetero-Rob]	16.59	5.15	23.17	24.97
Year F.E	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

Table 6.

First-stage regression results using wind speed.

Variable	PM10		PM25
	Pre-Covid19	Post-Covid19	Pre-Covid19	Post-Covid19
Model 1: Covariates Not Controlled
WindSpeed	−4.1466	−2.7242	−3.8745	−3.2803
	(0.9394)∗∗∗	(0.9370)∗∗∗	(0.5641)∗∗∗	(0.3867)∗∗∗
P>|t|	0.000	0.004	0.000	0.000
	[0.9799]∗∗∗	[0.9171]∗∗∗	[0.5456]∗∗∗	[0.4724]∗∗∗
	0.000	0.003	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	19.48	8.45	47.18	71.95
F-Stat [Hetero-Rob]	17.91	8.82	50.42	48.21
Model 2: GRDP & Population Controlled
WindSpeed	−4.2368	−2.7315	−3.9669	−3.2797
	(0.9471)∗∗∗	(0.9387)∗∗∗	(0.5688)∗∗∗	(0.3870)∗∗∗
P>|t|	0.000	0.004	0.000	0.000
	[0.9802]∗∗∗	[0.9218]∗∗∗	[0.5553]∗∗∗	[0.4752]∗∗∗
	0.000	0.003	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	20.01	8.47	48.63	71.83
F-Stat [Hetero-Rob]	18.68	8.78	51.03	47.63
Model 3: GRDP, Population, Precipitation & Temperature Controlled
WindSpeed	−3.4779	−2.8369	−3.4633	−3.2973
	(0.9382)∗∗∗	(0.9397)∗∗∗	(0.5623)∗∗∗	(0.3856)∗∗∗
P>|t|	0.000	0.003	0.000	0.000
	[0.9598]∗∗∗	[0.9240]∗∗∗	[0.5494]∗∗∗	[0.4788]∗∗∗
	0.000	0.002	0.000	0.000
Observations	1020	816	1012	816
F-Stat (Clustered)	13.74	9.11	37.93	73.14
F-Stat [Hetero-Rob]	13.13	9.43	39.74	47.43
Year F.E	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

As indicated in Table 5, more frequent winds originating from the Gobi Desert facilitate the transport of particulate matter from abroad, as well as from industrial regions typically located along South Korea's western coast. Conversely, slower wind speeds hinder the dispersion of air pollutants within the region, resulting in higher particulate matter concentrations. Previous research commonly employs both wind direction and speed as instrumental variables (Cho, 2024; Deryugina et al., 2019; Hu et al., 2022; Kwak et al., 2022; Zabrocki et al., 2022). However, following the onset of the COVID-19 pandemic, wind direction appears to be a less appropriate instrument, as air pollution from abroad has significantly diminished (Cha et al., 2023; Feng et al., 2022; Yin et al., 2021). Consequently, the use of wind speed as the sole instrument appears more valid in the post-COVID-19 period, as evidenced by the first-stage F-statistics presented in Table 6. In light of this, we provide the 2SLS results using both wind direction and wind speed, as well as wind speed alone.

We then present the 2SLS results using wind speed and direction as instrumental variables. The 2SLS results for the period prior to the COVID-19 pandemic are shown in Table 7, while those for the period following the pandemic are presented in Table 8.

Table 7.

Regression results using 2SLS (wind speed & direction) before COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	216.850	163.096	27.8325	19.8260	6.0957
	(83.1581)∗∗∗	(59.8472)∗∗∗	(13.0359)∗∗	(10.1082)∗∗	(2.4448)∗∗∗
P>|t|	0.009	0.006	0.033	0.050	0.013
	[83.4271]∗∗∗	[59.9638]∗∗∗	[12.6940]∗∗	[11.1203]∗	[2.2499]∗∗∗
P>|t|	0.009	0.007	0.028	0.075	0.007
Observations	1020	1020	1020	1020	1020
PM25	383.820	293.487	45.8102	33.9607	10.5623
	(110.459)∗∗∗	(79.5542)∗∗∗	(17.3488)∗∗∗	(13.6648)∗∗	(3.2343)∗∗∗
P>|t|	0.001	0.000	0.008	0.013	0.001
	[116.791]∗∗∗	[84.3087]∗∗∗	[18.2721]∗∗	[14.6871]∗∗	[3.2420]∗∗∗
P>|t|	0.001	0.000	0.012	0.021	0.001
Observations	1012	1012	1012	1012	1012
Model 2: GRDP & Population Controlled
PM10	237.386	178.252	31.1794	19.9296	8.0249
	(75.0319)∗∗∗	(54.0671)∗∗∗	(11.7358)∗∗∗	(9.4204)∗∗	(2.1766)∗∗∗
P>|t|	0.002	0.001	0.008	0.034	0.000
	[78.9578]∗∗∗	[56.9319]∗∗∗	[11.7030]∗∗∗	[10.7017]∗	[2.1796]∗∗∗
P>|t|	0.003	0.002	0.008	0.063	0.000
Observations	1020	1020	1020	1020	1020
PM25	423.073	322.232	52.2472	34.8113	13.7817
	(100.729)∗∗∗	(72.7628)∗∗∗	(15.6908)∗∗∗	(12.8354)∗∗∗	(2.8789)∗∗∗
P>|t|	0.000	0.000	0.001	0.007	0.000
	[111.721]∗∗∗	[80.5919]∗∗∗	[17.3934]∗∗∗	[14.2827]∗∗	[3.1054]∗∗∗
P>|t|	0.000	0.000	0.003	0.015	0.000
Observations	1012	1012	1012	1012	1012
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	274.098	206.932	34.2188	23.4836	9.4638
	(91.2396)∗∗∗	(66.0127)∗∗∗	(13.9918)∗∗	(11.2742)∗∗	(2.6862)∗∗∗
P>|t|	0.003	0.002	0.014	0.037	0.000
	[95.7115]∗∗∗	[69.2217]∗∗∗	[13.8621]∗∗	[13.0006]∗	[2.6334]∗∗∗
P>|t|	0.004	0.003	0.014	0.071	0.000
Observations	1020	1020	1020	1020	1020
PM25	496.526	380.839	57.5112	41.8104	16.3649
	(121.131)∗∗∗	(88.0326)∗∗∗	(18.3957)∗∗∗	(15.1340)∗∗∗	(3.5006)∗∗∗
P>|t|	0.000	0.000	0.002	0.006	0.000
	[133.992]∗∗∗	[97.1137]∗∗∗	[20.4324]∗∗∗	[17.0126]∗∗	[3.7241]∗∗∗
P>|t|	0.000	0.000	0.005	0.014	0.000
Observations	1012	1012	1012	1012	1012
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively. Over-identification tests, conducted using the Hansen J statistic, reveal no evidence of over-identification at the 1 % significance level.

Table 8.

Regression results using 2SLS (wind speed & direction) after COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	259.142	173.756	57.7451	19.3730	8.2677
	(354.117)	(216.461)	(36.1024)	(103.418)	(9.1245)
P>|t|	0.464	0.422	0.110	0.851	0.365
	[277.534]	[168.509]	[31.0962]∗	[81.8949]	[7.0009]
P>|t|	0.350	0.302	0.063	0.813	0.238
Observations	816	816	816	816	816
PM25	226.615	151.191	49.0126	19.0177	7.3940
	(296.009)	(180.124)	(26.9111)∗	(88.7626)	(7.5725)
P>|t|	0.444	0.401	0.069	0.830	0.329
	[234.442]	[142.985]	[23.7458]∗∗	[69.3859]	[5.9914]
P>|t|	0.334	0.290	0.039	0.784	0.217
Observations	816	816	816	816	816
Model 2: GRDP & Population Controlled
PM10	189.434	123.353	47.2840	12.2557	6.5412
	(342.963)	(207.948)	(33.4398)	(102.300)	(8.8393)
P>|t|	0.581	0.553	0.157	0.905	0.459
	[265.307]	[158.746]	[28.7018]∗	[80.6678]	[6.6785]
P>|t|	0.475	0.437	0.099	0.879	0.327
Observations	816	816	816	816	816
PM25	172.929	111.808	41.0249	13.9686	6.1278
	(291.917)	(176.723)	(25.9879)	(88.4352)	(7.4783)
P>|t|	0.554	0.527	0.114	0.874	0.413
	[227.329]	[136.852]	[22.7815]∗	[68.8392]	[5.8203]
P>|t|	0.447	0.414	0.072	0.839	0.292
Observations	816	816	816	816	816
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	180.535	118.667	45.8334	9.6211	6.4142
	(335.771)	(203.533)	(32.5982)	(100.327)	(8.6573)
P>|t|	0.591	0.560	0.160	0.924	0.459
	[255.122]	[153.255]	[27.9430]	[76.6351]	[6.5272]
P>|t|	0.479	0.439	0.101	0.900	0.326
Observations	816	816	816	816	816
PM25	164.658	107.390	40.4668	10.8108	5.9903
	(292.513)	(177.025)	(26.0569)	(88.6572)	(7.4936)
P>|t|	0.573	0.544	0.120	0.903	0.424
	[223.841]	[135.195]	[22.7200]∗	[67.0123]	[5.8005]
P>|t|	0.462	0.427	0.075	0.872	0.302
Observations	816	816	816	816	816
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively. Over-identification tests, conducted using the Hansen J statistic, reveal no evidence of over-identification at the 1 % significance level.

Before the COVID-19 pandemic, increases in particulate matter (PM10 and PM2.5) driven by wind patterns were associated with a rise in the number of mental health patients. Specifically, Model 1 indicates that a 1 μg per cubic meter increase in PM10 led to a 1.2 percent increase in depression patients, a 0.5 percent increase in sleep disorder patients, a 0.4 percent increase in ADHD patients, and a 0.9 percent increase in OCD patients, resulting in an overall increase of 1.0 percent relative to the mean. Similarly, a 1 μg per cubic meter increase in PM2.5 resulted in a 2.1 percent increase in depression patients, a 0.9 percent increase in sleep disorder patients, a 2.1 percent increase in ADHD patients, and a 1.5 percent increase in OCD patients, with a total increase of 1.8 percent. Notably, the statistical significance remains consistent between clustered standard errors and heteroskedasticity-robust standard errors, demonstrating the robustness of our findings. Furthermore, the estimated magnitudes and standard errors remain consistent regardless of the choice of control variables. Despite the significant effect of particulate matter on mental health outcomes prior to the pandemic, this relationship is no longer statistically significant after the onset of COVID-19, with the exception of sleep disorder patients.

We also present the two-stage least squares (2SLS) results using wind speed as the instrumental variable. The 2SLS results for the period before the COVID-19 pandemic are shown in Table 9, while those for the period after the pandemic are presented in Table 10.

Table 9.

Regression results using 2SLS (wind speed) before COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	397.645	308.721	44.3270	33.8649	10.7322
	(140.297)∗∗∗	(103.940)∗∗∗	(19.9323)∗∗	(15.6085)∗∗	(3.9704)∗∗∗
P>|t|	0.005	0.003	0.026	0.030	0.007
	[159.786]∗∗	[118.651]∗∗∗	[22.8592]∗	[17.3180]∗	[4.3675]∗∗
P>|t|	0.013	0.009	0.052	0.051	0.014
Observations	1020	1020	1020	1020	1020
PM25	433.683	336.298	48.4718	37.1722	11.7412
	(122.342)∗∗∗	(88.7431)∗∗∗	(18.8361)∗∗∗	(14.9026)∗∗	(3.5520)∗∗∗
P>|t|	0.000	0.000	0.010	0.013	0.001
	[137.558]∗∗∗	[100.121]∗∗∗	[21.4112]∗∗	[16.2447]∗∗	[3.8343]∗∗∗
P>|t|	0.002	0.001	0.024	0.022	0.002
Observations	1012	1012	1012	1012	1012
Model 2: GRDP & Population Controlled
PM10	445.792	343.755	52.1753	35.6811	14.1810
	(135.746)∗∗∗	(101.054)∗∗∗	(18.9442)∗∗∗	(14.8945)∗∗	(3.9965)∗∗∗
P>|t|	0.001	0.001	0.006	0.017	0.000
	[156.765]∗∗∗	[116.294]∗∗∗	[22.5718]∗∗	[16.7651]∗∗	[4.5069]∗∗∗
P>|t|	0.004	0.003	0.021	0.033	0.002
Observations	1020	1020	1020	1020	1020
PM25	483.506	372.516	56.6719	38.9466	15.3714
	(113.349)∗∗∗	(82.6216)∗∗∗	(17.1819)∗∗∗	(14.0703)∗∗∗	(3.2364)∗∗∗
P>|t|	0.000	0.000	0.001	0.006	0.000
	[132.752]∗∗∗	[96.5042]∗∗∗	[20.7483]∗∗∗	[15.7884]∗∗	[3.6964]∗∗∗
P>|t|	0.000	0.000	0.006	0.014	0.000
Observations	1012	1012	1012	1012	1012
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	555.032	430.905	60.5930	45.7395	17.7947
	(187.780)∗∗∗	(141.384)∗∗∗	(24.6400)∗∗	(19.8277)∗∗	(5.6165)∗∗∗
P>|t|	0.003	0.002	0.014	0.021	0.002
	[212.143]∗∗∗	[158.940]∗∗∗	[29.0369]∗∗	[22.0941]∗∗	[6.1780]∗∗∗
P>|t|	0.009	0.007	0.037	0.038	0.004
Observations	1020	1020	1020	1020	1020
PM25	565.830	438.909	61.8916	46.8952	18.1339
	(137.476)∗∗∗	(100.965)∗∗∗	(20.1760)∗∗∗	(16.6961)∗∗∗	(3.9615)∗∗∗
P>|t|	0.000	0.000	0.002	0.005	0.000
	[158.245]∗∗∗	[115.656]∗∗∗	[24.2540]∗∗	[18.6327]∗∗	[4.4132]∗∗∗
P>|t|	0.000	0.000	0.011	0.012	0.000
Observations	1012	1012	1012	1012	1012
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

Table 10.

Regression results using 2SLS (wind speed) after COVID19.

Variable	Total	Depression	Sleep Disorder	ADHD	OCD
Model 1: Covariates Not Controlled
PM10	276.052	182.714	56.8372	27.1766	9.3238
	(376.718)	(230.191)	(38.0780)	(109.891)	(9.7796)
P>|t|	0.464	0.427	0.136	0.805	0.340
	[308.752]	[191.139]	[34.8718]	[85.5574]	[8.2206]
P>|t|	0.371	0.339	0.103	0.751	0.257
Observations	816	816	816	816	816
PM25	229.258	151.742	47.2026	22.5698	7.7433
	(302.505)	(184.058)	(27.4508)∗	(90.7408)	(7.7442)
P>|t|	0.449	0.410	0.086	0.804	0.317
	[248.143]	[153.399]	[25.3452]∗	[70.6448]	[6.5438]
P>|t|	0.356	0.323	0.063	0.749	0.237
Observations	816	816	816	816	816
Model 2: GRDP & Population Controlled
PM10	223.019	142.633	48.7248	23.4523	8.2089
	(367.952)	(223.054)	(35.8534)	(109.399)	(9.5706)
P>|t|	0.544	0.523	0.174	0.830	0.391
	[296.002]	[180.487]	[32.6952]	[84.9141]	[7.9504]
P>|t|	0.451	0.429	0.136	0.782	0.302
Observations	816	816	816	816	816
PM25	185.745	118.794	40.5811	19.5326	6.8369
	(299.368)	(181.223)	(26.6240)	(90.6986)	(7.6763)
P>|t|	0.535	0.512	0.127	0.829	0.373
	[240.910]	[146.989]	[24.3686]∗	[70.3976]	[6.3999]
P>|t|	0.441	0.419	0.096	0.781	0.285
Observations	816	816	816	816	816
Model 3: GRDP, Population, Precipitation & Temperature Controlled
PM10	200.120	128.745	46.3637	17.4344	7.5765
	(353.466)	(214.154)	(34.2636)	(105.454)	(9.1763)
P>|t|	0.571	0.548	0.176	0.869	0.409
	[279.303]	[170.760]	[31.0535]	[79.5685]	[7.5397]
P>|t|	0.474	0.451	0.135	0.827	0.315
Observations	816	816	816	816	816
PM25	172.176	110.768	39.8896	14.9999	6.5185
	(298.462)	(180.608)	(26.5647)	(90.4729)	(7.6516)
P>|t|	0.564	0.540	0.133	0.868	0.394
	[235.881]	[144.282]	[24.1360]∗	[68.2756]	[6.3068]
P>|t|	0.465	0.443	0.098	0.826	0.301
Observations	816	816	816	816	816
Year F.E	Yes	Yes	Yes	Yes	Yes
Month F.E	Yes	Yes	Yes	Yes	Yes
Regional F.E	Yes	Yes	Yes	Yes	Yes

Notes: Clustered standard errors are presented in (parentheses), whereas heteroskedasticity-robust standard errors appear in square [brackets]. ∗,∗∗,∗∗∗ denote statistical significance at the 10 %, 5 % and 1 % levels, respectively.

The statistical significance and magnitude of the estimated coefficients are generally consistent with those presented in Table 7, Table 8, although the magnitudes have increased for both the PM10 and PM2.5 regressions. This is likely because using wind speed alone as an instrument better satisfies the relevance condition compared to using wind speed and direction jointly as instruments. For PM10, the magnitudes for Model 1 indicate a 1.8 percent increase in total patients, a 2.2 percent increase in depression patients, a 0.9 percent increase in sleep disorder patients, a 2.1 percent increase in ADHD patients, and a 1.5 percent increase in OCD patients. Similarly, for PM2.5, the magnitudes for Model 1 reflect a 2.0 percent increase in total patients, a 2.4 percent increase in depression patients, a 0.9 percent increase in sleep disorder patients, a 2.3 percent increase in ADHD patients, and a 1.7 percent increase in OCD patients. Similarly, the statistical significance remains consistent between clustered standard errors and heteroskedasticity-robust standard errors. Additionally, both the magnitude and statistical significance of the estimates remain stable regardless of the inclusion of control variables, showing the robustness of our 2SLS estimates.

Lastly, we examine which demographic groups, categorized by age and sex, are more affected by increases in particulate matter. Separate 2SLS regression analyses (following Model 1) were conducted for each age group and sex, using wind speed as the instrumental variable. Given the positive direction of the effect, a one-tailed test was employed, restricting the null hypothesis to values greater than zero. Groups were identified as affected if the p-value was less than 10 percent. Table 11 presents the 2SLS results for the period prior to the pandemic, while Table 12 displays the results for the period following the pandemic.

Table 11.

Affected age groups and sex by mental diseases before COVID19.

Depression					ADHD
	Male		Female			Male		Female
Age	PM10	PM25	PM10	PM25	Age	PM10	PM25	PM10	PM25
0∼9	Affected	Affected	Affected	Affected	0∼9	Affected	Affected	Affected	Affected
10∼19	Affected	Affected	Affected	Affected	10∼19
20∼29	Affected	Affected	Affected	Affected	20∼29	Affected	Affected	Affected	Affected
30∼39	Affected	Affected	Affected	Affected	30∼39	Affected	Affected	Affected	Affected
40∼49	Affected	Affected	Affected	Affected	40∼49	Affected	Affected	Affected	Affected
50∼59	Affected	Affected	Affected	Affected	50∼59	Affected	Affected	Affected	Affected
60∼69	Affected	Affected	Affected	Affected	60∼69	Affected	Affected
70∼79	Affected	Affected	Affected	Affected	70∼79	Affected	Affected	Affected	Affected
80∼	Affected	Affected	Affected	Affected	80∼	Affected	Affected	Affected	Affected

Sleep Disorder					OCD
	Male		Female			Male		Female
Age	PM10	PM25	PM10	PM25	Age	PM10	PM25	PM10	PM25
0∼9					0∼9
10∼19			Affected	Affected	10∼19	Affected	Affected	Affected	Affected
20∼29	Affected	Affected			20∼29	Affected	Affected	Affected	Affected
30∼39	Affected	Affected			30∼39	Affected	Affected	Affected	Affected
40∼49					40∼49	Affected	Affected	Affected	Affected
50∼59				Affected	50∼59	Affected	Affected	Affected	Affected
60∼69	Affected	Affected	Affected	Affected	60∼69	Affected	Affected	Affected	Affected
70∼79	Affected	Affected	Affected	Affected	70∼79
80∼	Affected	Affected	Affected	Affected	80∼	Affected	Affected	Affected	Affected

Notes: One tail test (the null hypothesis: H₀ > 0) is conducted. We declared “Affected” if the statistical significance is below 10 % levels.

Table 12.

Affected age groups and sex by mental diseases after COVID19.

Depression					ADHD
	Male		Female			Male		Female
Age	PM10	PM25	PM10	PM25	Age	PM10	PM25	PM10	PM25
0∼9					0∼9
10∼19					10∼19
20∼29					20∼29
30∼39					30∼39
40∼49					40∼49
50∼59					50∼59
60∼69	Affected	Affected	Affected	Affected	60∼69
70∼79				Affected	70∼79	Affected	Affected
80∼					80∼			Affected	Affected

Sleep Disorder					OCD
	Male		Female			Male		Female
Age	PM10	PM25	PM10	PM25	Age	PM10	PM25	PM10	PM25
0∼9					0∼9
10∼19	Affected	Affected	Affected	Affected	10∼19
20∼29	Affected	Affected			20∼29
30∼39	Affected	Affected	Affected	Affected	30∼39	Affected	Affected
40∼49	Affected	Affected			40∼49	Affected	Affected	Affected	Affected
50∼59			Affected	Affected	50∼59	Affected	Affected
60∼69	Affected	Affected	Affected	Affected	60∼69	Affected	Affected	Affected	Affected
70∼79	Affected	Affected	Affected	Affected	70∼79	Affected	Affected	Affected	Affected
80∼	Affected	Affected	Affected	Affected	80∼

Notes: One tail test (the null hypothesis: H₀ > 0) is conducted. We declared “Affected” if the statistical significance is below 10 % levels.

According to Table 11, nearly all age groups and sexes are affected by the increase in particulate matter. While the effects are not statistically significant on average following the onset of the COVID-19 pandemic (as shown in Table 8, Table 10), certain demographic groups continue to experience mental health impacts from particulate matter exposure. For depression and ADHD, the results suggest that older cohorts are more affected. In the case of sleep disorders, the impact extends from teenage students to older generations, whereas individuals in their prime working years appear to be more affected by OCD. These findings indicate that air pollution continues to exert adverse mental health effects on specific segments of the population.

5. Discussion

The 2SLS regression analysis, utilizing wind direction and speed as instrumental variables, suggests that more frequent winds from the Gobi Desert facilitate the transport of particulate matter from various regions, both domestically and internationally, while slower winds impede its dispersion, resulting in higher concentrations. However, following the onset of COVID-19, wind direction becomes a less effective instrument, whereas wind speed emerges as a more relevant indicator, as reflected in the increase in the first-stage F-statistic when wind speed is used as the sole instrument (see Table 5, Table 6). Before the pandemic, elevated particulate matter concentrations significantly contributed to a higher incidence of mental health conditions, including depression, sleep disorders, ADHD, and OCD (see Table 7, Table 9). Post-pandemic, these effects weaken, with the exception of sleep disorders, which remain statistically significant (see Table 8, Table 10). A demographic analysis further reveals that, despite the overall decline in statistical significance post-pandemic, certain groups continue to be affected (see Table 12).

The lack of statistically significant results following the onset of the pandemic is likely attributable to the reduction in particulate matter concentrations (see Fig. 1). Although the COVID-19 pandemic negatively affected mental health outcomes due to frequent lockdowns, this should not be a problem for our 2SLS results. Our 2SLS estimates specifically capture the effects of increased PM10 or PM2.5 driven by wind properties. A decrease in wind speed by 1 m per second or an additional hour of wind originating from the Gobi Desert is unlikely to produce immediate changes in mental health outcomes (exclusion restriction), including depression, sleep disorders, ADHD, and OCD.

Our research makes several key contributions. A key methodological contribution of this study is the use of wind speed and direction as instrumental variables to address endogeneity in the relationship between air pollution and mental health. While previous research has employed wind direction and speed as instruments (Cho, 2024; Deryugina et al., 2019; Hu et al., 2022; Kwak et al., 2022; Zabrocki et al., 2022), most studies have measured wind direction using a 360-degree scale. In contrast, this study introduces Wind Direction, defined as the total number of hours per month during which winds originate from the Gobi Desert. This measure is particularly relevant for estimating the causal effect of air pollution on mental health in South Korea, where both domestic and transboundary pollution sources frequently contribute to elevated particulate matter concentrations (Bhardwaj et al., 2019; Jia & Ku, 2019; Jung et al., 2021; Kim, 2019; Lee et al., 2013; Oh et al., 2015; Park & Hwang, 2017; Park & Shin, 2017; Yoo et al., 2020). As Wind Direction from the Gobi Desert increases air pollution levels, our instrumental variable satisfies the monotonicity condition (see Table 5), a criterion that would not have been adequately met using a conventional 360-degree scale. Indeed, by leveraging exogenous variations in pollution through the instrumental variables approach, this study enables more precise causal inference. This methodological framework strengthens the reliability of our findings and offers a model for future research in environmental health.

Second, it offers a more comprehensive and detailed analysis of the structural changes in air pollution and their impact on mental health. We provide empirical evidence of the shifts in air pollution patterns before and after the onset of COVID-19, highlighting how these changes influence mental health outcomes. By distinguishing between the effects of PM2.5 and PM10, our study also identifies which type of particulate matter has a greater association with mental health conditions. Additionally, we extend the analysis to examine how different demographic groups—categorized by age and sex—are disproportionately affected by increases in particulate matter. This nuanced approach not only enhances understanding of the population-level health burden but also underscores the importance of targeted policy interventions to protect vulnerable groups.

Third, while the adverse physical health effects of air pollution are well-documented, the mental health impacts of pollutant exposure have received comparatively less attention. This study addresses this gap by investigating the causal relationship between air pollution and mental health outcomes, utilizing clinical data on patients diagnosed with depression, sleep disorders, ADHD, and OCD in South Korea. Furthermore, by focusing on less frequently examined mental disorders—such as ADHD and OCD—which are known to have strong genetic determinants (Albayrak et al., 2008; Biederman & Faraone, 2005; den Braber et al., 2016; Luo et al., 2019; Mattheisen et al., 2015; Tarver et al., 2014; Thapar et al., 2013), this study expands the scope of research on the societal costs associated with deteriorating air quality.

Lastly, this study contributes to the literature by examining a country with relatively higher pollution levels using a representative dataset. Since much of the existing research has been conducted in Western countries (see, for example, Bishop et al., 2023; Chay & Greenstone, 2003; Currie et al., 2009; Currie & Neidell, 2005; Persico & Marcotte, 2022), where air pollution levels are comparatively lower (Burroughs Peña & Rollins, 2017; Chung et al., 2011; Sánchez-Triana et al., 2021; Smith & Mehta, 2003), it is crucial to estimate the causal effects of air pollution in regions experiencing more severe pollution as the effect of air pollution is likely to be more pronounced in heavily polluted areas than in those with relatively cleaner air (Arceo et al., 2016; Qiu et al., 2022; Xue et al., 2023; Xu et al., 2025). In this context, our study addresses a gap by focusing on a country with more significant air pollution. Additionally, mental health research based on clinical data often lacks representativeness, as it is typically limited to specific cases within certain hospitals (see, for example, Bernardini et al., 2019; Gao et al., 2017; Lu et al., 2020; Szyszkowicz et al., 2010; Tao et al., 2023). By utilizing nationwide patient data, this study provides a more comprehensive analysis of the causal relationship between air pollution and mental health outcomes.

This study presents several policy recommendations. To mitigate the mental health risks associated with air pollution, targeted public health interventions should be implemented for vulnerable populations, particularly children and the elderly. These interventions should include promoting protective measures, such as encouraging the use of masks on high-pollution days and installing indoor air filtration systems in schools, hospitals, and nursing homes. Additionally, public health campaigns should aim to raise awareness of the psychological effects of air pollution while expanding access to mental health services in regions with high pollution levels. Next, given the transboundary nature of air pollution, particularly the transport of particulate matter, policymakers should also strengthen regional cooperation with neighboring countries, including China, Mongolia, Taiwan, and Japan, as well as among regional provincial governments, to develop joint air quality management strategies. Bilateral and multilateral agreements on pollution reduction, data-sharing on meteorological and air quality trends, and coordinated emission control policies are essential to effectively addressing cross-border air pollution and its associated health impacts.

This study has two primary limitations. First, the limited number of air pollution monitoring centers in South Korea makes it difficult to capture particulate matter concentrations at the second-tier administrative district level. This constraint reduces the ability to account for finer spatial variations in air pollution, which would allow for a more precise estimation of its effects on mental health. Second, conducting a regional heterogeneity analysis was challenging. Particulate matter concentrations exhibit substantial seasonal fluctuations, with higher levels observed during winter and spring and relatively lower levels during summer and fall. During periods of high air pollution, regional variations in pollution levels are more pronounced, whereas during periods of low air pollution, regional variations are minimal. While the instrumental variables remain valid, as wind speed and direction continue to correlate with air pollution levels across seasons, analyzing regional heterogeneity—such as variations in the impact of air pollution on mental health across regions with different levels of economic development or disparities in healthcare resources—would require focusing on specific months of the year. However, restricting the analysis to certain time periods may not be statistically adequate. A more comprehensive examination of regional heterogeneity would be feasible if there were greater regional variations in air pollution levels throughout the year. Despite these limitations, this study show the causal effect of air pollution on mental health in South Korea, particularly on the less examined mental disorder (i.e., ADHD and OCD) across different demographic groups, while effectively addressing endogeneity concerns.

6. Conclusion

In conclusion, this study provides robust evidence that air pollution significantly impacts mental health outcomes, with the adverse effects diminishing following the onset of the COVID-19 pandemic. By employing a 2SLS regression approach using wind speed and direction as instrumental variables, the analysis effectively addresses endogeneity concerns, enhancing the reliability of causal inferences. The findings underscore the importance of considering demographic disparities, as certain groups remain disproportionately affected despite overall declines in pollution levels. Furthermore, the study bridges a critical gap in environmental health research by examining the effects of air pollution on less frequently studied mental health conditions such as ADHD and OCD. These insights emphasize the need for targeted policy interventions to mitigate the mental health risks associated with air pollution, thereby safeguarding vulnerable populations and informing future public health strategies.

CRediT authorship contribution statement

Jae Il Cho: Writing – review & editing, Writing – original draft, Formal analysis, Conceptualization. Kyungsun Kim: Writing – review & editing, Writing – original draft, Formal analysis, Conceptualization.

Ethical Statement

Hereby, I/Jae Il Cho/consciously assure that for the manuscript/Shadows in the Air: Mental Health Vulnerabilities Under PM10 and PM2.5, Before and After COVID-19/the following is fulfilled.

• 1)This material is the authors' own original work, which has not been previously published elsewhere.
• 2)The paper is not currently being considered for publication elsewhere.
• 3)The paper reflects the authors' own research and analysis in a truthful and complete manner.
• 4)The paper properly credits the meaningful contributions of co-authors and co-researchers.
• 5)The results are appropriately placed in the context of prior and existing research.
• 6)The data collected by Health Insurance Review & Assessment Service is publicly available for research purposes.
• 7)All sources used are properly disclosed (correct citation). Literally copying of text must be indicated as such by using quotation marks and giving proper reference.
• 8)All authors have been personally and actively involved in substantial work leading to the paper and will take public responsibility for its content.

The violation of the Ethical Statement rules may result in severe consequences.

I agree with the above statements and declare that this submission follows the policies outlined in the Guide for Authors and in the Ethical Statement.

Declaration of interest statement

We do not have any financial and personal conflicts of interest.

Data availability

Data will be made available on request.