Differential association between ambient PM_2.5 constituents and depression and anxiety among Indian adults

Summary

Ambient fine particulate matter (PM_2.5) exposure influences mental disorders, yet the differential effects of its chemical constituents remain poorly understood in low-and middle-income countries. We examined associations between PM_2.5 components with depression and anxiety among 34,802 Indian adults from the National Mental Health Survey (2015–16). We integrated satellite data and chemical transport modeling at 1-km resolution to assign one-year PM_2.5 component exposures to participants’ residential addresses. Mental disorders were assessed using standardized diagnostic interviews. Using logistic mixed-effect regression, we found that an interquartile range increase in PM_2.5 raised the odds of depression and anxiety by 8% (7–9) and 2% (1–3), respectively. Secondary inorganic aerosols and carbonaceous aerosols showed stronger associations than other constituents. The cumulative effect of PM_2.5 components exceeded that of total mass, indicating that relying solely on PM_2.5 mass underestimates mental health effects. Our findings support targeted emission controls that focus on sectors producing the most toxic species, resulting in greater health benefits.

Graphical abstract

Highlights

• •Long-term ambient PM_2.5 exposure was associated with depression and anxiety in Indian adults
• •Secondary inorganic aerosols and carbonaceous aerosols showed stronger associations
• •Cumulative effect of PM_2.5 components exceeds that of total PM_2.5 mass
• •Associations were stronger among middle-aged adults, low-income groups, and urban residents

Public health; Anxiety in health technology; Environmental science; Environmental health; Neuroscience

Introduction

Mental well-being is a vital part of overall health.¹ Mental disorders are a leading cause of the global disease burden,² posing serious public health challenges, especially in low- and middle-income countries (LMICs). According to the Global Burden of Disease (GBD) 2021, depression and anxiety were the top contributors to Years Lived with Disability (YLDs), accounting for 56.3 and 42.5 million YLDs, respectively.² In India, nearly one in seven individuals is affected by mental disorders, with both prevalence and economic burden increasing since 1990.³ Given the heavy toll of these disorders, significant gaps in awareness, diagnosis, and treatment still exist. It is therefore essential to identify the modifiable risk factors for depression and anxiety, especially those that can be addressed through population-level interventions.

Besides the conventional determinants such as genetics, lifestyle behaviors, socioeconomic status, family history, and medical conditions, environmental exposures are significantly recognized as key contributors to mental health.⁴ Among these, ambient fine particulate matter (PM_2.5) poses one of the greatest global health threats, contributing to 4.7 million (95% UI: 3.5–5.8) deaths annually.⁵ Historically, PM_2.5 has been strongly associated with cardiopulmonary morbidity and mortality,^6,7 but emerging evidence also suggests a potential detrimental impact on mental health outcomes.⁸ PM_2.5, often coated with neurotoxic chemicals,⁹ penetrates the brain through the blood-brain barrier, olfactory bulb, and gut-brain axis, adversely affecting the central nervous system.⁸ Several cohort and cross-sectional studies, mostly from Western countries and China, have reported a positive link between short- and long-term PM_2.5 exposure and increased risk of depression and anxiety.^10,11,12,13 Although systematic reviews and meta-analyses support a consistent link with depression, findings regarding anxiety remain inconclusive.^14,15 However, the biologically relevant exposure window remains unknown, and recent exposure may aggravate acute systemic inflammation or exacerbate chronic conditions. In contrast, long-term exposure may contribute through chronic oxidative stress, neuroinflammation, or the development of chronic disease.¹⁶

PM_2.5 is a heterogeneous mixture of distinct chemical constituents, mainly elemental carbon, organic carbon, sulfate, nitrate, and ammonium, emitted from multiple sources, such as vehicular emissions, residential fuel combustion, industrial emissions, fossil fuel combustion, agricultural activity, and secondary atmospheric reactions. While most existing epidemiological studies have predominantly focused on total PM_2.5 mass,^10,11,16,17 the differential impacts of its constituents remain poorly understood. Toxicological evidence reported that PM_2.5 constituents differ markedly in their physicochemical properties and neurotoxicity,¹⁸ suggesting that some may pose greater mental health risks than others.

A key methodological challenge lies in multicollinearity, as many PM_2.5 species share common sources and are therefore highly correlated. Such collinearity can inflate variance estimates and may obscure true associations. Several statistical frameworks have been developed to address this issue. These include ridge regression, which stabilizes estimates in the presence of strong correlations by introducing a small bias.¹⁹ Mixture models, such as Weighted Quantile Sum (WQS) regression, quantile-based g-computation, and Bayesian Kernel Machine Regression (BKMR), estimate the joint effect of correlated pollutants and identify major contributors.^20,21,22,23 Alternatively, two-stage residual models regress each component on total PM_2.5 to derive residuals representing independent variability, which are then incorporated into outcome models to estimate component-specific associations.²⁴ Compared with mixture approaches, this residualized model provides interpretable coefficients and allows direct comparison with total PM_2.5 effects.

Although emerging studies from China have adopted such mixture frameworks for mental-health outcomes.²⁵ No comparable national-scale analyses exist for India, where PM_2.5 levels are among the highest globally and emission sources differ substantially from those in high-income countries. Therefore, it is crucial to identify the most toxic components to design evidence-based and source-targeted interventions, rather than relying solely on aggregated PM_2.5 mass.

In this analysis, we filled these knowledge gaps using the latest nationally representative mental health data and high-resolution exposure data from India. First, we examined the long-term associations between ambient PM_2.5 exposure and depression and anxiety among Indian adults. Second, we assessed whether specific chemical components of PM_2.5 have stronger associations with mental health outcomes than the total PM_2.5 mass. Finally, we evaluated effect modification by key socio-demographic factors. Our findings are intended to help develop more targeted air pollution control policies that improve mental health and reduce health disparities in LMICs.

Results

Characteristics of the study population

The NMHS survey contacted 39,532 eligible adults aged 18 years and older, and 34,802 (88%) were interviewed from October 2, 2015, to July 10, 2016 (Figure S1). Respondents aged 18–29 years comprised the largest age group (34%), with more than half being women (52%). Nearly 75% of respondents were married, 24% were illiterate, and 69% were from rural areas. Other sociodemographic characteristics of the NMHS participants are shown in Table 1.

Table 1.

Descriptive characteristics of the NMHS study population

Characteristic	N = 34,802a
Age Group
18–29	11,848 (34%)
30–39	7,062 (20%)
40–49	5,854 (17%)
50–59	4,448 (13%)
60 and above	5,590 (16%)
Education
Illiterate	8,510 (24%)
Primary	6,160 (18%)
Secondary	5,722 (16%)
High School	6,493 (19%)
Higher Education	7,917 (23%)
Gender
Female	18,217 (52%)
Male	16,585 (48%)
Marital Status
Married	25,980 (75%)
Never Married	6,678 (19%)
Widowed/Divorced/Separated	2,144 (6.2%)
Occupation
Not Working	18,002 (52%)
Working	16,800 (48%)
Region
Central	5,462 (16%)
Eastern	5,668 (16%)
North-Eastern	5,455 (16%)
Northern	6,403 (18%)
Southern	5,538 (16%)
Western	6,276 (18%)
Place of residence
Rural	23,957 (69%)
Urban Non-Metro	6,601 (19%)
Urban Metro	4,244 (12%)
Season of data collection
Pre-monsoon	10,651 (31%)
Monsoon	198 (0.6%)
Post-monsoon	10,046 (29%)
Winter	13,907 (40%)
Wealth Quintiles
Lowest	6,992 (20%)
Second	6,849 (20%)
Middle	6,865 (20%)
Fourth	7,003 (20%)
Highest	7,093 (20%)
Age (years)	36 (25, 51)

^aNote: Demographic, socioeconomic, and residential characteristics of participants from the National Mental Health Survey of India (2015–2016). Data presented as n (%) for categorical variables and median (Q1, Q3) for continuous variables. N = 34,802 adults aged ≥18 years from 720 clusters across 12 states. Variables include age group, education level, gender, marital status, occupation, geographic region, place of residence, season of data collection, wealth quintiles, and age distribution.

The weighted prevalence of the current depressive disorder among adults was 2.68% (2.65–2.71). The prevalence was relatively higher among females (2.97%), in the 40–49 age group (3.65%), those who were illiterate (3.62%), belonged to lower wealth quintiles (3.42%), those who were widowed, divorced, or separated (5.24%), or resided in urban metros (5.19%). The prevalence of depressive disorder showed significant variation across different regions of India, with the highest rates in Eastern India (4.54%), followed by Southern India (3.91%). The prevalence varied according to seasons and exposure levels of PM_2.5 and diurnal temperature ranges (Table S1).

Regarding current anxiety disorder, the weighted prevalence (2.95%; 2.92–2.98) is slightly higher than that of current depressive disorder (2.68%; 2.65–2.71) and shows significant variation across factors such as age, gender, education, occupation, marital status, wealth quintiles, and season, as well as differing exposure levels to PM_2.5 and diurnal temperature ranges. Eastern India has the highest prevalence of anxiety disorder (4.69%), followed by Western India (3.45%). The prevalence patterns are consistent with those observed in depressive disorder (Table S2). The prevalence of anxiety disorder subtypes ranges from PTSD (0.24%) to agoraphobia (1.62%) (Table S3).

Heterogeneity in exposure to PM_2.5 and its constituents

Figure 1 shows the spatial distribution of annual average PM_2.5 exposure across NMHS clusters in India, along with the corresponding mass fractions of major chemical components and sectoral contributions. We found the annual average PM_2.5 concentration across the study population was 44.3 μg m⁻³, with an IQR of 34.5–49 μg m⁻³ (Table S4). The primary contributors were organic carbon (OC), sulfate (SO₄²⁻), ammonium (NH₄⁺), nitrate (NO₃⁻), and elemental carbon (EC), with mean concentrations of 8.8, 5.6, 2.6, 2.5, and 1.7 μg m⁻³, respectively. Other species, including remaining inorganic ions, soil, water molecules, and unspecified species, accounted for an additional 21.6 μg m⁻³.

Figure 1.

Spatial distribution of annual PM_2.5 exposure along with the mass fractions of species and sectoral contributions at the geolocation of NMHS clusters

PM_2.5 concentrations at 1-km resolution across 720 clusters in 12 Indian states. (A) Northern Region (Punjab and Uttar Pradesh), (B) Western Region (Rajasthan and Gujarat), (C) North-Eastern Region (Assam and Manipur), (D) Central Region (Madhya Pradesh and Chhattisgarh), (E) Eastern Region (West Bengal and Jharkhand), and (F) Southern Region (Tamil Nadu and Kerala). Circle size represents PM_2.5 concentration (range: 21.8–96.2 μg/m³; mean ± SD: 44.3 ± 13.5 μg/m³). Pie charts within circles show mass fractions of major PM_2.5 chemical components (elemental carbon, organic carbon, sulfate, nitrate, ammonium, soil, and other species) and sectoral contributions from WRF-CMAQ simulations (transport, domestic, road dust, agricultural residue burning, industry, power, and other sectors). Data represent annual averages for the study period (2015–2016). N = 720 clusters. See also Figures S1 and S2, and Table S4.

The correlation matrices among PM_2.5 constituents showed strong coherence (Figure S2). Total PM_2.5 mass was highly correlated with OC, EC, NH₄⁺, and NO₃⁻ (r = 0.78 to 0.90). Strong correlations were also observed between EC and OC (r = 0.89), which shared common sources, and between NH₄⁺ and SO₄²⁻ (r = 0.81), which showed secondary particle formation. NO₂ and O₃ showed weak correlations with PM_2.5 species because of their distinct atmospheric chemistry.

Association of PM_2.5 and its composition with mental health

In the fully adjusted model, long-term exposure to PM_2.5 mass and its major chemical components had statistically significant positive associations with depression (Figure 2A). A per-IQR increase in total PM_2.5 was associated with 8% higher odds of depression (OR: 1.08; 95% CI: 1.07–1.09). Among the components, the strongest associations were observed for secondary inorganic aerosols such as SO₄²⁻ (1.21; 1.18–1.23), NH₄⁺ (1.17; 1.15–1.18), and NO₃⁻ (1.09; 1.07–1.10), followed by carbonaceous aerosols such as OC (1.04; 1.03–1.05) and EC (1.03;1.02–1.05). Soil and other species also showed positive associations, with an OR of 1.03 and 1.07, respectively.

Figure 2.

Association between exposure to PM_2.5 and its components and mental health outcomes

Odds ratios (ORs) with 95% confidence intervals (CIs) for (A) Depression and (B) Anxiety per interquartile range (IQR) increase in one-year average PM_2.5 and its component concentrations. Logistic mixed-effects regression models (N = 34,802 participants from 720 clusters, 9,666 households). All associations shown are statistically significant (p < 0.05) except OC, SO₄²⁻, soil, and other species for anxiety. Models were adjusted for age group, gender, education, occupation, marital status, wealth quintiles, residential type, season of data collection, and diurnal temperature range. Random intercepts were included at the cluster and household levels to account for hierarchical clustering. N = 34,802 participants from 720 clusters. See also Figures S3–S11, Tables S5, and S6.

For anxiety, the associations were weaker than those observed for depression. NO₃⁻and EC showed the highest associations, with ORs of 1.14 (1.12–1.15) and 1.10 (1.08–1.11), respectively, followed by NH₄⁺ (1.05; 1.04–1.07), whereas several other components, such as OC, SO₄²⁻, and soil, showed null or minimal associations (Figure 2B).

Sensitivity analysis of co-pollutant

After adjusting for gaseous co-pollutants such as NO₂ and O₃, the associations between all PM_2.5 components and depression remained highly consistent with the primary estimates (Tables S5).

Effect estimates for SO₄²⁻, NO₃⁻, NH₄⁺, EC, OC, soil, and other species showed minimal changes and retained statistical significance. In case of anxiety, the key components such as NO₃⁻, EC, and NH₄⁺ remained statistically significant, with slight reductions in magnitude; while the associations for OC, SO₄²⁻, soil, and other species showed attenuations and loss of precision (Table S6). These findings suggest strong robustness for depression and partial attenuation for anxiety.

Exposure-response relationship

The exposure-response curve showed non-linear relationships between annual PM_2.5 and its major constituents and the odds of depression and anxiety (Figure 3). Total PM_2.5, EC, and OC showed a U-shaped pattern with steeper positive slopes above approximately 50 μg m⁻³ (PM_2.5), 2 μg m⁻³ (EC), and 10 μg m⁻³ (OC). At concentrations above about 4 μg m⁻³ and 3 μg m⁻³, respectively, we observed significant positive associations for both NO₃⁻ and NH₄⁺, with higher slopes for anxiety than for depression. SO₄²⁻ was positively associated with depression across the entire concentration range, while the anxiety curve fluctuated and was not statistically significant (Figure 2). The other species showed a clear positive association for depression at higher concentrations. These findings suggest that PM_2.5 constituents, particularly those associated with combustion and secondary aerosol formation, contribute to adverse mental health outcomes.

Figure 3.

Exposure-response relationships between one-year average PM_2.5 and its components and the log-odds of depression and anxiety

Non-linear exposure-response curves for depression (blue) and anxiety (red) across PM_2.5 and its major chemical components. The solid line represents model-predicted log-odds from generalized additive mixed models (GAMMs) with penalized cubic splines; shaded regions show 95% CIs. Rug plots along the x axis show the distribution of pollutant concentrations across study participants. Models were adjusted for age group, gender, education, occupation, marital status, wealth quintiles, residential type, season of data collection, and diurnal temperature range. Random intercepts were included at the cluster and household levels. N = 34,802 participants from 720 clusters.

We highlight two additional key points from this analysis. First, anxiety consistently showed greater sensitivity to higher pollutant concentrations compared to depression among Indian adults. Second, the slope of the exposure-response curve becomes steeper when PM_2.5 levels exceed 50 μg m⁻³.

Cumulative effect of PM_2.5 composition

We defined “cumulative effect” as the mass fraction weighted sum of effect sizes for PM_2.5 components. This approach (see STAR Methods) estimates the expected change in mental health outcomes per 10 μg m⁻³ increase in PM_2.5 mass, as described in our previous work.²⁴

The cumulative PM_2.5 components showed stronger associations with both depression and anxiety than total PM_2.5 mass. For depression, for every 10-μg m⁻³ increase in cumulative exposure to PM_2.5 constituents (1.32; 1.29–1.34) was higher than the OR (1.06; 1.05–1.06) using PM_2.5 mass as a proxy. While for anxiety, the corresponding OR was 1.15(1.11–1.18) for cumulative constituents versus 1.01(1.00–1.02) for total PM_2.5 mass.

Sensitivity analysis of the mixture effect of PM_2.5 composition

Each decile increase in the mixture was associated with an OR of 1.04 (0.99–1.10) for depression and 1.03 (0.97–1.10) for anxiety. The relative contributions were highest for SO₄²⁻ (34%), NH₄⁺ (28%), and Soil (19%) for depression, and soil (54%), NO₃⁻(30%), and EC (9.6%) for anxiety (Figure S3).

Subgroup analysis

We examined seven potential effect modifiers: age group, education, residence type, marital status, wealth quintiles, seasons, and regions, which significantly altered the association between PM_2.5 exposure and mental health outcomes. Table 2 displays the odds ratios (ORs) for depression and anxiety linked to a per IQR increase in PM_2.5, varied by subgroups. Several key points were notable. The associations with both depression and anxiety were strongest among individuals aged 40–49 years compared to other age groups.

Table 2.

Odds ratio of depression and anxiety associated with a Per-IQR increase in PM_2.5 by subgroups

	OR (Depression)	OR (Anxiety)
Age Group
18-29	0.94 (0.93–0.96)	0.95 (0.94–0.97)
30-39	1.06 (1.05–1.08)	1.02 (1.01–1.03)
40-49	1.16 (1.15–1.18)	1.12 (1.1–1.13)
50-59	1.14 (1.12–1.16)	1.04 (1.02–1.05)
≥60	1.12 (1.11–1.14)	0.99 (0.97–1)
Education
Higher-Education	0.94 (0.92–0.95)	1 (0.99–1.01)
High-School	1.06 (1.04–1.07)	1.07 (1.06–1.09)
Secondary	1.09 (1.07–1.11)	1.02 (1–1.03)
Primary	1.1 (1.09–1.12)	1.03 (1.02–1.04)
Illiterate	1.13 (1.12–1.15)	0.98 (0.97–1)
Marital Status
Married	1.07 (1.06–1.08)	1.02 (1.01–1.03)
Never Married	1.06 (1.04–1.07)	0.99 (0.97–1)
Others	1.19 (1.17–1.21)	1.1 (1.08–1.12)
Region
Central	0.81 (0.79–0.82)	0.88 (0.86–0.89)
Eastern	1.08 (1.06–1.09)	1.01 (1–1.03)
North-Eastern	0.89 (0.87–0.91)	0.69 (0.67–0.71)
Northern	0.82 (0.81–0.84)	0.92 (0.9–0.93)
Southern	1.01 (0.99–1.03)	0.72 (0.7–0.73)
Western	0.87 (0.85–0.88)	1.02 (1–1.03)
Residential Type
Rural	0.99 (0.98–1)	0.93 (0.92–0.94)
Urban-metro	1.31 (1.3–1.33)	1.19 (1.18–1.21)
Urban-non-metro	0.95 (0.94–0.97)	0.87 (0.86–0.89)
Season of data collection
Post-Monsoon	1.24 (1.22–1.25)	1.12 (1.1–1.13)
Pre-monsoon	0.86 (0.85–0.87)	0.75 (0.74–0.77)
Winter	0.99 (0.97–1)	0.98 (0.97–0.99)
Wealth Quintiles
Highest	0.94 (0.93–0.95)	0.92 (0.91–0.93)
Fourth	0.99 (0.98–1.01)	0.95 (0.93–0.96)
Middle	1.1 (1.08–1.11)	1.02 (1–1.03)
Second	1.14 (1.12–1.15)	1.08 (1.06–1.09)
Lowest	1.18 (1.16–1.19)	1.11 (1.1–1.13)

Effect modification of the association between one-year average PM_2.5 exposure and mental health outcomes across sociodemographic strata. Odds ratios (ORs) with 95% confidence intervals (CIs) are shown for a per-interquartile range (IQR = 14.44 μg/m³) increase in PM_2.5 concentration. Subgroups include age group, education level, marital status, geographic region, residential type, season of data collection, and wealth quintiles.

Statistical analysis: Logistic mixed-effects regression models with interaction terms (N = 34,802 participants from 720 clusters). Models were adjusted for age group, gender, education, occupation, marital status, wealth quintiles, residential type, season of data collection, and diurnal temperature range. Random intercepts were included at the cluster and household levels. Stratified estimates are presented for subgroups showing statistically significant effect modification (p for interaction <0.05). See also Figures S4–S11 for component-specific subgroup analyses.

We observed a clear socioeconomic gradient, with the lowest income group showing a stronger link between PM_2.5 exposure and both depression and anxiety. For example, the OR for depression per IQR increase in PM_2.5 exposure ranged from 1.18 (1.16–1.19) in the lowest income group to 0.94 (0.93–0.95) in the highest. Urbanicity also affected these associations, with residents of metropolitan areas being more susceptible to both outcomes compared to residents of rural and non-metropolitan areas. Educational attainment also influences these relationships: individuals with no formal education had the strongest link with depression, while for anxiety, the highest association was among those with a high school education.

Marital status was another key factor, with divorced, separated, or widowed individuals showing increased susceptibility to both outcomes linked to PM_2.5 exposure. These associations varied across regions, with stronger links observed in eastern India for depression related to PM_2.5 and its components. Conversely, the western region showed the strongest links with anxiety, especially for EC, OC, NO₃⁻, NH₄⁺, and soil. Additionally, seasonal variation was evident, with the post-monsoon season showing stronger associations compared to other seasons. Subgroup analyses for PM_2.5 and its components are detailed in Figures S4–S11, displaying similar patterns to those observed for total PM_2.5 mass.

Discussion

To date, no epidemiological studies in India have thoroughly examined the links between both total PM_2.5 mass and its chemical components with mental health outcomes, especially depression and anxiety. Using a unique population-based mental health survey across 12 Indian states, we found that long-term exposure to ambient PM_2.5 and several key chemical constituents was associated with poor mental health among Indian adults, with stronger and more consistent associations observed for depression (see Figure 2). Secondary inorganic aerosols (SO₄²⁻, NO₃⁻, and NH₄⁺) and carbonaceous aerosols (EC and OC) had stronger associations with depression and anxiety, and these associations remained robust after adjusting for gaseous co-pollutants. These findings suggest that PM_2.5 toxicity varies by species, with some elements more likely to cause adverse neurological and psychological effects. We applied a two-stage modeling approach that considers the covariation among components, allowing us to isolate the impact of each species and reduce confounding from co-occurring pollutants. Further, we complemented this approach with WQS mixture analysis. This convergence across modeling approaches strengthens confidence in our conclusions while acknowledging that future work could leverage non-linear mixture methods such as BKMR. These results offer important evidence for developing targeted air pollution control measures that focus on the most toxic species.

The observed exposure-response curves revealed non-linear patterns, including U-shaped relationships at lower concentrations. These patterns warrant cautious interpretation. Such shapes may reflect data sparsity and higher measurement error at very low concentrations, which can introduce instability in spline-based estimates. In addition, low-exposure areas in our study tend to differ socioeconomically and environmentally from high-exposure areas, including higher socioeconomic status, more green space, reduced noise, and different behavioral or psychosocial contexts, raising the possibility of residual confounding. Heterogeneity in species composition at lower concentrations may also contribute to these patterns rather than true protective effects. These caveats highlight the need for longitudinal studies with more precise exposure characterization.

Our findings are consistent with a growing body of literature that links ambient PM_2.5 exposure to adverse mental health outcomes, including depression and anxiety.^10,11,12,13 Large cohort studies and meta-analyses have reported similar associations in the general population,^{13,14,15,16,17} with some studies reporting null or inverse relationships.^26,27 Such heterogeneity may arise from differences in study design, geographical coverage, population vulnerability, exposure measurement, and outcome assessment. Recent studies from China reported that prolonged exposure to air pollution intensifies the mental health burden,²⁸ which aligns with our study. While prior epidemiological research has focused on total PM_2.5 mass,^10,11,16,17 with relatively few studies exploring the role of its chemical composition. However, two recent studies from China have examined these associations, indicating that carbonaceous aerosols, heavy metals, NH₄⁺, and SO₄²⁻ were the key drivers of neurological outcomes.^29,30 Our results are consistent with these patterns, showing stronger associations for SO₄²⁻, NO₃⁻, NH₄⁺, EC, and OC, thereby extending this evidence to South Asia.

Moreover, our analysis revealed that the cumulative association of PM_2.5 components for every 10-μg m⁻³ increase in ambient PM_2.5 was significantly greater than the effect estimate of total PM_2.5 mass alone, implying that some less-toxic species may dilute the overall mass effect. Therefore, relying solely on PM_2.5 mass could substantially underestimate the magnitude of these associations. However, this interpretation assumes additive effects, which may not hold if components interact synergistically or antagonistically, as a limitation. This potential interaction among components can be further explored in future longitudinal studies.

Several biological plausibility supports our findings. Toxicological research has shown that PM_2.5 and its reactive species can reach the brain through the blood-brain barrier, the olfactory nerve, or the gut-brain axis, triggering oxidative stress, neuroinflammation, alterations in gut microbiota, and systemic inflammation.^8,31,32 Secondary inorganic aerosols such as NO₃⁻ and SO₄²⁻ may induce oxidative stress³³ and mitochondrial dysfunction in the central nervous system,³⁴ while NH₄⁺ may induce free radical production and astrocyte swelling in the brain.^35,36 EC has been identified as neurotoxic,³⁷ and OC contains polycyclic aromatic hydrocarbons that disrupt the immune system and thus damage our CNS.³⁸ These exposures upregulate proinflammatory cytokines (e.g., IL-1β, IL-6, and TNF-α),³⁹ disrupt the hypothalamic-pituitary-adrenal axis,⁴⁰ and alter cortisol regulation,⁴¹ leading to depression-like behaviors.⁴² Furthermore, depression and anxiety may occur as a secondary response to respiratory or cardiac conditions linked to PM_2.5 exposure.¹⁶

India’s exceptionally high PM_2.5 concentrations provide critical insight into exposure-response relationships less observable in lower-exposure regions. However, the limited overlap with air quality levels in Europe and North America warrants caution in extrapolating to cleaner areas. Although toxicological research showed that the key PM_2.5 components may exert neurobiological effects even at low concentrations, the slope and shape of the exposure-response curve may differ elsewhere. Multi-country analyses with harmonized exposure frameworks are necessary to assess global generalizability.

Subgroup analyses revealed stronger associations among middle-aged adults, the lowest-income groups, and urban metropolitan residents, reflecting a combined biological and social vulnerability. This pattern was echoed by previous studies,²⁸ with an inverted U-shaped age pattern in air pollution-related mental health risks, with middle-aged groups being more vulnerable, which may be due to cumulative life stressors, occupational exposures, and caregiving responsibilities. Urban metropolitan residents may be disproportionately affected due to high ambient exposure, compounded psychological stressors, lower quality of life, and disrupted sleep.⁴⁵ A population-based study from the United States reported that residents of highly polluted neighborhoods had higher distress than those in cleaner areas.⁴⁶ These findings also align with the socioeconomic gradient we observed, where lower-income groups showed stronger associations, indicating that disadvantaged people often face greater stressors, higher exposures, and reduced access to healthcare.²⁸ Further, educational attainment acts as a modifier, with illiterate individuals being susceptible to depression, and those with a high school education being most affected by anxiety. This disparity may be due to limited employment opportunities, reduced income, and diminished social support.⁴⁷ Social context also played a role, with individuals who were divorced, separated, or widowed showing stronger associations, which could be attributed to differences in social support and life stability.⁴⁷ Finally, regional heterogeneity was also observed, with depression being stronger in the Eastern region and anxiety in the Western region, possibly reflecting differences in pollutant composition, emission source apportionment, meteorological conditions, and population characteristics.^48,49 Seasonal modification was also observed, with stronger associations during the post-monsoon period, likely due to crop residue burning, festive emissions, and shifts in pollutant composition.⁵⁰

In the real world, individuals are exposed to multiple environmental stressors concurrently. Emerging evidence suggests that extreme heat, temperature variability, and other air pollutants contribute to physiological distress^43,44 and may interact synergistically with PM_2.5 through shared biological pathways, including oxidative stress, systemic inflammation, circadian and neuroendocrine disruption, and autonomic imbalance. Although our study focused on PM_2.5 compositions, future research should adopt multi-exposure frameworks to evaluate the combined and interactive effects of air pollutants, heat, and social stressors on mental health.

Our study has several strengths. This is the first nationwide study in India exploring the associations between ambient PM_2.5 components and mental health outcomes (depression and anxiety) using a large, population-based survey. The nationally representative survey, with its large sample size, provided us with adequate statistical power and enhanced the generalizability of the findings. Furthermore, the mental health outcomes, based on a structured diagnostic strategy considered equivalent to a clinical diagnosis, provide household-level analysis and overcome limitations that require diagnoses by specialists for specific outcomes.

Our findings have several important policy implications. First, considering PM_2.5 mass as an exposure metric may underestimate the “true” effect of air pollution in India; species-resolved evidence is crucial. Second, because the most toxic components map clearly onto specific sources, targeted emissions controls could yield greater health benefits. Bottom-up emission inventory-driven source attribution at the NMHS clusters (summarized in Figure 1) indicates that the domestic sector (solid fuel combustion) is a key source of total PM_2.5, highlighting the need for accelerated clean-fuel transitions and improved combustion efficiency. These high-risk components also relate to major emission sectors, with SO₄²⁻, NO₃⁻, and NH₄⁺ mainly affected by traffic, industry, and agriculture-related precursor emissions, and EC and OC largely originating from fossil fuel use and residential biomass burning. For secondary inorganic aerosols, coordinated precursor control of NO_x, NH₃, and SO₂ is critical, consistent with recent evidence that precursor-focused strategies substantially reduce PM_2.5 toxicity.^51,52 Third, the exposure-response curves indicate that exposure must be reduced below levels at which the slopes steepen to achieve significant health benefits. Fourth, socioeconomic status emerged as a modifier in the association between PM_2.5 components and depression and anxiety. Targeted interventions are crucial for protecting vulnerable groups.

In conclusion, our analysis is the first of its kind in India, providing strong epidemiological evidence that exposures to total PM_2.5 and its constituents are associated with depression and anxiety among adults, and the association varies with exposure to individual species. Importantly, the overall effect of air pollution on mental health outcomes may be underestimated if PM_2.5 mass is used as an exposure metric. Despite the limitations, we strongly believe that our interpretations will hold true in a general sense, contributing evidence to the discussion table. While we acknowledge and recommend cohort studies to establish causal pathways for these associations, we anticipate refined analysis with the ongoing NMHS-2 across all states and UTs, along with geo-tagged participant data, will provide a longitudinal framework to strengthen causal understanding of the mental health impacts of air pollution.

Limitations of the study

This study has several limitations that warrant cautious interpretation. First, the cross-sectional design restricts causal inference and prevents establishing temporal directionality. Second, we assigned exposure based on participants’ residential addresses and assumed that this was the primary exposure for each individual, which may not fully capture the personal exposure. We did not account for exposure to indoor pollution sources due to a lack of data, which is particularly relevant in India due to the widespread use of biomass fuel. Third, the PM_2.5 chemical compositions were estimated by combining modeled mass fractions of each species with satellite-derived PM_2.5 at 1 km resolution. Although this integration approach reduces spatial misalignment, it may introduce exposure misclassification if the mass fraction of chemical species varies sharply within a 36 km grid cell, for example, near dense industrial or traffic sources. Such exposure error is expected to be non-differential with respect to mental health outcomes and, therefore, may bias results toward the null. Fourth, even after controlling for multiple covariates, we could not fully rule out the possibility of unmeasured confounders, such as stress, physical activity, and substance use, which may increase outcome uncertainty. Fifth, although the MINI used in the NMHS is a validated and diagnostic instrument, interviewer bias cannot be completely excluded in a large-scale survey. To minimize this, investigators underwent rigorous, multi-level training under close supervision, followed by supervised field assessments, inter-rater evaluations, and 5% re-interviews for quality control. Because the MINI assessed current disorders (past two weeks for depression, past six months for generalized anxiety disorder, and past month for other anxiety disorders), substantial recall bias is unlikely, although some degree of recall error may persist, particularly for anxiety disorders. Finally, the exposure-response curves we observed were mostly nonlinear and non-monotonic; this may be due to residual confounding, exposure error, or true heterogeneity, which warrants further investigation in longitudinal cohort studies with repeated exposure measurements.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Sagnik Dey (sagnik@cas.iitd.ac.in).

Materials availability

This study did not generate new datasets.

Data and code availability

• •Deidentified NMHS participant data will be made available on the submission of the protocol by the investigator. The proposal will be reviewed by the Publication Oversight Committee, and the appropriate decision will be communicated. The signed access agreement adheres to the publication and dissemination policy of the National Mental Health Survey. All relevant reports are available from http://indianmhs.nimhans.ac.in. The data dictionary will be provided only for the approved proposals.
• •All original code has been deposited at figshare.com and is now publicly available. DOIs are listed in the key resources table.
• •Temperature datasets,⁵⁶ global NO₂ LUR data,⁵⁷ and global ozone data⁵⁸ are available from respective repositories cited in the article.
• •PM_2.5 datasets from Katoch et al.,⁵⁴ speciated PM_2.5 fractions from WRF-CMAQ,⁵⁵ and any additional information required to reanalyze the data reported in this article are available from the lead contact upon request.

Acknowledgments

The work is supported by a research grant from the Clean Air Fund to IIT Delhi. P.K. acknowledges the doctoral fellowship provided by the University Grants Commission. The National Mental Health Survey (NMHS) was funded by the Ministry of Health and Family Welfare, Government of India, and was implemented and coordinated by NIMHANS, Bengaluru, INDIA, in collaboration with state partners. The NMHS collaborators group includes Pathak K., Singh L. K., Mehta R. Y., Ram D., Shibukumar T. M., Kokane A., Lenin Singh R. K., Chavan B. S., Sharma P., Ramasubramanian C., Dalal P. K., Saha P. K., Deuri S. P., Giri A. K., Kavishvar A. B., Sinha V. K., Thavody J., Chatterji R., Akoijam B. S., Das S., Kashyap A., Ragavan V. S., Singh S. K., Misra R., and investigators as listed in the NMHS Summary report available from http://indianmhs.nimhans.ac.in. We express our sincere gratitude to all the respondents and their family members across the surveyed states for their cooperation in the conduct of the National Mental Health Survey. We also thank Dr. Andrew Larkin and Prof. Susan Anenberg for providing the NO₂ data and the reviewers for their insightful suggestions, which helped us to improve the article.

Author contributions

P.K. contributed to the conceptualization of the study, data curation, formal analysis, investigation, methodology, software, visualization, writing the original draft, and review and editing of the article. S.D. contributed to the conceptualization, data curation, methodology, supervision, and review and editing of the article. A.K. contributed to the conceptualization, methodology, supervision, and review and editing of the article. S.G. contributed to the formal analysis, methodology, software, and review and editing of the article. G.N., G.G., V.B., and M.V. contributed to data curation, investigation, and review and editing of the article. S.D., G.N., G.G., V.B., and M.V. have verified the underlying data. All authors critically revised the article and approved the final version.

Declaration of interests

We declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
National Mental Health Survey of India (NMHS 2015–16, Phase 1)	Gururaj et al.53	https://indianmhs.nimhans.ac.in/
Satellite-derived gap-filled PM2.5 (1 km × 1 km)	Katoch et al.54	https://doi.org/10.1021/acs.est.3c03355
PM2.5 speciated mass fractions (WRF v3.9.1 – CMAQ v5.3.1)	Singh et al.55	https://doi.org/10.1021/acsearthspacechem.1c00170
Global daily near-surface air temperature (1-km resolution)	Zhang et al.56	https://doi.org/10.5194/essd-14-5637-2022
Global daily Land Use Regression (LUR) model for NO2 at 50 m × 50 m	Larkin et al.57	https://doi.org/10.3389/fenvs.2023.1125979
Global land daily surface Ozone (10-km resolution)	Wang et al.58	https://doi.org/10.1038/s41597-025-05990-x
Software and algorithms
R software (v4.3.2)	R Core Team	https://www.r-project.org
mgcv package (v1.9-0)	Wood59	CRAN: https://cran.r-project.org/package=mgcv
gWQS package (v3.0.5)	Carrico et al.60	CRAN: https://cran.r-project.org/package=gWQS
Code to reproduce the results of this study	Authors (This Study)	https://doi.org/10.6084/m9.figshare.30774533

Experimental model and study participant details

Study design and population

Our analysis utilized data from the National Mental Health Survey of India (NMHS 2015-16),⁵³ the first nationwide representative survey on mental health burdens in India. NMHS-1 was conducted between October 2, 2015, and June 10, 2016, across 12 states (Punjab, Uttar Pradesh, Rajasthan, Gujarat, Assam, Manipur, Madhya Pradesh, Chhattisgarh, West Bengal, Jharkhand, Tamil Nadu, and Kerala), covering all six regions of the country. The survey aimed to estimate the prevalence, burden, and sociodemographic correlates of different mental disorders in India and identify treatment gaps, along with the disability associated with these disorders.

A multi-stage, stratified, random cluster sampling method was used to obtain national and state-level representativeness. Participants aged 18 and older were interviewed through door-to-door surveys. According to the 2011 India census, each inhabited village represented a rural cluster, while each census enumeration block represented an urban cluster (metro or non-metro). The number of rural and urban clusters was selected randomly using the probability proportional to size (PPS). From each cluster, 15 households were selected systematically, and all eligible residents who had lived in the household for at least six months were interviewed after obtaining informed consent.

Psychiatric disorders were assessed electronically using handheld tablets. In total, 39,532 eligible adults were contacted, and 34,802 (88%) were interviewed from 9,666 households across 720 clusters in 80 taluks of 43 districts across 12 states in India. The survey instruments were validated at four levels: (1) at NIMHANS with clinical outpatients, (2) during the pilot study, (3) before the main survey, and (4) through reinterview of 5% of participants by the respective state team. Individual participants within the selected household were considered the unit of analysis. Additional methodological details are available in the NMHS 2015–2016 technical report and elsewhere.^61,62,63

Method details

Exposure assessment

We estimated ambient PM_2.5 mass and its major chemical components using a hybrid approach. We first employed high-resolution (1 km × 1 km) satellite-derived PM_2.5 concentrations using the validated dataset developed by Katoch et al. (2023).⁵⁴ Briefly, the surface-level PM_2.5 concentrations were derived from the daily aerosol optical depth (AOD) product retrieved at a 1 km by 1 km spatial resolution by the Moderate Resolution Imaging Spectroradiometer (MODIS). However, spatial gaps arising from cloud cover and high surface albedo were addressed using an extreme gradient boosting algorithm. This machine-learning approach integrated various meteorological, land-use, and emission-related predictors, including temperature, relative humidity, surface pressure, precipitation, cloud fraction, albedo, boundary layer height, wind components, fire counts, vegetation indices, elevation, population density, and road proximity, to convert gap-filled AOD data into daily PM2.5 estimates at the same spatial scale. Comparisons with coincident measurements from the Central Pollution Control Board (CPCB) ground monitoring stations confirmed the high quality of the satellite-PM_2.5 data across all seasons, based on cross-validation results (R² = 0.86; RMSE = 23.46 μg m⁻³) on a daily scale.⁵⁴

Next, we estimated PM_2.5 components by integrating satellite-derived PM_2.5 data with speciated mass fractions simulated by the WRF v3.9.1–CMAQ v5.3.1 modeling framework, as described in Singh et al. (2021).⁵⁵ The chemical transport model was run at 36 km × 36 km resolution with 25 vertical layers using ERA5 meteorology, and the GAINS-ASIA sectoral emission inventory, based on government-reported sectoral energy consumption data for 2016. Additional emissions, including ammonia, shipping, and transboundary contributions from the neighboring countries like Bangladesh, Bhutan, Nepal, Myanmar, Pakistan, Sri Lanka, and parts of Afghanistan and China, were provided by the ECLIPSE (version 5) inventory, and boundary conditions were obtained from the CAM-chem global model. As reported in Singh et al. (2021),⁵⁵ the simulated PM_2.5 concentrations showed strong agreement against ground-based observations (coefficient of determination, R² = 0.81; index of agreement = 0.94).

We estimated the daily mass concentration of each component i (M_i) using the scaling methodology following our previous work²⁴:

$$M_i=\left(\frac{M_{i,model}}{{PM}_{2.5,model}}\right)\times {PM}_{2.5,satellite}$$ (Equation 1)

Where M_{i, model} denotes the modeled mass of PM_2.5 components ‘i’, PM_{2.5, model} is the modeled total PM_2.5 mass, and PM_{2.5, satellite} is satellite-derived PM_2.5 concentration. To integrate the model outputs with the 1 km × 1 km satellite-derived PM_2.5 data, we first bilinearly regridded the 36 km modeled mass fraction to a 1 km × 1 km satellite grid. These interpolated mass fractions were then multiplied by the satellite PM_2.5 estimates to produce daily mass concentrations of each component at 1 km resolution. This integration strategy is based on the established understanding that the PM_2.5 chemical component is spatially more homogenous than total PM_2.5 mass, making it appropriate to apply coarse-resolution speciation fractions to high-resolution satellite-derived PM_2.5.

For the present study, we first aggregated the 1 km daily PM_2.5 estimates to the survey cluster level to construct area-specific long-term PM_2.5 exposure for both mental health outcomes (depression and anxiety). We then linked the full 1 km × 1 km daily PM_2.5 mass and components concentrations to the geocoded residential addresses of each NMHS participant across 12 Indian states. Using these data, we calculated the one-year average PM_2.5 mass and each chemical component before each participant’s interview date, which served as the temporal anchor for exposure assignment. This one-year exposure metric is an indicator of long-term exposure, consistent with the chronic nature of depression and anxiety. All exposure metrics were assigned based on participants’ residential coordinates.

Assessment of depression and anxiety

Depressive and anxiety disorders were diagnosed using the Mini-International Neuropsychiatric Interview (MINI)-Adult Version 6.0.0 (October 2010), a structured diagnostic interview based on International Classification of Diseases, Tenth Revision (ICD-10) diagnostic criteria for research purposes.⁶⁴ The MINI was administered by trained non-clinical interviewers who underwent extensive training and supervision by qualified mental health professionals to ensure diagnostic reliability across states. Using the reporting of participants’ behaviors and symptoms, the MINI algorithm generates diagnoses compatible with clinical diagnoses. It is therefore a structured diagnostic instrument, different from symptom checklists or self-reported symptom lists, or a screening instrument.

For this study, diagnostic criteria for current disorders (depression and anxiety) at the time of the survey interview were determined using the structured module of MINI. Participants categorized as currently experiencing depression were those who met ICD-10 diagnostic criteria for a depressive episode or recurrent depressive disorder (F32-F33), encompassing mild, moderate, or severe episodes with or without psychotic symptoms. Whereas, participants identified as currently experiencing anxiety were those diagnosed with any one of the following anxiety disorders: agoraphobia (F40.0), social phobia (F40.1), panic disorder (F41.0), generalized anxiety disorder (GAD, F41.1), obsessive-compulsive disorder (OCD, F42.0-F42.8), and post-traumatic stress disorder (PTSD, F43.1).⁶³ All diagnoses were coded dichotomously, with ‘1’ indicating their presence and ‘0’ denoting their absence.

Covariates

NMHS provided a sociodemographic questionnaire to collect information on several individual and household-level variables, adapted from the 2011 India Census household questionnaire,⁶¹ which were identified as potential confounders in our analysis based on prior literature.^{10,13,16,17,29,30}

The following individual-level variables were considered for the health outcome: age group (“18–29”, “30–39”, “40–49”, “50–59” and “60 and above”), gender (“Male,” “Female”), education (“Higher Education,” “High School,” “Secondary,” “Primary,” and “Illiterate”), and marital status (“Married,” “Never Married,” and “Widowed/Divorced/Separated”). The household-level variable was household income, which is categorized into five quintiles (“Highest,” “Fourth,” “Middle,” “Second,” and “Lowest”), and the cluster-level variable was the resident type (“Rural,” “Urban Metro,” and “Urban Non-Metro”). Additionally, we incorporated the seasonal category following the definition set by the India Meteorological Department into the analysis at the time of the survey. Due to the limited number of observations during the monsoon season (Jun-Sep), we combined it with the post-monsoon (Oct-Nov) season to ensure that all seasons have a roughly similar number of observations, thereby reducing sampling bias.

We obtained daily maximum and minimum temperature data with a spatial resolution of 1 km × 1 km⁵⁶ to account for any potential confounding effect related to heat. We calculated the diurnal temperature range (DTR) as the difference between the daily maximum and minimum temperatures. To be consistent with the main exposure data, we calculated one-year average maximum, minimum temperature, and DTR for all clusters before the NMHS interview and then adjusted them into the model. All these variables were included in the analysis based on their significance with their respective health outcomes. Further, we considered region (“Central,” “Eastern,” “North-Eastern,” “Northern,” “Southern,” and “Western”) as a potential effect modifier on depression and anxiety, along with the other relevant covariates, for each IQR rise in ambient PM_2.5 exposure were further examined by stratified analysis. Additionally, as a sensitivity analysis, we also estimate ambient nitrogen dioxide (NO₂)⁵⁷ and ozone (O₃)⁵⁸ concentrations to account for potential confounding by co-pollutants. Similar to the PM_2.5 metrics, we linked daily NO₂ and O₃ estimates to each participant’s geocoded residential address and calculated one-year averages before the interview date.

Quantification and statistical analysis

Weighted prevalence estimates for current depressive and anxiety disorders were derived using survey weights to increase national representativeness and to correct for potential bias due to unequal probability of selection inherent in the multi-stage PPS sampling design of the NMHS. Sampling weights were derived as the product of design weights (based on district and taluka selection probabilities) and the individual response rate weight. A detailed explanations for weight estimation are provided in the NMHS report and elsewhere.^61,65 All prevalence estimates are reported with 95% CI.

We employed logistic mixed-effects regression models to examine the association between air pollution exposure and mental health outcomes (depression and anxiety) among Indian adults. All models incorporated individual-level sampling weights to account for the complex survey design and included random effects at both the cluster and household levels to address hierarchical clustering. This modeling strategy was used because multiple individuals were sampled within households, which makes intra-household correlation non-negligible. We compared alternative model structures (only cluster-level and only household-level random effects) and found that effect estimates were stable, but the model including both levels of clustering provided the best model fit based on Akaike Information Criterion, ensuring the most appropriate variance estimation.

The model was run to estimate the direct association between each individual PM_2.5 component and mental health outcomes while adjusting for potential covariates. To address potential multicollinearity among PM_2.5 components, we used a two-stage model approach, as described in our previous work.²⁴

Stage 1: For each component, total PM_2.5 mass was regressed on an individual component, and the residual PM_2.5 was extracted, representing the remaining variability of total PM_2.5 mass not explained by the component. This isolates the independent association of each component while eliminating the influence of other members of the total PM_2.5 mass.

$${\left({ResidualPM}_{2.5}\right)}_i={PM}_{2.5}-{\hat{\delta }}_0-\widehat{{\delta }_1}{X}_i$$

Where X_i is the individual PM_2.5 component, and ${\hat{\delta }}_0$, ${\hat{\delta }}_1$ represents the intercept and slope estimates from the first-stage regression model.

Stage 2: We fitted a logistic mixed-effects model to assess the relationship between the PM_2.5 component and binary health outcomes (depression and anxiety) adjusted for residual PM_2.5 components, clustering, sampling weights, and relevant confounders.

$$logit\left\{\mathit{Pr}\left(Y_{ij}=1|u_i\right)\right\}={\beta }_0+{\beta }_1{X}_i+{\beta }_2{\left({ResidualPM}_{2.5}\right)}_i+\sum _k{\gamma }_k{C}_{ijk}+b_i$$

where Y_ijrepresented the binary outcome (depressive and anxiety disorders) for the j^th individual in i^th cluster, ${\left({ResidualPM}_{2.5}\right)}_i$ is the PM_2.5 residual from stage I at i^th cluster, C_ijk represented k^th individual and household-level confounders for j^th individual in i^th cluster, and b_i is the i^th cluster-specific random intercept with $b_i\sim N\left(0,{\sigma }_b^2\right)$. The results from the models were reported as odds ratios (ORs) with 95% confidence intervals (CIs) per interquartile range (IQR) increase in pollutant concentration.

To identify subpopulations that might be particularly susceptible, we performed subgroup analyses for one-year exposure windows based on age group, education, marital status, wealth quintiles, residence type, season of data collection, and region by interaction regression model technique. The statistical significance of effect modification was assessed by testing the regression coefficient corresponding to the interaction term as zero or different from zero. Stratified estimates were reported only for those coefficients of interaction that were statistically significant. We fitted the penalized cubic smoothing spline of the pollutant in the second-stage model to examine the potential non-linear association with mental health outcomes. The models were adjusted for the same covariates as our main model.

To estimate the “cumulative association” of PM_2.5 components on mental health outcomes, we computed a composite regression coefficient by summing the individual component-specific estimates, weighted by their respective mass fractions:

$${\hat{\beta }}_{cum}=\sum _{i=1}^k{m}_i{\hat{\beta }}_i$$

where ${\hat{\beta }}_{cum}$ is the cumulative regression coefficient, m_i is the mass fraction of the i^th component, and ${\hat{\beta }}_i$ is the estimated regression coefficient for the i^th PM_2.5 component. The standard error of ${\hat{\beta }}_{cum}$ is calculated assuming independence among the regression coefficients of each component:

$$V\left({\hat{\beta }}_{cum}\right)=\sum _{i=1}^k{m}_i^{2}V\left({\hat{\beta }}_i\right)$$

Some sensitivity analyses were performed to test the robustness of our results. First, to evaluate the potential confounding by gaseous pollutants, we re-estimated our second-stage models after additionally adjusting for ambient NO₂ and O₃. These gaseous pollutants were included only in sensitivity analyses and not in primary models. Second, as PM_2.5 is a composite mixture of multiple species, it often shows strong correlations with itself. We used weighted quantile sum (WQS) regression to estimate the joint mixture effect of the PM_2.5 component for each decile increase, while identifying the relative contribution of each component within the PM_2.5 mixture. The WQS models were run separately for depression and anxiety with 100 bootstrap samples, and all the weights were set to be positive.

All the statistical analyses were conducted in R, version 4.3.2, using the “mgcv” (version 1.9–0)⁵⁹ and “gWQS” (Version 3.0.5)⁶⁰ R packages, with a p value of <0.05, are considered statistically significant.

Published: January 29, 2026