Complements, Not Substitutes? Health Systems, Institutional Quality, and the PM2.5–NCD Mortality Nexus—A Global Panel Analysis

Abstract

This study investigates whether Universal Health Coverage (UHC) and institutional quality—proxied by the Corruption Perceptions Index (CPI)—moderate the association between ambient PM2.5 exposure and adult non-communicable disease (NCD) mortality. Using a country-year panel of 157 countries over 2012–2021, we estimated two-way fixed-effects panel regressions with Driscoll–Kraay standard errors. A consistency check confirmed the expected positive association between PM2.5 and air-pollution-attributable mortality (β = 0.247, p = 0.002). However, the direct effect of PM2.5 on NCD mortality was not significant in any specification. Critically, the interaction terms PM2.5 × UHC (β = 0.000485, p = 0.215) and PM2.5 × CPI (β = 0.000838, p = 0.067) were not statistically significant, indicating that neither health-system capacity nor institutional quality significantly moderates the PM2.5–NCD mortality relationship at the global level. These findings contrast with prior studies documenting moderating effects in specific contexts—under-five mortality and West Africa—suggesting that such effects may be context-dependent rather than universally applicable. While UHC showed a positive association with NCD mortality (β = 0.056, p < 0.001) likely reflecting improved diagnosis, reporting, and broader development processes, CPI was negatively associated (β = −0.033, p = 0.029), suggesting that higher institutional quality is associated with lower NCD mortality. Our findings are consistent with the view that emissions reduction and health-system strengthening should be pursued as complements rather than substitutes, implying a dual strategy for low- and middle-income countries: invest in health systems and institutional quality while simultaneously pursuing cleaner energy, transport, and industrial policies.

1. Introduction

Ambient fine particulate matter (PM2.5) is consistently ranked among the leading environmental risk factors for premature death worldwide, with the Global Burden of Disease framework attributing millions of annual deaths to long-term exposure through cardiovascular disease, ischemic heart disease, stroke, and chronic respiratory illness [1]. Yet the translation of a given level of ambient exposure into realised mortality is not mechanical. The same concentration of PM2.5 can produce very different health consequences depending on whether a population has timely access to diagnosis, chronic-disease management, and emergency care, and depending on whether public institutions can enforce air-quality standards, allocate health resources efficiently, and resist the diversion of public funds intended for health systems.

This investigation asks a deceptively simple question: do health-system capacity and institutional quality moderate the relationship between ambient PM2.5 exposure and non-communicable disease (NCD) mortality across countries? If health-system and institutional quality act as buffers, then policies that strengthen Universal Health Coverage and reduce corruption should be understood not only as health-sector reforms in their own right but also as elements of environmental-health policy—complementary to, rather than substitutes for, emissions reduction.

The question matters for the scope of this journal because it sits precisely at the intersection of environmental sustainability, public health, and governance that defines the social and economic dimensions of sustainable development. It also matters for policy: if the moderating effect is strong, then investment in health-system resilience is a cost-effective complement to costly and slow-to-implement emissions reductions, particularly for low- and middle-income countries that cannot rapidly decarbonize their energy and transport systems.

We use a country-year panel of up to 208 countries, with a final analytical sample of 157 countries after accounting for missing data, over 2012–2021, drawn from the Sustainable Development Report/SDG Index database [2] and combined with fixed-effects panel regression to test three hypotheses: (H1) a consistency check to verify that our PM2.5 indicator and data produce the expected positive association with air-pollution-attributable mortality (tested separately to avoid circularity with the GBD model); (H2) that the association between PM2.5 and NCD mortality is moderated by Universal Health Coverage (UHC); and (H3) that this moderation also operates through institutional quality, proxied by the Corruption Perceptions Index (CPI), as a robustness and mechanism check. Contrary to expectations derived from prior studies in specific contexts, our results indicate that neither UHC nor CPI significantly moderates the PM2.5–NCD mortality relationship, suggesting that the buffering effect is context-dependent rather than universal.

2. Conceptual Background

We review the existing literature on the relationship between ambient PM2.5 exposure and non-communicable disease (NCD) mortality and examine the theoretical and empirical basis for hypothesising that health-system capacity and institutional quality may moderate this relationship. The review proceeds in five sections. First, we synthesise the epidemiological evidence establishing PM2.5 as a risk factor for NCD mortality, paying particular attention to the disease categories most robustly linked to chronic exposure. Second, we examine the literature on health systems and institutional quality as determinants of health outcomes, distinguishing between direct effects on population health and potential buffering effects. Third, we review the sparse but growing body of evidence on the moderating role of health-system and institutional capacity in the environmental health nexus. Fourth, we develop the theoretical framework that informs the study’s hypotheses, drawing on institutional economics, the weak-versus-strong sustainability debate, and health-systems resilience theory. Fifth, we propose the study’s hypotheses, drawing on institutional economics, the weak-versus-strong sustainability debate, and health-systems resilience theory.

2.1. PM2.5 and Non-Communicable Disease Mortality: Epidemiological Foundations

A large and methodologically diverse epidemiological literature establishes the causal pathway between long-term exposure to ambient fine particulate matter (PM2.5) and mortality from cardiovascular and respiratory diseases. The Global Burden of Disease framework attributes millions of annual deaths to PM2.5 exposure, with cardiovascular diseases—particularly ischaemic heart disease and stroke—accounting for the majority of this burden [1]. Studies from multiple world regions have consistently reported positive associations between long-term PM2.5 exposure and all-cause and cardiovascular mortality. The ELAPSE project, pooling data from eight European cohorts, confirmed these associations even at low pollution levels, suggesting no safe threshold for exposure [3,4].

Systematic reviews have further refined understanding by examining the differential toxicity of PM2.5 chemical constituents. A recent meta-analysis found that short-term exposure to carbon-containing constituents (black carbon and organic carbon) and metal iron showed the most consistent associations with mortality and morbidity outcomes related to all-cause, natural, cardiovascular, and respiratory diseases among different PM2.5 constituents. Meanwhile, only long-term exposure to nitrate ions and silicon in PM2.5 showed significant positive associations with cardiovascular and natural mortality outcomes. Additionally, subgroup and meta-regression analyses highlight that geographical region and age group are primary sources of heterogeneity [5].

The pathophysiological mechanisms by which PM2.5 exposure results in adverse cardiovascular outcomes have also been examined. As reviewed by Krittanawong et al. [6], PM2.5 is hypothesised to function via three primary mechanisms: increased oxidative stress, activation of the inflammatory pathway of the immune system, and stimulation of the autonomic nervous system. These pathways ultimately promote endothelial dysfunction, atherosclerosis, and systemic inflammation, which can lead to cardiovascular events [6].

For the purposes of this study, several features of this literature are noteworthy. First, the epidemiological evidence is strongest for cardiovascular and respiratory mortality, which are the disease categories that constitute the bulk of adult NCD mortality and the focus of our outcome measure. Second, effect sizes are moderate and vary systematically across populations, suggesting that exposure–response relationships may be modified by contextual factors—including, potentially, the capacity of health systems to manage chronic disease. Third, much of the evidence derives from cohort studies in high-income settings; generalisability to low- and middle-income countries, where PM2.5 levels are often higher and health systems weaker, remains an important research frontier.

2.2. Health Systems and Institutional Quality as Determinants of Health Outcomes

Health-system capacity and institutional quality are widely associated with population health outcomes in the global health literature. Universal Health Coverage (UHC) is defined as ensuring that all people receive the health services they need, of sufficient quality, without financial hardship, and SDG indicator 3.8.1 tracks progress through a service coverage index built from tracer indicators across maternal and child health, infectious disease, non-communicable diseases, and service capacity and access [7,8]. This index is strongly associated with broader measures of population health; across countries, higher UHC service coverage is linked to higher life expectancy even after adjustment for national income and education [8]. More recent work using an effective coverage framework also argues that measuring whether services are aligned with need and of sufficient quality is central to understanding health-system performance under UHC [7].

Evidence for the health benefits of UHC is strongest for maternal and child health and for communicable disease control. In 60 low- and middle-income countries, a one-unit increase in the UHC index was associated with a 1.2% lower risk of infant death, and an estimated 15.5 million infant deaths were averted between 2000 and 2019 because of UHC gains [9]. Summary coverage indices in maternal and child health are also inversely correlated with under-five and neonatal mortality, supporting the value of broad service coverage as a determinant of survival [10]. At the same time, UHC progress does not automatically eliminate inequality: coverage among the poorest quintiles was 1% to 66% lower than national averages in 52 countries, and recent LMIC evidence suggests that later stages of UHC expansion can benefit richer groups more unless lower-income populations are explicitly prioritised [8,9].

The relationship between UHC and NCD mortality is more complex. Cross-country evidence suggests that higher UHC service coverage is associated with lower total DALYs overall, but that this association is concentrated in communicable, maternal, neonatal, and nutritional conditions, with little to no significant association for non-communicable diseases or injuries [11]. Likewise, the GBD 2019 effective coverage analysis found that many countries perform poorly on NCD-related effective coverage indicators relative to maternal, child, and communicable conditions, despite NCDs accounting for a growing share of potential health gains [7]. This means that cross-sectional associations between stronger UHC performance and population-level NCD mortality are not straightforward causal estimates [8].

A key reason is the epidemiological transition. As countries reduce premature mortality from infectious and early-life conditions and improve survival, populations age and a larger share of deaths shifts toward chronic non-communicable diseases [12]. In the Americas, reductions in DALY rates between 2000 and 2019 would have lowered the disease burden, but these gains were offset by a 28% increase due to population ageing and a 22% increase due to population growth [13]. Similar concerns are reported for low- and middle-income settings, where rapid population ageing increases the prevalence of NCDs and disability while many health systems still face a persistent burden of infectious disease, creating a double burden that strains service capacity [14]. Country examples such as India and Nigeria reflect this pattern: both are described as facing rising NCD burdens alongside continuing communicable disease challenges and important gaps in data, primary care integration, and health-system readiness [14].

Institutional quality is also consistently associated with better health outcomes. In a panel of 158 countries from 2001 to 2020, institutional quality measured through control of corruption, voice and accountability, political stability, rule of law, regulatory quality, and government effectiveness was associated with lower infant mortality and higher life expectancy, even alongside GDP, education, health expenditure, and urbanisation [15]. A separate panel study of 105 countries similarly found that stronger institutional quality increased life expectancy and reduced infant mortality [16]. In West Africa, governance quality across 15 countries was significantly associated with improved life expectancy and lower infant, under-five, and maternal mortality, with regulatory quality, government effectiveness, and control of corruption appearing especially influential [17]. Related evidence from developing countries also links better governance indicators to improved quality of life more broadly [18].

Several mechanisms help explain why institutional quality matters for health. Governance affects how effectively states make and implement policy, allocate budgets, regulate systems, and manage accountability, transparency, participation, integrity, and capacity [19]. Better governance appears to improve health partly by making public spending more effective and more equitably distributed [20]. It also matters for corruption control and for aligning stakeholders and institutions around UHC goals [19]. WHO-oriented governance work therefore treats governance as a core health-system function rather than a peripheral political variable.

Most of the direct evidence on institutional quality and health still focuses on infant mortality, maternal mortality, and life expectancy rather than adult NCD mortality [15,16,20]. Even so, the available evidence supports the view that governance likely matters for NCD outcomes through different pathways, including chronic disease management, primary care organisation, effective regulation, and the capacity to adapt services to ageing populations and changing disease profiles [7]. Overall, the literature supports the claim that stronger health systems and better institutions are associated with better population health, but the specific relationship between UHC and NCD mortality should be interpreted cautiously because it is shaped by epidemiological transition, population ageing, disease composition, and the limits of cross-sectional comparison [7,12,13].

2.3. The Moderation Hypothesis: Buffering Environmental Health Risks

The evidence supports a moderation hypothesis in which health-system capacity conditions how environmental exposures translate into observed health outcomes, rather than affecting health only through direct pathways [21,22]. This framing is also consistent with institutional theory, which treats institutions as durable constraints that shape how societies respond to risks and allocate resources [21,23].

In a panel analysis of 145 countries from 2009 to 2017, environmental degradation was linked to worse health outcomes, and air pollution exposure was significantly associated with higher mortality and DALYs [21,22]. The same study found that stronger healthcare infrastructure, accessibility, and quality helped mitigate some pollution-related risks, but these improvements were insufficient to offset the long-term adverse effects of CO₂ emissions [21]. The interaction terms between CO₂ emissions and healthcare resilience were statistically insignificant, which suggests that even well-functioning health systems cannot fully counteract the health consequences of environmental degradation. This implies that health-system strengthening should be understood as a complement to, not a substitute for, environmental regulation and emissions reduction.

A related result appears in West Africa, where an ecological panel study of 11 countries from 2015 to 2020 tested whether health expenditure modified the association between ambient PM2.5 and premature male NCD mortality [22,23]. Mean PM2.5 exposure in that sample was 60.3 µg/m³, more than twelve times the WHO guideline level, and mean premature NCD mortality was 23.4% [22,23]. In pooled models, higher PM2.5 was associated with lower NCD mortality, but this association disappeared after country fixed effects were introduced, indicating substantial confounding by country-level development factors rather than any protective effect of pollution [22,23]. Stratified analyses showed that the negative PM2.5–mortality association was significant only in low-expenditure settings and attenuated to null in higher-expenditure contexts, supporting effect modification by health-system capacity in this regional sample. However, the same study found no evidence that health expenditure acted as a mediator, so the supported claim is modification rather than mediation.

The hypothesis that health-system capacity may moderate the health effects of environmental exposure is grounded in the health-systems resilience literature [21]. Three mechanisms are plausible. First, health systems with strong primary care can detect early signs of pollution-related cardiovascular and respiratory disease—such as hypertension or reduced lung function—through routine screening, enabling earlier intervention. Second, effective chronic disease management—including access to antihypertensives, lipid-lowering agents, and asthma/COPD medications—can reduce mortality among those already exposed. Third, emergency and acute care capacity can reduce case fatality from acute pollution-triggered events such as myocardial infarction and stroke. These mechanisms are not specific to child health or to West Africa; they are general features of health systems that could plausibly operate across contexts. However, whether they do so in practice, and whether the broad UHC service coverage index adequately captures these capacities, is an empirical question that our study addresses.

In any case, we acknowledge that the UHC service coverage index is a broad, aggregate measure that includes tracer indicators for maternal and child health, infectious diseases, and non-communicable diseases [7,8]. While it is the most comprehensive and harmonised global measure available for a panel spanning 157 countries over a decade, it may not capture NCD-specific service capacity with the precision that would be ideal. For example, it does not separately measure the availability of antihypertensive medications, the density of cardiologists, or the quality of cancer screening programmes. This is a limitation, which we discuss in Section 5.4. Nevertheless, if UHC service coverage—as a broad measure of health-system capacity—does not moderate the PM2.5–NCD mortality relationship, it is unlikely that more specific NCD-focused measures would show a strong buffering effect unless health systems are specifically oriented towards chronic disease management.

2.4. Research Gaps and Positioning of the Present Study

Despite growing interest in the moderation hypothesis, the empirical evidence remains limited in several important respects. The clearest gap concerns adult NCD mortality: the available evidence directly examining this outcome is regionally restricted, with one ecological panel study focusing on premature NCD mortality among males aged 30–70 years in 11 West African countries between 2015 and 2020 [22]. By contrast, the broader global evidence examines pollution-related mortality and DALYs across countries, rather than adult NCD mortality specifically [21].

A second gap concerns geographic scope and external validity. The West African study provides evidence that health-system capacity can modify the observed association between ambient PM2.5 and premature NCD mortality, but its regional ecological design limits the extent to which that result can be generalised beyond the countries studied [22]. The global study covers 145 countries and explicitly tests whether stronger healthcare systems reduce pollution-induced mortality and DALYs, but it does so using broader health outcomes rather than disease-specific adult NCD mortality [21].

Current evidence also does not clearly separate health-system capacity from broader institutional quality within the same empirical framework. The environmental health studies supplied here test moderation using healthcare-system indicators or health expenditure measures rather than a parallel institutional-quality moderator [21,22]. Institutional theory provides a basis for expecting broader governance structures to shape how societies absorb and respond to shocks, because institutions influence uncertainty, transaction costs, and long-run performance, but this contribution is conceptual rather than an empirical test of environmental-health moderation [23].

These gaps position the present study to make a more specific contribution. It can extend the literature by testing the moderation hypothesis on adult NCD mortality in a broader cross-national setting and by distinguishing whether any buffering effect operates through the health-system channel or through a more general institutional channel. In this way, the study speaks directly to a literature that currently combines global analyses with non-specific outcomes and regional analyses with narrower outcome specificity and limited generalisability.

Our study is therefore designed as a test of generalisability of previously documented context-specific effects, rather than assuming their universality. This framing acknowledges that the moderation effect may be context-dependent, and that testing its external validity is a contribution in its own right. In the absence of such tests, policy recommendations derived from context-specific studies risk being overgeneralised to settings where they may not hold. A null result at the global level is therefore meaningful: it prevents overgeneralisation and highlights the need for context-specific policy guidance.

2.5. Hypotheses

The theoretical framework for this study posits that the relationship between PM2.5 exposure and mortality is not purely mechanical, but can be conditioned by health-system and institutional capacity. This expectation is consistent with institutional theory, which treats institutions as durable constraints that shape performance by influencing uncertainty, transaction costs, and the organisation of collective action [23]. Applied to environmental health, this logic implies that countries with stronger health systems and more effective institutional arrangements may be better able to detect, manage, and treat pollution-related disease burdens, thereby attenuating the health consequences of a given level of exposure [21,22].

At the same time, the framework recognises limits to buffering. Strong-sustainability arguments hold that some environmental functions cannot be readily replaced by technological systems and that medium-term limits to substitution are real [24]. Consistent with this view, global panel evidence indicates that better healthcare infrastructure, accessibility, and quality can mitigate some pollution-related risks, but are insufficient to offset the long-term adverse effects of environmental degradation in all cases [21].

This tension between buffering and its limits motivates the hypotheses tested in this study. H1: PM2.5 concentration is positively associated with the age-standardised death rate attributable to air pollution within countries over time. This serves as a consistency check for the exposure indicator and is consistent with cross-national evidence linking air pollution exposure to higher mortality and DALYs [21]. H2: The association between PM2.5 and adult NCD mortality is weaker in settings with greater health-system capacity. This hypothesis is grounded in evidence that healthcare systems can mitigate some pollution-related risks globally [21] and that health-system capacity modifies the observed PM2.5–premature NCD mortality association in West Africa [22]. However, the generalisability of this effect to adult NCD mortality at the global level has not been tested, and our study examines whether this pattern holds in a broader cross-national setting. H3: The attenuating pattern should also appear when broader institutional quality is used as a moderator, if buffering reflects state capacity beyond the health sector alone. This hypothesis is theoretically grounded in institutional economics [23], though the empirical evidence base is more limited than for H2; we therefore present it as a plausible extension worth testing rather than as a strongly established empirical expectation.

3. Materials and Methods

3.1. Data

We constructed a country-year panel dataset spanning 2012–2021, drawing on variables from the Sustainable Development Report 2024 database [2]. This temporal window was selected for two reasons. First, the Corruption Perceptions Index (CPI) underwent a methodological revision in 2012, rendering earlier years non-comparable with subsequent observations. Second, WHO-reported NCD mortality data in the SDG Index database are complete through 2021 in the most recent release, with later years not yet available. The resulting unbalanced panel comprises 208 countries and up to 2080 potential country-year observations, subject to the missing-data patterns detailed in Section 3.3.

All variables, their definitions and coding are summarised in Table 1.

Table 1. Variables and definitions/coding sources.

Role	Variable	Definition/Coding
Independent (focal)	PM2.5	Annual mean concentration of fine particulate matter (μg/m3), country-year
Dependent (H1, consistency check)	Air-pollution mortality	Age-standardised death rate attributable to air pollution (per 100,000)
Dependent (H2, H3—focal outcome)	NCD mortality	Age-standardised probability of dying from cardiovascular disease, cancer, diabetes, or chronic respiratory disease between ages 30–70 (%)
Moderator (H2)	UHC index	Universal Health Coverage index of service coverage (0–100)
Moderator (H3)	CPI	Corruption Perceptions Index (0–100, higher = cleaner)
Control	Gov. health/education spending	General government expenditure on health and education (% of GDP)
Control	Poverty ($3/day)	Poverty headcount ratio at $3.00/day (%), proxy for income level
Control	Electricity access	Percentage of population with access to electricity
Control (robustness)	Urban slums	Percentage of urban population living in slums

Note: PM2.5: fine particulate matter; NCD: non-communicable disease; UHC: Universal Health Coverage; CPI: Corruption Perceptions Index; SDG: Sustainable Development Goal.

The UHC service coverage index aggregates 14 tracer indicators across four categories: reproductive, maternal, newborn and child health; infectious diseases; non-communicable diseases; and service capacity and access [7,8]. While it is the most comprehensive and harmonised global measure of health-system coverage available for a panel of this scope, its composition includes indicators beyond NCD-specific services.

The Corruption Perceptions Index (CPI) is a perception-based measure of public sector corruption, compiled by Transparency International. It is the most widely used cross-national measure of corruption and has consistent annual coverage across our panel window (2012–2021). However, we acknowledge its limitations as a subjective indicator; alternative objective measures of governance effectiveness, such as the Worldwide Governance Indicators, are not available in the SDG Index database.

We acknowledge that the 10-year window does not allow for meaningful lagged exposure specifications (e.g., 5- or 10-year lags), which would be preferable for capturing the cumulative effects of chronic PM2.5 exposure on NCD mortality. However, longer panels are not feasible with the current data due to the CPI methodology change in 2012 and the availability of harmonised NCD mortality data. Future research with longer panels should explore lagged exposure specifications.

3.2. Outcome Variables

Our study employs two distinct dependent variables, serving different analytical purposes. The air-pollution-attributable mortality rate—an age-standardised death rate (per 100,000) estimated by the Institute for Health Metrics and Evaluation (IHME) within the Global Burden of Disease (GBD) framework—is used exclusively for a consistency check (H1). This variable is itself the product of a comparative risk assessment model that embeds an assumed exposure–response function between PM2.5 and mortality [1]. Using it as the dependent variable to test whether that relationship is moderated by health-system or institutional quality would be partially circular, as any moderating effect would need to have been incorporated into the upstream GBD model to manifest in the output. We therefore restrict its use to confirming that our PM2.5 indicator captures the expected health risk.

For the focal analyses (H2 and H3), we use the age-standardised NCD mortality rate—defined as the probability of dying between ages 30 and 70 from cardiovascular disease, cancer, diabetes, or chronic respiratory disease, expressed as a percentage [10]. This variable is derived from WHO vital registration and epidemiological surveillance data, and is not pre-adjusted for institutional or health-system characteristics. It thus provides an appropriate outcome measure for testing whether UHC and CPI moderate the PM2.5–NCD mortality relationship.

3.3. Missing Data

Some country-year observations were excluded due to missing values in the key dependent and independent variables required for the main models. Specifically, observations with missing values on ncd_mortality_pct, pm25_ugm3, uhc_index, corruption_perceptions_index, gov_health_edu_spend_pctgdp, or electricity_access_pct were excluded from the analytical sample. This listwise deletion resulted in an initial analytical sample of 1610 country-year observations across 165 countries for the descriptive statistics (Table 2). For the regression models, which include log_poverty as a control, additional observations with missing poverty data were excluded, resulting in a final analytical sample of 1485 observations across 157 countries. The reduction is primarily attributable to missing data in the Corruption Perceptions Index (which is not reported for all countries by Transparency International) and in the poverty indicator (which is not consistently reported for high-income countries).

Table 2. Descriptive statistics of variables included in the analysis.

Variable	N	Mean	SD	Min	Max
NCD mortality rate (%)	1610	19.35	6.70	6.90	40.60
PM2.5 (μg/m3)	1610	22.15	13.63	3.70	94.70
UHC service coverage index	1610	65.98	17.09	23.00	92.00
Corruption Perceptions Index	1610	43.23	18.94	8.00	92.00
Gov. health + education spending (% GDP)	1610	7.67	3.62	0.14	22.89
Poverty headcount at $3/day (%)	1485	15.44	21.08	0.00	95.74
Access to electricity (%)	1610	81.62	27.92	4.20	100.00

3.4. Empirical Strategy

3.4.1. Core Specification

Our primary analytical approach is a two-way fixed-effects panel regression. The core specification is:

$${NCD}_{it} = {\beta }_0 + {\beta }_1{PM2.5}_{it} + {\beta }_2{UHC}_{it} + {\beta }_3\left({PM2.5}_{it} \times {UHC}_{it}\right) + {\beta }_4{X}_{it} + {\alpha }_i + {\gamma }_t + u_{it}$$

where:

• ${NCD}_{it}$ is the age-standardised adult NCD mortality rate for country i in year t;
• ${PM2.5}_{it}$ is the annual mean concentration of fine particulate matter;
• ${UHC}_{it}$ is the Universal Health Coverage service-coverage index;
• $X_{it}$ is a vector of time-varying controls (government health and education spending as % of GDP, poverty headcount ratio, electricity access, and, in robustness checks, urban slum prevalence);
• ${\alpha }_i$ are country fixed effects, absorbing all time-invariant country characteristics (geography, baseline epidemiological profile, historical institutional legacy);
• ${\gamma }_t$ are year fixed effects, absorbing global shocks common to all countries in a given year (e.g., global economic cycles, the COVID-19 pandemic in 2020–2021);
• $u_{it}$ is the idiosyncratic error term.

All continuous independent variables are mean-centred prior to the construction of interaction terms to facilitate interpretation of main effects [25].

3.4.2. Testing the Hypotheses

H1 (consistency check) is tested using an analogous specification with air-pollution-attributable mortality as the dependent variable, without interaction terms (a direct-effect model). A positive and statistically significant coefficient for PM2.5 would confirm that our indicator captures the expected health risk.

H2 (focal hypothesis) is tested via the coefficient on the interaction term PM2.5 × UHC in the core specification. A negative and statistically significant coefficient would indicate that higher UHC coverage attenuates the PM2.5–NCD mortality association.

H3 (mechanism/robustness) is tested by re-estimating the H2 specification with CPI in place of UHC, and then with both moderators included simultaneously. We also test the joint significance of both interaction terms when included together.

3.4.3. Inference and Standard Errors

Preliminary diagnostic testing (reported in Section 4.2) indicated the presence of both heteroscedasticity and first-order autocorrelation within panels. To address these violations, we employ Driscoll–Kraay standard errors for our main models [26], which are robust to heteroscedasticity, autocorrelation up to a specified lag, and cross-sectional dependence. For models where Driscoll–Kraay correction is not applied, we report standard errors clustered at the country level.

3.4.4. Robustness Checks

We address several anticipated threats to inference through a combination of design choices and robustness checks.

Reverse causality and omitted confounding. Countries with higher UHC and CPI are generally wealthier and may have lower PM2.5. Country fixed effects absorb time-invariant confounding, while the poverty control absorbs some time-varying confounding. Nevertheless, we explicitly frame all reported associations as conditional correlations rather than identified causal effects, consistent with the observational nature of our design.

Multicollinearity. We compute variance inflation factors (VIFs) for all specifications and report them. If VIFs exceed conventional thresholds (>10), we estimate a reduced-control specification alongside the full model.

Measurement error in NCD mortality. For countries with weak vital registration systems, WHO-modelled NCD mortality estimates are subject to measurement error. Country fixed effects absorb any time-invariant component of this error specific to a country’s statistical system. We discuss this limitation explicitly in Section 5.

COVID-19 pandemic (2020–2021). Year fixed effects absorb the global component of the COVID-19 shock. As an additional robustness check, we re-estimate all main models excluding 2020–2021 to confirm that results are not driven by pandemic-era anomalies in reported mortality.

Functional form. The Ramsey RESET test is used to detect potential non-linearity in the relationship between PM2.5 and NCD mortality. If non-linearity is detected, we explore alternative specifications including cubic splines for PM2.5.

Structural stability. We test for structural stability by including interaction terms between the COVID-19 period (2020–2021) and all independent variables. A non-significant joint test would support pooling all years in the main specification.

4. Results

4.1. Descriptive Statistics

The final analytical sample comprises an unbalanced panel of 1610 country-year observations across 165 countries for the descriptive statistics (Table 2). For the regression models, which include log_poverty as a control, the sample is 1485 observations across 157 countries.

Table 2 presents descriptive statistics for all variables included in the analysis.

The sample exhibits substantial cross-country variation across all variables, as indicated by the ranges and standard deviations. Mean adult NCD mortality is 19.4%, ranging from 6.9% to 40.6%. Mean PM2.5 concentration is 22.2 μg/m³, with a maximum of 94.7 μg/m³—substantially above the WHO Air Quality Guideline of 5 μg/m³. Mean UHC service coverage is 66.0 (range: 23–92), while the mean CPI score is 43.2 (range: 8–92), indicating wide variation in both health-system capacity and institutional quality.

4.2. Model Specification and Diagnostic Testing

Following the methodological approach of Spanos [27], we subjected our empirical model to a comprehensive battery of diagnostic tests to ensure statistical adequacy before proceeding to substantive inference. Table 3 summarises the results.

Table 3. Diagnostic tests for model specification.

Test	Statistic	p-Value	Conclusion
Panel specification
Hausman (FE vs. RE)	χ2(13) = 175.77	<0.001	Use fixed effects
Functional form
Ramsey RESET	F(3, 1466) = 10.52	<0.001	Non-linearity detected
Splines (joint significance)	F(4, 1310) = 0.97	0.423	Linear specification retained
Multicollinearity
Variance Inflation Factor (mean)	3.37	–	No multicollinearity
Residual diagnostics
Shapiro–Wilk (normality)	W = 0.89	<0.001	Non-normal residuals
Breusch–Pagan (heteroscedasticity)	χ2(1) = 3.17	0.075	Homocedasticity
Wooldridge (autocorrelation)	F(1, 1322) = 321.86	<0.001	First-order autocorrelation
Structural stability
COVID-19 interactions (joint test)	F(3, 156) = 1.31	0.273	Coefficients stable

Note: All models estimated with country and year fixed effects. Robust standard errors used throughout; Driscoll–Kraay correction applied in Model 2.

Panel specification: The Hausman test strongly rejected the null hypothesis that the random effects model is consistent (χ²(13) = 175.77, p < 0.001), indicating that country-level fixed effects are the appropriate specification. This decision is further supported by the F-test for joint significance of the country fixed effects (F(156, 1416) = 284.23, p < 0.001), confirming substantial unobserved country-level heterogeneity.

Functional form: The Ramsey RESET test indicated potential non-linearity in the relationship between PM2.5 and NCD mortality (F(3, 1466) = 10.52, p < 0.001). To explore this, we estimated a model with cubic splines for PM2.5 using three interior knots at the 25th, 50th, and 75th percentiles of the distribution. The spline terms were not jointly significant (F(4, 1310) = 0.97, p = 0.423), and the Akaike Information Criterion was virtually identical between the two specifications. Given that the spline terms were not jointly significant, we retain the linear specification on grounds of parsimony.

Multicollinearity: Variance inflation factors were well below the conventional threshold of 10 (mean VIF = 3.37; maximum VIF = 6.32 for UHC), ruling out multicollinearity as a concern for coefficient stability.

Residual diagnostics: The Shapiro–Wilk test rejected normality of the residuals (W = 0.89, p < 0.001). While non-normal residuals can indicate model misspecification, we have already explored alternative specifications (splines, interactions, and structural breaks) without finding evidence of substantial misspecification. Given the large sample size (n = 1485), the Central Limit Theorem ensures asymptotic normality of the estimators. The Breusch–Pagan residual diagnosis indicated weak evidence of heteroscedasticity (χ²(1) = 3.17, p = 0.075), which we address through robust standard errors (clustered and Driscoll–Kraay). The Wooldridge test for serial correlation was significant (F(1, 1322) = 321.86, p < 0.001), indicating the presence of first-order autocorrelation. To address this, we employ Driscoll–Kraay standard errors for the main model (Model 2), which are robust to both heteroscedasticity and autocorrelation [26].

Structural stability: A test for structural stability using COVID-19 period interactions was not significant (F(3, 156) = 1.31, p = 0.273), indicating that the estimated coefficients are stable across the pre-COVID (2012–2019) and COVID (2020–2021) periods. This supports the pooling of all years in the main specification.

4.3. Hypothesis 1 (Consistency Check): PM2.5 and Air Pollution Mortality

Before testing the moderation hypotheses, we conduct a consistency check to confirm that our PM2.5 indicator captures the expected association with health outcomes. Table 4 presents the results of the fixed-effects specification with air-pollution-attributable mortality as the dependent variable.

Table 4. Fixed effects panel regression: PM2.5 and air-pollution-attributable mortality (H1 consistency check).

Variable	Coefficient	Clustered SE	t	p-Value
PM2.5	0.247	0.078	3.19	0.002
Gov. health + education spending (% GDP)	0.210	0.309	0.68	0.498
Poverty headcount at $3/day (%)	3.463	1.578	2.20	0.030
Access to electricity (%)	0.018	0.081	0.22	0.823
Year fixed effects (2013–2021; 2012 omitted)
2013	−1.576	0.325	−4.84	<0.001
2014	−3.209	0.526	−6.10	<0.001
2015	−3.905	0.682	−5.72	<0.001
2016	−5.628	0.871	−6.46	<0.001
2017	−6.807	0.999	−6.81	<0.001
2018	−8.568	1.133	−7.56	<0.001
2019	−9.495	1.268	−7.49	<0.001
2020	−12.385	1.373	−9.02	<0.001
2021	−12.750	1.486	−8.58	<0.001
Constant	72.182	8.437	8.55	<0.001

The coefficient for PM2.5 is positive and highly significant (β = 0.247, p = 0.002), indicating that a 1 μg/m³ increase in ambient PM2.5 concentration is associated with an increase of 0.25 deaths per 100,000 population attributable to air pollution, within countries over time. This finding is consistent with the established epidemiological literature and provides a useful consistency check for our PM2.5 indicator, supporting its use in the subsequent analyses.

The year fixed effects are all negative and highly significant, confirming the global downward trend in air-pollution-attributable mortality over the 2012–2021 period. The magnitude of the decline accelerates over time (from −1.57 in 2013 to −12.75 in 2021), consistent with global progress in air quality regulation and reductions in ambient pollution concentrations in many regions.

The poverty coefficient is positive and significant (β = 3.463, p = 0.030), suggesting that higher poverty levels are associated with higher air-pollution-attributable mortality within countries. This is consistent with the literature on environmental inequality and the disproportionate exposure of low-income populations to air pollution.

A potential concern is that PM2.5 is significantly associated with air-pollution-attributable mortality (as shown here) but not with NCD mortality in the subsequent models. This is not a logical inconsistency. The two dependent variables differ fundamentally in their nature and construction. Air-pollution-attributable mortality is a modelled output from the GBD framework that embeds an assumed exposure–response relationship between PM2.5 and mortality; the association is therefore expected by construction. NCD mortality, by contrast, is an observed, age-standardised outcome that reflects the cumulative effect of multiple risk factors—tobacco use, diet, alcohol consumption, physical inactivity, and other determinants—among which ambient PM2.5 is only one contributor. The absence of a significant direct effect of PM2.5 on NCD mortality in the fixed-effects specification reflects the multifactorial nature of NCD mortality, the relatively short 10-year panel window, and the fact that fixed effects absorb time-invariant between-country variation.

4.4. Hypothesis 2: The Moderating Role of Universal Health Coverage

Model 2 tests the central hypothesis of this study: whether Universal Health Coverage (UHC) service coverage attenuates the association between PM2.5 and NCD mortality. We include the centred interaction term PM2.5 × UHC, along with the main effects of both variables and the full set of controls. To address the detected autocorrelation, we estimate this model with Driscoll–Kraay standard errors. Table 5 presents the results.

Table 5. Fixed effects panel regression: PM2.5 × UHC interaction (H2).

Variable	Coefficient	Driscoll–Kraay SE	t	p-Value
PM2.5 (centred)	0.004	0.007	0.54	0.537
UHC (centred)	0.056	0.008	7.48	<0.001
PM2.5 × UHC	0.000	0.000	1.33	0.215
Gov. health + education spending (% GDP)	0.037	0.048	0.78	0.454
Log(Poverty + 1)	0.232	0.089	2.61	0.028
Access to electricity (%)	0.035	0.005	6.49	<0.001
Year fixed effects (2013–2021; 2012 omitted)
2013	−0.321	0.017	−19.32	<0.001
2014	−0.479	0.017	−28.14	<0.001
2015	−0.582	0.021	−27.68	<0.001
2016	−0.844	0.018	−46.53	<0.001
2017	−1.148	0.023	−50.07	<0.001
2018	−1.255	0.020	−62.25	<0.001
2019	−1.514	0.022	−69.44	<0.001
2020	−1.704	0.043	−39.71	<0.001
2021	−2.197	0.041	−53.40	<0.001
Constant	17.142	0.865	19.82	<0.001

Hypothesis 2 is not confirmed. The interaction term between PM2.5 and UHC is not statistically significant (β = 0.000485, p = 0.215). This finding indicates that, in our global sample, the quality and coverage of health systems does not significantly modify the association between ambient PM2.5 exposure and NCD mortality. The buffering effect documented in previous studies—specifically Liu et al. [28] for under-five mortality and Barrow et al. [22] for West Africa—does not appear to generalise to adult NCD mortality at the global level.

The main effect of UHC is positive and highly significant (β = 0.056, p < 0.001). This finding requires careful interpretation. Three explanations are plausible. First, improved diagnosis and reporting: countries with stronger health systems have more complete vital registration and cause-of-death coding, leading to higher reported NCD mortality. Second, reverse causality: countries with higher NCD burdens may have invested more in health systems in response. Third, composition effects: the UHC index includes maternal and child health indicators, which may correlate with development processes that also increase NCD risk (urbanisation, dietary changes, and reduced physical activity). Importantly, because our dependent variable is age-standardised (the probability of dying between ages 30 and 70), the positive coefficient cannot be straightforwardly attributed to population ageing. We therefore interpret this coefficient not as a causal effect of UHC on NCD mortality, but as reflecting broader development and health-system characteristics.

The poverty coefficient is significant with Driscoll–Kraay standard errors (β = 0.232, p = 0.028), suggesting that higher poverty levels are associated with higher NCD mortality within countries. This is consistent with the literature on socioeconomic gradients in health.

The coefficient for electricity access remains positive and significant (β = 0.035, p = 0.001), reflecting the broader development process in which electrification is accompanied by urbanisation, lifestyle changes, and increased exposure to risk factors for non-communicable diseases.

The year fixed effects are all negative, large in magnitude, and highly significant, confirming a strong and steady global decline in age-standardised NCD mortality over the 2012–2021 period. The magnitude of the decline accelerates over time (from −0.321 in 2013 to −2.197 in 2021), consistent with global progress in cardiovascular disease management, cancer screening, and chronic disease treatment [1].

4.5. Hypothesis 3: The Moderating Role of Institutional Quality

Model 3 tests whether institutional quality, proxied by the Corruption Perceptions Index (CPI), moderates the association between PM2.5 and NCD mortality. We estimate the same fixed-effects specification, substituting CPI for UHC as the moderator. Table 6 presents the results.

Table 6. Fixed effects panel regression: PM2.5 × CPI interaction (H3).

Variable	Coefficient	Driscoll–Kraay SE	t	p-Value
PM2.5 (centred)	0.006	0.007	0.74	0.478
CPI (centred)	−0.033	0.012	−2.59	0.029
PM2.5 × CPI	0.001	0.001	2.08	0.067
Gov. health + education spending (% GDP)	0.043	0.048	0.89	0.395
Log(Poverty + 1)	0.088	0.078	1.13	0.287
Access to electricity (%)	0.043	0.004	9.15	<0.001
Year fixed effects (2013–2021; 2012 omitted)
2013	−0.288	0.015	−18.64	<0.001
2014	−0.381	0.010	−37.29	<0.001
2015	−0.477	0.011	−42.69	<0.001
2016	−0.725	0.100	−71.01	<0.001
2017	−0.999	0.007	−131.81	<0.001
2018	−1.134	0.009	−122.94	<0.001
2019	−1.372	0.013	−101.50	<0.001
2020	−1.531	0.043	−35.62	<0.001
2021	−2.024	0.043	−46.85	<0.001
Constant	16.554	0.756	21.89	<0.001

Hypothesis 3 is not confirmed. The interaction term between PM2.5 and CPI is not statistically significant (β = 0.001, p = 0.067). This finding indicates that general institutional quality, as measured by perceived corruption, does not significantly moderate the association between ambient PM2.5 exposure and NCD mortality. The buffering effect hypothesised in H3 does not find empirical support in this global panel.

The main effect of CPI is negative and significant at the 5% level (β = −0.033, p = 0.029). This suggests that, within countries over time, improvements in institutional quality (reductions in perceived corruption) are associated with lower NCD mortality. This finding is consistent with the broader literature on the health effects of governance and institutional quality [19,20]. It suggests that the association between institutional quality and health outcomes operates independently of air pollution exposure, possibly through mechanisms such as more efficient health spending, better enforcement of health regulations, and greater public trust in health systems.

The year fixed effects follow the same pattern as in Models 1 and 2, confirming the global downward trend in NCD mortality.

4.6. Comparison of Models and Robustness

Table 7 summarises the key coefficients across the three models, highlighting the consistency of the findings.

Table 7. Comparison of key coefficients across models.

Variable	Model 1 (Direct)	Model 2 (UHC)	Model 3 (CPI)
PM2.5	0.247 ** (0.078)	0.005 (0.007)	0.006 (0.007)
Moderator	–	0.056 *** (0.008)	−0.033 * (0.012)
PM2.5 × Moderator	–	0.000 (0.000)	0.001 (0.001)
Gov. spending	0.210 (0.309)	0.037 (0.048)	0.043 (0.048)
Log(Poverty)	3.463 * (1.578)	0.232 * (0.089)	0.088 (0.078)
Access electricity	0.018 (0.081)	0.035 *** (0.005)	0.043 ** (0.005)
R2 (within)	0.433	0.296	0.295
N	1485	1485	1485
Standard errors	Clustered (country)	Driscoll–Kraay	Driscoll–Kraay

Note: Standard errors in parentheses. * p < 0.05; ** p < 0.01; *** p < 0.001.

The comparison across models reveals several consistent patterns:

• PM2.5 is not significantly associated with NCD mortality in any specification. This null effect is robust to model specification, the inclusion of interaction terms, and the correction for autocorrelation.
• Neither UHC nor CPI significantly moderates the PM2.5–NCD mortality relationship. Both interaction terms are small in magnitude and statistically non-significant.
• UHC and CPI have significant main effects on NCD mortality, but in opposite directions. UHC is positively associated, which likely reflects improved diagnosis and reporting, reverse causality, and broader development processes rather than a causal effect. CPI is negatively associated, consistent with the view that better institutional quality directly improves health outcomes.
• The effect of poverty is sensitive to standard error specification; it is significant only with Driscoll–Kraay correction in Model 2 (β = 0.232, p = 0.028), suggesting a weak but detectable association between poverty and NCD mortality.
• Access to electricity consistently predicts higher NCD mortality, reflecting the epidemiological transition associated with economic development.
• The year fixed effects are stable across models, confirming the robustness of the temporal trend.

4.7. Summary of Findings

Table 8 provides a summary of the hypothesis tests and their corresponding findings.

Table 8. Summary of hypothesis tests.

Hypothesis	Finding
H1 (consistency check): PM2.5 is positively associated with air-pollution-attributable mortality	Confirmed: β = 0.247, p = 0.002
H2 (focal): UHC moderates the PM2.5–NCD mortality relationship	Not confirmed: interaction term p = 0.215
H3 (mechanism/robustness): CPI moderates the PM2.5–NCD mortality relationship	Not confirmed: interaction term p = 0.067
UHC main effect	Confirmed: improved diagnosis, reporting, and broader development processes
CPI main effect	Confirmed: negative and significant (reflecting institutional quality)
Global trend	Confirmed: NCD mortality declined significantly 2012–2021

5. Discussion

The central finding of this study is clear: neither Universal Health Coverage (UHC) nor institutional quality, as measured by the Corruption Perceptions Index (CPI), significantly moderates the association between ambient PM2.5 exposure and adult non-communicable disease (NCD) mortality in our global panel of 157 countries over 2012–2021. This result stands in contrast to two prior studies that found evidence of moderating effects: Liu et al. [28], who documented that public service provision moderates the PM2.5–under-five mortality relationship, and Barrow et al. [22], who found that health-system capacity conditions the air pollution–NCD mortality nexus in West Africa. Our findings suggest that the moderating effect documented in these specific contexts may not generalise to adult NCD mortality at the global level, at least within the 2012–2021 period.

This result is consistent across the specifications examined. The non-significance of the interaction terms holds across multiple specifications: with and without controls, with UHC and CPI as alternative moderators, and with both clustered and Driscoll–Kraay standard errors. The consistency of this null finding across models strengthens our confidence in its reliability, though we acknowledge the limitations discussed in Section 5.4, including the limited within-country variation in UHC and CPI.

Several factors may explain why the moderation effect documented in prior studies does not generalise to our global adult sample.

First, the outcome variable differs substantially. Liu et al. [28] examined under-five mortality, a health outcome that is more immediately sensitive to environmental insults and more directly responsive to public health interventions. Child mortality from respiratory infections, for example, can respond quickly to improved healthcare access or reductions in acute pollution episodes. Adult NCD mortality, by contrast, is the cumulative result of decades of exposure to multiple risk factors—including tobacco use, diet, alcohol consumption, and physical inactivity—among which air pollution is only one contributor. The moderating effect of health-system quality may be more detectable for acute or childhood outcomes than for chronic adult conditions, where the signal-to-noise ratio is lower and the lag between exposure and outcome is longer.

Second, the geographic scope differs. Barrow et al. [22] focused exclusively on West Africa, a region with particular epidemiological, institutional, and environmental characteristics. West African health systems face specific challenges—limited infrastructure, shortages of trained personnel, and high out-of-pocket expenditures—that may make the moderating effect of health-system capacity more salient. The effect they identified may be specific to that context and may not hold in a global sample that includes high-income countries with well-established health systems and lower pollution levels, as well as middle-income countries with heterogeneous institutional quality. In other words, the moderation effect may be context-dependent rather than universal, and our global sample may be too heterogeneous to detect an effect that operates primarily in specific regional or income-group contexts.

Third, the methodological approach differs. Both prior studies use cross-sectional or short-panel designs, while our study uses a within-country fixed-effects design that exploits variation over a decade. This methodological choice is more conservative and less prone to omitted-variable bias, but it also requires sufficient within-country variation in the moderator (UHC and CPI) to identify the interaction effect. If UHC and CPI change slowly over time—as they typically do, given the institutional nature of these variables—the within-country variation may be limited, making it harder to detect a moderating effect even if one exists. This is a classic trade-off between internal validity (fixed effects) and statistical power, and it suggests that the moderation effect may be more detectable in cross-sectional designs that exploit between-country variation, even at the cost of greater potential for omitted-variable bias.

Fourth, the time horizon differs. Our 10-year panel (2012–2021) may be insufficient to capture the cumulative effects of both PM2.5 exposure and health-system strengthening on adult NCD mortality. The health effects of chronic PM2.5 exposure manifest over decades, not years, and the improvements in health-system capacity that could plausibly moderate these effects also unfold over extended periods. A longer panel, extending back to the 1990s, might reveal moderating effects that are not detectable in our shorter window.

Although the moderation hypotheses are not supported, the main effects of UHC and CPI offer interesting and interpretable findings that contribute to our understanding of the determinants of NCD mortality.

UHC is positively and significantly associated with NCD mortality (β = 0.056, p < 0.001). This finding requires careful interpretation. Three explanations are plausible. First, improved diagnosis and reporting: countries with stronger health systems have more complete vital registration and cause-of-death coding, leading to higher reported NCD mortality. Second, reverse causality: countries with higher NCD burdens may have invested more in health systems in response. Third, composition effects: the UHC index includes maternal and child health indicators, which may correlate with development processes that also increase NCD risk (urbanisation, dietary changes, and reduced physical activity). Importantly, because our dependent variable is age-standardised (the probability of dying between ages 30 and 70), the positive coefficient cannot be straightforwardly attributed to population ageing. We therefore interpret this coefficient not as a causal effect of UHC on NCD mortality, but as reflecting broader development and health-system characteristics. This interpretation is supported by the positive and significant coefficient for access to electricity (β = 0.035, p = 0.001), which also reflects the development process—urbanisation, lifestyle changes, and increased exposure to behavioural risk factors for NCDs. Together, these findings underscore that the relationship between health-system development and NCD mortality is shaped by demographic and epidemiological transitions [29] that are not easily captured by simple cross-sectional associations.

Conversely, CPI is negatively and significantly associated with NCD mortality (β = −0.033, p = 0.029). This suggests that improvements in institutional quality—reductions in perceived corruption—are associated with lower NCD mortality within countries over time. This finding is consistent with the broader literature on the health effects of governance [30,31] and suggests that the association between institutional quality and health outcomes is independent of air pollution exposure. Possible mechanisms include more efficient allocation of health resources, better enforcement of health regulations, reduced diversion of public funds intended for health services, and greater public trust in health institutions, which may improve health-seeking behaviour and treatment adherence. The significance of this effect, even in a conservative fixed-effects specification with Driscoll–Kraay standard errors, reinforces the importance of governance as a determinant of population health.

We interpret this pattern as consistent with the view that emissions reduction and health-system strengthening are complements rather than substitutes. However, we emphasise that this is an interpretation, not a demonstration: the non-significant interaction terms indicate that we did not detect a buffering effect under our data and specification. The non-significant interaction terms do not prove that health systems and environmental quality are complements; they demonstrate that we did not detect a buffering effect under our data and specification. This interpretation is therefore qualified and should be understood as one plausible reading of the evidence.

5.1. Theoretical Implications: Weak vs. Strong Sustainability

Our findings also speak to the weak-versus-strong sustainability debate [32]. The moderation hypothesis—that institutional capacity can buffer the health effects of environmental degradation—is consistent with a weak-sustainability perspective, in which human and institutional capital can substitute for environmental quality. If health systems and institutional quality could effectively buffer the health effects of air pollution, then investments in these areas might be seen as substitutes for emissions reductions, at least in the short to medium term.

Our results, however, do not support this substitution hypothesis at the level of adult NCD mortality: the buffering effect of UHC and CPI is not statistically detectable. This does not mean that institutional capacity is irrelevant—indeed, the direct effects of UHC and CPI on mortality are significant and meaningful. But it suggests that institutional capacity cannot simply “compensate” for air pollution exposure in the short to medium term. The health burden of PM2.5 appears to persist even in countries with stronger health systems and better governance, at least within the time horizon of our analysis.

This finding has implications for the weak-versus-strong sustainability framing. From a strong sustainability perspective, natural capital (including environmental quality) and human/institutional capital are complements rather than substitutes. Our results are more consistent with this view: emissions reduction and health-system strengthening should be pursued as complementary strategies, not as substitutes. Investments in health systems and governance improve health outcomes directly, but they do not eliminate the need for air pollution control. This is a critical policy message: the absence of a buffering effect should not be interpreted as a rationale for deprioritising either environmental or health-system investments. As Ayres [24] noted, there are practical limits to substitution between natural and technological capital, and our findings align with that perspective in the context of environmental health.

5.2. Policy Implications

If the moderation effect is not generalisable, what are the policy implications for countries seeking to reduce the health burden of air pollution?

First, investments in health-system strengthening and institutional quality remain essential, not because they buffer the health effects of air pollution, but because they directly improve health outcomes (as reflected in the CPI main effect) and are part of broader development processes that reduce NCD mortality (as reflected in the UHC main effect). The direct effects of UHC and CPI on NCD mortality are significant and should not be overlooked. Strengthening health systems and reducing corruption are valuable policy objectives in their own right, independent of their role in environmental health.

Second, the null moderation finding carries a clear and specific policy implication: because health-system capacity does not appear to buffer the health effects of PM2.5 on adult NCD mortality, emissions reduction cannot be substituted by health-system strengthening. This implies a hierarchy of strategies: while health-system investments are valuable for their direct effects on health outcomes, they should not be viewed as a substitute for air pollution control. Emissions reduction remains the primary strategy for reducing the health burden of PM2.5, and health-system strengthening should be understood as a complementary, not alternative, policy.

Third, for low- and middle-income countries, this implies a dual strategy: continue to strengthen health systems and institutional quality while simultaneously pursuing emissions reductions through cleaner energy, transportation, and industrial policies. These should be understood as complementary strategies, not substitutes. Countries cannot afford to choose between environmental and health-system investments; both are necessary to address the complex, multi-factorial determinants of NCD mortality.

Fourth, the context-dependence of the moderation effect, suggested by our findings in comparison with prior studies, implies that policy recommendations should be tailored to specific contexts. The conditions under which health systems can buffer the health effects of air pollution may vary by region, income level, disease outcome, and institutional context. Future policy-oriented research should identify these conditions to provide more targeted guidance.

To test whether the absence of a moderating effect masks heterogeneity across country groups, we conducted subgroup analyses by income level (LMIC vs. HIC) and by region (Sub-Saharan Africa). The results are presented in Appendix Table A1. In all subgroups, the interaction term PM2.5 × UHC was not statistically significant (LMIC: β = 0.000167, p = 0.832; HIC: β = −0.003457, p = 0.486; Sub-Saharan Africa: β = −0.000569, p = 0.655). These results reinforce the conclusion that the moderation effect is not detectable even within specific country groups. However, we acknowledge that other sources of heterogeneity—such as pollution levels or stages of epidemiological transition—were not tested due to data limitations.

Our subgroup analyses (Appendix Table A1) provide further support for the context-dependence interpretation. If the moderating effect were a general phenomenon, we would expect to observe it across different country groups. However, the interaction term PM2.5 × UHC was not significant in any subgroup—neither in low- and middle-income countries (LMIC: β = 0.000167, p = 0.832) nor in high-income countries (HIC: β = −0.003457, p = 0.486), nor even in Sub-Saharan Africa (β = −0.000569, p = 0.655), the region where Barrow et al. [22] previously documented a moderating effect. The absence of a detectable effect across all subgroups tested is consistent with the view that the moderation effect is context-dependent rather than universal.

5.3. Comparison with Previous Studies

How do our findings fit with the existing evidence? Liu et al. [28] and Barrow et al. [22] provided important initial evidence that institutional and health-system capacity can moderate the health effects of air pollution. Our study extends this line of inquiry by testing the generalisability of this finding to a global sample and to adult NCD mortality.

The absence of a moderating effect in our sample does not invalidate the earlier findings; rather, it suggests that the moderation effect is context-dependent. It may be stronger for child health [28] than for adult NCD mortality, because child health outcomes are more immediately sensitive to environmental insults and more directly responsive to public health interventions. It may be stronger in regions with particularly weak health systems [22], where the marginal benefit of health-system strengthening is greater. It may be more detectable in cross-sectional designs than in within-country fixed-effects designs, which are more conservative and require greater within-country variation in the moderator.

Our study contributes to the literature by suggesting that the moderation effect may not be universal. This is an important finding because it prevents overgeneralisation of results from specific contexts and highlights the need for context-specific policy recommendations. It also opens up new research questions: under what conditions does the moderation effect operate? What thresholds of institutional quality or health-system capacity are required for a buffering effect to emerge? Are there specific disease outcomes or population subgroups for which the moderation effect is stronger? It is important to recognise that the comparison between our findings and those of Liu et al. [28] and Barrow et al. [22] is not a direct test of the same hypothesis in different settings, but a comparison across substantially different designs. Liu et al. [28] examined under-five mortality using a cross-sectional panel-threshold model, which exploits between-country variation and allows for non-linear threshold effects. Barrow et al. [22] used an ecological panel design in West Africa with mediation analysis, focusing on male premature NCD mortality and health expenditure as a moderator. Our study, by contrast, uses a within-country fixed-effects design with a global sample, focusing on adult NCD mortality for both sexes and using UHC service coverage as the moderator. These differences in outcome, geography, and method mean that the absence of a moderating effect in our study does not directly contradict these prior findings; rather, it suggests that the effect is context-dependent and may not generalise beyond the specific conditions under which it was originally observed.

Although the focal hypotheses were not confirmed, the consistency check (H1) confirmed that PM2.5 is strongly and significantly associated with air-pollution-attributable mortality (β = 0.247, p = 0.002). This result is important for two reasons.

First, it demonstrates that our PM2.5 indicator captures the expected health risk, confirming its consistency and supporting its use in the subsequent analyses. The coefficient is of plausible magnitude and is consistent with the established epidemiological literature on the health effects of particulate matter [1,5,6]. This consistency check is essential for interpreting the null results for the direct effect of PM2.5 on NCD mortality: the absence of a significant direct effect is not due to a problem with the PM2.5 indicator, but rather reflects the specificities of the outcome measure and the time horizon.

Second, it reinforces that PM2.5 is indeed a health hazard. The absence of a direct effect on NCD mortality in the fixed-effects specification should not be interpreted as evidence that PM2.5 is harmless. Rather, it reflects the specificities of the outcome measure (age-standardised NCD mortality) and the time horizon (10 years), which may be insufficient to capture the cumulative effects of chronic exposure. The consistency check confirms that PM2.5 is associated with health outcomes in the expected direction, and the null direct effect on NCD mortality should be interpreted cautiously, as discussed above.

5.4. Limitations

Several limitations should be considered when interpreting these findings, and they should be understood in the context of the study’s contributions rather than as grounds for dismissal.

First, the observational, country-level design precludes causal inference. Associations at the country level may not reflect individual-level relationships (the ecological fallacy), and we cannot capture individual-level exposure to PM2.5 or individual susceptibility to its health effects. While the fixed-effects specification and the battery of diagnostic tests strengthen the internal validity of our estimates, residual confounding from time-varying unobservables cannot be ruled out. Therefore, the null finding for the moderation effects may reflect these ecological limitations rather than the true absence of a buffering effect. We interpret our results as associations that are robust to a range of specifications, not as definitive evidence of the absence of a moderating effect.

Second, the 10-year panel window (2012–2021) may be too short to capture the long-term cumulative effects of PM2.5 exposure on NCD mortality, which manifest over decades. A longer panel, extending back to the 1990s, might reveal associations that are not detectable in our shorter window. The period was chosen for data comparability reasons (CPI methodology change in 2012, NCD data coverage), but this choice imposes constraints on the analysis. This is particularly relevant given that the direct effect of PM2.5 on NCD mortality was not significant in our specifications, while the consistency check (H1) confirmed the expected association with GBD-derived air-pollution-attributable mortality—the latter being a modelled outcome that already incorporates an assumed exposure–response relationship.

Third, measurement error in NCD mortality—particularly for countries with weak vital registration systems—may attenuate the estimated coefficients. The WHO-modelled NCD mortality data used in this study are the best available global estimates, but they rely on statistical modelling for countries with incomplete registration systems. Country fixed effects absorb time-invariant measurement error, but time-varying measurement error remains a concern. We did not conduct sensitivity analyses excluding countries with poor data quality because there is no widely accepted threshold for data quality in this context, and such exclusions would substantially reduce the sample and generalisability. This limitation is common to global health research and should be kept in mind when interpreting the results.

Fourth, the statistical power to detect interaction effects in fixed-effects panel models is partly determined by the within-country variation in the moderators. UHC and CPI are slow-moving institutional variables, meaning that their variation within countries over a decade is limited. This reduces the precision with which interaction effects can be estimated and may contribute to the non-significance of the interaction terms. While the sample size (N = 1485) is large relative to typical macro-panels, and the significant H1 effect confirms that the design has power to detect direct effects, the power to detect interactions with slow-moving variables is inherently constrained. Future research with longer panels or alternative identification strategies (e.g., instrumental variables) may be better suited to detect such effects if they exist.

Fifth, the UHC service coverage index is a broad aggregate measure that may not capture NCD-specific service capacity with precision. The index aggregates 14 tracer indicators, many of which are maternal and child health services only tangentially related to NCD management [7,8]. This means that our test of the moderation hypothesis using the UHC index is a conservative one: the index may not capture the specific dimensions of health-system capacity—such as access to antihypertensives, chronic disease management programmes, or specialist care—that would be most relevant to buffering the health effects of PM2.5 on NCD mortality. If a buffering effect exists but operates through NCD-specific services, our use of the broad UHC index may underestimate it. Future research with NCD-specific health-system indicators is needed to test this possibility.

Sixth, the CPI is a perception-based measure of corruption, not an objective measure of governance effectiveness. While it is the standard cross-national measure of corruption and has consistent coverage across our panel, its subjective nature is a limitation. Objective measures of governance effectiveness, such as the Worldwide Governance Indicators (WGI), are not available in the SDG Index database and could not be included without merging with another dataset. Future research should test the robustness of our findings using alternative measures of institutional quality.

Seventh, while we conducted subgroup analyses by income level and region, we did not test for heterogeneity by pollution levels or by stages of epidemiological transition. These dimensions of heterogeneity may be relevant, as the buffering effect of health systems could be more pronounced in high-pollution settings or in countries at earlier stages of epidemiological transition. Future research should explore these sources of heterogeneity.

Eighth, the diagnostic tests revealed non-normality of residuals and autocorrelation, which we addressed through robust standard errors (Driscoll–Kraay). These are corrections, not solutions, and they imply that the standard errors are conservative but may still be affected by the underlying violations. The non-normality of residuals, in particular, is common in large samples and is mitigated by the Central Limit Theorem, but it is a limitation worth acknowledging.

Ninth, the study is limited to the specific variables and time period available in the SDG Index database. Other potentially relevant variables—such as healthcare expenditure per capita, hospital bed density, or specific air quality regulations—are not included. Future research should incorporate these variables to provide a more comprehensive test of the moderation hypothesis.

5.5. Future Research Directions

The limitations identified above point to several priorities for future research.

First, subnational and individual-level analyses are needed to address the ecological fallacy inherent in country-level designs. Within-country variation in PM2.5 exposure, health-system capacity, and institutional quality would allow for more precise estimates and the inclusion of subnational fixed effects, reducing the risk of aggregation bias. Such analyses could also incorporate individual-level data on exposure, susceptibility, and health outcomes, providing a more direct test of the moderation hypothesis.

Second, longer panel windows—extending back to the 1990s or earlier—would better capture the cumulative effects of chronic PM2.5 exposure on adult NCD mortality, which manifest over decades. This would require addressing changes in data availability and methodology over time (e.g., the CPI methodology change in 2012), but would provide a more complete picture of the long-term health effects of air pollution and the role of institutional factors.

Third, more precise measurement of health-system capacity and institutional quality is needed. Future research should employ NCD-specific health-system indicators—such as the availability of essential NCD medicines, the density of cardiologists and oncologists, or the coverage of cancer screening programmes—to provide a more precise test of the moderation hypothesis for adult NCD mortality. Similarly, alternative measures of institutional quality—such as the Worldwide Governance Indicators, the Rule of Law Index, or the Varieties of Democracy (V-Dem) indicators—could test the robustness of the CPI findings and examine whether different dimensions of governance (regulatory quality, rule of law, government effectiveness) have differential moderating effects. Alternative sources of NCD mortality data—such as verbal autopsy studies or high-quality vital registration data from specific countries—would also help reduce measurement error concerns.

Fourth, quasi-experimental approaches—such as difference-in-differences or regression discontinuity designs—could be applied to specific policy changes (e.g., the introduction of UHC reforms, anti-corruption campaigns, or air quality regulations) to provide more credible causal estimates of the moderation effect. While observational panel designs are valuable for global descriptive analysis, quasi-experimental approaches would complement them by providing evidence from specific policy interventions.

Fifth, systematic exploration of heterogeneity is essential to clarify the conditions under which the moderation effect operates. Our global sample may be too heterogeneous to detect effects that are specific to certain contexts. Future research should test for heterogeneity by region, income group, pollution levels, and stages of epidemiological transition, as well as by disease subtype (cardiovascular vs. respiratory mortality) and outcome metric (years of life lost or disability-adjusted life years). Subgroup analyses could reveal patterns that are masked in the pooled sample and help identify the specific conditions—if any—under which health systems can buffer the health effects of air pollution.

6. Conclusions

This study tested whether health-system capacity (proxied by the UHC service coverage index) and institutional quality (proxied by the CPI) moderate the association between ambient PM2.5 exposure and adult NCD mortality in a global panel of 157 countries over 2012–2021. Using two-way fixed-effects panel regressions with Driscoll–Kraay standard errors, we tested three hypotheses.

The consistency check (H1) was confirmed: PM2.5 is positively and significantly associated with air-pollution-attributable mortality (β = 0.247, p = 0.002), a result consistent with the established epidemiological literature, reinforcing that PM2.5 remains a significant health hazard. However, the direct effect of PM2.5 on age-standardised NCD mortality was not statistically significant in any specification—a finding that likely reflects the cumulative, long-term nature of PM2.5 health effects, the multifactorial aetiology of NCD mortality, and the relatively short 10-year panel window.

Critically, neither H2 nor H3 was confirmed. The interaction terms between PM2.5 and UHC (β = 0.000485, p = 0.215) and between PM2.5 and CPI (β = 0.000838, p = 0.067) were not statistically significant. These findings indicate that, contrary to prior studies documenting moderating effects in specific contexts—Liu et al. [28] for under-five mortality and Barrow et al. [22] for West Africa—the buffering effect of health-system and institutional quality does not generalise to adult NCD mortality at the global level within the 2012–2021 period.

Despite the absence of moderating effects, the main effects of UHC and CPI offer important insights. UHC is positively associated with NCD mortality (β = 0.056, p < 0.001). This finding likely reflects a combination of improved diagnosis and reporting, reverse causality, and broader development processes, rather than a causal effect of UHC on mortality. Because the dependent variable is age-standardised, population ageing is not a sufficient explanation. CPI is negatively associated with NCD mortality (β = −0.033, p = 0.029), suggesting that better institutional quality is associated with lower NCD mortality, independent of air pollution exposure [30,31].

These findings carry clear policy implications. The absence of a detectable buffering effect implies that health-system strengthening cannot substitute for emissions reduction; the two strategies should be understood as complements, not substitutes. For low- and middle-income countries, this implies a dual strategy: strengthening health systems and institutional quality while simultaneously pursuing cleaner energy, transport, and industrial policies. The context-dependence of the moderation effect—suggested by the contrast between our global null findings and prior context-specific studies—also implies that policy recommendations should be tailored to specific regional and institutional contexts.

In conclusion, this study contributes to the literature by suggesting that the moderating effect of health-system and institutional quality on the PM2.5–NCD mortality relationship may not be universal. While prior studies have documented such effects in specific contexts [22,28], our global analysis suggests that these effects may not generalise to adult NCD mortality at the planetary scale. This finding is consistent with the view that emissions reduction and health-system strengthening are complements, not substitutes, in the pursuit of improved population health. However, we emphasise that this is an interpretation, not a demonstration: the non-significant interaction terms indicate the absence of a detectable substitution effect under our data and specification, not definitive proof of complementarity.