Numerous literature documents the health consequences of environmental degradation, encompassing air and water pollution, resource depletion, and climate change. The environmental burden of disease is substantial, with ambient air pollution alone implicated in millions of premature deaths annually. The challenge in cross-national work is separating the environmental signal from the developmental context in which it sits, since the wealth that lowers some forms of pollution also extends life through channels unrelated to the environment.

We therefore frame the analysis around population-health outcomes that are plausibly environment-sensitive—life expectancy and tuberculosis incidence—and treat health-system capacity as a control rather than as a competing explanation to be ignored.

The scale of the burden is what makes its cross-national measurement worth the difficulty. Ambient and household air pollution together rank among the leading environmental risk factors in global burden-of-disease accounting, implicated in millions of premature deaths each year and in a substantial share of the disease burden from cardiovascular and respiratory conditions ([1]; [8]). Such estimates are built up from individual-level dose-response relationships applied to population exposures; the complementary question, and the one a country panel can address, is whether the burden they imply is visible in the movement of aggregate health outcomes—whether countries carrying heavier environmental burdens have measurably shorter lives, and whether reductions in that burden within a country track gains in longevity over time. The value of the cross-national view is not that it improves on the epidemiology but that it tests whether the epidemiological signal aggregates to the national scale at which environmental policy is set.

The most developed strand of this literature concerns particulate air pollution and mortality. Fine particulate matter is robustly associated with cardiovascular and respiratory disease and with elevated mortality in cohort and time-series studies ([11]), and air pollution is plausibly linked to respiratory infections including tuberculosis. The cross-national question is whether these micro-level associations aggregate to a measurable population-level relationship between a country’s particulate burden and its life expectancy, net of income and medical care—and, more demandingly, whether changes in pollution within a country track changes in longevity over time.

The mechanisms are well characterized at the individual level. Fine particulate matter penetrates deep into the respiratory tract and enters the bloodstream, contributing to systemic inflammation, atherosclerosis, and impaired pulmonary function, and it is associated with elevated cardiovascular and respiratory mortality in large cohort and time-series studies ([11]). Particulate exposure has also been linked to susceptibility to respiratory infection, which supplies a plausible pathway to tuberculosis incidence, although that population-level association is weaker and more heavily mediated by the social determinants that dominate infectious-disease transmission. The strength of the individual-level evidence for the mortality channel is what makes particulate pollution the natural focal stressor for a longevity outcome, and it is why we treat the particulate-life-expectancy association as the study’s central estimate rather than as one association among six.

What the micro-level evidence cannot settle is whether the relationship is identified within countries or only across them. Cross-sectional comparisons of polluted and clean nations are confounded by everything that distinguishes rich from poor countries, so a negative pollution-longevity correlation in the cross-section is compatible with a purely developmental explanation. The more demanding question is whether a country that reduces its particulate burden over time experiences a corresponding gain in life expectancy, net of the income and medical resources that also change. This is precisely the question our fixed-effects specification is designed to answer, and it is the margin on which most cross-national environmental associations weaken.

Our design speaks to exactly this question, and the particulate-pollution-life-expectancy association is the study’s central estimate.

The relationship between income and health is summarized by the Preston curve, in which life expectancy rises steeply with income at low levels and flattens at high ones ([12]). Because income drives both environmental quality and health-system capacity, the Preston relationship is the principal confounder in any environment-health analysis. The environmental Kuznets curve adds a parallel income-degradation relationship ([6]; [14]), and the STIRPAT framework situates environmental impact within population, affluence, and technology ([3]; [17]). Together these motivate controlling for income, education, and—crucially—health-system capacity, so that environmental associations are not simply proxying for medical resources.

The Preston relationship is not merely a control to be included but the central inferential obstacle. Because income shifts the entire mapping from environment to health—cleaning the air, funding physicians and hospitals, and raising education simultaneously—any environmental coefficient estimated without holding development constant is liable to be development in disguise ([2]). The response adopted here is to condition on income and its square, on education, and, unusually for cross-national environment-health work, on two direct measures of health-system capacity: physicians per capita and health expenditure. Including medical capacity as a control means the environmental associations are estimated net of the very resources through which affluence most plausibly lengthens life independently of the environment, so that what remains is less easily dismissed as a proxy for development. The environmental Kuznets curve and STIRPAT traditions ([3]; [4]; [6]; [14]; [17]) supply the parallel logic on the environmental side, in which income drives degradation through offsetting scale, composition, and technique effects, and they motivate treating income as a shape to be estimated rather than a linear nuisance.

Beyond air pollution, the depletion of energy, mineral, and forest resources and the withdrawal of freshwater constitute a broader environmental burden whose health consequences are more diffuse and less directly tied to human exposure. Resource depletion acts on health partly through the degradation of the ecosystems on which subsistence and livelihoods depend, partly through the pollution that extraction and combustion generate, and partly as a marker of an extractive development path with its own health profile. Forest loss, for example, is associated with altered disease ecologies and with the loss of provisioning services, while energy depletion often accompanies the fossil-intensive activity that also raises local pollution. These channels are real but indirect, and the expectation is that resource-depletion stressors carry a smaller and less robust health cost than direct-exposure pollutants.

Freshwater withdrawal occupies an ambiguous position in this scheme. Water is fundamental to health, yet national withdrawal volumes are governed largely by agriculture, industry, and climate rather than by any direct health-relevant exposure, and withdrawals tend to rise with the economic development that also lengthens life. A cross-national withdrawal measure therefore behaves less as a stressor than as a development indicator, a point that becomes important when interpreting its association with life expectancy. Including it alongside the other stressors is nonetheless useful, both because water is a central environmental dimension and because its anomalous behavior illustrates the more general hazard of reading a development-correlated environmental measure as a health effect. This multi-stressor treatment—relating each environmental dimension to each health outcome rather than summarizing them—parallels the multi-outcome logic increasingly adopted in cross-national environmental analysis and is what allows the direct-exposure and resource-accounting stressors to be distinguished rather than merged.

The methodological literature on cross-national environment-health analysis converges on a single caution: the relationship is confounded, dynamic, and potentially bidirectional, so that a partial correlation is weak evidence of an environmental effect. Confounding by income and institutions is the first-order concern, addressed here by conditioning on development and medical capacity. Persistence is the second: both health outcomes and environmental conditions adjust slowly, so contemporaneous associations may over- or understate the relationships that would obtain in equilibrium, which motivates dynamic estimation even where the resulting estimates are imprecise. Reverse causation is the third, and the most conceptually interesting: healthier, longer-lived populations are also wealthier and better organized and may demand and obtain cleaner environments, so an environment-to-health reading cannot simply be assumed.

Our identification strategy responds to each of these in turn without overclaiming. We absorb unobserved country and year heterogeneity through two-way fixed effects and allow for cross-sectional dependence with Driscoll-Kraay standard errors ([5]); we probe persistence with a system-GMM estimator ([13]) that we report but do not lean upon, given the near-unit-root behavior of life expectancy; and we test direction directly, relating lagged health outcomes to current environmental stressors and running panel Granger tests in both directions. Neither device identifies causal direction: fixed effects absorb only heterogeneity that is constant within a country, leaving time-varying confounders untouched, and panel Granger tests bear on temporal precedence in the conditional mean rather than on causation. We deliberately do not report an external-instrument estimate, because no credible exclusion restriction is available for a national environmental stressor with respect to national health, and a weak or invalid instrument would purchase the appearance of causal identification at the cost of its substance. The coefficients are accordingly interpreted as conditional associations—strengthened where they survive within-country variation and the directional tests—and this graduated posture, claiming association, testing direction, and declining to manufacture an instrument, is central to how the results should be read.

Two gaps motivate this study. First, cross-national environment-health work often examines a single stressor or a single outcome, so the comparative association of different environmental stressors with different health outcomes is rarely assessed within a common framework that also controls for health-system capacity; as a result, a carbon or air-pollution finding is easily reported as an environmental one, and the medical resources that confound the relationship are often left out. Second, much of the evidence is cross-sectional, leaving open whether environmental associations with health survive within countries over time—the margin on which the developmental confound is hardest to sustain. We address both by relating six environmental stressors to two health outcomes and a composite in a delineated design, controlling for income, education, and medical capacity, and by re-estimating every relationship within countries and testing its direction.

We use a country-year panel of 70 countries observed from 2009 to 2020, comprising 721 observations and averaging 10.3 years per country, compiled from World Bank indicators. The sample spans all four World Bank income groups, providing variation across the full development range and supporting both cross-country and within-country estimation. Countries enter the panel if the World Bank reports all six environmental stressors, both health outcomes, and every control variable in at least one year of the window, and country-years with any missing analysis variable are dropped, so the estimation sample is complete by construction and no imputation is used. The 70 retained countries supply 721 of the 840 country-years the 2009-2020 window allows, or 85.8 per cent; coverage ranges from twelve years for 32 countries to two years for the least-covered one. Appendix A lists every country with its income group, region, and years observed.

The World Bank indicators offer internationally comparable environmental, economic, and health series with the broad country coverage needed to identify the relationships of interest, and the panel structure—repeated observation of the same countries—supports estimation both across countries and within them over time. The inclusion of all four income groups is deliberate: variation across the full development range is what makes it possible to separate the environmental associations from the developmental context, since a sample confined to affluent countries would offer little leverage on the confound at the heart of the analysis. The panel is close to but not perfectly balanced, averaging 10.3 of a possible twelve years per country, and all national-total variables are expressed in per-capita or intensity terms before analysis so that estimated relationships reflect environmental intensity rather than country size. Because the complete-case rule is applied to the same variable set for every country, coverage is uneven across the income range: the 36 high-income countries supply 92.4 per cent of their possible country-years, the 19 upper-middle-income countries 82.9 per cent, the 10 lower-middle-income countries 74.2 per cent, and the 5 low-income countries 73.3 per cent. The cross-income-group readings below therefore rest on denser data at the top of the income distribution than at the bottom, and the low-income group contributes 44 of the 721 observations. In counts, 119 of the 840 possible country-years are excluded: 33 of 432 in the high-income group, 39 of 228 in the upper-middle-income group, 31 of 120 in the lower-middle-income group, and 16 of 60 in the low-income group.

Each environmental stressor enters as a distinct independent variable and each health outcome as a distinct dependent variable (Table 1 and Table 2). Life expectancy is oriented so that higher values denote better health and tuberculosis incidence so that higher values denote worse health; the composite Health Index is sign-oriented so that higher values denote better health. Emissions are expressed per capita and income as the natural log of GDP per capita, mean-centered, with a squared term for the EKC. Crucially, we control for health-system capacity—physicians per capita and health expenditure—so that environmental associations are estimated net of medical resources.

Table 1. Independent and control variables.

Table 2. Dependent variables (population-health outcomes).

The composite Health Index is formed by standardizing each health outcome, orienting each so that higher values denote better health, and averaging with equal weights; it is reported as a descriptive summary rather than as the primary outcome, since the delineated stressor-by-outcome estimates are where the analysis does its work. Two design choices warrant emphasis. First, health-system capacity enters as a control rather than as an outcome: physicians per capita and health expenditure are the principal channels through which affluence lengthens life independently of the environment, so conditioning on them is what makes the residual environmental associations interpretable. GDP per capita, health expenditure, and physicians are treated as confounders, on the argument that their levels are set by the development path rather than by environmental degradation. Where that argument fails—where a control responds to environmental degradation—the variable is post-exposure, and conditioning on it can move the environmental coefficient in either direction, through collider as well as mediator paths, rather than only toward zero ([9]). The estimates are accordingly not claimed as a lower bound on the environmental burden, and Section 5.4 reports the central estimate with and without the health-system controls. Second, income enters in logs with a squared term, following the environmental Kuznets and Preston logics, so that the strongly non-linear relationship between income and both environment and health is captured rather than forced into a line.

Every series is taken from the World Bank’s World Development Indicators ([16]), and the indicator codes and transformations are as follows. Particulate pollution is adjusted savings: particulate emission damage (NY.ADJ.DPEM.GN.ZS), in per cent of GNI; the World Bank builds it from Global Burden of Disease estimates of mortality attributable to ambient PM2.5, ambient ozone, and household PM2.5 from solid-fuel cooking, valued as foregone labour income, with non-benchmark years extrapolated from mortality trends. Energy, mineral, and net forest depletion are the companion adjusted-savings series (NY.ADJ.DNGY.GN.ZS, NY.ADJ.DMIN.GN.ZS, NY.ADJ.DFOR.GN.ZS), also in per cent of GNI. Greenhouse-gas emissions are total emissions in kilotonnes of CO₂ equivalent (EN.ATM.GHGT.KT.CE) divided by total population (SP.POP.TOTL) and multiplied by 10⁶, giving kilograms of CO₂ equivalent per capita. Freshwater withdrawals are annual total withdrawals as a per cent of internal renewable resources (ER.H2O.FWTL.ZS), a ratio that exceeds 100 where withdrawal draws on external or non-renewable sources. Life expectancy at birth (SP.DYN.LE00.IN) is in years and tuberculosis incidence (SH.TBS.INCD) per 100,000 people. Among the controls, GDP per capita is in current US dollars (NY.GDP.PCAP.CD), entered as its natural logarithm, mean-centred on the estimation sample, together with the square of that centred term; current health expenditure is in per cent of GDP (SH.XPD.CHEX.GD.ZS); physicians are per 1,000 people (SH.MED.PHYS.ZS); compulsory education is its duration in years (SE.COM.DURS); and population (SP.POP.TOTL) enters in natural logarithms. No price deflator is applied: the only monetary variable, GDP per capita, enters in logarithms and mean-centred, every other variable is a ratio, rate, or physical intensity, and no aggregation is performed beyond the per-capita and per-cent-of-GNI normalisations carried by the source series themselves.

Coefficients are therefore read in the units of the source series. A within-country particulate coefficient of −2.849 on life expectancy says that a one-percentage-point increase in particulate emission damage as a share of GNI accompanies a fall of 2.849 years in life expectancy; a one-standard-deviation increase in the measure (0.67 percentage points) corresponds to about 1.9 years. An energy-depletion coefficient of −0.271 in the pooled model is per percentage point of GNI, and a greenhouse-gas coefficient is per kilogram of CO₂ equivalent per capita, so that a one-tonne increase in per-capita emissions corresponds to 1,000 times the tabulated value.

We report descriptive statistics, a correlation matrix, and variance-inflation factors, then estimate for each health outcome a multiple linear regression on the full set of six environmental stressors and controls, with country-clustered standard errors. We re-estimate the full model with two-way fixed effects and Driscoll-Kraay standard errors to identify associations from within-country variation and to allow for cross-sectional dependence. To make the pattern explicit, we tally, for each estimator, the environmental associations that harm health versus those that appear protective, and we display the full sign pattern.

For robustness we retain all six stressors throughout and add a specification test of fixed against random effects in cluster-robust auxiliary-regression form ([7]; [10]), a Blundell-Bond system-GMM estimator ([13]) with the lagged dependent variable and the full stressor set (instruments collapsed; variables standardized), reverse-causality and panel-Granger tests for every stressor-outcome pair to assess temporal precedence in each direction, and a set of lagged and moving-average exposure specifications for the central particulate association. We do not report an external-instrument estimate, as a credible exclusion restriction is unavailable; coefficients are interpreted as conditional associations.

The logic of this graduated sequence is to let the reader see how the central associations behave as successive identifying assumptions are relaxed. The pooled regression establishes the conditional associations in the full cross-section and time series; the two-way fixed-effects specification removes all stable between-country differences and identifies the associations from within-country variation alone, the more demanding test; the Driscoll-Kraay standard errors guard against the cross-sectional dependence that national environmental and health series are likely to exhibit; and the dynamic and directional tests probe persistence and the direction of the relationship. Rather than resting the conclusions on any single estimator, we report the sign and significance pattern across estimators and tally, for each, the environmental associations that harm health against those that appear protective, so that the coherence—or fragility—of the pattern is visible directly rather than asserted.

Table 3 summarizes the sample, which spans the full range of national income, environmental quality, and population health.

The spread is wide on every dimension. Life expectancy ranges from the low fifties to the mid-eighties, tuberculosis incidence from a handful of cases per hundred thousand to several hundred, and GDP per capita across more than two orders of magnitude, so the sample contains the developmental heterogeneity on which identification depends. The environmental stressors are similarly dispersed, and their skewness—several are bounded below at zero and have long right tails—reflects the concentration of resource depletion and heavy pollution in particular countries and years.

Table 3. Descriptive statistics (721 country-year observations).

Table 4 reports Pearson correlations. No variance-inflation factor exceeds six (maximum 5.1), so multicollinearity does not threaten the estimates.

Table 4. Correlation matrix.

Variables:1 = Life expectancy(years); 2 = Incidence of TB(per100k); 3 = GHG emissions(per capita); 4 = Energy depletion(%GNI); 5 = Mineral depletion(%GNI); 6 = Particulate pollution; 7 = Freshwater withdrawals; 8 = Forest depletion(%GNI); 9 = log GDP per capita; 10 = Physicians(per1k); 11 = Health expenditure.

Table 5 reports the coefficient of every environmental stressor and control on each health outcome, each from a separate country-clustered regression. Environmental degradation is associated with worse health: higher particulate pollution, greenhouse-gas emissions, and energy and forest depletion are each associated with lower life expectancy. The particulate-pollution coefficient is the largest and most precisely estimated, and the life-expectancy model fits closely (R² = 0.93), reflecting the strong role of the income and health-system controls against which the environmental associations are net. Table 6 and Table 7 report the full regressions for life expectancy and tuberculosis incidence.

Table 5. Multiple linear regression coefficients: each environmental stressor on each health outcome (pooled).

Each column is a separate regression; country-clustered standard errors. *** p < 0.001, ** p < 0.01, * p < 0.05.

The signs are as the framework anticipates. Among the stressors, particulate pollution carries by far the largest life-expectancy penalty, and greenhouse-gas emissions and energy and forest depletion each enter negatively, so that the direct-exposure and resource-accounting stressors align in harming longevity. The controls behave as the developmental logic predicts: income enters strongly and positively, and health expenditure positively, confirming that the environmental associations are estimated against a backdrop in which development and medical capacity are the dominant correlates of longevity. That the environmental coefficients remain significant net of these controls is the first indication that they are not simply proxying for development. The tuberculosis model fits less well and yields few significant stressor associations in the pooled specification, consistent with infectious-disease incidence being governed more heavily by social and programmatic factors than by the environmental stressors measured here.

Table 6. Full regression for life expectancy.

R² = 0.925; adj. R² = 0.924; F = 100.7; N = 721.

Table 7. Full regression for tuberculosis incidence.

R² = 0.526; adj. R² = 0.518; F = 7.2; N = 721.

Table 8 re-estimates the full model under two-way fixed effects. The central result survives: higher particulate pollution remains associated with lower life expectancy within countries over time, net of country and year effects and of health-system capacity, and energy depletion is associated with higher tuberculosis incidence. That the particulate-life-expectancy association holds within countries—where most cross-national environmental associations weaken—is the strongest evidence in the study. Driscoll-Kraay standard errors yield the same qualitative pattern.

Table 8. Fixed-effects coefficients: each environmental stressor on each health outcome (within-country).

Country and year effects; country-clustered standard errors. *** p < 0.001, ** p < 0.01, * p < 0.05.

The contrast between the pooled and within-country estimates is the analytical heart of the study. Removing all stable between-country differences strips out precisely the developmental correlates that make cross-sectional environmental associations difficult to interpret, so the associations that survive are identified from changes within countries over time. That the particulate-life-expectancy association is among the survivors—attenuated in magnitude, as within estimators typically are, but retaining its sign and significance—is far harder to attribute to confounding than a cross-sectional correlation would be. The appearance of an energy-depletion association with tuberculosis within countries, absent in the pooled model, further suggests that some resource-depletion channels operate through year-to-year variation that the cross-section obscures. The health-system controls do not drive the estimate, and they do not attenuate it uniformly. Dropping physicians and health expenditure moves the within-country particulate coefficient from −2.849 to −3.092, and removing all controls moves it to −3.523; in the pooled specification the same coefficient moves in the opposite direction, from −3.670 with the health-system controls to −3.472 without them. Conditioning on medical capacity is therefore not a device for producing a conservative estimate, and the coefficients are not interpreted as bounds.

The associations are predominantly harmful, as the environmental-burden framework predicts. Counting significant stressor-outcome cells across the three reported outcomes, the pooled model yields 6 harming against 2 apparently protective; the fixed-effects model 4 and 2; and the Driscoll-Kraay model 6 and 3. Restricted to the two primary outcomes, life expectancy and tuberculosis incidence, the same tallies are 4 and 1, 2 and 1, and 4 and 2; the difference is the composite index, which adds harming cells for energy depletion and particulate pollution and one apparently protective cell for freshwater withdrawals. The harming associations concentrate on life expectancy and are led by particulate pollution and resource depletion; the apparently protective associations are confined to freshwater withdrawals, which enters positively for life expectancy and the composite and negatively for tuberculosis incidence, and this reflects the fact that water withdrawals rise with the agricultural and industrial development that also extends life, and should not be read as a health benefit of water extraction. Figure 2 displays the sign pattern for all three outcomes under pooled and fixed-effects estimation.

Read through the framework, the pattern is coherent: the environmental stressors most directly tied to human exposure—airborne particulates above all—carry the clearest health cost, and that cost is visible even within countries over time and net of the medical resources that might otherwise mask it.

Figure 2. Sign map of environmental stressor-health associations for all three health outcomes under pooled and fixed-effects estimation. ✕ = harming health; ✓ = protective.

The sign tally is therefore not merely a count but a summary of where the environmental-burden logic holds: the harmful associations cluster on the mortality outcome and on the stressors closest to human exposure, the apparently protective associations are confined to the development-correlated water measure, and the pattern is stable enough across estimators to be read as a coherent signal rather than as noise.