The proposition that education improves environmental outcomes draws on human-capital theory [1] and on a large body of work linking schooling to pro-environmental attitudes, behaviours, and policy support [2][3]. Human-capital theory holds that schooling raises the productive, cognitive, and civic capabilities of a population, and each of these capabilities has a plausible environmental corollary. Cognitively, education improves the capacity to process complex, probabilistic, and temporally distant information—precisely the character of climate risk—so that better-educated populations are argued to perceive environmental hazards more accurately and to discount the future less steeply. Civically, education is associated with greater political participation and with support for the collective institutions, from regulation to public investment, through which environmental goods are provided. Technically, an educated labour force adopts and operates cleaner and more efficient technologies more readily, lowering the emissions intensity of a given level of output.

Taken together, these mechanisms imply a demand-side and a supply-side channel. On the demand side, educated citizens want cleaner environments and vote, consume, and organize accordingly; on the supply side, educated workers and firms are better able to deliver cleaner production. Both point in the same direction: more education should be associated with less environmental degradation. The micro-level evidence is broadly consistent with this expectation, though it is nuanced: quasi-experimental work exploiting compulsory-schooling reforms finds that additional education causally raises several pro-environmental behaviours [2], and schooling has been shown to yield “green returns” in the form of more environment-friendly conduct [3], even as the effect is stronger for knowledge-based actions than for stated concern or willingness to pay [2].

The macro-level evidence, however, is mixed and sensitive to the outcome considered and to whether income is held constant, reflecting precisely the entanglement of education with development that this study foregrounds. In long-run panel data for OECD economies, advanced (tertiary) human capital is associated with substantially lower per-capita CO₂ [4]; yet other cross-national studies find the association weak, absent, or contingent on income and on the measure of human capital used. Because education and income co-evolve, isolating an educational association from a developmental one is the central empirical challenge of the cross-national literature. A society becomes more educated as it grows richer, and it emits more carbon per capita as it grows richer; the two movements are confounded at the level of the country. We address this in three ways: by controlling for income directly, by distinguishing cross-country from within-country variation, and by treating the slow-moving, stock-like character of educational attainment explicitly when interpreting within-country estimates. The result is a design in which the education signal is estimated net of affluence rather than in place of it.

A recurring theme in this literature is that education’s environmental influence is indirect. Rather than lowering emissions by itself, education is thought to shift the conditions that determine emissions: it raises support for and adoption of renewable energy, accelerates the move away from coal, improves environmental quality through regulation and cleaner production, and builds the research and human-capital base needed to decarbonize. On this view, the appropriate way to test the education-climate relationship is not only to estimate direct associations but also to identify the channels through which education operates. A direct regression of carbon on education asks whether educated countries emit less; a mediation model asks why, and answers by locating the intermediate outcomes that carry the relationship.

The renewable-energy share is a natural candidate mediator. Education plausibly promotes renewable adoption—through public support for energy transition, through the skilled labour and engineering capacity that renewable deployment requires, and through the research base that lowers the cost of clean technology—and renewable adoption plausibly lowers carbon by displacing fossil fuels in final energy use. Cross-national evidence for OECD countries indicates that the extent to which renewable-energy consumption translates into environmental quality itself depends on education, consistent with education operating on the environment through the energy mix [7]. Part of education’s association with carbon should therefore run through the energy mix. If the mediator carries most of the relationship, the direct education-carbon path will be small once the mediator is included, and the indirect path—the product of the education → renewables and renewables → carbon effects—will be the locus of the action.

This motivates a formal mediation framework, in which the total association of education with carbon is decomposed into a direct component and an indirect component operating through renewable energy [9][10]. The classic causal-steps logic [9] and the resampling-based inference for indirect effects [10] together provide a way to estimate the education → renewables path, the renewables → carbon path, and the residual direct path, and to test whether the product of the first two is distinguishable from zero. A finding of a significant indirect effect with little direct effect would support the view that education greens carbon outcomes by greening the energy system rather than by acting on carbon directly—the indirect-only case in the taxonomy of mediation, and the sharpest possible statement of the hypothesis that education’s environmental value is real but mediated.

The role of income makes the environmental Kuznets curve (EKC) directly relevant. The EKC posits an inverted-U relationship between income and degradation: pollution rises in the early stages of development and falls once income is high enough that societies demand and can afford environmental quality [5]. If education rises with income, then part of any education-environment association may simply be the downward-sloping arm of the EKC operating through affluence rather than through schooling as such. Controlling for income, and for its square, is therefore essential to prevent the EKC from masquerading as an education effect. Recent work embeds education directly within the EKC and finds that the expansion of higher education has increasingly offset the per-capita emissions associated with income growth, suggesting that education carries an environmental effect distinct from affluence [11].

The empirical status of the EKC is, however, fragile. The inverted-U is frequently undetectable, is sensitive to the pollutant, the functional form, and the sample, and often fails to reappear when the specification is varied [12][13]. This fragility cuts two ways for the present study. It means we cannot lean on the EKC as a settled account of the income-environment relationship, but it also means that income is treated as a control rather than a hypothesis. Log income enters linearly in the pooled models and with a squared term in the fixed-effects models, and the pooled models are re-estimated with the squared term as a sensitivity check (Section 6), so that any curvature is absorbed without committing the analysis to a particular shape.

The STIRPAT framework offers a complementary decomposition. Building on the IPAT identity, STIRPAT expresses environmental impact as a stochastic function of population, affluence, and technology [6][14], and it situates both education and the energy mix within the technology term. In this accounting, education is not a fourth factor standing outside the identity but a determinant of the technology multiplier—the efficiency and cleanliness with which a given population and level of affluence are converted into environmental impact. This is exactly the intermediating role that the mediation design formalizes: education shapes technology, technology shapes impact. The STIRPAT tradition thus justifies controlling for population and affluence, anticipates that affluence will carry a positive association with per-capita carbon that competes with any mitigating role of education, and locates education’s influence precisely where the mediation analysis looks for it.

Ecological modernization theory holds that advanced societies can decouple environmental harm from economic growth through knowledge, technology, and institutions [8]. On this account, modernization is not inherently at war with the environment; beyond a certain level of development, the same capacities that drive growth—scientific knowledge, technological sophistication, and institutional capacity—can be turned to environmental ends, so that richer, more knowledge-intensive societies begin to clean up rather than degrade further. Recent empirical tests of the theory’s core claim—that technological innovation decouples environmental impact from growth—use cross-national environmental-technology patent data and find qualified support, underscoring that the mechanism runs through the greening of technology rather than through growth itself [15]. Education is a central enabler in this story, both as a source of the knowledge that makes cleaner production possible and as a driver of the institutional and civic capacity that turns that knowledge into policy. Ecological modernization therefore predicts precisely the pattern this paper investigates: an inverse education-degradation relationship that runs through the greening of technology and the energy system rather than through a mechanical reduction of emissions.

Two gaps remain in the empirical literature, and this study is organized around them. First, education is usually represented by a single measure, so the comparative environmental association of attainment, enrollment, spending, and research output is rarely assessed together. Different facets of education plausibly bear on different environmental margins—spending on the financing of clean-energy capacity, advanced human capital on the skills of decarbonization, research output on the technological frontier—and collapsing them into one index averages these distinct channels into a single, less informative number. Second, the widely asserted indirect character of the education-climate link is seldom tested as an explicit mediation; it is asserted in theory and left unmodeled in estimation. This study addresses both gaps by estimating each education measure against each climate indicator and by formally testing the education → renewables → carbon pathway.

We use a country-year panel of 97 countries observed from 2013 to 2020, comprising 402 observations, drawn from the World Development Indicators, where the dependent indicators fall under the World Bank’s Climate Change theme. The panel spans every region and income level, from low-income to high-income economies and from Sub-Saharan Africa to Oceania, which gives the cross-country dimension wide gradients in both education and environmental performance that the design exploits. The panel is short and unbalanced, averaging 4.1 years per country; this is an important feature for interpretation, because educational attainment changes slowly and a short panel offers limited within-country variation. We therefore treat cross-country variation as the primary source of identification for slow-moving education stocks and interpret within-country estimates with corresponding caution. This is a design choice made in advance and stated openly, not a post hoc rationalization of weak within-country results.

Each education measure enters as a distinct independent variable and each climate-change indicator as a distinct dependent variable (Table 1 and Table 2). The six education measures—master’s attainment, bachelor’s attainment, tertiary enrollment, the advanced-education share of the labour force, government education spending, and scientific output—capture different facets of a country’s educational endowment, from the stock of highly qualified people to the flow of resources into schooling and the output of its research base. Four of the five environmental indicators—per-capita CO₂, PM2.5, forest depletion, and the coal share of electricity—are oriented so that higher values denote greater degradation; the renewable-energy share is oriented so that higher values denote improvement. Keeping this orientation convention in mind is essential to reading the coefficient tables because the expected sign of an environment-improving education effect is negative on the four degradation indicators and positive on the renewable share.

Scientific output is expressed per capita to remove country-size scaling, so that a large country does not appear research-intensive merely by virtue of its population. Income is entered as the natural log of GDP per capita, mean-centered, linearly in the pooled and mediation models and with a squared term in the fixed-effects models; a pooled sensitivity analysis adds the squared term (Section 6).

The renewable-energy indicator is the WDI series Renewable energy consumption (% of total final energy consumption) (EG.FEC.RNEW.ZS), the measure used for Sustainable Development Goal indicator 7.2. It counts all renewable sources in final energy use—hydropower, wind, solar, geothermal, and solid, liquid, and gaseous biofuels and renewable waste—including the traditional use of biomass for cooking and heating. It is therefore a measure of renewable final-energy consumption rather than of clean-electricity generation, and in lower-income economies a high value can reflect reliance on traditional biomass rather than modern renewable deployment. Results for the renewable share are interpreted accordingly.

Table 1.Independent and control variables.

Table 2.Dependent variables (climate change and environmental indicators).

We first report descriptive statistics, a correlation matrix, and variance-inflation factors to characterize the sample and to confirm that multicollinearity among the education measures and income is within tolerable bounds. We then estimate, for each climate indicator, a multiple linear regression on the full set of six education measures and controls, with country-clustered standard errors [16], mapping which education-indicator associations are significant in the pooled country-year sample. Because pooled estimates combine between- and within-country variation, they are described as pooled associations, and the two components are separated explicitly in Section 5.4. The full set of education measures is retained in every specification, so that no result depends on a favorable choice of which facet of education to include; the delineated design is carried through the estimation rather than imposed only at the descriptive stage. Each estimator yields 30 education-indicator coefficients (six measures × five outcomes), which form a multiple-testing family. Significance markers in the tables are unadjusted; we additionally compute Benjamini-Hochberg false-discovery-rate q-values across each family of 30 [17] and treat cells that are nominally significant but do not survive the adjustment as exploratory.

To probe whether these associations operate within countries as well as between them, we re-estimate the full model with two-way fixed effects and with Driscoll-Kraay standard errors [16]; we emphasise at the outset that, because educational attainment is near-constant within countries over this short panel, the within-country estimates are underpowered for the slow-moving stocks and should be read as a limitation rather than as a contradiction of the cross-country results. Fixed effects difference out everything that is constant within a country, including most of the variation in attainment, so the estimator is being asked to identify an education effect from the small year-to-year movements that remain—a demand the data cannot meet over four-odd years. To separate the between- and within-country components directly, we also estimate a within-between hybrid model [18][19], in which each education measure and control enters both as its country mean and as the deviation from that mean, with year effects and country-clustered standard errors; the coefficient on the country mean is the between-country association and the coefficient on the deviation is the within-country association.

Because the inverse relationship is expected to be indirect, we test mediation explicitly. Using renewable energy as the mediator, we estimate the education → renewables path (a), the renewables → carbon path conditional on education (b), and the direct education → carbon path (c′), and we obtain the indirect effect (a × b) with a country-clustered bootstrap confidence interval [10]. We report mediation both for a composite education index and for government education spending, the channel most directly tied to the financing of renewable adoption. The mediation analysis rests on the standard sequential-ignorability assumptions, and education spending, the renewable share, and CO₂ are measured in the same year, so the models cannot establish temporal ordering among them; we treat the resulting indirect association as evidence consistent with the hypothesized channel rather than as a fully identified causal mechanism, and we return to this point in the limitations.

The mediation models are three linear regressions estimated by ordinary least squares on the full analytic sample of 402 country-year observations from 97 countries:

M_it = a₀ + a∙X_it + γ₁′C_it + ε_1,it (1)

Y_it = b₀ + c′∙X_it + b·M_it + γ₂′C_it + ε_2,it (2)

Y_it = c₀ + c·X_it + γ₃′C_it + ε_3,it (3)

where X is government education spending (% of GDP), M is the renewable-energy share (% of final energy consumption), Y is per-capita CO₂, and C contains log GDP per capita, log population, and unemployment; the other five education measures do not enter the mediation models. Standard errors are clustered by country. The indirect association is a × b and the direct association is c′; because Equations (1)-(3) share the same covariates and sample, the total association decomposes exactly as c = c′ + a × b. The 95% confidence interval for a × b is the percentile interval from 2000 cluster-bootstrap draws, each of which resamples the 97 countries with replacement, keeping all years of a sampled country together, and re-estimates Equation (1) and Equation (2). The same models are estimated with the composite education index in place of X.

We also report Hausman tests comparing fixed- and random-effects specifications [20] and reverse-causality checks. We do not estimate system GMM: with an average of 4.1 years per country and nearly time-invariant education stocks, a dynamic panel estimator is not credibly identified, and we prefer to say so rather than report fragile dynamics. Reporting a dynamic-panel specification on a panel this short would convey a precision the data do not support; the honest description of the identification available here is cross-sectional and mediation-based, and we let the analysis rest on that footing.

Table 3 summarizes the sample, which spans a wide range of educational development and environmental performance. The spread is the engine of the cross-country design: per-capita carbon ranges from near zero to 44.82 tonnes, the renewable share from 0 to 97 percent, and PM2.5 exposure from 4.90 to 88.79 µg/m³, so the between-country gradients that the regressions exploit are wide rather than compressed. The education measures show comparably broad variation, with master’s attainment ranging from 0 to 30.65 percent of the population aged 25 and over, and government education spending from under 1 to 8.60 percent of GDP. This dispersion is what makes it possible to detect margin-specific education effects that would be invisible in a narrower, more homogeneous sample.

Table 3. Descriptive statistics (402 country-year observations).

Table 4 reports Pearson correlations among the outcomes, education measures, and controls. Two features stand out. First, the education measures correlate with one another and with income, as expected of variables that co-evolve with development; bachelor’s attainment, tertiary enrollment, and scientific output are all strongly correlated with log GDP per capita (0.75, 0.74, and 0.90, respectively). This is the collinearity the design must manage, and it is the reason income is controlled directly rather than left to contaminate the education coefficients. Second, the raw correlations already foreshadow the regression pattern: the renewable share is negatively correlated with per-capita carbon (−0.43) and with the coal share (−0.36), while advanced-education labour and education spending are among the education measures least entangled with income (correlations with log GDP of 0.26 and 0.46), which is part of why their environmental associations survive the income control.

After expressing scientific output per capita, no variance-inflation factor exceeds nine (maximum 8.6), and the residual income-education correlation is handled by direct control. The VIF diagnostic matters because the delineated design keeps all six education measures in the model simultaneously; a maximum VIF below the conventional threshold confirms that the coefficients can be separately estimated despite the correlations among the regressors, so that the margin-specific pattern in Table 5 is not an artifact of collinearity.

Table 4.Pearson correlation matrix.

Variables: 1 = CO₂ per capita (t); 2 = PM2.5 (µg/m³); 3 = Forest depletion (% GNI); 4 = Coal electricity (% gen.); 5 = Renewable energy (% final); 6 = Master’s attainment (% 25+); 7 = Bachelor’s attainment (% 25+); 8 = Tertiary enrollment (% gross); 9 = Advanced-educ. labor (%); 10 = Govt. education spending (% GDP); 11 = Scientific articles (per capita, log); 12 = log GDP per capita; 13 = log Population; 14 = Unemployment (%).

Table 5 reports the coefficient of every education measure and control on every climate indicator, each from a separate country-clustered regression. Reading the table with the orientation convention in mind, the pooled pattern matches the inverse-relationship expectation and concentrates where theory predicts. Government education spending is strongly associated with a higher renewable share (4.679, p < 0.001); advanced-education labor and tertiary enrollment with lower PM2.5 (−0.381 and −0.147, both p < 0.05); and master’s attainment and advanced labor with a higher renewable share (0.802 and 0.581, both p < 0.05). These are exactly the energy-composition and air-quality margins anticipated by H2, and the two strongest channels—spending and advanced human capital—are those predicted by H5. After Benjamini-Hochberg adjustment across the 30 education cells, only the spending coefficient on the renewable share remains significant (q = 0.022); the other nominally significant cells have q-values of about 0.13 and are treated as exploratory. Income, meanwhile, carries the anticipated positive association with per-capita carbon (2.070, p < 0.01), competing with education on that margin, while explaining the strong fit for CO₂ (R² = 0.51) and PM2.5 (R² = 0.60); the model fits the income-governed outcomes well precisely because income does much of the work there.

The delineated design earns its keep in the differences across measures. The education measures do not move together: master’s attainment raises the renewable share, while bachelor’s attainment is negatively signed on it (−0.735, p < 0.05), and scientific output—strongly correlated with income—shows no independent environmental association once income is controlled. Collapsing these into a single education index would average away the very heterogeneity that identifies which facet of education bears on which outcome. The negative bachelor’s coefficient on the renewable share, set against the positive master’s coefficient, is a case in point: an aggregate measure would net these to something close to zero and report, misleadingly, that attainment has no bearing on the energy mix, when in fact the two levels of attainment pull in opposite directions on that margin.

Table 6 and Table 7 report the full regressions for per-capita CO₂ and for the renewable-energy share—the carbon outcome and the principal channel—showing coefficients, standard errors, and significance for every predictor. The contrast between them is the empirical heart of the paper: on carbon, no education measure attains significance and income dominates (2.070, p < 0.01), while the model still fits well (R² = 0.511) because income carries the outcome; on the renewable share, four of the six education measures are significant—master’s (0.802, p < 0.05), bachelor’s (−0.735, p < 0.05), advanced labour (0.581, p < 0.05), and spending (4.679, p < 0.001)—and income is not (−6.271, n.s.). Education’s signal lives on the energy-composition margin, not the carbon margin, exactly as H2 anticipates, and the weak direct carbon coefficient is the clue that motivates the mediation analysis rather than an embarrassment to be explained away.

Table 5.Multiple linear regression coefficients: each education measure on each climate indicator (pooled country-year associations).

Each column represents a separate regression; country-clustered standard errors are used; significance levels are unadjusted for multiple testing, and only Govt. education spending → Renew survives Benjamini-Hochberg adjustment (q = 0.022). Orientation: CO₂/PM2.5/ForDepl/Coal higher = worse; Renew higher = better. ***p < 0.001, **p < 0.01, *p < 0.05.

Table 6.Full regression for per-capita CO₂ emissions (pooled).

Table 7.Full regression for the renewable-energy share (pooled).

Table 8 re-estimates the full model under two-way fixed effects. Consistent with H4, the pooled associations largely do not appear within countries over this short panel. We interpret this as a consequence of limited within-country variation in slow-moving education stocks rather than as evidence against the inverse relationship: educational attainment changes by fractions of a percentage point per year, so a 4.1-year window leaves little signal for fixed-effects identification, and the few within-country coefficients that reach significance should be treated cautiously. The one within-country coefficient that is robustly significant—scientific output on per-capita carbon (0.426, p < 0.01) and on the coal share (3.969, p < 0.01)—is most naturally interpreted as a development-and-industrialization dynamic rather than as an education effect on the environment, since research output tracks the growth of an industrial, energy-intensive economy. The cross-country evidence remains the appropriate basis for inference about education stocks.

Table 8. Fixed-effects coefficients: each education measure on each indicator (within country).

Country and year effects; country-clustered standard errors; controls are log GDP per capita and its square, log population, and unemployment. ***p < 0.001, **p < 0.01, *p < 0.05.

Table 9 separates the two sources of variation directly with the within-between hybrid model. The between-country coefficients reproduce the pooled pattern: education spending (5.274, p < 0.001), master’s attainment (0.847, p < 0.05), and advanced-education labor (0.670, p < 0.05) are associated with a higher renewable share and bachelor’s attainment with a lower one (−0.773, p < 0.05), while tertiary enrollment (−0.157, p < 0.05) and advanced-education labor (−0.423, p < 0.05) are associated with lower PM2.5; bachelor’s attainment on per-capita CO₂ (0.112, p < 0.05) and scientific output on PM2.5 (2.408, p < 0.05) are the between-country associations in the worsening direction. The within-country coefficients are almost uniformly insignificant, the exceptions being master’s attainment on per-capita CO₂ (−0.077, p < 0.05) and advanced-education labor on forest depletion (0.016, p < 0.05). A between estimator fitted to the 97 country means gives the same picture for the renewable share (spending 4.883, p < 0.001; master’s 1.174, p < 0.01; advanced labor 0.627, p < 0.05; bachelor’s −0.752, p < 0.05) and for advanced-education labor on PM2.5 (−0.484, p < 0.05). The education associations are therefore concentrated in the between-country component, as H4 anticipates.

Table 9.Within-between hybrid model: between-country and within-country coefficients of each education measure on each indicator.

Each column is a single regression in which every education measure and control (log GDP per capita, log population, unemployment) enters as its country mean (Panel A) and as the deviation from that mean (Panel B); year effects; country-clustered standard errors; N = 402 country-years, 97 countries. Orientation: CO₂/PM2.5/ForDepl/Coal higher = worse; Renew higher = better. ***p < 0.001, **p < 0.01, *p < 0.05.

The pooled results support the inverse-relationship hypothesis. Counting nominally significant education-indicator cells across the five indicators, the pooled model yields 5 associations that improve the environment (less degradation or more renewables) against only 1 that worsens it; the within-country model, underpowered for slow-moving stocks, yields 0 and 3. With 30 tests per estimator, these tallies are exploratory summaries of direction: after Benjamini-Hochberg adjustment, the pooled model retains one improving cell (spending on the renewable share) and the within-country model none. The improving associations are not scattered: they concentrate in the renewable-energy share and in air quality, exactly the energy-composition and quality margins through which education is theorized to act, and they are led by education spending and advanced human capital. Figure 2 displays the full sign pattern, with environment-improving associations in one shade and worsening associations in another.

The pattern is consistent with the view that education is inversely related to climate degradation, but that it acts on the composition of the energy system and the quality of the environment rather than uniformly on every indicator. The 5-to-1 tally in the pooled model and the 0-to-3 tally within countries are two views of the same fact: the inverse relationship is a between-country phenomenon over this panel, and the within-country estimator, deprived of variation, produces a handful of coefficients whose signs are as likely to reflect development dynamics as environmental ones. Where education is weakest is precisely on per-capita carbon, the margin most strongly governed by income—motivating the mediation analysis that follows. That the direct carbon association is weak is therefore not an anomaly to be explained away but the central clue: it points toward an indirect pathway in which education reaches carbon through the energy mix rather than on its own.

Figure 2. Sign map of education-indicator associations: environment-improving (inverse relationship) and worsening cells, under pooled and fixed-effects estimations.

The central question is whether education’s association with carbon is direct or indirect. Figure 3 reports the mediation for government education spending, with the renewable-energy share as the mediator and per-capita CO₂ as the outcome, controlling for income, population, and unemployment. Education spending is associated with a higher renewable share (path a = 4.30, p < 0.01), and a higher renewable share is associated with lower per-capita carbon (path b = −0.040, p < 0.05). The resulting indirect effect is negative and statistically significant (a × b = −0.173; 95% bootstrap CI [−0.373, −0.019], excluding zero), while the direct association of education spending with carbon is not statistically significant (c′ = −0.167, p = 0.51). Because c = c′ + a × b, the total association is −0.339 (p = 0.22), of which the indirect component accounts for about half and the direct component for the other half. The indirect association is distinguishable from zero and the direct association is not, but the two are similar in magnitude, so the direct association is imprecisely estimated rather than shown to be absent. The result therefore supports the indirect component of H3 but not its expectation of little direct effect, and it does so on the spending channel that H5 identified as the one most tightly linked to the financing of clean-energy capacity.

The structure of the result is worth stating precisely. Both constituent paths carry the signs the framework expects—higher spending accompanies a higher renewable share, and a higher renewable share accompanies lower carbon—so their product, the indirect association, is negative and its bootstrap interval excludes zero. The direct path, by contrast, is statistically indistinguishable from zero but not small: at −0.167 it is close to the indirect component (−0.173), and the total association (−0.339) divides roughly equally between them. The renewable-share channel therefore accounts for part of the association of spending with carbon, leaving a residual association of comparable size that the data estimate imprecisely. Because spending, the renewable share, and CO₂ are measured in the same year, the decomposition describes how the associations are arranged rather than a sequence in which spending raises the renewable share and thereby lowers carbon.

Figure 3. Mediation of the education-carbon relationship through renewable energy (government education spending channel). Indirect association is significant; direct association is not significant; total association c = c′ + a × b = −0.339.

The composite education index shows the same directional pattern—education is associated with a higher renewable share, which is associated with lower carbon—but its indirect effect is not statistically distinguishable from zero, because the index blends fast-moving, policy-sensitive spending with slow-moving attainment stocks that are bound to income. The reason is instructive rather than disappointing: averaging across facets dilutes the sharp spending signal with noisier, more confounded components. This is the delineated design speaking again—the mediation is visible on the specific channel where theory locates it and blurs when that channel is averaged with others. Together with the cross-country pattern, the mediation is consistent with the paper’s thesis that part of education’s association with climate outcomes runs through the composition of the energy system.