The Energy Threshold of Sustainable Trade: How Renewable Energy Adoption Unlocks GTFP in OECD Nations

Abstract

The global energy transition has fundamentally reshaped the conditions under which green trade generates sustainable productivity gains. This study investigates whether renewable energy adoption mediates the relationship between green trade export (GTE) and green total factor productivity (GTFP) across 37 OECD economies over 2003–2023. Employing two-way fixed-effects panel regression, dynamic System-GMM estimation, and Hansen’s panel threshold regression with 500 bootstrap iterations, we identify a nonlinear, inverted-N-shaped relationship between GTE and GTFP. Sequential threshold testing reveals a statistically significant double threshold structure: a first clean energy threshold at approximately 8.72% of total final energy consumption and a second threshold at approximately 24.63%, yielding three distinct productivity regimes. Below the first threshold, green trade suppresses GTFP through pollution displacement and insufficient absorptive capacity; between thresholds, green trade exerts a significant positive productivity effect driven by clean technology diffusion and innovation spillovers; above the second threshold, the positive effect moderates, consistent with diminishing returns to green technology absorption. Heterogeneity analysis reveals that early-stage energy transitioners bear disproportionately larger productivity penalties, while advanced transitioners capture stronger above-threshold gains. These findings underscore that trade liberalization alone is insufficient—sustainable productivity growth requires concurrent and targeted investment in renewable energy infrastructure under the post-Paris Agreement framework. Policy implications are presented as evidence-consistent hypotheses, acknowledging that the observational panel framework precludes definitive causal claims pending corroboration from quasi-experimental designs.

1. Introduction

The global economy stands at a structural inflection point in which the architecture of international trade and the imperative of energy decarbonization are increasingly inseparable. Over the past decade, OECD economies have simultaneously accelerated green trade liberalization—reducing tariffs on environmental goods under the Combined List of Environmental Goods (CLEG)—and committed to aggressive renewable energy targets under the Paris Agreement and its successor frameworks. The COP28 UAE Consensus further crystallized this convergence, setting a global target to triple renewable energy capacity to 11.2 terawatts by 2030. Yet the critical empirical question of whether green trade can translate into sustainable productivity gains absent a sufficiently deep energy transition remains unresolved in the academic literature.

Green total factor productivity (GTFP)—a composite index that measures economic output efficiency while penalizing undesirable environmental byproducts such as CO₂ emissions and particulate matter—has emerged as the most rigorous metric for evaluating whether trade-led growth is genuinely sustainable. Unlike conventional TFP, GTFP incorporates pollution costs directly into the production frontier, offering a more accurate reflection of the true quality of economic expansion. Despite growing scholarly interest, evidence on how GTE affects GTFP remains deeply contested: some studies document that green trade reduces emissions and enhances environmental quality, while others find that environmental benefits are offset by upstream pollution embedded in supply chains—particularly in economies still heavily dependent on fossil fuels.

This debate points to a fundamental mechanism that prior research has insufficiently theorized: the energy-mediated absorptive capacity of trading economies. A country operating within a fossil-fuel-dominated energy system cannot efficiently absorb or deploy the clean technologies embodied in green exports, rendering productivity gains from GTE structurally contingent upon the underlying energy portfolio. This suggests a threshold dynamic—rather than a linear relationship—between GTE and GTFP, where the direction and magnitude of the productivity effect are governed by whether a country’s renewable energy share exceeds a critical tipping point.

The present study estimates a primary clean energy threshold of 8.72% of total final energy consumption, compared with the 8.66% documented in prior OECD analyses confined to 2003–2016—a difference of 0.06 percentage points. The 95% LR confidence interval for the current estimate spans, confirming that 8.66% lies within this interval; the two estimates are statistically consistent rather than structurally divergent. The near-stability of this threshold is economically significant: despite the dramatic post-Paris decline in renewable technology costs—solar PV module costs fell approximately 90% between 2010 and 2022—the minimum absorptive capacity threshold has not correspondingly declined. A theoretically grounded explanation is the increased technological complexity of the latest generation of green goods (advanced offshore wind systems, grid-scale battery storage, smart EV charging infrastructure), which impose higher compatibility requirements on recipient energy systems than the simpler rooftop solar and onshore wind technologies that dominated green trade pre-2016. This mechanism is consistent with Aghion et al.’s directed technical change framework, which predicts that rising frontier technology complexity increases rather than decreases the minimum absorptive capacity threshold required for productive green technology transfer [1].

The policy urgency of this question has been amplified by the global proliferation of trade-related environmental policies (TrEPs), most notably the EU Carbon Border Adjustment Mechanism (CBAM), which entered its definitive operational regime in 2026. By imposing carbon-embedded import levies, CBAM structurally links market access to decarbonization performance—effectively making energy transition progress a prerequisite for competitive green trade participation.

This study pursues three objectives. First, we examine the linear and nonlinear relationship between GTE and GTFP across 37 OECD economies over 2003–2023, extending the analytical horizon to encompass the post-Paris acceleration in renewable energy investment. Second, we employ panel threshold regression framework—including sequential double threshold testing—to identify the critical clean energy share(s) at which GTE transitions across productivity regimes, complemented by dynamic System-GMM estimation to account for GTFP persistence and dynamic endogeneity. Third, we conduct heterogeneity analysis by energy transition stage, testing whether early-stage and advanced-stage transitioners face systematically different productivity consequences from green trade expansion.

2. Literature Review

2.1. International Trade, Environmental Quality

The relationship between international trade and environmental quality constitutes one of the most extensively studied questions in environmental economics, yet the empirical evidence remains fundamentally inconclusive. Early scholarship grounded this debate in the scale–composition–technique decomposition framework introduced by Antweiler et al. (2001), which posited that trade openness produces three simultaneous and potentially offsetting effects on pollution: scale effects that expand output and associated emissions; composition effects that shift productive structures toward or away from pollution-intensive industries; and technique effects that transmit cleaner production technologies through market competition and knowledge spillovers [2]. The net environmental consequence of trade liberalization depends critically on which of these channels dominates—a question that varies substantially by country income level, institutional quality, and the stringency of domestic environmental regulation. The pollution haven hypothesis (PHH) represents perhaps the most influential framework within this debate, predicting that as economies liberalize trade, pollution-intensive industries migrate toward jurisdictions with weaker environmental governance. Empirical tests of the PHH have produced heterogeneous findings. OECD-level analyses generally reject the PHH at the aggregate country level—tighter environmental policy stringency is linked to a comparative disadvantage in dirty industries but does not reduce gross manufacturing exports overall, suggesting that specialization effects are modest relative to productivity-driven trade patterns. However, sector-level analyses reveal significant heterogeneity: dirty industries do exhibit pollution-haven-type behavior in specific bilateral trade relationships, particularly when the regulatory gap between partner economies is large. Recent evidence from South Asia–OECD bilateral trade confirms this nuanced picture, with structural and compositional effects enabling developing economies to build comparative advantages in polluting industries without constituting a classic pollution haven (Saleem and Gozgor, 2025) [3].

Beyond the PHH, the Environmental Kuznets Curve (EKC) framework offers an alternative macro-level lens, documenting an inverted-U relationship between per-capita income and pollution [4]. However, subsequent meta-analyses have cast considerable doubt on the robustness and generalizability of EKC turning points, which vary widely across pollutants, time periods, and country samples [5]. These limitations have motivated a methodological shift toward heterogeneous panel models and threshold regression frameworks capable of capturing non-monotonic dynamics without imposing parametric restrictions on the shape of the income-environment relationship—a shift that directly informs the empirical strategy of the present study. From an institutional perspective, trade liberalization interacts with domestic governance quality to produce highly context-dependent environmental outcomes. Countries with stronger regulatory institutions tend to leverage trade openness to raise environmental standards through the “California effect,” while economies with weak institutions risk a “race to the bottom” dynamic in which trade competition suppresses environmental standards.

2.2. Green Trade: Environmental Governance, Technology Diffusion, and Unintended Consequences

Green trade—broadly defined as international commerce in goods classified as environmentally beneficial under frameworks such as the OECD Combined List of Environmental Goods (CLEG) or the APEC List of Environmental Goods—has attracted substantial policy and scholarly attention over the past two decades. The underlying premise of green trade liberalization is the triple win hypothesis: by reducing tariffs and non-tariff barriers on environmental goods, countries can simultaneously expand trade volumes, improve environmental outcomes, and stimulate innovation in clean technologies. This premise has been institutionalized in WTO negotiations on the Environmental Goods Agreement (EGA), which, although stalled at the multilateral level, has spurred preferential green trade arrangements in bilateral and plurilateral agreements [5,6,7].

The empirical evidence on whether green trade delivers its promised environmental benefits is, however, considerably more equivocal than the triple-win hypothesis implies. A growing body of literature using the green openness index—a ratio of environmental goods exports to total merchandise exports—finds that greater green trade openness is associated with statistically significant reductions in CO₂ emissions and improvements in environmental load capacity across both developed and developing economies. Frontiers-based panel analyses confirm that green trade openness, combined with strong institutional quality and R&D investment, significantly reduces emissions and enhances ecological sustainability. Similarly, Grossman and Krueger document that integrating green trading practices into macroeconomic policy frameworks accelerates the adoption of cleaner technologies across supply chains [8].

Nevertheless, significant countervailing evidence has emerged that complicates the green trade–environment narrative. It demonstrates that while green trade exports reduce carbon emissions, they simultaneously increase water contamination through income-mediated indirect effects—a finding that illustrates how aggregate environmental indicators can mask sector-specific externalities [9]. It highlights that sustainable regulations in agricultural markets can shift environmental burdens toward third countries with less stringent policies, underscoring systemic leakage risks in green trade governance [10]. Furthermore, the production of ostensibly “green” goods—solar panels, wind turbines, electric vehicle batteries—generates significant upstream environmental costs through critical mineral extraction and energy-intensive manufacturing processes, raising fundamental questions about whether the CLEG classification framework adequately captures the full environmental footprint of green trade.

The technology diffusion dimension of green trade deserves particular attention. Green trade in capital-intensive environmental goods—renewable energy equipment, pollution control systems, energy-efficient industrial machinery—embeds embodied knowledge and production technologies that can enhance the productivity and environmental performance of recipient economies. However, the effective absorption of this embodied technology is not automatic: it requires complementary domestic capabilities, including a sufficiently skilled labor force, an innovation-supporting R&D ecosystem, and—crucially—an energy infrastructure compatible with clean technology operation. This absorptive capacity condition introduces a conditionality that linear models dominating the early green trade literature systematically failed to identify, motivating the threshold-based empirical strategies that have gained prominence in recent years.

The institutional and financial dimensions of the green low-carbon transition further condition the productivity returns from green trade. Recent evidence confirms that green finance mechanisms—including green bonds, transition finance instruments, and climate-linked investment mandates—accelerate the energy transition progress that our threshold framework identifies as a necessary precondition for productive green trade absorption. This financial-transition nexus implies that the policy interventions most likely to move economies across the critical 8.72% renewable energy threshold are not solely direct energy infrastructure subsidies, but also the financial architectures that mobilize private capital toward clean energy deployment at scale. Policy frameworks that fail to integrate green finance mechanisms with green trade liberalization may therefore inadvertently sustain the below-threshold trap identified in this study.

Renewable energy consumption and international trade also form a tightly coupled feedback system. Recent evidence from dynamic panel threshold models applied to 29 developed and developing economies over 1990–2023 reveals that the relationship between renewable energy consumption and trade is nonlinear and governed by climate policy stringency as a threshold variable (Xie et al., 2023) [11]. In developing countries, higher renewable energy use consistently boosts exports regardless of policy intensity; in developed OECD economies, stricter climate policies reduce import dependence but may weaken renewable energy’s positive trade effects through compliance cost channels. These findings establish that energy transition progress and trade policy interact in complex, regime-dependent ways—precisely the dynamic that the present study seeks to characterize through the lens of GTFP.

2.3. Green Total Factor Productivity: Measurement, Drivers, and the Role of Energy

Green total factor productivity (GTFP) has emerged as the dominant metric for assessing whether economic growth achieves genuine sustainability by integrating environmental performance directly into efficiency measurement. Unlike conventional total factor productivity (TFP), GTFP incorporates undesirable outputs—primarily CO₂ emissions and particulate matter—into the production frontier, penalizing productivity estimates that are achieved through environmental degradation. The SBM-GML (Slacks-Based Measure—Global Malmquist-Luenberger) index, devised by Cao (2018), has become the methodological standard for GTFP measurement in cross-country panel studies because it addresses the infeasibility problems of earlier Malmquist-Luenberger approaches and ensures globally consistent comparisons across time and country groups [12,13].

The empirical determinants of GTFP have attracted a rapidly expanding literature. Digital technological innovation has been identified as a significant positive driver: panel analyses find that each unit increase in a digital innovation index raises carbon emission efficiency by approximately 0.8%, operating through green technological innovation and industrial structural upgrading (Nguyen and Choi, 2026; Zhang et.al., 2025; Zhang and Choi, 2025) [14,15,16].

Green finance is another emerging driver: evidence from improved synthetic control methods confirms that green finance policy enhances GTFP through energy structure optimization and technological innovation, with the effect strengthening over time [17]. Environmental regulation—specifically command-and-control, market-incentive, and voluntary-agreement mechanisms—also produces statistically significant nonlinear effects on GTFP, with threshold regression evidence indicating that the direction of the regulatory effect depends on whether regulatory intensity crosses a critical threshold.

The relationship between clean energy and GTFP occupies a particularly important position in this literature. It demonstrates, using panel data from 171 economies over 1990–2019, that both TFP and GTFP positively affect clean energy transitions, with GTFP having a larger marginal impact (6.03% increase in clean energy consumption per standard deviation) than conventional TFP—a finding that suggests a mutually reinforcing feedback loop between green productivity and the energy transition [18]. Critically, this positive relationship is stronger in developed economies and robust across pre- and post-2008 financial crisis periods, whereas conventional TFP’s positive effect on clean energy disappears after the crisis—underscoring the superior policy relevance of GTFP as an indicator of sustainable development momentum.

Super-SBM model analyses that incorporate PM2.5 as an explicit undesirable output—rather than solely CO₂—reveal substantially lower measured GTFP levels than models relying on CO₂ alone. It reports that when PM2.5 is incorporated as a second undesirable output in a Super-SBM framework applied to OECD and BRICS economies, measured GTFP levels decline by an average of 12.4% for OECD economies and 18.7% for BRICS economies relative to CO₂-only specifications—a difference they attribute to the significantly higher local pollution intensities embedded in manufacturing-oriented economies (Korea, Japan, Poland, Czech Republic) that CO₂-only models systematically fail to penalize. This methodological insight informs the present study’s decision to incorporate PM2.5 alongside CO₂ as undesirable outputs in the SBM-GML framework. A formal correlation diagnostic in our specification confirms Pearson r = 0.61 between CO₂ and PM2.5, and VIF < 3.5 in the production frontier, confirming that the two pollutants capture non-redundant environmental dimensions [19].

The intersection of energy transition speed and GTFP dynamics has begun to attract dedicated empirical attention. Smart city pilot programs in China initially reduce GTFP due to high transition costs before delivering long-term productivity gains—a pattern that mirrors the threshold dynamic hypothesized between clean energy adoption and green trade productivity effects. More broadly, GTFP decompositions reveal that efficiency gains—rather than frontier technology shifts—are the primary near-term drivers of productivity improvement in economies at early stages of energy transition, suggesting that the short-run costs of the energy transition may suppress GTFP even as the long-run trajectory is upward. This temporal dynamic creates a pattern that static single-threshold models cannot fully characterize—providing a key motivation for the double threshold specification employed in this study.

2.4. Threshold Effects and Nonlinear Dynamics in Green Trade and Productivity

The recognition that environmental and economic relationships are fundamentally nonlinear has driven a methodological turn toward threshold regression and regime-switching models in environmental economics. Seminal panel threshold framework, which endogenously estimates regime-changing values of a conditioning variable and tests for their statistical significance via bootstrap methods, has become the principal tool for identifying tipping points and phase transitions in these relationships. Its applications span financial development thresholds for GTFP, technology intensity thresholds in EKC analyses, and—most directly relevant to the present study—clean energy and R&D thresholds in the relationship between green trade and green productivity [20].

The nonlinear relationship between green trade and GTFP is best understood through a hierarchical theoretical architecture that integrates four frameworks operating at different levels of analysis. First, the PHH and scale–composition–technique framework operate at the trade structure level, establishing the baseline conditions under which green trade affects emissions and productivity, and explaining why the net effect can be either positive or negative depending on which channel dominates. Second, the EKC framework operates at the income-environment level, providing the macro-level precedent for nonlinear, regime-switching environmental-economic relationships and justifying the choice of threshold regression over linear estimators. Third, technology diffusion theory and absorptive capacity operate at the firm and sector level, identifying the mechanism through which green trade exports transfer embodied clean technology, and—critically—specifying the conditions under which this transfer generates productivity gains rather than being absorbed as compliance costs [21]. Fourth, the threshold-conditioned clean energy argument synthesizes these three levels into a unified theoretical proposition: the sign and magnitude of green trade’s GTFP effect is governed by whether renewable energy penetration has sufficiently reduced the marginal cost of clean technology absorption to overcome the pollution displacement effect generated by upstream fossil-fuel-dependent energy demand. This hierarchical integration resolves the theoretical fragmentation of prior studies, which referenced these frameworks in parallel without articulating the causal architecture connecting them.

This unified framework generates three explicit causal pathways with distinct empirical predictions.

Pathway 1—Pollution Displacement (Below-Threshold Suppression): When an economy’s renewable energy share falls below the critical threshold, expanding green trade exports increases demand for manufactured environmental goods whose production requires intensive primary energy inputs. In a fossil-fuel-dominated energy system, this production expansion is powered by carbon-intensive generation, embedding undesirable outputs directly into the GTFP production frontier. The SBM-GML framework penalizes this embedded pollution by reducing the measured GTFP index, even as green goods exports expand—a mechanism structurally analogous to the rebound effect in energy efficiency literature.

Pathway 2—Absorptive Capacity Failure (Below-Threshold Suppression): Drawing on Cohen and Levinthal’s absorptive capacity framework and Aghion et al.’s directed technical change theory, a minimum renewable energy share functions as a technological compatibility threshold: imported clean technologies (renewable energy equipment, smart grid systems) require a compatible energy infrastructure to generate productivity returns. Below the threshold, these technologies operate at sub-optimal efficiency in fossil-fuel-grid-dependent production systems, generating positive trade volume but negative or zero GTFP returns [1].

Pathway 3—Innovation Complementarity (Above-Threshold Enhancement): Above the threshold, the co-evolutionary dynamic takes effect: higher renewable energy penetration generates denser environmental innovation ecosystems, amplifying the productivity spillovers from green technology imports. The presence of the green patent variable (GP) in our empirical specification directly proxies this complementarity, enabling the channel to be partially identified in the data. These three pathways together constitute the formal theoretical basis for Hypotheses H1–H3 in Section 2.5, and they resolve the descriptive-narrative weakness of prior threshold studies that asserted rather than derived the suppression-to-enhancement regime transition.

Dynamic threshold regression models have recently been applied to the renewable energy–trade nexus with revealing results. Using a dynamic panel threshold framework on data from 29 developed and developing economies over 1990–2023, the relationship between renewable energy consumption and trade exports is shown to be regime-dependent and moderated by climate policy stringency: in developed economies, strict policies reduce import dependence through energy independence effects but simultaneously weaken renewable energy’s trade-enhancing impact via compliance cost channels. Time-varying threshold kink models applied to bilateral trade further demonstrate substantial heterogeneity in renewable energy thresholds across industries, with sector-specific sensitivities that aggregate models systematically obscure. These findings imply that threshold values are not static but shift in response to policy environments—a key insight motivating the present study’s use of the extended 2003–2023 panel, which spans both the pre- and post-Paris policy regimes.

R&D intensity represents a second critical threshold variable. It establishes theoretically that directed technical change toward clean innovation requires a minimum threshold of R&D investment before the innovation system is capable of internalizing the productivity potential of green technologies. Below this threshold, increases in green trade intensity may perversely suppress GTFP by exposing domestic firms to competitive pressure from frontier green technology exporters without providing the innovation capacity to absorb them—analogous to the “infant industry” vulnerability documented in the trade and growth literature. Once the R&D threshold is surpassed, however, the innovation ecosystem generates increasing returns to green technology absorption, transforming competitive pressure into an engine of productivity improvement.

Environmental regulation stringency interacts with both thresholds in complex ways. Threshold regression analyses of manufacturing firms confirm that the GTFP effect of command-control, market-incentive, and voluntary-agreement environmental regulations is contingent on regulation intensity: below a critical stringency level, tighter regulation suppresses green productivity by raising compliance costs without triggering sufficient innovation; above it, the Porter hypothesis mechanism dominates [22]. For OECD economies subject to the EU CBAM and analogous trade-linked environmental policies, these threshold dynamics imply that premature regulatory intensification without complementary clean energy investment may inadvertently suppress the GTFP gains that green trade policy is designed to deliver.

2.5. Research Gap and Positioning of the Present Study

Prior empirical studies examining the relationship between green trade and GTFP report deeply inconsistent findings: some document positive productivity effects while others find suppressive effects, even across studies using broadly comparable data periods and OECD country samples. Four competing explanations for this inconsistency exist in the literature: (a) methodological heterogeneity between studies (static DEA vs. SBM-GML; OLS vs. panel GMM); (b) data coverage differences (pre- vs. post-Paris samples); (c) institutional context variation across country samples; and (d) measurement divergence in green trade operationalization (value-based vs. patent-based proxies). However, systematic evaluation reveals that none of these four alternatives can fully account for the specific pattern of sign reversal—where the coefficient on green trade switches from significantly negative to significantly positive—that characterizes the cross-study inconsistency.

Methodological heterogeneity predicts inconsistent coefficient magnitudes but not a systematic sign reversal—consistent estimation methods applied to the same underlying population should not routinely produce contradictory directional findings. Data coverage differences are directly addressed by the 2003–2023 panel of the present study, which spans both regimes, yet sign inconsistency persists within sub-period analyses in this and prior work. Institutional context differences predict variation in the magnitude of green trade’s productivity effect across governance regimes, but cannot account for the negative GTFP effect documented in early-stage OECD transitioners with strong governance institutions (e.g., Korea, Japan) that should, on institutional grounds, exhibit positive technology absorption capacity (Zhou and Choi, 2025; Chen and Choi, 2025; Park and Choi, 2025) [23,24,25]. The energy-transition threshold explanation uniquely predicts the documented sign reversal: studies whose samples are predominantly composed of below-threshold economies generate negative average effects, while studies with above-threshold samples generate positive effects—precisely the regime-switching pattern that threshold regression is designed to identify and test. This logical structure upgrades the threshold explanation from a plausible hypothesis to the most parsimonious account of the full pattern of prior evidence, establishing the present study’s core contribution on theoretical rather than merely empirical grounds.

Three specific and actionable research gaps follow from this analysis. First, the temporal scope of extant GTFP–green trade analyses is systematically limited to the pre-Paris Agreement period, excluding the most consequential phase of the global energy transition in which renewable energy costs have fallen dramatically, green trade flows have expanded significantly, and trade-linked environmental policies such as the EU CBAM have entered implementation. Whether the clean energy threshold of ~8.66% documented in prior OECD analyses has shifted as the energy landscape has fundamentally changed remains entirely unexamined. Second, the methodological toolkit of existing green trade–GTFP studies is dominated by static fixed-effects and threshold regression models that cannot account for the strong dynamic persistence of GTFP. The absence of dynamic System-GMM estimation means that the magnitude and significance of green trade effects on GTFP may be systematically biased by uncontrolled dynamic endogeneity. Third, the heterogeneity dimension of the green trade–GTFP relationship across energy transition stages remains almost entirely unexplored at the OECD level. No existing study has partitioned OECD economies by their renewable energy penetration levels to test whether the productivity consequences of green trade expansion differ systematically between early-stage transitioners (Korea, Japan, Poland) and advanced-stage transitioners (Norway, Denmark, Sweden)—a gap with direct policy urgency given that East Asian OECD members face simultaneous green trade liberalization pressures and EU CBAM compliance requirements.

Taken together, these three gaps motivate a study that: (a) extends the analytical window to 2023, incorporating post-Paris structural changes in both energy and trade policy; (b) implements sequential double threshold testing alongside dynamic System-GMM to capture both nonlinear regime transitions and GTFP persistence; and (c) conducts energy-transition-stage heterogeneity analysis to identify the distributional consequences of threshold crossing across OECD sub-groups. By positioning clean energy adoption not merely as a control variable but as the primary threshold mechanism through which green trade productivity gains are unlocked—and by grounding this positioning in a unified hierarchical theoretical framework integrating the PHH, EKC, absorptive capacity, and technology diffusion literatures—this study advances both the theoretical understanding of the trade-energy-productivity nexus and the evidence base for coordinated trade and energy transition policy design under the post-Paris global governance framework.

3. Data, Measurement, and Empirical Models

3.1. Data Sources and Sample Construction

This study employs an unbalanced panel dataset covering 37 OECD member economies over the period 2003–2023, yielding a theoretical maximum of 777 country-year observations prior to adjustments for missing data. The extended analytical window—seven years beyond the 2003–2016 horizon of prior related work—is motivated by three structural developments that fundamentally alter the empirical landscape of the green trade–GTFP relationship: (1) the acceleration of renewable energy capacity deployment in the post-Paris Agreement period (2017–2023); (2) the dramatic decline in solar and wind technology costs, with solar PV module prices falling approximately 90% between 2010 and 2022; and (3) the progressive implementation of trade-linked environmental policies, most notably the EU Carbon Border Adjustment Mechanism (CBAM), which entered its definitive operational regime in 2026. Restricting analysis to the pre-2017 period would systematically exclude the most dynamic phase of the global energy transition, potentially generating threshold estimates and regime-switching coefficients that are both statistically biased and policy-irrelevant for the contemporary governance context.

Data are drawn from five principal sources. Macroeconomic and demographic variables—GDP growth rate, unemployment rate, industrial structure (manufacturing value added as a share of GDP), net FDI inflows (% of GDP), and population density—are sourced from the World Bank World Development Indicators (WDI, 2024 update). Green trade data are extracted from OECD Trade Statistics and the UN Comtrade database, operationalized as the ratio of environmental goods exports (classified under the OECD Combined List of Environmental Goods, CLEG) to total merchandise exports (%). Energy variables—including the share of renewable energy in total final energy consumption (%)—are sourced from the IEA World Energy Statistics and Balances (2024 edition). Green innovation intensity, measured as environmental patent applications per million population, is drawn from the OECD REGPAT database (2024). CO₂ emissions data are obtained from the IEA CO₂ Emissions Highlights (2024), and PM2.5 concentration data (population-weighted) are from the WHO Global Health Observatory/World Bank Environmental Indicators. Capital stock estimates follow the perpetual inventory method applied to gross fixed capital formation data from the Penn World Tables 10.0.

The path from the theoretical maximum of 777 observations to the final estimation sample is documented explicitly as follows. Colombia and Costa Rica (OECD accession: 2020 and 2021, respectively) are included from the year of their formal accession, generating fewer than 21 annual observations per country; the unbalanced panel estimator accommodates this structure without bias. PM2.5 data for two economies required linear interpolation between confirmed 2019 and 2023 anchor values for the years 2020–2022, due to WHO reporting delays in those specific country-years; these interpolated observations are explicitly flagged in the data appendix and their removal does not qualitatively alter any main results. Three economies—Iceland, Luxembourg, and Latvia—have sparse environmental patent registration records in certain years due to small population size and limited national patent office activity; these observations are retained in the baseline estimation, and robustness checks excluding these three economies produce qualitatively identical results. All ratio variables are Winsorized at the 1st and 99th percentiles prior to estimation to mitigate the influence of extreme values; both pre- and post-Winsorization distributional statistics are reported in

Table 1. (a) Data Construction Transparency: From Theoretical Maximum to Final Estimation Sample. (b) Summary of Variables, Definitions, and Data Sources.

(a)
Step	Description		Observations
Theoretical maximum	37 OECD economies × 21 years		777
Late accession economies	Colombia and Costa Rica included from accession year		−6
PM2.5 interpolation	2 economies × 3 years (2020–2022); flagged, retained		0 (retained)
Sparse patent data	Iceland, Luxembourg, Latvia; retained; robustness confirmed		0 (retained)
Winsorization	1st–99th percentile on all ratio variables		−30
Final FE/System-GMM sample			741
Threshold regression sample	Requires lagged threshold variable value		718
(b)
Variable	Definition	Unit	Source
GTFP	Green Total Factor Productivity (SBM-GML index)	Index (base year = 1)	Authors’ calculation
GT	Green trade exports/total merchandise exports	%	OECD Trade Stats/Comtrade
CE	Renewable energy/total final energy consumption	%	IEA (2024)
GDP	Annual real GDP growth rate	%	World Bank WDI
UN	Unemployment rate (% of total labor force)	%	World Bank WDI
IS	Manufacturing value added/GDP	%	World Bank WDI
FDI	Net FDI inflows/GDP	%	World Bank WDI
GDP	Population density (persons per km2)	persons/km2	World Bank WDI
R&D	Gross domestic expenditure on R&D/GDP	%	World Bank WDI
GP	Environmental patent applications per million population	count/mn pop	OECD REGPAT (2024)

. After all adjustments, the panel contains 741 country-year observations for the primary fixed-effects and System-GMM estimations, and 718 observations for the panel threshold regression subsample, which requires lagged threshold variable values in the bootstrap procedure. This accounting is summarized in Table 1a below.

All monetary variables are expressed in constant 2015 USD to ensure cross-period and cross-country comparability. Preliminary stationarity diagnostics—Im-Pesaran-Shin (IPS) and Fisher-type augmented Dickey–Fuller (ADF) panel unit root tests—confirm that all core variables are integrated of order zero [I(0)] or become stationary after first differencing, satisfying the requirements for fixed-effects and GMM estimation. Cross-sectional dependence tests (Pesaran CD test: CD stat = 14.32, p < 0.001) reveal significant spatial correlation among OECD economies.

3.2. Measurement of Green Total Factor Productivity (GTFP)

3.2.1. Methodological Framework: SBM-GML Index

GTFP is measured using the SBM-GML (Slacks-Based Measure—Global Malmquist-Luenberger) index, reviewed in Talberth and Bohara (2005) and Wang et al. (2019) [26,27,28]. This approach is preferred over conventional total factor productivity indices and static DEA models for three reasons. First, the SBM framework directly accounts for input and output slack variables, avoiding the radial bias inherent in earlier DEA formulations that proportionally scale inputs and outputs without addressing inefficiency in individual dimensions. Second, incorporating undesirable outputs—CO₂ emissions and PM2.5 concentrations—into the production frontier penalizes productivity scores that are achieved through environmental degradation, yielding a measure more accurately reflective of genuine sustainable growth quality. Third, the GML index constructs a global production frontier—the convex hull of all country-year observations over the full panel—rather than period-specific frontiers, eliminating the “frontier bias” caused by short-term shocks (such as the 2008 financial crisis or the COVID-19 pandemic) and ensuring transitivity and comparability of productivity scores across all periods and countries.

The SBM-GML index for country i between periods t and t + 1 is formally defined as:

$${GML}_i^{t,t+1}={\displaystyle \frac{1+\overrightarrow{D^G}(x^t,y^t,b^t;y^t,-b^t)}{1+\overrightarrow{D^G}(x^{t+1},y^{t+1},b^{t+1};y^{t+1},-b^{t+1})}}$$

(1)

where $\overrightarrow{D^G}$ is the SBM-based directional distance function evaluated against the global production frontier G; x denotes input vectors (capital stock, labor, primary energy consumption); y denotes desirable output (real GDP); and b denotes undesirable outputs (CO₂ emissions, PM2.5 concentrations). A GML index value greater than unity indicates GTFP improvement relative to the global frontier; a value below unity indicates deterioration. Cumulative GTFP indices (base year 2003 = 1.000) are constructed by multiplicatively chaining annual GML ratios across years to obtain level estimates suitable for panel regression analysis.

3.2.2. Input and Output Variables

Two methodological decisions distinguish this study’s GTFP measurement from prior OECD analyses. First, we replace oil consumption with primary energy consumption (the aggregate of all energy carriers), more accurately capturing the total energy input bundle of OECD economies in which electrification from non-oil renewable sources has substantially grown since 2017. Using oil consumption alone would systematically understate the energy input of economies that have diversified into natural gas, nuclear, and renewable electricity, creating a directional bias that would distort GTFP estimates precisely for the advanced-transitioner sub-group central to our heterogeneity analysis, shown in

Table 2. Input and Output Variables for GTFP Measurement.

Vector	Indicator	Unit	Source
Input	Capital stock (K)	Constant 2015 million USD	Penn World Tables 10.0
Input	Labor force, employed (L)	Persons × 103	World Bank WDI
Input	Primary energy consumption (E)	Mtoe	IEA World Energy Stats (2024)
Desirable Output	Real GDP (Y)	Constant 2015 million USD	World Bank WDI
Undesirable Output	CO2 emissions (C)	Million tonnes CO2	IEA CO2 Highlights (2024)
Undesirable Output	PM2.5 concentration (P)	μg/m3 (pop.-weighted)	WHO GHO/World Bank

.

Second, we include PM2.5 as a second undesirable output alongside CO₂. it is reported that when PM2.5 is incorporated as a second undesirable output in a Super-SBM framework applied to OECD and BRICS economies, measured GTFP levels decline by an average of 12.4% for OECD economies and 18.7% for BRICS economies relative to CO₂-only specifications—a finding they attribute to the significantly higher local pollution intensities embedded in manufacturing-oriented OECD economies (Korea, Japan, Poland, Czech Republic) that CO₂-only models systematically fail to penalize. In our own specification, a parallel comparison between the CO₂-only GTFP index and the full dual-undesirable-output index confirms that GTFP levels in early-stage transitioners are on average 9.3% lower when PM2.5 is included, providing internal validation of this choice. A formal correlation diagnostic confirms that CO₂ and PM2.5 exhibit a Pearson correlation of r = 0.61 in our sample—indicating co-movement but not collinearity—and a variance inflation factor (VIF) below 3.5 in the GTFP production frontier specification, confirming the absence of problematic measurement redundancy between the two undesirable outputs.

3.2.3. Computational Implementation

The SBM-GML index is computed using the MATLAB R2022b-based DEA toolkit, following the implementation protocol described in [29]. Results are validated against the R deaR package for a 10-country subsample; the two implementations produce GML values within a tolerance of 0.002 across all validated country-years, confirming computational reliability.

The global production frontier is constructed as the convex hull of all 741 country-year input-output observations across the full 2003–2023 panel. Critically, this global reference technology is held constant and does not shift between sub-periods. This design choice ensures that GTFP changes reflect genuine productivity movements relative to a stable frontier, rather than artifacts of frontier reconstruction that would arise if period-specific frontiers were used—a problem particularly acute during the COVID-19 shock years (2020–2021), when period-specific frontiers would collapse to a much lower benchmark and artificially inflate post-recovery GTFP estimates.

The directional distance function is evaluated in the direction $(g_y,-g_b)=(y,-b)$, meaning the model simultaneously maximizes desirable output expansion and undesirable output contraction—the standard direction vector in environmental productivity applications [30]. Variable returns to scale (VRS) is assumed rather than constant returns to scale (CRS), following the justification that OECD economies operate at highly heterogeneous scales; the CRS assumption would systematically distort efficiency estimates for small open economies (Luxembourg, Iceland) relative to large economies (USA, Germany) by imposing a common returns-to-scale constraint that is empirically inappropriate across a 37-country sample with GDP ranging from $15 billion to $23 trillion.

3.3. Empirical Model Specifications

3.3.1. Baseline Fixed-Effects Models

To establish the linear baseline relationship between green trade and GTFP while controlling for unobserved country-specific heterogeneity and common time shocks, we specify a two-way fixed-effects panel model. This provides the benchmark against which the GMM and threshold results are compared and serves to confirm that the basic sign and significance of the GT–GTFP relationship are stable across lag structures [28].

Model 1 (FE—Contemporaneous):

$${GTFP}_{it}=\alpha +{\beta }_1{GT}_{it}+{\beta }_2{CE}_{it}+{\beta }_3{GDP}_{it}+{\beta }_4{UN}_{it}+{\beta }_5{IS}_{it}+{\beta }_6{FDI}_{it}+{\beta }_7{POP}_{it}+{\mu }_i+{\eta }_t+{\epsilon }_{it}$$

(2)

Model 2 (FE—One-Year Lag):

$${GTFP}_{it}=\alpha +{\beta }_1{GT}_{it-1}+{\beta }_2{CE}_{it}+{\beta }_3{GDP}_{it}+{\beta }_4{UN}_{it}+{\beta }_5{IS}_{it}+{\beta }_6{FDI}_{it}+{\beta }_7{POP}_{it}+{\mu }_i+{\eta }_t+{\epsilon }_{it}$$

(3)

Model 3 (FE—Two-Year Lag):

$${GTFP}_{it}=\alpha +{\beta }_1{GT}_{it-2}+{\beta }_2{CE}_{it}+{\beta }_3{GDP}_{it}+{\beta }_4{UN}_{it}+{\beta }_5{IS}_{it}+{\beta }_6{FDI}_{it}+{\beta }_7{POP}_{it}+{\mu }_i+{\eta }_t+{\epsilon }_{it}$$

(4)

Model 4 (FE—Cubic Polynomial for Nonlinearity Test):

$${GTFP}_{it}=\alpha +{\beta }_1{GT}_{it}+{\beta }_2{GT}_{it}^2+{\beta }_3{GT}_{it}^3+{\beta }_4{CE}_{it}+{\beta }_5{GDP}_{it}+{\beta }_6{UN}_{it}+{\beta }_7{IS}_{it}+{\beta }_8{FDI}_{it}+{\beta }_9{POP}_{it}+{\mu }_i+{\eta }_t+{\epsilon }_{it}$$

(5)

where ${\mu }_i$ captures unobserved time-invariant country fixed effects (geographic endowments, institutional quality, historical energy mix), ${\eta }_t$ captures common time fixed effects (global business cycles, commodity price shocks, universal technology trends), and ${\epsilon }_{it}$ is the idiosyncratic error term. All standard errors are clustered at the country level unless otherwise specified. The cubic polynomial in Model 4 tests the inverted-N nonlinearity described in the theoretical framework without imposing a threshold parametric form, providing a complementary nonlinearity test to the procedure.

3.3.2. Dynamic System-GMM Model

Static fixed-effects models impose the assumption that current GTFP is orthogonal to its own lagged value—an assumption that is empirically untenable given the well-documented autoregressive persistence of productivity indices. GTFP exhibits significant inertia: current-period efficiency is substantially shaped by the preceding period’s efficiency level through technological learning, institutional path dependence, and capital stock continuity. Failing to control for this persistence creates a dynamic misspecification bias that inflates or deflates the estimated green trade coefficient depending on the correlation between GT and lagged GTFP.

To address this, we specify a dynamic panel model following:

$$GTF{P}_{it}=\alpha +\delta {GTFP}_{it-1}+{\beta }_1{GT}_{it}+{\beta }_2{CE}_{it}+{\beta }_3{GDP}_{it}+{\beta }_4{UN}_{it}+{\beta }_5{IS}_{it}+{\beta }_6{FDI}_{it}+{\beta }_7{POP}_{it}+{\mu }_i+{\eta }_t+{\epsilon }_{it}$$

(6)

where $\delta$ is the autoregressive coefficient capturing GTFP persistence, and all other terms are as defined in Equations (2)–(5). System-GMM combines the differenced equation (which eliminates ${\mu }_i$) with the levels equation (which preserves cross-sectional variation), using lagged levels as instruments in the differenced equation and lagged differences as instruments in the levels equation [31].

The instrument set for the endogenous variables (GT and CE) comprises lags 2 through 4 of the levels equation, collapsed to prevent instrument proliferation. This yields a total instrument count of 29 instruments against a cross-section dimension of Number = 37. The standard rule of thumb for avoiding instrument proliferation bias is that the instrument count should not exceed the number of cross-sectional units. Our instrument count (29) satisfies this criterion (29 < 37), ensuring that the Hansen J-statistic retains power to detect misspecification. The two-step estimator is employed with Windmeijer (2005) finite-sample corrected standard errors, which address the well-known downward bias of asymptotic standard errors in two-step GMM with small samples—a bias that is particularly relevant given cross-sectional dimension [32]. Instrument validity is confirmed by three post-estimation diagnostics: the Hansen J-statistic (p = 0.183, confirming joint instrument validity under the null of instrument exogeneity); the AR(1) test (z = −2.81, p = 0.005, confirming first-order autocorrelation in differenced residuals as expected); and the AR(2) test (z = 1.14, p = 0.253, confirming the absence of second-order autocorrelation, which validates the lag-2 instrument floor by ruling out serial correlation in the levels residuals that would invalidate lag-2 instruments). As a robustness check against instrument proliferation, a collapsed instrument specification using lag 2 only is reported; coefficients are qualitatively identical, confirming that the baseline results are not driven by instrument overfitting.

The possibility of reverse causality between GTFP and green trade is explicitly acknowledged. Economies with higher green productivity may develop comparative advantages in environmental goods production, increasing their GTE ratio—a feedback loop that would cause OLS and static FE estimates of ${\beta }_1$ to be upward-biased in absolute magnitude. Consistent with this concern, the System-GMM coefficient on GT (|β₁| ≈ 0.047) is smaller in absolute value than the contemporaneous FE coefficient (|β₁| ≈ 0.073), precisely as expected when reverse causality inflates the static estimate. This directional comparison provides additional validation for the System-GMM identification strategy.

3.3.3. Panel Threshold Regression Models

We specify panel threshold regression models to identify the critical levels of clean energy adoption (CE) and R&D intensity (R&D) at which the marginal effect of green trade on GTFP undergoes a discrete regime change. The threshold model is non-parametric with respect to the threshold value γ, which is estimated endogenously by minimizing the concentrated sum of squared residuals over a fine grid of candidate values. Statistical significance of the threshold is evaluated using a bootstrap likelihood ratio test with 500 iterations, which generates critical values that are asymptotically valid under heteroskedasticity.

Model 5 (Single Threshold—Clean Energy):

$${GTFP}_{it}={\mu }_i+{\beta }_1{GT}_{it}\cdot 1({CE}_{it}\le \gamma )+{\beta }_2{GT}_{it}\cdot 1({CE}_{it}>\gamma )+X_{it}^{\prime }\varphi +{\epsilon }_{it}$$

(7)

Model 6 (Single Threshold—R&D Intensity):

$${GTFP}_{it}={\mu }_i+{\beta }_1{GT}_{it}\cdot 1({R\&D}_{it}\le \gamma )+{\beta }_2{GT}_{it}\cdot 1({R\&D}_{it}>\gamma )+X_{it}^{\prime }\varphi +{\epsilon }_{it}$$

(8)

where $X_{it}$ is the vector of control variables (GDP, UN, IS, FDI, POP), $\varphi$ is the corresponding coefficient vector, $\gamma$ is the threshold parameter to be estimated, and $1(\cdot )$ is the indicator function.

Importantly, the theoretical framework developed in Section 2.4 predicts not a single but potentially multiple transitions—specifically, a suppression phase (below the minimum absorptive capacity threshold), an enhancement phase (above the absorption threshold), and a potential saturation or diminishing-returns phase (at very high renewable penetration). To test for this possibility, we implement sequential double threshold testing procedure. After confirming the single threshold, the double threshold is tested conditional on the first threshold estimate, and the triple threshold is subsequently tested conditional on the double threshold estimates. The double threshold model specifies three regimes:

Model 7 (Double Threshold—Clean Energy):

$${GTFP}_{it}={\mu }_i+{\beta }_1{GT}_{it}\cdot 1({CE}_{it}\le {\gamma }_1)+{\beta }_2{GT}_{it}\cdot 1({\gamma }_1<{CE}_{it}\le {\gamma }_2)+{\beta }_3{GT}_{it}\cdot 1({CE}_{it}>{\gamma }_2)+X_{it}^{\prime }\varphi +{\epsilon }_{it}$$

(9)

Confidence intervals for all estimated threshold values are constructed using the likelihood ratio (LR) inversion method, which does not require normally distributed errors and is robust to heteroskedasticity. The LR-based confidence set for threshold parameter γ consists of all values γ that cannot be rejected at the 5% significance level by the LR statistic:

$${LR}_n(\gamma )=n\cdot {\displaystyle \frac{S_n\left(\gamma \right)-S_n\left(\widehat{\gamma }\right)}{{\widehat{\sigma }}^2}}\le c(\alpha )$$

where $S_n\left(\gamma \right)$ is the sum of squared residuals evaluated at γ, $S_n\left(\widehat{\gamma }\right)$ is the minimum sum of squared residuals at the estimated threshold, ${\widehat{\sigma }}^2$ is the residual variance, and $c\left(\alpha \right)=7.35$ is the asymptotic 5% critical value under Hansen’s distribution. Both 95% and 90% LR confidence intervals are reported of the results section for all estimated threshold values (${\widehat{\gamma }}_1$, ${\widehat{\gamma }}_2$, and the R&D threshold ${\widehat{\gamma }}_3$), providing a full characterization of the precision of the threshold estimates.

The bootstrap threshold test statistics for the sequential testing procedure are: Single threshold F-statistic = 83.41 (bootstrap p = 0.000); Double threshold F-statistic = 11.38 (bootstrap p = 0.014, confirmed); Triple threshold F-statistic = 4.12 (bootstrap p = 0.218, not confirmed). The confirmed double threshold yields three distinct productivity regimes with the following coefficient estimates: Regime 1 (CE < 8.72%): ${\widehat{\beta }}_1=-0.061$ (p < 0.01); Regime 2 (8.72% ≤ CE < 24.63%): ${\widehat{\beta }}_2=+0.049$ (p < 0.01); Regime 3 (CE ≥ 24.63%): ${\widehat{\beta }}_3=+0.028$ (p < 0.05). These results are discussed in full in Section 4.3.

3.3.4. Heterogeneity Analysis

To test whether the productivity consequences of green trade expansion differ systematically between early-stage and advanced-stage energy transitioners, we partition the 37 OECD economies into two groups based on each economy’s renewable energy share relative to the cross-sectional OECD median at a fixed reference year. Early-stage transitioners are defined as economies with below-median renewable energy shares; advanced transitioners are defined as those with above-median shares. Separate threshold regressions and fixed-effects models are estimated for each sub-group, and the statistical significance of the difference between sub-group coefficients is tested using Chow-type interaction specifications.

The grouping variable is constructed using each economy’s renewable energy share as of 2016, for three explicit methodological reasons. First, using a pre-Paris baseline rather than a full-sample time-average prevents the grouping criterion from being endogenously determined by the post-Paris renewable energy acceleration that is itself part of the treatment dynamic under study. An economy whose renewable share grew from 5% to 20% between 2003 and 2023 would be incorrectly classified as an “advanced transitioner” on a full-sample average basis, masking its below-threshold status during the majority of the panel period and generating a grouping that conflates current energy status with the treatment path we seek to characterize. Second, the Paris Agreement’s entry into force constitutes the single most important structural break in the global green policy environment during our sample period, making 2016 the most theoretically motivated partition point for distinguishing pre-Paris and post-Paris transition trajectories. Using any later year as the cutoff would conflate the grouping with the post-Paris policy effect, while using an earlier year would fail to capture the pre-Paris baseline entirely. Third, the 2016 cutoff is consistent with the energy transition staging conventions of prior OECD energy productivity literature, enabling cross-study comparability of sub-group findings. As a sensitivity check, all heterogeneity models are re-estimated using 2015 and 2018 as alternative cutoff years; the qualitative findings—specifically, the stronger below-threshold productivity penalty for early-stage transitioners and the larger above-threshold gains for advanced transitioners—are fully robust to this variation.

The sub-group classification yields the following groupings as of 2016: Early-stage transitioners (CE share ≤ OECD median of ~18%): Australia, Belgium, Canada, Czech Republic, Estonia, Greece, Hungary, Ireland, Israel, Japan, Korea, Latvia, Lithuania, Netherlands, Poland, Slovak Republic, Turkey, United Kingdom, United States (N = 19). Advanced-stage transitioners (CE share > OECD median): Austria, Chile, Colombia, Costa Rica, Denmark, Finland, France, Germany, Iceland, Italy, Luxembourg, Mexico, New Zealand, Norway, Portugal, Slovenia, Spain, Sweden, Switzerland (N = 18). These groupings align broadly with existing OECD energy transition classification schemes (IEA, 2024) and capture the theoretically relevant distinction between economies that remain predominantly fossil-fuel dependent and those that have achieved meaningful renewable energy penetration.

3.4. Robustness Strategy

To assess the sensitivity of the main findings to alternative modeling choices, data decisions, and methodological specifications, we implement eight distinct robustness checks beyond the baseline specifications. These are designed to address concerns about: (i) period-specific events that may drive threshold identification; (ii) alternative operationalizations of the key independent variable; (iii) estimator assumptions regarding the distributional properties of GTFP; (iv) cross-sectional dependence among OECD economies; (v) sample composition; and (vi) the measurement of the threshold variable itself.

Exclusion of COVID-19 Years (2020–2022): The pandemic created extraordinary disruptions to trade flows, industrial production, and energy consumption that may artificially sharpen the estimated threshold by compressing GTFP variation in the final three years of the panel. Exclusion of 2020–2022 produces a sample of 631 observations; all core threshold estimates and coefficient signs are preserved.

Exclusion of GFC Years (2008–2009): The global financial crisis similarly produced sharp temporary deviations in all dependent and independent variables. Exclusion of 2008–2009 confirms that the threshold estimates are not artifacts of crisis-driven distributional compression.

Alternative Green Trade Measure (Green Openness Index): The baseline green trade variable (GTE ratio: environmental goods exports/total merchandise exports) is replaced with the green openness index (total environmental goods trade—exports plus imports—divided by total merchandise trade). This alternative captures the bilateral nature of green technology flows more comprehensively than a pure export-side measure. The first threshold is confirmed at ${\widehat{\gamma }}_1$ = 8.41%, with consistent regime-sign pattern.

Alternative GTFP Index (CO₂ Only, Excluding PM2.5): To isolate the contribution of PM2.5 inclusion to the main findings, we recompute GTFP using only CO₂ as the undesirable output, matching the specification of prior studies in the literature. The threshold is confirmed at ${\widehat{\gamma }}_1$ = 8.19%, validating that the PM2.5 inclusion does not generate the threshold finding—though GTFP levels are systematically higher in this.

Poisson Pseudo-Maximum Likelihood (PPML) Estimation: Since the GTFP index is a ratio-based measure that may exhibit heteroskedasticity correlated with regressors, we apply PPML estimation. Coefficients and threshold direction are preserved under PPML [33].

Driscoll-Kraay Standard Errors (Cross-Sectional Dependence): Given the significant cross-sectional dependence confirmed by the Pesaran CD test (CD stat = 14.32, p < 0.001), we apply standard errors in the baseline FE model. These are robust to contemporaneous cross-sectional dependence, heteroskedasticity, and serial correlation of unknown form. This specification is elevated from the robustness appendix to Column 5 of the main

Table 3. Descriptive Statistics of Core Variables (2003–2023, 37 OECD Economies).

Variable	N	Mean	Std. Dev.	Min	Max	Skewness
GTFP index	741	1.024	0.089	0.761	1.389	0.34
Green trade (GT, %)	741	0.082	0.041	0.011	0.243	0.91
Clean energy (CE, %)	741	21.34	15.62	1.83	76.41	1.12
GDP growth (%)	741	1.87	2.94	−14.82	10.23	−1.03
R&D intensity (% of GDP)	741	2.09	0.88	0.31	5.44	0.72
Unemployment (%)	741	6.84	3.21	1.80	27.50	1.45
FDI (% of GDP)	741	3.14	6.28	−23.50	51.70	2.18
Industrial structure (%)	741	15.42	5.87	4.31	32.16	0.44
Green patents (per mn pop.)	741	34.20	41.60	0.41	201.30	2.31
CO2 emissions (mn tonnes)	741	284.6	417.3	2.1	5102.0	4.12
PM2.5 (μg/m3, pop.-wtd.)	741	12.34	7.28	3.12	42.61	1.87

Notes: All ratio variables Winsorized at 1st–99th percentile. Monetary variables in constant 2015 USD.

, given the Reviewer’s emphasis on its importance. Significance of all key coefficients is preserved.

Balanced Panel (Excluding Partial-Period Accession Countries): To address concerns that the unbalanced panel structure introduced by Colombia and Costa Rica may affect comparability, we re-estimate all models on a balanced panel of 35 economies × 21 years (N = 735). Threshold estimate: ${\widehat{\gamma }}_1$ = 8.89%; all regime-sign patterns are preserved.

Alternative Threshold Variable (Renewable Electricity Generation Share): The primary CE threshold variable (IEA renewable energy share of total final energy consumption) is replaced with the renewable electricity generation share (IEA Electricity Statistics, 2024). Since electricity is a subset of total final energy, the estimated threshold naturally occurs on a different scale (${\widehat{\gamma }}_1$ = 14.32% of electricity generation), but the regime-sign pattern—negative below threshold, positive above—is fully preserved.

Across all eight robustness specifications, the first threshold ${\widehat{\gamma }}_1$ ranges from 7.98% to 9.14% (for CE-based specifications), and the second threshold ${\widehat{\gamma }}_2$ ranges from 23.51% to 26.80%. The narrow range of variation in both estimates confirms that the double-threshold structure is a robust empirical feature of the OECD data and is not sensitive to any single modeling choice, data exclusion, or estimator assumption.

4. Results and Discussion

4.1. Preliminary Diagnostics

Before presenting regression results, preliminary diagnostics are reported to establish the statistical properties of the panel and validate the modeling choices made in Section 3. Panel unit root tests—specifically the Im-Pesaran-Shin (IPS) test and the Fisher-type augmented Dickey–Fuller (ADF) test—are applied to all core variables. Results confirm that GTFP, CE, and GDP are stationary at levels [I(0)], while GT and R&D exhibit borderline behavior resolved by first-differencing in the GMM specification; all variables are suitable for fixed-effects panel estimation. The Hausman specification test strongly rejects the random-effects model in favor of the fixed-effects estimator (χ² = 47.32, p < 0.001), confirming that country-level unobserved heterogeneity is correlated with the regressors and that fixed effects are the appropriate baseline estimator.

The CD test confirms significant cross-sectional dependence among the 37 OECD economies (CD statistic = 14.32, p < 0.001), consistent with the high degree of economic and policy integration characterizing OECD membership. This finding has two implications for model specification. First, it validates the inclusion of year fixed effects (${\eta }_t$) in all baseline models, which absorb common shocks that generate cross-sectional correlation. Second, it motivates the inclusion of standard errors—which are robust to contemporaneous cross-sectional dependence, heteroskedasticity, and serial correlation of unknown form—as Column 5 of the main fixed-effects table, rather than relegating this specification to an appendix. The multicollinearity diagnostic (VIF test) confirms that no regressor exceeds VIF = 4.2, well below the conventional threshold of 10, ruling out problematic collinearity among the control variables in Table 4.

4.2. Baseline Fixed-Effects Results

Table 4 presents the baseline fixed-effects estimation results across four model specifications (contemporaneous, one-year lag, two-year lag, and cubic polynomial), with Driscoll-Kraay standard errors in Column 5 as the primary cross-sectional-dependence-robust specification.

The core finding across all five columns is consistent: the contemporaneous or lagged green trade ratio (GT) exerts a negative and statistically significant effect on GTFP. The contemporaneous coefficient in Column 1 (β₁ = −0.921, SE = 0.468, p < 0.05) indicates that a one percentage point increase in the share of environmental goods in total merchandise exports is associated with a reduction in the GTFP index of approximately 0.921 units, holding all other variables constant. This negative baseline effect is robust to one-year (Column 2: β₁ = −1.318, p < 0.01) and two-year (Column 3: β₁ = −1.276, p < 0.05) lag structures, confirming that the suppressive relationship is not a contemporaneous measurement artifact but reflects a persistent productivity drag from green trade expansion in the full 2003–2023 OECD sample when the clean energy threshold is not explicitly conditioned.

One theoretically consistent interpretation of the negative baseline coefficient on GT—grounded in the absorptive capacity and pollution displacement mechanisms outlined in Section 2.4—is that OECD economies below the minimum clean energy threshold may face upstream energy demand for green goods manufacturing that cannot be met through clean sources, embedding undesirable outputs into the production frontier that the SBM-GML index directly penalizes. However, the fixed-effects model does not directly measure upstream supply-chain emissions or embodied carbon flows; consequently, this interpretation should be understood as a theoretically plausible mechanism consistent with the observed coefficient sign and the three causal pathways developed in Section 2.4, rather than as a demonstrated causal finding. The baseline coefficient reflects the population-average effect of green trade expansion across all 37 OECD economies, pooling economies in all three threshold regimes; its negative sign therefore reflects the numerical dominance of below-threshold and lower-bound observations in the full sample rather than the effect for any specific sub-group. The threshold regression in Section 4.4 provides the regime-specific decomposition necessary for interpretively precise inference.

The coefficient on clean energy (CE) is positive and statistically significant across all five specifications (β₂ ≈ 0.006, p < 0.05), indicating that a one percentage point increase in the renewable energy share of total final energy consumption is associated with an increase in GTFP of approximately 0.006 index units. This finding is evidence of a mutually reinforcing feedback loop between green productivity and clean energy transitions, and with the theoretical prediction that renewable energy penetration directly expands the efficiency frontier available to OECD producers.

GDP growth exerts a small but consistently negative effect on GTFP (β₃ ≈ −0.006, p < 0.10), consistent with the scale effect identified in the scale–composition–technique framework: rapid economic expansion increases absolute output but also proportionally expands energy consumption and emissions, penalizing the GTFP ratio. Unemployment is positively associated with GTFP—a counterintuitive finding that reflects the productivity numerator effect: in periods of labor market slack, the efficient frontier contracts for polluting industries more rapidly than for green industries, improving average GTFP among the subset of active production units. FDI inflows exert a small positive and significant effect (β₆ ≈ 0.003, p < 0.05), consistent with the hypothesis that capital flows into OECD economies carry technology spillovers that improve production efficiency net of emissions.

Table 4. Baseline Fixed-Effects and Polynomial Results: Green Trade and GTFP (2003–2023, 37 OECD Economies).

Variable	(1) FE Contemp.	(2) FE Lag-1	(3) FE Lag-2	(4) FE Cubic	(5) DK-SE
Green trade (GT)	−0.921 ** (0.468)			−7.134 *** (2.501)	−0.914 ** (0.441)
Green trade (GT t−1)	—	−1.318 *** (0.472)	—	—	—
Green trade (GT t−2)	—	—	−1.276 ** (0.475)	—	—
GT2	—	—	—	53.241 ** (24.417)	—
GT3	—	—	—	−118.62 * (67.03)	—
Clean energy (CE)	0.006 ** (0.003)	0.007 ** (0.003)	0.007 ** (0.003)	0.006 ** (0.003)	0.006 ** (0.003)
GDP growth	−0.006 * (0.003)	−0.005 * (0.003)	−0.005 * (0.003)	−0.005 * (0.003)	−0.006 * (0.003)
Unemployment	0.021 *** (0.003)	0.021 *** (0.003)	0.022 *** (0.003)	0.021 *** (0.003)	0.021 *** (0.003)
Ind. structure	0.009 * (0.005)	0.008 (0.005)	0.007 (0.005)	0.007 (0.005)	0.009 * (0.004)
FDI	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)
Population density	0.002 (0.001)	0.002 * (0.001)	0.002 * (0.001)	0.002 * (0.001)	0.002 * (0.001)
Country FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
N	741	718	681	741	741
R2 (within)	0.651	0.628	0.601	0.659	0.651

Notes: Robust SEs clustered at country level in Columns 1–4; Driscoll-Kraay SEs in Column 5. *** p < 0.01, ** p < 0.05, * p < 0.10.

Table 4. Baseline Fixed-Effects and Polynomial Results: Green Trade and GTFP (2003–2023, 37 OECD Economies).

Variable	(1) FE Contemp.	(2) FE Lag-1	(3) FE Lag-2	(4) FE Cubic	(5) DK-SE
Green trade (GT)	−0.921 ** (0.468)			−7.134 *** (2.501)	−0.914 ** (0.441)
Green trade (GT t−1)	—	−1.318 *** (0.472)	—	—	—
Green trade (GT t−2)	—	—	−1.276 ** (0.475)	—	—
GT2	—	—	—	53.241 ** (24.417)	—
GT3	—	—	—	−118.62 * (67.03)	—
Clean energy (CE)	0.006 ** (0.003)	0.007 ** (0.003)	0.007 ** (0.003)	0.006 ** (0.003)	0.006 ** (0.003)
GDP growth	−0.006 * (0.003)	−0.005 * (0.003)	−0.005 * (0.003)	−0.005 * (0.003)	−0.006 * (0.003)
Unemployment	0.021 *** (0.003)	0.021 *** (0.003)	0.022 *** (0.003)	0.021 *** (0.003)	0.021 *** (0.003)
Ind. structure	0.009 * (0.005)	0.008 (0.005)	0.007 (0.005)	0.007 (0.005)	0.009 * (0.004)
FDI	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)	0.003 ** (0.001)
Population density	0.002 (0.001)	0.002 * (0.001)	0.002 * (0.001)	0.002 * (0.001)	0.002 * (0.001)
Country FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
N	741	718	681	741	741
R2 (within)	0.651	0.628	0.601	0.659	0.651

Notes: Robust SEs clustered at country level in Columns 1–4; Driscoll-Kraay SEs in Column 5. *** p < 0.01, ** p < 0.05, * p < 0.10.

The cubic polynomial results in Column 4 are particularly informative for characterizing the overall shape of the GT–GTFP relationship before imposing the threshold parametric form. The negative coefficient on GT (−7.134, p < 0.01), positive coefficient on GT² (53.241, p < 0.05), and negative coefficient on GT³ (−118.62, p < 0.10) collectively confirm an inverted-N-shaped relationship: the productivity effect of green trade is initially negative, transitions to positive at intermediate trade intensity levels, and then moderates at very high trade intensity. The two implied turning points—estimated at GT ≈ 6.8% and GT ≈ 19.2% of total merchandise exports—correspond to the theoretical stage transitions predicted by the unified causal framework, providing parametric corroboration of the threshold regression findings reported in Section 4.4.

4.3. Dynamic System-GMM Results

Table 5. Two-Step System-GMM Dynamic Panel Results: GT and GTFP (2003–2023).

Variable	(1) Sys-GMM	(2) Sys-GMM Lag-1	(3) Sys-GMM + Patents
GTFP (t−1)	0.638 *** (0.042)	0.624 *** (0.044)	0.601 *** (0.046)
Green trade (GT)	−0.713 ** (0.341)	—	−0.681 ** (0.336)
Green trade (GT t−1)		−0.924 *** (0.338)	—
Clean energy (CE)	0.004 ** (0.002)	0.005 ** (0.002)	0.004 ** (0.002)
Green patents (GP)	—	—	0.0008 ** (0.0004)
GDP growth	−0.004 * (0.002)	−0.004 * (0.002)	−0.003 (0.002)
Unemployment	0.014 *** (0.003)	0.014 *** (0.003)	0.013 *** (0.003)
FDI	0.002 * (0.001)	0.002 * (0.001)	0.002 * (0.001)
Year FE	Yes	Yes	Yes
Instrument count	29	29	29
AR(1) p-value	0.005	0.000	0.000
AR(2) p-value	0.253	0.318	0.402
Hansen J p-value	0.183	0.239	0.284
N	718	681	718

Notes: Two-step System-GMM with Windmeijer (2005) [32] finite-sample corrected SEs. Instruments: lags 2–4 of GT and CE, collapsed. *** p < 0.01, ** p < 0.05, * p < 0.10.

presents the System-GMM results across three specifications: the baseline dynamic model (Column 1), a one-year lagged green trade specification (Column 2), and an augmented model incorporating green patent intensity (Column 3). The System-GMM estimator addresses two limitations of the static FE results: the dynamic misspecification bias from omitting lagged GTFP, and the potential endogeneity of both GT and CE due to reverse causality.

The autoregressive coefficient on lagged GTFP is positive and highly significant across all three columns (δ ≈ 0.60–0.64, p < 0.01), confirming that GTFP exhibits strong persistence—approximately 60% of the productivity level in one year is directly transmitted to the next year through capital stock continuity, institutional path dependence, and technology learning. This finding validates the concern motivating the GMM approach: static FE models that omit this autoregressive term produce misspecified residuals and potentially biased trade coefficients.

After controlling for GTFP persistence and endogeneity, the negative effect of GT on GTFP remains statistically significant (Column 1: β₁ = −0.713, SE = 0.341, p < 0.05), though smaller in absolute magnitude than the static FE estimate (−0.921). This attenuation is consistent with two expected corrections: first, the GMM estimator removes the upward bias from reverse causality (productive economies with higher GTFP accumulate green trade advantages, inflating the static FE coefficient magnitude); second, by controlling for lagged GTFP, the GMM specification estimates the short-run effect of green trade on productivity net of persistence, while the static FE estimate conflates short-run and medium-run effects.

The post-estimation diagnostics confirm the validity of the System-GMM specification. The Hansen J-statistic (p = 0.183) does not reject the null of joint instrument validity, confirming that the collapsed lag-2 to lag-4 instrument set is uncorrelated with the error term. The AR(1) test (z = −2.81, p = 0.005) confirms the expected first-order autocorrelation in the differenced residuals—a mechanical property of differencing, not a violation. The AR(2) test (z = 1.14, p = 0.253) confirms the absence of second-order autocorrelation in the differenced residuals, validating the use of lag-2 instruments by ruling out serial correlation in the levels residuals that would invalidate the lag-2 instrument floor. The total instrument count is 29—below the cross-sectional dimension, satisfying rule of thumb against instrument proliferation. All these diagnostics are reported directly in the main table rather than footnotes, following reviewer recommendation.

Column 3 introduces green patent intensity (GP) as an additional control, capturing the innovation ecosystem dimension of the theoretical Pathway 3 (innovation complementarity) developed in Section 2.4. The positive and significant coefficient on GP (β = 0.0008, p < 0.05) indicates that a one unit increase in environmental patent applications per million population is associated with a 0.0008 index point improvement in GTFP, independently of the clean energy threshold mechanism. The introduction of GP does not materially alter the magnitude or significance of the GT coefficient (−0.681, p < 0.05), confirming that the green trade suppression effect is not explained away by innovation ecosystem differences—the two channels operate in parallel rather than being collinear.

The directional comparison between the static FE coefficient (|β₁| = 0.921) and the System-GMM coefficient (|β₁| = 0.713) is consistent with the expected upward bias from reverse causality in the static estimator. This comparison provides suggestive but not conclusive evidence of reverse causality: economies with higher GTFP may develop comparative advantages in environmental goods production, feeding back into higher GT ratios and inflating the static FE estimate magnitude. The System-GMM specification reduces this bias through the instrument strategy, but the observational nature of the analysis precludes complete causal identification. Researchers seeking to establish definitive causal estimates should employ quasi-experimental variation, such as country-year-specific CLEG tariff liberalization events or renewable energy auction assignments, as instruments for green trade flows.

4.4. Panel Threshold Regression Results

4.4.1. Clean Energy Threshold (Single and Double)

Table 6. Panel Threshold Test Statistics and Likelihood Ratio Confidence Intervals.

Threshold Variable	Estimate ($\stackrel{\mathit{ˆ}}{\mathit{\gamma }}$)	Test Type	F-Statistic	Bootstrap p	95% LR-CI	90% LR-CI
Clean energy (${\widehat{\gamma }}_1$)	8.72%	Single	83.41	0.000 ***	[7.31%, 10.15%]	[7.68%, 9.84%]
Clean energy (${\widehat{\gamma }}_2$, 2nd)	24.63%	Double	11.38	0.014 **	[21.87%, 27.44%]	[22.53%, 26.71%]
Clean energy (triple)		Triple	4.12	0.218	Not confirmed	—
R&D intensity (${\widehat{\gamma }}_3$)	0.66% of GDP	Single	58.17	0.000 ***	[0.51%, 0.79%]	[0.54%, 0.76%]

Notes: 500 bootstrap iterations; LR-CI = likelihood ratio inversion method (Hansen, 1999). *** p < 0.01, ** p < 0.05.

presents the panel threshold regression results with clean energy share (CE) as the threshold variable. The sequential bootstrap tests confirm a statistically significant double threshold structure, yielding three distinct productivity regimes.

The single threshold test strongly rejects the null of linearity (F = 83.41, bootstrap p = 0.000), with a threshold estimate of ${\widehat{\gamma }}_1=8.72\mathrm{\%}$. Conditional on this first threshold, the double threshold test also rejects the null of a single threshold (F = 11.38, bootstrap p = 0.014), with a second threshold estimate of ${\widehat{\gamma }}_2=24.63\mathrm{\%}$. The triple threshold test fails to reject the null of a double threshold (F = 4.12, bootstrap p = 0.218), confirming that the double threshold model is the appropriate specification. The three-regime model decomposition yields:

• Regime 1 (CE < 8.72%): β₁ = −0.061 (SE = 0.019, p < 0.01)—green trade suppresses GTFP
• Regime 2 (8.72% ≤ CE < 24.63%): β₂ = +0.049 (SE = 0.016, p < 0.01)—green trade enhances GTFP
• Regime 3 (CE ≥ 24.63%): β₃ = +0.028 (SE = 0.013, p < 0.05)—green trade moderately enhances GTFP

The first threshold estimates of 8.72% is marginally higher than the 8.66% reported in prior OECD analyses covering 2003–2016—a difference of 0.06 percentage points. The 95% LR confidence interval for the current estimate ([7.31%, 10.15%]) fully contains the prior estimate of 8.66%, confirming that the two estimates are statistically consistent and that the near-equivalence reflects genuine parameter stability rather than sampling variation. The economic significance of this near-stability is interpretively important: despite the dramatic post-Paris decline in renewable technology costs—solar PV module prices fell approximately 90% between 2010 and 2022—the minimum absorptive capacity threshold for productive green trade integration has not fallen. This finding is consistent with Aghion et al.’s (2016) directed technical change framework, which predicts that rising frontier technology complexity increases rather than decreases the minimum absorptive capacity requirement [1]. The latest generation of green goods commanding premium export status—advanced offshore wind systems, grid-scale battery storage, smart EV charging infrastructure, industrial electrolysis for green hydrogen—impose higher energy system compatibility requirements than the simpler rooftop solar and onshore wind technologies that dominated CLEG-classified trade in the pre-2016 period. The stability of the threshold across periods, despite the profound changes in the technology composition of green trade, constitutes a substantively important and policy-relevant finding in its own right.

The second threshold at 24.63% captures a diminishing-returns transition: above this level, the marginal productivity gain from an additional unit of green trade is positive but smaller than in Regime 2 (β₃ = +0.028 vs. β₂ = +0.049). One theoretically consistent account of this attenuation, compatible with the innovation complementarity pathway (Pathway 3, Section 2.4), is that at very high renewable penetration levels (above ~24.63%), advanced-stage OECD transitioners approach the global green technology frontier and face diminishing marginal returns to further green technology imports—additional green goods imports provide smaller incremental productivity gains because the most complementary technologies have already been absorbed. However, this interpretation remains a hypothesis consistent with the coefficient pattern rather than a causally identified finding; the attenuation in Regime 3 could alternatively reflect composition effects as advanced transitioners shift toward more complex and less tradeable frontier technologies. As of 2023, the Regime 3 sub-group includes Norway (CE ≈ 71%), Iceland (CE ≈ 76%), Sweden (CE ≈ 58%), and Denmark (CE ≈ 43%)—economies that are simultaneously technology frontier producers and domestic deployers of green goods, complicating the absorption interpretation.

4.4.2. R&D Intensity Threshold

The R&D intensity threshold is statistically confirmed at the 1% level. Below this threshold, the coefficient on GT is negative and significant (β = −0.038, p < 0.05), consistent with the absorptive capacity failure mechanism identified in Pathway 2 of Section 2.4: economies with insufficient innovation investment cannot translate imported green technologies into productivity gains, and the competitive pressure from frontier green exporters may actually suppress domestic GTFP by displacing less-productive green industry incumbents. Above the R&D threshold, the coefficient on GT becomes positive and significant (β = +0.041, p < 0.01), that above-threshold innovation ecosystems generate increasing returns to green technology absorption. Of the 37 OECD economies, 11 fall below the 0.66% R&D threshold in 2023—concentrated among the Central and Eastern European members (Slovakia, Hungary, Latvia, Poland) and non-European members with historically lower R&D intensities (Chile, Colombia, Costa Rica, Mexico, Turkey)—identifying the sub-population most at risk of below-threshold green trade productivity penalties.

4.4.3. Interaction of Clean Energy and R&D Thresholds

To test whether the clean energy and R&D thresholds operate independently or interact, we specify an augmented threshold model that conditions simultaneously on both threshold variables. The joint threshold test confirms that the clean energy threshold retains its significance conditional on R&D threshold status (F = 74.18, bootstrap p = 0.000), and the R&D threshold retains its significance conditional on clean energy threshold status (F = 41.33, bootstrap p = 0.002). This independence of threshold effects implies that policymakers cannot substitute energy transition investment for R&D investment, or vice versa, in order to unlock the above-threshold GTFP gains from green trade—both channels must be simultaneously addressed. Economies below both thresholds (low CE and low R&D) face the largest below-threshold productivity penalties: for this sub-group, the coefficient on GT is −0.089 (p < 0.01), nearly 1.5 times the magnitude of the full-sample below-threshold penalty. This finding has direct relevance for the design of OECD development cooperation and EU neighborhood policy directed at Central and Eastern European member states.

4.5. Heterogeneity Analysis by Energy Transition Stage

Table 7. Heterogeneity Analysis by Energy Transition Stage (2016 CE Cutoff).

	Early-Stage Transitioners (N = 19)	Advanced-Stage Transitioners (N = 18)	Chow Test
FE: GT coefficient	−1.847 *** (0.521)	−0.287 (0.218)	F = 7.83 **
Threshold estimate ($\widehat{\gamma }$)	7.83%	26.41%	—
Below-threshold β	−0.074 *** (0.021)	—	—
Above-threshold β	+0.031 ** (0.014)	+0.067 *** (0.019)	—
GMM: GT coefficient	−1.214 *** (0.398)	−0.241 (0.187)	—
GTFP persistence (δ)	0.581 *** (0.051)	0.671 *** (0.048)	—
Country/Year FE	Yes	Yes	—
Observations	399	342	—

Notes: *** p < 0.01, ** p < 0.05, Sub-group classification based on 2016 CE share relative to OECD median (~18%). Chow test for equality of GT coefficient across sub-groups.

presents the sub-group threshold regression and fixed-effects results by energy transition stage, defined by the 2016 renewable energy share cutoff (Section 3.3.4). The results reveal pronounced asymmetry in the productivity consequences of green trade expansion across the two sub-groups.

For early-stage transitioners (below-median CE share in 2016; N = 19 economies): the baseline FE coefficient on GT is −1.847 (p < 0.01), substantially larger in absolute magnitude than the full-sample coefficient (−0.921), confirming that the aggregate negative effect is disproportionately driven by economies that entered the sample period with underdeveloped renewable energy infrastructure. The single threshold for this sub-group is estimated at 7.83%—lower than the full-sample threshold of 8.72%—implying that early-stage transitioners face a slightly lower threshold barrier, potentially reflecting a lower floor of absorptive capacity requirements for the simpler green technologies (efficiency improvements, waste management systems) that dominate their green trade composition. Below this sub-group threshold, β₁ = −0.074 (p < 0.01); above it, β₂ = +0.031 (p < 0.05). The smaller above-threshold positive effect for early-stage transitioners (compared to advanced transitioners) is consistent with the lower innovation complementarity available in less developed renewable ecosystems.

For advanced-stage transitioners (above-median CE share in 2016; N = 18 economies): the baseline FE coefficient on GT is −0.287 (p = 0.18, not significant), confirming that once clean energy penetration is sufficiently advanced at the group level, the average green trade effect is statistically indistinguishable from zero. The threshold structure for this sub-group shows a confirmed single threshold at 26.41%—corresponding to the Regime 2/3 transition in the full-sample model—above which β₂ = +0.067 (p < 0.01), a positive effect nearly 2.2 times larger than the Regime 2 coefficient for early-stage transitioners (0.031). This finding confirms the theoretical prediction of Pathway 3 (innovation complementarity): advanced transitioners not only escape the below-threshold productivity penalty but capture meaningfully larger productivity gains from green trade once their innovation and energy systems are co-evolved.

The sub-group asymmetry provides evidence consistent with the hypothesis that energy transition stage is a binding conditioning variable for green trade productivity outcomes. This pattern should, however, be interpreted with appropriate caution: the sub-group partition is based on a cross-sectional split rather than random assignment, and unobserved institutional or structural differences between early-stage and advanced-stage transitioners may partially account for the observed coefficient differences. Institutional quality (e.g., regulatory efficiency, government effectiveness) correlates with both renewable energy share and GTFP, and while country fixed effects absorb time-invariant institutional factors, time-varying institutional improvements may confound the sub-group comparison. The Chow test for coefficient equality between the two sub-groups rejects the null at the 5% level (F = 7.83, p = 0.006), providing statistical support for the heterogeneity claim while acknowledging the observational limitations of the sub-group design.

4.6. Robustness Checks

Table 8. Robustness Summary: First Threshold Estimate and GT Regime Coefficients Across Eight Specifications.

Check	Specification	${\stackrel{\mathit{ˆ}}{\mathit{\gamma }}}_1$	Below-Threshold β	Above-Threshold β	Double Threshold ${\stackrel{\mathit{ˆ}}{\mathit{\gamma }}}_2$
Baseline	Full sample, FE + Threshold	8.72%	−0.061 ***	+0.049 ***	24.63% **
1	Exclude COVID-19 (2020–2022)	8.49%	−0.058 ***	+0.047 ***	23.51% **
2	Exclude GFC (2008–2009)	9.14%	−0.063 ***	+0.051 ***	25.18% **
3	Green openness index (GT alt.)	8.41%	−0.054 **	+0.043 ***	24.07% **
4	GTFP: CO2 only (no PM2.5)	8.19%	−0.047 **	+0.038 ***	22.84% *
5	PPML estimation	7.98%	−0.052 **	+0.041 ***	23.91% **
6	Driscoll-Kraay SE	8.72%	−0.061 ***	+0.049 ***	24.63% **
7	Balanced panel (N = 35, T = 21)	8.89%	−0.065 ***	+0.052 ***	26.80% **
8	Alt. threshold: elec. gen. share	14.32% †	−0.059 ***	+0.046 ***	38.17% †*

Notes: † Estimates on the electricity generation share scale (not comparable to CE total final energy consumption scale). *** p < 0.01, ** p < 0.05, * p < 0.10.

summarizes the key coefficient estimates and threshold values across all eight robustness specifications described in Section 3.4. The results confirm the stability of the main findings across all sensitivity checks.

The first threshold estimate ranges from 7.98% to 9.14% across CE-based specifications—a bandwidth of just 1.16 percentage points—confirming that the threshold identification is not sensitive to any single methodological choice. The second threshold similarly ranges from 23.51% to 26.80%, confirming the double-threshold structure as a robust empirical property of the 2003–2023 OECD panel. The regime-sign pattern (negative below threshold → positive in Regime 2 → positive but smaller in Regime 3) is preserved across all eight checks without exception.

4.7. Synthesis and Implications for the Theoretical Framework

The empirical results, taken together, provide evidence consistent with all three causal pathways specified in the theoretical framework of Section 2.4. Pathway 1 (Pollution Displacement) predicts that below-threshold economies embed carbon-intensive upstream production into their GTFP frontier; the large negative and significant Regime 1 coefficient (β₁ = −0.061) is consistent with this prediction, and the finding that the penalty is larger for early-stage transitioners (β = −0.074)—who rely more heavily on fossil fuels for manufacturing energy—is consistent with a pollution displacement mechanism. Pathway 2 (Absorptive Capacity Failure) predicts that below-threshold imported clean technologies underperform in fossil-fuel-grid-dependent systems; the R&D threshold result—where the GT effect is most negative for economies below both the CE and R&D thresholds simultaneously (β = −0.089)—provides additional evidence consistent with this pathway, since low R&D intensity is a direct proxy for low absorptive capacity. Pathway 3 (Innovation Complementarity) predicts that above-threshold economies with denser innovation ecosystems generate larger GTFP gains from green trade; the positive and significant coefficient on green patent intensity in the GMM Column 3 model (β = +0.0008, p < 0.05), and the substantially larger above-threshold gains for advanced-stage transitioners (β₂ = +0.067 vs. +0.031), are consistent with this complementarity mechanism.

It is important to note that each of these observations provides evidence consistent with the three pathways, not definitive causal proof of each mechanism separately. The degree of overlap between the CE threshold, R&D threshold, and the patent-productivity correlation means that disentangling the independent contribution of each channel requires micro-level data on firm-level green technology adoption, energy mix, and productivity—a priority direction for future research. The macroeconomic panel evidence presented here establishes the existence and stability of the regime-switching pattern and confirms that all three theorized channels are directionally operative, while appropriately acknowledging the limitations of aggregate observational inference.

5. Conclusions and Policy Implications

This study examined whether renewable energy adoption governs the relationship between GTE and GTFP across 37 OECD economies over 2003–2023. Using two-way fixed-effects panel regression, dynamic System-GMM, and Hansen’s panel threshold regression with 500 bootstrap iterations, four findings emerge.

The baseline GT–GTFP relationship is negative and significant in the full pooled sample, with the coefficient attenuation under GMM consistent with reverse causality bias in the static estimator. Sequential threshold testing confirms a double threshold structure: a first clean energy threshold at 8.72% of total final energy consumption and a second at 24.63%, delineating three productivity regimes—suppression below 8.72%, enhancement between thresholds, and diminishing-returns enhancement above 24.63%. The first threshold is near-identical to the 8.66% reported in prior 2003–2016 OECD analyses—confirming stability despite the post-Paris transformation of green technology composition—consistent with directed technical change prediction that rising frontier technology complexity maintains minimum absorptive capacity requirements. Heterogeneity analysis confirms early-stage transitioners bear disproportionately larger below-threshold penalties while advanced transitioners capture substantially larger above-threshold gains.

Policy implications are presented as evidence-consistent hypotheses, not causal prescriptions. Three directions are supported by the evidence. First, green trade liberalization and renewable energy investment are strategic complements: OECD economies at or below the 8.72% CE threshold—including several Central and Eastern European members—should prioritize clean energy infrastructure investment as a prerequisite for productive green trade participation. The EU CBAM’s design is structurally consistent with this threshold logic. Second, green finance mechanisms—green bonds, transition finance instruments, and climate-linked investment mandates—are among the most effective instruments for accelerating threshold-crossing, and should be explicitly coordinated with green trade policy rather than treated as separate instruments. Third, the independent R&D threshold (0.66% of GDP) confirms that both energy and innovation channels must be simultaneously activated; neither substitutes for the other.

Key limitations include observational identification constraints, OECD-only scope that excludes major green manufacturing hubs such as China—whose producer-side threshold mechanisms differ fundamentally from the import-side absorptive capacity framework applied here—and sectoral aggregation that masks product-level heterogeneity. Future research should employ quasi-experimental designs exploiting CBAM implementation schedules and renewable energy auction assignments to establish causal estimates of threshold-crossing effects on GTFP.