The impact of climate risk on forestry ecological total factor productivity: evidence from double machine learning

1 Introduction

Climate change is reshaping natural ecosystems and economic production systems worldwide at an unprecedented pace. According to the World Meteorological Organization, 11,778 weather-, climate- and water-related disasters were recorded globally between 1970 and 2021, causing more than 2 million deaths and cumulative economic losses of US$4.3 trillion (World Meteorological Organization, 2023). The increasing frequency and intensity of extreme heat, heavy rainfall and persistent drought expose production systems to simultaneous shocks from physical damage, operational uncertainty and distorted resource allocation (Wang et al., 2018). Identifying the effects of climate risk on total factor productivity has therefore become a central concern in climate economics and green-development research.

Climate-related risk is not a unitary concept. The standard classification distinguishes physical risk from transition risk. Physical risk refers to direct risks arising from climate events and changing climatic conditions and can be divided into event-driven acute physical risk and chronic physical risk caused by long-term shifts in climate patterns. Transition risk arises mainly from changes in policies and regulations, production technologies, energy systems and market preferences during the transition to a low-carbon economy (IFRS Foundation, 2023). This study focuses exclusively on physical climate risk, measured using the CPRI constructed from the numbers of extreme-cold, extreme-heat, extreme-rainfall and extreme-drought days. The index primarily captures exposure to event-driven climate extremes; the drought measure also reflects the persistence and accumulation of climate shocks within a year. It does not directly measure chronic physical risks such as changes in long-term mean temperature or precipitation patterns, sea-level rise or shifts in forest climate zones, nor does it include transition risks induced by carbon policy, low-carbon technologies or market transformation. Unless otherwise stated, “climate risk” hereafter therefore refers to the composite physical climate risk represented by the CPRI and dominated by exposure to acute extreme events.

Forestry depends more strongly on climatic conditions and is more directly exposed to climate risk than conventional industrial sectors. Tree assets are fixed in particular locations, and production cycles often span years or decades, making it difficult for operators to avoid extreme-weather shocks by relocating capacity or rapidly substituting inputs. Forestry also supports carbon sequestration and sink enhancement, water conservation, biodiversity protection and the provision of ecosystem products. Changes in its production efficiency therefore affect not only industrial output but also ecological security and the effectiveness of climate governance. Ecological total factor productivity in forestry is consequently an important measure of resource-allocation efficiency, technological progress and the quality of green development in forestry. A systematic assessment of how climate risk affects forestry TFP can reveal how climate change alters efficiency in natural-resource sectors and provide empirical evidence for strengthening forestry climate adaptation, safeguarding regional ecological security and supporting China’s dual-carbon goals.

Existing studies provide a foundation for understanding the relationship between climate risk and production efficiency, but important gaps remain. Macroeconomic evidence shows that climate risk significantly affects cross-country TFP and that institutional quality moderates this relationship (Song et al., 2023). Firm-level studies suggest that climate-policy uncertainty can reduce corporate TFP by discouraging research and development, compressing free cash flow or weakening supply-chain resilience (Ren et al., 2022; Zhang et al., 2025), but can also generate positive effects through green finance and green transformation (Sun and Liu, 2025). Studies in agriculture and ecology further show that extreme weather affects agricultural green TFP and the resilience of forest productivity (Song et al., 2022; Lyu et al., 2025; Zheng et al., 2026). However, research has largely focused on the macroeconomy, manufacturing, carbon-intensive industries or agriculture, with limited sector-specific empirical evidence for climate-sensitive forestry. Evidence is also limited on whether forest-disaster management and the allocation of public attention to environmental issues transmit the effects of climate risk to forestry TFP. Moreover, the mixed direction of previous findings and the contextual and nonlinear factors underlying this divergence remain insufficiently explained.

China provides a representative setting in which to examine these questions. Its vast territory and diverse climatic zones produce substantial interprovincial variation in climate-risk exposure, enabling spatial identification of heterogeneous climate shocks. China has also made the development of forestry carbon sinks an important component of its dual-carbon goals and ecological-civilization program. High-quality forestry development is therefore central to increasing ecosystem carbon-sink capacity, stabilizing the supply of ecosystem products and safeguarding regional ecological security. Examining the effects of climate risk on forestry TFP using provincial panel data from China can thus extend sectoral evidence on the economic consequences of climate risk and inform forestry climate-adaptation policy.

To address these theoretical and practical questions, we use panel data for 30 Chinese provinces from 2012 to 2023, measure physical climate risk dominated by exposure to acute extreme events with the composite physical climate risk index (CPRI) (Wang et al., 2018), and apply a double machine learning framework to examine its effects on ecological total factor productivity in forestry, transmission pathways and boundary conditions (Chernozhukov et al., 2018). Building on a systematic review of the relevant literature, this study makes three main contributions.

First, this study makes a theoretical contribution. Previous research on climate risk and green total factor productivity has concentrated on the macroeconomy, manufacturing firms or agricultural production, typically explaining productivity changes through research and development, corporate cash flow, institutional quality, green finance or digital transformation. These findings do not directly explain how climate shocks affect forestry production systems, in which natural assets are immobile, production cycles are long and both economic and ecological outputs are generated. We conceptualize forestry as a socioecological production system with high physical exposure and climate sensitivity and develop an integrated framework linking climate shocks, resource allocation and production governance, adaptive capacity, and forestry TFP. The framework brings direct effects, transmission through production and social governance, structural exposure, external network connections and regional adaptive capacity into a common analytical structure. It therefore tests not only whether climate risk affects forestry TFP, but also how climate shocks enter forestry production systems and why their effects vary across regional contexts.

Second, this study makes an empirical contribution. Forestry research has extensively examined forest carbon-sink efficiency, ecosystem productivity, the economic value of land and the spatiotemporal evolution of productivity, but has provided relatively limited identification of the relationship between composite physical climate risk and forestry ecological TFP. We combine the CPRI, which captures exposure to extreme cold, extreme heat, extreme rainfall and extreme drought, with a measure of forestry ecological TFP that incorporates labor, capital, land and energy inputs as well as desirable and undesirable outputs, thereby providing provincial-level evidence from China. We further identify two internal transmission pathways, the control gap for forest diseases, insect pests and rodents and the allocation of public attention to pollution issues, and test boundary conditions arising from structural exposure, trade-network connections and regional adaptive capacity. Beyond documenting the average effect of climate risk, our results show that forestry productivity losses arise not only from direct damage to trees and infrastructure, but also from gaps in disaster-management coverage, the allocation of social-governance resources and regional capacity to absorb shocks.

Third, the study has policy value. Rather than offering a generic call to reduce climate risk, it identifies specific points of intervention for forestry climate-adaptation policy. The transmission effect through the control gap for forest diseases, insect pests and rodents indicates a need to strengthen disaster monitoring, control coverage and emergency resource mobilization under extreme climatic conditions. The results for the allocation of public attention to pollution issues indicate that continuous and accurate environmental communication must be maintained during disasters to prevent sustained crowding-out of general pollution-control issues. Differences in structural exposure, external network connections and adaptive capacity further indicate that forestry climate-adaptation policy should be tailored to regional industrial and carbon-emission profiles, embeddedness in trade networks, and the conditions for forestry stewardship and ecological governance. These recommendations correspond directly to our empirical results and do not extend to untested arrangements involving specific insurance contracts or financial instruments.

2 Literature review

2.1 Climate risk and total factor productivity

The relationship between climatic conditions and economic productivity is a central topic in climate economics. Early cross-country studies primarily used annual temperature variation to identify macroeconomic effects. Dell et al. (2012) found that heat shocks significantly reduce economic growth in low-income countries and may simultaneously affect agriculture, industry and social stability. Burke et al. (2015) further identified a global nonlinear relationship between temperature and economic output. These studies established the empirical basis for linking climate shocks to productivity, but their focus on macroeconomic aggregates provides limited insight into production processes in specific natural-resource sectors.

Subsequent research shifted toward industries and firms. Using data on Chinese manufacturing firms, Zhang et al. (2018) found that high temperatures reduce firm-level total factor productivity and induce factor reallocation. From a transition-risk perspective, Ren et al. (2022) showed that climate-policy uncertainty affects corporate TFP through research and development and cash-flow channels. Related studies report that environmental-policy uncertainty and climate risk influence corporate green investment, rural land-factor allocation and energy efficiency under transition policies (Hu et al., 2023; Wan et al., 2025; Wei and Liu, 2025; Li et al., 2026). In agriculture, Song et al. (2022) found that changes in temperature, precipitation and humidity affect China’s agricultural green total factor productivity through output, input and structural channels, while Zhou et al. (2024) documented pronounced regional variation in the agricultural TFP effects of climate change. Studies using composite climate-risk measures likewise identify significant links between climate risk and TFP or green TFP, with institutional quality, greening and digitalization potentially altering the magnitude of these effects (Song et al., 2023; Sun and Liu, 2025). Overall, substantial evidence now spans the macroeconomic, corporate and agricultural levels, but forestry, with its long production cycles and strong dependence on nature, remains understudied.

2.2 Forestry productivity and its climatic determinants

Research on forestry productivity has followed three main strands: efficiency measurement, carbon-sink outputs and regional variation. Lin and Ge (2019) incorporated forest carbon sinks and forestry output value into a productivity-evaluation system and used the SBM and Malmquist–Luenberger indices to assess forestry eco-economic efficiency in China, showing that both economic and environmental conditions affect forestry productivity. Lyu et al. (2025) developed an indicator framework based on the resilience of forest-ecosystem productivity and transition risk; Zheng et al. (2026) examined how climate change affects the productivity gains from greening through ecological projects; and Ribbers et al. (2026) argued that existing management policies for boreal forests insufficiently account for climate risk. These studies broaden the ecological dimensions of forestry-output evaluation and forest management, but focus mainly on productivity measurement, resilience assessment, spatiotemporal evolution or management policy rather than systematically examining the effects of composite physical climate risk on forestry productivity.

Research in ecology and forest economics further shows that climate change alters both ecological outputs and the economic value of forests. Hanewinkel et al. (2013) found that shifts in tree-species suitability caused by changes in temperature and precipitation could substantially reduce the economic value of forest land. A global synthesis by Seidl et al. (2017) showed that warming, drought, wind disturbance, insects and pathogens may have mutually reinforcing effects. These studies demonstrate the high climatic sensitivity of forest production. However, forest net primary productivity, land value and carbon-sink efficiency are not equivalent to ecological total factor productivity in forestry, which jointly accounts for labor, capital, land and energy inputs and for desirable and undesirable outputs. The effect of composite physical climate risk on forestry TFP therefore warrants examination from the perspectives of the production frontier and factor allocation.

2.3 Mechanisms linking climate risk to forestry productivity

Existing studies generally explain the productivity effects of climate risk through direct damage, factor allocation and adaptive capacity. Extreme temperatures, drought and heavy rainfall directly damage trees, roads and protective infrastructure and alter the overwintering, reproduction and dispersal conditions of harmful organisms. Seidl et al. (2017) reported that warmer and drier conditions intensify fire, drought and insect disturbances, whereas warmer and wetter conditions may increase wind and pathogen disturbances. Adaptation research also indicates that climate extremes, financial vulnerability and adaptation gaps may reinforce one another (Touch et al., 2025). When additional disaster pressure exceeds existing monitoring and control capacity, insufficient management coverage may not only damage trees and ecological functions but also force operators to redirect capital and labor from long-term cultivation and technological upgrading to emergency response. Although previous research establishes links between climate and forest disturbance, few studies treat the control-coverage gap as a testable mechanism connecting climate risk to forestry TFP.

Public attention represents another potential social-governance channel. Zhou and Ding (2023) found that online environmental attention can promote corporate environmental behavior through responses from consumers, investors and governments, indicating that public attention can generate informal pressure for environmental governance. Climate-adaptation policy, climate-risk communication and climate-information services may also improve adaptive behavior and production resilience (Gao et al., 2026; Jahan et al., 2026; Tamru et al., 2025). However, little research has examined whether information about emergency warnings, casualties and property losses during extreme climate events crowds out sustained public attention to general pollution issues, or whether such reallocation of attention affects forestry production efficiency. The control gap for forest diseases, insect pests and rodents and the allocation of public attention to pollution issues therefore represent production-governance and social-governance mechanisms, respectively, complementing explanations based solely on direct physical damage or corporate financial channels.

In summary, three substantive gaps remain. First, research on climate risk and productivity has focused primarily on the macroeconomy, manufacturing, and agriculture, leaving limited systematic evidence for forestry, a sector characterized by long production cycles, strong dependence on natural capital, and the joint production of economic and ecological outputs. Second, forestry-productivity research has concentrated largely on carbon-sink efficiency, ecological productivity, and their spatial and temporal evolution, while the effect of composite physical climate risk on forestry ecological TFP remains insufficiently examined. Third, although previous studies have discussed the direct damage caused by climatic shocks and the role of adaptive capacity, limited research has jointly examined transmission through forest-disaster management gaps and the allocation of public attention, together with boundary conditions associated with structural exposure and regional adaptive capacity. To address these substantive questions, we use panel data for 30 Chinese provinces from 2012 to 2023 and combine the composite physical climate risk index with forestry ecological TFP to examine the direct effect, transmission mechanisms, and boundary conditions of climate risk. The contribution of the study therefore lies primarily in providing sector-specific evidence for forestry and in clarifying the relevant mechanisms and regional differences, rather than merely in applying a particular econometric method.

3 Theoretical framework and hypotheses

3.1 An integrated framework of climate shocks, resource allocation, and adaptive capacity

The IPCC climate-risk framework conceptualizes climate risk not as a consequence of extreme climate events alone, but as the dynamic interaction among climatic hazards, the exposure of economic and ecological systems, and systemic vulnerability. Vulnerability encompasses both the sensitivity of a production system to shocks and its capacity to respond to, absorb and adapt to them. Forestry is both an economic production system and a natural ecosystem. Tree assets are spatially fixed, production cycles are long and output depends strongly on temperature and water availability, resulting in high physical exposure and climate sensitivity. At the same time, pest control, access to information, environmental governance and resource-mobilization capacity determine whether climate shocks can be absorbed and addressed promptly (Intergovernmental Panel on Climate Change, 2022).

Within a production-theory framework, ecological total factor productivity in forestry captures the overall efficiency with which inputs such as labor, capital, land and energy are transformed into economic and ecological outputs. Climate risk may alter this transformation in three ways. First, it damages trees, roads and water infrastructure, reducing attainable output for a given set of inputs. Second, it increases expenditure on monitoring, disaster prevention and recovery, redirecting capital, labor and equipment from long-term cultivation and technological upgrading to emergency response. Third, it increases operational uncertainty, affecting technological investment, factor mobility and production organization. Climate shocks may therefore shift the production frontier inwards and reduce technical and allocative efficiency relative to a given frontier.

Adaptive capacity determines whether these shocks can be absorbed. Resilience theory defines adaptive capacity as the ability of a socioecological system to maintain its core functions, reorganize resources and learn and adjust following external disturbance (Nelson et al., 2007). Accordingly, we divide the process through which climate risk affects forestry TFP into three interconnected levels. The first is the direct effect of climate shocks on the production frontier and allocative efficiency. The second is their internal transmission through production- and social-governance systems, represented, respectively, by the control gap for forest diseases, insect pests and rodents and the allocation of public attention to pollution issues. The third comprises boundary conditions that determine the strength of the effect. Capital-market exposure to energy-intensive industries and regional carbon-emission intensity represent the structural exposure of provincial economies; trade openness represents connections to external production networks and markets; and forestry stewardship, the digital economy and environmental-protection intensity capture different dimensions of shock detection, resource mobilization and ecological-governance capacity.

The hypotheses that follow therefore do not derive from independent theories. H1 tests the overall direct effect of climate shocks; H2a and H2b test the two internal transmission pathways through production and social governance; H3a and H3b test how regional structural exposure moderates the marginal effect of climate risk; H4 tests nonlinear regime shifts arising from external network connections; and H5a–H5c test differences in effects under varying adaptive capacities. Together, these hypotheses address a unified question: how climate shocks enter forestry production systems, how they are transmitted through resource-allocation and governance processes, and why their effects differ across regions.

3.2 Direct effect of climate risk on forestry TFP

Total factor productivity captures the overall efficiency with which a production system transforms resources into output for a given set of factor inputs. Changes in TFP arise from technological progress, improvements in technical efficiency and better resource allocation (Ahmad et al., 2026; Song et al., 2023). As an external physical shock, climate risk affects the determinants of TFP through multiple channels by damaging production factors, disrupting technological accumulation and distorting expectations about resource allocation.

Forestry production depends strongly on natural conditions, including land, sunlight and water. Tree assets are spatially fixed and production cycles commonly span years or decades, making it difficult to avoid external risks by relocating capacity or substituting inputs as industrial sectors can. Extreme heat, persistent drought and heavy rainfall directly damage tree stocks and forestry infrastructure such as roads and water facilities, increasing the inputs required to maintain a given level of output and thereby reducing total factor productivity (Zheng et al., 2026; Lyu et al., 2025). Beyond direct physical damage, climate risk can have persistent effects on forestry TFP by distorting expectations. Greater climate risk increases operational uncertainty, prompting forestry operators to reduce long-term investment in technological improvement, equipment renewal and large-scale operations and instead retain liquid funds as a precaution. This reallocation from productive investment to defensive reserves persistently constrains the technological accumulation that drives TFP growth (Ren et al., 2022; Hao and Yu, 2026). In regions with persistently high climate risk, lower expected returns to forestry also encourage labor and capital outflows, further reducing allocative efficiency (Song et al., 2023). Together, these three pathways imply a negative net effect of climate risk on forestry TFP.

H1: Composite physical climate risk dominated by exposure to acute extreme events significantly suppresses ecological total factor productivity in forestry.

3.3 Transmission through forestry pest and rodent management

Within the integrated framework, the control gap for forest diseases, insect pests and rodents is an internal transmission channel arising from insufficient production-governance capacity. Climate risk not only directly damages trees and infrastructure but also alters the occurrence, spread and control of forest pests and diseases. Abnormally high temperatures, persistent drought and heavy rainfall may weaken tree physiology, alter the overwintering and reproductive environments of harmful organisms, and impede monitoring, field operations and coordinated interregional control. When existing control capacity cannot keep pace with additional risk, the overall control rate for forest diseases, insect pests and rodents declines, expanding the share of risk that remains ineffectively managed.

To correspond directly to the forestry production process, we define the control gap for forest diseases, insect pests and rodents (M1) as 1 minus the overall control rate for forest diseases, insect pests and rodents. A higher value indicates that a larger share is not covered by the existing control system. A wider management gap increases tree mortality, timber-volume losses and damage to ecological functions. It also forces operators to redirect capital, labor and equipment from long-term cultivation and technological upgrading to emergency response and post-disaster recovery, thereby reducing both output performance and allocative efficiency. The conceptual object, sectoral boundary and causal chain of this variable thus all relate directly to forestry TFP.

H2a: Climate risk has a negative indirect effect on forestry TFP by widening the control gap for forest diseases, insect pests and rodents.

3.4 Transmission through the allocation of public attention to pollution issues

The allocation of public attention to pollution issues is an internal transmission channel arising from the reallocation of social-governance resources. Public attention is finite. Following an extreme climate event, demand for information may concentrate on weather warnings, casualties, property losses and emergency response, crowding out sustained attention to general pollution-control issues. Public attention to pollution may promote investment in pollution control, implementation of ecological protection and adoption of green production practices through information diffusion, social oversight and local environmental-governance responses, thereby benefiting forestry TFP. A climate-risk-induced decline in attention to general pollution issues may therefore create an additional negative pathway.

To reduce sensitivity to idiosyncratic variation in a single keyword, we replace the original single-keyword measure based on “environmental pollution” with the composite index of public attention to pollution issues (M2). Specifically, we add 1 to the annual provincial Baidu search indices for “environmental pollution” and “haze,” take their natural logarithms, standardize each keyword within the full sample, and then calculate their equally weighted mean. This index captures online public search attention to general pollution issues and is not equivalent to specialized attention to climate risk, forest protection or forestry ecological governance.

We therefore interpret M2 as a proxy for the allocation of public attention to general pollution issues, rather than as a comprehensive measure of public environmental awareness. The theory does not presume that climate risk necessarily increases all forms of environmental attention; instead, it tests whether sudden risks crowd out attention to other issues. Given the restricted set of keywords, this pathway provides exploratory evidence. Future studies could incorporate continuously available multisource data on “extreme weather,” “forest protection,” “forest pests,” “forestry carbon sinks,” and related terms.

H2b: Climate risk has a negative indirect effect on forestry TFP by reducing public search attention to general pollution issues.

3.5 Moderating role of structural exposure

Within the integrated framework, both capital-market exposure to energy-intensive industries and regional carbon-emission intensity are defined as structural exposure of the regional economic system, rather than as general industrial structure or financial-development capacity. Capital-market exposure to energy-intensive industries is measured as the market capitalization of listed firms in the six major energy-intensive industries relative to the total market capitalization of listed firms in the province. A higher value indicates that the value of listed companies, investment expectations and related economic activity are more closely tied to energy-intensive industries. As climate risk rises, fluctuations in these industries’ operations, changes in capital valuations and adjustments in resource demand may be transmitted to forestry through local investment and factor markets. Climate pressure may also induce capital reallocation toward green production and adaptation projects. W1 therefore captures changes in the strength of climate-shock transmission through regional capital and industrial networks, with the direction determined empirically. It is not equivalent to the general level of industrialization or the complete provincial industrial structure.

H3a: Capital-market exposure to energy-intensive industries significantly moderates the effect of climate risk on forestry TFP.

Regional carbon-emission intensity is measured as regional carbon dioxide emissions divided by regional gross domestic product and is natural-log-transformed in the regressions. Higher carbon-emission intensity generally indicates lower energy efficiency, stronger dependence on conventional energy pathways, and greater environmental-governance and emission-reduction pressures. Following a climate shock, such regions may face simultaneous demands for resources for energy supply, environmental governance and disaster recovery, altering the resources available for forestry production, ecological restoration and technological upgrading. Conversely, stronger transition pressure may prompt local governments to accelerate investment in energy-saving technologies and ecological governance. W2 therefore tests whether carbon-emission characteristics alter the marginal effect of climate risk, with the direction again determined empirically. The index does not represent financial development or the concentration of financial resources in high-carbon activities, and we do not interpret it as a moderating role of the financial system.

H3b: Regional carbon-emission intensity significantly moderates the effect of climate risk on forestry TFP.

3.6 Threshold effect of external network connections

Within the integrated framework, trade openness captures a region’s connectedness to external production networks and markets. The effect of climate risk on forestry TFP may not be linear and uniform across all conditions, but may change structurally with key conditioning variables. We therefore use trade openness as a threshold variable and apply the panel-threshold model of Hansen (1999) to test for critical transitions in the effect of climate risk.

Openness alters both the channels through which climate shocks are dispersed and the scope of their transmission. Imported equipment and technologies, diversified export markets and cross-regional cooperation can provide alternative sources of production inputs, alleviate shortages caused by localized climate shocks, and facilitate the diffusion of forestry technology and adaptive-management experience, producing risk-diversification and technological-spillover effects. Conversely, as regions become more deeply embedded in domestic and international supply chains, climate shocks may be transmitted to local forestry production through raw-material prices, transport costs, external demand and disruptions to trading partners’ production, amplifying supply-chain risk. Supply-chain networks may therefore enhance post-disaster recovery through partner support while also broadening the propagation of climate shocks (Todo et al., 2015; Sun et al., 2024).

The marginal contributions of these two channels need not change proportionally with openness. At low levels of openness, technological learning, input substitution and market diversification generated by additional trade connections may be relatively prominent. As embeddedness in trade networks increases, the effects of supply-chain dependence, external-demand volatility and cross-regional risk transmission may gradually intensify. The effect of climate risk on forestry TFP may therefore change structurally once openness crosses a critical value.

H4: Through both risk-diversification and technological-spillover channels and the transmission of supply-chain shocks, trade openness affects the relationship between climate risk and forestry TFP. Once openness crosses a specific threshold, the direction or magnitude of the climate-risk effect may change, producing a nonlinear threshold effect.

3.7 Heterogeneity in adaptive capacity

Within the integrated framework, forestry management and protection capacity, the digital-economy foundation and environmental-governance intensity primarily represent distinct dimensions of adaptive capacity, although they may also create new network exposures. A higher density of forestry stewardship personnel can strengthen routine patrols, disaster detection and emergency response, but also indicates greater reliance on continuous workforce organization and on-site operations. Transport disruption and suspended operations during extreme weather may therefore cause more pronounced efficiency losses, leaving the direction of the between-group difference to be determined empirically. The digital economy improves access to meteorological information and resource mobilization, but also deepens cross-regional production networks and market connections, allowing climate shocks to propagate through supply chains, energy systems and digital infrastructure. Stronger environmental governance can improve systemic recovery through pollution control, ecological restoration and environmental enforcement. We therefore test heterogeneity in adaptive capacity across three dimensions: forestry production stewardship, information and resource mobilization, and ecological governance.

H5a: The effect of climate risk on forestry TFP differs with forestry management and protection capacity.

H5b: The effect of climate risk on forestry TFP differs with the level of digital-economy development.

H5c: Stronger environmental-protection intensity buffers the adverse effect of climate risk on forestry TFP.

4 Research design

4.1 Sample and data sources

The sample comprises panel data for 30 Chinese provinces from 2012 to 2023, excluding Tibet because of missing forestry data, and contains 360 province-year observations. The primary sources are the China Forestry Statistical Yearbook, China Forestry and Grassland Statistical Yearbook, China Statistical Yearbook, China Energy Statistical Yearbook, China Environmental Statistical Yearbook, provincial statistical yearbooks, and the climate-risk-index database published by the National Climate Center of the China Meteorological Administration. Industrial-pollution data are obtained from successive editions of the China Environmental Statistical Yearbook and supplemented with provincial environmental bulletins. The value of forest ecosystem products is drawn from assessments published by the National Forestry and Grassland Administration. The overall control rate for forest diseases, insect pests and rodents comes from the China Forestry Statistical Yearbook and the China Forestry and Grassland Statistical Yearbook. The underlying search data for the composite index of public attention to pollution issues are obtained from the Baidu Index platform.

4.2 Variable measurement

4.2.1 Outcome variable: ecological total factor productivity in forestry

We measure ecological total factor productivity in forestry using a slack-based super-efficiency EBM (Epsilon-Based Measure) model. Proposed by Tone and Tsutsui (2010), the EBM model integrates radial efficiency and non-radial slack variables within a unified framework and jointly accommodates desirable and undesirable outputs, making it suitable for measuring production efficiency under environmental constraints. Previous studies of agricultural production efficiency and green TFP commonly include labor, capital, land and energy inputs alongside desirable outputs and undesirable outputs such as pollutant emissions (Ahmad et al., 2026; Song et al., 2022). The use of the EBM model to estimate provincial forestry TFP is therefore grounded in established methodological and indicator choices.

For inputs, we follow the forestry-production process and the green-efficiency literature and include five categories: labor, capital, land, forest ecosystem products and energy. Labor input is measured by the number of employees in the forestry system at year end; capital input by completed investment in forestry fixed assets; land input by the area of forestry land; forest-ecosystem-product input by the value of forest ecosystem products; and energy input by allocating total energy consumption according to forestry gross output as a share of regional gross domestic product. These indicators encompass conventional factor inputs while accounting for forest ecosystem products and energy use, thereby providing broad coverage of the forestry production-efficiency process.

For outputs, we specify one desirable output and four undesirable outputs. The desirable output is gross forestry output at constant prices. The undesirable outputs are emissions of SO₂, smoke and dust, solid waste and wastewater, with industrial pollutant emissions allocated according to forestry gross output as a share of regional GDP. Because current statistics do not provide direct provincial emissions data for the forestry sector, allocating pollutant emissions according to sectoral output shares is a common approximation in industry-level green-efficiency studies. It captures the environmental constraints on forestry production while maintaining interprovincial comparability (Ahmad et al., 2026; Song et al., 2022).

4.2.2 Core explanatory variable: composite physical climate risk

We use the composite physical climate risk index (CPRI) to measure provincial physical climate risk. Following the general approach of Wang et al. (2018), who characterize regional climate risk using multiple categories of extreme weather and climate events, and given the available provincial component data, we construct the CPRI from four dimensions: extreme low-temperature days (LTD), extreme high-temperature days (HTD), extreme rainfall days (ERD) and extreme drought days (EED). LTD is the number of days on which the daily minimum temperature falls below the 10th percentile of the local historical distribution; HTD is the number of days on which the daily maximum temperature exceeds the 90th percentile; ERD is the number of days on which daily precipitation exceeds the 95th percentile; and EED is the number of extreme-drought days identified using the standardized precipitation index. After dimensionless transformation, the four components are equally weighted to form an annual composite index, with higher values indicating greater composite physical climate risk in a province-year. Conceptually, LTD, HTD, and ERD primarily capture event-driven acute physical risk, whereas EED captures extreme drought and its persistent and cumulative physical pressure within a year. The CPRI should therefore be interpreted as a composite physical-risk measure dominated by exposure to acute extreme events, rather than a complete measure of all climate-related risks. It excludes transition risks arising from carbon-mitigation policies, substitution by low-carbon technologies, energy-system adjustment and changing market preferences. It also does not directly measure chronic physical risks such as long-term mean-temperature trends, long-term changes in precipitation patterns or shifts in forest climate zones. Compared with a single temperature or precipitation indicator, the CPRI more comprehensively captures the annual physical climate shocks facing forestry production across multiple types of extreme events. Because its distribution is right-skewed, we use its natural logarithm in the empirical analysis, denoted lnCPRI.

4.2.3 Mediating, moderating, and threshold variables

The control gap for forest diseases, insect pests and rodents (M1) is measured as 1 minus the overall control rate for forest diseases, insect pests and rodents; a higher value indicates a larger share that is not effectively managed. The composite index of public attention to pollution issues (M2) is constructed from annual provincial Baidu search indices for the keywords “environmental pollution” and “haze.” We add 1 to each keyword’s search index, take the natural logarithm, standardize the two series separately, and calculate their equally weighted mean. The moderators are capital-market exposure to energy-intensive industries (W1) and regional carbon-emission intensity (W2). W1 is the market capitalization of listed firms in the six major energy-intensive industries divided by the total market capitalization of listed firms in the province. W2 is regional carbon dioxide emissions divided by regional gross domestic product and is natural-log-transformed in the regressions. The threshold variable, trade openness, is measured as total import and export trade divided by regional gross domestic product (Table 1).

Table 1. Components of the ecological total factor productivity in forestry index.

4.2.4 Control variables

Following common specifications in studies of climate risk, green TFP and production efficiency (Song et al., 2023; Ren et al., 2022; Sun and Liu, 2025), the controls include economic-development level, urbanization rate, industrial structure, labor force, innovation level, environmental-regulation intensity, energy structure, and trade openness. Economic-development level is measured as the natural logarithm of GDP per capita; urbanization rate as the share of permanent urban residents; industrial structure as primary-industry output relative to GDP; labor force as the natural logarithm of the number of employed persons at year end; innovation level as the natural logarithm of the number of patents granted; environmental-regulation intensity as government environmental-protection expenditure relative to GDP; energy structure as clean-energy consumption relative to total energy consumption; and trade openness as total import and export trade relative to regional gross domestic product.

4.3 Descriptive statistics and correlation analysis

Table 2 reports descriptive statistics for the main variables. Forestry TFP (EBM efficiency) has a mean of 0.757 and a standard deviation of 0.281, indicating substantial interprovincial variation. The mean lnCPRI is 3.803, with values ranging from 3.251 to 4.435. The control gap for forest diseases, insect pests and rodents has a mean of 0.226 and a standard deviation of 0.213, indicating that, on average, approximately 22.6% of the area affected by pests and rodents was not covered by control measures in the same period and that interprovincial variation was pronounced. After standardization and aggregation, the composite index of public attention to pollution issues has a mean of 0 and a standard deviation of 0.940. Capital-market exposure to energy-intensive industries has a mean of 0.243 and a standard deviation of 0.205. Natural-log-transformed regional carbon-emission intensity has a mean of 9.474, a standard deviation of 0.645, a minimum of 7.524 and a maximum of 10.876. W2 has 360 observations, consistent with the other core variables.

Table 2. Descriptive statistics for the main variables (N = 360).

Table 3 reports Pearson correlation coefficients for the core variables. The correlation between lnCPRI and forestry TFP is −0.130 and is significant at the 5% level. The control gap for forest diseases, insect pests and rodents is significantly negatively correlated with forestry TFP, whereas the composite index of public attention to pollution issues is significantly positively correlated with forestry TFP, consistent in direction with the subsequent mechanism tests. The largest absolute correlation among the core variables is 0.700, between W1 and W2.

Table 3. Pearson correlation matrix for the core variables.

N = 360. *, , and * indicate significance at the 1, 5, and 10% levels, respectively. Definitions of Y, D, M1, M2, W1, W2, and Q are provided in Supplementary Table 1.

4.4 Model specification

4.4.1 Baseline model: double machine learning

To identify the effect of climate risk on forestry TFP, we use a partially linear regression model within the double machine learning framework. The empirical model specifications used in this study are presented in Equations 1–5:

Y it = θ D it + g ( X it , μ i , λ t ) + ε it (1)

Here, Yit denotes forestry TFP in province i and year t; Dit is the core explanatory variable, lnCPRI; Xit is the set of control variables; μi and λt denote province and year fixed effects, respectively; εit is the stochastic error term; and θ is the coefficient of interest capturing the effect of climate risk. Province fixed effects control for time-invariant interprovincial differences, including geographical endowments and baseline forestry resources. Year fixed effects control for macroeconomic policies and common climate-trend shocks. Standard errors are clustered at the province level.

We use DML not simply because the model contains numerous control variables, but primarily to reduce functional-form misspecification bias that may arise in conventional linear regression. Climate-risk exposure and forestry TFP are jointly affected by economic development, urbanization, industrial structure, openness, labor, innovation, environmental regulation and energy structure. Economic development, urbanization and innovation investment may exhibit diminishing marginal effects or stage-dependent patterns, while industrial structure, energy structure and environmental regulation may interact in complex ways with regional development conditions. If a conventional regression omits relevant nonlinear terms or interactions from its prespecified functional form, the estimated climate-risk coefficient may be biased by model misspecification. Our nuisance models contain 8 continuous controls, 29 province indicators and 11 year indicators, yielding 48 candidate features. This is not a conventional ultra-high-dimensional setting in which the number of variables exceeds the sample size. Instead, DML allows learners such as random forests to flexibly estimate the conditional expectations in the outcome equation E(Y|X) and treatment equation E(D|X), reducing reliance on researcher-specified nonlinearities and interactions. Cross-fitting reduces overfitting bias arising from fitting and estimation on the same sample, while the Neyman-orthogonal score attenuates the influence of nuisance-model estimation errors on the core climate-risk coefficient. The practical purpose of DML here is therefore to improve the robustness of target-parameter estimates to functional-form misspecification and regularization bias under complex confounding, rather than to use machine learning to predict forestry TFP or to claim that DML automatically resolves all unobserved confounding or endogeneity. We use random forests as the primary algorithm and replace them with support vector regression, gradient boosting, Lasso and elastic net in algorithm-substitution tests. DML uses 5-fold cross-fitting, repeated 20 times, with a random seed of 42 to improve reproducibility and stability.

To evaluate the finite-sample performance of the nuisance learners, we report out-of-sample R², root mean squared error (RMSE), and mean absolute error (MAE) for both the outcome and treatment equations. These statistics are calculated from the cross-fitted predictions and therefore evaluate predictive performance on observations that were not used to train the corresponding nuisance models. They are used as transparent learner diagnostics rather than as goodness-of-fit measures for the DML target-effect model or as universal validity thresholds.

For the random-forest learner, the outcome equation yields an out-of-sample R² of 0.0002, an RMSE of 0.1639, and an MAE of 0.1051. The corresponding treatment equation yields an out-of-sample R² of 0.3458, an RMSE of 0.1345, and an MAE of 0.1022. Elastic Net produces similar results: the out-of-sample R² values are 0.0003 for the outcome equation and 0.3457 for the treatment equation, with RMSE and MAE values nearly identical to those obtained using random forests. The corresponding diagnostic results for RF and Elastic Net are reported in Supplementary Table 2. The treatment equation therefore has meaningful out-of-sample predictive power, whereas the outcome equation provides little incremental predictive improvement relative to a mean-prediction benchmark.

The specific difference between conventional TWFE and DML lies in how observed confounding is controlled. A conventional TWFE specification generally imposes a prespecified linear and additive form for the covariates, whereas DML retains province and year fixed effects while allowing the nuisance learners to estimate nonlinear covariate effects and interactions more flexibly. The estimates may therefore differ when the treatment and outcome equations contain nonlinear structures that a conventional parametric specification does not capture. Baiardi and Naghi (2024) show that the value added by machine learning in causal inference depends on the presence of such nonlinear nuisance relationships. Clarke and Polselli (2026) further extend DML to static panel settings with fixed effects, providing direct methodological support for using DML when the functional form of confounding is uncertain. We therefore use DML to reduce sensitivity to functional-form misspecification, not to claim that it is universally superior to conventional TWFE.

4.4.2 Mediation analysis

We evaluate the indirect effect of climate risk through each mediator using three quantities. First, path ( a ) is estimated from the mediator equation:

M it = α D it + g ( X it , μ i , λ t ) + υ it (2)

Second, path ( b )and the direct effect ( c ‘ ) are estimated from an outcome equation that includes climate risk and the corresponding mediator simultaneously:

Y it = θ ′ D it + β M it + g ( X it , μ i , λ t ) + ε it (3)

The indirect effect is defined as the coefficient product ( ab ). We conduct formal inference on this product using the Delta method and a province-clustered covariance matrix. The variance of the indirect effect is calculated as:and its standard error is ( SE ( ab ) = Var ( ab ) ) . The corresponding z-statistic, two-sided p-value, and confidence interval are reported. Thus, evidence of an indirect effect is based on formal inference for the coefficient product rather than solely on the separate significance of paths ( a ) and ( b ).

[ Var ( ab ) = b 2 Var ( a ) + a 2 Var ( b ) + 2 abCov ( a , b ) , ]

4.4.3 Moderating effects

Moderating effects were identified using mean-centered interaction terms. The model was specified as follows:

Y it = θ D it + ρ ( D it × W it ) + η W it + g ( X it , μ i , λ t ) + ε it (4)

Here, Wit is the moderator and ρ is the coefficient of the interaction term. A value of ρ > 0 indicates that the variable mitigates the adverse effect of climate risk on forestry TFP, whereas ρ < 0 indicates that it amplifies this effect. Within the DML framework, we treated the interaction term as the treatment and included the core explanatory variable, moderator and control variables jointly in the orthogonalization procedure.

4.4.4 Endogeneity sensitivity test

To examine the sensitivity of the baseline estimates to potential endogeneity, we employ a DML-PLIV specification. We replace one-period-lagged climate risk with an interaction between annual ENSO intensity and predetermined provincial historical temperature-anomaly sensitivity as the instrument for contemporaneous climate risk. Lagged climate risk is not retained as an instrument because earlier climate shocks may directly affect current forestry TFP through persistent forest-stand damage, soil and moisture conditions, pest accumulation, and recovery expenditure.

Provincial historical temperature-anomaly sensitivity is estimated using monthly provincial temperature data and Niño3.4 sea-surface-temperature anomalies for 1981–2011. For each province, we regress the absolute temperature anomaly on ENSO intensity while controlling for a linear time trend and standardize the resulting provincial coefficient. The instrument is defined as:

[ Z it = ENS O t × Sensitivit y i . ]

The instrument exploits cross-provincial differences in historical exposure to a common external ENSO shock. The provincial sensitivity measure is estimated entirely before the 2012–2023 study period and is therefore predetermined relative to the outcome sample. Province fixed effects absorb time-invariant provincial characteristics, including the main effect of historical sensitivity, while year fixed effects absorb common annual shocks, including the main effect of ENSO intensity.

4.4.5 Threshold effects

To examine nonlinearities in the effect of climate risk on forestry TFP, we applied Hansen’s panel threshold model, specified as follows:

Y it = β 1 D it I ( q it ≤ γ ) + β 2 D it I ( q it > γ ) + δ X it + μ i + λ t + ε it (5)

Here, qit is the threshold variable, γ is the threshold value to be estimated, and β1 and β2 denote the effects of climate risk on forestry TFP on either side of the threshold. The threshold was determined by grid search over the candidate interval after 15% trimming. Its significance was assessed using 300 residual Bootstrap replications, and its confidence interval was constructed from the profile likelihood-ratio statistic of Hansen (1999).

4.4.6 Heterogeneity analysis and between-group difference tests

We examined heterogeneity across three dimensions: forestry management and protection capacity, the digital economy, and environmental-protection intensity. Forestry management and protection capacity was measured as the ratio of forestry-system employees to forestland area in 2011. The 30 provinces were divided into high and low groups at the cross-province median, ensuring that grouping preceded the 2012–2023 study period. For the digital economy and environmental-protection intensity, provinces were divided at the median of their sample-period means.

For each heterogeneity dimension, the high- and low-group DML models are estimated separately using the same control variables, fixed effects, cross-fitting procedure, and province-clustered inference. We test the null hypothesis [ H 0 : β H = β L , ]. Because the two groups contain non-overlapping provinces, the coefficient estimates are based on independent sets of province-level clusters. The standard error of the coefficient difference is calculated as

[ SE ( β H − β L ) = S E H 2 + S E L 2 . ]

The corresponding test statistic isand the two-sided p-value is 2 [ 1 − Φ ( ∣ z ∣ ) ].

[ z = β H − β L SE ( β H − β L ) , ]

5 Empirical results and analysis

5.1 Baseline regression results

Table 4 reports the baseline estimates of the effect of climate risk on forestry TFP. Under two-way fixed effects, all five learning algorithms, random forest, support vector regression, gradient boosting, Lasso and elastic net, yielded significantly negative coefficients for lnCPRI. The consistent signs and significance across the five algorithms indicate that the baseline findings in Table 4 do not depend on a particular nuisance learner, supporting H1. Importantly, these estimates capture the composite physical climate risk represented by the CPRI, particularly exposure to extreme cold, heat, rainfall and drought events, and cannot be directly extrapolated to climate transition risk. Thus, H1 supports an adverse effect of physical climate shocks on forestry production efficiency, not an effect of carbon policy, technological transition or market transition on forestry production efficiency.

Table 4. Baseline regression results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote significance at the 1, 5, and 10% levels, respectively.

Using the random-forest estimate with two-way fixed effects as the baseline, an increase of 1 sample standard deviation in lnCPRI (0.187) was associated with an estimated decline of approximately 0.066 in forestry TFP (0.352 × 0.187). After province and year fixed effects were included, the absolute coefficients increased for all learners in Table 4; for example, the random-forest estimate changed from −0.210 to −0.352. This pattern indicates a stronger within-sample negative association after accounting for time-invariant provincial characteristics and common year shocks. However, the change in coefficient magnitude alone does not establish that any specification is necessarily closer to the true causal effect.

The out-of-sample diagnostics show different predictive performance across the two nuisance equations. For random forest, the outcome equation yields an out-of-sample R² of 0.0002, an RMSE of 0.1639, and an MAE of 0.1051, whereas the treatment equation yields an out-of-sample R² of 0.3458, an RMSE of 0.1345, and an MAE of 0.1022. Elastic Net produces very similar results. These diagnostics are reported transparently as measures of nuisance-model prediction rather than as criteria for validating the target-effect estimate.

As a conventional benchmark, the TWFE model yields a coefficient of 0.0515 on lnCPRI [SE = 0.0435, p = 0.238; 95% CI = (−0.0341, 0.1371)]. The difference between the TWFE and DML estimates indicates that the estimated climate-risk relationship is sensitive to functional-form specification. This comparison provides empirical motivation for allowing nonlinearities and interactions in the nuisance functions.

5.2 Robustness tests

Table 5 first assesses the robustness of the baseline results in four ways. In column (1), replacing the dependent variable with SBM efficiency yielded a coefficient of −0.503 (p = 0.019), indicating that the adverse effect of climate risk on forestry TFP does not depend on a particular efficiency measure. In column (2), replacing 5-fold with 10-fold cross-fitting yielded a coefficient of −0.345 (p = 0.036), broadly consistent with the baseline estimate. In column (3), restricting the sample period to 2016–2023 yielded a coefficient of −0.365 (p = 0.055), significant at the 10% level and consistent in sign with the full-sample result. In column (4), winsorizing the main variables at 1% yielded a coefficient of −0.354 (p = 0.043), indicating that the baseline result was not driven by extreme values. All four estimates were negative; three were significant at the 5% level and one at the 10% level, with coefficients ranging from −0.503 to −0.345.

Table 5. Robustness test results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote significance at the 1, 5, and 10% levels, respectively. Column (5) uses ln[(LTD+HTD+ERD)/3]. Column (6) uses the natural logarithm of the entropy-weighted composite index, with entropy weights of 0.262 for LTD, 0.083 for HTD, 0.416 for ERD and 0.239 for EED.

To further assess whether the findings depended on the component composition and equal weighting of the CPRI, we conducted two additional tests using alternative climate-risk measures. First, EED is identified using the standardized precipitation index and thus displays pronounced persistence and accumulation, whereas LTD, HTD and ERD primarily count extreme-event days identified from temperature- or precipitation-percentile thresholds. We therefore excluded EED, averaged the remaining three components with equal weights and took the natural logarithm. In column (5), the alternative index yielded an estimated coefficient of −0.359 (p = 0.022), which remained negative and significant at the 5% level. Thus, the baseline result was not driven solely by EED, whose measurement rule differs from those of the other components. Second, to assess the effect of assigning equal weights to all four components, we retained the original component scores and derived entropy weights from the information dispersion of each indicator. The resulting weights for LTD, HTD, ERD, and EED were 0.262, 0.083, 0.416, and 0.239, respectively, and we took the natural logarithm of the entropy-weighted composite index. The estimated coefficient in column (6) was −0.337 (p = 0.009), significant at the 1% level. Both estimates were broadly consistent with the baseline regression in sign and magnitude, indicating that the central finding does not depend entirely on the EED component or the equal-weight aggregation of the four components.

5.3 Mediation results

Table 6 reports the regression results for the two proposed transmission pathways, For the control gap for forest diseases, insect pests and rodents, the estimated effect of climate risk on the mediator, path ( a ), is 0.209 and is significant at the 5% level. When climate risk and the control gap are included simultaneously in the outcome equation, the estimated coefficient on the control gap, path ( b ), is −0.541 and is significant at the 1% level, while the direct effect of climate risk, ( c ′ ), is −0.244 and is significant at the 10% level. The estimated indirect effect through the control gap is ab = −0.113, with a Delta-method standard error of 0.064 and a p-value of 0.078.

Table 6. Mediation effect test results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

For the public-attention pathway, the estimated effect of climate risk on the composite index of public attention to pollution issues, path ( a ), is −1.553 and is significant at the 1% level. When climate risk and public attention are included simultaneously in the outcome equation, the estimated coefficient on public attention, path ( b ), is 0.096 and is significant at the 1% level, while the direct effect of climate risk, ( c ′ ), is −0.222 and is significant at the 10% level.

The estimated indirect effect through public attention is (ab = −0.149), with a Delta-method standard error of 0.053, a p-value of 0.005, and a 95% confidence interval of approximately (−0.253, −0.045). The confidence interval excludes zero, indicating a statistically significant negative indirect effect through the allocation of public attention to pollution issues.

5.4 Moderation results

Column (1) of Table 7 reports the moderating effect of capital-market exposure to energy-intensive industries. The interaction coefficient was 0.904 (p < 0.05), indicating that this measure of capital-market exposure altered the marginal effect of climate risk on forestry TFP. This finding does not imply that expanding energy-intensive industries yields ecological benefits. A more cautious interpretation is that shared characteristics, including the sectoral composition of listed firms, existing infrastructure and regional capacity to respond to risk, may be associated with the short-term recovery of forestry production after climate shocks.

Table 7. Moderating effect test results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

In column (2), the interaction coefficient for ln (regional carbon-emission intensity) was 0.211 (p < 0.10), indicating that the carbon emissions burden per unit of regional output significantly altered the marginal effect of climate risk on forestry TFP. Because this variable contains no data on financial assets, credit or deposits, the result cannot be interpreted as a liquidity-buffering effect of the financial system. It indicates only that high- and low-carbon regions may differ in their combined industrial characteristics, energy infrastructure and risk exposure; it does not imply that increasing carbon emissions intensity improves forestry efficiency.

5.5 Endogeneity sensitivity test results

Table 8 reports the results of the instrumental-variable endogeneity test. In the first stage, the interaction between ENSO intensity and historical temperature-anomaly sensitivity is positively and significantly associated with contemporaneous physical climate risk (β = 0.064, SE = 0.020, p < 0.01). The first-stage cluster-robust F-statistic is 10.579, slightly above the conventional rule-of-thumb threshold of 10, and the partial R² is 0.085. These results indicate that the instrument has meaningful explanatory power for contemporaneous physical climate risk and does not exhibit an obvious weak-instrument concern.

Table 8. Endogeneity sensitivity test results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote significance at the 1, 5, and 10% levels, respectively. The first-stage cluster-robust F-statistic and partial R² assess instrument relevance.

In the second stage, the estimated coefficient on lnCPRI is −1.391 (SE = 0.759, p = 0.067). The coefficient remains negative and is statistically significant at the 10% level. Its direction is consistent with the baseline DML estimate, providing supplementary evidence that the negative relationship between physical climate risk and forestry TFP persists under the instrumental-variable specification.

5.6 Threshold effect results

Table 9 reports the threshold estimates and test results for trade openness. The estimated threshold is 0.146, with an F-statistic of 7.670 and a bootstrap p-value of 0.070 based on 300 residual-bootstrap replications. The null hypothesis of no threshold effect is therefore rejected at the 10% significance level. The 90% confidence interval for the threshold, constructed using the profile likelihood-ratio statistic of Hansen (1999), is (0.071, 0.292), which contains the estimated threshold. Since trade openness is measured as the ratio of total imports and exports to regional gross domestic product, the estimated threshold corresponds to a trade-openness level of approximately 14.6% of regional GDP.

Table 9. Threshold effect test results.

Province-clustered robust standard errors are reported in parentheses. *, , and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

The regime-specific estimates indicate that the relationship between physical climate risk and forestry TFP differs across the threshold. When trade openness is at or below 0.146, the estimated coefficient on lnCPRI is 0.181 (p = 0.021), indicating a positive and statistically significant association. When trade openness exceeds 0.146, the coefficient becomes negative at −0.102, although it is not statistically significant at conventional levels (p = 0.115). The difference between the two regime-specific coefficients is 0.283, with a province-clustered robust standard error of 0.123. The corresponding Wald test yields a z-statistic of 2.300 and a p-value of 0.021, indicating that the two coefficients differ significantly at the 5% level. Taken together, these results provide evidence of a nonlinear change in the climate risk–forestry TFP relationship across different levels of trade openness.

5.7 Heterogeneity analysis

Table 10 reports heterogeneous estimates along three dimensions, forestry management and protection capacity, the digital economy and environmental protection intensity, together with formal tests of between-group coefficient differences.

Table 10. Heterogeneity analysis results.

Province-clustered robust standard errors are reported in parentheses; *, , and * denote significance at the 1, 5, and 10% levels, respectively.

Forestry management and protection capacity is measured as the number of forestry-system employees per unit of forest land in 2011, and the 30 provinces are divided into high and low groups at the median. The results show that the coefficient on physical climate risk is −0.616 (p = 0.015) in the high-capacity group and −0.032 (p = 0.821) in the low-capacity group. The coefficient difference between the high and low groups is −0.584, with a p-value of 0.045. These results indicate that regions with higher densities of forestry management and protection personnel exhibit greater efficiency sensitivity to climate shocks, possibly because forestry production in these regions depends more heavily on continuous field operations and organizational coordination.

The coefficients on physical climate risk are −0.638 and −0.034 in the high- and low-digital-economy groups, respectively, with a between-group difference of −0.605 (p = 0.040). This finding indicates that a stronger digital foundation does not automatically translate into greater adaptive capacity in forestry. Regions with more developed digital economies may also have more complex cross-regional factor flows and supply-chain linkages. However, we do not directly test these specific transmission channels and therefore do not interpret the result as evidence that the digital economy itself exacerbates physical climate risk. The corresponding coefficients are −0.032 and −0.634 in the high- and low-environmental-protection-intensity groups, respectively, with a between-group difference of 0.602 (p = 0.008). The forestry-management result provides support for H5a. The digital-economy difference supports H5b, while the environmental-protection result supports H5c.

6 Conclusion and policy implications

6.1 Main findings

Using panel data for 30 Chinese provinces from 2012 to 2023, this study applies double machine learning to examine how composite physical climate risk dominated by exposure to acute extreme events affects ecological total factor productivity in forestry. The results show a significant negative effect on forestry TFP. Physical climate risk also exerts negative indirect effects by widening the control gap for forest diseases, insect pests and rodents and by crowding out public search attention to general pollution issues, further reducing forestry TFP. Capital-market exposure to energy-intensive industries and regional carbon-emission intensity significantly modify the marginal effect of physical climate risk; the former is not equivalent to general industrial structure, and the latter does not represent the financial system. The threshold and heterogeneity analyses further reveal nonlinear and regionally differentiated effects. These conclusions apply only to the physical climate risk captured by the CPRI and cannot be directly extrapolated to transition risk.

6.2 Policy implications

To avoid extending statistical relationships to untested policy instruments, the recommendations below are organized around the baseline effect, the internal transmission pathways and the identified boundary conditions. They identify directions and potential intervention points for forestry climate adaptation; they do not imply that any specific policy instrument has received a direct causal evaluation in this study.

First, coordinate the monitoring of physical climate risk with that of forestry production efficiency. The baseline and alternative-measure results show that composite physical climate risk, primarily comprising exposure to extreme cold, heat, rainfall and drought, is significantly associated with lower forestry TFP. These indicators can therefore be monitored jointly with forestry TFP, tree damage, disruptions to production operations and progress in ecological restoration. When physical climate risk rises sharply, forest-damage assessments and preparations for production recovery can be initiated promptly. The monitoring results should be used to identify risk exposure and resource needs; estimated sample coefficients should not be converted mechanically into uniform administrative targets or fixed funding-allocation ratios.

Second, narrow the control gap for forest diseases, insect pests and rodents. The pathway tests indicate that physical climate risk may further suppress forestry TFP by widening this gap, supporting broader disaster monitoring and control coverage. Meteorological-anomaly warnings can be linked more closely to the monitoring of harmful forest organisms; cross-regional surveillance, quarantine and coordinated control can be strengthened; and monitoring equipment, professional control teams and emergency-response resources can be prioritized in regions with larger control gaps. The overall control rate and the corresponding gap should also be tracked over time to determine whether additional climate pressure is exceeding existing control capacity. Because this study does not test deductibles, coinsurance, differentiated premium rates or other insurance-contract arrangements, the findings do not support institutional recommendations concerning those instruments.

Third, maintain environmental information provision and public oversight during climate disasters. The pathway results show that higher physical climate risk is associated with lower public search attention to general pollution issues, whereas such attention is positively associated with forestry TFP. Alongside weather warnings and reports of casualties and property losses, public agencies can therefore continue to disclose information on pollution control, forest damage, the spread of diseases, insect pests and rodents, and progress in ecological restoration. This may prevent general environmental-governance issues from being displaced for prolonged periods by short-term emergency information. Because this study uses search indices only for “environmental pollution” and “haze,” the recommendation is limited to information provision and public communication concerning general pollution issues; it should not be generalized to comprehensive public environmental awareness or used to justify specific financial, insurance or industrial policies.

Fourth, differentiate climate adaptation according to regional structural exposure and adaptive capacity. The moderation, threshold and heterogeneity results show that the magnitude of the effect of physical climate risk varies with capital-market exposure to energy-intensive industries, regional carbon-emission intensity, trade openness and regional adaptive capacity. These indicators may inform risk diagnosis and policy prioritization, but they do not justify a uniform forestry climate-adaptation program. Regions with greater structural exposure, weaker environmental-governance capacity or limited disaster-recovery capacity may require earlier allocation of resources for ecological restoration, pollution control and forestry emergency preparedness. The proxies must nevertheless be interpreted within their measurement boundaries: capital-market exposure to energy-intensive industries does not represent the complete industrial structure; regional carbon-emission intensity is not a measure of the financial system; the trade-openness threshold is a sample-specific regime boundary rather than an optimal level of openness or an administrative standard; and the between-group result for forestry management and protection capacity is sensitive to model choice. These indicators are therefore more appropriate for regional risk screening than for rigid administrative thresholds.

6.3 Limitations and future research

This study has five main limitations. First, the control gap for forest diseases, insect pests and rodents is constructed as the complement of the overall control rate. It captures incomplete control coverage but cannot distinguish among diseases, insect pests and rodents or directly measure control quality, disaster severity or operator inputs. The composite index of public attention to pollution issues combines two keywords but does not fully cover physical climate risk, forest protection or ecological governance in forestry. Future research could combine more granular data on areas affected by harmful forest organisms, control inputs and control outcomes with multi-keyword, multi-platform information on “extreme weather,” “forest protection,” “forest diseases and insect pests,” and “forestry carbon sinks.” Second, capital-market exposure to energy-intensive industries covers only listed companies and is not equivalent to the complete provincial industrial structure; regional carbon-emission intensity likewise does not measure financial development. Future studies could use more direct measures, such as the output share of energy-intensive industries. Third, the undesirable outputs used to estimate forestry TFP are approximated by allocating industrial pollution emissions according to forestry output as a share of regional GDP. More granular data on forestry-specific pollution emissions would improve measurement precision. Fourth, the CPRI combines indicators of extreme cold, heat, rainfall and drought and therefore primarily captures event-driven physical climate risk. Although the drought component reflects some persistence and accumulation of climate shocks, the index does not directly measure chronic physical risks such as long-term trends in mean temperature, shifts in long-term precipitation patterns, sea-level rise or migration of forest climate zones. Nor does it include transition risks arising from climate policy, low-carbon technologies or changes in market structure. Future research could construct separate measures of acute physical risk, chronic physical risk and transition risk and compare their effects on ecological total factor productivity in forestry. Fifth, the analysis is limited to the provincial level. County- or forest-farm-level data are needed to test the micro-level mechanisms further.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this manuscript, generative AI tools were used only for language polishing, grammar checking, and improving readability. The authors carefully reviewed and edited all AI-assisted content and take full responsibility for the final content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/ffgc.2026.1930608/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Hao Luo, Leipzig University, Germany

Reviewed by: Wenxiu Hu, Xi’an University of Technology, China

An Zeng, Chinese Academy of Sciences (CAS), China