## wpiea2025098-print-pdf - 2.1 Data Sources ## Source details **Canonical URL:** [wpiea2025098-print-pdf - 2.1 Data Sources](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025098-print-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025098-print-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025098-print-pdf.pdf.json) --- ### Data and sample construction - Combined Juhász et al. (2023) IP database with bilateral trade flows (Gaulier and Zignago, 2010). - Final dataset: HS6-digit products spanning 156 countries between 2009 and 2022; 5018 products in the balanced panel. - Green products list: compiled from six sources; final list contains 869 HS6-digit products. - IP counts: country-product-year stock of active protectionist IPs (GTA red evaluation) converted to first differences (∆IP) to capture IP shocks. - Reporting-lag rule: keep only policies announced and published by GTA within the same calendar year. - Missing active-IP values filled with zero if the country has any positive IP in 2009–2022. - Caveats: GTA records only policies affecting commercial interests; database starts in 2009; Juhász et al. (2023) focus on national-level activities; IP indicators are not measures of intensity (e.g., subsidy value). ### Motives, instruments, and coverage - Motive assignment: LLM-based classification (Evenett et al., Forthcoming) trained on NIPO stated motives (available since 2023); focus on four motives: climate mitigation, strategic competitiveness, geopolitical/national security, GVC resilience. - Reporting on motives: approximately 60% of IPs can be assigned at least one stated motive; among IPs with stated motive(s) in 2018–2022: strategic competitiveness ≈ 50%, climate mitigation = 34%, GVC resilience = 16%, geopolitical/national security = 1%. - GTA evaluation categories: red (protectionist), amber (ambiguous), green (liberalizing); analysis restricts to red (protectionist) IPs. - Policy instruments: GTA maps to 66 instruments aggregated into eight groups per UN MAST: export barriers, import barriers, domestic subsidies, export incentives, FDI measures, public procurement, local content, others. - Descriptive prevalence: - Domestic subsidies and export incentives are most commonly used in both AEs and EMs. - One export incentive in EMs targets on average almost 400 products; one local content measure in AEs targets close to 200 products on average. - Roughly 70–80% of protectionist domestic subsidies are classified as IPs (Rotunno and Ruta (2024) reference). - Around 10% of IPs in 2009–2022 were removed. - Aggregate policy share: IPs rose from under 25 percent before 2017 to over 35 percent during 2017–2022. ### Trade competitiveness and RCA measures - Main outcome: revealed comparative advantage (RCA) proxied by the Balassa index; alternative RCA from Vollrath (1991) used for robustness. - RCA formula structure: RCA_cpt = (X_cpt / sum_p X_cpt) / (sum_c X_cpt / sum_cp X_cpt). - Export data: CEPII BACI (Gaulier and Zignago, 2010). - Panel: balanced panel of 156 countries and 5018 products for 2009–2022. - Dependent variable for LP-DiD: rca_{c,p,t+h} = ln(RCA_{c,p,t+h} + 10^{−3}); approximately 42% of observations in the clean sample have zero RCA. ### Empirical method and sample construction - Main empirical approach: local projection difference-in-differences (LP-DiD) to address bias from DiD-TWFE under staggered treatment and dynamic effects. - Clean-sample treatment definition (equation (1) in source): - Treatment = first-time IP shock D_{c,p,t} where ∆IP_{c,p,t} > 0 and D_{c,p,t−j} = 0 for 1 ≤ j ≤ L. - Clean controls require D_{c,p,t−j} = 0 for −H ≤ j ≤ L. - Baseline stabilization lag L = 5 (robustness with L = 3); baseline controls include two lags of the dependent variable (robustness with three lags). - Fixed effects: country-product α_{c,p}, country-year δ_{c,t}, product-year ρ_{p,t}; standard errors clustered at the country-product level. - Final clean sample size: approximately 30,000 observations in the treatment group and close to 10 million observations in the control group. - Instrument-specific clean-sample refinement to study heterogeneous effects by instrument i ∈ [1,8]; focus empirically on domestic subsidies (14,658 treated units) and export incentives (7,519 treated units). ### LLM ensemble for motive classification (methodological appendix highlights) - Base model: RoBERTa-Large finetuned in an ensemble of ten runs; full finetuning with AdamW lr = 1×10^{−5}, 8 epochs, weight decay = 0.01, warm-up 10% of steps. - Classification threshold: weighted probability > 60% for assigning motive. - Ensemble performance (selected metrics): - Climate Change Mitigation: Ensemble RoBERTa-Large Accuracy / Macro-F1 = 0.97 / 0.94. - Strategic Competitiveness: 0.92 / 0.90. - Geopolitical Concerns: 0.95 / 0.88. - GVC Resilience: 0.96 / 0.90. ### Robustness checks (overview) - Seven alternative scenarios tested, including: - Excluding China — results robust. - Changing stabilization lag L from 5 to 3. - Controlling for third lag of ln(RCA + 10^{−3}) and ΔNonIP. - Using alternative RCA adjusted for imports: ln(RCA_exports + 10^{−3}) − ln(RCA_imports + 10^{−3}). - Using ln(RCA + 1) instead of ln(RCA + 10^{−3}). - Excluding units treated in 2020 to avoid Covid-specific confounding. - Using all GTA subsidies. - General robustness: main findings are generally robust; exception: export incentives’ medium-term positive effect is less pronounced in some exercises. ### Key empirical findings — average effects and margins - Average effect: - Products targeted by IPs experience a 5.6% higher increase in trade competitiveness than non-targeted products three years after the introduction of the IP. - Margin decomposition: - Short-term increase in competitiveness mainly driven by the intensive margin (increase in exports of products already in the country’s export basket). - Medium-term increase driven by the extensive margin (export participation of previously non-exporting products). - Extensive-margin result: targeted products experience a higher probability to start exporting three years after treatment (conditional sample: Export_{c,p,t−1} = Export_{c,p,t−2} = 0). - Intensive-margin estimates are statistically insignificant but mirror average-effect shape. ### Heterogeneity by initial competitiveness and instruments - Initial competitiveness (interaction Treated × (RCA_{c,p,t−1} > 1)): - Previously competitive products (RCA_{c,p,t−1} > 1) experience a large short-term boost in competitiveness after the IP shock; the positive association peaks after two years and then declines and becomes statistically insignificant at four years, though estimated magnitude can remain as high as 9 percent. - Initially non-competitive products (RCA_{c,p,t−1} ≤ 1) show an initial decline in RCA and then a gradual, statistically insignificant increase over the horizon considered. - Policy-instrument heterogeneity (domestic subsidies vs export incentives): - Domestic subsidies: associated with a short-term 5 percent increase in competitiveness for targeted relative to non-targeted products, which fades over time. - Export incentives: associated with an initial 1 percent decline in competitiveness, followed by longer-term improvements. - Interpretation: domestic subsidies produce short-term boosts; export incentives appear to encourage sustained medium-term gains but may conflict with WTO rules and risk retaliation. ### Green vs. non-green products - Dynamics and effect sizes: - IPs targeting green products increase RCA by about 20 percent after 4 years. - IPs targeting non-green products show only a mild short-term increase in RCA and smaller/insignificant medium-term effects. - Timing and margin composition differences: - For green products, the positive association manifests mostly in the medium-term; for non-green products, the positive association appears only in the short term. - For green products, long-run RCA gains are mainly driven by extensive margin (previously non-competitive products entering exports); non-green products’ effects are largely driven by initially competitive products and intensive margin dynamics. - Instrument interaction: - Stronger positive association between IPs (both subsidies and export incentives) and RCA when IPs target green products, particularly at longer horizons. - Domestic subsidies: for green products insignificant short-term and positive medium-term; for non-green products small temporary improvement turning negative medium-term. - Export incentives: both green and non-green products receive boosts, with larger and more pronounced effects for green products. ### Cross-product spillovers and green value chains - Value chains studied: wind turbines, photovoltaic panels, electric vehicles (mapping from Rosenow and Mealy (2024)); products assigned to stages: raw materials, processed materials, subcomponents, end products. - Findings on spillovers: - IPs targeting upstream products are associated with larger improvements in the RCA of products that use these upstream products relative to IPs targeting the same stage. - IPs targeting downstream products yield similar effects as those targeting products at the same value chain stage. - Interpretation: upstream IPs may alleviate capacity constraints and benefit downstream products through reductions in input costs. - Instrument composition across stages: - Domestic subsidies account for the majority of IPs in all stages, more prevalent in initial stages (raw materials, over 70%) compared to downstream stages (between 40 and 60%); composition may partially explain differential impacts. ### Limitations, scope, and causality caveats - Partial assessment: analysis compares relative performance of targeted versus non-targeted products; does not assess overall welfare gains or absolute desirability of IPs. - Full welfare assessment requires structural, general-equilibrium analysis and information on IP sizes and fiscal costs. - Causality caveats: - LP-DiD mitigates biases relative to DiD-TWFE under staggered treatment but does not fully eliminate endogeneity concerns (selection bias, reverse causality, endogenous timing of treatment). - Results are informative of expected effects of IPs on trade competitiveness but do not necessarily establish fully causal relationships. ### Policy implications and research agenda - Substantive implications: - On average, protectionist IPs are positively associated with improvements in competitiveness of targeted products, but effects are heterogeneous by product characteristics and policy instruments. - Policymakers face trade-offs: domestic subsidies deliver short-term gains; export incentives tend to deliver medium-term gains but may provoke international objections and retaliatory measures. - IPs targeting green products show particularly strong medium-term gains and promote entry of previously non-exporting green products. - Cautions: - IPs should be handled with care due to potential general-equilibrium spillovers, fiscal costs, and international policy consistency concerns. - Countries must weigh costs and benefits in general equilibrium, ensure consistency with international rules, and prioritize multilateral policy cooperation. - Research agenda: - Incorporating general-equilibrium channels and fiscal cost information into analyses of IPs is a fruitful avenue for future research. *Source: wpiea2025098-print-pdf - Section 2.1 and related excerpts (IMF Working Paper).* ### 2.1 Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ### wpiea2025098-print-pdf - 2.1 Data Sources ### Data and sample - Combined Juhász et al. (2023) novel database of industrial policies (IPs) with trade flows (Gaulier and Zignago, 2010). - Final dataset: HS6-digit products spanning 156 countries between 2009 and 2022. - Green products list: compiled from six different sources; final list contains 869 HS6-digit products. - Business-press evidence: number of articles mentioning IP rose from less than 1000 times in 1990 to more than 18000 times in 2019. - Most commonly stated IP motive: strategic competitiveness, constituting over 35% of IPs in 2023. ### Methods - Empirical approach: local projection difference-in-differences (LP-DiD) to explore dynamics between IPs and trade competitiveness (Balassa revealed comparative advantage, RCA). - Motivation for LP-DiD: alleviates bias from DiD-TWFE under staggered treatment and dynamic effects (e.g., previously treated units serving implicitly as controls can bias estimates). - Unit of analysis: country-product pairs that may receive multiple IP treatments at different points in time. - Complementary approaches and sensitivity analyses described in appendices (e.g., robustness exercises excluding China, alternative lag structures, alternative RCA measures). ### Key empirical findings: average effects and margins - On average, products targeted by IPs experience a greater increase in trade competitiveness (RCA) compared to non-targeted products. - Short-term increase in competitiveness mainly driven by the intensive margin (increase in exports of products already in the country’s export basket). - Medium-term increase driven by the extensive margin (export participation of previously non-exporting products). ### Heterogeneity by product characteristics and instruments - Initial competitiveness: - Positive association between IPs and competitiveness is mostly driven by products that were globally competitive prior to the IP announcement. - Policy instruments (focus on the two most popular instruments): - Domestic subsidies: associated with short-term improvement in trade competitiveness, with effects vanishing after a few years. - Export incentives: associated with larger medium-term improvements in trade competitiveness of targeted products compared to non-targeted ones. - Green vs. non-green products (three notable differences): - Timing: For green products, the positive association manifests mostly in the medium-term; for non-green products, the positive association appears only in the short term. - Margin composition: For green products, long-run RCA gains are mainly driven by products that had not yet established comparative advantage in the global market (extensive margin); this contrasts with non-green products. - Instrument interaction: Stronger positive association between IPs (both subsidies and export incentives) and RCA when IPs target green products, particularly at longer horizons. ### Cross-product spillovers and value chains - Focused value chains: wind turbines, photovoltaic panels, and electric vehicles (mapping between HS6-digit products and production stages from Rosenow and Mealy (2024)). - Suggestive OLS evidence: IPs targeting upstream products are associated with larger improvements in the RCA of products that use these upstream products compared to IPs targeting the same value chain stage; IPs targeting downstream products yield similar effects as those targeting products at the same value chain stage. - Interpretation: Upstream IPs may alleviate capacity constraints and benefit downstream products through reductions in input costs. ### Limitations and scope - Analysis provides a partial assessment of IPs’ impact on competitiveness: - Compares relative performance of targeted versus non-targeted products; does not assess overall welfare gains or absolute desirability of IPs. - Full welfare assessment would require a structural, general-equilibrium analysis and information on IP sizes and fiscal costs. - Causality caveats: - LP-DiD reduces certain biases relative to DiD-TWFE but does not fully eliminate endogeneity concerns (selection bias, reverse causality, endogenous selection into time of treatment). - Results are informative of expected effects of IPs on trade competitiveness but do not necessarily establish a fully causal relationship. ### Contribution to literature - Cross-country empirical analysis of IPs’ effects on trade competitiveness, complementing country-specific case studies and firm-/industry-level analyses. - Novelty: - Analyzes dynamics using LP-DiD under staggered treatment. - Documents differential dynamics for green versus non-green products. - Examines temporal patterns by policy instrument (short-term vs medium-term effects). *Source: wpiea2025098-print-pdf - Section 2.1 Data Sources (IMF Working Paper).* ### 2.1 Data Sources ### 2.1 Data Sources ### Data on Industrial Policies - Country-product level IP counts are from Juhász et al. (2023), who classify whether policy announcements in the Global Trade Alert (GTA) database over the period 2009-2022 qualify as IPs using machine learning. - GTA collects state policy measures and credible announcements that discriminate against foreign commercial interests. - Juhász et al. (2023) focus on national-level economic activities and define IP as “goal-oriented state action. The purpose is to shape the composition of economic activity. Specifically: industrial policy seeks to change the relative prices across sectors or direct resources towards certain selectively targeted activities (e.g., exporting, R&D), to shift the long-run composition of economic activity”. - GTA provides, for each policy: implementing jurisdiction, HS codes of targeted products, type of instrument used, and GTA evaluation. GTA evaluation categories: red, amber, and green. - Red: protectionist; almost certainly discriminate against foreign commercial interests. - Amber: ambiguous (“likely involve discrimination against foreign commercial interests”). - Green: liberalizing. - Analysis restricts attention to IPs with a red GTA evaluation (protectionist IPs). - Reporting-lag adjustment: only keep policies announced and published by GTA within the same calendar year to avoid inflated counts in earlier years. - For each country-product-year, number of active IPs is counted as IPs that are announced but not yet removed, with starting point the announcement year (to account for anticipation). - Missing values of active IP counts are filled with zero if the country has ever had a strictly positive number of active IPs in at least one product during 2009-2022. - The active IP count is a stock variable; empirical analysis uses the year-to-year change (first difference) to capture IP shocks (a flow variable). Descriptive statistics in Section 2.2 use counts of announced IPs. - Data caveats: - EMDEs may have less transparent and granular policy disclosure and reporting standards, possibly underestimating IPs relative to AEs. - GTA-derived IP data records only policies affecting commercial interests. - Database starts in 2009; may underestimate pre-2009 stock of IPs for countries that were active before 2009. - Juhász et al. (2023) focus on national-level activities; decentralized implementations and indirect incentives (e.g., subsidized bank loans) may not be fully captured. - IPs in Juhász et al. (2023) are indicator variables and not measures of intensity (e.g., subsidy value). - Complementary evidence: - Approximately a third of policies in the New Industrial Policy Observatory (NIPO) have an associated subsidy value. - Positive correlation at country-product level in 2023 between count of IPs and the log of the subsidy value: 0.52, suggesting counts approximate size of IP values. - Reporting and prevalence notes: - Roughly 70-80% of protectionist domestic subsidies are classified as IPs (Rotunno and Ruta (2024) reference). - Among all GTA policies (IP and non-IP), approximately 80% are protectionist and 15% are liberalizing; more than 35% of policies are liberalizing for FDI measures, import and export barriers. - Around 10% of the IPs in the 2009-2022 period were removed. - Coverage and sample: - Juhász et al. (2023) exclude sub-national (e.g., province-level in China) policies. ### Data on Trade Competitiveness - Main metric: revealed comparative advantage (RCA), proxied by the Balassa index. - RCA definition provided: - RCA_cpt = (X_cpt / sum_p X_cpt) / (sum_c X_cpt / sum_cp X_cpt) (formula structure as in source). - Variables: c is country, p is product, t is year, X is export value. - Interpretation: RCA > 1 implies the country exports a relatively higher share of the product compared to the world average. - Alternative RCA: Vollrath (1991) measure accounting for both exports and imports used for robustness (Appendix F.4). - Export value data from CEPII BACI database (Gaulier and Zignago, 2010): bilateral trade flows for 233 countries and 5018 products at HS 6-digit level. - Merging RCA data with IP counts yields a balanced panel of 156 countries and 5018 products for 2009-2022. - For more details on the Balassa index, see https://unctadstat.unctad.org/EN/RcaRadar.html (reference in source). ### IP Motives - Motive assignment follows the machine learning algorithm in Evenett et al. (Forthcoming), trained using stated motives from NIPO (available since 2023); one IP can have more than one motive. - Focus on four most common IP motives: climate mitigation, strategic competitiveness, geopolitical concerns or national security, and GVC resilience. - Since 2023, GTA experts assign stated motives to policy interventions by collecting/reviewing official statements or direct quotes from senior officials. ### Classification of Green vs. Non-Green Products - Objective: distinguish products related to the green transition (“green products”) versus “non-green products”. - Final list of green products contains 869 HS 6-digit products, compiled from six sources covering raw materials, intermediate inputs, and final products critical for the green transition. - Sources and key features: - IMF Climate Change Dashboard - Low Carbon Technologies (LCT): based on Pigato et al. (2020); includes solar panels, wind turbines, energy-efficient lighting, HVAC systems, insulation materials, filters and scrubbers, components for deployment of renewable energy, and products contributing to sustainable environmental management and pollution control. - IMF Climate Dashboard - Environmental Goods: includes connected goods (directly serve environmental protection) and adapted goods (made more environmentally friendly). Starting point OECD/Eurostat (1999); 108 product codes were added over time mainly in adapted goods. - Note: some environmental goods may lack equivalent HS codes; some HS codes may include non-environmental goods. - Rosenow and Mealy (2024): HS 6-digit products in three green value chains: solar PV, wind turbines, and electric vehicles (EVs); classified into raw materials, processed materials, sub-components, and end products; used to analyze industrial upgrading along green value chains (Section 4.3.1). - Kowalski and Legendre (2023): critical raw materials used intensely in green transition technologies (li-ion batteries, fuel cells, wind energy, electric traction motors, photovoltaics); classification taken from Bobba et al. (2020); Annex B details products in the Inventory at HS2007 code. - Goldschlag et al. (2020): link CPC patent classification to HS product codes using machine learning; classify green patent codes as CPC code Y02; define HS codes as green if weight on Y02 > 50 percent. - Mealy and Teytelboym (2022): dataset replicated from WTO core list, APEC list of environmental goods, and OECD illustrative product list; products classified as green have endorsement by many WTO/APEC members or environmental benefits determined by OECD countries. ### Descriptive Statistics (preview from 2.2) - The paper provides descriptive statistics of protectionist IPs by stated motive, country income group, policy instrument, and GTA evaluation. IP refers to protectionist IP unless explicitly comparing by GTA evaluation. - Evolution of announced IP counts (2009-2022): - Clear surge in announced IPs since 2017. - Number of announced IPs increases by eight times between 2017 and 2022. - Approximately 60% of IPs can be assigned to at least one stated motive. - For IPs with non-missing motive, about 13% have more than one motive; such IPs are counted multiple times (once per motive). - Among IPs with stated motive(s) in 2018-2022: - Approximately half motivated by strategic competitiveness. - Climate mitigation: 34%. - GVC resilience: 16%. - Geopolitical concerns or national security: 1% (very uncommon). - AE vs. EM patterns (classification based on IMF’s World Economic Outlook): - Both AEs and EMDEs actively implemented IPs; pervasive use predates 2009 in some large EMDEs. - AEs: number of IPs rose from around 100 in 2017 to over 1000 in 2022. - EMDEs: added 350 interventions between 2017 and 2022. - Share of recently implemented active IPs by AEs rose since 2017. - IP motives are more diverse in AEs; in EMs strategic competitiveness represents 70% of IPs with an assigned motive. - Low share of IPs with geopolitical/national security motives in both AEs and EMs. - Uptick in IPs for GVC resilience since 2020 in EMs and AEs. - Aggregate policy share: - Before 2017, IPs accounted on average for less than 25 percent of the total count of policies in GTA. - This rose to over 35 percent during the 2017-2022 period. - Notes on figures (as described in source): - Figure 1: Evolution of announced IP counts by stated motive; y-axis counts announced IPs adjusted for reporting lags; an announced IP with n stated motives is counted n times. - Figure 2: Evolution by AE vs. EM; LICs omitted in figure because they account for 1% of announced IPs, but LICs are included in empirical sample. - Policy instruments: - GTA assigns each policy to one of 66 instruments; authors aggregate into eight groups per UN MAST classification: export barriers, import barriers, domestic subsidies, export incentives, FDI measures, public procurement measures, local content measures, and others. - Observations: - Close to 80% of domestic subsidies take the form of financial-related measures (state loan, financial grant, loan guarantee). - Production subsidy accounts for 2.24% of domestic subsidies IPs in the data. - Approximately 90% of export incentives IPs are financial-related (trade finance, financial assistance in foreign market). - Figure 3 (described): shows share of eight policy instruments out of total IPs by AEs and EMs in 2018- (figure cut-off in source). *Source: IMF Working Paper (sections 2.1–2.2 as provided).* ### 2022. Panel (a) counts each policy once, even if the policy targets multiple HS products. Panel (b) ### wpiea2025098-print-pdf - 2022. Panel (a) counts each policy once, even if the policy targets multiple HS products. Panel (b) ### Breakdown by policy instrument (2018–2022) - Panel (a) counts each policy once, even if the policy targets multiple HS products; Panel (b) counts each policy as many times as the number of targeted HS products. - Key empirical facts: - Domestic subsidies and export incentives are the most commonly used policy instruments in both AEs and EMs. - EMs use a more diverse set of policy instruments than AEs. - EMs tend to use more trade barriers, whereas AEs tend to use more local content measures. - Differences between panels are explained by differences in the average number of products each policy instrument targets (see Appendix D.3): - One export incentive in EMs targets on average almost 400 products, the highest number across all policy instruments and country income groups. - For AEs, one local content measure targets close to 200 products on average. - Table 1 (as presented in the source) shows policy-instrument-level treated-unit counts in the clean sample and appears as: - Export barriersImport barriersDomestic subsidiesExport incentivesLocal content 52255851464875191171 ### Breakdown by GTA evaluation (2018–2022) - GTA evaluation categories: red = protectionist, amber = ambiguous, green = liberalizing. - Panel (a) counts each policy once; Panel (b) counts each policy as many times as the number of targeted HS products. Results are highly consistent across panels. - Key findings: - The majority of IPs are protectionist: almost all IPs in AEs and close to 80% of IPs in EMs are protectionist. - EMs implement a larger share of liberalizing IPs: around 20%, compared to virtually zero in AEs. - Additional patterns (Appendix D.4): - A large share of policies classified as import barriers and FDI is liberalizing, particularly in EMs. - EMs conduct a higher share of liberalizing domestic subsidies than AEs. ### Empirical strategy (LP-DiD framework and sample construction) - Main econometric method: LP-DiD (local-projection difference-in-differences) to address bias from DiD-TWFE with staggered treatment and dynamic/heterogeneous effects. - Clean-sample construction (equation (1) in source): - Treatment: first-time IP (D_{c,p,t} = 1) with D_{c,p,t−j} = 0 for 1 ≤ j ≤ L. - Controls: clean controls require D_{c,p,t−j} = 0 for −H ≤ j ≤ L. - D_{c,p,t} is an indicator that the first difference of the number of active IPs (∆IP_{c,p,t} = IP_{c,p,t} − IP_{c,p,t−1}) is greater than zero. IP_{c,p,t} is the stock of active IPs in year t; ∆IP is the IP “shock”. - Parameter choices and robustness: - Baseline stabilization lag L = 5; robustness with L = 3 (Appendix F.2). - Horizon parameter H appears in the clean-sample definition; choice of L faces a bias–variance trade-off. - Baseline specification controls for two lags of the dependent variable following Chudik and Pesaran (2015); robustness tested with three lags. - Additional sample restrictions: - Focus on protectionist IPs (red GTA evaluation): restrict treated units to those treated only by red IPs and not by green or amber at t. - Final clean sample size: approximately 30,000 observations in the treatment group and close to 10 million observations in the control group. - Dependent variable construction and technical choices: - rc a_{c,p,t+h} = ln(RCA_{c,p,t+h} + 10^{−3}). The small constant 10^{−3} allows inclusion of zero RCA observations. - Choice of 10^{−3} justified by mean RCA = 1.29 in the sample; robustness also tested by adding 1 instead of 10^{−3} (Appendix F.5). - Approximately 42% of observations in the clean sample have zero RCA. - Specification (baseline dynamic equation, equation (2) in source) includes: - Treated_{c,p,t} indicator for IP shock at t. - Two lags of ln(RCA + 10^{−3}) and two lags of the non-IP shock (∆nonIP_{c,p,t}). - Fixed effects: country-product α_{c,p}, country-year δ_{c,t}, product-year ρ_{p,t}. - Instrument-specific clean-sample refinement (equation (3) in source) to study heterogeneous effects by policy instrument i ∈ [1,8]: - Treatment group for instrument i: first-time IP_i (D_{i,c,p,t} = 1), not treated by any other instrument at the same time (D_{−i,c,p,t} = 0), and not treated in preceding L periods by any instrument (D_{c,p,t−j} = 0 for 1 ≤ j ≤ L). - Control group remains the clean controls as in equation (1). - Focus for heterogeneous-instrument analysis: domestic subsidies and export incentives (highest number of treated units in the final clean sample). ### Results — All products - Average effect (Figure 5): - Products targeted by IPs experience a 5.6% higher increase in trade competitiveness than non-targeted products three years after the introduction of the IP. - Solid line in the figure is estimated percent change in RCA + 10^{−3}; dashed lines are 90% confidence intervals; standard errors clustered at the country-product level. - Decomposition into extensive vs intensive margins (Section 4.1.2; Figure 6): - Extensive margin: probability to start exporting conditional on not being an exporter in the previous two years. - Targeted products experience a higher probability to start exporting compared to non-targeted products in the medium run (three years after treatment). - Intensive margin: change in ln(RCA) conditional on already being an exporter. - Intensive-margin dynamics are statistically insignificant but inherit the shape of the average effect. - Estimation details: - Extensive-margin regression uses Export_{c,p,t+h} (indicator = 1 if export value > 0 at t+h) and restricts sample to Export_{c,p,t−1} = Export_{c,p,t−2} = 0. - Intensive-margin estimates follow the baseline dynamic specification replacing rca_{c,p,t+h} with ln(RCA_{c,p,t+h}) for −2 ≤ h ≤ 4. - Heterogeneity by product’s initial competitiveness (Section 4.1.3; Figure 7): - Regression includes interaction Treated × (RCA_{c,p,t−1} > 1) to capture differential effects by initial RCA. - Findings: - Previously competitive products (RCA_{c,p,t−1} > 1) experience a large short-term boost in competitiveness after the IP shock; the positive association peaks after two years and then declines and becomes statistically insignificant at four years, though the estimated magnitude remains as high as 9 percent. - Initially non-competitive products (RCA_{c,p,t−1} ≤ 1) show an initial decline in RCA and then a gradual, albeit statistically insignificant, increase over the horizon considered. - Interpretation offered in the source: - Initially uncompetitive products may face short-term adjustment costs to become globally competitive; improvements for such products may require country fundamentals (e.g., high human capital). - Targeting products with existing comparative advantage can yield more immediate results and lower risk of failures; targeting low-initial-RCA products may be justified if there are potential dynamic gains (e.g., green transition products). - Heterogeneity by policy instrument (Section 4.1.4; Figure 8): - Focus instruments (clean-sample treated-unit counts in final clean sample): - Domestic subsidies: 14658 treated units. - Export incentives: 7519 treated units. - Empirical findings: - Domestic subsidies are associated with a short-term 5 percent increase in competitiveness for targeted relative to non-targeted products, which fades over time (Panel (a)). - Export incentives yield an initial 1 percent decline in competitiveness, followed by longer-term improvements (Panel (b)). - Interpretation in the source: - Short-term boost from domestic subsidies may be attractive to policymakers with short horizons; export incentives appear to provide more sustained benefits over the medium to long term. - Export incentives encourage firms to compete in global markets, potentially yielding medium- to long-term gains; however, many export incentives are prohibited under WTO rules and may provoke retaliatory measures, which could undermine benefits. - The paper focuses empirically on domestic subsidies and export incentives because they have the highest numbers of treated units in the final clean sample; results for other policy instruments are reported in Appendix E.1. - Additional empirical notes: - Regression standard errors clustered at the country-product level. - Figures and results reference GTA (2022), Juhász et al. (2023), and the author’s calculations. *Italic: Source — wpiea2025098-print-pdf (content units and excerpts as provided).* ### Appendix E.2 presents results for initially competitive vs. initially uncompetitive products by policy ### Appendix E.2 presents results for initially competitive vs. initially uncompetitive products by policy instrument ### Initially competitive vs. initially uncompetitive products (by policy instrument) - Results by policy instrument are similar to the ones for overall IPs discussed in Section 4.1.3. - There is a positive association for initially competitive products in the short run and for initially uncompetitive products in the longer horizon. - Figure 8 (Domestic subsidies; Export incentives) shows estimated percent change in RCA+ 10⁻³ for domestic subsidies (Panel (a)) and export incentives (Panel (b)), with 90% confidence intervals and standard errors clustered at the country-product level. ### Green vs. Non-Green Products (Section 4.2) - Definition: Green products are 6-digit HS92 products that are critical to the green transition. Section 2.1.4 provides the detailed green products list. - Motivation: The current IP wave has a growing share of green IPs; two primary justifications for IPs targeting green products are: - the novelty of green technologies (low carbon technologies are new and compete with established ones); and - emission externalities (private benefits of LCTs are lower than social benefits, implying under-provision without policy). - Empirical findings: - IPs targeting green products increase RCA by about 20 percent after 4 years (green line in Figure 9). - IPs targeting non-green products are associated with only a mild short-term increase in RCA, with smaller and insignificant effects in the medium term. - Evidence suggests IPs targeting green products have a more prominent impact on competitiveness in the longer horizon than IPs targeting non-green products. - Role of initial competitiveness (Figure 10): - Previous findings on initial competitiveness in Section 4.1.3 are mainly driven by non-green products. - IPs are positively associated with long-run gains in RCA for green products that have not yet established comparative advantage in the global market (i.e., initially uncompetitive green products). - Policy instrument heterogeneity by product type (Figure 11): - The association between IP and RCA is generally more positive when targeting green products for both domestic subsidies and export incentives, particularly in the longer horizon. - Domestic subsidies: - For green products: insignificant in the short term and positive in the medium term. - For non-green products: small temporary improvement in RCA that turns negative in the medium term. - Export incentives: - Both green and non-green products experience boosts after IP treatment, but the effect for green products is more significant and pronounced. ### Cross-product spillovers along the green value chain (Section 4.3.1) - Motivation: Positive long-run association between IPs and RCA of green products motivates examining spillovers along the green value chain and effects on non-targeted products. - Data: Use dataset from Rosenow and Mealy (2024) that maps 6-digit HS codes to three major green value chains: wind turbines, solar panels, and electric vehicles. Products are assigned to one of four value chain stages: raw materials, processed materials, subcomponents, end products. No product is assigned to more than one green value chain. - Empirical specification: local projection framework estimating rca_{c,p,t+h} on downstream/upstream IP shocks (∆DownIP_{c,v(p),s(p),t}, ∆UpIP_{c,v(p),s(p),t}), lags of these shocks, lags of the dependent variable, non-IP shocks and their lags, with fixed effects (country-product, product-year, country-value chain-year). - Outcomes of interest are β_d and β_u, capturing effects of downstream (upstream) IPs relative to IPs targeting products within the same stage of the value chain. - Note: ∆DownIP + ∆UpIP + ∆OwnIP = ∆IP and ∆IP is absorbed by country-value chain-year fixed effect; hence estimates are relative to IPs targeting the same stage. - Findings (Figure 12): - IPs targeting more upstream products are associated with stronger improvements in RCA relative to those targeting products at the same value chain stage. - IPs targeting more downstream products have similar effects as IPs targeting products within the same value chain stage. - Interpretation: Upstream IPs may alleviate capacity constraints and benefit downstream products through reductions in input costs. - Instrument composition note: - Domestic subsidies account for the majority of IPs in all stages of production, and are more prevalent in the initial stages of the value chain (raw materials, over 70%) compared to more downstream stages (between 40 and 60%). Some differences in impact may be partially attributable to composition effects. ### Robustness checks (Section 4.3.2) - Seven alternative scenarios tested; main points: - Excluding China from the main sample (IPs in China not well represented by GTA) — results robust to exclusion of China. - Changing number of stabilization lags from 5 years to 3 years when constructing the clean sample — trade-off: smaller L increases number of "first-time IP" treatments (more inclusion) but increases bias. - Additionally controlling for third lag of ln(RCA+ 10⁻³) and non-IP shock. - Using alternative RCA measure that also accounts for imports. - Using ln(RCA+ 1) instead of ln(RCA+ 10⁻³) as main dependent variable. - Excluding units treated in 2020 due to concern IPs announced during 2020 may be Covid-specific. - Using all subsidies in the GTA database to address classification concerns. - Overall robustness: Results are generally robust to main findings. The only exception is export incentives, for which the medium-term positive effect is less pronounced in certain exercises, possibly due to relatively low number of export incentives IPs in the data. ### Conclusion (Section 5) — substantive findings and policy implications - Dataset: covers 156 countries and 5018 products in 2009-2022. - Main findings: - On average, a positive link between IPs and improvements in competitiveness of targeted products, with heterogeneous effects across products and policy instruments. - Product-level heterogeneity: - Positive link between IPs and a product’s RCA is mostly driven by products that were previously competitive. - Product characteristics such as relation to the green transition affect timing: green products experience larger medium-term improvements in RCA following IPs compared to non-green products. - Instrument heterogeneity: - Domestic subsidies associated with short-term improvements in competitiveness of targeted products. - Export incentives associated with medium-term improvements. - Suggestive evidence of cross-product spillovers along value chains. - Policy implications and cautions: - IPs should be handled with care; their nuanced effects point to limited use case. - Analysis provides a partial picture and does not fully account for general equilibrium effects, such as cross-sectoral reallocations, retaliatory measures by other countries, and fiscal costs. - Countries must carefully weigh costs and benefits of IPs in general equilibrium, ensure consistency of IPs with international rules, and prioritize multilateral policy cooperation. - IPs can entail significant fiscal costs, amplifying debt sustainability concerns. - Research agenda: - Incorporating general equilibrium channels in analysis of IPs is a fruitful avenue for future research. *Source: IMF Working Paper — Appendix E.2 and Sections 4.2–5 as provided in the supplied content.* ### References ### References ### Methodological Appendix — LLM Ensemble Approach for Classifying IP Motives - Overview: - Adopted pretrain-finetune paradigm leveraging RoBERTa-Large (Liu et al. (2019)) as base model. - Task: assign a stated motive to industrial policies (IPs) described in GTA and NIPO datasets; annotated subset from NIPO used as training/validation. - Motivation: LLMs provide contextual comprehension beyond bag-of-words; ensemble used to mitigate training instability (McCoy et al. (2020)). - Model architecture and training (Step 0 — Model Construction; Step 2 — Finetuning): - Base model: RoBERTa-Large generates 1,024-dimensional vector representations for each input token. - Classification head: fully connected layer (1,024×1,024), ReLU activation, dropout layer (dropout rate: 0.1), final classification layer (1,024×2), softmax output. - Full finetuning: all model weights (RoBERTa’s pre-trained weights and classification head) updated during training. - Model size referenced: 355 million parameters (RoBERTa-Large manageable on a single RTX 8000 GPU). - Optimizer and hyperparameters: - AdamW optimizer with learning rate of 1×10^−5. - Train for 8 epochs. - Weight decay of 0.01. - Linear learning rate scheduler with warm-up; warm-up proportion set to 10% of total training steps. - Loss function for classification: cross-entropy loss - L = − Σ_{i=1}^N w_i · y_i log(ˆy_i), with class weights w_i to account for label imbalance. - Ensemble technique: - RoBERTa-Large finetuned ten times; each run uses a randomly initialized classification head and randomized batch order. - For production classification, feed policy title and description to each of the ten finetuned models, obtain ten probability sets, compute weighted average of probabilities, and classify the policy as having the stated motive if the weighted probability exceeded 60%. - Text preprocessing (Step 1): - Remove all non-Unicode characters and redundant escape sequences; replace non-English characters with English counterparts when possible. - Truncate input text at 512 tokens (RoBERTa-Large maximum context length). - Train/test split: class-stratified 80% training sample, 20% testing sample; split performed once and fixed for ensemble process. - Ensemble production criteria: - Final classification threshold: weighted probability > 60%. ### Validation and Comparative Performance - Comparison approaches evaluated: - Bag-of-Words: TF-IDF + logistic regression. - In-Context Learning (Few-shot) with models: Llama3-8b-instruct, Qwen2-8b-instruct, GPT-3.5 Turbo. - Finetuning instruction-following LLMs using Low-Rank Adaptation (LoRA) and finetuning the instruction-following models (Llama3-8b-instruct, Qwen2-8b-instruct, GPT-3.5-Turbo). - RoBERTa-Large finetuning and RoBERTa-Large ensemble. - Evaluation metrics: - Accuracy = (TP + TN) / (TP + TN + FP + FN). - F1 = 2 × (Precision × Recall) / (Precision + Recall). - Macro-F1 = 1/2 (F1_positive class + F1_negative class). - Performance on motive: Climate Change Mitigation (Table B.1; Accuracy / Macro-F1): - TF-IDF+logistic: 0.88 / 0.81 - Llama3-8b-instruct (In-Context Learning): 0.87 / 0.75 - Qwen2-8b-instruct (In-Context Learning): 0.88 / 0.82 - GPT-3.5 Turbo (In-Context Learning): 0.87 / 0.81 - Llama3-8b-instruct (Finetuning): 0.91 / 0.85 - Qwen2-8b-instruct (Finetuning): 0.88 / 0.80 - GPT-3.5-Turbo (Finetuning): 0.94 / 0.90 - RoBERTa-Large (Finetuning): 0.94 / 0.90 - Ensemble RoBERTa-Large: 0.97 / 0.94 - Performance of LLM ensemble on other motives (Table B.2; Accuracy / Macro-F1): - Strategic Competitiveness: 0.92 / 0.90 - Geopolitical Concerns: 0.95 / 0.88 - GVC Resillence: 0.96 / 0.90 ### Appendix sections and robustness checks described in the source - Online Appendix A: Classifying IP Motives — describes the LLM ensemble approach in detail (Steps 0–3) and validation experiments. - Appendix A.1: Algorithm in detail — model construction, text preprocessing, finetuning, ensemble and production steps (includes specific model architecture and hyperparameters as above). - Appendix A.2: Validation — comparison with TF-IDF, in-context learning (Llama3-8b-instruct, Qwen2-8b-instruct, GPT-3.5 Turbo), and finetuning (LoRA, finetuned instruction-following LLMs); details for prompts and evaluation metrics. - Appendix B: Comparing IMF’s Environmental Goods and LCT Products — presents Figure B.2: Environmental Goods vs. LCT Products. - Appendix C: Classification of Policy Instruments — presents figures for Export Barriers, Import Barriers, Domestic Subsidies, Export Incentives, FDI Measures, Public Procurement Measures, Local Content Measures, Others; and Figure B.2: Composition of Broad Policy Instruments (2018-2022) with note: Sources: GTA (2022), Juhász et al. (2023), and author’s calculations. Notes include: y-axis represents the share of disaggregated policy instrument out of each broad group in 2018-2022, adjusted for reporting lags. - Appendix D: Additional Descriptive Statistics — includes Figures B.3–B.8 on evolution of announced IPs over time, by country income group (AE/EM/LIC per IMF’s World Economic Outlook), average number of targeted products by policy instrument (2018-2022), and breakdowns by GTA evaluation (protectionist/ambiguous/liberalizing). - Appendix E: Additional Results — includes Figures B.9–B.10 on effects of IP on ln(RCA + 10^{−3}) for other policy instruments, and the role of initial RCA for Domestic Subsidies vs. Export Incentives (distinguishing products with RCA_{c,p,t−1} > 1 and RCA_{c,p,t−1} ≤ 1). - Appendix F: Robustness checks — extensive robustness analyses and alternative specifications, including: - F.1 Excluding China (Figures B.11–B.14): effects on ln(RCA + 10^{−3}), by initial RCA, green vs. non-green products, policy instrument breakdowns. - F.2 Alternative number of stabilization lags (L = 3) (Figures B.15–B.18). - F.3 Controlling for the third lag (Figures B.19–B.22): regressions additionally control for the third lag of ln(RCA + 10^{−3}) and ΔNonIP. - F.4 Alternative RCA measure adjusted for imports (Figures B.23–B.25): RCA measure adjusted as ln(RCA_exports + 10^{−3}) − ln(RCA_imports + 10^{−3}). - F.5 Using ln(RCA + 1) as dependent variable (Figures B.26–B.29). - F.6 Excluding Covid-19 period (Figures B.30–B.33): clean sample excludes units treated in 2020. - F.7 All subsidies in GTA (Figures B.34): effect of all GTA subsidies on RCA (and RCA adjusted for imports); clean sample excludes units treated in 2020. - Notes repeated across figures: - Sources: GTA (2022), Juhász et al. (2023), and author’s calculations. - Standard errors are clustered at the country-product level. - Confidence bands shown are 90% confidence intervals where specified. - The list of green products is described in Section 2.1.4 (referenced in figure notes). *Source: wpiea2025098-print-pdf - References (IMF Working Paper content provided in the source PDF).* --- _Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025098-print-pdf.pdf_