## Annexes 1–3. Data Sources; Empirical Frameworks; Additional Results

## Source details

**Canonical URL:** [Annexes 1–3. Data Sources; Empirical Frameworks; Additional Results](https://www.imf.org/-/media/files/publications/sdn/2025/english/sdnea2025002-s001.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/sdn/2025/english/sdnea2025002-s001.pdf.md)
- [Structured JSON version](/-/media/files/publications/sdn/2025/english/sdnea2025002-s001.pdf.json)

---

### Annex 1 — Data Sources and Sample Coverage
- Industrial Policies Data
  - Source: Juhász and others (2023) using machine learning on Global Trade Alert (GTA) descriptions.
  - Definition: IP = “…goal-oriented state action. The purpose is to shape the composition of economic activity.”
  - Scope and coverage:
    - Approximately 18,000 IPs covering 109 countries during 2009–22.
    - European Union–level policies allocated to all member countries.
    - Data mergeable to original GTA to extract targeted products/sectors, instruments, and GTA expert protectionist evaluation.
  - Instruments identified (seven categories): Export barriers; Import barriers; Domestic subsidies; Export incentives; Restrictions on foreign direct investment; Procurement policies; Local content requirements.
  - SDN focus: protective domestic subsidies, protective export incentives, and policies that lift import barriers (liberalizing import policies).
  - Aggregation: product/sector × country × year.
  - Time coverage: 2009–22. Country coverage: 109 countries (varies by exercise).

- Revealed Comparative Advantage (RCA) Data
  - Source: CEPII BACI (Gaulier and Zignago 2010 construction).
  - Aggregation: HS6 product code–country level.
  - RCA definition: export share of product in country’s total exports divided by world export share of product; RCA > 1 indicates global competitiveness.
  - Time coverage: 1995–2022; analysis focuses on 2009–22.
  - Country coverage: all countries with available trade and IP data (100+ countries).

- Firm-Level Data
  - Source: Bureau van Dijk (BvD) Orbis global database.
  - Variables: total factor productivity, labor compensation, age, total assets, sales, operating revenue (gross output), tangible fixed assets, liabilities, earnings before interest and taxes, cash flow.
  - Transformations: all data in constant 2010 US dollars; TFP computed following Ackerberg, Caves, and Frazer (2015).
  - Time coverage for econometric analysis: 2009–21 (driven by IP data).
  - Country coverage for empirical analysis: 42 countries (AEs and EMEs) with available TFP data.

- Patent Data (INPACT-S)
  - Source: INPACT-S database (Labelle and others 2024), built from Worldwide Patent Statistical Database global autumn 2021 edition.
  - Construction: patent applications by country of origin to application authority at IPC 4-digit for 1980–2019; transforms IPC → ISIC Rev 3 (2-digit) via Lybbert and Zolas (2022).
  - Dimensions: 91 patent authorities, 213 countries of origin, 40 years, 31 ISIC codes.
  - Time coverage (econometric analysis): 1980–2019; analysis focuses on 2009–19.
  - Country coverage: countries with available IP and patent information for the period.

### Annex 2 — Empirical Frameworks (Overview)
- Baseline LP-DiD for IPs and product competitiveness
  - Method: local projection difference-in-differences (LP-DiD) following Dube and others (2024).
  - Horizons: h ∈ [0, 4].
  - Dependent variable: y_{c,p,t+h} = ln(RCA_{c,p,t+h} + 10^{-3}). Baseline chosen because mean RCA in sample = 1.29.
  - Main regressor: IP treatment dummy TREAT_{c,p,t}; coefficient β_h.
  - Controls/FEs: up to two lags of dependent and non-IP shock (Chudik and Pesaran (2015) rule-of-thumb T^{1/3}=14^{1/3}≈2.41), Δnon-IP_{c,p,t}, α_{c,p}, δ_{c,t}, ρ_{p,t}.
  - Standard errors clustered at country-product level.
  - Clean sample selection: first-time IP clean treatment D_{c,p,t}=1 and D_{c,p,t−j}=0 for 1 ≤ j ≤ L; clean controls D_{c,p,t−j}=0 for −h ≤ j ≤ L. Baseline L = 5; robustness L = 3.

- Heterogeneity: initially competitive products
  - Indicator: lagged RCA > 1.
  - Regression includes interaction TREAT × 1{RCA_{c,p,t−1} > 1}.
  - Interpretation: β_h = effect for initially uncompetitive products; β_h + γ_h = effect for initially competitive products.

- Heterogeneity: policy instruments
  - Two approaches: instrument-specific clean samples (restrict to first-time IP of instrument i and not treated by others) and separate regressions by instrument i with TREAT_{i,c,p,t}.

- Green value chain analysis
  - Data: Rosenow and Mealy (2024) mapping HS6 → three green value chains (wind turbines, solar panels, electric vehicles); stages: raw materials; processed materials; subcomponents; end products.
  - Method: OLS local projections.
  - Key regressors: ΔDownstream_{c,v(p),s(p),t} and ΔUpstream_{c,v(p),s(p),t}; controls include non-green IP shock, non-IP shock, two lags, and two lags of dependent variable.
  - Fixed effects: country-product, country-year, product-year, value chain-year, country-value chain.
  - Coefficients of interest: β_d (downstream effect), β_u (upstream effect).

- Firm-level local projections (baseline)
  - Outcomes Y_{f,t}: payroll, capital, TFP quantity, value added.
  - Main variables: ΔnP_{i,c,t}^{k} and nP_{i,c,t}^{k} by instrument k ∈ {subsidies, export incentives, trade barriers, local-content requirements, other} and GTA evaluation T ∈ {protectionist (red), likely protectionist (amber), liberalizing (green)}.
  - Controls/FEs: α_f, α_{c,t}, α_{i,t}, two lags of dependent and independent variables, X_{i,c,t}.
  - SEs clustered at country-industry level.
  - Spillovers: upstream exposure UPT_{s,c,t} and downstream exposure DNT_{s,c,t} built using GTAP input-output shares i_{c,s'→s} and counts of protectionist IPs in other sectors.

- Heterogeneity at firm and country levels
  - Firm heterogeneity: interact ΔIPs with firm characteristic terciles Q_q (age, cash flow/assets) to estimate β_h^{q,k}.
  - Country heterogeneity: interact ΔIPs with country characteristic Z_{c,t−1} (trade openness, credit market conditions, business environment, average schooling); impacts evaluated at 10th/25th and 75th/90th percentiles; CIs via delta method.

- Threats to identification and robustness (firm-level)
  - Baseline mitigations: rich FEs, lags, pre-trend visual checks for three years prior.
  - Alternative IP intensity index: nP_{i,c,t}^{Giedk} = Σ_{product q ∈ i_c} (TREAT_{q,i_c} / TREAT_{i_c}) · 1{policy U affects product q}.
  - Complementary methods:
    1. LP-DiD at firm level with clean treatment/control.
    2. Two-stage least squares IV: instrument ΔnP_{i,c,t}^{k} with ΔnP_{i',c',t}^{k'} in other countries weighted by bilateral trade shares nV_{i,c,t}^{k} = Σ_{c'≠c} (TotalTrade_{c,c',t} / Σ_{c'≠c} TotalTrade_{c,c',t}) · ΔnP_{i',c',t}^{k'}.
  - Preferred instruments: protectionist subsidies instrumented by protectionist subsidies in other countries/industries; liberalizing trade barriers instrumented similarly; protectionist export incentives instrumented by total protectionist IPs in other countries/industries.
  - Additional robustness checks: excluding China; adding firm-level controls (reduces country coverage); dropping firms with abnormal growth; removing Spain, Italy, France; reweighting by inverse firm counts and firm size; controlling for three lags.

- Industry-level frameworks
  - Baseline parallels firm-level with outcomes at country-industry level: log Y_{i,c,t+h} − log Y_{i,c,t−1} on ΔnP_{i,c,t}^{k,GTA} with analogous controls and SE clustering at country-industry.
  - Industry heterogeneity: median markups and median external financial dependence (EFD) computed from ORBIS; interactions with IP changes to estimate γ_h^{k,EFD} and γ_h^{k,Markup}.
  - Calculation of impact in high-distortion industries: β_h^{k} · nP( p75 − p25 ) + γ_h^{k,EFD} · nP( p75 − p25 ) · FFD_{p75} + γ_h^{k,Markup} · nP( p75 − p25 ) · MKT_{p75}. CIs via delta method; clustering by country and industry.

- Industry-level patenting: baseline, climate-application, and IV strategy
  - Dependent transform: log(10^{-6} + PAT_{c,s,t+h}) − log(10^{-6} + PAT_{c,s,t−1}).
  - Main regressors: ΔLiberalizing_{c,s,t}, ΔProtectionist_{c,s,t}, ΔAnnounced_{c,s,t}, ΔN_{non-IP,c,p,t}.
  - FEs: country-sector λ_{c,s}, country-year λ_{c,t}, sector-year λ_{s,t}. SEs clustered at country level.
  - IV strategy: instrument ΔProtectionist_{c,s,t} with changes in protectionist IPs in other sectors/countries weighted by political distance (UN vote patterns); first-stage F-statistic > 10 for all horizons with country-sector clustered SEs and for all horizons except 4 and 5 with country clustered SEs; Hansen J-test supports exogeneity.
  - Liberalizing IPs instrumented by (1) changes in liberalizing IPs in other sectors/countries weighted by political distance and (2) weighted by dummy identifying main trade partners (>90 percent of country’s trade). First-stage F-statistics > 10 for all horizons under both clustering schemes; Hansen J-test supports exogeneity.
  - Pre-trend checks: horizons −3 to −1.

- Comparing IPs and structural reforms
  - Approach: Rajan and Zingales (1998) style contrasts, interacting structural policies and IPs with industry distortion measures (FFD, MARKUP).
  - Regression includes ψ_{FD,SP}, ψ_{Markup,SP}, ψ_{FD,IP}, ψ_{Markup,IP} interaction terms; comparison plots compute policy impacts over difference between 75th and 25th percentiles. CIs via delta method; SEs clustered at country-industry.

### Annex 3 — Additional Results (Selected Robustness and Findings)
- Robustness exercises — Industrial Policies and Export Competitiveness
  - Assessing impact of all GTA subsidies (IP-classified or not): results comparable to subsidies classified as IPs.
  - Using alternative RCA measures (Vollrath 1991 controlling for imports): results robust.

- Firm-level robustness highlights
  - IP trade intensity index (restricting to IPs with non-missing HS codes): within this sample, baseline measure and IP trade intensity index yield similar results; across broader sample, relationships between instruments and firms’ value added, capital, and TFP broadly the same.
  - IV strategy (two-stage least squares):
    - Strong first stages for average firm analysis.
    - Domestic subsidies on capital by firm financial constraints: first-stage F-statistic = 8.6.
    - Export incentives on TFP by firm age: first-stage F-statistic = 3.7.
    - Comparison to baseline: value added, TFP, and capital responses to protectionist export incentives and liberalizing trade restrictions broadly consistent; protectionist subsidies lose significance in IV specification (coefficients for capital remain positive); firm heterogeneity coefficients directionally aligned but larger magnitudes.
  - LP-DiD at firm level (clean treatment/control):
    - Average firm responses broadly consistent with baseline; domestic subsidies on capital become positive but insignificant.
    - Firm heterogeneity: domestic subsidies on capital by financial constraints aligned with baseline; export incentives now increase TFP of both young and old firms similarly, with evidence that targeted young firms underperformed targeted old firms (stronger relative impact on young firms).
  - Sensitivity checks (excluding China; removing Spain/Italy/France; winsorizing at 1 and 99 percentiles): results broadly in line with baseline.
  - Adding firm-level controls, more lags, and removing input-output exposure controls: results broadly in line with baseline (sample reduced when adding firm-level controls).

- Sectoral-level robustness highlights
  - IV strategy: all first stages strong; domestic subsidies on value added positive but significant only at last horizon; export incentives on value added positive but insignificant; export incentives on advanced economies (AEs) positive at larger magnitude but insignificant starting horizon 1.
  - LP-DiD strategy: domestic subsidies on VA negative upon impact and positive in medium term; export incentives on VA similar shape as baseline but less significant; export incentives positive impact on AE materializes in first horizon then goes to zero.

- Economic effect of industrial policies on patent filing
  - Decomposition: subsidies and export incentives drive positive effect of protectionist IPs on patent filing by foreign inventors.
  - Quantitative effects:
    - An additional protectionist subsidy is associated with a 2.0 percent increase in the number of received foreign patent applications in the first year but does not last beyond the second year.
    - Export incentives associated with a temporary effect of the same size that materializes in the second year.
  - Robustness: results robust to IV specifications using instruments described above.

*Source: sdnea2025002-s001 - Annex 2. Empirical Frameworks (IMF Staff Discussion Note Online Annex).*

### Annex 1. Data Sources and Sample Coverage

### Annex 1. Data Sources and Sample Coverage

### Industrial Policies Data
- Source: Juhász and others (2023), applying machine learning to descriptions in the Global Trade Alert (GTA) project to build a cross-country database of industrial policies (IPs).
- Definition used: IP is defined as “...goal-oriented state action. The purpose is to shape the composition of economic activity.”
- Scope and coverage:
  - Approximately 18,000 IPs covering 109 countries during 2009–22.
  - European Union–level policies are allocated to all member countries.
  - Data can be merged to the original GTA database to extract policy-related characteristics such as targeted products and sectors, instruments used, and whether the policy is protectionist as evaluated by GTA experts.
- Instruments identified (seven broad categories):
  - Export barriers (for example, export bans and quotas)
  - Import barriers (for example tariffs and import licensing)
  - Domestic subsidies (such as state loans, loan guarantees, and production subsidies)
  - Export incentives (such as tax-based incentives and trade financing)
  - Restrictions on foreign direct investment (for example, ownership requirements and foreign direct investment screening decisions)
  - Procurement policies
  - Local content requirements
- Focus in the Staff Discussion Note (SDN): mostly on three instruments that account for the bulk of IPs—protective domestic subsidies, protective export incentives, and policies that lift import barriers (liberalizing import policies).
- Data aggregation for analysis: typically aggregated at the product/sector, country, and year level.
- Time Coverage: 2009–22.
- Country Coverage: 109 countries (number used in each exercise varies with availability of other data sources).

### Revealed Comparative Advantage (RCA) Data
- Source: CEPII BACI (see Gaulier and Zignago 2010 for dataset construction details).
- Construction and definition:
  - Database reconciles data reported by almost 150 countries to the United Nations Statistics Division, collated via Comtrade; reconciles reporting when both exporter and importer report the flow.
  - Aggregated at the HS6 product code–country level.
  - Product RCA for a given country defined as the export share of the product in the country’s total exports, divided by the world’s export share of the product. A value greater than 1 indicates the country is a globally competitive producer.
- Time Coverage: data available for 1995–2022; analysis focuses on 2009–22.
- Country Coverage: covers all countries with available trade and IP data (100+ countries).

### Firm-Level Data
- Source: Bureau van Dijk (BvD) Orbis global database.
- Notes on Orbis:
  - Largest cross-country firm-level database; reports data for all industries and for both private and public firms.
  - Collects and harmonizes data from various sources (in particular, publicly available national company registries).
  - Coverage of firms varies by country, industry, over time, and across variables due to differing national filing requirements.
- Dataset construction and variables:
  - Cleaning steps follow Kalemli-Özcan and others (2015) and Gopinath and others (2017).
  - Variables used: total factor productivity, labor compensation, age, total assets, sales, operating revenue (gross output), tangible fixed assets, liabilities, earnings before interest and taxes, and cash flow.
  - All data transformed into constant 2010 US dollars.
  - Total factor productivity computed following Ackerberg, Caves, and Frazer (2015).
- Time Coverage (econometric analysis): 2009–21 (driven by IP data coverage).
- Country Coverage (empirical analysis): 42 countries (advanced economies and emerging market economies) with available TFP data.

### Patent Data Used in Cross-Country Analysis
- Source: International Patent and Citations across Sectors (INPACT-S) database (see Labelle and others 2024), built from Worldwide Patent Statistical Database global autumn 2021 edition.
- Construction and mapping:
  - Compute number of patent applications from a country of origin (residence of the inventor or owner of the technology) to an application authority at the International Patent Classification (IPC) 4-digit level for 1980–2019.
  - Account for both the applicant and the inventor.
  - Use concordance tables developed by Lybbert and Zolas (2022) to transform IPC codes into industry codes—ISIC Rev 3, at 2-digit.
- Resulting database dimensions:
  - 91 patent authorities, 213 countries of origin, 40 years, and 31 ISIC codes.
- Advantages of INPACT-S highlighted:
  - More comprehensive view of global patent activity by encompassing a wider array of patent authorities.
  - Incorporates consistently constructed data on both cross-border and domestic patents.
- Time Coverage (econometric analysis): INPACT-S spans 1980–2019; analysis focuses on 2009–19 (driven by IP data coverage).
- Country Coverage: empirical analysis focuses on all countries with available IP and patent information for the period of analysis.

*Annex 1. Data Sources and Sample Coverage — Staff Discussion Notes Online Annexes for Industrial Policies: Handle with Care.*

### Annex 2. Empirical Frameworks

### Annex 2. Empirical Frameworks

### Baseline empirical specification: Industrial Policies and the Competitiveness of Targeted Products
- Method: local projection difference-in-differences (LP-DiD) following Dube and others (2024); used to estimate dynamic effect of industrial policy (IP) treatment on revealed comparative advantage (RCA) for country c and product p over horizons h ∈ [0, 4].
- Dependent variable: y_{c,p,t+h} = ln(RCA_{c,p,t+h} + 10^{-3}); 10^{-3} added to include zero-RCA country-product pairs. Baseline chosen because mean value of RCA in the sample is very small, equals to 1.29.
- Main independent variable: IP treatment dummy TREAT_{c,p,t}; coefficient of interest β_h.
- Controls and fixed effects:
  - up to two lags of the dependent variable and of the non-IP shock (choice follows Chudik and Pesaran (2015) rule-of-thumb T^{1/3}=14^{1/3}≈2.41),
  - non-IP shock Δnon-IP_{c,p,t} (first difference of stock of non-IP GTA policies),
  - country-product fixed effect α_{c,p},
  - country-year fixed effect δ_{c,t},
  - product-year fixed effect ρ_{p,t}.
- Standard errors clustered at the country-product level.
- Clean sample selection for LP-DiD:
  - clean treatment: first-time IP, D_{c,p,t} = 1 and D_{c,p,t−j} = 0 for 1 ≤ j ≤ L;
  - clean controls: D_{c,p,t−j} = 0 for −h ≤ j ≤ L.
- Stabilization lag L: baseline L = 5; robustness test for L = 3.
- Motivation: LP-DiD addresses bias from staggered treatment timings and previously treated units contaminating control groups.

### Heterogeneity: Targeted products
- Definition: initially competitive products have lagged RCA > 1; interpretation: country exports relatively higher share of product compared to world.
- Regression includes interaction of TREAT with indicator (RCA_{c,p,t−1} > 1).
- Interpretation of coefficients:
  - β_h: estimated effect for initially uncompetitive products,
  - β_h + γ_h: estimated effect for initially competitive products.
- Same controls and FEs as baseline.

### Heterogeneity: Policy instruments
- Two modifications:
  1. Instrument-specific clean sample i:
     - treatment group restricted to first-time IP i (D_{i,c,p,t} = 1) up to L preceding periods and not treated by any other instruments −i at the same time (D_{−i,c,p,t} = 0).
     - clean controls as in baseline.
  2. Regression estimated separately for each policy instrument i; main independent variable becomes TREAT_{i,c,p,t} (positive IP shock under instrument i).
- Purpose: isolate effects of distinct IP instruments.

### Green value chain
- Data: Rosenow and Mealy (2024) product mapping at six-digit HS code to three green value chains: wind turbines, solar panels, electric vehicles.
- Value chain stages: raw materials; processed materials; subcomponents; end products. No product assigned to more than one green value chain.
- Method: ordinary least squares local projection framework.
- Key variables:
  - ΔDownstream_{c,v(p),s(p),t}: first difference of active downstream green IPs for the value chain v(p) and stage s(p) of product p (captures downstream green IP shock).
  - ΔUpstream_{c,v(p),s(p),t}: upstream green IP shock.
  - Controls: non-green IP shock, non-IP shock, two lags of these shocks, two lags of dependent variable.
  - Fixed effects (FFFE): country-product, country-year, product-year, value chain-year, country-value chain.
- Outcomes of interest: β_d (effect of downstream green IPs relative to green IPs targeting products within same stage), β_u (effect of upstream green IPs).

### Industrial policies and firm-level economic performance (baseline)
- Method: local projection (Jordà 2005) tracking outcomes for firm f over horizons (T = {2011,...,2018}).
- Outcomes Y_{f,t}: payroll, capital, productivity measured as total factor productivity quantity, and value added.
- Main independent variables:
  - ΔnP_{i,c,t}^{k} (change in stock) and nP_{i,c,t}^{k} (stock) for IP instrument k ∈ {subsidies, export incentives, trade barriers, local-content requirements, other instruments} and GTA evaluation T ∈ {protectionist (red), likely protectionist (amber), liberalizing (green)}.
- Controls:
  - firm fixed effect α_f,
  - country-year α_{c,t},
  - industry-year α_{i,t},
  - two lags of dependent and independent variables (Chudik and Pesaran guidance),
  - X_{i,c,t} additional controls.
- Standard errors clustered at the country-industry level.
- Spillovers: constructed upstream exposure UPT_{s,c,t} and downstream exposure DNT_{s,c,t} using input-output shares i_{c,s'→s} from GTAP and counts of protectionist IPs in other sectors.

### Firm heterogeneity
- Interaction of IP changes with firm characteristics Z_{f,t−1} (age, cash flow to assets ratio) terciles Q_q to estimate heterogeneous β_h^{q,k}.
- Specification includes same lags, controls, and fixed effects as baseline.
- Coefficients β_h^{q,k} capture dynamic relationship between IPs and firm outcomes by tercile.

### Country heterogeneity
- Interact IP changes with country characteristic Z_{c,t−1} (Budina and others (2023) indicators: trade openness, credit market conditions, business environment; or average schooling year from UN HDI).
- Coefficients of interest: β_h^{k} Z̄ + γ_h^{k} measure impact of an additional IP in average firm for country characteristic level Z̄.
- SDN evaluates impacts at 10th/25th and 75th/90th percentiles of Z_{c,t} distribution.
- Confidence intervals calculated using the delta method.

### Threats to identification and robustness exercises (firm-level)
- Baseline inclusion of rich FEs and lags addresses omitted variables and past dynamics; pre-trend checks visually inspected for three years prior to policy to detect catching-up.
- Main independent variable counts number of IPs targeting each industry; robustness constructs IP intensity index measuring share of trade in industry affected by IPs:
  - nP_{i,c,t}^{Giedk} = Σ_{product q ∈ i_c} (TREAT_{q,i_c} / TREAT_{i_c}) · 1{policy U affects product q}
- Complementary methods implemented:
  1. LP-DiD at firm level: restrict sample to clean treatment (first treated in three years) and clean controls (never directly exposed) to avoid contamination.
  2. Two-stage least squares IV: instrument change in IPs of type k implemented in country c and industry i with IPs of type k' implemented by trade partners c' ≠ c in other industries i' ≠ i, weighted by bilateral trade shares:
     - nV_{i,c,t}^{k} = Σ_{c'≠c} (TotalTrade_{c,c',t} / Σ_{c'≠c} TotalTrade_{c,c',t}) · ΔnP_{i',c',t}^{k'}
- Preferred instruments:
  - protectionist subsidies instrumented by protectionist subsidies in other countries and industries,
  - liberalizing trade barriers instrumented by liberalizing trade barriers in other countries and industries,
  - protectionist export incentives instrumented by total protectionist IPs in other countries and industries.
- Additional robustness checks:
  1. excluding China,
  2. adding firm-level controls (reduces country coverage),
  3. dropping firms with abnormal growth in each horizon,
  4. removing countries over-represented in Orbis: Spain, Italy, France,
  5. reweighting regressions by inverse of number of firms in each country and by firm size,
  6. controlling for three lags of dependent and independent variables.

### Industry-level economic performance
- Baseline regression parallels firm-level specification, outcomes at industry i in country c:
  - log Y_{i,c,t+h} − log Y_{i,c,t−1} regressed on ΔnP_{i,c,t}^{k,GTA} and dynamics.
- Main outcomes: industry value added, productivity, capital, payroll, allocative efficiency (payroll and allocative efficiency following IMF (2024)).
- Controls and FEs: analogous to firm-level; standard errors clustered at country-industry level.
- Main coefficients β_h^{k,kv} capture dynamic correlations between industry performance and IP types.

### Industry heterogeneity
- Steps:
  1. Calculate median markups and median external financial dependence (EFD) across firms in each country-industry in ORBIS; compute long-term averages to smooth yearly fluctuations.
  2. Regress medians on industry and country FEs; industry FEs are the measures used for heterogeneity.
  3. Regress industry outcomes on IP changes interacted with industry distortion measures (median EFD, median markups).
- Regression includes additional interaction terms γ_h^{k,EFD} and γ_h^{k,Markup} capturing differential effects in distorted industries.
- Impact of IPs in high-distortion industries computed as:
  - β_h^{k} · nP( p75 − p25 ) + γ_h^{k,EFD} · nP( p75 − p25 ) · FFD_{p75} + γ_h^{k,Markup} · nP( p75 − p25 ) · MKT_{p75}
  - where nP( p75 − p25 ) is difference between 75th and 25th percentiles of IP distribution; FFD_{p75} and MKT_{p75} denote 75th percentiles of EFD and markups respectively.
- Robustness: comparisons using standard deviation of policy variables; confidence intervals via delta method; clustering by country and industry.

### Threats to identification and robustness exercises (industry-level)
- Same strategies as firm-level: rich FEs and lags; LP-DiD and LP-IV robustness checks; excluding China; alternative variable using all GTA subsidies.
- Note: subnational or high-frequency financial designs used in some studies are unfeasible given SDN data features.

### Industry-level patenting: baseline, climate-application, and IV strategy
- Baseline specification (local projections, Jordà 2005) for patent stock P_{c,s,t} aggregated after 1980; distinguishes patents by local vs foreign inventors.
- Dependent transform: log(10^{-6} + PAT_{c,s,t+h}) − log(10^{-6} + PAT_{c,s,t−1}) to handle zeros.
- Main independent variables:
  - ΔLiberalizing_{c,s,t} (change in liberalizing industrial policies),
  - ΔProtectionist_{c,s,t} (change in protectionist IPs),
  - ΔAnnounced_{c,s,t} (announced industrial policies),
  - ΔNon-IP trade policies ΔN_{non-IP,c,p,t}.
- Fixed effects: country-sector λ_{c,s}, country-year λ_{c,t}, sector-year λ_{s,t}. Standard errors clustered at country level.
- Coefficients of interest:
  - β_h^{2}: effect of additional protectionist IP on patent filing,
  - β_h^{1}: effect of liberalizing additional IP.
- Pre-trend checks for horizons −3 to −1 to assess endogeneity from targeting/timing.

- Application to climate-related protectionist IPs:
  - Regression distinguishes climate-related and non-climate-related IPs for both liberalizing and protectionist categories.
  - Coefficients of interest: β_h^{2,g} and β_h^{2,ng} for protectionist climate-related and non-climate-related IPs respectively.
  - Pre-trends (horizons −1 to −3) are not significantly different from zero at the 90 percent significance level for analyzed specifications.

- Instrumental variable (IV) strategy for patenting analysis:
  - Instruments change in protectionist IPs with changes in protectionist IPs in other sectors and countries weighted by political distance (UN vote patterns).
  - First-stage specification reported; instrument relevance: associated F-statistic above 10 for all horizons with country-sector clustered SEs, and for all horizons except 4 and 5 when using country clustered SEs.
  - Exogeneity supported by Hansen J-test.
  - Liberalizing industrial policies instrumented by two instruments: (1) changes in liberalizing IPs in other sectors and countries weighted by political distance; (2) changes in liberalizing IPs in other sectors and countries weighted by dummy identifying country’s main trade partners (those representing more than 90 percent of country’s trade).
  - First-stage F-statistics above 10 for all horizons under country-sector and country-clustered SEs; Hansen J-test supports exogeneity.
  - Rationale: tit-for-tat dynamics and political distance create exogenous variation in other countries’ sectoral IPs.

### Comparing IPs and structural reforms
- Approach follows Rajan and Zingales (1998) contrasting IPs and structural policies across industries with different levels of distortions.
- Regression includes interaction terms:
  - ψ_{FD,SP} · FFD_{i} · FFD_{c,t−1} (financial development structural policies),
  - ψ_{Markup,SP} · MARKUP_{i} · DOV_{c,t−1} (governance),
  - ψ_{FD,IP} · FFD_{i} · ΔnP_{i,c,t}^{R_kd} and ψ_{Markup,IP} · MARKUP_{i} · ΔIP_{i,c,t}^{R_kd} (IPs interacted with industry distortions),
  - plus dynamic IP terms and usual lags and controls.
- Comparison metrics:
  - plot ψ_{FD,SP}·FFD·(p75−p25)·FFD_{(p75−p25)} versus ψ_{FD,IP}·FFD·IP(p75−p25),
  - plot ψ_{Markup,SP}·MARKUP·(p75−p25)·DOV_{(p75−p25)} versus ψ_{Markup,IP}·MARKUP·nP(p75−p25).
- Coefficient intervals via delta method; standard errors clustered at country-industry level.

*Source: sdnea2025002-s001 - Annex 2. Empirical Frameworks (IMF Staff Discussion Note Online Annex).*

### Annex 3. Additional Results

### Annex 3. Additional Results

### Table 3.1. Robustness Exercises — Industrial Policies and Export Competitiveness
- Assessing the impact of all Global Trade Alert (GTA) subsidies
  - Description: Run a regression that tracks the impact on an additional subsidy (whether it is classified as industrial policy [IP] or not) on value added.
  - Additional details: —
  - Comparison to Baseline: Results are comparable to those found for subsidies classified as IPs.

- Using alternative RCA measures
  - Description: Run regressions using a revealed comparative advantage (RCA) measure that controls for imports (Vollrath 1991).
  - Additional details: —
  - Comparison to Baseline: Results are robust to the use of this alternative RCA measure.

### Industrial Policies at the Firm level
- IP trade intensity index
  - Description: Construct IP trade intensity index based on the share of each industry’s total trade that is affected by IPs. This robustness check restricts the analysis to IPs with non-missing Harmonized System (HS) codes.
  - Comparison to Baseline:
    - Within the sample of IPs with non-missing HS codes, using baseline measure or the IP trade intensity index does not significantly change the results.
    - Within the broader sample, the relationship between each instrument and firms’ value added, capital, and total factor productivity (TFP) is broadly the same.

- IV strategy (two-stage least squares)
  - Description: Instrument the count of IPs in a country-sector with the count of IPs that trading partners implemented in other industries.
  - Additional details:
    - There are strong first stages for average firm analysis.
    - Domestic subsidies on capital by firm financial constraints has a first-stage F-statistic of 8.6.
    - Export incentives on TFP by firm age has first-stage F-statistic of 3.7.
  - Comparison to Baseline:
    - Results for average firm: The behavior of value added, TFP, and capital in response to protectionist export incentives and liberalizing trade restrictions is broadly consistent with the baseline results. Protectionist subsidies lose significance on all outcomes at all horizons in the IV specification, although coefficients for capital are still positive.
    - Firm heterogeneity: The direction of coefficients is broadly aligned with Baseline although with much larger magnitudes.

- Local projections difference-in-differences (LP-DiD) strategy
  - Description: Restrict sample to clean treatment (firms that received IPs for the first time in three years) and clean control (firms that did not receive IPs for the span of six consecutive years). Replace main independent variable with a dummy that equals one if IPs were implemented.
  - Comparison to Baseline:
    - Results for average firm: The behavior of value added, TFP, and capital in response to each IP instrument is broadly consistent with Baseline, except for the impact of domestic subsidies on capital stock which becomes positive but insignificant.
    - Firm heterogeneity:
      - Domestic subsidies on capital by firm financial constraints remains broadly aligned with Baseline.
      - Export incentives now increase TFP of both young and old firms in similar magnitudes, but there is evidence that targeted young firms were underperforming targeted old firms, suggesting stronger relative impact of export incentives on young firms.

- Sensitivity of results to sample restrictions
  - Description: (1) Drop observations from China. (2) Remove Spain, Italy, and France, one at a time, given large share of observations in the regression sample from these countries. (3) Winsorize sample at 1 and 99 percentiles of [log Y_ft+h − log Y_ft−1] for each horizon h.
  - Comparison to Baseline: Results broadly in line with Baseline.

- Adding firm-level controls; additional lags; remove input-output exposure controls
  - Description: (1) Control for total assets, leverage ratio, and cash flow to assets ratio. (2) Include three lags of dependent and independent variables to test sensitivity of the rule of thumb of Chudik and Pesaran (2015). (3) Remove controls for input-output exposure.
  - Additional details: Reduces the sample of countries, as some emerging market developing economies do not report this information.
  - Comparison to Baseline: Results broadly in line with Baseline.

- Changing the weighting scheme
  - Description: Weight regressions by: (1) the inverse of the number of firms in the sample for each country; (2) firms sales in the previous year.
  - Comparison to Baseline: Results broadly in line with Baseline.

- Assessing the impact of all GTA subsidies (duplicate entry)
  - Description: Run a regression that tracks the impact on an additional subsidy (whether it is classified as IP or not) on value added.
  - Comparison to Baseline: Results are comparable to those found for subsidies classified as IPs.

### Industrial Policies at the Sectoral Level
- IV strategy
  - Description: Same as firm-level robustness exercise.
  - Comparison to Baseline:
    - All first stages are strong.
    - Domestic subsidies on value added is still positive, but only significantly positive at the last horizon.
    - Domestic subsidies on advanced economies (AEs) remains broadly consistent with Baseline.
    - Export incentives on value added now becomes positive but insignificant throughout all horizons.
    - Export incentives on AEs becomes positive at larger magnitude but insignificant starting in horizon 1.

- LP-DiD strategy
  - Description: Same as firm-level robustness exercise.
  - Comparison to Baseline:
    - Domestic subsidies on VA is negative upon impact and positive with larger magnitude in the medium term.
    - Domestic subsidies on AE is positive but insignificant.
    - Export incentives on VA have same shape as Baseline but less significance.
    - Positive impact of export incentives on AE now materializes in the first horizon and goes to zero in the medium term.

- Excluding China
  - Description: Same as firm-level robustness exercise.
  - Comparison to Baseline: Results broadly in line with Baseline.

- Different specifications for the comparison with structural policies
  - Description: Include all structural policies (Financial Development, Governance, Regulatory Burden of Doing Business) interacted with their respective distortions in a single equation or doing one at a time in separate regressions.
  - Comparison to Baseline: Comparison between IPs and structural policies are aligned with Baseline.

### Economic Effect of Industrial Policies on Patent Filing
- Decomposition of protectionist IPs by type
  - Description: Subsidies and export incentives definitions taken from GTA’s classification. The specification is similar to the IV strategy described above with the instrumental variable considering the use of the same type of IP in other countries and sectors.
  - Comparison to Baseline / Results:
    - Results are robust to using the same instruments as the ones described above.
    - Subsidies and export incentives drive the positive effect of protectionist IPs on patent filing by foreign inventors.
    - An additional protectionist subsidy is associated with a 2.0 percent increase in the number of received foreign patent applications in the first year but does not last beyond the second year.
    - Export incentives are associated with a temporary effect of the same size that only materializes in the second year.

*Source: Authors’ calculations.*

---


_Source: https://www.imf.org/-/media/files/publications/sdn/2025/english/sdnea2025002-s001.pdf_