## wpiea2024044-print-pdf — Sections 2.1, 3.1–4.3, Appendix Table A2

## Source details

**Canonical URL:** [wpiea2024044-print-pdf — Sections 2.1, 3.1–4.3, Appendix Table A2](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024044-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024044-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024044-print-pdf.pdf.json)

---

### Construction of the Embodied Technology Imports Indicator (Section 2.1)
- Definition:
  - ETI_ig,t = (sum_{j ∈ J_ig,t} (M_ijg,t · S_jg,t)) / (sum_j (M_ijg,t)), where
    - J_ig,t = set of trading partners of country i in year t for goods of type g,
    - M_ijg,t = dollar value of imports of good g from partner j to country i in year t,
    - S_jg,t = relative technology score of country j in year t.
  - S_jg,t = (P_jg,t / max_j {P_jg,t}) × 100.
- Interpretation and properties:
  - ETI ranges between 0 and 100 and captures the average technological quality of imports of good g relative to the technological leader at each period in percentage points.
  - An increase in ETI between t and t+1 means country i is importing a mix of varieties closer in quality to those produced by the technology leader in period t.
  - ETI is scale-free and focuses on composition of imports.
  - Scoring relative to the leader (not ranks) yields finer quantification of technology leadership.

- Patent stock and depreciation:
  - Patent stock P_jg,t = sum of all patents for good category g granted up to time t in country j, discounted at a 5% annual rate.
  - Authors note earlier specification using patents in the 20 years preceding t without depreciation; similar results with 10- or 5-year windows.

- Patent counting and attribution:
  - Data: Spring 2022 PATSTAT, patent period 1947-2020, 90 patent-issuing authorities.
  - Attribution:
    - Count patent families by earliest national application year.
    - Attribute family to countries where registered within three years of first application (accounts for publication lags and production/export origin).
    - EP (EPO) applications attributed to individual European patent offices where post-grant fees are paid.
    - Attribution window experiments (one year) yield similar results.
  - Fractional counting:
    - Split each patent equally across reported IPC4 classes (e.g., 0.1 per IPC4 if patent has 10 IPC4 classes).
  - Mapping to products:
    - Use probabilistic IPC4→SITC Rev. 2 crosswalk (Lybbert & Zolas 2014), convert Rev. 2→Rev. 1 via UN crosswalk; final counts at SITC1 5-digit level.
  - Trade merge:
    - Merge patent allocations with COMTRADE trade values in USD for SITC1 commodities as reported by exporting partners.

- Aggregation and baseline ETI:
  - Aggregate SITC1 into six classes: “Manufacturing and Mining Machinery”, “Agriculture Machinery”, “Other Machinery”, “Other Intermediates”, “Other Equipment”, and “Final Goods” (residual).
  - Baseline ETI for all machinery combines “Agriculture Machinery”, “Manufacturing and Mining Machinery”, and “Other Machinery”.
  - Weights in ETI: total dollar value of commodities in the three machinery classifications.
  - Robustness:
    - ETI weighted by family size or by one-year forward citations highly correlated with baseline: correlation coefficients 0.884 (family size) and 0.899 (citations).

### Validation exercises and stylized facts (Section 2.1 summary)
- Validation targets:
  - Share of world exports of machinery, Economic Complexity Index (ECI), imports/exports of IP services (WTO-OECD Balanced Trade in Services: BPM5 1995-2012 “Royalties and license fees”; BPM6 2005-2021 “Charges for the use of intellectual property n.i.e.”).
- Key empirical relationships:
  - Positive correlation between technology score S_jg,t and share of world exports of machinery (binned scatter with 20 bins); robust to dropping observations with share > 0.05.
  - Correlation between country rankings by ECI and by S_jg,t = 0.72 for 1995-2020.
  - Higher S_jg,t associated with higher exports of IP services (plots excluding US and Japan; full plots show stronger correlations).
  - ETI for all machinery positively correlates with total IP imports as percent of GDP.
- Stylized facts:
  - Fact 1: Long-run ETI trends driven by leaders’ patenting; EMDE average ETI ~40 over 1970-2020 with decrease in 1980s and increase in 2010s; time fixed-effects recommended.
  - Fact 2: Inverse U-shaped ETI vs GDP PPP p.c.; among EMDEs relation is linear and upward-sloping.
  - Fact 3: Growth in ETI precedes GDP PPP p.c. growth in EMDEs (five-year averages, residualized by five-year fixed-effects).
  - Fact 4: ETI positively correlated with higher external financial openness (Chinn-Ito index), controlling for five-year period fixed-effects.

### Data and empirical approach (Section 3.1–3.2)
- Capital flows liberalization episodes:
  - De jure liberalization measured by Chinn-Ito index (KAOPEN): unbalanced panel of 181 countries, 1971-2020, index range −1.93 to 2.31.
  - Episode definition: year-on-year increase in KAOPEN ≥ 95th percentile of changes = 0.254 (KAOPEN rounded at third decimal); almost 90% of observations see no year-on-year change.
  - “Clean” reform episodes: no KAOPEN change in previous five years; first five years after joining panel considered unclean.
  - Episodes identified: 123 events total; 90 in EMDEs.
  - Average increase in KAOPEN for events: 0.96; median change: 1.07 (97th percentile).
  - Episodes cluster in 1990s and 2000s; controls exclude non-clean large changes.

- Trade restrictions control:
  - Use MATR (Measure of Aggregate Trade Restrictions) to control for trade liberalization; MATR covers 157 countries up to 2019.

- Macroeconomic data:
  - IMF WEO: real GDP p.c. PPP, imports/exports of goods and services, capital flow measures (FDI, portfolio, other investment).
  - Use lifetime cumulative capital flows and inverse hyperbolic sine transformation to handle dispersion; transformation defined over entire real line.

- Empirical method:
  - LP-DiD adaptation (Dube et al. 2023) for non-absorbing multiple treatments.
  - Core equation: y_{i,t+h} − y_{i,t−1} = β_h ∆D_{i,t} + δ^h_t + Σ γ^h_j y_{i,t−j} + Σ η^{h,′}_j x_{t−j} + ε^h_{c,t}, where ∆D_{i,t} = 1 if liberalization episode at t.
  - Clean control condition: require no treatment in preceding L periods; chosen L = 5 based on stabilization condition (β̂_{L+s} within 90% CI of β̂_L).
  - Outcomes: log GDP PPP p.c., asinh cumulative capital inflows (total, FDI, portfolio, other), log ETI.
  - Controls: two lags of GDP PPP p.c. growth; robust SEs clustered at country level.
  - Alternative inference: Arellano & Bond (1991) and Driscoll & Kraay (1998) SEs reported in appendices.

- Treatment-control balance:
  - Pre-treatment differences significant only in GDP PPP p.c. and KAOPEN; mitigated by including GDP growth lags and enforcing clean control condition.
  - Conditional on covariates and clean condition, no significant pre-trend differences.

### Main results — Capital flows, Technology adoption, Output (Sections 4.1–4.3)
- KAOPEN response:
  - After events, KAOPEN exhibits a permanent increase of about one unit.
  - This increase = almost three standard deviations of KAOPEN changes using full sample and 2.8 standard deviations within EMDEs.
  - Events are large, rarely reversed; no significant pre-event trends in ten years before episodes (first five years unclean).

- Capital flows (lifetime cumulative inflows, asinh transformed):
  - Transformed lifetime gross capital inflows increase by 0.24 over the medium term and 0.28 in the long-run.
  - These represent 0.9 to 1.1 times the standard deviation of changes in transformed capital flows for all EMDEs and 1.1 to 1.3 times the average within-country standard deviation.
  - Equivalent to increases of 28% and 33% of lifetime capital flows over the medium and long-term respectively.
  - Composition: increase driven by FDI; lifetime FDI inflows increase by 21% five years after the event.
  - Other investment and portfolio flows: increases not statistically significant.

- Technology adoption (ETI, log):
  - ETI increases by 7% five years after liberalization and by 9% in the long-run.
  - Effect sizes equal:
    - 30% to 40% of the standard deviation of ETI changes for all EMDEs over the sample.
    - 45% to 57% of the average within-country standard deviation of ETI changes.
  - Interpretation: liberalization → capital inflows (notably FDI) → improvement in technology embodied in imports.

- Output (PPP Real GDP per capita):
  - Output increases by 9% and 12% five and ten years after the event respectively.
  - Equivalences: 1.3 and 1.8 standard deviations of GDP changes in the sample; 2.5 and 3.4 average within-country standard deviations.
  - Sequence observed: liberalization → capital inflows (FDI) → technology upgrading (ETI) → higher output.

### Robustness, alternative specifications, and mechanisms
- Robustness checks:
  - Narrower event set excluding simultaneous trade liberalizations (pure MATR components): data reduced (events down 40%, sample down ~13%); main results hold with larger capital flow effects and similar ETI and output effects.
  - Full set of controls (IMF programs, trade openness, net capital flows) reduces event pool by 25%; estimates less precise but similar in direction, magnitude, significance.
  - Capital account restriction episodes: restrictions partially reversed within five years; no significant effect on FDI or ETI; significant decrease in portfolio inflows and GDP (GDP decrease comparable to liberalization-induced increase).
  - ETI robustness:
    - Lifetime IP imports increase by 20% five years after event.
    - ETI increases observed for agriculture machinery, manufacturing & mining machinery, and other machinery; strongest for manufacturing machinery.
    - ETI adjusted for family size and for forward citations produce similar significant increases (patent-quality adjustments).
  - Econometric robustness:
    - Arellano-Bond GMM (minimum lags = 4, collapse instruments) yields point estimates within 90% CIs of baseline, often smaller.
    - Driscoll & Kraay SEs (bandwidth = 5) yield more precise estimates for most horizons; smaller bandwidths closer to baseline SEs.
    - If stabilization lag L = 0, estimates biased toward 0 due to contamination from previously treated units.

- Mechanisms (interpretation):
  - FDI appears central: liberalizations produce immediate FDI increases; restrictions do not raise FDI.
  - Three channels by which FDI may promote technology adoption:
    - FDI raises GDP p.c., leading to upgrading of capital goods.
    - FDI alleviates firms’ credit constraints enabling purchase of better capital goods.
    - Foreign investors transmit expertise and knowledge, improving local firms’ capital purchases.
  - Disentangling channels requires firm-level data or structural modeling.

### Appendix Table A2 — Clean capital flow liberalization episodes (inventory and balance)
- Clean episodes definition: KAOPEN increase > 0.254 occurring > five years from any prior KAOPEN change.
- Inventory: 123 events listed; selected sample entries include ARG 1993; AUS 1984, 2012; BRA 1998; CHN 1993; IND entries not listed in excerpt; many others through 2020 (full list in Appendix Table A2).
- Balance (EMDEs) — descriptive statistics after residualizing for year fixed-effects (means and standard deviations/standard errors in parentheses):
  - ∆log(GDP): column (1) = 0.031 (0.064); column (2) = 0.036 (0.058); column (3) = 0.002 (0.003)
  - ∆log(GDP, PPP): column (1) = 0.032 (0.064); column (2) = 0.036 (0.058); column (3) = 0.002 (0.003)
  - ∆log(GDP, PPP, p.c.): column (1) = 0.010 (0.065); column (2) = 0.019 (0.056); column (3) = 0.013*** (0.003)
  - ∆log(ETI), baseline: column (1) = 0.010 (0.311); column (2) = 0.005 (0.373); column (3) = -0.010 (0.018)
  - ∆MATR: column (1) = 0.017 (0.368); column (2) = -0.019 (0.687); column (3) = -0.041 (0.035)
  - ∆KAOPEN: column (1) = -0.022 (0.129); column (2) = -0.047 (0.229); column (3) = -0.023** (0.011)
  - ∆FDI/GDP, pp: column (1) = -0.038 (4.585); column (2) = 0.060 (2.160); column (3) = -0.056 (0.231)
  - ∆Net Capital Inflows, asinh: column (1) = -0.434 (17.686); column (2) = 0.064 (17.475); column (3) = 0.074 (1.232)
  - ∆Trade Openness: column (1) = -0.005 (0.250); column (2) = -0.031 (0.863); column (3) = -0.042 (0.036)
  - ∆Exports/GDP, goods, pp: column (1) = -0.834 (19.284); column (2) = -1.007 (24.766); column (3) = -0.290 (1.296)
  - ∆Imports/GDP, goods, pp: column (1) = -0.089 (12.935); column (2) = -1.367 (33.743); column (3) = -1.769 (1.453)
  - ∆No. IMF programs: column (1) = 0.128 (0.352); column (2) = 0.166 (0.387); column (3) = 0.025 (0.024)
  - Observations: column (1) = 1,061; column (2) = 1,149; column (3) = 2,210
- Interpretation: conditional on covariates and clean control condition, pre-trend differences are not statistically significant except for ∆KAOPEN and ∆log(GDP, PPP, p.c.), addressed via controls.

*Italic: Source: wpiea2024044-print-pdf — Sections 2.1, 3.1–4.3, Appendix Table A2*

### 2.1    Construction of the Embodied Technology Imports Indicator

### 2.1 Construction of the Embodied Technology Imports Indicator

### Indicator definition and interpretation
- The Embodied Technology Imports indicator (ETI) for country i and goods of type g is defined as:
  - ETI_ig,t = (sum_{j ∈ J_ig,t} (M_ijg,t · S_jg,t)) / (sum_j (M_ijg,t)), where
    - J_ig,t denotes the set of trading partners of country i in year t for goods of type g,
    - M_ijg,t is the dollar value of imports of goods of type g from partner j to country i in year t,
    - S_jg,t is the relative technology score of country j in year t.
- The relative technology score S_jg,t is built from patent stocks as:
  - S_jg,t = (P_jg,t / max_j {P_jg,t}) × 100.
- Properties and interpretation:
  - ETI ranges between 0 and 100 and captures the average technological quality of imports of good g relative to the technological leader at each period in percentage points.
  - An increase in ETI between t and t+1 means that in t+1 country i is importing a mix of varieties closer in quality to those produced by the technology leader in period t.
  - The ETI is scale-free: it captures composition of imports rather than mechanically increasing with import volume.
  - Defining the score relative to the leader (rather than order ranks) allows more precise quantification of technology leadership.

### Patent stock construction and depreciation
- Patent stock P_jg,t is built as the sum of all patents for good category g granted up to time t in the patent office of country j, discounted at a 5% annual rate to reflect duration of patent protection.
- The authors note a previous iteration where P_jg,t was the sum of patents registered in country g in the 20 years preceding period t without depreciation; results were similar when using shorter intervals like 10 or 5 years.

### Counting and attributing patents
- Source dataset:
  - Spring 2022 edition of PATSTAT maintained by the European Patent Office, covering 90 patent-issuing authorities.
  - Patent period used: 1947-2020.
  - Endpoints chosen to reflect availability of the Chinn-Ito Index (1971-2020) and to allow at least 20 years to build the patent stock available in the first year and up to four lags of the resulting indicator.
- Attribution rules:
  - For each application authority, count families of patents granted in each year.
  - Attribute each patent family to the earliest year an application appears in a national patent authority.
  - Attribute the family to countries in which it gets registered within three years of the first application date (to account for publication lags and to identify where the invention is produced and exported from).
  - Patents registered directly at the EPO (application authority listed as “EP”) are attributed to individual European patent offices where corresponding post-grant fees are paid.
  - The authors experimented with attribution windows (within one year) with similar results.
- Fractional patent counting:
  - Since patents usually apply to multiple commodities, fractional counts split each patent among the goods it applies to.
  - For each family, obtain IPC4 (International Patent Classification at the 4-digit level) classes reported by PATSTAT and attribute an equal fraction of each patent to each IPC4 class.
  - Example: count 0.1 patents for each IPC4 code of a patent classified in 10 IPC4 classes.
- Mapping patents to product classes:
  - Use a probabilistic crosswalk of IPC4 codes into SITC Revision 2 codes compiled by Lybbert & Zolas (2014) on US patents; this crosswalk fractionally attributes each IPC4 to multiple SITC codes with probability weights.
  - Translate IPC4→SITC Rev. 2 into IPC4→SITC Rev. 1 using a UN Statistics Division crosswalk; when Rev. 2 codes match multiple Rev. 1 codes, probability weights are partitioned equally among matched subclasses.
  - Final attribution allows patent counts at SITC1 5-digit level (lowest level of product classification used).
- Trade data:
  - Merge PATSTAT-derived patent allocations with COMTRADE trade values in USD for SITC1 commodities imported by all sample countries as reported by their trade partners (exporter countries).
  - For low-income countries, trade partners for technologically advanced goods will likely be advanced economies with higher data collection capabilities.

### Aggregation and baseline ETI focus
- Aggregate SITC1 commodities into six large classes for interpretation: “Manufacturing and Mining Machinery”, “Agriculture Machinery”, “Other Machinery”, “Other Intermediates”, “Other Equipment”, and “Final Goods” (residual).
  - “Other Equipment” includes optical and measurement, medical, and transportation equipment.
  - “Other Intermediates” includes unfinished intermediates.
- Baseline: ETI for all machinery combines “Agriculture Machinery”, “Manufacturing and Mining Machinery”, and “Other Machinery”.
  - For each year, undepreciated fractional count of granted patent families applying to SITC1 commodities relevant to these aggregations builds technology scores of sending countries.
  - Weights used in ETI are the total dollar value of all commodities falling into the three machinery classifications.
- Robustness checks:
  - Results are robust to ETI variants that weigh families by family size (an ex-ante quality measure counting countries a patent is registered in) or that use one-year forward citations (an ex-post quality measure).
  - Correlation coefficients for baseline measure:
    - 0.884 with ETI corrected by family size.
    - 0.899 with the citation-corrected ETI.

---

### Validation exercises (summary of validation strategy and key numeric relationships)
- Validation targets:
  - Exports of machinery (share of world exports of machinery).
  - Economic Complexity Index (ECI) of Hausmann et al. (2014).
  - Imports and exports of services related to the use of intellectual property (IP) from WTO-OECD Balanced Trade in Services dataset (BPM5: 1995-2012 item “Royalties and license fees”; BPM6: 2005-2021 item “Charges for the use of intellectual property n.i.e.”).
  - Measures reported in millions of current US dollars; shares of GDP obtained by dividing by nominal GDP in dollars when relevant.
- Key empirical correlations and findings:
  - Share of world exports of machinery plotted against technology score S_jg,t shows a clear positive correlation (binned scatter with 20 bins); results robust to dropping observations with share of world exports higher than 0.05.
  - Correlation between country rankings by ECI and rankings by technology score S_jg,t is 0.72 for 1995-2020.
  - Higher technology score S_jg,t is associated with higher exports of intellectual property services (both BPM5 and BPM6 measures). Plots exclude US and Japan to avoid dwarfing other observations; full plots in Appendix A show even stronger correlations.
  - ETI for all machinery positively correlates with total IP imports as a percentage of the importer’s GDP, suggesting association with knowledge flows and adoption of new technologies.
  - The authors verified similar patterns when considering narrower commodity groupings (e.g., only agriculture or manufacturing machinery).

---

### Stylized facts on the ETI (Facts 1–4)
- Fact 1: Long-run trends in technology adoption are driven by technology leaders’ activity.
  - Over 1970-2020, cross-country average ETI for EMDEs remained relatively stable around an average of 40, with a gradual decrease over the 1980s and a more recent increase in the 2010s.
  - Long-run changes reflect large increases in patenting in technology-leading countries:
    - Example: large increase in patenting in Japan during the 1980s not matched by equivalent increases in imports from Japan.
    - Recent increase related to increased patenting activity in China concurrent with the large increase in worldwide trade with China.
  - Because ETI is relative to the country with the highest stock of granted patents at each point in time, the index can drop across the board when the leader has far more patents than followers.
  - Recommendation: include time fixed-effects to account for long-run trends across countries.
- Fact 2: Inverse U-shaped relationship between ETI and GDP per capita.
  - Binned scatter of ETI against PPP real GDP per capita shows an inverse U-shape.
  - Interpretation: along the development path, countries first import and adapt increasingly sophisticated foreign technologies, then move to produce their own advanced machinery and import relatively less sophisticated machinery than they produce.
  - ETI is an adequate measure of technology diffusion primarily in a developing country context; when restricting to EMDEs the relation between ETI and GDP PPP p.c. is linear and upward-sloping (available on request).
- Fact 3: Growth in ETI is followed by GDP growth in EMDEs.
  - Increases in ETI usually precede growth in real GDP PPP p.c.
  - Evidence: binned scatter of average five-year GDP PPP p.c. growth against lagged five-year ETI growth (both averaged to absorb short-term fluctuations); observations residualized by five-year fixed-effects.
- Fact 4: ETI is positively correlated with higher external financial openness.
  - In the cross section of EMDEs, higher Chinn-Ito index values correspond to higher levels of the ETI.
  - Evidence shown as binned scatter of average ETI against average Chinn-Ito index for five-year periods; five-year period fixed-effects are partialled out.
  - Suggests technology adoption comes together with higher external finance openness; the rest of the paper examines this relation in depth.

_Italic: Source: wpiea2024044-print-pdf - 2.1 Construction of the Embodied Technology Imports Indicator_

### 3.1    Data

### 3.1    Data

### Capital flows liberalization episodes
- Focus: de jure external financial liberalization defined as removal of legislated barriers to international capital flows.
- Indicator: Chinn-Ito index (Chinn & Ito 2008), which aggregates four dummy variables encoding:
  - presence of multiple exchange rates,
  - restrictions on current account transactions,
  - restrictions on capital account transactions,
  - surrender of export proceeds.
- Index coverage and scale:
  - Available for an unbalanced panel of 181 countries in the period 1971-2020.
  - Index ranges between −1.93 and 2.31, with higher values denoting more de jure openness to capital flows (less restrictions).
- Episode definition:
  - A capital flow liberalization episode is a country-year pair with an increase in the Chinn-Ito index greater than or equal to the 95th percentile of observed year-on-year changes in the sample.
  - Threshold corresponds to a year-on-year change in KAOPEN of 0.254.
  - KAOPEN rounded at the third decimal digit; almost 90% of observations see no year-on-year change.
  - “Clean” reform episodes: those that do not see other change in KAOPEN in the previous five years (motivated by LP-DiD methodology).
  - First five years after a country joins the panel are considered unclean due to possible unobserved prior KAOPEN changes.
- Episodes identified:
  - 123 capital flow liberalization events in total.
  - 90 of the 123 events occur in EMDEs.
  - Average increase in KAOPEN for these events: 0.96.
  - Median change in KAOPEN: 1.07, corresponding to the 97th percentile of year-on-year changes in the full sample.
- Empirical note on episodes:
  - Episodes cluster in the 1990s and 2000s.
  - Capital controls are often gradually lifted, resulting in multiple staggered increases in KAOPEN; non-clean episodes are not used as controls but may be included among treated units when appropriate.

### Trade Restrictions
- Control for trade liberalization episodes using the MATR (Measure of Aggregate Trade Restrictions) built by Estefania-Flores et al. (2022), which also builds on the IMF AREAER dataset but focuses on trade-restricting measures.
- MATR use rationale:
  - Shares some components with Chinn-Ito (current account restrictions, surrender requirements).
  - Included to ensure ETI results are not driven solely by increases in trade openness.
  - ETI is scale-free and varies with composition of trade partners for specific goods; changes in MATR should not mechanically affect ETI if trade restrictions are similar across partners.
- MATR coverage: available for 157 countries and for all years in the sample up to 2019.

### Macroeconomic data
- Source: IMF World Economic Outlook.
- Variables pulled:
  - real GDP per capita in PPP terms,
  - imports and exports of goods and services in constant dollars,
  - measures of capital flows: FDI inflows and outflows, overall capital inflows and outflows, portfolio inflows and outflows, “other investment” inflows and outflows.
- Treatment of volatile flows:
  - Use lifetime cumulative capital flows (smoother series and better pre-period benchmark than yearly flows at t−1).
  - Apply inverse hyperbolic sine transformation to cumulative flows to handle high cross-sectional dispersion; monotonic transformation similar to natural logarithm but defined over entire real line and concave for positive values.

### Empirical approach (3.2)
- Method: Local projection difference in differences (LP-DiD) adapting equations (24) and (25) in Dube et al. (2023) to non-absorbing multiple treatments.
- Core estimating equation (conceptual):
  - Dependent variable: cumulative growth in outcome y between t−1 and t+h: y_{i,t+h} − y_{i,t−1} = β_h ∆D_{i,t} + δ^h_t + Σ γ^h_j y_{i,t−j} + Σ η^{h,′}_j x_{t−j} + ε^h_{c,t}.
  - Main regressor: ∆D_{i,t}, dummy indicating liberalization episode at time t.
  - Controls: time fixed effects δ^h_t, pre-treatment lags of change in outcome, vector of predetermined control variables x_{t−j}.
- Sample restriction (clean condition):
  - For treatment at time t, require no unit included was treated at any of preceding L periods:
    - Treated: ∆D_{i,t} = 1; ∆D_{i,t−j} = 0 ∀ j ∈ {1,...,L}.
    - Clean controls: ∆D_{i,t} = 0; ∆D_{i,t−j} = 0 ∀ j ∈ {1,...,L}.
  - Purpose: ensure control units are not experiencing delayed treatment effects from prior liberalizations.
- Stabilization periods L:
  - L is the “number of stabilization periods” — years required for treatment effects to stabilize.
  - Condition for causal recovery: true treatment effects β_h are constant after L periods (β_h = β_{h+s} for all s > 0 and h ≥ L).
  - Chosen L = 5 based on analysis of time path of estimated coefficients and condition that β̂_{L+s} fall in the 90% confidence interval for β̂_L.
  - Note: number of treated units drops for large post-treatment horizons; check condition up to post-treatment period with comparable observations.
- Outcome variables and transformations:
  - log of GDP PPP per capita,
  - inverse hyperbolic sine of cumulative capital inflows (total, FDI, portfolio, other investment),
  - log of Embodied Technology Imports (ETI) measure.
  - Transformations facilitate interpretation as percentage cumulative increases at horizon.
  - When using inverse hyperbolic sine, report cumulative percentage increase in neighborhood of sample average.
- Controls and inference:
  - Include two lags of GDP PPP p.c. growth in all regressions.
  - Robust standard errors clustered at country level.
  - Address potential Nickell (1981) bias via alternative Arellano & Bond (1991) specification in Appendix B.
  - Appendix B also reports Driscoll & Kraay (1998) standard errors to account for cross-sectional correlation.

### Treatment and Control Balance (3.3)
- Comparison of treated vs never-treated countries along multiple dimensions to detect unconditional violations of parallel trends.
- Significant pre-treatment differences found only in:
  - GDP PPP per capita,
  - capital openness (KAOPEN).
- Mitigations:
  - Use GDP PPP per capita growth as a control.
  - Enforce clean control condition (no KAOPEN changes in prior L = 5 years) to eliminate pre-treatment differences in KAOPEN behavior across treatment and control units.
- Result:
  - Conditional on covariates and clean control condition, no significant differences in pre-trends across treatment and control units.
  - Appendix A.3 provides balance table and construction details.

### Main Results (overview)
- Sample: 123 events identified; 90 correspond to liberalizations in EMDEs.
- KAOPEN response:
  - After events, KAOPEN exhibits a permanent increase of about one unit.
  - This increase equals almost three standard deviations of KAOPEN changes using full sample and 2.8 standard deviations of changes within EMDEs.
  - On average, events are not followed by a reversal and represent a large change in capital account openness.
  - No significant pre-event trends in the ten years before the episodes (note: first five years are unclean by construction).

### Capital Flows (4.1)
- Variable treatment: gross capital inflows defined as liability flows net of repayments; expressed in billions of US dollars; inverse hyperbolic sine applied to cumulative lifetime total.
- Estimated effects (from Figure 11 summary):
  - Transformed lifetime gross capital inflows increase by 0.24 over the medium term and 0.28 in the long-run.
  - These represent 0.9 to 1.1 times the standard deviation of changes in transformed capital flows for all EMDEs and 1.1 to 1.3 times the average within-country standard deviation of transformed capital flows changes.
  - Equivalent to increases of 28% and 33% of lifetime capital flows over the medium and long-term respectively.
- Composition of inflows:
  - Increase in lifetime capital flows is driven by Foreign Direct Investment (FDI).
  - Lifetime FDI inflows increase by 21% five years after the event.
  - Increases in other investment flows and portfolio flows are not statistically significant.
- Robustness notes:
  - Qualitatively similar results obtained using period-by-period inflows and using natural log instead of inverse hyperbolic sine (results available upon request).
  - If stabilization lags set to 0, estimated treatment effects are biased toward 0 due to previously treated units contaminating controls (consistent with methodological discussion).

### Technology adoption (4.2)
- Outcome: Embodied Technology Imports (ETI) indicator (natural logarithm used).
- Estimated effects (from Figure 12 summary):
  - ETI increases by 7% five years after liberalization event.
  - ETI increases by 9% in the long-run.
  - Effect size equivalences:
    - 30% to 40% of the standard deviation of ETI changes for all EMDEs over the sample period.
    - 45% to 57% of the average within-country standard deviation of ETI changes.
- Interpretation:
  - Capital account liberalization episodes are followed by larger capital inflows, particularly via FDI, and by an improvement in technology embodied in imports.
  - Appendix B provides further supporting results using import of IP and alternative ETI definitions (goods aggregation and patent counting).
  - Results on capital flows suggest the improvement in technology adoption is related to increased FDI.

*Source: IMF working paper chapter section 3.1–4.2 (content unit: wpiea2024044-print-pdf - 3.1 Data).*

### 4.3    Output

### 4.3    Output

### Main findings on output response to financial liberalization
- Financial liberalization has a positive impact on PPP Real GDP per capita levels over the medium to long-term.
- Output increases by 9% and 12% five and ten years after the event respectively.
- These increases are equivalent to 1.3 and 1.8 standard deviations of GDP changes in the sample and to 2.5 and 3.4 of the average within country standard deviation of GDP changes.
- The tighter identification strategy used in the analysis allows the authors to find this positive and significant effect in output levels (as conjectured by Henry 2007).
- The results provide suggestive evidence that the increase in capital flows allows for investment in more modern technology and a consequent increase in output.

### Evidence linking financial liberalization, capital flows, technology adoption, and output
- Financial liberalization episodes are followed by heightened capital inflows and by a level shift in PPP real GDP p.c., suggesting a long-lasting effect of opening up to financial flows.
- The ETI (Embodied Technology Imports) measure of technology adoption increases following liberalizations; the paper reports a technological upgrade amounting to forty percent of a standard deviation increase over the 5 to 10 years following a positive change in the Chinn-Ito index.
- The sequence observed is: liberalization → capital inflows (notably FDI) → technology upgrading (ETI) → higher output.

### Robustness checks and alternative specifications
- Pre-treatment lags are presented to inspect pre-treatment differences; these are not systematic after conditioning for lagged GDP growth and the dependent variable.
- Inclusion of lagged variables (GDP growth and outcome variables) is used to control for expectations formed based on past outcomes.
- Arellano & Bond (1991) methodology is used to show that inclusion of lags of the dependent variable does not cause Nickell’s (1981) bias.
- Driscoll & Kraay (1998) standard errors are employed with unchanged results.

### Simultaneous reforms and narrower event definitions (Section 5.1)
- Episodes that also ease trade restrictions (based on the pure trade components of MATR) are excluded to form a narrower set of liberalization events.
- Data availability for MATR reduces the number of events by 40%, and the rest of the overall sample by about 13%.
- Results hold under this narrower definition: larger effects on capital flows and a similar effect on technology adoption and output are found.

### Full set of controls (Section 5.2)
- A larger set of controls is considered: IMF programs, trade openness, and changes in net capital flows, together with lags of output growth and the dependent variable.
- Including all these controls reduces the pool of events by 25%, which produces less precise estimates, but results remain in the same direction, magnitude and significance as under the baseline for both technology adoption and GDP.

### Capital account restriction episodes (Section 5.3)
- Capital account restrictions tend to be partially reversed within five years, with initial sizes similar to liberalizations.
- No significant effect is found on FDI inflows or technology adoption during restriction episodes.
- There is a significant decrease in portfolio inflows and in GDP following restriction episodes; the GDP decrease is comparable to the increase found after liberalizations.
- Capital account restrictions do not mirror liberalization episodes and affect output through different channels; constraints in restriction episodes do not seem to affect technology adoption in either direction.
- A mildly negative response of other investment (which includes bank credit) in the case of capital flow restrictions may indicate that FDI provide more than just financing.

### Interpretation and mechanisms (Section 6 Discussion)
- Opening to external finance plays an important role in upgrading countries’ capital stocks.
- FDI flows appear crucial: liberalization episodes see a significant increase in FDI that manifests right after reforms, while FDI do not react in the case of restrictions.
- Three plausible channels by which FDI promotes technology adoption:
  - FDI raises GDP per capita, which leads to upgrading of capital goods (stylized fact: in EMDEs, GDP per capita and ETI are positively correlated).
  - FDI alleviates firms’ credit constraints, allowing purchase of better capital goods.
  - Foreign investors transmit expertise and knowledge, improving capital purchases by local firms.
- Disentangling these mechanisms requires firm-level data on technology adoption and capital flows by type.

### Data and index construction (Section 7 Conclusion)
- The paper constructs a new index of technology adoption (ETI) covering 181 countries (of which 155 are EMDEs) over the period 1970-2020.
- The ETI measures the sophistication of machinery imports based on a technology score for exporters.
- The technology score correlates strongly with alternative technology measures, notably the Economic Complexity Index, and with imports of intellectual property for the subset with such data.
- The ETI is particularly useful for analyzing technology in EMDEs where adoption of foreign technologies is the main path to upgrading the capital stock, and in these countries the ETI also predicts further growth.

### Conclusions and future research avenues
- Financial liberalization in EMDEs leads to significant increases in technology adoption (ETI) and to higher PPP real GDP p.c. in the medium to long-term.
- Results are robust to controlling for contemporaneous removal of trade barriers, broader short-run economic conditions, and alternative standard error estimation methods.
- Two avenues for future research:
  - Investigate the proposed causal chain (capital inflows → technology upgrading → higher output) further via mediation analysis or firm-level data, or via structural modeling.
  - Use the ETI to study responses to changes in countries’ political alignment to assess the impact of geopolitical fragmentation on technology adoption.

*Source: wpiea2024044-print-pdf - 4.3    Output*

### 1996. Appendix Table A2 reports all the “clean” events that we include in our sample.

### Table A2: Clean Capital Flow Liberalization Episodes (Appendix)

### Clean episodes (inventory)
- The appendix reports "clean" capital flow liberalization episodes used in the analysis. A clean treatment is defined as an increase in KAOPEN larger than 0.254 which occurs more than five years away from any previous changes in KAOPEN.
- Selected episodes (country — episode years):
  - ARG Argentina — 1993
  - AUS Australia — 1984, 2012
  - AUT Austria — 1991
  - AZE Azerbaijan — 2002
  - BEL Belgium — 1990
  - BGD Bangladesh — 1992
  - BRA Brazil — 1998
  - BRB Barbados — 1993
  - BWA Botswana — 1987
  - CAF Central African Republic — 1991
  - CHL Chile — 1976, 1995
  - CHN China — 1993
  - CIV Côte d’Ivoire — 1993
  - CMR Cameroon — 1993
  - COD Democratic Republic of the Congo — 1997
  - COG Congo, Republic of — 1995
  - COL Colombia — 1990, 2004
  - CPV Cabo Verde — 2017
  - CRI Costa Rica — 2011
  - CYP Cyprus — 1993, 2003
  - DMA Dominica — 2005
  - DNK Denmark — 1988
  - DOM Dominican Republic — 1991
  - DZA Algeria — 1988
  - EGY Egypt — 1994
  - ESP Spain — 1993
  - FIN Finland — 1991
  - FRA France — 1990
  - GBR United Kingdom — 1979
  - GHA Ghana — 1994
  - GNQ Equatorial Guinea — 1984, 1995
  - GRC Greece — 1992
  - GRD Grenada — 1993, 2005
  - GTM Guatemala — 2001
  - HND Honduras — 2008
  - HRV Croatia — 2003
  - HTI Haiti — 1997, 2003
  - HUN Hungary — 1993
  - IRL Ireland — 1978, 1992
  - IRN Iran — 2001
  - ISL Iceland — 1992, 2016
  - ISR Israel — 1993
  - ITA Italy — 1982, 1990
  - JAM Jamaica — 1992
  - JOR Jordan — 1979, 1995
  - JPN Japan — 1979
  - KAZ Kazakhstan — 2019
  - KHM Cambodia — 2015
  - KOR Korea — 1978, 1988, 2008
  - LAO Lao P.D.R. — 1995
  - LKA Sri Lanka — 1977, 1992
  - LVA Latvia — 2003
  - MAR Morocco — 1986, 1993
  - MDG Madagascar — 1997
  - MDV Maldives — 1996
  - MLT Malta — 1994, 2004
  - MMR Myanmar — 2020
  - MNG Mongolia — 2008
  - MUS Mauritius — 1984, 1993
  - MWI Malawi — 1995, 2013
  - MYS Malaysia — 2008
  - NER Niger — 1984, 1995
  - NGA Nigeria — 1986, 1997
  - NIC Nicaragua — 1990
  - NOR Norway — 1993
  - NPL Nepal — 1995
  - NZL New Zealand — 1976, 1984
  - OMN Oman — 2003
  - PHL Philippines — 1985, 1992
  - PNG Papua New Guinea — 2005
  - POL Poland — 1994, 2002, 2015
  - PRT Portugal — 1988
  - ROU Romania — 1992, 2002
  - RUS Russia — 2009
  - RWA Rwanda — 1994, 2010
  - SDN Sudan — 1987
  - SGP Singapore — 1978
  - SLE Sierra Leone — 2005, 2016
  - SLV El Salvador — 1992
  - SUR Suriname — 2016
  - SVK Slovak Republic — 2003
  - SWE Sweden — 1993
  - TCD Chad — 1993
  - THA Thailand — 2019
  - TJK Tajikistan — 2008
  - TTO Trinidad and Tobago — 1992
  - TUN Tunisia — 1992
  - TUR Turkey — 1989, 2008
  - UZB Uzbekistan — 2018
  - VNM Vietnam — 1993, 2008
  - WSM Samoa — 1993
  - ZAF South Africa — 1993
  - ZWE Zimbabwe — 1994, 2009

### Balance between treatment and control (EMDEs)
- Sample construction:
  - Comparison between countries that received a clean treatment (increase in KAOPEN > 0.254 and > five years from previous changes) and countries that never saw an increase in KAOPEN.
  - Controls exclude countries that see a year-on-year increase in KAOPEN larger than 0.254.
  - Time fixed-effects are included to residualize means because events are clustered over time.
- Key descriptive-statistics results (means and differences after residualizing for year fixed-effects; standard deviations or standard errors in parentheses):
  - ∆log(GDP): column (1) = 0.031 (0.064); column (2) = 0.036 (0.058); column (3) = 0.002 (0.003)
  - ∆log(GDP, PPP): column (1) = 0.032 (0.064); column (2) = 0.036 (0.058); column (3) = 0.002 (0.003)
  - ∆log(GDP, PPP, p.c.): column (1) = 0.010 (0.065); column (2) = 0.019 (0.056); column (3) = 0.013*** (0.003)
  - ∆log(ETI), baseline: column (1) = 0.010 (0.311); column (2) = 0.005 (0.373); column (3) = -0.010 (0.018)
  - ∆MATR: column (1) = 0.017 (0.368); column (2) = -0.019 (0.687); column (3) = -0.041 (0.035)
  - ∆KAOPEN: column (1) = -0.022 (0.129); column (2) = -0.047 (0.229); column (3) = -0.023** (0.011)
  - ∆FDI/GDP, pp: column (1) = -0.038 (4.585); column (2) = 0.060 (2.160); column (3) = -0.056 (0.231)
  - ∆Net Capital Inflows, asinh: column (1) = -0.434 (17.686); column (2) = 0.064 (17.475); column (3) = 0.074 (1.232)
  - ∆Trade Openness: column (1) = -0.005 (0.250); column (2) = -0.031 (0.863); column (3) = -0.042 (0.036)
  - ∆Exports/GDP, goods, pp: column (1) = -0.834 (19.284); column (2) = -1.007 (24.766); column (3) = -0.290 (1.296)
  - ∆Imports/GDP, goods, pp: column (1) = -0.089 (12.935); column (2) = -1.367 (33.743); column (3) = -1.769 (1.453)
  - ∆No. IMF programs: column (1) = 0.128 (0.352); column (2) = 0.166 (0.387); column (3) = 0.025 (0.024)
  - Observations: column (1) = 1,061; column (2) = 1,149; column (3) = 2,210
- Interpretation and adjustments:
  - No significant differences across most variables except ∆KAOPEN and ∆log(GDP, PPP, p.c.).
  - Treated countries exhibit larger restrictions in capital flow measures in periods preceding treatment (more than five years before liberalization).
  - Event-study graphs show KAOPEN appears larger and marginally significant only for estimates concerning 10 periods before the event for the full sample.
  - GDP PPP per capita shows a different trend in the two to three years before events; to address this, the analysis includes two lags of GDP PPP per capita growth as controls to match treatment and control units.

### Robustness — Technology adoption (summary of findings)
- Main additional tests:
  - Outcome variables: intellectual property (IP) imports and alternative variations of the Embodied Technology Imports (ETI).
  - Patent-count adjustments: raw count, adjusted by patent family size, adjusted by forward citations.
- Key quantitative result:
  - Lifetime IP imports increase by 20% five years after the event (Panel a).
- Decomposition by machinery type:
  - Agriculture machinery, manufacturing and mining machinery, and other machinery ETIs all exhibit increases after liberalization events.
  - Results are consistently significant and strongest for manufacturing machinery over the analyzed horizon.
- Patent-quality adjustments:
  - ETI adjusted for family size and ETI adjusted by forward citations both show statistically significant increases in the ETI similar to baseline estimates.
- Data limitations:
  - Number of events decreases dramatically after five years, reducing precision of estimates beyond that horizon.

### Robustness — Alternative econometric assumptions
- Arellano-Bond GMM (xtabond2, Roodman 2009) implementation:
  - Potentially endogenous: lags of the dependent variable (gmm style).
  - Exogenous: year fixed effects, event variable, and GDP lags when GDP is not the main dependent variable (iv style).
  - Minimum number of lags for instruments set at 4 to avoid AR(2)/AR(3) test failures; collapse option used to limit instrument proliferation.
  - Result: point estimates are within the 90% confidence bands of baseline; often smaller point estimates suggesting sample restriction compensates Nickell bias concerns.
- Driscoll & Kraay (1998) standard errors:
  - Bandwidth chosen = 5 to mirror stabilization lags.
  - Findings: Driscoll-Kraay yields more precise estimates for most horizons relative to baseline; smaller bandwidths bring estimated standard errors closer to baseline.
- Overall robustness conclusion:
  - Results are robust to Arellano-Bond GMM and Driscoll-Kraay standard errors. Estimates generally remain within baseline confidence intervals and retain qualitative features.

*Do Capital Inflows Spur Technology Diffusion? Evidence from a New Technology Adoption Index — Working Paper No. WP/2024/044*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024044-print-pdf.pdf_