## wp18117

## Source details

**Canonical URL:** [wp18117](https://www.imf.org/-/media/files/publications/wp/2018/wp18117.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp18117.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp18117.pdf.json)

---

### I. Main objective and key claim
- Main empirical objective: identify a brief list of key determinants of cross-country economic growth while addressing endogeneity (causation vs. correlation) problems.
- Standard IV limitations highlighted:
  - “blunt instrument” problem (multiple studies use the same instrument for different endogenous variables) as in Bazzi and Clemens (2013).
  - weak instruments producing misleading estimates.
- Proposed solution: instrument each endogenous variable with the average of the same variable in neighboring marine and land countries (variable-specific, time-varying “sharp” instruments).
  - Claimed properties: sharp (variable-specific), strong (produce strong correlations), less prone to bluntness.
- Key empirical claim: export sophistication is the only robustly significant determinant of growth among standard determinants (human capital, trade, financial development, institutions); in presence of export sophistication, other standard growth determinants mostly become statistically insignificant.
- Additional finding: export orientation of domestic production (exporting sophisticated tradables) matters more than domestic production per se or mere specialization in manufacturing.

### II. Export sophistication: measurement, data, and stylized facts
- EXPY definition: export-share weighted average of sophistication levels of a country’s export basket; sophistication of a good measured by weighted average of real GDP per capita of countries exporting that good.
- Data sources and coverage:
  - real GDP per capita: Penn World Tables 9.0 (1960-2014);
  - years of schooling: Barro and Lee (2013);
  - EXPY: World Trade Flows data (Feenstra and Romalis 2014) for 1962-2014, extended to 2014 in this paper.
- Alternative proxies and correlations with EXPY:
  - share of manufacturing in goods’ exports (correlation with EXPY about 60 percent);
  - real manufacturing exports per capita (correlation with EXPY about 75 percent).
- Structural EXPY (S-EXPY): constructed to discount artificially high EXPY of high-income commodity exporters (appendix).
- Stylized associations:
  - Export sophistication has a strong positive association with 5-year ahead real GDP per capita growth conditional on initial GDP per capita.
  - Alternative proxy (log real manufacturing exports per capita) also shows strong positive correlation with 5-year ahead growth conditional on initial income.
- Contextual example: many oil-exporting countries grow episodically on oil income but lack sustained productivity growth because of low export sophistication (Cherif and Hasanov 2016).

### III. Sharp-instrument methodology and identification strategy
- Blunt-instrument illustration: using same instrument z for different endogenous variables can render at least one IV invalid due to correlation with omitted components.
- Sharp instrument implementation:
  - instrument country i’s variable with the average of that same variable in country i’s neighboring countries.
  - instruments are variable-specific, time-varying, and expected to correlate strongly with country variables.
- Exclusion restriction rationale: neighbors’ variables affect country growth mainly via the country’s own variables.
- Threats to validity and mitigations:
  - Direct neighbor spillovers (imitation, regional competition, common language, cultural proximity): controlled by including average real GDP per capita or real GDP in neighbors.
  - Large-neighbor dominance: alternative instruments tested — median of neighbors’ variables and inverse-real-GDP weighted averages.
  - Shared time-invariant traits among neighbors: controlled using latitude, ethnic fractionalization, Sub-Saharan Africa dummy, and fixed-effect IV regressions.
- Robustness checks emphasized: control for neighbor GDP, use median/inverse-GDP weighted neighbor averages, geographic/fractionalization controls, run fixed-effect IV.

### IV. Main empirical results (IV, OLS, FE, GMM) — key statistics and magnitudes
- OLS/FE associations:
  - Example interpretation: a 10 percent increase in export sophistication associated with about 0.2-0.3 percent increase in annual growth in OLS/FE.
  - OLS Table 1 example coefficients:
    - Log export sophistication: 0.027*** (column 1 OLS; robust s.e. (0.004)); OLS range 0.012 to 0.029 with corresponding s.e.
    - Log real GDP per capita: -0.010*** to -0.015*** in OLS; in FE ranges include -0.044*** to -0.052*** and as low as -0.097*** in some FE columns.
    - Years of schooling: 0.003*** to 0.004*** in OLS; insignificant or negative in FE.
    - Credit to private sector (% of GDP): negative in OLS (e.g., -0.007** column 1) and in FE in some columns.
  - Sample sizes/examples: observations vary (e.g., 1,592; 1,226; 1,376; 1,333; 748; 609; 601). Adjusted R-squared examples: 0.082 (column 1 OLS) up to 0.301 (some FE specs).
- IV using neighbor averages (Table 2) — export sophistication:
  - IV coefficients (log EXPY) examples:
    - 0.072*** (column 1 IV; s.e. (0.009));
    - 0.063*** (column 2 IV; s.e. (0.009));
    - ranges 0.065*** to 0.077*** with s.e. (0.013), (0.012);
    - up to 0.105*** and 0.148*** in later specifications and 0.154** in an IV-FE column.
  - Magnitude interpretation: a 10 percent increase in EXPY increases average annual growth in the next 5 years by about 0.6-0.7 percent in baseline IV specifications.
  - Comparison: IV coefficients on EXPY are about two to three times larger than OLS/FE estimates, indicating a large downward bias in OLS/FE.
  - Initial real GDP per capita (log) IV coefficients remain significantly negative, e.g., -0.024*** (column 1), -0.020*** (column 2); FE-IV columns show larger magnitudes (e.g., -0.116***).
- Instrument strength and diagnostics (selected exact statistics from tables):
  - Cragg-Donald F stat examples: 347.0, 136.2, 40.7, 12.9, 37.9, 9.6, 4.4, 0.8, 1.4.
  - Kleibergen-Paap F stat examples: 267.5, 104.8, 38.4, 10.8, 24.8, 7.7, 2.5, 1.0, 1.3.
  - Kleibergen-Paap LM test p-values: 0.00 for many IV specs; 0.02, 0.11 in some.
  - Conditional Likelihood Ratio (CLR) confidence set for coefficient on EXPY (column 1): Lower CLR bound 0.06, Upper CLR bound 0.09; CLR p-value for H0: Beta_EXPY=0 equals 0.00.
  - Hansen J-test p-value example: 0.17.
  - Some specifications show weak-instrument concerns where KP Wald statistics are small.
- IV control-variable findings:
  - When instrumenting other determinants with neighbor averages, EXPY and initial income remain robustly significant in most specifications.
  - Years of schooling and trade are not statistically significant in IV (columns 3-4).
  - Credit to private sector and law and order show negative IV coefficients in some columns (credit: -0.023*; law and order: -0.016**).
  - In IV regression excluding EXPY (column 7), only years of schooling and initial income remain significant; with EXPY included (column 8), years of schooling becomes insignificant, law and order negative at 10 percent, while EXPY and initial income are strongly significant with larger EXPY coefficients.
- Robustness and interpretation summary:
  - Neighbor-average IVs generally pass strength tests in many specifications and produce consistently positive, economically meaningful effects of EXPY on subsequent growth.
  - CLR-based inference indicates EXPY coefficient positive and statistically significant in baseline IVs (example CLR range 0.06 to 0.09); point estimates larger in other specifications.
  - Authors interpret results as implying improvements in human capital, trade, financial development, or institutions raise growth primarily to the extent they raise export sophistication.

### V. Robustness checks and alternative measures
- Alternative proxies for export sophistication:
  - manufactures exports as share of merchandise exports; log real manufacturing exports per capita.
  - Both proxies significant and positive in baseline regressions (Table 3, columns 2 and 4).
  - Including EXPY and another manufacturing measure produces quasi-multicollinearity and insignificant coefficients; weak instrument tests indicate weak instruments.
- Manufacturing production controls:
  - Manufacturing value added (% of GDP) significant when EXPY included (Table 3, columns 6-7).
  - Log real manufacturing value added per capita not significant in those columns (columns 8-9).
  - With other proxies, manufacturing value added (% of GDP) becomes insignificant while log real manufacturing value added per capita coefficient is negative (columns 10-11).
  - Overall: export orientation matters; producing manufacturing without exporting may not be sufficient for long-run growth.
- Spillover effects from neighbors:
  - Controlled for average log real GDP (or real GDP per capita) in neighboring countries; instrumented neighbors’ GDP by neighbors’ neighbors’ averages.
  - Results examples:
    - Spillover (neighbors’ real GDP per capita) significant at 10 percent (Table 4, column 1); excluding schooling becomes insignificant (column 2).
    - In baseline (Table 4, column 2) average real GDP of neighbors positive and strongly significant while EXPY no longer significant (column 3); excluding schooling yields positive and significant EXPY (Table 4, column 4).
    - Using median or weighted averages for instruments makes spillover coefficient insignificant for real GDP per capita (columns 5-6 and 9-10); real GDP of neighbors estimates similar across weighting schemes (columns 7-8 and 11-12).
  - Conclusion: spillovers do not invalidate EXPY as a key growth determinant.
- Additional controls (Table 5, Table 6):
  - Tested: investment to GDP ratio, national saving rate, FDI, government consumption, Gini coefficient, corruption.
  - EXPY coefficient range across robustness checks: 0.05-0.08 (text summary); specific table examples include 0.084***, 0.065***, 0.043**, 0.109***.
  - Using Ding and Hadzi-Vaskov (2017) measure yields a robust positive estimate larger than previous regressions (Table 5, column 2).
  - S-EXPY produces a strong positive coefficient two to three times smaller than original EXPY (Table 5, column 1).
- GMM-system estimation (Table 6 and Table 7):
  - GMM instruments: lags of explanatory variables plus neighbor averages as exogenous IVs and sharp instruments as GMM instruments.
  - EXPY highly significant and positive in all GMM specifications; magnitudes around 0.05-0.1 consistent with IV.
  - Level-equation GMM (GMM-LEV) with fewer instruments avoids weak instruments and produces EXPY similar to IV (Table 7, column 7).
  - Difference-equation GMM specifications show weak instruments; identification appears to come from level-equation cross-country variation.
  - Tests: J-test suggests overidentifying restrictions not valid in some specs; Hahn, Ham, and Moon (2011) test indicates neighbor-average instruments valid (p-value ≈ 0.7).
- Panels and instrument variants:
  - 10-year panels: EXPY statistically significant in most estimations; coefficients similar to 5-year panels.
  - Alternative instrument aggregations (median, inverse-GDP weights) broadly unchanged relative to simple averages (Tables 9-10); inverse-GDP weights reduce influence of large neighbors.
  - Weak instrument tests unfavorable in a few estimations but overall robustness holds.

### VI. Quantitative examples and selected exact values
- Representative estimated effects of log EXPY on average annual growth (examples listed verbatim):
  - 0.036*** (0.011)
  - 0.038*** (0.011)
  - 0.076*** (0.010)
  - 0.085*** (0.013)
  - 0.062*** (0.010)
  - 0.074*** (0.013)
  - 0.081*** (0.020)
  - 0.061*** (0.016)
  - 0.031** (0.013)
  - 0.269 (0.444)
  - 0.157** (0.072)
- Log real GDP per capita coefficients examples:
  - -0.018*** (0.004)
  - -0.025*** (0.004)
  - -0.030*** (0.005)
  - -0.021*** (0.004)
  - -0.062** (0.026)
  - -0.135** (0.059)
- Selected table statistics (preserved exactly where provided):
  - Cragg-Donald F stat examples (Table 3): 40.7 (column 1), 109.7 (column 2), 2.6 (column 3), 13.0 (column 4).
  - Kleibergen-Paap F stat examples (Table 3): 38.4 (column 1), 100.0 (column 2), 1.7 (column 3).
  - Table 6 GMM examples: EXPY coefficients around 0.061**, 0.067***, 0.078***, 0.053***, 0.098***.
  - Table 7 examples: EXPY coefficients include 0.052***, 0.059***, 0.084***, 0.111***, 0.098**.
- Significance codes preserved: *** p<0.01, ** p<0.05, * p<0.1.
- Observations examples across panels/specs: 1,080; 1,089; 944; 1,216; 724; 596; 627; 618; 365; 306; 301; 363; 1,590; 1,216; 1,369; 1,319; 748; 606; 598; 983.

### VII. Identification limitations and diagnostics summary
- Including several endogenous explanatory variables increases weak-instrument risk because endogenous variables and instruments may be correlated.
- Fixed-effects IV (differenced equation) especially suffers from weak instruments—harder to predict growth rates using neighbors’ variables than levels.
- Reported test statistics reference Stock and Yogo (2005) for 2-3 endogenous variables while regressions sometimes include up to 6 endogenous variables; thresholds used are relatively conservative.
- CLR-based inference used to handle potential weak-instrument problems for single endogenous variable; CLR results show positive, significant EXPY in baseline IVs.

### VIII. Interpretation, mechanisms, and policy implications
- Interpretation:
  - Export sophistication is a robust, positive, and economically meaningful determinant of long-run growth across many specifications and estimation methods.
  - Identification stronger in level-equation (cross-country variation) estimations than in difference-equation (time-series variation); weak instruments recur when differencing or including many endogenous variables.
  - Mechanisms: cross-border diffusion, mimetic forces, regional “pioneer” dynamics—improvements in export sophistication diffuse regionally and may underlie observed growth patterns.
  - Standard determinants matter insofar as they raise export sophistication.
- Policy implications and recommendations:
  - Prioritize expansion of the tradable sector and the sophistication of exports.
  - Design purpose-specific policies to develop tradable sophisticated sectors rather than relying solely on a general “enabling environment.”
  - Recognize constraints in developing countries (skills, resources) and create targeted interventions to foster more sophisticated tradable outputs.
  - The neighbor-average instrument technique can be applied in other empirical contexts facing the blunt-instrument problem to generate variable-specific, dynamic instruments.

*Source: wp18117 - References (IMF working paper content provided).*

### References .............................................................................................................

### wp18117 - References

### I. Introduction
- Main empirical objective: identify a brief list of key determinants of cross-country economic growth while addressing endogeneity (causation vs. correlation) problems.
- Standard approach limitation: instrumental variable (IV) estimation requires instruments that are both strong (correlated with endogenous regressor) and valid (uncorrelated with growth equation residual). Many growth studies face:
  - the “blunt instrument” problem (multiple studies use the same instrument for different endogenous variables) as highlighted by Bazzi and Clemens (2013), implying at least one IV estimation may be invalid;
  - weak instruments producing misleading estimates.
- Proposed solution: instrument each endogenous variable with the average of the same variable in neighboring marine and land countries (variable-specific, time-varying “sharp” instruments), argued to be:
  - sharp (variable-specific),
  - strong (produce strong correlations), and
  - less prone to bluntness across studies.
- Key empirical claim: export sophistication is the only robustly significant determinant of growth among standard determinants (human capital, trade, financial development, institutions); in presence of export sophistication, other standard growth determinants mostly become statistically insignificant.
- Additional finding: export orientation of domestic production (exporting sophisticated tradables) matters more than domestic production per se or mere specialization in manufacturing.

### II. Export sophistication: measurement and stylized facts
- EXPY definition: export-share weighted average of sophistication levels of a country’s export basket; sophistication of a good measured by weighted average of real GDP per capita of countries exporting that good.
- Data sources and coverage:
  - real GDP per capita: Penn World Tables 9.0 (1960-2014);
  - years of schooling: Barro and Lee (2013);
  - EXPY: World Trade Flows data (Feenstra and Romalis 2014) for 1962-2014, extended to 2014 in this paper;
  - alternative proxies: share of manufacturing in goods’ exports and real manufacturing exports per capita (correlation with EXPY about 60 and 75 percent, respectively);
  - structural EXPY (S-EXPY) constructed to discount artificially high EXPY of high-income commodity exporters (see Appendix).
- Stylized empirical associations:
  - Export sophistication has a strong positive association with 5-year ahead real GDP per capita growth conditional on initial GDP per capita (visualized in Figure 1).
  - Alternative proxies (log real manufacturing exports per capita) also show strong positive correlations with 5-year ahead growth conditional on initial income.
- Contextual example: many oil-exporting countries grow episodically on oil income but lack sustained productivity growth because of low export sophistication (Cherif and Hasanov 2016).

### III. Sharp vs. blunt instrument (methodology)
- Blunt-instrument illustration: if two true determinants x and x′ exist, using the same instrument z for x in one study and for x′ in another can render at least one IV invalid because z may correlate with omitted variable components of the other specification.
- Sharp instrument approach:
  - instrument country i’s variable with the average of that same variable in country i’s neighboring countries;
  - advantages: variable-specific instruments, time-varying, broad applicability, expected strong correlation with country variable.
- Exclusion restriction rationale: neighbors’ variables affect country growth mainly via the country’s own variables (if valid).
- Threats to validity and mitigation strategies:
  - Direct neighbor spillovers (e.g., imitation, regional competition, common language, cultural proximity) could invalidate instruments; the authors control for average real GDP per capita or real GDP in neighboring countries as proxies for spillovers.
  - Large neighbors dominating simple averages: alternative instrument constructions tested — median of neighbors’ variables and weighted averages with weights inversely proportional to real GDP (reduces influence of big neighbors).
  - Shared time-invariant traits among neighbors (geography, climate, ethnicity): controlled using latitude, ethnic fractionalization, dummy for Sub-Saharan Africa, and fixed-effect IV regressions.
- Robustness checks proposed: control for neighbor GDP, use median and inverse-GDP weighted neighbor averages, include geographic/fractionalization controls, run fixed-effect IV.

### IV. Empirical results (overview and key statistics)
- OLS and fixed effects (FE) correlations:
  - Regressions of 5-year average annual growth on log initial real GDP per capita and log export sophistication (EXPY) plus single additional standard determinants typically yield statistically significant coefficients with expected signs (law and order exception in some specifications; private sector credit shows negative coefficient).
  - Interpretation example: a 10 percent increase in export sophistication associated with about 0.2-0.3 percent increase in annual growth in OLS/FE.
  - OLS Table 1 coefficients (selected):
    - Log export sophistication: 0.027*** (column 1 OLS; robust s.e. (0.004)); range across OLS columns shown from 0.012 to 0.029 with corresponding standard errors.
    - Log real GDP per capita: -0.010*** to -0.015*** in OLS; in FE regressions ranges include -0.044*** to -0.052*** and as low as -0.097*** in some FE columns.
    - Years of schooling: 0.003*** to 0.004*** in OLS; insignificant or negative in FE.
    - Trade (% of GDP): OLS coefficients positive (e.g., 0.017*** in column 1) but mixed in other specifications.
    - Credit to private sector (% of GDP): negative in OLS (e.g., -0.007** column 1) and in FE in some columns.
  - Sample sizes and model fit examples:
    - Observations vary across specifications (e.g., 1,592; 1,226; 1,376; 1,333; 748; 609; 601).
    - Adjusted R-squared example: 0.082 (column 1 OLS), up to 0.301 (some FE specs).
- IV estimation using neighbor averages as instruments (Table 2) — key findings:
  - Export sophistication (log) IV coefficients:
    - 0.072*** (column 1, IV; s.e. (0.009));
    - 0.063*** (column 2, IV; s.e. (0.009));
    - 0.065*** to 0.077*** in other IV columns with standard errors reported (e.g., (0.013), (0.012));
    - Reaching as high as 0.105*** and 0.148*** in later specifications and 0.154** in an IV-FE column.
  - Magnitude interpretation: a 10 percent increase in export sophistication (EXPY) increases the average annual growth rate in the next 5 years by about 0.6-0.7 percent in baseline IV specifications (column-level results corroborate this magnitude).
  - Comparison to OLS/FE: IV coefficients on export sophistication are about two to three times larger than OLS/FE estimates, suggesting a large downward bias in OLS/FE.
  - Initial real GDP per capita (log) IV coefficients remain significantly negative, e.g., -0.024*** (column 1), -0.020*** (column 2), with some larger magnitudes in FE-IV columns (e.g., -0.116***).
- Instrument strength and weak-instrument diagnostics (selected statistics from Table 2):
  - Cragg-Donald F stat examples: 347.0, 136.2, 40.7, 12.9, 37.9, 9.6, 4.4, 0.8, 1.4 (varies by specification).
  - Kleibergen-Paap F stat examples: 267.5, 104.8, 38.4, 10.8, 24.8, 7.7, 2.5, 1.0, 1.3.
  - Kleibergen-Paap LM test p-value examples: 0.00 for many IV specs; 0.02, 0.11 in some.
  - Conditional Likelihood Ratio (CLR) confidence set for the coefficient on export sophistication (robust to weak instrument for one endogenous variable; column 1): Lower CLR bound 0.06, Upper CLR bound 0.09; CLR p-value for H0: Beta_EXPY=0 equals 0.00.
  - Hansen J-test p-value example: 0.17 (test of overidentifying restrictions in one IV specification).
  - Some specifications show potential weak-instrument concerns (notably where KP Wald statistic is much smaller and related null hypotheses about t-test size cannot be rejected at conventional thresholds).
- IV control-variable findings:
  - When controlling for other determinants one at a time (all instrumented with neighbor averages), export sophistication and initial income remain robustly significant in most specifications.
  - Years of schooling and trade are not statistically significant in IV (columns 3-4).
  - Credit to private sector and law and order show negative IV coefficients in some columns (credit: -0.023*; law and order: -0.016**), though with lower significance than EXPY or initial income; these negative signs may reflect nonlinearities or vulnerabilities associated with excessive credit or institutional dynamics.
  - In an IV regression excluding EXPY (column 7), only years of schooling and initial income remain significant, but coefficient on initial income drops sharply; with EXPY included (column 8), years of schooling becomes insignificant, law and order negative at 10 percent significance, while EXPY and initial income are strongly significant with larger EXPY coefficients.
- Robustness and interpretation summary:
  - Neighbor-average IVs generally pass strength tests in many specifications and produce a consistently positive and economically meaningful effect of export sophistication on subsequent growth.
  - CLR-based inference indicates the coefficient on EXPY is positive and statistically significant in baseline IVs (range 0.06 to 0.09 in one CLR result; point estimates larger in other specifications).
  - Authors interpret results as suggesting that improvements in human capital, trade, financial development, or institutions would raise growth primarily to the extent they raise export sophistication (i.e., the channel of export sophistication is central).

*Source: wp18117 - References (IMF working paper content provided).*

### 0.15 on export sophistication. In these two regressions, we no longer obtain favorable statistics

### wp18117 - 0.15 on export sophistication. In these two regressions, we no longer obtain favorable statistics

### Instrument strength and identification
- Including several endogenous explanatory variables in the growth regression likely produces the weak instrument problem because endogenous variables and instruments are probably correlated, resulting in weak identification.
- The coefficient on export sophistication is statistically significant at the 1 percent level in the specification that includes multiple endogenous variables.
- Fixed effects IV regression (differenced equation, dependent variable = change in growth rate) also suffers from the weak instrument problem, suggesting it is harder to predict endogenous variables that are growth rates rather than levels using neighbors’ variables as instruments.
- Note: reported test statistics are based on 2-3 endogenous variables from Stock and Yogo (2005) while the regressions have a total of 6 endogenous variables; thus thresholds used are relatively conservative.

### Robustness: alternative proxies for export sophistication
- Alternative proxies tested: manufactures exports as a share of merchandise exports; log real manufacturing exports per capita.
- Using the baseline regression:
  - Both the share of manufacturing exports and log real manufacturing exports per capita have significant and positive coefficients (Table 3, columns 2 and 4).
  - Including EXPY and another measure of manufacturing exports produces quasi-multicollinearity and insignificant coefficients (columns 3 and 5); weak instrument tests indicate weak instruments in these regressions.
- Adding controls for manufacturing production:
  - Manufacturing value added (% of GDP) is statistically significant in regressions with EXPY (Table 3, columns 6-7).
  - Log real manufacturing value added per capita is not statistically significant in those columns (columns 8-9).
  - With other proxies for export sophistication, manufacturing value added (% of GDP) becomes statistically insignificant while log real manufacturing value added per capita coefficient is negative (columns 10-11), appearing to pick up the effect of initial income.
- Overall: export sophistication proxies have positive and statistically significant estimates across regressions, suggesting export orientation matters for growth and that producing manufacturing (potentially sophisticated) goods without exporting them may not be sufficient to increase long-run growth.

### Spillover effects from neighbors
- Controlled for average logarithm of real GDP (or real GDP per capita) in neighboring countries to capture spillovers and mitigate potential exclusion restriction violations.
- Instruments:
  - For neighbors’ GDP/GDP per capita, used average GDP/GDP per capita in neighbors’ neighbors.
  - Used neighbors’ lagged real GDP per capita to instrument for initial income.
- Results:
  - Spillover measured by neighbors’ real GDP per capita is statistically significant at the 10 percent level (Table 4, column 1); excluding schooling it becomes statistically insignificant (column 2).
  - In the baseline (Table 4, column 2), coefficient of spillover measured by average real GDP of neighbors is positive and strongly significant while EXPY is no longer significant (column 3).
  - Excluding schooling yields a positive and statistically significant coefficient on EXPY (Table 4, column 4).
- Instrument weighting schemes:
  - Coefficient for spillover measured by real GDP per capita of neighbors is statistically insignificant when using median or weighted averages of neighbors for instruments (columns 5-6 and 9-10).
  - Estimates using real GDP of neighbors as the spillover effect are similar across weighting schemes (columns 7-8 and 11-12).
  - Tests indicate instruments are not weak when using real GDP as the spillover measure.
- Conclusion: the spillover effect, even if present, does not invalidate the finding that export sophistication is a key growth determinant.

### Additional controls and robustness (Table 5, Table 6)
- Additional explanatory variables tested in baseline regression: investment to GDP ratio, national saving rate, FDI, government consumption, Gini coefficient, corruption.
- Export sophistication coefficient:
  - Varies in the range of 0.05-0.08 across these robustness checks and remains statistically significant in line with previous results.
  - Using IMF’s Ding and Hadzi-Vaskov (2017) measure of export sophistication yields a robust and positive estimate larger than previous regressions (Table 5, column 2).
  - Structural EXPY (S-EXPY, corrects for share of commodities) produces a strong positive coefficient that is two to three times smaller in magnitude than original EXPY (Table 5, column 1).
- GMM-system estimation (Table 6):
  - Standard GMM instruments: two lags of explanatory variables (levels in difference equation, DIF) and one lagged difference in level equation (LEV).
  - Added neighbor averages of explanatory variables as exogenous IV instruments and experimented with sharp instruments as GMM instruments.
  - Export sophistication has a highly significant and positive coefficient in all GMM specifications; magnitude remains relatively stable around 0.05-0.1 and is similar to IV regressions.
- Dissecting GMM (Table 7):
  - Level-equation GMM (GMM-LEV) with a smaller/collapsed number of instruments (Table 7, column 7) does not suffer from weak instruments and produces a positive and statistically significant coefficient on EXPY similar to IV estimates.
  - Using own explanatory variables and sharp instruments as GMM instruments results in weak instruments (columns 8-9).
  - Identification appears to come from sharp instruments used as IV instruments (column 5) where EXPY is statistically significant at 10 percent, but weak instruments remain a concern.
  - Tests: J-test (column 7) suggests overidentifying restrictions are not valid; Hahn, Ham, and Moon (2011) test indicates instruments based on neighbors’ averages are valid (p-value ≈ 0.7).
  - Difference-equation GMM specifications (columns 1-5) show weak instruments; level-equation (cross-country variation) produces parameter identification and favorable test statistics.
- Fixed effects:
  - If fixed effects are correlated with neighbors’ variables, differencing is required to remove fixed effects; however, differenced/time-series variation yields weak instruments and non-robust parameter estimates.
  - Table 2, column 9 and Table 4, column 12 (not reproduced here) considering fixed effects report coefficients on export sophistication similar to other estimations.

### Panels, instrument variants, and additional checks
- Using neighbors’ averages instrument set in 10-year panel regressions: export sophistication statistically significant in most estimations; coefficients similar in magnitude; weak instrument tests suggest instruments are as strong as in 5-year panel in several estimations.
- Alternative instrument aggregation: median and weighted mean (weights = inverse of real GDP) produce broadly unchanged results relative to simple averages (Tables 9-10).
  - Using inverse-GDP weights reduces the influence of the size of neighboring countries.
  - Although weak instrument tests are unfavorable in a few estimations, overall robustness holds.

### Key quantitative findings (selected exact values from tables and text)
- Coefficient range for log export sophistication across robustness checks: 0.05–0.08 (text summary).
- Specific reported coefficients and statistics (examples from tables):
  - Table 3 (selected): Log export sophistication coefficients include 0.065*** (column 1), 0.053 (column 2), 0.209 (column 3), 0.063*** (column 4), 0.083*** (column 5).
  - Log real GDP per capita coefficients in Table 3 vary, e.g., -0.024*** (column 1), -0.012*** (column 2), -0.026* (column 3).
  - Cragg-Donald F stat examples (Table 3): 40.7 (column 1), 109.7 (column 2), 2.6 (column 3), 13.0 (column 4).
  - Kleibergen-Paap F stat examples (Table 3): 38.4 (column 1), 100.0 (column 2), 1.7 (column 3).
  - Table 5 (selected): Log export sophistication coefficients include 0.084*** (column 1), 0.065*** (column 2), 0.043** (column 4), 0.109*** (column 5).
  - Table 6 (GMM) examples: Log export sophistication coefficients around 0.061**, 0.067***, 0.078***, 0.053***, 0.098*** (columns shown).
  - Table 7 (Dissecting GMM): Log export sophistication coefficients include 0.052*** (column 1), 0.059*** (column 2), 0.084*** (column 3), 0.111*** (column 4), 0.098** (column 5).
- Robust standard errors and significance: robust standard errors reported in parentheses; significance codes: *** p<0.01, ** p<0.05, * p<0.1.

### Interpretation and implications (as stated in the source)
- Export sophistication is a robust, positive, and economically meaningful determinant of long-run growth across multiple specifications, measures, and estimation methods.
- Export orientation appears important for growth; producing manufacturing and potentially sophisticated goods without exporting them may not be sufficient to increase long-run growth.
- Identification is stronger in level-equation (cross-country variation) estimations than in difference-equation (time-series variation) estimations; weak instruments are a recurrent concern when differenced variables or many endogenous variables are included.

*Source: wp18117 - 0.15 on export sophistication. In these two regressions, we no longer obtain favorable statistics*

### conclusion remains the same (Table 10). Export sophistication is still a key determinant in

### wp18117 - conclusion remains the same (Table 10). Export sophistication is still a key determinant in

### Main empirical findings
- Export sophistication stands out as an important and robust determinant of growth across multiple IV specifications.
- Representative estimated effects of log export sophistication on average annual growth:
  - 0.036*** (0.011)
  - 0.038*** (0.011)
  - 0.076*** (0.010)
  - 0.085*** (0.013)
  - 0.062*** (0.010)
  - 0.074*** (0.013)
  - 0.081*** (0.020)
  - 0.061*** (0.016)
  - 0.031** (0.013)
  - 0.269 (0.444)
  - 0.157** (0.072)
- Log real GDP per capita typically enters negatively in growth regressions:
  - -0.018*** (0.004)
  - -0.025*** (0.004)
  - -0.030*** (0.005)
  - -0.021*** (0.004)
  - -0.062** (0.026)
  - -0.135** (0.059)
- Other covariates show mixed or weak robustness:
  - Years of schooling: small positive or insignificant estimates (e.g., 0.003*** (0.001), 0.002 (0.001))
  - Trade (% of GDP): sometimes negative and occasionally significant (e.g., -0.042** (0.021) in one specification)
  - Credit to private sector (% of GDP): mixed signs and generally insignificant (e.g., -0.033 (0.025), -0.021* (0.013))
  - Law and order: mixed signs and occasional significance (e.g., -0.018** (0.007), -0.013* (0.008))

### Instrument strategy and diagnostics
- Instruments: each standard factor of growth is instrumented by its average in neighboring countries (neighbor averages) to create “sharp” instruments intended to avoid the “blunt instrument” problem.
- Instrument strength and diagnostic statistics reported across specifications:
  - Examples of Cragg-Donald F stat: 1.9; 9.7; 147.6; 260.4; 313.1; 260.4; 260.4
  - Examples of Kleibergen-Paap F stat: 2.5; 10.4; 103.7; 192.8; 232.7; 72.8
  - Kleibergen-Paap LM test p-values: 0.01; 0.00; 0.00; 0.00; 0.00; 0.00
  - Hansen J-test p-values where reported: 0.00; 0.03; 0.47; 0.31; 0.52; 0.11; 0.02
  - CLR bounds for H0: Beta_EXPY=0: Lower CLR bound 0.04; Upper CLR bound 0.12; H0: Beta_EXPY=0 | CLR p-value 0.00
- Observations in different panels and specifications include: 1,080; 1,089; 944; 1,216; 724; 596; 627; 618; 365; 306; 301; 363; 1,590; 1,216; 1,369; 1,319; 748; 606; 598; 983.

### Methodology details
- Instrument construction: neighbor-country averages of each endogenous standard growth determinant are used as instruments to generate variable-specific, dynamic instruments that are plausibly exogenous.
- The paper opens up the GMM approach and addresses concerns raised by Bazzi and Clemens (2013) regarding “black box” GMM.
- Multiple instrument sets are evaluated (median, weighted mean, and other variants) and robustness is assessed across 5-year and 10-year panels.

### Export-sophistication measures and data
- Primary export-sophistication index used: EXPY, constructed from product-level PRODY following Hausmann, Hwang, and Rodrik (2007).
- Alternative/adjusted index: S-EXPY (structural EXPY) excludes small islands and economies with population less than 2 million and scales down country per-capita GDP contributions by (1 − Commodity_Share_j) to adjust for primary commodity/export composition biases.
- Data sources:
  - Trade data from United Nations Commodity Trade Statistics Database (COMTRADE), converted to World Trade Flow databases based on 4-digit SITC rev.2 using Feenstra and Romalis’s (2014) codes.
  - Per-capita real GDP data from Penn World Tables 9.0, covering 182 countries.
  - Sample coverage: 1962 to 2014.

### Interpretation and mechanisms
- Correlations among neighboring countries’ variables are consistent with mimetic forces and cross-border diffusion of technologies and policies.
- Regional “pioneer” dynamics are plausible: a distant pioneer can achieve high sustained growth and trigger regional spillovers with neighboring countries benefiting from diffusion of policies and experience.
- While standard growth determinants are not uniformly robust in the regressions, they may matter insofar as they help improve export sophistication.

### Policy implications and recommendations
- If export sophistication is the main engine of growth, development strategies should prioritize:
  - Expansion of the tradable sector and the sophistication of exports.
  - Purpose-specific policies to develop tradable sophisticated sectors rather than relying solely on a general “enabling environment.”
- Recognize constraints in developing countries (skills, resources) and design targeted interventions to foster more sophisticated tradable outputs.
- The proposed neighbor-average instrument technique can be applied to other empirical contexts facing the blunt instrument problem, offering variable-specific, dynamic, and plausibly valid instruments.

*Source: IMF Working Paper wp18117 (conclusion and Tables 8–10, Appendix on EXPY/S-EXPY and data).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp18117.pdf_