## wpiea2021192-print-pdf

## Source details

**Canonical URL:** [wpiea2021192-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021192-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021192-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021192-print-pdf.pdf.json)

---

### I. Introduction and key contributions
- Constructs a novel sentiment-enhanced corruption perception index (SECPI) for 111 countries from 2005-18.
- SECPI combines:
  - the number of corruption-related Financial Times (FT) articles over 2005–18, and
  - in-depth sentiment analysis of these articles using the Loughran-McDonald financial dictionary.
- Advantages of SECPI relative to CPI (Transparency International) and CoC (Control of Corruption, Kaufmann et al. 2010):
  - Reacts quickly to current events (e.g., corruption investigations, election years).
  - Can be updated at higher frequency due to daily FT publishing.
  - Much lower cost to construct and update relative to survey-based indexes.
- Correlation with existing measures:
  - Correlation coefficients between SECPI and CPI/CoC are around -0.9 (negative sign reflects index construction: more negative perception → SECPI increases while CoC and CPI decrease).

### II. Data sources and sample coverage
- FT online archive over 2005–18: over 800,000 articles.
- Corruption-related selection via keywords starting with “corrupt” and “brib” (e.g., “corruption”, “corrupted”, “bribery”).
- Selected database:
  - 27,935 articles covering 182 countries (47 countries have at least one selected article every year).
  - Benchmark sample restricted to countries with at least one article per year in at least seven of the 14 years → results in 111 countries.
- Limitations noted:
  - FT-only coverage may reflect FT editors’ opinions and focus on big events; may underrepresent local/petty corruption.
  - Country coverage is uneven; some low-income countries have fewer than ten corruption-related articles during 2005–18.
  - Database contains only English articles.

### III. Summary statistics (by country group)
- Advanced economy
  - # of articles: 1,420,344
  - # of corruption articles: 10,431
  - Ratio (%): 0.73
- Emerging market
  - # of articles: 477,953
  - # of corruption articles: 22,028
  - Ratio (%): 4.61
- Low income countries
  - # of articles: 53,182
  - # of corruption articles: 4,731
  - Ratio (%): 8.90

### IV. Sentiment analysis methods and validation
- Approach: dictionary-based sentiment analysis using the Loughran-McDonald Financial dictionary as benchmark.
- Rationale: includes economic and financial terminology appropriate for FT news articles.
- Validation and robustness:
  - Alternative lexicons used for robustness checks: Harvard-IV, Henry’s Financial dictionary, Qdap dictionaries, and variants using negative-only lists.
  - Human audits used to validate automatic sentiment classifications.
- Sentiment scoring notes:
  - Scores calculated by matching lexicon words to article text and aggregating matched-word sentiment values.
  - Benchmark uses negative word list from Loughran-McDonald; alternative indexes constructed using other dictionaries and positive/negative lists.

### V. Construction of the SECPI (four steps)
- Step One:
  - Select corruption-related FT articles; restrict benchmark to countries with at least one article per year in at least seven of 14 years → 111 countries.
- Step Two:
  - For each article i, calculate benchmark sentiment score s_i by aggregating sentiment of individual words with sentiment scores (negative-word list of Loughran-McDonald).
- Step Three:
  - For country j at time t, aggregate sentiment scores of corruption-related articles and weight by total number of FT articles mentioning country j to adjust for uneven FT coverage.
  - Formula (verbatim symbol description retained in source): s_{t,j} = (sum over i of s_i * 1{article i at time t mentions country j}) / (sum over i of 1{article at time t mentions country j}).
- Step Four:
  - Convert s_{t,j} into a percentile I_{t,j} across all scores in all 14 years:
    - I_{t,j} = P_s( P( s < s_{t,j} ) )
  - Index bounded between zero and one: one = most negative corruption perception; zero = least negative corruption perception.
  - Percentile calculated using scores across all 14 years to capture cross-sectional and time variation.
  - Note: index values will slightly change when new articles are added because percentiles are recalculated.

### VI. Robustness results and alternative lexicons (correlations)
- Correlations between benchmark SECPI and alternative lexicon-based indexes across countries and time:
  - NegativityQDAP: 0.99
  - NegativityHE: 0.90
  - NegativityGI: 0.99
  - SentimentQDAP: 0.29
  - SentimentLM: 0.98
  - SentimentHE: 0.70
  - SentimentGI: 0.65
- General finding: benchmark SECPI is highly correlated with most alternative indexes except SentimentQDAP.

### VII. Human audits and classification findings
- Articles grouped into five categories from most negative (Category 1) to least negative (Category 5).
- Audit of 147 articles summarized (Table 3):
  - Category 1: Average Scores 0.140, Count 2
  - Category 2: Average Scores 0.109, Count 15
  - Category 3: Average Scores 0.066, Count 109
  - Category 4: Average Scores 0.066, Count 17
  - Category 5: Average Scores 0.057, Count 4
- Country-linking human audit:
  - Among 1,070 articles in the human audits, 830 articles are linked to the correct countries, 22 are wrongly linked and others can be linked to not just one country.
- Box 1 refinements for country extraction:
  - Built a dataset of countries and their major locations.
  - Use the most popular location-country pair when city names are shared by many countries.
  - Allow only one country linked to one article; link the most frequently mentioned country.

### VIII. Comparison with existing indexes, elections, and event responsiveness
- SECPI correlations:
  - SECPI negatively correlates with the CoC and the CPI across country groups; correlation coefficients are around -0.9 for each pair.
  - SECPI compared with NIC (Hlatshwayo, et al. 2018) yields a correlation coefficient around 0.65.
- Event responsiveness:
  - SECPI identifies long-term trends similar to other indexes but is more responsive to individual events in the near term.
  - Often spikes in election years.
    - The average value of the index during a legislative or executive election year is 0.05 higher than outside election years (controlling for country-specific effects).
  - No empirical evidence that the size of the spike depends on the overall level of corruption.
  - Some evidence that election-related spikes are higher (lower) when countries are experiencing lower (higher) growth than recent historical trends.
- Case study patterns:
  - Election-based spikes can coincide with regime change and permanent shifts in corruption perceptions.
  - Large-scale corruption investigations can produce sharp increases in SECPI during the year of the investigation and correlate with drops in stock market returns or FDI inflows.
  - Economic downturns can raise focus on corruption; SECPI can rise during crises and then revert as stability is restored.
  - Sustained sequences of investigations can lead to gradual increases in SECPI and be associated with dramatic declines in growth linked to political paralysis.

### IX. Macro-relevance: correlations and local projection methodology
- Structural correlations:
  - Negative corruption sentiment is significantly correlated with lower Regulatory Quality; lower Ease of Doing Business (Starting a Business); lower Global Competitiveness Index; weaker Checks and Balances; lower Prosperity Index; and higher Poverty.
  - Several correlations remain significant even after controlling for income per capita (e.g., Starting a Business, Regulatory Quality, Global Competitiveness Index, Polity Index, and the Prosperity Index).
- Local projection approach and model details:
  - Use Jorda (2005) local projections at an annual frequency with country and year fixed effects.
  - Lags p = 2 and horizon h = 5 years.
  - Create a “shock” dummy equal to one when corruption sentiment is more than one standard deviation above the mean (country-specific mean and standard deviation).
  - Controls: battery of possible controls covering prices and exchange rates; volatility; external and domestic demand; and structural variables (e.g., bureaucratic quality, democratic accountability, religious tensions); more than 25 controls plus transformations (levels, two lags, first differences, lagged first differences) and lagged dependent variables.
  - Variable selection: Belloni, et al.’s (2014) Lasso-based procedure used to select controls and address model uncertainty/endogeneity.
  - Robustness: include executive and legislative elections, both leads and lags of shocks; standard errors clustered at the country level.
  - Panel covers 111 countries and the years of 2007–18.
- Annex I: list of variables for local projections includes Real PPP GDP; Real Per Capita Growth; Current Account/GDP; General Government Expenditure/GDP; Net FDI/GDP; Exchange Rate Depreciation; Inflation Rate; Public External Debt/GDP; Total Debt/GDP; Population 10Y Growth; Polity Score; Checks & Balances Index; Bureaucracy Quality; Democratic Accountability; Ethnic Tensions; Religious Tensions; Monetary Union Membership, Dummy; and many other macro and structural indicators (sources as listed).

### X. Quantitative macroeconomic impacts and recovery dynamics
- Full panel GDP per capita growth impact:
  - Periods with negative corruption sentiment at least one standard deviation above countries’ mean are associated with an average drop in GDP per capita growth of 0.65 percentage point in t+1 (up from a hit of almost 0.50 in year 0).
  - The effect dissipates after t+1.
- Heterogeneity by country group:
  - Emerging market and low-income countries (EMLICs) see larger effects compared to advanced economies.
  - In advanced economies the impact of negative sentiment is not significant.
  - EMLICs also experience related falls in household consumption and private investment in percent of GDP.
- Comparison with related literature:
  - Hlatshwayo, et al. (2018) find shocks calibrated to be double the size considered here (two vs. one standard deviation) are associated with lower real GDP per capita growth of two percentage points, roughly double the size of this paper’s findings.
  - Ugur’s (2014) meta-analysis of 29 studies finds corruption has significant effects on low income countries while tending to have an insignificant effect in broader income groups.
- Recovery dynamics:
  - Economic growth recovers in the medium run after the negative corruption sentiment shock hits.
  - Recovery could reflect adoption of forceful anti-corruption measures; successful measures may improve business climate and rebuild confidence.
  - If anticorruption measures are unsuccessful, recovery could be prolonged.
  - Economic recovery is shown to be stronger in advanced economies than in EMLICs.

### XI. Conclusions and policy implications
- SECPI construction and features:
  - Sentiment-enhanced corruption perception index based on FT news articles over 2005–18 using sentiment analysis.
  - SECPI complements and correlates with existing corruption perception measures.
  - Highly sensitive to current events (e.g., corruption investigations and elections).
  - Can be made available at a higher frequency with relatively low costs.
- Policy-relevant findings:
  - SECPI correlates with governance, business climate, and poverty variables.
  - Corruption perception, as captured by SECPI, has sizable negative impact on growth, especially in emerging market and low-income countries.
  - High-frequency SECPI provides a useful tool for investigating short-term impacts of corruption on economies.
  - Techniques used can be applied to extract opinions on corruption from diverse data sources, including local newspapers and social media posts.

*Source: wpiea2021192-print-pdf - References (SECPI construction, data, methods, robustness, and macroeconomic findings).*

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

### wpiea2021192-print-pdf - References

### I. Introduction and key contributions
- Constructs a novel sentiment-enhanced corruption perception index (SECPI) for 111 countries from 2005-18.
- SECPI combines:
  - the number of corruption-related Financial Times (FT) articles over 2005–18, and
  - in-depth sentiment analysis of these articles using the Loughran-McDonald financial dictionary.
- Advantages of SECPI relative to CPI (Transparency International) and CoC (Control of Corruption, Kaufmann et al. 2010):
  - Reacts quickly to current events (e.g., corruption investigations, election years).
  - Can be updated at higher frequency due to daily FT publishing.
  - Much lower cost to construct and update relative to survey-based indexes.
- Correlation with existing measures:
  - Correlation coefficients between SECPI and CPI/CoC are around -0.9 (negative sign reflects index construction: more negative perception → SECPI increases while CoC and CPI decrease).

### II. Data sources and sample coverage
- FT online archive over 2005–18: over 800,000 articles.
- Corruption-related selection via keywords starting with “corrupt” and “brib” (e.g., “corruption”, “corrupted”, “bribery”).
- Selected database:
  - 27,935 articles covering 182 countries (47 countries have at least one selected article every year).
  - Benchmark sample restricted to countries with at least one article per year in at least seven of the 14 years → results in 111 countries.
- Limitations noted:
  - FT-only coverage may reflect FT editors’ opinions and focus on big events; may underrepresent local/petty corruption.
  - Country coverage is uneven; some low-income countries have fewer than ten corruption-related articles during 2005–18.
  - Database contains only English articles.

### III. Summary statistics (Table 1)
- By country group:
  - Advanced economy
    - # of articles: 1,420,344
    - # of corruption articles: 10,431
    - Ratio (%): 0.73
  - Emerging market
    - # of articles: 477,953
    - # of corruption articles: 22,028
    - Ratio (%): 4.61
  - Low income countries
    - # of articles: 53,182
    - # of corruption articles: 4,731
    - Ratio (%): 8.90

### IV. Sentiment analysis methods
- Approach: dictionary-based sentiment analysis using the Loughran-McDonald Financial dictionary as benchmark.
- Rationale for dictionary choice: includes economic and financial terminology appropriate for FT news articles.
- Validation and robustness:
  - Alternative lexicons used for robustness checks: Harvard-IV, Henry’s Financial dictionary, Qdap dictionaries, and variants using negative-only lists.
  - Human audits used to validate automatic sentiment classifications.
- Notes on sentiment scoring:
  - Sentiment scores calculated by matching lexicon words to article text and aggregating matched-word sentiment values.
  - Benchmark uses negative word list from Loughran-McDonald; alternative indexes constructed using other dictionaries and positive/negative lists.

### V. Construction of the SECPI (four steps)
- Step One:
  - Select corruption-related FT articles; restrict benchmark to countries with at least one article per year in at least seven of 14 years → 111 countries.
- Step Two:
  - For each article i, calculate benchmark sentiment score s_i by aggregating sentiment of individual words with sentiment scores (negative-word list of Loughran-McDonald).
- Step Three:
  - For country j at time t, aggregate sentiment scores of corruption-related articles and weight by total number of FT articles mentioning country j to adjust for uneven FT coverage.
  - Formula (verbatim symbol description retained in source): s_{t,j} = (sum over i of s_i * 1{article i at time t mentions country j}) / (sum over i of 1{article at time t mentions country j}).
- Step Four:
  - Convert s_{t,j} into a percentile I_{t,j} across all scores in all 14 years:
    - I_{t,j} = P_s( P( s < s_{t,j} ) )
  - Index bounded between zero and one: one = most negative corruption perception; zero = least negative corruption perception.
  - Percentile calculated using scores across all 14 years to capture cross-sectional and time variation.
  - Note: index values will slightly change when new articles are added because percentiles are recalculated.

### VI. Robustness results and alternative lexicons (Table 2)
- Correlations between benchmark SECPI and alternative lexicon-based indexes across countries and time:
  - NegativityQDAP: 0.99
  - NegativityHE: 0.90
  - NegativityGI: 0.99
  - SentimentQDAP: 0.29
  - SentimentLM: 0.98
  - SentimentHE: 0.70
  - SentimentGI: 0.65
- General finding: benchmark SECPI is highly correlated with most alternative indexes except SentimentQDAP.

### VII. Macroeconomic impacts and empirical findings
- SECPI is negatively associated with growth and structural variables:
  - A more negative perception of corruption (higher SECPI) is associated with lower GDP per capita growth.
  - When corruption perception is at least one standard deviation above the mean, GDP per capita growth is on average 0.65 percentage point lower cumulatively by the second year.
  - The immediate impact is 0.5%.
- Heterogeneity by country income:
  - The negative growth effect is especially significant in emerging market and low-income countries.
  - In emerging market and low-income countries, household consumption and private investment in percent of GDP also fall when SECPI indicates more negative corruption perception.
- Consistency with literature:
  - Findings align with established channels where corruption/perceptions of corruption harm tax compliance, public spending efficiency, increase borrowing costs, reduce trust and financial market participation, and adversely affect investment and productivity.

### VIII. Methodological notes, validations, and limitations
- Human audits: used to design guidance and validate sentiment categorization; articles grouped into five categories from most negative to least negative (taxonomy described in source).
- Validation across dictionaries and human reading indicates limited type I and type II errors for the keywords-based selection strategy.
- Limitations reiterated:
  - FT coverage uneven across countries and biased toward large events and English-language reporting.
  - SECPI is not an IMF-approved or recommended measure; presented for research purposes.

_italic Source: wpiea2021192-print-pdf - References (SECPI construction, data, methods, robustness, and macroeconomic findings)._

### 1. The articles on confirmed corruption with big economic or political impacts.

### 1. The articles on confirmed corruption with big economic or political impacts.

### Classification and human audits
- Articles were grouped into five categories based on the negativity of corruption identified, from category 1 (most negative) to category 5 (least negative).
- The paper audited 147 articles and grouped them into the five categories.
- Table 3. Summary of Human Audits in Quantifying Sentiment
  - Category 1: Average Scores 0.140, Count 2
  - Category 2: Average Scores 0.109, Count 15
  - Category 3: Average Scores 0.066, Count 109
  - Category 4: Average Scores 0.066, Count 17
  - Category 5: Average Scores 0.057, Count 4
- SentimentQDAP uses the dictionary built by Hu and Liu (2004) based on customer reviews, which contains a different positive word list than others.
- Country-linking human audit for FT articles:
  - Among 1,070 articles in the human audits, 830 articles are linked to the correct countries, 22 are wrongly linked and others can be linked to not just one country.
- Box 1 refinements for country extraction:
  - Built a dataset of countries and their major locations.
  - Use the most popular location-country pair when city names are shared by many countries.
  - Allow only one country linked to one article; link the most frequently mentioned country.

### Comparison with existing indexes
- SECPI correlations:
  - SECPI negatively correlates with the CoC and the CPI across country groups.
  - Correlation coefficients are around -0.9 for each pair.
  - SECPI compared with NIC (Hlatshwayo, et al. 2018) yields a correlation coefficient around 0.65.
- Cross-group similarity:
  - All three indexes show that average corruption perception is less negative in advanced economies (AEs) than in emerging market (EMs) and low-income countries (LICs).

### Elections, event responsiveness, and case studies
- SECPI characteristics:
  - Identifies long-term trends similar to other indexes but is more responsive to individual events in the near term.
  - Reflects most of the change during the year of singular events (e.g., start of a corruption investigation or change in political leadership), whereas other indexes reflect changes more gradually and/or with a lag.
  - Often spikes in election years.
    - The average value of the index during a legislative or executive election year is 0.05 higher than outside election years (controlling for country-specific effects).
    - Election-year spikes appear when corruption is a key campaign issue and receives additional news coverage.
    - No empirical evidence that the size of the spike depends on the overall level of corruption.
    - Some evidence that election-related spikes are higher (lower) when countries are experiencing lower (higher) growth than recent historical trends.
- Case study findings (Box 2) — illustrative patterns:
  - Election-based spikes can coincide with regime change and permanent shifts in corruption perceptions.
  - Large-scale corruption investigations can produce sharp increases in SECPI during the year of the investigation and correlate with drops in stock market returns or FDI inflows.
  - Economic downturns can raise focus on corruption as a cause of fiscal and revenue problems; SECPI can rise during crises and then revert as stability is restored.
  - Sustained sequences of investigations can lead to gradual increases in SECPI and be associated with dramatic declines in growth linked to political paralysis.

### Macro-relevance: correlations and local projection methodology
- Structural correlations:
  - Negative corruption sentiment is significantly correlated with lower Regulatory Quality; lower Ease of Doing Business (Starting a Business); lower Global Competitiveness Index; weaker Checks and Balances; lower Prosperity Index; and higher Poverty.
  - Pearson’s correlation tests for these relationships are statistically significant.
  - Several correlations remain significant even after controlling for income per capita (e.g., Starting a Business, Regulatory Quality, Global Competitiveness Index, Polity Index, and the Prosperity Index).
- Local projection approach and model details:
  - Use Jorda (2005) local projections at an annual frequency with country and year fixed effects.
  - Lags p = 2 and horizon h = 5 years.
  - Create a “shock” dummy equal to one when corruption sentiment is more than one standard deviation above the mean (country-specific mean and standard deviation).
  - Controls: battery of possible controls covering prices and exchange rates; volatility; external and domestic demand; and structural variables (e.g., bureaucratic quality, democratic accountability, religious tensions); more than 25 controls plus transformations (levels, two lags, first differences, lagged first differences) and lagged dependent variables.
  - Variable selection: Belloni, et al.’s (2014) Lasso-based procedure used to select controls and address model uncertainty/endogeneity.
  - Robustness: include executive and legislative elections, both leads and lags of shocks; standard errors clustered at the country level.
  - Panel covers 111 countries and the years of 2007–18.

### Quantitative macroeconomic impacts
- Full panel GDP per capita growth impact:
  - Periods with negative corruption sentiment at least one standard deviation above countries’ mean are associated with an average drop in GDP per capita growth of 0.65 percentage point in t+1 (up from a hit of almost 0.50 in year 0).
  - The effect dissipates after t+1.
- Heterogeneity by country group:
  - Emerging market and low-income countries (EMLICs) see larger effects compared to advanced economies.
  - In advanced economies the impact of negative sentiment is not significant.
  - EMLICs also experience related falls in household consumption and private investment in percent of GDP.
- Comparison with related literature:
  - Hlatshwayo, et al. (2018) find shocks calibrated to be double the size considered here (two vs. one standard deviation) are associated with lower real GDP per capita growth of two percentage points, roughly double the size of this paper’s findings.
  - Ugur’s (2014) meta-analysis of 29 studies finds corruption has significant effects on low income countries while tending to have an insignificant effect in broader income groups.
- Recovery dynamics:
  - Economic growth recovers in the medium run after the negative corruption sentiment shock hits.
  - Recovery could reflect adoption of forceful anti-corruption measures; successful measures may improve business climate and rebuild confidence.
  - If anticorruption measures are unsuccessful, recovery could be prolonged.
  - Economic recovery is shown to be stronger in advanced economies than in EMLICs.

### Conclusions and implications
- SECPI construction and features:
  - Sentiment-enhanced corruption perception index based on FT news articles over 2005–18 using sentiment analysis.
  - SECPI complements and correlates with existing corruption perception measures.
  - Highly sensitive to current events (e.g., corruption investigations and elections).
  - Can be made available at a higher frequency with relatively low costs.
- Policy-relevant findings:
  - SECPI correlates with governance, business climate, and poverty variables.
  - Corruption perception, as captured by SECPI, has sizable negative impact on growth, especially in emerging market and low-income countries.
  - High-frequency SECPI provides a useful tool for investigating short-term impacts of corruption on economies.
  - Techniques used can be applied to extract opinions on corruption from diverse data sources, including local newspapers and social media posts.

*Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021192-print-pdf.pdf*

### References

### wpiea2021192-print-pdf - References

### Key bibliographic sources (selected)
- Achury, Carolina, Christos Koulovatianos, and John Tsoukalas. Political economics of external sovereign defaults. No. 508. Center for Financial Studies (CFS), 2015.
- Agarwal, Apoorv, Fadi Biadsy, and Kathleen Mckeown. "Contextual phrase-level polarity analysis using lexical affect scoring and syntactic n-grams." In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pp. 24-32. 2009.
- Aghion, Philippe, Ufuk Akcigit, Julia Cagé, and William R. Kerr. "Taxation, corruption, and growth." European Economic Review 86 (2016): 24-51.
- Akitoby, Bernardin, and Thomas Stratmann. "The value of institutions for financial markets: evidence from emerging markets." Review of World Economics 146, no. 4 (2010): 781-797.
- Belloni, Alexandre, Victor Chernozhukov, and Christian Hansen. "Inference on treatment effects after selection among high-dimensional controls." The Review of Economic Studies 81, no. 2 (2014): 608-650.
- Burscher, Bjorn, Rens Vliegenthart, and Claes H. de Vreese. "Frames beyond words: Applying cluster and sentiment analysis to news coverage of the nuclear power issue." Social Science Computer Review 34, no. 5 (2016): 530-545.
- Cambria, Erik, Björn Schuller, Yunqing Xia, and Catherine Havasi. "New avenues in opinion mining and sentiment analysis." IEEE Intelligent systems 28, no. 2 (2013): 15-21.
- Campos, Nauro F., Ralitza D. Dimova, and Ahmad Saleh. "Whither corruption? A quantitative survey of the literature on corruption and growth.” Discussion Paper No. 5334 (Bonn: Institute for the Study of Labor), 2010.
- Cavoli, Tony, and John K. Wilson. "Corruption, central bank (in) dependence and optimal monetary policy in a simple model." Journal of Policy Modeling 37, no. 3 (2015): 501-509.
- Deng, Shuyuan, Atish P. Sinha, and Huimin Zhao. "Adapting sentiment lexicons to domain-specific social media texts." Decision Support Systems 94 (2017): 65-76.
- Eshbaugh-Soha, Matthew. "The tone of local presidential news coverage." Political Communication 27, no. 2 (2010): 121-140.
- Fayad, Ghada, Chengyu Huang, Yoko Shibuya, and Peng Zhao. "How do member countries receive IMF policy advice: results from a state-of-the-art sentiment index." (2020).
- Fazekas, Mihály, and Gábor Kocsis. "Uncovering high-level corruption: cross-national objective corruption risk indicators using public procurement data." British Journal of Political Science 50.1 (2020): 155-164.
- González-Bailón, Sandra, and Georgios Paltoglou. "Signals of public opinion in online communication: A comparison of methods and data sources." The ANNALS of the American Academy of Political and Social Science 659, no. 1 (2015): 95-107.
- Grimmer, Justin, and Brandon M. Stewart. "Text as data: The promise and pitfalls of automatic content analysis methods for political texts." Political analysis 21.3 (2013): 267-297.
- Guiso, Luigi, Paola Sapienza, and Luigi Zingales. “Cultural biases in Economic Exchange?” The Quarterly Journal of Economics, Vol. 124 (3) (2009): 1095-1131.
- Gupta, Sanjeev, Hamid Davoodi, and Rosa Alonso-Terme. "Does corruption affect income inequality and poverty?" Economics of Governance 3, no. 1 (2002): 23-45.
- Hlatshwayo, Sandile, Anne Oeking, Manuk Ghazanchyan, David Corvino, Ananya Shukla, and Lamin Y. Leigh. The Measurement and Macro-Relevance of Corruption: A Big Data Approach. No. 2018/195. International Monetary Fund, 2018.
- Hopkins, Daniel J., and Gary King. "A method of automated nonparametric content analysis for social science." American Journal of Political Science 54, no. 1 (2010): 229-247.
- Hu, Minqing, and Bing Liu. "Mining and summarizing customer reviews." In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 168-177. 2004.
- Huang, Haizhou, and Shang-Jin Wei. "Monetary policies for developing countries: The role of institutional quality." Journal of International Economics 70, no. 1 (2006): 239-252.
- Kaufmann, Daniel. "Can corruption adversely affect public finances in industrialized countries." Brookings Institution, April 19 (2010).
- Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi (2010). "The Worldwide Governance Indicators: Methodology and Analytical Issues". World Bank Policy Research Working Paper No. 5430.
- Krosnick, Jon A. "Survey research." Annual review of psychology 50.1 (1999): 537-567.
- Liu, Bing. Sentiment analysis: Mining opinions, sentiments, and emotions. Cambridge university press, 2020.
- Loughran, Tim, and Bill McDonald. "When is a liability not a liability? Textual analysis, dictionaries, and 10‐Ks." The Journal of Finance 66, no. 1 (2011): 35-65.
- Manning, Christopher, and Hinrich Schutze. Foundations of statistical natural language processing. MIT press, 1999.
- Mithe, Ravina, Supriya Indalkar, and Nilam Divekar. "Optical character recognition." International journal of recent technology and engineering (IJRTE) 2, no. 1 (2013): 72-75.
- Monroe, Burt L., Michael P. Colaresi, and Kevin M. Quinn. "Fightin'words: Lexical feature selection and evaluation for identifying the content of political conflict." Political Analysis 16, no. 4 (2008): 372-403.
- Nguyen, Thien Hai, and Kiyoaki Shirai. "Topic modeling based sentiment analysis on social media for stock market prediction." In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 1354-1364. 2015.
- Nguyen, Thien Hai, Kiyoaki Shirai, and Julien Velcin. "Sentiment analysis on social media for stock movement prediction." Expert Systems with Applications 42, no. 24 (2015): 9603–9611.
- Oliveira, Nuno, Paulo Cortez, and Nelson Areal. "The impact of microblogging data for stock market prediction: Using Twitter to predict returns, volatility, trading volume and survey sentiment indices." Expert Systems with Applications 73 (2017): 125-144.
- Pang, Bo, and Lillian Lee. "A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts." arXiv preprint cs/0409058 (2004).
- Rajkumar, Andrew Sunil, and Vinaya Swaroop. "Public spending and outcomes: Does governance matter?" Journal of Development Economics 86, no. 1 (2008): 96-111.
- Sapienza, Paola, and Luigi Zingales. "A trust crisis." International Review of Finance 12, no. 2 (2012): 123-131.
- Soroka, Stuart N. "The gatekeeping function: Distributions of information in media and the real world." The Journal of Politics 74, no. 2 (2012): 514-528.
- Soroka, Stuart N., Dominik A. Stecula, and Christopher Wlezien. "It's (change in) the (future) economy, stupid: economic indicators, the media, and public opinion." American Journal of Political Science 59, no. 2 (2015): 457-474.
- Soroka, Stuart, Lori Young, and Meital Balmas. "Bad news or mad news? Sentiment scoring of negativity, fear, and anger in news content." The ANNALS of the American Academy of Political and Social Science 659, no. 1 (2015): 108-121.
- Tanzi, Vito, and Hamid Davoodi. "Corruption, public investment, and growth." In The welfare state, public investment, and growth, pp. 41-60. Springer, Tokyo, 1998.
- Turney, Peter D. "Thumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews." arXiv preprint cs/0212032 (2002).
- Ugur, Mehmet. "Corruption's direct effects on per‐capita income growth: a meta‐analysis." Journal of Economic Surveys 28, no. 3 (2014): 472-490.
- Van Atteveldt, Wouter, Jan Kleinnijenhuis, Nel Ruigrok, and Stefan Schlobach. "Good news or bad news? Conducting sentiment analysis on Dutch text to distinguish between positive and negative relations." Journal of Information Technology & Politics 5, no. 1 (2008): 73-94.
- Wilson, Theresa, Janyce Wiebe, and Paul Hoffmann. "Recognizing contextual polarity in phrase-level sentiment analysis." In Proceedings of human language technology conference and conference on empirical methods in natural language processing, pp. 347-354. 2005.
- Young, Lori, and Stuart Soroka. "Affective news: The automated coding of sentiment in political texts." Political Communication 29, no. 2 (2012): 205-231.

### Annex I. List of Variables for Local Projections
- The sources of the data are in parentheses.
- Real PPP GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Real Per Capita Growth (World Economic Outlook)
- Current Account/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- General Government Expenditure/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- General Government Interest Expenditure/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- General Government Revenues/ GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Government Consumption/ GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Government Investment/ GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Net FDI/ GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Exports/ GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Imports/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Trade Openness 10Y Average (IMF Vulnerability Exercise Fiscal Model Database)
- Trading Partner Growth (IMF Vulnerability Exercise Fiscal Model Database)
- Partner Import Demand (IMF Vulnerability Exercise Fiscal Model Database)
- Real Effective Exchange Rate (IMF Vulnerability Exercise Fiscal Model Database)
- Exchange Rate Depreciation (IMF Vulnerability Exercise Fiscal Model Database)
- Exchange Rate, End of Period (IMF Vulnerability Exercise Fiscal Model Database)
- PPP Exchange Rate (IMF Vulnerability Exercise Fiscal Model Database)
- Terms of Trade Inflation (IMF Vulnerability Exercise Fiscal Model Database)
- Inflation Rate (World Economic Outlook)
- Reserves Growth, LCU (IMF Vulnerability Exercise Fiscal Model Database)
- Remittances/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Public External Debt/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Public External Debt/Exports (IMF Vulnerability Exercise Fiscal Model Database)
- Public Debt/ Revenues (IMF Vulnerability Exercise Fiscal Model Database)
- Public Debt/GDP, 5y Change (IMF Vulnerability Exercise Fiscal Model Database)
- Total Debt/GDP (IMF Vulnerability Exercise Fiscal Model Database)
- Growth Deviation from 5y Avg (IMF Vulnerability Exercise Fiscal Model Database)
- RGDP Growth Volatility (IMF Vulnerability Exercise Fiscal Model Database)
- Terms of Trade Volatility (IMF Vulnerability Exercise Fiscal Model Database)
- Exchange Rate Volatility (IMF Vulnerability Exercise Fiscal Model Database)
- Inflation Volatility (IMF Vulnerability Exercise Fiscal Model Database)
- Population 10Y Growth (IMF Vulnerability Exercise Fiscal Model Database)
- Population 5Y Growth (IMF Vulnerability Exercise Fiscal Model Database)
- Population (IMF Vulnerability Exercise Fiscal Model Database)
- Nat. Disaster Impact Growth (IMF Vulnerability Exercise Fiscal Model Database)
- Polity Score (Center for Systemic Peace)
- Checks & Balances Index (IMF Vulnerability Exercise Fiscal Model Database)
- Bureaucracy Quality (ICRG)
- Democratic Accountability (ICRG)
- Ethnic Tensions (ICRG)
- Investment Prof (ICRG)
- Law & Order (ICRG)
- Military in Politics (ICRG)
- Religious Tensions (ICRG)
- Socioeconomic Conditions (ICRG)
- Monetary Union Membership, Dummy (IMF Vulnerability Exercise Fiscal Model Database)

- 19 Country Data Online, The PRS Group, Inc., www.prsgroup.com

*Source: wpiea2021192-print-pdf - References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021192-print-pdf.pdf_