## Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact

## Source details

**Canonical URL:** [Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact](https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022094-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2022/english/wpiea2022094-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2022/english/wpiea2022094-print-pdf.pdf.json)

---

### I. Introduction: purpose and scope
- Public procurement constitutes about 12 percent of global GDP or 11 trillion USD per year.
- Losses through procured spending are estimated to be about 10-20 percent, even in countries with relatively high integrity of their procurement systems in the European Union (Hafner et al., 2016).
- Main contribution: assess whether red flags of corrupt behaviors have an impact on prices of procured goods and services by estimating the impact of corruption risks (assessed through red flags) on relative prices (i.e., comparing prices paid with reference prices).
- Dataset: over 1.5 million contracts across five pilot countries; captures from 15 to 55 percent of total procured spending in each country.
- Five pilot countries analyzed: Georgia, Indonesia, Paraguay, Romania, and Uganda.

### II. Methodology: measuring corruption risks with objective data
- Corruption Risk Index (CRI) construction:
  - Seven objective indicators (“red flags”) are mapped and used to compute a composite CRI.
  - CRI is a simple average of the seven red flags after normalization.
  - Contract-level scores: 0 (lowest risk), 0.5 (medium risk), 1 (highest risk) for each red flag.
  - Example: for single bidding the contract-level score is 0 or 1; aggregate example: Romania 31 percent of contracts were awarded in tenders with a single bidder.
- The seven red flags:
  - Single bidder contracts.
  - Non-open procedures.
  - Lack of publication of call for tenders.
  - Period for submitting bids (short and excessively long submission periods flagged).
  - Period for selecting the winning bid (short or long decision periods flagged).
  - Spending concentration (by organization, by year).
  - Share of suppliers registered in jurisdictions offering limited company and banking transparency (tax-haven indicator).
- Scoring and aggregation: CRI equals the average of the seven normalized red-flag scores for a contract or aggregation unit.

### III. Data coverage and implementation details
- Data sources: government web portals for public procurement across the five pilot countries.
- Coverage: over 1.5 million contracts; captures from 15 to 55 percent of total procured spending in each country.
- Automation and processing:
  - Automated web crawler scraping HTML, XML, CSV; tools include Python and Java.
  - Publications transposed into a uniformly structured data template; linked publications to reconstruct each tendering cycle.
  - Manual cross-checking and standardization of buyers’ and suppliers’ names.
  - For Indonesia and Uganda: multi-step token-based string-matching algorithm for missing product codes.
  - Full technical documentation and codes located at: https://github.com/digiwhist/backend.

### IV. Strengths and limitations of the CRI
- Strengths:
  - CRI more robust than individual components because corruption typically exploits multiple vulnerabilities simultaneously.
  - Predominantly data-driven and informed by established theories of corruption.
  - Validity regressions (Appendix II) identify definitions aligned with corruption outcomes; selection of high-risk procedure types driven by association with single bidding.
  - Provides granular guidance for policy makers to explore sectoral vulnerabilities and factors driving price differentials.
- Limitations and caveats:
  - CRI measures corruption risks, not confirmed instances of corruption; associations with relative prices may reflect structural market features.
  - Quantitative approach does not assess effects on quality of procured goods and services.
  - Does not include capacity assessment due to data limitations and subjectivity concerns.
  - Cross-country comparisons should be interpreted cautiously because of procedural differences and idiosyncrasies.

### V. Regression methodology and specifications for price impact estimation
- Dependent variable: relative price (RP) = actual contract value divided by initially estimated contract value (or via savings if available).
- Five main regression models:
  - Model 1: CRI only.
  - Models 2–5: add controls for product market (CPV division, location, contract value), organizational framework (buyer type), and year.
  - Models 1–3 restrict RP to between 0.5 and 1.5.
  - Additional conservative regressions restrict RP to 0.5–1.
  - Model 5 allows quadratic specification for CRI (non-linearities).
  - Model 4 chosen as main prediction model (most robust, widest controls).
- Controls included across specifications: Year controls; Contract Value (100 quantiles); CPV division; Buyer type; Buyer location.
- Standard errors: robust, clustered over buyers. Significance notation: *** p<0.01, ** p<0.05, * p<0.1.
- Caveat: association ≠ causality; regressions control for market specificities via product market fixed effects.

### VI. Main regression results (selected coefficients and fit statistics)
- Example selected coefficients reported for Georgia (various models):
  - CRI: 0.276*** (standard error (0.004))
  - CRI: 0.315*** (0.004)
  - CRI: 0.312*** (0.004)
  - CRI: 0.312*** (0.004)
  - CRI: 0.222*** (0.012)
  - (CRI)^2: 0.116*** (0.016) — included in quadratic model.
- Sample sizes and fit (Georgia example across columns):
  - Observations: 188,472; 188,472; 188,472; 188,414; 188,414.
  - R-squared: 0.15; 0.20; 0.21; 0.21; 0.21.
- Robust standard errors in parentheses. Coefficients remain significant and largely same across specifications; non-linear models add little improvement relative to complexity.

### VII. Country-level elasticities and price impacts (Model 4 elasticities)
- Price elasticity to CRI (Elasticities taken from model 4):
  - Paraguay: 0.395
  - Georgia: 0.314
  - Romania: 0.314
  - Uganda: 0.10
  - Indonesia: 0.07
- Price impact of CRI increase (Percent) (an additional red flag (1/7 points increase on the CRI score) increases prices by 1/7 times the elasticity):
  - Paraguay: 5.5
  - Georgia: 4.5
  - Romania: 4.4
  - Uganda: 1.4
  - Indonesia: 1.0
- Illustration provided: Romania—an additional red flag (1/7 points increase on the CRI score) is predicted to increase prices by 4.4 percentage points (0.307*(1/7)*100=4.4).

### VIII. Alternative specifications and component analysis (selected examples)
- Alternative regressions estimate individual red-flag effects:
  - Georgia (RP, 0.5<RP≤1) selected coefficients:
    - 1.singleb: 0.149*** (0.00111)
    - 1.nocft: -0.0192*** (0.00311)
    - w_ycsh40: 0.0567*** (0.00364)
    - Observations: 188,414; 188,414; 130,722.
    - R-squared: 0.4; 0.06; 0.07.
- Trade-off: data-driven weighting of CRI components may improve fit but reduces interpretability and even-handedness.
- Consistency with prior studies: positive impacts of CRI on infrastructure prices; low-corruption countries show muted or insignificant cost impacts.
- Country-specific main CRI regression highlights (selected statistics):
  - Paraguay (Table 12): CRI coefficients 0.378*** (0.0245); 0.380*** (0.0201); 0.371*** (0.0191); 0.386*** (0.0162); 0.554*** (0.0563). CRI2 -0.306*** (0.0958). Observations: 25,597. R-squared: 0.09, 0.164, 0.168, 0.242, 0.243.
  - Uganda (Table 14): CRI coefficients 0.0845*** (0.024); 0.0863*** (0.0224); 0.0913*** (0.0218); 0.0996*** (0.0191); 0.0703 (0.0475). CRI2 0.0347 (0.0527). Observations: 41,394. R-squared: 0.02, 0.041, 0.066, 0.103, 0.104.
  - Romania (Table 16): CRI coefficients 0.325*** (0.0423); 0.312*** (0.0348); 0.311*** (0.0327); 0.307*** (0.0331); 0.491*** (0.0155). CRI2 -0.437*** (0.0323). Observations: 247,750. R-squared: 0.0898, 0.131, 0.139, 0.159, 0.167.
  - Indonesia (Table 18): CRI coefficients 0.0800*** (0.00546); 0.101*** (0.00721); 0.0699*** (0.00546); 0.0700*** (0.00546); CRI -0.0632*** (0.0139). CRI2 0.2079*** (0.01867). Observations: 655,861; R-squared: 0.014, 0.058, 0.142, 0.142, 0.145.
- Single bidding and bidding-structure measures often positive and significant in alternative specifications (e.g., Indonesia 1.corrbid and 2.corrbid positive).

### IX. Validation strategy and red-flag thresholds (Appendix II highlights)
- Two validation frameworks:
  1. Logit for single bidding: Z_i = α + Σ_{j=1}^{7} β_j X_{j,i} + Σ_{j=1}^{n} β_j C_{j,i} + ε_i. Result: positive and significant coefficient for each of the seven indicators in logit regressions.
  2. OLS (fixed effects) for supplier contract share: S_i = α + Σ_{j=1}^{7} β_j X_{j,i} + Σ_{j=1}^{n} β_j C_{j,i} + ε_i. Regressions limited to suppliers with more than 4 contracts per year; robustness checks with 3+ and 10+ contracts.
- Time-related red-flag cutoffs (exact thresholds by country):
  - Submission period thresholds (Table 7 examples):
    - Georgia: High risk Less than 6 days; Medium risk Less than 13 days; Not a red flag More than 13 days.
    - Indonesia: High risk 0 to 7 days; Medium risk 8 to 14 days; Not a red flag More than 14 days.
    - Paraguay: High risk Less than 13 days or 31 to 47 days; Medium risk 13 to 30 days; Not a red flag 16 to 20 days or more than 47 days.
    - Romania: High risk 30 to 33 days if procedure type is open or negotiated with publication and 9 to 14, and 65 to 378 days for the rest; Not a red flag Less than 30 days and more than 33 days for open and negotiated with publication procedure types and less than 9 and more than 14 for the rest.
    - Uganda: High risk Less than 17 days; Medium risk 17 to 41 days; Not a red flag More than 41 days.
  - Decision period thresholds (Table 8 examples):
    - Georgia: High risk Less than 14 days or more than 25 days; Not a red flag 14 to 25 days.
    - Indonesia: High risk Less than 4 days; Medium risk 5 to 11 days or more than 25 days; Not a red flag 11 to 25 days.
    - Paraguay: High risk 0 to 22 days; Medium risk 23 to 64 days; Not a red flag More than 64 days.
    - Romania: High risk Less than 32 days; Medium risk 33 to 53 days; Not a red flag More than 50 days.
    - Uganda: High risk 1 day or more than 14 days; Not a red flag 2 to 14 days.

### X. Procedure-type classifications and validation findings
- Procedure-type classification method: text searches of administrative procedure names; regression-based classification into high risk, medium risk, not a red flag based on links to single bidding and spending concentration.
- Selected country procedure-type classifications (exact names preserved):
  - Georgia high risk red flag examples: 1. e-Procurement Procedure (GEO); 2. e-Procurement Procedure (GEO) via price list.
  - Romania high risk red flag examples: 1. Negotiated; 2. Negotiated without publication.
  - Paraguay high risk red flag examples: 1. Direct contracting; 2. Other.
  - Indonesia high risk red flag examples (original language): 1. e-Lelang Pemilihan Langsung; 2. e-Penunjukan Langsung; 3. e-Seleksi Langsung; 4. Lelang Pemilihan Langsung – Pascakualifikasi Satu File - Harga Terendah Sistem Gugur.
  - Uganda high risk red flag: 1. Restricted.
- Validation logit results (selected coefficients, Table 9 excerpts):
  - Medium risk red flag /1:
    - Georgia: 0.128*** (0.016)
    - Indonesia: 0.418** (0.191)
    - Paraguay: 0.952*** (0.037)
    - Romania: 1.081*** (0.018)
    - Uganda: 0.316*** (0.114)
  - High risk red flag /1:
    - Georgia: 0.542*** (0.052)
    - Indonesia: 0.582** (0.290)
    - Paraguay: 1.373*** (0.040)
    - Uganda: 1.662*** (0.100)
  - Pseudo-R2 by country: Georgia 0.05; Indonesia 0.10; Paraguay 0.13; Romania 0.12; Uganda 0.47.
  - Observations (selected quoted): Georgia 2004; Indonesia 0364; Paraguay 7401; Romania 8064; Uganda 3544? (tables contain exact counts).

### XI. Descriptive statistics and red-flag prevalence (Appendix I excerpts)
- Number of observations and sample years:
  - Georgia: 202,299; 2011-2019
  - Romania: 620,261; 2007-2020
  - Indonesia: 682,070; 2012-2018
  - Paraguay: 142,878; 2010-2020
  - Uganda: 47,641; 2016-2020
- Number of buyers:
  - Georgia: 2,833
  - Romania: 9,710
  - Indonesia: 4,146
  - Paraguay: 434
  - Uganda: 190
- Number of suppliers:
  - Georgia: 18,203
  - Romania: 47,533
  - Indonesia: 93,292
  - Paraguay: 13,277
  - Uganda: 10,810
- CRI descriptive moments (selected exact values as reported):
  - Georgia: Mean 0.9; Standard Deviation 0.2; Missing Rate 3.69%
  - Romania: Mean 0.7; Standard Deviation 0.5; Missing Rate 43.73%
  - Indonesia: Mean 0.98; Standard Deviation 3.11; Missing Rate 1.17%
  - Paraguay: Mean 0.7; Standard Deviation 75.1; Missing Rate 70.87%
  - Uganda: Mean 8,527.8; Standard Deviation 1,668,938.0; Missing Rate 3.21%
- CRI percentiles (10th / 90th, selected):
  - Georgia: 10th 0.2; 90th 0.6
  - Romania: 10th 0.0; 90th 0.5
  - Indonesia: 10th 0.1; 90th 0.5
  - Paraguay: 10th 0.1; 90th 0.4
  - Uganda: 10th 0.2; 90th 0.7
- Selected prevalence and summary statistics:
  - "% single bidding = 0" 49.0% 68.3% 99.9% 41.2% 31.4%
  - "% single bidding = 15" 1.0% 31.5% 0.1% 19.7% 68.6%
  - No CFT (% nocft): Georgia 98.6%; Romania 75.9%; Indonesia 99.5%; Paraguay 93.4%; Uganda 45.4%
  - Submission period means and s.d. (selected): Submission period Georgia Mean 11.3, Standard Deviation 7.1; Romania Mean 1.6, Standard Deviation 69.0.
  - Decision period means and s.d. (selected): Decision period Georgia Mean 19.0, Standard Deviation 13.2; Romania Mean 155.0, Standard Deviation 162.2; Indonesia Mean 12.2, Standard Deviation 9.9; Paraguay Mean 46.4, Standard Deviation 35.0; Uganda Mean 7.5, Standard Deviation 12.5.
  - Tax haven Contract Share Mean (selected): Georgia Mean 0.4; Romania Mean 0.3; Indonesia Mean 0.7; Paraguay Mean 0.0; Uganda Mean 0.6.
  - Tax haven Contract Share Missing rate (%) 0.11% (Georgia), 0.62% (Romania), 1.78% (Indonesia), 6.92% (Paraguay), 0.39% (Uganda).

### XII. Policy implications and recommendations
- Practical tool: Corruption Cost Tracker (CCT) to identify corruption risk sources by sector, region, or organization type and inform reform measures addressing red flags.
- Key quantitative takeaway: an additional red flag leads to a price increase of 1 to 5 percent (country-dependent).
- Recommended measures informed by red-flag analysis:
  - Reassess submission periods flagged as vulnerabilities; compare against peer countries to determine justification or need for revision.
  - Prioritize development of web portals and machine-readable public procurement databases to increase transparency and enable monitoring.
  - Promote adoption of e-procurement systems (particularly in low-income countries) to reduce discretion and transaction costs and curb procurement vulnerabilities.
- Use cases:
  - Track corruption risks and their impact in procured medical supplies (relevant for COVID-19 related procurement scrutiny).
  - Monitor in real time whether corruption risks decline following measures to strengthen oversight.
  - Rapid updates feasible as countries publish more procurement contracts; the CCT can be adapted and customized by users.

### XIII. Further work and extensions
- Authors working on adding additional cost types such as cost overruns and higher-quality pricing information (e.g., unit prices of standardized goods) for more granular assessment.
- Suggest exploration of interactions among CRI components, potentially via machine learning approaches, to capture complex dynamics at the country level.
- Caution reiterated on cross-country comparability due to differences in data quality, regulatory prescriptions, and data scope.

*IMF Working Paper: Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact*

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

### wpiea2022094-print-pdf - References .............................................................................................................

### I. Introduction: purpose and scope
- Public procurement constitutes about 12 percent of global GDP or 11 trillion USD per year.
- Losses through procured spending are estimated to be about 10-20 percent, even in countries with relatively high integrity of their procurement systems in the European Union (Hafner et al., 2016).
- The paper’s main contribution is to assess whether red flags of corrupt behaviors have an impact on prices of procured goods and services by estimating the impact of corruption risks (assessed through red flags) on relative prices (i.e., comparing prices paid with reference prices).
- The dataset for the five pilot countries includes over 1.5 million contracts, capturing from 15 to 55 percent of total procured spending in each country.
- Five pilot countries analyzed: Georgia, Indonesia, Paraguay, Romania, and Uganda.

### II. Methodology: measuring corruption risks with objective data
- Approach overview:
  - Seven objective indicators (“red flags”) are mapped and used to compute a composite Corruption Risk Index (CRI).
  - CRI is a simple average of the seven red flags after normalization.
  - Each contract is scored on each red flag with discrete values: 0 for lowest corruption risk, 0.5 for medium risk, and 1 for the highest risk.
  - Scores are averaged across contracts for categories and across the seven red flags to compute the CRI.
- The seven red flags and their intended interpretation:
  - Single bidder contracts — contract awarded in a tender where only one bidder participated; a straightforward gauge of limited competition.
  - Non-open procedures — use of restricted or exceptional procedure types that lead to uncompetitive tenders.
  - Lack of publication of call for tenders — limited publication that can lead to uncompetitive tenders and deviates from transparency norms.
  - Period for submitting bids — both short and excessively long submission periods can be associated with corruption risks; assessment takes into account country-specific features.
  - Period for selecting the winning bid — short or long decision periods can be indicative of unfair assessment or manipulation.
  - Spending concentration (by organization, by year) — high concentration can signal favoritism or abuse of dominant market positions to extract rents.
  - Share of suppliers registered in jurisdictions offering limited company and banking transparency — such registrations can facilitate secrecy and illicit payments.
- Scoring and aggregation:
  - Contract-level scores: 0, 0.5, 1 for each red flag.
  - Example: for single bidding the contract-level score is 0 or 1; the contract-level frequency aggregates to a country or sector score (e.g., Romania 31 percent of contracts were awarded in tenders with a single bidder).
  - CRI equals the average of the seven normalized red-flag scores for a contract or aggregation unit.

### III. Data coverage and implementation details
- Data sources: government web portals for public procurement across the five pilot countries.
- Coverage: over 1.5 million contracts across Georgia, Indonesia, Paraguay, Romania, and Uganda; captures from 15 to 55 percent of total procured spending in each country.
- The seven red flags are selected because they are observable and consistently calculable across the publicly available procurement datasets used.
- The CRI enables cross-country, cross-sectoral, and time-series comparisons, with caveats regarding procedural differences and idiosyncrasies.

### IV. Strengths and limitations of the CRI
- Strengths:
  - CRI is more robust than individual components because corruption typically exploits multiple vulnerabilities simultaneously.
  - Predominantly data-driven and informed by established theories of corruption.
  - Validity regressions (Appendix II) determine which definitions of risky categories align most closely with the adopted definition of corruption; selection of high-risk procedure types is driven by association with single bidding.
  - Provides granular guidance for policy makers to explore sectoral vulnerabilities and factors driving price differentials.
- Limitations and caveats:
  - CRI measures corruption risks, not confirmed instances of corruption; observed associations with relative prices indicate potential corrupt behaviors but may reflect structural market features instead.
  - Quantitative approach does not assess how corruption risks affect the quality of procured goods and services.
  - Does not include capacity assessment due to data limitations and potential subjectivity of capacity evaluations.
  - Cross-country comparisons are possible but should be interpreted with caution because of differences in procurement procedures and indicator development.

### V. Implementation notes and examples from the literature table
- CRI construction details:
  - Red flags are defined using a cardinal order (low-medium-high risk) and assigned values 0, 0.5, and 1 respectively.
- Examples and relevant numeric points from literature and table entries (as presented):
  - Single bidder, number of bidders, and recurrent contract awards are widely available in many countries and used as indicators of level of competition.
  - Some listed empirical studies and contexts (from the table): Paraguay 2004-2007; Italy 2000-2005; EU 2008-2012; EU 2006-2010; Argentina 1996-1997; Indonesia 2003-2004; Sweden 1990-1998; Romania 2008-2012.
  - The authors note the need to harmonize definitions for some indicators and highlight availability issues for price and quantity data across countries.

### VI. Next analytical step: estimating price impacts
- The paper proceeds to estimate the impact of the CRI and individual red flags on relative prices (prices paid compared to reference prices) and to evaluate whether observed corruption risks translate into measurable price differentials across the five pilot countries.
- Appendices provide:
  - Appendix I: Data Collection Methodology and Descriptive Statistics.
  - Appendix II: Regressions Underpinning the Validity of the CRI.
  - Appendix III: Additional Price Regressions (country-specific: Paraguay, Uganda, Romania, Indonesia).

*Source: IMF Working Paper content provided in the PDF extract.*

### 9. We assess the budget implications of corruption risks, by estimating the impact of CRI (and other

### 9. We assess the budget implications of corruption risks, by estimating the impact of CRI (and other

### Methodology and regression specifications
- Regressions link the size of discounts offered by the winning firm compared to the auction reference price on corruption risks while controlling for year, contract value, main market, buyer location, and buyer type on the contract level.
- Relative prices are calculated as actual contract values divided by the initially estimated contract value of the tender (or through savings, if available directly in the dataset).
- Five main regression models estimated:
  - Model 1: CRI as the only independent variable.
  - Models 2–5: include controls for product market (CPV division, location, and contract value), organizational framework (buyer type) and year.
  - Models 1–3 restrict relative prices to between 0.5 and 1.5.
  - Additional conservative regressions restrict relative prices to 0.5–1 (cropping the upper end).
  - Model 5 allows a quadratic specification for CRI to capture non-linearities.
  - Model 4 chosen as the main prediction model for all countries (most robust, widest range of controls, typically highest explanatory power).
- Note: coefficients remain significant and largely the same size across specifications; non-linear models add little improvement relative to complexity.
- Caveat: association between CRI and higher relative prices is not sufficient to imply causality; higher prices can reflect structural market or country-specific circumstances. Regressions control for market specificities using product market fixed effects.

### Main regression results (selected coefficients and statistics)
- Dependent variable: relative price (RP) defined as the ratio of actual contract value and normal value (i.e., the value underpinned by the reference price, based on standard market prices).
- Selected coefficient estimates (as reported for Georgia, Model columns):
  - CRI: 0.276*** (standard error (0.004))
  - CRI: 0.315*** (0.004)
  - CRI: 0.312*** (0.004)
  - CRI: 0.312*** (0.004)
  - CRI: 0.222*** (0.012)
  - (CRI)^2: 0.116*** (0.016) — included in model with quadratic specification.
- Controls and indicators included across specifications:
  - Year controls: ✔ ✔ ✔ ✔
  - Contract Value (100 quantiles): ✔, ✔ ✔
  - CPV division: ✔ ✔ ✔ ✔
  - Buyer type: ✔ ✔ ✔
  - Buyer location: ✔ ✔ ✔
- Sample sizes and fit:
  - Observations: 188,472; 188,472; 188,472; 188,414; 188,414 (across columns)
  - R-squared: 0.15; 0.20; 0.21; 0.21; 0.21
- Robust standard errors in parentheses. Clustered over buyers. *** p<0.01, ** p<0.05, * p<0.1.

### Country-level impacts and elasticities (model 4 elasticities and price impacts)
- Price elasticity to CRI (Elasticities taken from model 4):
  - Paraguay: 0.395
  - Georgia: 0.314
  - Romania: 0.314
  - Uganda: 0.10
  - Indonesia: 0.07
- Price impact of CRI increase (Percent) (Everything equal, an additional red flag (1/7 points increase on the CRI score) increase prices by 1/7 times the elasticity):
  - Paraguay: 5.5
  - Georgia: 4.5
  - Romania: 4.4
  - Uganda: 1.4
  - Indonesia: 1.0
- Illustration: in Romania, an additional red flag (1/7 points increase on the CRI score) is predicted to increase prices by 4.4 percentage points (0.307*(1/7)*100=4.4).

### Alternative specifications and component analysis
- Alternative regressions explore how individual CRI indicators affect relative prices to identify specific factors contributing to overpricing.
- Trade-off noted: letting data dictate relative weights of CRI components may improve fit but would reduce CRI’s ease of interpretation and even-handedness.
- Findings consistent with prior studies (e.g., Fazekas and Tóth, 2017) showing positive impacts of CRI on infrastructure prices across countries, with low-corruption countries showing muted or insignificant cost impacts.
- Example alternative-regression results (selected coefficients for Georgia, dependent variable RP, 0.5<RP≤1):
  - 1.singleb: 0.149*** (0.00111)
  - 1.nocft: -0.0192*** (0.00311)
  - w_ycsh40: 0.0567*** (0.00364)
  - Observations: 188,414; 188,414; 130,722
  - R-squared: 0.4; 0.06; 0.07
- Regression controls include contract values, buyer type, buyer location, market, and tender year. Robust standard errors in parentheses. Clustered over buyers. *** p<0.01, ** p<0.05, * p<0.1.

### Policy implications and recommendations
- The methodology provides a user-friendly Corruption Cost Tracker (CCT) to identify corruption risk sources by sector, region, or organization type, and to inform reform measures addressing identified red flags.
- Key quantitative takeaway: an additional red flag leads to a price increase of 1 to 5 percent (country-dependent).
- Practical anti-corruption measures informed by red-flag analysis:
  - Reassess submission periods where they are flagged as vulnerabilities; compare against peer countries to determine if country-specific factors justify current lengths or whether submission periods should be revised to reduce corruption risks.
  - Prioritize development of web portals and machine-readable public procurement databases to increase transparency and enable monitoring.
  - Promote adoption of e-procurement systems (particularly in low-income countries) to reduce discretion and transaction costs, and to curb procurement vulnerabilities.
- Use cases:
  - Track corruption risks and their impact in procured medical supplies (relevant for COVID-19 related procurement scrutiny).
  - Monitor in real time whether corruption risks decline following measures to strengthen oversight.
  - Rapid updates feasible as countries publish more procurement contracts; the CCT can be adapted and customized by users.

### Further work and extensions
- Authors working on adding additional cost types such as cost overruns and higher-quality pricing information (e.g., unit prices of standardized goods) for more granular assessment of corruption effects.
- Suggest exploration of interactions among CRI components, potentially via machine learning approaches, to capture complex dynamics at the country level.
- Caution on cross-country comparability: non-negligible differences remain in data quality, regulatory prescriptions, and data scope, limiting direct comparisons of response curves across countries.

*IMF Working Paper: Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact*

### 12. https://doi.org/10.3917/ncae.022.0001.

### Appendix I–II: Data Collection Methodology, Descriptive Statistics, and Validation of the Corruption Risk Index (CRI)

### Data collection methodology
- Data source: contract award announcements (publication mandatory) and, when available, calls for tenders from government-run electronic procurement platforms.
- Automated web crawler implemented to scrape HTML, XML, and CSV outputs; tools include Python and Java.
- Publications transposed into a uniformly structured data template; conversions include numbers, dates, and enumeration values.
- Linked publications to reconstruct each tendering cycle: Call for Tenders → Contract Award → payments / completion announcements; modifications or cancellations are accounted for.
- Reconciling linked records to create a single best image of each public tender, including reconciliation of conflicting information and filling empty fields from related notices.
- Manual cross-checking with source publications followed by standardization of buyers’ and suppliers’ names.
- For Indonesia and Uganda: multi-step token-based string-matching algorithm used for observations with missing tender product codes, matching tender title, lot title, and/or product description to relevant product codes.
- Full technical documentation and codes located at: https://github.com/digiwhist/backend.

### Descriptive statistics and red-flag prevalence (Table 6, selected exact values)
- Number of observations and sample years:
  - Georgia: 202,299; 2011-2019
  - Romania: 620,261; 2007-2020
  - Indonesia: 682,070; 2012-2018
  - Paraguay: 142,878; 2010-2020
  - Uganda: 47,641; 2016-2020
- Number of buyers:
  - Georgia: 2,833
  - Romania: 9,710
  - Indonesia: 4,146
  - Paraguay: 434
  - Uganda: 190
- Number of suppliers:
  - Georgia: 18,203
  - Romania: 47,533
  - Indonesia: 93,292
  - Paraguay: 13,277
  - Uganda: 10,810
- Corruption Risk Index (CRI) descriptive moments (selected):
  - Georgia: Mean 0.9; Standard Deviation 0.2; Missing Rate 3.69%
  - Romania: Mean 0.7; Standard Deviation 0.5; Missing Rate 43.73%
  - Indonesia: Mean 0.98; Standard Deviation 3.11; Missing Rate 1.17%
  - Paraguay: Mean 0.7; Standard Deviation 75.1; Missing Rate 70.87%
  - Uganda: Mean 8,527.8; Standard Deviation 1,668,938.0; Missing Rate 3.21%
- CRI percentiles (10th / 90th, selected):
  - Georgia: 10th 0.2; 90th 0.6
  - Romania: 10th 0.0; 90th 0.5
  - Indonesia: 10th 0.1; 90th 0.5
  - Paraguay: 10th 0.1; 90th 0.4
  - Uganda: 10th 0.2; 90th 0.7
- Single bidding prevalence (selected exact strings from table):
  - "% single bidding = 0" 49.0% 68.3% 99.9% 41.2% 31.4%
  - "% single bidding = 15" 1.0% 31.5% 0.1% 19.7% 68.6%
  - "% single bidding = Missing" 0.0% 0.2% 0.0% 39.2% 0.0%
- Procedure-type red-flag prevalence snippets:
  - "% corr_proc = 0" 23.6% 83.7% 10.3% 19.2% 96.2%
  - "% corr_proc = 1" 75.4% 16.3% 47.6% 19.8% 1.8%
  - "% corr_proc = 2" 1.0% ... 41.9% 60.8% 2.0%
  - "% corr_proc = Missing" 0.0% ... 0.2% 0.3% 0.0%
- Submission period and decision period red-flag summary stats (selected means and s.d.):
  - Submission period: Georgia Mean 11.3, Standard Deviation 7.1; Romania Mean 1.6, Standard Deviation 69.0
  - Decision period: Georgia Mean 19.0, Standard Deviation 13.2; Romania Mean 155.0, Standard Deviation 162.2; Indonesia Mean 12.2, Standard Deviation 9.9; Paraguay Mean 46.4, Standard Deviation 35.0; Uganda Mean 7.5, Standard Deviation 12.5
- No CFT (% nocft) prevalence:
  - Georgia 98.6%
  - Romania 75.9%
  - Indonesia 99.5%
  - Paraguay 93.4%
  - Uganda 45.4%
- Tax-haven supplier and contract-share indicators (selected):
  - Foreign Supplier not in a tax haven: 0.2% (Georgia), 0.6% (Romania)
  - Foreign Supplier in a tax haven: 0.0% (Georgia), 0.0% (Romania)
  - Local Supplier: 99.8% (Georgia), 99.4% (Romania)
  - Tax haven Contract Share (w_ycsh/proa_ycsh) Mean (selected): Georgia Mean 0.4; Romania Mean 0.3; Indonesia Mean 0.7; Paraguay Mean 0.0; Uganda Mean 0.6
  - Tax haven Contract Share Missing rate (%) 0.11% (Georgia), 0.62% (Romania), 1.78% (Indonesia), 6.92% (Paraguay), 0.39% (Uganda)

### Validation strategy for the CRI and red flags (Appendix II)
- Conceptual approach: corruption defined as lack of competition favoring a connected bidder; two key outcomes used for validation:
  - Single bidding (indicator of lack of competition)
  - Supplier contract share (S): repeated favoritism of the same supplier
- Two regression frameworks employed:
  1. Logit model for single bidding:
     - Model: Z_i = α + Σ_{j=1}^{7} β_j X_{j,i} + Σ_{j=1}^{n} β_j C_{j,i} + ε_i
     - Where Z_i is the log of the probability of single bidding, X are red flags, C are control variables (contract values in real terms, market types based on product codes, buyer types, tender year).
     - Result: positive and significant coefficient for each of the seven indicators in the logit regressions, indicating they collectively explain reduced competitiveness.
  2. OLS (fixed effects) model for supplier contract share:
     - Model: S_i = α + Σ_{j=1}^{7} β_j X_{j,i} + Σ_{j=1}^{n} β_j C_{j,i} + ε_i
     - Regressions limited to suppliers with more than 4 contracts per year (robustness tests with 3+ and 10+ contracts; results broadly unchanged).
     - Rationale for cutoffs: exclude trivially concentrated suppliers (e.g., single-contract suppliers with 100 percent concentration).
- Two time-related red flags (submission period and decision period):
  - Data-driven approach: split periods into deciles; identify a decile as the norm (usually longest submission period or decision period closest to average).
  - For each decile, compute deviation from the norm and test whether deviations are associated with higher single bidding probability or higher supplier contract share.
  - Aim is correlation across contexts rather than causal identification.

### Thresholds for submission and decision periods (data-driven red-flag cutoffs)
- Submission Period Threshold Red Flags (Table 7, exact thresholds by country):
  - Georgia: High risk Less than 6 days; Medium risk Less than 13 days; Not a red flag More than 13 days
  - Indonesia: High risk 0 to 7 days; Medium risk 8 to 14 days; Not a red flag More than 14 days
  - Paraguay: High risk Less than 13 days or 31 to 47 days; Medium risk 13 to 30 days; Not a red flag 16 to 20 days or more than 47 days
  - Romania: High risk 30 to 33 days if procedure type is open or negotiated with publication and 9 to 14, and 65 to 378 days for the rest; Not a red flag Less than 30 days and more than 33 days for open and negotiated with publication procedure types and less than 9 and more than 14 for the rest.
    - Note: medium risk for Romania not included by choice based on validity regressions.
  - Uganda: High risk Less than 17 days; Medium risk 17 to 41 days; Not a red flag More than 41 days
- Decision Period Threshold Red Flags (Table 8, exact thresholds by country):
  - Georgia: High risk Less than 14 days or more than 25 days; Not a red flag 14 to 25 days
  - Indonesia: High risk Less than 4 days; Medium risk 5 to 11 days or more than 25 days; Not a red flag 11 to 25 days
  - Paraguay: High risk 0 to 22 days; Medium risk 23 to 64 days; Not a red flag More than 64 days
    - Footnote: while more than 64 days might appear excessive for simple procurement, regressions associated medium risk (23-64 days) with higher single bidding and spending concentration.
  - Romania: High risk Less than 32 days; Medium risk 33 to 53 days; Not a red flag More than 50 days
  - Uganda: High risk 1 day or more than 14 days; Not a red flag 2 to 14 days

### Procedure-type red flags and validation findings
- Method: identify all administrative procedure types (text searches in laws, secondary legislation, tender documents); use the same regression frameworks (single bidding logit and supplier share OLS) with controls; classify procedure types as:
  - High risk red flag: significant and positive link with single bidding and/or spending concentration (e.g., direct awards without expectation of competition)
  - Medium risk red flag: significant but smaller effects (e.g., invitation tenders where invited bidders are expected to compete)
  - Not a red flag: statistically indistinguishable from reference category (open procedure)
- Selected procedure-type red-flag classifications (Table 11, country excerpts; exact procedure names preserved):
  - Georgia
    - High risk red flag: 1. e-Procurement Procedure (GEO); 2. e-Procurement Procedure (GEO) via price list
    - Medium risk red flag: 1. Electronic Tender (SPA); 2. Electronic Tender (SPA) via price list; 3. Simplified Electronic Tender (SPA); 4. Simplified Electronic Tender (SPA) via price list.
    - Not a red flag: Donor electronic procurement procedure (DEP); Electronic Tender (DAP); Electronic Tender Without Reverse Auction (NAT); Electronic Tender Without Reverse Auction (NAT) via price list; Simplified Electronic Tender Without Reverse Auction (NAT); Simplified Electronic Tender Without Reverse Auction (NAT) via price list; Simplified Electronic Tender (DAP); Simplified Two Stage Electronic Tender (MEP); Two Stage Electronic Tender (MEP); Two Stage Electronic Tender (MEP) via price list
  - Romania
    - High risk red flag: 1. Negotiated; 2. Negotiated without publication
    - Medium risk red flag: 1. Open; 2. Approaching bidders; 3. Competitive dialog; 4. Negotiated with publication; 5. Restricted
  - Paraguay
    - High risk red flag: 1. Direct contracting; 2. Other
    - Medium risk red flag: 1. Open within threshold
    - Not a red flag: 1. Open auction; 2. Limited tendering
  - Indonesia (selected exact procedure names preserved in original language)
    - High risk red flag includes: 1. e-Lelang Pemilihan Langsung; 2. e-Penunjukan Langsung; 3. e-Seleksi Langsung; 4. Lelang Pemilihan Langsung – Pascakualifikasi Satu File - Harga Terendah Sistem Gugur
    - Medium risk red flag includes: 1. e-Lelang Sederhana, e-Lelang Umum; 2. e-Seleksi Umum; 3. Lelang Sederhana - Pascakualifikasi Satu File - Harga Terendah Sistem Gugur; 4. Lelang Sederhana - Prakualifikasi Dua File - Kualitas dan Biaya; 5. Lelang Sederhana - Prakualifikasi Dua File - Sistem Nilai; 6. Lelang Sederhana - Prakualifikasi Satu File - Biaya Terendah; 7. Lelang Umum - Pascakualifikasi Dua File - Sistem Nilai; 8. Lelang Umum - Pascakualifikasi Dua File - Sistem Umur Ekonomis, Lelang Umum - Pascakualifikasi Satu File - Harga Terendah Sistem Gugur; 9. Lelang Umum - Prakualifikasi Dua File – Kualitas, Lelang Umum - Prakualifikasi Dua File - Kualitas dan Biaya
  - Uganda
    - High risk red flag: 1. Restricted
    - Medium risk red flag: 1. Approaching Bidders
    - Not a red flag: 1. Open; 2. Negotiated; 3. Negotiated without publication; 4. Negotiated with publication

### Key regression validation results (selected exact coefficients and model diagnostics)
- Logit validation using single bidding (Table 9, selected coefficients; standard errors in parentheses):
  - Medium risk red flag /1:
    - Georgia: 0.128*** (0.016)
    - Indonesia: 0.418** (0.191)
    - Paraguay: 0.952*** (0.037)
    - Romania: 1.081*** (0.018)
    - Uganda: 0.316*** (0.114)
  - High risk red flag /1:
    - Georgia: 0.542*** (0.052)
    - Indonesia: 0.582** (0.290)
    - Paraguay: 1.373*** (0.040)
    - Uganda: 1.662*** (0.100)
  - Other reported items (selection):
    - Pseudo-R2 by country (bottom row): Georgia 0.05; Indonesia 0.10; Paraguay 0.13; Romania 0.12; Uganda 0.47
    - Observations by country (bottom row): Georgia 2004; Indonesia 0364; Paraguay 7401; Romania 8064; Uganda 3544? (note: table contains exact observation counts in source; preserve as quoted in full records)
  - Controls included: contract values, buyer type, market, and tender year.
  - Significance notation: *** p<0.01, ** p<0.05, * p<0.1
- OLS validation using supplier contract share (Table 10, selected coefficients; standard errors in parentheses):
  - Dependent variable: Supplier contract share, more than 4 contracts per year.
  - Example coefficients (first row):
    - Georgia: 0.0693*** (0.001)
    - Indonesia: 0.122*** (0.022)
    - Paraguay: 0.00609*** (0.001)
    - Romania: 0.0141*** (0.001)
    - Uganda: -0.0524*** (0.007)
  - R2 by country (bottom row): Georgia 0.23; Indonesia 0.19; Paraguay 0.25; Romania 0.32; Uganda 0.08
  - Observations by country (bottom row): Georgia 13989; Indonesia 12519; Paraguay 8675; Romania 76239; Uganda 47702? (table includes exact observation counts in source)
  - Controls included: contract values, buyer type, market, and tender year.
  - Significance notation: *** p<0.01, ** p<0.05, * p<0.1

### Empirical conclusions reported in the appendix
- The seven red-flag indicators are positively and significantly associated with higher probability of single bidding across countries in the logit models.
- The red flags are also positively and significantly associated with higher supplier contract share (concentration) in fixed-effects OLS models for suppliers with more than 4 contracts per year; robustness checks with 3+ and 10+ contracts yield similar results.
- Time-related red flags (short submission periods and anomalous decision periods) are, in most cases, associated with a significant increase in single bidding and higher spending concentration on specific bidders.
- Procedure-type classifications derived from regressions provide a data-driven decomposition into high-risk, medium-risk, and non-risk procedure types; direct awards without competition represent the highest risk.

*Source: IMF Working Paper — Appendix I and Appendix II content as provided from the supplied PDF extract.*

### 10. Lelang Umum - Prakualifikasi Dua File - Sistem Nilai

### 10. Lelang Umum - Prakualifikasi Dua File - Sistem Nilai

### Red flags and cross-border considerations
- The last two indicators tested: the lack of published tenders and the residency in tax heavens of suppliers, are tested using two regression models and show positive and significant coefficients, reinforcing their inclusion among seven red flags.
- For suppliers registered in jurisdictions that avoid disclosing ownership details, the Tax Justice Network company and financial secrecy metric is used as the objective indicator.
- Notes on interpretation:
  - Single bidding and concentration of procurement spending on specific bidders are treated as proxy indicators of corruption but can arise from non-corrupt conditions; methodology seeks to minimize measurement error.
  - Co-occurrence analysis: association between single bidding/high spending concentration and known methods for favoring connected bidders (direct awards, short advertisement periods) is used to lower measurement error.
  - Expectation: positive correlation across corruption risk indicators but imperfect fit, implying corruption can occur without simple indicators and vice versa.

### Validation insights and measurement caveats
- Empirical estimations broadly support the selected red flags, but counter-intuitive results indicate data quality issues and the need for composite indicators rather than reliance on single indicators.
- Example: in Romania, some risk indicators that behave as expected in the single bidder regression are negative significant in the supplier contract share regression—possibly due to organization ID reliability issues making the latter regression noisier.
- Single bidding regressions are considered more reliable and provide sufficient evidence for indicator validity (reference to Fazekas and Kocsis, 2020).
- Recommendation implied: rely on a broad composite indicator (CRI) given potential divergence across different validation outcomes.

### Paraguay — Main and alternative regression results (selected statistics)
- Dependent variable: relative price (RP) defined as the ratio of actual contract value and normal value.
- Main results (Table 12): CRI coefficients
  - CRI 0.378*** (0.0245)
  - CRI 0.380*** (0.0201)
  - CRI 0.371*** (0.0191)
  - CRI 0.386*** (0.0162)
  - CRI 0.554*** (0.0563)
- Additional statistic:
  - CRI2 -0.306*** (0.0958)
- Controls included: Year controls; Contract Value (100 quantiles); CPV division; Buyer type; Buyer location (varies by specification).
- Observations: 25,597; in alternative models Observations 23,551.
- R-squared values: 0.09, 0.164, 0.168, 0.242, 0.243 (by specification).
- Alternative specification (Table 13) selected coefficients:
  - 1.singleb 0.0851*** (0.00375)
  - 1.nocft 0.0229** (0.00935)
  - w_ycsh4 0.0573*** (0.0145)
- Observations in alternative specs: 23,398; 23,551; 22,859.
- R-squared in alternative specs: 0.24; 0.17; 0.17.

### Uganda — Main and alternative regression results (selected statistics)
- Dependent variable: relative price (RP) defined as the ratio of actual contract value and normal value.
- Main results (Table 14): CRI coefficients
  - CRI 0.0845*** (0.024)
  - CRI 0.0863*** (0.0224)
  - CRI 0.0913*** (0.0218)
  - CRI 0.0996*** (0.0191)
  - CRI 0.0703 (0.0475)
- Additional statistic:
  - CRI2 0.0347 (0.0527)
- Controls included: Year controls; Contract Value (100 quantiles); CPV division; Buyer type; Buyer location.
- Observations: 41,394; alternative Observations 35,793.
- R-squared values: 0.02, 0.041, 0.066, 0.103, 0.104.
- Alternative specification (Table 15) selected coefficients:
  - 1.singleb 0.0330*** (0.0109)
  - 1.nocft 0.0524*** (0.00957)
  - w_ycsh4 0.0148** (0.00618)
- Observations in alternatives: 35,793; 35,793; 22,334.
- R-squared in alternative specs: 0.09; 0.11; 0.09.

### Romania — Main and alternative regression results (selected statistics)
- Dependent variable: relative price (RP) defined as the ratio of actual contract value and normal value.
- Main results (Table 16): CRI coefficients
  - CRI 0.325*** (0.0423)
  - CRI 0.312*** (0.0348)
  - CRI 0.311*** (0.0327)
  - CRI 0.307*** (0.0331)
  - CRI 0.491*** (0.0155)
- Additional statistic:
  - CRI2 -0.437*** (0.0323)
- Controls included: Year controls; Contract Value (100 quantiles); Contract type; CPV division; Buyer type; Buyer location.
- Observations: 247,750; alternative Observations 233,946.
- R-squared values: 0.0898, 0.131, 0.139, 0.159, 0.167.
- Alternative specification (Table 17) selected coefficients:
  - 1.singleb 0.111*** (0.0038)
  - 99.singleb 0.0557*** (0.00772)
  - 1.nocft 0.0218*** (0.00375)
  - w_ycsh4 -0.00293 (0.00603)
- Observations in alternatives: 233,946; 233,946; 159,139.
- R-squared in alternative specs: 0.20; 0.09; 0.09.

### Indonesia — Main and alternative regression results (selected statistics)
- Dependent variable: relative price (RP) defined as the ratio of actual contract value and normal value.
- Main results (Table 18): CRI coefficients
  - CRI 0.0800*** (0.00546)
  - CRI 0.101*** (0.00721)
  - CRI 0.0699*** (0.00546)
  - CRI 0.0700*** (0.00546)
  - CRI -0.0632*** (0.0139)
- Additional statistic:
  - CRI2 0.2079*** (0.01867)
- Controls included across specifications: Year controls; Contract Value (100 quantiles); Contract type; CPV division; Buyer type; Buyer location (varies by model).
- Observations: 655,861; 654,590; 654,590; 654,262; 654,262.
- R-squared values: 0.014, 0.058, 0.142, 0.142, 0.145.
- Alternative specification (Table 19) selected coefficients:
  - 1.singleb 0.0256*** (0.00555)
  - 1.corrbid 0.0388*** (0.00134)
  - 2.corrbid 0.0546*** (0.00191)
  - 1.nocft 0.0149** (0.00681)
  - w_ycsh4 0.0222*** (0.00246)
- Observations in alternatives: 654,262; 654,262; 654,262; 244,579.
- R-squared in alternative specs: 0.14, 0.19, 0.14, 0.14.
- Note on bidding structure: Model 2 uses an alternative definition of bidding structure where 1.corr_bid corresponds to 12 to 22 bidders and 2.corr_bid corresponds to 1 to 11 bidders.

### Cross-country empirical takeaways
- CRI (composite corruption risk indicator) shows consistently positive and often statistically significant associations with relative price (RP) across multiple country specifications (Paraguay, Uganda, Romania, Indonesia), though magnitudes and significance vary by model and country.
- Some models include CRI2 with positive or negative coefficients (e.g., Paraguay CRI2 -0.306***; Romania CRI2 -0.437***; Indonesia CRI2 0.2079***), indicating robustness checks and alternative formulations can yield different directional effects.
- Single bidding (1.singleb) and related bidding structure measures (1.corrbid, 2.corrbid) often show positive and significant coefficients in alternative specifications, reinforcing their relevance as tendering outcome indicators.

*Source: Authors Assessment. IMF WORKING PAPERS Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022094-print-pdf.pdf_