Do ghosts exist? Corruption in the Colombian education systempubdocs.worldbank.org/en/603631530034379544/Fergusson-Harker-Molina... · Do ghosts exist? Corruption in the Colombian

Do ghosts exist?Corruption in the Colombian education system

Leopoldo Fergusson† Arturo Harker† Carlos Molina‡

†Universidad de los Andes ‡Pearson Institute (University of Chicago)

World Bank ABCDE ConferenceJune 24-25, 2018

Ghosts exist

⇑ ⇒ ⇑ $ transfers ⇒ ⇑ Capture (corruption)

Ghosts exist


Ghosts exist


Ghosts exist


This project

I Political alignment (same party, RDD estimates) betweenlocal (municipality mayor) and regional (departmentalgovernor) levels:

1. Ghosts increase, and quite substantially.2. Larger effects in:

I “Autonomous municipalities”, more discretion for funds’allocation.

3. Additional findings:I More subsequent electoral fraud, improved incumbency

advantage, no increases in education quantity or quality.

I Consistent with corruption:

1. deviating resources for the reproduction of the clientelisticnetwork (more valuable if network connected in higherechelons),

2. with an efficiency cost for voters,3. and entrenching the clientelistic machine.

Related literature and contribution

1. Corruption in developing countries and efficiency costs(Olken & Pande, 2012)

2. Clientelism beyond particularism and implications forcorruption(E.g. Maiz & Requejo, 2001; Singer, 2009)

3. Political alignment beyond fiscal transfer consequences(a long list)

4. Risks of incentive and fixed-rule schemes in financing publicgoods with weak oversight.(Among others, Acemoglu, Fergusson, Robinson, Romero, andVargas (2016); Bold, Kimenyi, Mwabu, Ng’ang’a, andSandefur (2013); Behrman, Parker, Todd, and Wolpin (2015))

Contents

Context

Data

Results

Conclusion: next steps

Public education: institutional details

I Most money spent is from central government transfers(88%).

I The rest, from regional government (3%) and localgovernment (9%), is only for investment.

I Enrollment is key in central government $:

Account % totalresources

Transferred to Distribution criteria

Payroll 90% Paid directly byMinistry

# Teachers =f(enrollment)

Quality-enrollment(calidad-matrıcula)

5% Regional EducationSecretaries

performance,poverty, enrollment

Quality-access(calidad-gratuidad)

5% Schools enrollment

Public education: institutional detailsEducation Secretary and School Principals

I Regional Education Secretary (one per department, withrelevant exceptions below).

I Monitors funds managed by schools in the jurisdiction.I Hiring, promotion and firing of teachers.

I Nominally tenure track system based on open calls.I If open calls fail (which may “accidentally”), then provisional

direct hiring.I Payroll, especially direct provisional hiring, linked to local

electoral support for mayor → Classic clientelistic patronage.

I School Principals draft reports used to project enrollment.

The political landscape

Audit: census of schools

I Audit study in 2012 financed and managed by the nationalgovernment (Ministry of Education).

I Audit firms implementing the study were competitivelyselected.

I Ministry functionaries sought to protect from cooptation:no local auditors.

I Comprehensive audit that nearly reached the goal:I 8,167,051 from 8,679,035 students were audited (94.1%).

I Detailed verification:I Face to face verification.I Complementary documentary evidence.

I ResultsI 148,410 ghosts.

Ghosts and alignment in the Colombian territory

=0

(0,1%]

(1,10%]

(10,20%]

(20,24%]

Missing

=0

(0,1%]

(1,10%]

(10,20%]

(20,24%]

Missing

Descriptive StatisticsParties not aligned Parties aligned

Mean Median S.D. Min Max Mean Median S.D. Min Max

A. All SampleGhosts (%) 1.35 0.00 4.66 0.00 96.84 1.40 0.06 5.38 0.00 100.00Men 52.4 52.3 7.5 0.0 100.0 51.7 51.8 7.4 0.0 78.3Minors 90.9 96.4 12.1 1.7 100.0 90.1 95.7 12.3 32.6 100.0Secondary 26.2 28.4 25.0 0.0 100.0 29.5 33.2 23.9 0.0 100.0Rural 76.9 100.0 40.8 0.0 100.0 67.1 100.0 45.9 0.0 100.0Strata 0-2 97.0 100.0 6.8 35.3 100.0 97.1 99.7 7.0 42.0 100.0Extended school day 41.6 0.0 48.2 0.0 100.0 21.3 0.0 39.2 0.0 100.0No diploma 1.2 0.0 6.6 0.0 100.0 3.1 0.0 11.3 0.0 100.0Students 253 164 283 3 2640 320 233 318 4 2880Teachers 22 11 27 1 165 27 17 29 1 194

B. Close races (<10%)Ghosts (%) 1.31 0.08 5.05 0.00 93.12 1.61 0.13 4.80 0.00 56.19Men 52.4 52.1 6.8 6.1 84.6 51.6 51.7 8.7 0.0 78.3Minors 88.3 92.3 12.3 31.3 100.0 89.5 96.6 13.8 32.6 100.0Secondary 31.0 35.2 21.8 0.0 100.0 27.3 30.5 25.5 0.0 100.0Rural 74.7 100.0 41.3 0.0 100.0 73.3 100.0 43.2 0.0 100.0Strata 0-2 98.5 99.8 3.8 63.5 100.0 96.8 100.0 8.1 42.0 100.0Extended school day 23.9 0.0 40.7 0.0 100.0 21.2 0.0 39.8 0.0 100.0No diploma 1.9 0.0 7.9 0.0 80.0 2.2 0.0 12.0 0.0 100.0Students 320 268 261 6 1640 300 193 331 4 2880Teachers 28 21 26 1 120 23 14 27 1 194

Municipality level

Manipulation?McCrary (2008) density test

0.0

0.5

1.0

1.5

2.0D

ensity

−1.0 −0.8 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0

Vote Share Margin of Victory for Governor’s Candidate

Note: Discontinuity estimate: 0.0268 (s.e. 0.2647)

Balance on observable variablesSchool level

Men

Minors

Belonging to ethnic group

High school

With disabilities

In rural areas

In stratum 0, 1 or 2

Extended school day

Contracted service

Teachers without diploma

Coordinators

Total students

Total teachers

−2 −1 0 1 2Standardized coefficient

Note: Observations within CCT bandwidth and 95% confidence interval.

Balance on observable variablesMunicipal level

Fiscal dependency

Per capita property tax

Investment on education

SGP on education

Coca presence

Displaced rate

Births

Mortality rate

Notaries

Area

Elevation

Distance to Bogota

Car theft

Number of schools

Rural population

Urban population

Total population

−2 −1 0 1 2Standardized coefficient

Note: Observations within CCT bandwidth and 95% confidence interval.

Main results: ghosts (%)Graphical analysis

0

1

2

3

Ghosts

−0.15 −0.10 −0.05 0.00 0.05 0.10 0.15Vote Share Margin of Victory for Governor’s Candidate

1−30 31−100 101−200

Observations

0

1

2

3

4

Ghosts

−0.25 −0.20 −0.15 −0.10 −0.05 0.00 0.05 0.10 0.15 0.20 0.25Vote Share Margin of Victory for Governor’s Candidate

1−30 31−100 101−200

Observations

Note: Local polynomial regression. Left: linear fit. Right: quadratic fit. Observationswithin CCT bandwidth displayed. Bin selection method: mimicking variance, evenlyspaced using spacings estimators.

Full sample

Main resultsRobustness to bandwidth choice

500

1,000

1,500

2,000

Observ

ations w

ithin

banw

idth

−1.0

0.0

1.0

2.0

3.0G

overn

or−

Mayor

alig

nm

ent

50 60 70 80 90 100 110 120 130 140 150

Bandwidth as proportion of optimal bandwidth (percentage)

Estimator CI 10% CI 5% Observations

Main results: ghosts (%)Nonparametric estimators: CCT with optimal bandwidth

(1) (2) (3) (4) (5) (6)

Panel A. No controlsAlignment 1.402** 1.272* 1.163 1.670** 1.510* 1.463*

(0.692) (0.708) (0.730) (0.748) (0.783) (0.861)

Observations 4,383 4,383 4,383 4,383 4,383 4,383Bandwidth 0.130 0.118 0.105 0.242 0.197 0.146Obs. in bandwidth 1338 1249 1091 2211 1926 1522

Kernel Triangular Epanechnikov Uniform Triangular Epanechnikov UniformLocal polynomial Order 1 1 1 2 2 2

Panel B. ControlsAlignment 1.336** 1.514** 1.660** 1.131** 1.206* 1.363**

(0.561) (0.708) (0.659) (0.514) (0.696) (0.682)

Student controls X X XSchool controls X X XTeacher controls X X XMunicipality and party controls X XObservations 3,809 3,809 3,809 3,809 3,809 3,809Bandwidth 0.101 0.140 0.164 0.127 0.0772 0.0781Obs. in bandwidth 1044 1422 1633 1323 749 755

Notes: In Panel B, regressions are weighted using a triangular kernel and assuming a linear polynomial.Bias corrected coefficients and robust standard errors clustered at municipality level (Calonico, Cattaneo,& Titiunik, 2014). Optimal bandwidth in all columns. * 10%, ** 5%, ***1%.

Extensive margin Municipal level

The impact of autonomy

“You can take advantage of the certified municipality as a channelfor good and bad things, but you can certainly take a lot ofadvantage.”

(Former Education Secretary)

Heterogeneous effectAutonomous municipality

(1) (2)

Alignment 1.301** 1.363*(0.601) (0.772)

Alignment × Autonomy 1.786*(0.935)

Observations 4,383 4,383Bandwidth 0.119 0.119Obs. in bandwidth 1338 1338

Notes: Bias-corrected coefficients and robust standard errors clus-tered at municipality level (Calonico et al., 2014). Optimal bandwidthin all columns. Default * 10%, ** 5%, *** 1%.

Other implications?

I Do these places receive better/more or worse/less?I Not better, nor more.

I Does the strategy pay off electorally?I Party electoral disadvantage in Colombia, but less for aligned

parties. (Fergusson, Querubın, Ruiz, & Vargas, 2017).

I Does it entrench clientelistic machines?I We observe increases in subsequent vote buying.

Results

Next steps

1. Mayor future career prospects.

2. Examine role of local councilors:“The city is divided by neighborhoods controlled bycouncilors.”(Secretario de Educacion, identity hidden)

I Council members play a key role in the mayor-local leader link.I Examine schools in local councilor influence areas within

municipalities, with variation in mayor-councilor alignment.

References I

Abadie, A., Athey, S., Imbens, G. W., & Wooldridge, J. (2017). When should you adjust standard errors forclustering? (Tech. Rep.). NBER Working Paper No. 24003.

Acemoglu, D., Fergusson, L., Robinson, J. A., Romero, D., & Vargas, J. F. (2016). The perils of high-poweredincentives: Evidence from colombia’s false positives (Tech. Rep.). NBER Working Paper # 22617.

Behrman, J. R., Parker, S. W., Todd, P. E., & Wolpin, K. I. (2015). Aligning learning incentives of students andteachers: results from a social experiment in mexican high schools. Journal of Political Economy, 123(2),325–364.

Bold, T., Kimenyi, M., Mwabu, G., Ng’ang’a, A., & Sandefur, J. (2013). Scaling up what works: Experimentalevidence on external validity in kenyan education.

Calonico, S., Cattaneo, M. D., & Titiunik, R. (2014). Robust nonparametric confidence intervals forregression-discontinuity designs. Econometrica, 82(6), 2295–2326.

Fergusson, L., Querubın, P., Ruiz, A., & Vargas, J. (2017). The real winner’s curse (Documentos Cede No.2017-5). Universidad de los Andes-Cede.

Gelman, A., & Imbens, G. (2017). Why high-order polynomials should not be used in regression discontinuitydesigns. Journal of Business & Economic Statistics(just-accepted).

Maiz, R., & Requejo, R. (2001). Clientelism as a political incentive structure for corruption. In Trabajo presentadoen la conferencia de primavera del european consortium for political research, grenoble.

McCrary, J. (2008, February). Manipulation of the running variable in the regression discontinuity design: Adensity test. Journal of Econometrics, 142(2). Retrieved fromhttp://eml.berkeley.edu/~jmccrary/mccrary2006 DCdensity.pdf

Olken, B. A., & Pande, R. (2012). Corruption in developing countries. Annual Review of Economics, 4(1),479-509. Retrieved fromhttp://www.annualreviews.org/doi/abs/10.1146/annurev-economics-080511-110917 doi:10.1146/annurev-economics-080511-110917

Singer, M. (2009). Buying voters with dirty money: The relationship between clientelism and corruption.

http://eml.berkeley.edu/~jmccrary/mccrary2006_DCdensity.pdf

http://www.annualreviews.org/doi/abs/10.1146/annurev-economics-080511-110917

Descriptive StatisticsMunicipality level

Mean Median S.D. Min Max

Ghosts (%) 1.266 1.266 2.346 0 23.00Fiscal dependency 0.719 0.719 0.182 0.0948 0.981Per capita property tax 25.25 25.25 37.04 0 344.3Coca presence 57.58 57.58 346.6 0 4,846Notaries 0.748 0.748 0.765 0 7Displaced rate 336.8 336.8 524.1 0 3,649Rural population 13,096 13,096 13,766 247 134,910Urban population 22,776 22,776 56,267 159 534,956Total population 35,872 35,872 61,843 1,015 535,642Births 519.4 519.4 947.7 11 7,765Mortality rate 24.71 24.71 9.964 7.770 63.16Area 6.229 6.229 1.134 2.833 11.09Elevation 754.2 754.2 1,565 2 25,221Distance to Bogota 390.4 390.4 190.8 49.91 1,271Car theft 2.280 2.280 9.497 0 103Number of schools 2.042 2.042 6.387 0 61Investment on education 6.380 6.380 19.96 0.0496 162.2SGP on education 0.155 0.155 0.154 0 0.910

Back

Extensive margine: Prob(ghosts>0)School level

(1) (2) (3) (4) (5) (6)

Alignment 0.0653 0.0700 0.0490 0.127 0.101 0.0197(0.186) (0.193) (0.193) (0.206) (0.210) (0.217)

Observations 4,383 4,383 4,383 4,383 4,383 4,383Bandwidth 0.112 0.0995 0.0928 0.159 0.155 0.144Obs. in bandwidth 1193 1041 969 1584 1560 1455


Notes: Mean of dependent variable is equal to 0.444 and standard deviation equal to 0.497. Biascorrected coefficients and robust standard errors clustered at municipality level (Calonico et al., 2014).Optimal bandwidth in all columns. * 10%, ** 5%, ***1%.

Extensive margine: Prob(ghosts>0)Municipality level

(1) (2) (3) (4) (5) (6)

Alignment -0.00202 0.0228 0.0390 0.0337 0.0452 0.0146(0.0326) (0.0365) (0.0526) (0.0277) (0.0305) (0.0628)

Observations 332 332 332 332 332 332Bandwidth 0.0901 0.0786 0.0747 0.109 0.101 0.119Obs. in bandwidth 81 68 65 98 91 109


Notes: Mean of dependent variable is equal to 0.886 and standard deviation equal to 0.318. Biascorrected coefficients and robust standard errors clustered at municipality level (Calonico et al., 2014).Optimal bandwidth in all columns. * 10%, ** 5%, ***1%.

Back

Main results: ghosts (%)Graphical analysis: All sample

0

1

2

3

Ghosts

−0.50 −0.40 −0.30 −0.20 −0.10 0.00 0.10 0.20 0.30 0.40Vote Share Margin of Victory for Governor’s Candidate

1−30 31−100 101−200

Observations

−1

0

1

2

3

Ghosts

−0.50 −0.40 −0.30 −0.20 −0.10 0.00 0.10 0.20 0.30 0.40Vote Share Margin of Victory for Governor’s Candidate

1−30 31−100 101−200

Observations

Note: Local polynomial regression. Left: linear fit. Right: quadratic fit. Bin selectionmethod: mimicking variance, evenly spaced using spacings estimators.

Back

Aggregating schools: ghosts (%)Municipal level regressions

(1) (2) (3) (4) (5) (6)

Alignment 3.872* 3.931* 2.232 5.270* 5.413* 3.786(2.382) (2.380) (1.919) (2.851) (2.867) (2.566)



Notes: Mean of dependent variable is 1.266 and standard deviation is 2.346. Bias corrected coefficientsand robust standard errors clustered at municipality level (Calonico et al., 2014). Optimal bandwidth inall columns. * 10%, ** 5%, ***1%.

Back

A falsification testPredicting prior election results (1997, 2000, 2003 and 2007)

−0.3

−0.2

−0.1

0.0

0.1

0.2S

hare

Marg

in o

f V

icto

ry (

prior

ele

ctions)

−0.15 −0.10 −0.05 0.00 0.05 0.10 0.15Share Margin of Victory

1−30 31−100 101−200

Observations

Note: Local polynomial regression. Left: linear fit. Right: quadratic fit. Observations within CCT bandwidth

displayed. Bin selection method: mimicking variance, evenly spaced using spacings estimators. By year House

Back

A falsification testPredicting prior election result

Panel A. 1997 Panel B. 2000

−0.6

−0.4

−0.2

0.0

0.2

0.4

Sh

are

Ma

rgin

of

Vic

tory

(p

rio

r e

lectio

ns)


1−30 31−100 101−200

Observations

−0.5

0.0

0.5

Sh

are

Ma

rgin

of

Vic

tory

(p

rio

r e

lectio

ns)


1−30 31−100 101−200

Observations

Panel C. 2003 Panel D. 2007

−0.4

−0.2

0.0

0.2

0.4

Sh

are

Ma

rgin

of

Vic

tory

(p

rio

r e

lectio

ns)


1−30 31−100 101−200

Observations

−0.4

−0.2

0.0

0.2

0.4

Sh

are

Ma

rgin

of

Vic

tory

(p

rio

r e

lectio

ns)


1−30 31−100 101−200

ObservationsBack

Note: Local polynomial regression. Left: linear fit. Right: quadratic fit. Observations within CCT bandwidthdisplayed. Bin selection method: mimicking variance, evenly spaced using spacings estimators.

Balance on observable variablesNo effect on vote share for key party in prior elections of lower chamber

0.10

0.15

0.20

0.25

0.30

0.35V

ote

share

low

er

cham

ber

2010

−0.15 −0.10 −0.05 0.00 0.05 0.10 0.15Vote Share Margin of Victory for Governor’s Candidate

1−30 31−100 101−200

Observations

Note: Observations within CCT bandwidth and 95% confidence interval. Back

Political alignment and tests scoresSABER 11

(1) (2) (3) (4) (5) (6)2012 2013 2014 2012− 2010 2013− 2010 2014− 2010

Panel A. Dependent variable is the score of SABER 11: Language

Alignment -0.251 -0.359* -0.132 -0.105 -0.384* -0.117(0.311) (0.216) (0.235) (0.302) (0.212) (0.229)

Observations 1,897 1,950 1,947 1,722 1,708 1,674Bandwidth 0.110 0.0663 0.0826 0.108 0.106 0.112Obs. in bandwidth 553 261 390 529 525 574Mean dependent 0 0 0 0.0632 0.0652 0.0915Std. dev. dependent 1 1 1 0.773 0.785 0.756

Panel B. Dependent variable is the score of SABER 11: Math

Alignment -0.101 0.000931 -0.0289 -0.0368 -0.222 -0.0570(0.294) (0.194) (0.231) (0.173) (0.271) (0.205)

Observations 1,897 1,950 1,947 1,722 1,708 1,674Bandwidth 0.0933 0.105 0.0926 0.125 0.0848 0.111Obs. in bandwidth 437 521 437 648 394 566Mean dependent 0 0 0 0.0520 0.0620 0.0795Std. dev. dependent 1 1 1 0.769 0.813 0.740

Coverage Rate

(1) (2) (3) (4) (5)Basic Primary Media Secondary Total

Panel A. Dependent variable is coverage rate in 2012

Alignment -0.0602 -0.0244 -0.120 -0.132 -0.0769(0.101) (0.0827) (0.0951) (0.146) (0.104)

Observations 329 329 329 329 329Bandwidth 0.144 0.151 0.127 0.128 0.141Obs. in bandwidth 131 138 118 118 128Mean dependent 0.878 0.858 0.345 0.656 0.857Std. dev. dependent 0.200 0.199 0.152 0.209 0.196

Panel B. Dependent variable is coverage rate in 2013

Alignment -0.0779 -0.0178 -0.134 -0.134 -0.0920(0.112) (0.0904) (0.101) (0.149) (0.117)

Observations 329 329 329 329 329Bandwidth 0.144 0.165 0.132 0.127 0.142Obs. in bandwidth 132 153 121 117 129Mean dependent 0.874 0.845 0.354 0.676 0.857Std. dev. dependent 0.209 0.205 0.156 0.211 0.205

Party future prospects?

(1) (2) (3) (4) (5) (6)

Dependent variable is, in next election, ...Run... Win... Vote share..

Wint=0 0.0296* 0.0198 -0.0418** -0.0429** -0.0132* -0.0193**(0.0154) (0.0159) (0.0180) (0.0179) (0.00787) (0.00805)

Alignment 0.00826 0.00329 -0.00490(0.0209) (0.0215) (0.0104)

Wint=0 × Alignment 0.0525** 0.00564 0.0322**(0.0259) (0.0280) (0.0133)

Observations 5,488 5,425 5,488 5,425 5,488 5,425R-squared 0.586 0.588 0.172 0.172 0.443 0.444

Notes: Regression include party fixed effects and year fixed effects. Bias corrected coefficientsand robust standard errors clustered at municipality level (Calonico et al., 2014). Optimalbandwidth in all columns. * 10%, ** 5%, *** 1%.

Vote buying

(1) (2) (3) (4) (5) (6)

Risk of electoral Fraud (2018)

Alignment 0.492* 0.468* 0.366 0.524** 0.464** 0.577**(0.251) (0.262) (0.228) (0.255) (0.234) (0.272)



Notes: Mean for dependent variable is 1.21 and standard deviation equals 1.15. Alignment coefficientsand robust standard errors clustered at municipality level (Calonico et al., 2014). Optimal bandwidth in allcolumns. * 10%, ** 5%, *** 1%.

Back

Documents

Do ghosts exist? Corruption in the Colombian education systempubdocs.worldbank.org/en/603631530034379544/Fergusson-Harker-Molina... · Do ghosts exist? Corruption in the Colombian