How do Vouchers Work?Evidence from Colombia

Eric Bettinger, Case Western U

Michael Kremer, Harvard

Juan Saavedra, Harvard

7 June 2007World Bank

Educational Vouchers

• Controversial and Often Debated Educational Reform• Large-Scale Voucher Programs

– Chile & Colombia– US Sites (Milwaukee, Cleveland, Florida, Ohio)– Other Voucher-Like Programs

• Private Programs in US• Regulated Voucher Programs Internationally (e.g. Sweden, Japan,

The Netherlands)

• Motivation for Programs Differ Substantially Across Sites– Competition and Opportunity– Overcrowding

Evidence on Educational Vouchers

• International Evidence– Colombia– Chile

• US Evidence– New York City Scholarship Program– Milwaukee Voucher Program– Other Private/Public Programs

Welfare Implications of Vouchers

• Costs/Benefits to Students Directly Receiving the Voucher– Cost of Voucher– Change in Educational Outcomes

• Costs/Benefits to Students not Receiving the Voucher– Change in Peers– Competition– Change in Resources

Peer Effects and Vouchers

• Hsieh and Urquiola (2006)– Chilean Voucher Program– Key Finding was that Aggregate Outcomes Did Not

Change– Voucher Recipients Experienced a Positive Peer

Effect– Other “Left Behind” Students Experienced a Negative

Effect• Epple and Romano (1998)

– Model of Educational Vouchers– Peer Effects Need Not Be Zero Sum but Some

Students are Worse Off

Other Peer Effect Literature

• Sacerdote (2001)

• Duncan, Boisjoly, Kremer, Levy and Eccles (2005)

• Zimmerman (2003)

Simple Model with Peer Effects

• Consider the Following Educational Production Function

• Yi = Educational Outcome Student i• Xi = Socioeconomic Characteristics of Student i• = Average Characteristic of School• P = Program Participation

Model (cont.)

• Positive Voucher Effects but No Peer Effects

β1=0, β2>0

• Positive Voucher Effects are Solely Peer Effects

β1>0, β2=0

Voucher Comparisons

• Because of Randomization, Difference in Average Outcome of Voucher Lottery Winners and Losers Measures the Effect:

• Reported Effects in Angrist et. al. (2002, 2006)

21 )( loserwinner XX

Social Effects

• Depends on Values of β1 and β2

• If β1=0, β2>0, then social effect is β2

• If β1>0, β2=0, then no social effect in simple linear-in-means peer effect model

21 )( loserwinner XX

How does One Untangle Peer and Voucher Effects?

• Ideal Experiment:– Randomly Assign Voucher Students to Group

with “Worse” Peers– Measure Whether Voucher Effect is Present

When Peer Effect is Bad– If Voucher Effects are from Peer Effects, then

Voucher Effects Should Disappear or Even Be Negative

• Colombia PACES Program May Have Such an Experiment

Colombia’s PACES Program

• Objectives– Increase secondary school enrollment for poor

• Rules– Student must be entering 6th grade and under 15 years old– Students must provide evidence that they live in poor

neighborhood (strata 1 or 2)– Renewable through graduation unless student is retained in a

grade– Vouchers awarded by lottery if demand exceeds supply– Covered about 60% of fees

• Scope– 216 Municipalities have Participated– Over 125,000 secondary students have received support

Previous Work on Colombia

• Angrist, Bettinger, Bloom, King, Kremer (AER 2002)• Surveyed Voucher Applicants from Bogotá 1995 Lottery• Compared Voucher Lottery Winners and Losers• Effects after Three Years• Key Findings on Voucher Recipients:

– Increased Usage of Private Schools– Higher Educational Attainment– No Difference in Drop-out Rates– Less Grade Repetition– Higher Test Scores– Less Incidence of Teen-age Employment

More Previous Work on Colombia

• Angrist, Bettinger, Kremer (AER 2006)• Tracked Applicants to Bogotá 1995 Lottery

Using Administrative Records• Compared Voucher Lottery Winners and Losers• Effects after Six to Seven Years• Key Findings on Voucher Recipients:

– Increased Likelihood of Taking College Entrance Exam

– Improved Exam Performance in Math and Language

Other Details on Colombia’s PACES

• Students Had to be Accepted at a Private School PRIOR to the Voucher Lottery– Some Students Applied to Vocational Schools– Vocational Schools May Attract “Worse” Peers

• The Voucher was Portable in Theory but Not Practice– 1998 MIT Survey of Voucher Applicants Shows that

Voucher Winners Did Not Retain Voucher When Transferring Schools

Patterns of Attendance

School Applied To

Vocational

School Attended Three Years after Voucher Lottery

Winner

Vocational .5962

Non-Vocational .2692

Dropout .1346

Patterns of Attendance

School Applied To

Vocational

School Attended Three Years after Voucher Lottery

Winner Loser

Vocational .5962 .4264

Non-Vocational .2692 .4264

Dropout .1346 .1473

Patterns of Attendance

School Applied To

Vocational Non-Vocational

School Attended Three Years after Voucher Lottery

Winner Loser Winner Loser

Vocational .5962 .4264 .0369 .0626

Non-Vocational .2692 .4264 .7995 .7517

Dropout .1346 .1473 .1590 .1790

Voucher Winners and Vocational Schools

• Voucher Winners who Applied to Vocational Schools Stay in Vocational Schools

• Lack of Portability Creates Rigidity– Voucher Winners Stay Even if School is

Suboptimal

• Did Vocational Schools Have “Worse” Peers?

Were Vocational Schools Worse?

• WB Funded School Survey in 2006

• Attempted to Contact 300 Schools– Chose Schools with Most Voucher Applicants

Attending– Schools Represented 85 Percent of Voucher

Applicants

• Survey Gathered Significant Data on Peer and School Quality

Comparing Peer Quality

• Estimate for Sample of Students Who Applied to Vocational Schools Prior to Lottery

Wi = α + γVi +πZi+ ui

Wi = School characteristic of student i

Vi = Student i is voucher winner

Zi = Other Student Characteristics

• Key issue will be low sample size. We only know schools of attendance for about 200 voucher students who had wanted to attend vocational school.– Pooling Estimates

Diffs in Peer Quality Vocational School Applicants (1)

Losers' Mean

(2) Vouher Diff w/o Covars

(3) Voucher Diff W/ Covars

A. Academic Indicators of Peer Quality

Mean ICFES Score 46.50 (3.188)

-.1518 (.4234)

-.1863 (.4552)

Proportion of Entering Class who later take the ICFES

.7591 (.4225)

-.0429 (.0749)

-.0699 (.0852)

Proportion of Students Enrolling in College

.3141 (.3158)

-.0736* (.0421)

-.0897** (.0448)

Percentage of Students Who Dropped Out in Senior Yr

.0339 (.0339)

.0118* (.0063)

.0080 (.0065)

More Diffs in Peer Quality Vocational School Applicants (1)

Losers' Mean

(2) Vouher Diff w/o Covars

(3) Voucher Diff W/ Covars

B. Other Indicators of Peer Quality

Has Job Training Program .1517 (.3597)

.0894** (.042)

.0931** (.0448)

Has Tutoring for Disadvantaged Students

.2809 (.0451)

.0718 (.0557)

.0648 (.0569)

Fees 40752.6 (47212.4)

-8778.2 (5959.8)

-9497.9 (6151.0)

Specialized Facilities Available

.9367 (.3101)

-.0034 (.0157)

-.0015 (.0167)

How to Increase Power?

• Average Effect Sizes– Normalize Measures

• Standard Deviation Units• Monotonicity

– Estimate Effect Sizes Simultaneously• Fully Interacted Model• Similar to Seemingly Unrelated Regression

– Combine Coefficients• Kolmogorov-Smirnov

– If there is no Effect of a Program, the Family of T-Statistics Should Behave Like Normal Distribution

– Doesn’t Account for Correlation Between Outcomes

Average Effect Sizes

Category of School Characteristics

Difference by Voucher Status (std. dev units)

Standard Error

Number of Measures

Student College Entrance & Graduation Behavior

-.1671* .0942 8

Student Dropout Behavior .1188 .2272 3 Existence of Remedial or

Vocational Programs .2672** .1160 6

Total Fees Across All Categories

-.3870** .1757 2

School Appearance -.1589 .1329 4 School Facilities .1330** .0552 22 Teacher Characteristics -.4260** .1315 6

Family of T-Statistics Vocational Voucher Applicants

0.1

.2.3

De

nsity

-4 -2 0 2 4t

PDF for Vocational Schools

PDF of Standard Normal

Family of T-Statistics Non-Vocational Voucher Applicants

0.1

.2.3

.4D

en

sity

-4 -2 0 2 4t

PDF for Non-Vocational Schools

PDF of Standard Normal

Key Results on Peer Quality

• Among Voucher Applicants to Vocational Schools, Winners’ Peers Are– Less Likely to Attend College– Less Likely to Graduate– More Likely to Attend Remedial Programs– Attend Schools with Lower Fees

• Voucher Winners’ Schools Also Appear to Have– Less Qualified Teachers– More Facilities on Campus

• Same Differences are not Present Among Applicants to Non-Voucher Schools

Are Voucher Effects Still Present?Vocational Non-vocational Dependent Variable

(1) Losers' Means

(2) Regression-Adjusted Voucher Diff

(3) Losers' Means

(4) Regression-Adjusted Voucher Diff

A. Probability of Taking ICFES ID Match

.2875 (.4528)

.0586** (.0175)

ID & City Match

.2725 (.4454)

.0577** (.0174)

ID & Name Match

.2117 (.4087)

.0336** (.0164)

N 1200 2612 B. Performance Outcomes on the ICFES Math Score cond'l on

taking

42.39 (4.762)

.3094 (.3559)

Reading Score cond'l on taking

47.19 (5.450)

.3427 (.3962)

Math Score (Upper Bound Estimate)

-- 1.5270** (.3302)

Reading Score (Upper Bound Estimate)

-- 1.6949** (.3648)

N 319 874

Are Voucher Effects Still Present?Vocational Non-vocational Dependent Variable

(1) Losers' Means

(2) Regression-Adjusted Voucher Diff

(3) Losers' Means

(4) Regression-Adjusted Voucher Diff

A. Probability of Taking ICFES ID Match .2548

(.4364) .0607* (.0300)

.2875 (.4528)

.0586** (.0175)

ID & City Match .2520 (.4348)

.0494* (.0299)

.2725 (.4454)

.0577** (.0174)

ID & Name Match .1884 (.3915)

.0536* (.0281)

.2117 (.4087)

.0336** (.0164)

N 361 810 1200 2612 B. Performance Outcomes on the ICFES Math Score cond'l on

taking 41.46 (4.865)

.7661 (.6370)

42.39 (4.762)

.3094 (.3559)

Reading Score cond'l on taking

45.71 (5.951)

2.060** (.7804)

47.19 (5.450)

.3427 (.3962)

Math Score (Upper Bound Estimate)

-- 2.507** (.6156)

-- 1.5270** (.3302)

Reading Score (Upper Bound Estimate)

-- 4.364** (.7376)

-- 1.6949** (.3648)

N 87 256 319 874

Key Conclusions

• Among Students Who Originally Applied to Vocational Schools, Voucher Winners . . .– Attended Schools with Worse Observable

Characteristics– Had Higher Likelihood of Taking the ICFES

College Entrance Exam– Had Higher Test Scores on ICFES Exam

• Peer Effects cannot Explain Voucher Effects

Voucher Mechanisms?

• Incentives– Vouchers were Renewable Conditional on Passing– Emerging Literature on Incentives (e.g. Kremer, Miguel, and

Thornton 2005)

• Attending School that Students Value More– Vocational Schools Increasingly Focused on Emerging Service

Industry– Private Sector Has Adjusted Quickly to Changes in Labor Market– Differences in Apprenticeship Opportunities– Industrial versus Commercial Education Curricula

• Limited Commercial Education Slots in Public Sector• Private Commercial Schools Retain Students

In Conclusion

• Peer Effects are Important in Education• Disentangling Voucher and Peer Effects is Difficult

– Need Unique Experiment where Vouchers Assign Students to “Worse” Peers

– Applicants to Vocational Schools (PRIOR) to Colombian Voucher Lottery

• Vocational Voucher Winners Attend Schools with “Worse” Peers

• Vocational Voucher Winners Have Higher Academic Achievement Than Voucher Lottery Losers

• Peer Effects May Not Explain the Voucher Effects• Other Mechanisms are Plausible