ABSTRACT
Title of dissertation: DECENTRALIZATION AND EDUCATION:
AN EMPIRICAL INVESTIGATION
Felipe Barrera-Osorio, Doctor of Philosophy, 2003
Dissertation directed by: Professor Robert Schwab
Department of Economics
This thesis analyses, using a quasi-experimental approach, the relationship between de-
centralization and education in Colombia, where in 1991 decentralization of the State was
implemented. The thesis focuses on two relationships:  rst, the relationship between decen-
tralization and quality of education; and second, this relationship across individuals with
di erent incomes.
Theoretically, decentralization may increase the e ciency in the provision of education,
and therefore, we would expect an increment in educational quality in schools a ected by
decentralization. Furthermore, decentralization may create a more unequal distribution of
educational quality, and therefore, we expect that the impact of decentralization is asym-
metric with respect to income.
The thesis makes use of a new dataset that comes from two sources. First, data from the
Ministry of Education provide an important array of school characteristics. Second, data
from the ICFES, the institute in charge of administering standardized tests in Colombia,
provide test scores and characteristics of individuals.
We present three types of quasi-experimental models based on di erent control and treat-
ment groups. First, we estimate the e ect of decentralization on public schools, using private
ones as a comparison group. Second, we restrict the estimation to public schools, but now
the treatment group is comprised of schools in initially highly dependent departments, and
the control group of schools in departments with initially highly independent relationship
with the central government. Finally, the third model is a nested model of the  rst two. It
is a more  exible model allowing nationwide e ects and public school e ects.
The empirical results are mixed. We  nd a positive impact of decentralization in the  rst
two models. However, the third model presents a negative result. The results from the tests
on an asymmetrical impact of decentralization, depending on income, are mixed as well. In
the last two models, the results are symmetric. However, in the  rst model the results are
asymmetric, and interestingly, in favour of low-income individuals. That is, decentralization
increases the test scores for individuals at the left tail of the income distribution.
DECENTRALIZATION AND EDUCATION: AN EMPIRICAL INVESTIGATION
by
Felipe Barrera-Osorio
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial ful llment
of the requirements for the degree of
Doctor of Philosophy
2003
Advisory Committee:
Professor Robert Schwab, Chair
Professor Jonah Gelbach
Professor Wallace Oates
Professor Jennifer Rice
Professor Seth Sanders
c Copyright by
Felipe Barrera-Osorio
2003
Acknowledgements
Several professors at Maryland helped me during my very happy years at the University.
I would like to thank Darrell Hueth and Rudy Hommes. Darrell was brave enough to take
the advice of Rudy to help me get an opportunity to study at Maryland. Robert Schwab was
my teacher of microeconomics, at a moment in which I needed one like him. He read in detail
my dissertation, allowing me to  nd mistakes and holes. His guidance was critical in the
making of the thesis. Judith Hellerstein read the  rst drafts. Conversations with her were
always an exercise of coherence that allowed me to build stronger arguments. The members
of my committee not only provided very important comments to improve the thesis, but
they also made my defense a very pleasant and challenging exercise. To Professors Oates,
Rice and Sanders, my thankfulness.
Jonah Gelbach provided help not only with the ideas and technical problems, but also
with detailed comments about the form and writing of the thesis. He provided ideas that,
with boldness, I took as mine. He was able to combine the precise amount of hard criticism
with valuable encouragement. Literally, without his help, I would not have been able to
 nish. He is a great teacher and, probably more important, a good friend. Jonah, thank
you.
Vickie Fletcher, Mary Louise Snidow and Marie Speake helped me to make my transit
through the university a very smooth experience. I hope that one day I can return all the
favours that they did for me.
In the process of building the thesis, many people in Colombia gave me support at a
ii
moment in which it was not abundant. Eduardo Lora gave me an opportunity, and friendship,
that I will never forget. Juan Jose Echavarr ia brought me to Fedesarrollo the summer of
2000. During those two months I had the initial idea of the thesis. He paid me to think
and he gave me for free his ideas and insights on the problem. A the Ministry of Education,
Margarita L opez, Margarita Pe~na and Mauricio Castillo allowed me to get the necessary
data in order to do the thesis, and Carlos Pardo did the same with the data at the ICFES.
To all of them, my deepest gratitude.
The Central Bank of Colombia and Colciencias provided very important monetary sup-
port. Jos e Dar io Uribe and the committee of scholarships of the Central Bank was very
generous with me. In Colciencias, I would like to thank Erin Dell and Juan Ra ul Mendoza,
the two people in charge of my case, and Hern an Jaramillo for believing in me. All of them
helped me in a great way, although passive, to  nish this journey.
Rocio Osorio, my mother, was widowed at the age of 36. She, alone, made true the
dreams, whatever they were, of her three sons. She is always the most enthusiastic follower
of our enterprises. Mother, I just  nished one that is the start of several. Thank you.
My wife, Katja Vinha, did an enormous amount of this dissertation, in the real sense of
the sentence. She read several times my drafts, correcting my poor English. She criticized
weak arguments and she helped me preparing the defenses of the proposal and the thesis.
On top of all this, she tolerated all my moods when things were not as bright as they are
today. For all her intelligence, bravery and honesty, thanks. I will always  ght to deserve
you.
iii
Table of Contents
List of Tables.................................................................................................vi
1. Introduction................................................................................................1
2. Literature Review........................................................................................4
2.1 The trade-o between e ciency and equality............................................6
2.2 Equalization of Expenditures...................................................................20
2.3 Solving the equity-e ciency trade-o ......................................................26
2.4 A conclusion.............................................................................................32
3. Decentralization in Colombia......................................................................34
3.1 Introducion................................................................................................34
3.2 Decentralization in Colombia.....................................................................36
3.3 Education...................................................................................................42
3.4 Educational data........................................................................................47
3.5 Changes in characteristics of schools..........................................................58
4. Empirical strategy and estimation...............................................................73
iv
4.1 Introduction...................................................................................................73
4.2 Quality of education: Production Function...................................................75
4.3 A framework for estimation: Quasi-experiments...........................................79
4.4 Estimation.....................................................................................................89
5. Conclusions...................................................................................................111
v
List of Tables
3.1 Main Laws in the Nineties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Functions by level of government . . . . . . . . . . . . . . . . . . . . . . . . 65
3.3 Decentralization in numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4 Kindergarten . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.4 Kindergarten, cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5 Kindergarten. DD estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Primary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.6 Primary, cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.7 Primary. DD estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.8 Secondary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.8 Secondary, cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.9 Secondary. Di erence in di erence . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1 Models used in the Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 Model 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.3 Model 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
vi
4.4 Model 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.5 Summary statistics Model 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.6 Summary statistics Model 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.7 Regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.8 Total e ect of decentralization: Model 1 . . . . . . . . . . . . . . . . . . . . 107
4.9 Total e ect of decentralization: Model 2 . . . . . . . . . . . . . . . . . . . . 107
4.10 Total e ect of decentralization: Model 3 . . . . . . . . . . . . . . . . . . . . 107
4.11 Impact di erences across income groups: Model 1 . . . . . . . . . . . . . . . 108
4.12 Impact di erences across income groups: Model 2 . . . . . . . . . . . . . . . 109
4.13 Impact di erences across income groups: Model 3 . . . . . . . . . . . . . . . 110
vii
Chapter 1
Introduction
In 1971, the State Court of California, United States, ordered a general equalization of
expenditure on education in localities within the state. In simple terms, the state changed
from a highly decentralized regime in which each locality decides its level of expenditures
on education to a centralized one in which the state decides how much each locality should
spend. The objective was to introduce equal expenditures across localities.
In 1991, the country of Colombia enacted a new Constitution that changed several as-
pects of the political, economic and administrative order in the country. One of the most
fundamental changes was the transformation of the country?s public sector from a highly
centralized system to a more decentralized one in order to increase e ciency and coverage
in the provision of local public goods.
The above two quasi-experiments may imply an implicit trade-o between equity and
e ciency in the provision of services. Apparently, di erent degrees of decentralization are
1
used to accomplish di erent objectives. In California, the change from a decentralized system
to a centralized one was made to promote equality, while in Colombia, the change from a
centralized regime to a decentralized one was implemented to improve e ciency.
The problem analyzed in this thesis is a critical one in economics. Does a decentralized
system deliver more e ciency than a centralized one? This question relies on several branches
of economics: microeconomics and the formation of prices; public economics and the Tiebout
hypothesis of voting with the feet; and human capital theory and the production of quality
of education. The answer to the question is still open. Chapter 2 presents an overview of
di erent theories on the topic, as well as some empirical results for the United States.
The primary objective of this dissertation is to investigate the impact of decentralization
on education in Colombia, which implemented a decentralization reform during the nineties.
The two main hypotheses of the thesis are the following. First, that decentralization
increases the e ciency in the provision of education, and therefore, has positively impacted
standardized test scores. Second, that decentralization has impacted asymmetrically indi-
viduals with di erent incomes, and therefore, decentralization has impacted di erently the
scores of high- and low-income individuals.
The Colombian reform is discussed in Chapter 3. As argued in that chapter, Colombia
decentralized in a limited way, transferring only some of the expenditure and tax functions
to the localities. That chapter presents data on the impact of decentralization on inputs of
the schools. As it argues, probably the most important change of decentralization is on the
characteristics of the schools, such as teacher quality, and that change ultimately impacts
2
test scores.
Chapter 4 lays out the empirical strategy that is based on treatment and control es-
timations. Speci cally, the chapter analyses changes in test scores in those schools that
experienced decentralization and those that did not. Chapter 4 also presents estimation
results. Chapter 5 discusses general conclusions.
3
Chapter 2
Literature Review
This chapter focuses on the theoretical and empirical literature about decentralization. The
next part of the chapter analyses several theoretical papers that focus on the debate on
the trade-o between inequality and e ciency. Under which conditions will decentralization
deliver increased inequality? Under which conditions will decentralization induce e ciency?
Is there any regime in which the trade o does not exist?
The third section of the chapter discusses the empirical evidence in the United States.
Several states faced a court order for the equalization of expenditures, quasi-experiments that
allow researchers to investigate the e ects of such changes on e ciency and distribution. The
empirical papers in the literature primarily investigate the e ects on distribution, most likely
because the reforms targeted reduction of inequality. These papers ask whether states that
enacted equalization of educational expenditures were successful in reducing inequality and
whether the court order was necessary for the change. Another question in the empirical
4
arena is whether expenditure equalization has caused convergence to the ex ante expenditure
levels of localities with the lowest expenditure levels or the higher ones. Also, this part present
evidence derived from general equilibrium models and simulation experiments that support
parts of the empirical  ndings.
The fourth section of this chapter presents a discussion from the political economy view
on how to break the apparent link between e ciency and inequality while preserving a
decentralized regime. First, there is an extensive literature on vouchers, which can be used
to induce equality in education expenditures by giving households the power to decide from
where to consume educational services, so that decentralization of the educational system is
preserved. Second, some papers discuss a system in which state taxes that are distributed
in such a way as to e ect equalization expenditures while preserving local jurisdiction over
administration and the pattern of expenditure.
At this point, it is important to de ne decentralization. Decentralization is the transfer
of several political and economical functions from the central government to the local gov-
ernments. In the sphere of economics, these functions include the decisions on expenditure
and revenues. For example, a central government will decide the amount of expenditure
on education countrywide in a centralized regime, whereas in a decentralized system local
governments (e.g. the municipal government) will decide the amount of expenditure on
education for its own community. Likewise, in a central regime, the national government
will decide tax rates and bases; in contrast, a decentralized regime will give this power to
localities.
5
Clearly, there are several degrees of decentralizations, ones with more functions trans-
ferred to the localities, and others where the localities are more restricted in the expenditure
and revenue decisions. Colombia, as the next chapter discusses in depth, has the character-
istics of moderate decentralization. Part of the expenditure of localities is outlined by the
national government, and taxation decisions are also limited by some central regulation.
2.1 The trade-o between e ciency and equality
The theoretical papers of this section present arguments on the potential e ects of decentral-
ization and centralization on e ciency and distribution. This section has three subsections:
informational issues and sorting, supply-side considerations, and externalities and strati ca-
tion.
In the  rst set of papers on informational issues the main conclusion is that the informa-
tion held by local service providers suggest e ciency gains in the provision of the services in
contrast with a central government, which presumably does not have complete information.
The sorting argument on public service provision draws conclusions from several papers in
which people vote with their feet to  nd their preferred consumption bundle for public ser-
vices and taxes. Households will locate in those communities that provide their preferred
consumption bundle and will pay the price (tax) charged for these services in the community.
This type of e ciency argument embodies a correlation between high-income and high-level
provision of services. The second set of papers on supply-side considerations looks into the
problem of provision of services and the incentives that local suppliers have to give in order
6
to achieve the optimal level of services. The last set of papers on externalities and strati -
cation argues in favor of centralization based on the presence of local externalities. There
are e ciency gains from centralization if there are peer e ects in education. There may,
however, be general externalities that would favor decentralization on e ciency grounds.
The section ends with a discussion of this dichotomy.
2.1.1 Informational issues and sorting
Behind the idea that decentralization induces a more e cient allocation of resources is the
idea of F.A.Hayek [34] that a decentralized system will provided a better allocation of re-
sources than a centralized system because agents at the micro level have more and better
information than any central power. Therefore individual agents will make better decisions
for themselves than the ones that the central power may take in their behalf. Clearly, some
individuals are going to demand low levels of education and others high levels, depending
on their preferences for education.
More recent literature (Tiebout [59], Oates [53] and the general overview in IDB [6] and
Oates [52]) emphasizes the idea that a decentralized system will provide better public services
because, in a Tiebout-type model, people choose the location that delivers the optimal set of
taxes and expenditures, given their preferences. If the services or taxes are not the preferred
ones, individuals will vote with their feet, leaving the location for another one that provides
a preferred set of taxes and services. Given a population with heterogeneous preferences,
there will be several di erent bundles of taxes and services in equilibrium, and households
7
will leave and enter communities according to their preferences. There will be high-service,
high-tax and low-service, low-tax communities. In this case, e ciency refers to the concept
of optimal choice (for services and taxes). People are maximizing utility subject to the set
of taxes and services o ered, so that in equilibrium people do not leave the communities.1
Hamilton [31] develops Tiebout?s idea and incorporates housing price and local property
taxes into the analysis. In a perfectly decentralized system, housing prices implicitly re ect
local taxes and the type of services that each community receives. Owners of houses will
see increases in house prices as the quality of services provided in the community increases,
holding constant tax bills. On the other hand, low-income people will live in low rent places
with low levels of services. In these localities, tax rates may be high and the total revenue
from taxes low since the tax base is smaller than in rich communities. Therefore, there is a
link between services, taxes and prices of houses. Communities with high levels of services
probably will pay a larger amount of dollars in taxes, with housing prices re ecting the high
level of services and taxes. Owners will be compensated for the high tax payments with
higher housing prices and better quality of services.
Sonstelie and Portney [57] show that, if a community allocates resources to the provision
of local public goods such that it maximizes the value of the property within its boundaries,
the allocation is Pareto optimal. The implication is that a decentralized system, in which
the local community decides expenditures and taxes for local public goods, is e cient in a
Pareto sense since people internalize the e ect of expenditures and taxes through the prices
1Even in the case of low level migration, as Oates [52] argues, decentralization may induce a higher level
of e ciency because the information arguments still hold.
8
of property.
The trade o between e ciency and inequality is implicitly present in the two previous
arguments. People choose taxes, services and rents to maximize their utility, a fact that
implies e ciency. There is, however, a correlation between income and the chosen levels of
services, taxes and rent. High-income families will be in high quality localities, with high
rents and high collection of taxes, and the contrary for low-income people.2
It is important to point out that even if decentralization, based on the information argu-
ments, brings theoretical e ciency gains some authors argue that in practice decentralization
faces important problems, including technical constraints and corruption. Local authorities
may have a lower technical capacity than the central government in deciding expenditure,
assignation of resources, collection of taxes, etc.3 If this is true, it is possible to observe a
dip in the quality of services immediately after the decentralization starts.
With respect to corruption, local governments with new resources may be prone to this
problem. But, the decentralization can give more instruments to the community to exercise
checks and balances. Since local service providers control the resources and programs to
be implemented they also can be held responsible for the results. Moreover, under the de-
centralized system the relationship between the units providing services and the community
receiving them is direct. Therefore, the local community can \demand" better services from
2However, if the model incorporates not only people but also companies, the apparent correlation may
vanish. Firms tend to locate in places in which rents are low. If this is the case, the tax rate and the tax
collection over  rms can compensate for the low taxes of individuals. The community may be low-income
with a high total tax base and a high level/quality of services (Fischel [27]).
3For precisely that reason, however, decentralization will force local governments to increase their technical
capacities via learning by doing.
9
suppliers.
2.1.2 Supply-side considerations
So far, the three models reviewed until now present the hypothetical situations of a world
that either has or does not have perfect decentralization. However, it is quite di erent
to analyze a country that changes from one regime to another, say from a decentralized
one to a centralized one. In terms of North [51], a system that is in place may have some
institutions that are di cult to change. Once the regime changes, old institutions may remain
and create certain dynamics that are important to analyze. For instance, decentralization
creates dynamics in the price of property as we saw above. Once the system changes from
a decentralized regime to a centralized one, and since taxes somehow re ect previously the
quality (expenditure) of services, the change in the system decouples the relationship between
taxes and services. This fact is precisely the starting point of Hoxby [41], Fischel [28] and
Fischel [26]
Hoxby [41] uses the notion that behind a certain quality and amount of expenditure
on education there is an implicit cost, which is the tax that provides the revenues for the
expenditure. Hoxby calculates this implicit tax price for several states. In some states, for
which the court ordered a strong equalization of expenditures, the local authorities have to
raise more than one dollar in revenue to spend a dollar. In these states, actual expendi-
ture is discouraged because equalization favors the localities within the state that have low
expenditure in education.
10
In a similar argument, Fischel [28] and Fischel [26] argue that an equalization of expendi-
tures will fracture the link between services, taxes and capitalization of houses and therefore
individuals will want to introduce limits on property taxes.
Campbell and Fischel [16] study the particular case of New Hampshire?s 1992 elections
in which a candidate proposed property tax limits and a state income tax earmarked for
education. The political basis for this proposal was that the political system did not provide
the most preferred reforms for the average citizen, because some people who did not want
these reforms had a considerable amount of power in the political process. However, the
candidate who proposed the reforms did not win the election. This fact implies that the
median voter was against the reforms, and by a transitivity argument, the median voter was
against the court orders as well. Campbell and Fischel  nd that the  scal proposal was an
important factor deciding the individuals? vote among various other campaign issues.
In summary, when individuals choose localities, they are implicitly choosing a set of taxes
and services. If the taxes are local taxes (mainly taxes on property), the tax works as a fee
for the public local services. House prices and taxes will re ect quality of public services.
These relationships induce a potential positive correlation between income and quality of
services that is at the heart of the trade o between equality and e ciency.
It is important to notice that the demand side drives the previous arguments. The
e ciency argument is that individual will be maximizing utility in the consumption of ed-
ucation. However, it is not clear that the local provider of the services will in fact deliver
the optimal demand of services. Hanushek [33], Hanushek [32], Manski [47] and Hoxby [40]
11
present a fundamental critique of the empirical literature on education that treats quality
as the output of an educational production function whose inputs are individual and schools
characteristics. These articles argue that the objective function of the school is not clear. For
instance, it is possible that among several objectives of a public school are the \integration"
of di erent groups of society, creating civic conscience, or job stability for teachers, just to
mention a few candidates.4
Therefore, it is important to know what induces the local provider of the services to deliver
the \right" level of services, e.g. the one demanded by the community. Decentralization can
open space for the active and direct participation of each locality?s individuals in school
decisions. The clearest example of this is the board system in which parents of children in
the school can actively make decisions such as to change the principal, remove bad teachers,
etc. In contrast, in a centralized system, central actors (usually the minister of  nance
and the teachers unions) make decisions and local authorities are simply instruments in
ful lling these decisions. In a centralized system one problem is that the central government
is unlikely to make decisions that accommodate the requirements of each locality, since it
generally does not have adequate information.5 Moreover, decisions at the central level can
re ect the preferences of the people making them rather than the preferences of the general
population. For instance, the teachers union may be interested in introducing job stability
for teachers in each locality.
Hoxby [40] stresses the importance of the supply side of education in the e ciency dis-
4A more detailed discussion is undertaken in the empirical chapter.
5This argument was made by Hayek [34].
12
cussion. Decentralization provides the means for individuals to enforce the optimal provision
of services. The providers will, under this system, deliver the optimal level of services for
the locality based on its collective preferences.
Hoxby [40] presents a principal-agent model in which the principal is the community
and the agent is the local school. The essential insight of the model is that the principal
(the board of directors of the school) can induce the optimal quality (and minimization of
cost by the school) in the presence of incomplete observation due to the Tiebout-Hamilton
mechanism of capitalization. In a (perfect) Tiebout world, cost-minimizing schools prevail.
However, the model assumes that characteristics are veri able by the board of directors
of the school, which may or may not be the case. Hoxby assumes that the board cannot
observe either some parameter in the cost function of the school or the e ort of the schools,
but that Tiebout capitalization allows the board to infer the e ort that the school is making
to minimize costs. Black [14] presents an estimation that is in line with this argument. She
infers the value that parents place on the quality of education by looking at housing prices.
She compares the prices of houses in the same neighborhood (i.e. across the street from
each other) but in di erent school districts. In this way she is able to isolate the impact of
school quality from other sources of variation like di erences in tastes, characteristics of the
neighborhood, etc.
As we saw, capitalization is the key link between quality of services and prices of houses.
Individuals will pay higher taxes in return for better services and higher property values.
People who own houses will be willing to invest in better services since the value of their
13
property increases with service quality. People who want better services will pay for them
through taxes and higher rents. Under perfect decentralization, the board of directors can
infer the type of services that the school is providing by means of property values.6
When there is not perfect decentralization and the Tiebout mechanism does not hold
(for instance when a court orders an equalization of expenditures), the link between services,
taxes and housing prices is broken, and the local community cannot infer the level of e ort
by the school, inducing a loss in e ciency.7
As in the previous papers in which the emphasis is on the demand side, in Hoxby?s paper
the trade o between equity and e ciency is present. Decentralization can give the com-
munity instruments to induce e ciency from part of the provider of the services. However,
quality is higher in those places in which the tax base is higher.
In conclusion, there are several theoretical reasons why a decentralized system may induce
a higher level of e ciency than a centralized one. First, people will be able to consume
the optimal quality and quantity of education. Second, decentralization allows people to
induce the local provider of the services to deliver the preferred amount. By contrast, in
a centralized system the authority will target spending to match the preferences of some
agent (presumably the median voter). The rest of the people will need to accommodate
their demands to this level. Moreover, local providers will not have any interaction with
local communities: they will receive mandates from a central government and will not have
6One problem of the argument is the fact that property values re ect a complete vector of several local
amenities, and not only local quality of education.
7If education has, however, a positive externality, and education of the whole society enters in the utility
function of each locality, then there may be a case for equalization of expenditures on the grounds of e ciency
if the equalization actually increases the overall education of society. More discussion on this issue above.
14
any incentive to deliver the optimal choices of the local community. Furthermore, in the
worst-case scenario the objective function of the decision makers may not match the median
voter?s at all, but rather be that of their own group.8 However, the gains in e ciency from
decentralization come with a cost in general cases. Rich communities will spend more on
services than poor ones. If quality of education depends on expenditure, then a perpetual
circle of inequality is created.
2.1.3 Externalities and strati cation
Despite the apparent con ict between equity and e ciency and levels of decentralization,
some theoretical models that include local educational externalities predict that the trade-
o will not exist in equilibrium. Moreover, these types of models (mainly Benabou [10],
Benabou [9] and Benabou [11]) present theoretical arguments in favor of a centralized regime
since it can theoretically deliver both e ciency and equity.
Benabou [10] presents a model with human capital and  nancial asset accumulation
that allows the analysis of distributional aspects. The main aim of the paper is to see
the impact of funding sources (state or local) on the over overall economic surplus, student
achievement and equality of opportunity for education. In the model, individuals make three
main decisions: location, capital accumulation and consumption (which is correlated with
expenditure on the public good). The model consists of a simple individual maximization
8Clearly, under decentralization the local decision makers can also try to enforce their own objective
function. However, in this case the community may have more instruments to induce decisions towards the
best interest of the locality than in a centralized regime.
15
problem in which utility depends directly on the quality of education. The model explicitly
includes a positive externality of education, since the individual?s education quality depends
on the quality of education of the community where the individual lives. In other words,
the community?s stock of human capital impacts positively the individual?s accumulation
of education. Hoxby [39] presents a strong case in favor of peer e ect externalities. She
uses gender and race changes in adjacent cohorts as the source of variation to identify the
externality. She  nds that the achievement of classmates has an important e ect on the
achievement of individuals.
Equilibrium conditions specify the distribution of people across the city. There are two
cases, a strati ed equilibrium in which the rich and the poor individuals live in di erent areas,
and an integrated equilibrium in which the rich and the poor live in both communities. If
there are complementarities between families? human capital and the quality of education
of the community, there will be strati cation. If there are (small) imperfections in the
capital market, so that the opportunity cost of funds for poor families is higher (for instance,
di erent interest rates for poor people), there will also be strati cation. In addition, small
di erences in lifetime resources are su cient to cause strati cation. Given strati cation,
wealth increases in rich communities and decreases in poor communities. This strati cation
creates a dynamic force in favor of segregation.
Inequality in this model increases over time under general conditions. Rich parents invest
more in education, live in wealthier localities and live in communities with high human cap-
ital. Under local externalities, this combination will induce increases in the human capital
16
of the individuals, creating a virtuous cycle within the rich community. The contrary will
occur in the poor community. However, capitalization has an e ect in the income distribu-
tion between poor and rich communities. Since rent, or the price of property, in the rich
community is higher than in the poor community, the poor neighborhood faces low prices
that induce an income e ect in favor of the poor. Also, assuming that rich people own all
houses, the lower price of rent in the poor community will decrease the income of rich people.
Despite these e ects, they are not strong enough to break the strati ed, stable equilibrium.
Benabou presents a measure of e ciency in the presence of segregation that depends on
three margins. First, who bene ts more from an increase in the level of community quality?
Presumably, it bene ts more the rich families. Second, who bene ts more on the margin from
increases in quality? Given that quality is high in rich communities, and that presumably
the utility function is concave with respect to quality, the e ect of segregation is quite small
for rich communities and it implies an important loss for poor communities. Third, in what
type of community will a highly-educated family contribute more to increase the level of
education of the community? Presumably, in a poorer community. Therefore, under certain
conditions, segregation can induce losses in e ciency since the losses of the poor may be
greater than the gains of the rich.
Three main conclusions result from the di erent ways to  nance education. First, since
strati cation is ine cient in the presence of local externalities, and decentralization may
induce strati cation, decentralization may be ine cient. Moreover, decentralization also
induces inequality when expenditure on the public good and the mean level of human capital
17
of the community are complements because rich communities will spend more than poor ones.
Second, forced equalization of expenditure on schools may leave strati cation intact, while
reducing wealth since equalization brings e ciency costs. In this case, the movement to a
more centralized system will level the  eld by bringing down the rich rather than raising the
poor. Despite this, inequality will decrease in this scenario. Third, in the case where the
individual human capital and the community human capital level are negatively related in
the production function of education, equalization of expenditures may lead to integration
and there will be gains in e ciency and overall student achievement.
Benabou [9] presents another perspective on the same problem. In this model, education
is a public good. There are two types of skills, high-quality and low-quality. The cost of
acquiring high-quality skills depends on the proportion of people with high-quality skills in
the community: the greater the proportion of high-skill people in the community, the lesser
the cost of acquiring high-quality skills. The dynamic of the problem is as follows. Suppose
the economy starts in a symmetric equilibrium, i.e. both communities with the same number
of high-skill individuals. Now assume that, for some exogenous reason, the proportion of
high-skilled people increases by a small amount in one community. This community becomes
more attractive for all agents, so rents increase there. But as soon as rents increase, only
those who can bid for higher rents will do it. If there is a positive correlation between skills
and income, only high-skill individuals would be able to move to the community with a
greater proportion of high-skilled people. The only stable equilibrium is one in which the
high-ability individuals are concentrated in one place.
18
The previous two models center their analyses on the demand side and leave aside supply
considerations. Speci cally, expenditure on education depends directly on household con-
sumption. However, it is argued above that optimal educational spending depends not only
on the individual?s maximization problem, but also on the interaction between the commu-
nity and the local supplier of education. The objective function of the local service providers
is not clear, and in the two models optimal provision of services is taken as given.
Another key issue in the two models is the presence of local externalities, not only among
individual human capital levels and the average community level of human capital, but also
between educational expenditures and human capital.9 However, it is possible to argue that
both local and general externalities are important in the accumulation of human capital. A
general externality is one that enhances the productivity of all agents, independent of the
type and location of the agent. One example is computers. Once the stock of knowledge
in a society reaches a certain level, individuals start producing new technologies such as
computers. Computers then reach, via schools, all of society, enhancing productivity of all
individuals independent of location.
Another example is education. It is possible to argue that education has a general positive
externality, a ecting the whole population. A more educated teacher can provide better
education to individuals, independent of the skill level of these individuals. An increase in
the stock of education will lead to increases in the education of everyone in the community,
9The way in which the human capital accumulation function is modeled plays a critical role: the marginal
productivity of human capital is decreasing, but the total productivity has increasing returns on mean human
capital of the community.
19
including future teachers as well current ones.10
In the case at hand, suppose that decentralization, via e ciency gains, increases the aver-
age level of human capital by comparison to a centralized regime. If human capital presents
positive general externalities, decentralization makes everyone better o , even though in-
equality can increase or stay constant.
2.2 Equalization of Expenditures
Several states in the United States have been experimenting with the equalization of expendi-
tures in education. These experiments have drawn strong attention from several researchers.
The literature on this topic includes empirical work with actual data and general equilibrium
models that try to analyze the potential impacts of these types of measurements.
The empirical evidence is crucial in understanding the e ects of decentralization or cen-
tralization. In principle it is possible to use the experiments in US to answer several questions.
First, does a court order for the equalization of expenditures across localities actually deliver
equalization? Second, if it does induce equalization, what type of equalization is it? For in-
stance, do all localities spend the same amount as the one that was spending the least under
perfect decentralization or does equalization raise expenditures in localities that previously
had low expenditures? Third, what is the impact of equalization on income distribution?
Finally, what is the impact of equalization on the quality of education?
10Some educational externalities are local. For example, a community of scientists with very specialized
knowledge will have positive spillover e ects at the cutting edge of their  eld only among equally prepared
colleagues.
20
This section has three subsections. The  rst discusses the experience in California where
strong equalization was enacted; the second considers equalization in general; and the third
discusses simulation and general equilibrium models.
The  rst section presents empirical evidence based on the  rst case of court order equal-
ization in California. The strong equalization of expenditure caused several results such as
the lowering in the mean expenditure per pupil in the state. The second set of articles on
equalization in general shows that the general conclusion may be di erent from those ob-
served in California. Finally, the last section presents some general equilibrium models and
simulations that corroborates some of the empirical conclusion of the previous part.
2.2.1 The California case: strong equalization
California was one of the  rst states to which a court ordered equalization of expenditures on
education, and much of the empirical literature focuses on this case. Silva and Sontelie [55],
in their empirical work of the California case, point out that two e ects are present when
a court orders equalization of educational expenditures: an income e ect and a price e ect.
The income e ect tends to reduce expenditure when median income is less than the mean
income. The price e ect can be positive or negative, depending on whether state taxes are
more progressive than local taxes. If so, then the \price" for education decreases, inducing
increased expenditure.
Silva and Sontelie [55]  nd an overall decrease in the per-capita expenditure across lo-
calities in California. They argue that the reduction in expenditure can be explained by
21
two factors: the court ordered equalization and an increase in enrollment. With respect to
equalization, the authors  nd that the income e ect (which induced a lower expenditure)
dominates the price e ect (which is positive and induced an increase in expenditure). The
other part of the explanation in expenditure reduction is due to the high increases in enroll-
ment in California over the last decades. If enrollment increases, for any given amount of
expenditure, per-capita expenditure decreases.
In the already mentioned article by Hoxby [41], the author  nds that California had
a strong equalization scheme. Strong equalizations are ones in which the implicit ex post
price of one dollar of educational expenditure is greater than one. This induces reduction in
average per-pupil expenditure because the reduction in local expenditure is more than the
increase in state expenditure. For this reason, in California the equalization was towards
to bottom of the expenditure distribution. In other words, rich localities reduced their
expenditures, converging to the expenditures of poor localities.
2.2.2 Equalization considered more generally
Evans, Murray and Schwab [23] and Evans, Murray and Schwab [24] (henceforth referred to
as EMS) present a general overview of the literature on education and court decisions across
several states in the USA. The authors stress the tension between the goal of equalization
and local control of public schools since local control has led to signi cant di erences in
education spending. Basically poor school districts have little property wealth that they can
22
tax, so they impose a higher tax rate in order to compensate for the base.11 Despite this,
poorer districts spend less per student than wealthier ones.
The  rst empirical  nding of EMS is that, including all states that ordered equalization
of expenditures, real resources per student grew for the period 1972-92. Second, the amount
of money spent directly by localities from local revenues in education was possibly crowded
out by money that the locality received from the state. It is possible that the court orders
were a form of tax relief for low-income localities. A third empirical  nding of EMS is that
several measures of district spending inequality decreased between 1972 and 1992 for the
states that mandated equalization of expenditure. Moreover, the convergence in education
spending is not towards the bottom, since spending rose in the low spending districts.
On another front, recalling the argument of Fischel [26], if there exists equalization
of expenditures, the link between spending and preferences will be broken. Families will
put children in private schools, and the political support for public education will decline,
leading to lower spending in schools, tax revolts and even further decline in the quality of
public education. Empirically this implies that the causality runs from court mandates to
tax limitation. However, EMS  nd the contrary: in almost all states with court ordered
equalizations, tax limitations occurred before the court mandates.
EMS review some studies with respect to the impact of court mandates on education
outcomes. First, there is no clear relationship between expenditures and student perfor-
mance. Second, greater equality in spending was not accompanied by greater equality in
11Notwithstanding the argument, some inner cities have an important amount of capital which implies an
important tax base, as argued by Fischel [27].
23
measured student performance. In other words, the change towards a centralized regime did
not imply losses in e ciency, or at least, the evidence is not clear on the point. The previous
section of this chapter presented theoretical arguments why decentralization is more e cient
in the provision of services. By symmetry, it is then conceivable that the equalization of
expenditure would induce reduction in e ciency. The presented evidence does not support
this claim.
2.2.3 Simulation and general equilibrium models
Loeb [45] enriches the previous discussion with a political view of the problem. First, political
support for public education in a centralized regime declines because people may not be
consuming their preferred bundle. Second, rich districts do not have any incentive to support
a common pool of resources. In her model, changing from a decentralized system to a
centralized one will have two e ects. The  rst is the price e ect discussed above: if the tax
rate of the central government is higher than the one at the local level, individuals will be
paying a higher price for services. Second, since people are not consuming their preferred
bundles, individuals are worse o with a centralized system than with a decentralized one.
Therefore, in a vote for a centralized system versus a decentralized one, the latter will win.
Loeb uses data on education and income from Michigan to calculate the tax price by
locality. Based on these prices, and assuming some elasticities of income and price, she
simulates what would happen under di erent scenarios of expenditure  nancing. The result
of Loeb?s paper is that a system of state grants with unlimited local supplementation does
24
not provide equalization across districts, and average spending per student may go down.
By contrast, a system of state grants with limited supplementation leads to smaller variance
in spending, while providing some room for localities to be providing preferred bundles for
their communities.
Fern andez and Rogerson [25], using a general equilibrium model, present evidence that
is consistent with the evidence in EMS. The model is a two-period model in which agents
make three main decisions. Income of old people is determined by education obtained when
young. Given income, individuals decide where to live and then they decide public expen-
diture on education in that community. This expenditure determines the level of education
of the young people. A key characteristic of the model is the presence of capital market im-
perfections that constrain the expenditure on education more for poor individuals than for
rich ones. Two important parameters of the model are the elasticity of mean earnings with
respect to the quality of education and the cross elasticity of community public-education
expenditure with respect to community mean income. When the  rst elasticity is high, gains
from smoothing expenditures are outweighed by e ciency gains from local expenditure. The
main result of simulating the model is that the average income and education expenditure
as a fraction of consumption increases under central government control. The simulation
shows that local expenditures do not decrease.
In short, the empirical evidence is mixed, with two main conclusions. First, equalization
of educational expenditures decreases the variance of expenditures across localities. Second,
strong equalizations may induce a lowering of mean expenditure in the state.
25
2.3 Solving the equity-e ciency trade o 
An important aspect to address is, if the trade-o between e ciency and equality actually
holds in a (perfectly) decentralized system, how is it possible to induce a better distribution of
expenditures across localities without breaking the potential control that local communities
have over education decisions? This section is divided in two parts. The  rst part discusses
vouchers and the second state taxes and redistribution.
Vouchers are a system in which people have the option to consume services from other
localities, mainly from those that o er better-quality services than provided in their own
community. Grants, which are moneys distributed to localities from statewide taxes, can
also change the quality of services provided. Localities that are bound by budget restriction
to supply low-quality services, can with the aid of the grants supply better services. Localities
that are not bound by their budget constraint, will deliver the same level of services as before.
2.3.1 Vouchers
A potential solution to the equity-e ciency trade-o of decentralization is to subsidize ed-
ucational expenses via demand vouchers so that people can choose their preferred school,
regardless of the location of the household. One of the  rst essays to advocate this market-
oriented solution is Friedman [29]. He argues that vouchers give individuals the option to
pick schools according to their abilities, breaking the link between income and services in
the di erent localities, at least for those individuals that decide to change schools with the
voucher. On top of this e ect, vouchers increase the competition between schools, mainly
26
public ones, and therefore they also induce e ciency gains.12
The voucher solution has a potential problem with respect to the distribution of abilities.
Parents who actually exercise the option of the voucher likely will be either high-ability or
highly-motivated who may leave the public school system to attend private school or better
public ones. Under the presence of local externalities (like peer e ects), public schools that
are losing people will retain only lower-ability individuals, reducing the potential of positive
externalities in public schools.
This type of argument is examined in Epple and Romano [22]. Their model has two
important characteristics. First, students di er in abilities and income. Second, achievement
of students depends not only on their own ability, but also on the ability of their peers.
Again, as in the Benabou model, local externalities drive the model?s dynamics. The model
incorporates private and public schools, but it does not model the potential e ciency gains
from vouchers via competition among schools. The equilibrium is characterized by a strict
order of qualities across schools, with the public schools having the lowest quality of education
and quality in private schools ranked in ascending order by income (and perhaps by ability).
Vouchers pull out high-ability individuals from public schools, drawing down the mean ability
of people that stay in the public schools. High-income individuals subsidize low-income,
high-ability people. The overall welfare consequences of vouchers may be signi cant, since,
despite increases in low-income, high-ability individuals? gains, the majority of people in
public schools lose due to the introduction of vouchers. This implies important distributional
12However, this solution does not address the problem of voucher  nancing, and may imply another type
of ine ciency.
27
changes as well.
Manski [47] presents a review of recent reform proposals in the USA using a general
equilibrium model. One diagnosis of the problem is that \it is widely held that decision
making in (the public schools) is too concentrated in school district administrations and that
teachers and principals lack the incentives and authority to perform their jobs e ectively."
In other words, more decentralization is needed. Other reforms are \choice", in the sense
that vouchers correct incentives and give opportunities to people.
In a model with micro fundamentals, the author simulates the e ects of several programs,
among them vouchers. The benchmark scenario yields the following equilibrium character-
istics. First, expenditures are almost always higher in the private sector than in the public
one. Second, the fraction of highly-motivated students is higher in the private schools than in
the public ones. Third, within each community there exists a negative relationship between
public enrollment and income.
When a system of vouchers is introduced, tuition in the private schools falls as value
of vouchers increases. As a consequence of this relative price e ect, the fraction of people
in public schools decreases. Moreover, the fraction of highly-motivated people falls in both
types of schools, private and public since the people moving from public to private schools
have lower motivation than those already in private schools, but higher motivation than
those left in the public schools.
One recent study shows some evidence in favor of a voucher system. Paul Peterson [54]
takes advantage of three randomized programs in New York City, Dayton, Ohio and Wash-
28
ington, D.C. The experiment was similar in the three cities. The government announced
provision of a certain amount of vouchers for low-income people. All interested people could
apply, with the government using a lottery to pick the experimental group to whom the
vouchers were given. The control groups consist of the people who did not receive the
voucher. Since the vouchers were given randomly, presumably the treatment e ect can be
estimated consistently using di erence by treatment status. The most important empirical
result of the paper is that after two years of the program, African American students who
switched from public to private schools improved their standardized test scores more than
their contra-parts who remain in the public system.
An important limitation of the study is that, while a population treatment is estimable,
it is only the e ect for those who applied to the experiment. On top of this, the study has
some statistical problems. First, the program?s rate of attrition was very large (e.g., around
40% of people in the control and treatment group did not report results in the follow up
interview). Second, it seems that all the result are driven by a small group of students in a
speci c wave in New York that score higher than any other African-Americans in the sample.
2.3.2 State taxes and redistribution
Loeb [45] discusses another potential solution to the trade-o , where each locality attains
their preferred bundle for education with optimal local taxes and a state wide tax that is
distributed to localities progressively. This solution has the bene t that rich communities
consume their preferred levels of education and that poor communities will be able to increase
29
their expenditure on education. It is important to notice that this solution will not receive
political support from communities that will not enjoy transfers from the state. However, if
the choice is between a centralized system in which each community has to spend the same
amount in education and a decentralized system with a state tax distributed in a progressive
way, people may vote for the second option. Clearly, agenda-setting power is critical.
The combination of a state grant with vouchers can be quite complex. Nechyba [48],
Nechyba [49] and Nechyba [50] analyze not only a voucher system but also di erent  nancial
arrangements under decentralization within a general equilibrium framework. On one hand,
vouchers may increase e ciency (primarily in public schools), but on the other, this creates
an expected equity loss since highly motivate students will leave public schools for private
ones. The main aim of the papers is to see if, in fact, this tension exists. The main features
of Nechyba?s model are:
1. Parental perceptions of school output are a function of per-pupil expenditure and
average peer quality of the school
2. Quality enters directly into parents? utility function;
3. There are local educational externalities (i.e. peer e ects)
4. Local public schools are funded by local property taxes, chosen by majority rule vote
among local residents
5. There are block grants per pupil based on state income tax
30
6. Private schools are pro t maximizing institutions that set a minimum standard for
peer quality
7. There is perfect correlation between peer quality and income
8. There is housing capitalization
In the model?s equilibrium, higher-income communities tend to have higher spending levels
for schools, lower property tax rates with high-value bases, and property values are increas-
ing with the community wealth. High-income individuals will use vouchers to attend private
schools in low-income communities to avoid high property taxes in the high-income commu-
nities.13 This migration has two main e ects. On one hand, spending per pupil increases in
public schools because fewer students attend the public sector. On the other hand, vouchers
induce a lower peer quality in public schools in low-income areas. However, the expenditure
e ect may outweigh the second e ect (the drop in peer quality) in which case public schools
in poor areas are made better o . In this model, the introduction of a voucher system breaks
educational strati cation by income and wealth.
Now, if education is funded by states (the central government), expenditures per pupil
are equalized across communities, and peer e ects matter. Then, capitalization of education
quality into housing prices breaks down. People have to pay the same tax, regardless of
where they live. For this reason, the voucher system does not induce the migration e ect,
e.g., rich people moving to areas with lower taxes but good private schools. In this scenario,
13In this model everyone can have access to the vouchers. It is a universal program, in contrast to one
targeted to low-income people.
31
the equilibrium e ect of vouchers on housing prices is very limited.
Now consider the combination of local taxes with state grants. In this case, all measures
of school quality in the public sector deteriorate. The key driving force behind the result is
that the transfer via the grant is not binding the expenditure that localities may want to
implement. Because of this, the per student expenditures can remain the same as before the
grant, and high-ability individuals will migrate from poor, public schools to good, private
ones. In other worlds, the negative peer e ect is present and the expenditure per pupil in
the public school does not change.
Notice that in this model the principal gainers are the high-ability individuals who mi-
grate. In a mean tested program, however, the main gainers would be high-ability low-income
people who take advantage of the program, changing from public schools in poor localities
to private (or public) schools in high-income localities.
2.4 A conclusion
There are some reasons to think that decentralized schools might perform better than cen-
tralized ones. First, localities have better information than the central government. Second,
decentralization may induce a process of checks and balances on the local provider of the
service, increasing e ciency on the provision of services. Despite this, there are some rea-
sons to think that decentralization might imply e ciency losses on the ground of segregate
localities, local externalities, and economies of scale in the provision of certain services.
Two main conclusions can be draw from the empirical studies. On one hand, some studies
32
actually  nd that court orders induce equalization and increase expenditure in education for
poor districts (Evans, Murray and Schwab [23] and Evans, Murray and Schwab [24]),  ndings
that are supported within a General Equilibrium framework (Fern andez and Rogerson [25]).
On the other hand, other studies that look at particular cases (such as California)  nd
that equalization occurs towards the district with lowest expenditure; that rich districts
are induced to spend less; and that several localities will be receiving educational services
that are not equal to their preferred bundles of consumption (Silva and Sontelie [55] and
Hoxby [41]), evidence that is supported by the simulations of Loeb [45]. The households
that have not been able to consume their preferred levels of education are paying local taxes
that do not correspond to the services they would like to recieve (Fischel [26] and Campbell
and Fischel [16]).
In conclusion, the direction of impact of decentralization on test scores is thus an empirical
question. The main task of this dissertation is, therefore, to estimate an empirical model of
education to examine this impact.
33
Chapter 3
Decentralization in Colombia
3.1 Introduction
Like other several developing countries in South America, Colombia undertook decentral-
ization programs during the 1980?s and 1990?s.1 In July 1991, Colombia enacted a new
Constitution. This Constitution gave a new push to a process of decentralization that had
started some years earlier.2 In fact, the major decentralization reform in 1991 followed an
incipient one in 1986. The \quasi-reform" of 1986 created some space for local governments
(speci cally, municipalities) to create and promote their own programs. It did not, however,
give the necessary  nancial resources to implement such programs. The purpose of this
chapter is to describe the general characteristics of Colombia?s decentralization reforms and
their implications for education. This section follows heavily Ahmad and Baer [2]; Alesina,
1The majority of the processes are described in Ter-Minassian [58] and Bird and Vaillancourt [13]. For a
general description of other macroeconomic changes across South America, see Lora and Barrera [46].
2Bird [12] presents some of the early ideas and proposals for the case of Colombia.
34
Carrasquilla, and Echavarria [3]; and Borjas and Acosta [15]. A recent set of papers pre-
pared for an Income Mission for Colombia were very helpful, specially Acosta and Bird [1]
and Smart, Zapata and Chaparro [56].
The  rst part of this Chapter presents the most relevant laws and data on decentralization
in Colombia. The second part describes in more detail the implication of decentralization
on education and concentrates on the allocation of teachers and the administration of public
education. The third part presents the data used in the empirical strategy. The last part
sets out changes in the school characteristics before and after decentralization. It allows us
to have a  rst indicator of whether anything has actually changed as result of the reforms.
Two important results emerge in this Chapter. First, the new system gives an important
amount of broadly speci ed resources, such as education or health, to the local government
(e.g. departments and municipalities). Second, the local governments, with the new set
of rules, have the opportunity to raise new taxes. Relatively few local governments have
increased local taxation, either because of the new resources they are receiving or because
local governments can borrow against these resources.
Decentralization gave more functions and power in the provision of education to the
subnational levels of government. It also provided signi cant resources to the departmental
and municipal governments. An important share of this increase went to increased wages
for teachers. Decentralization of education, however, was only partial. According to Borjas
and Acosta [15], the power of the central union of teachers is still strong, and the union still
a ects many critical educational decision at all levels.
35
It is important to stress that Colombia?s decentralization did not give full power to local-
ities either in expenditure decisions or in tax collection. Signi cant increases in subnational
expenditure powers, however, did occur. Colombia?s decentralization thus provides an im-
portant opportunity to study the e ects of decentralization, if only a partial one.
Regarding the changes in inputs, all the statistics indicate that the relative teacher quality
improved in public schools after decentralization, which is in line with the expectation that
decentralization increases the quality of education.
3.2 Decentralization in Colombia
Colombia was been divided politically, even before decentralization, into three levels of power:
central government, departments, and municipalities. The main objective of decentraliza-
tion was to shift from a system in which all the main decisions were made by the central
government to one in which some important decisions are left to departments and munici-
palities. The reform of 1991 gave substantial expenditure responsibilities to municipalities
and departments and increased the amount of resources available for them from both cen-
tral revenues and from taxes collected at the local levels. However, general policies are still
determined by the central government.
For example, the majority of the transfers from the central government are earmarked,
mainly for education and health. Laws enacted in 1993 and 1994 complemented the Consti-
tution by describing the actual mechanisms and capacities of taxation available to di erent
levels of government. In some sense, the Colombian decentralization is similar to the recently
36
implemented welfare reform in the United States, in which the States have more capacity
to design and implement programs but with speci c designation of certain funds from the
Federal government (for instance, see Baicker [5]).
In addition, the new Constitution dictates some rules on revenue sharing, and these were
implemented by Law 60 of 1993 (see Table 3.1). There are four main systems of revenue
sharing: two funds with speci c destinations, a system of co- nancing funds, and a fund
from royalties. Transfers from the central government to the departments and municipalities
made through these four systems have increased from 2.4% of the GDP in 1990 to over 5.6%
by the end of the 90?s.
The  rst fund, the situado  scal, or SF, provides transfers to the departments for educa-
tion and health. Fifteen percent of the fund is distributed equally among the departments,
and the rest is distributed according to the population of each department. The law was
implemented in 1993 with a minimum obligation of 22.5% of the central government?s total
current revenues for the fund, increasing to 24.5% by 1996.
The participacion municipal, or PM, fund is for the municipalities, to be used mainly for
education, health and water programs. This fund was implemented in 1993 as well, and its
share of total revenues from the central government increased from 14% to 21% by 2000.
60% of the fund?s resources are to be distributed according to the number of poor people in
the municipality and 40% according to the population,  scal and administrative e ciency,
and progress in reducing poverty in the local region.
The system of co nancing funds, or CF, started in the eighties as a way to induce in-
37
vestment in social and infrastructure projects. After the new Constitution, there was an
explosion in the number of these funds that by 1996 accounted for almost 40% of all trans-
fers to the municipalities.
Finally, the central government imposed a tax on natural resources such as coal and oil,
with 60% of the revenues going to the departments and municipality in which the resources
are, (47.5% to department and 12.5% to the municipalities), 8% to the municipalities in
which there are ports that are used for exporting the resources, and 32% to the National
Royality Fund, or NRF. The NRF distributes its revenues to the rest of the departments
and municipalities. In principle, these resources have to be invested into priority projects,
including development plans made by the departmental and municipal governments.
On the tax front, the central government collects the majority of taxes, although there
is some space for the local government to impose and collect their own taxes. Major taxes,
such as the VAT, international trade, personal and corporate income taxes, are collected by
the national government. The departments tax alcoholic drinks, cigarettes and vehicles. The
municipalities? taxes include a property tax, a business tax, and a vehicle tax.
Local taxes (departments and municipalities) account for 20% of all taxes, while the
central government collects the other 80%. The major taxes at the central level are the
income taxes, VAT, and payroll taxes. At the department level, 65% of tax revenue comes
from the sale of liquor and beer, and the rest from other sources such as taxes on vehicles
and tobacco. In total, the Industry and Commerce Tax and the Property Tax represent 84%
of municipal taxes (Alesina, Carrasquilla, and Echavarria [3]).
38
Some local governments took the opportunities generated by decentralization to extend
their tax base and impose new taxes. The technical capacity of departments and munici-
palities in collecting and administrating taxes, however, ranges from very low to quite high,
depending on the region. Major cities have modern systems of collection, tracking and pun-
ishment for evasion. By contrast, small and poorer municipalities generally do not have high
technical capacity.
Table 3.3 presents overall data on transfers, taxes and expenditures, before and after
decentralization. The purpose of this table is to address an important question: Has there
been any impact from decentralization in the country? As described above, local governments
began to have, at least on paper, more expenditure functions, and the central government,
again on paper, was obligated to transfer more resources to the localities. It could be the case
that the program was only a \list of intentions", without any actual changes in the system
in reality. The data show, however, a clear and important e ect from decentralization on
transfers, taxes, and expenditures at local level.
The amount of transfers between 1990 and 1999 increased rapidly, from 2.4% of GDP in
1990 to 5.5% in 1999. With respect to taxes, the changes have been di erent at the municipal
and departmental levels. At the municipal level taxes increased from 0.8% of GDP in 1990
to 1.8% in 1999, whereas departmental taxes stayed at the same level, between 0.8% and 1%
during the same period. Finally, expenditures at both levels show an important increase: at
the municipal level, from 2.55% of GDP in 1990 to 6.82% in 1999, and at the departmental
level, from 1.64% in 1990 to 4.2% in 1999.
39
Overall, the data show a clear fact: transfers to local governments have increased in a
signi cant way, taxes at the municipal level have doubled since the beginning of decentral-
ization and expenditures by the local government have experienced an important increment
as well. The data are evidence that decentralization has occurred in the country.
3.2.1 Deepening of decentralization and community role
Regardless of the important amount of resources transferred to the municipalities and de-
partments from the central government, the question of how decentralization has altered the
decisions at the community level has not yet been answered. Has decentralization reached
the communities?
Before decentralization, the President used to appoint the mayors and governors nation-
wide. In 1986, as one of the  rst measures of decentralization, the popular election of mayors
was established. The Constitution of 1991 increased community participation in the election
of municipal mayors and departmental governors, for example, by allowing for the possibility
of recall.3 These instruments, along with the increased amount of resources commanded by
each lower governmental level, have given the community one of the most important mech-
anisms for participation. Currently, the community can interact with the local government
in several ways. For instance, there are community meetings in which the community can
discuss with the authorities topics ranging from the quality of schools and public services to
3El Tiempo, the main newspaper in the country, has in its archives  ve cases of recall of mayors. These
recalls may induce a deterrence e ect in that other mayors are more aware of the consequences of not ful lling
their campaign promises. In this case, decentralization may yield additional bene ts due to the more active
community participation.
40
problems of corruption in the government.
In terms of the power of the municipalities, even though the majority of the transfers
are earmarked for education and health (approximately 80% of the transfers), municipalities
still have some decision power over them. For instance, the law does not specify in which
speci c items the municipality has to spend the money for education.4 Moreover, around
20% of the transfers are at the free disposal of the municipality, giving them more decision
power.
Finally, as Acosta and Bird [1] argue, the departments loses with the implementation of
decentralization. Before decentralization, departments used to have several important roles,
especially with regard to the realization of large infrastructure projects and the provision
of security for the country. Now, the two important decision makers are the Nation, which
took some of the functions of the departments, and the municipality, which is the unit of
expenditure for public services.
In sum, there is evidence that decentralization has implied a more active role for the
community in the decision making process, not only in terms of electing mayors but also in
the discussion of policies. Municipalities now, along with the National government, determine
the policies of the country.
4Nonetheless, an important amount is designated for the operation expenditures, a designation that is
quite in exible.
41
3.3 Education
The two key laws concerning decentralization and education are Law 60 of 1993 and Law 115
of 1994. These two laws have di erent objectives. The law of 1993 gave more responsibilities
to the municipalities, whereas the law of 1994 assigned a greater role to the departments.
These two Laws are con icting with each other to some degree in the sense that Law 60 gave
more power to municipalities, while Law 115 took away some of this power and gave it to
the departments.
3.3.1 Functions across jurisdictions
As Borjas and Acosta [15] points out, the distribution of functions across the three levels of
government (municipal, departmental and national) is not well de ned. Each of the three
levels of government has a role in the administration of public education, as established by
Law 60 and Law 115 (see Table 3.2 reproduced from Borjas and Acosta). In the central
government, the Ministry of Education is in charge of educational issues. In theory the
central government is in charge of technical, curricular and pedagogical norms to be used
by the other levels. However, it also makes some decisions about the number of \national
teachers" and their wages, their pensions, and some infrastructure investments.
The Departmental Secretaries of Education are the agencies in charge of the educational
area at the departmental level. They plan, administer and coordinate teaching services, and
they decentralize these functions to the municipalities. The Secretaries prepare teachers,
administer the co- nancing funds, and invest in school infrastructure.
42
The municipalities administer the three levels of education, which are kindergarten, pri-
mary and secondary. They also maintain infrastructure and investment in schools, and they
inspect and supervise the provision of education.
In sum, there exist several duplications in the responsibilities among the di erent levels
of government. For instance, investment in infrastructure is decided by the three levels of
government, without a clear assignation of responsibilities of each one. Borjas and Acosta [15]
point out other problems in the system as well, such as poor mechanisms for monitoring and
the lack of appropriate data to evaluate performance of the local schools.
3.3.2 Teachers and wage allocation
The process of teacher allocation has experienced several changes during the last decade.
Before decentralization, the teachers? union (\FECODE") used to have an important say
in the location of teachers across the national territory. For instance, they in uenced the
decisions on the number of teachers in a municipality. After decentralization, the decisions
process is in uenced both by the power of the union and by that of the municipalities. First,
there are \national teachers", who are allocated to departments and municipalities according
to negotiations between FECODE and the Ministry. Second, despite this, departments and
municipalities can also hire teachers, but the new teachers cannot replace the ones allocated
by the negotiations between FECODE and the Ministry. In short, the union still has some
power in the hiring and allocation of teachers. However, the localities have started to have
some level of autonomy in the process since they can hire additional teachers besides the
43
ones allocated by central decisions.
In terms of teacher wages, Law 115 gave the teachers? union an important role in the
determination of wages. This power was counter-balanced by Law 60, which gave munic-
ipalities an important role in determining wages. At present, the bulk of teachers? wages
is determined in a negotiation between FECODE and the Ministry of Education, and it is
implicitly  xed according to the level of the teacher. As it is presented below, teachers have
can be classi ed in 17 di erent ranks, where each rank is determined by experience and
education. Wages increases with rank. The table of wage by rank is the one determined by
negotiation between the union and the Ministry of Education.
A potential reason for Law 115 of 1994 was pressure from the teachers? union (\FE-
CODE") to try to centralize education decisions as much as possible (Borjas and Acosta [15]
and Duarte [20]). The teachers? union has opposed decentralization from the beginning,
arguing that decentralization implies fewer resources for education. However, the evidence
clearly suggests the opposite. As mentioned above, the bulk of the transfers that decentral-
ization mandates are allocated to education and health. Public expenditure on education
grew from 3.1% of the GDP in 1991 to over 4.5% at the end of the 1990?s. An important
component of this increase has been due to increases in teachers? wages (real wages for public
teachers grew at a rate of 3%) (Borjas and Acosta [15]). Clearly, another hypothesis of why
the union opposes decentralization is that the process itself implies a loss of union power.
At the beginning of 2002, the Congress of Colombia approved Law 120, which reforms
the Laws 60 and 115. This law gives more functions and clear control to the municipalities
44
in the provision of education, especially in the hiring and wage determination of teachers. In
several articles in El Tiempo [60], the most important newspaper of the country, FECODE
has strongly opposed decentralization, arguing again that decentralization implies fewer re-
sources for education and that decentralization implies more control for municipalities in the
provision of education. Evidently, the union sees increased local control as a bad implication
of decentralization. In addition, the union claims that with the passing of power from central
to local government education will lose  nancial resources.
3.3.3 Studies of decentralization and education in Colombia
The majority of studies on decentralization in Colombia address the relationship between
local  nances and the  scal de cit.5 According to these studies, two features in the design
of the decentralization contributed to the sharp deterioration of Colombia?s  scal conditions
during the 1990?s. First, tasks were not clearly assigned to the di erent levels of government.
As presented above, education is a clear example. Second, the debt rules facing the local
governments were extremely  exible. This induced over-spending by the local governments,
 nanced by unsustainable loans, and leading in several cases to bailouts by the central
government.
With respect to education and decentralization, Borjas and Acosta [15], use data from
National Household Surveys to provide some trends on education before and after decen-
tralization. They analyze four main variables by comparing the pre-decentralization period
5The most relevant ones are Alesina, Carrasquilla, and Echavarria [3]; Echavarria, Renteria and
Steiner [21] and Alonso, Olivera and Fainboim [4]
45
(1990-94) to the post-decentralization period (1995-97). They  nd a signi cant increase in
wages for teachers versus other workers, especially for public school teachers. Second, Bor-
jas and Acosta  nd an increase in enrollment rates, mainly among younger children (5 to
6 years old), precisely the population that is most likely a ected by the decentralization.
However, it is di cult to disentangle the e ect of decentralization from an overall trend
that started in the 80s according to their data. Third, the authors observe a low migration
rate among teachers. Teachers migrated less frequently, with this di erence shrinking after
decentralization. According to these authors, decentralization should a ect relative teachers
salaries across the country. This e ect should induce migration until salaries are equalized.
Fourth, they show some convergence in some educational variables (for instance, enrollment)
across some cities, but again it is di cult to disentangle the e ect of decentralization from
an overall trend.
In conclusion, these authors  nd mixed evidence concerning decentralization and educa-
tional variables. A clear impact is found in the increase in teachers? wages. The improvement
on the others variables that Borjas and Acosta study may be explained by a prior tendency
and not necessarily by the decentralization.
The data that Borjas and Acosta use are aggregate data that come from household
surveys. The data are derived from the questions on the occupation of the person, and using
other data in the survey they derive income and migration patterns for di erent occupations.
The contribution of this thesis is to investigate the impact of decentralization on education
using another source of information that is richer than the data used previously since it
46
directly measure educational outcomes.
3.4 Educational data
The data come from two sources. First, the C600 and C100 is a general survey administered
to schools. It provides data on the characteristics of the schools inputs. Second, data from
the national institute for higher education (in Spanish, the ICFES, \Instituto Colombiano
para el Fomento de la Educacion Superior") provide standardized test scores and data on
characteristics of individuals, beside school identi ers. This section will describe in a high
degree of detail the construction of the  nal data set.
3.4.1 The survey of schools
The C600 and the C100 surveys provide general data on schools. The Ministry of Education
collects the data directly via a questionnaire to schools. The purpose of this survey is to
have educational data for evaluation purposes. However, the data have not been used in a
systematic way and until now have only been serving as the source of speci c statistics. Data
are available for the years 1990-1994 and 1996-1999. Data for 1995 are not available, due
to technical problems in the compilation at the Ministry of Education. Two main computer
formats of the data are available; presentation of the content of the data will follow the two
types of format.
47
1990-1994
For the period 1990-1994, the data are divided into four levels of schools, kindergarten,
primary, new primary, and secondary. Every level has four main sections.
New primary is a type of primary, except that a teacher has students for several grades
in the same classroom. The primary objective of this arrangement is to cope with small
cohorts in rural areas, where there are not many teachers.
Kindergarten generally has three grades and primary (New and regular)  ve. Every entry
in all sections of the data indicates the department and the municipality in which the school
operates, the type of zone (rural or urban), the sector of the school (private or public), the
schedule of the school (complete, morning, afternoon and night) and a seven-digit code that
identi es the school (the DANE code.) Every schedule within the school is treated as a
school itself (e.g., one school that provides primary in the morning and secondary in the
afternoon is like two schools)
The  rst section in the kindergarten level has data about number of registered children
in the school by sex, number of dropouts by sex, and number of children who pass to the
next grade by sex. All of these variables are by grade within the level, e.g., kindergarten,
grade 1.6 From these variables, three main variables are constructed: the total number of
registered children (which is the sum of registered students by grade and by sex), the total
number of dropouts (the sum of dropout by grade and by sex) and the total number of
children who pass (the sum of passing children by grade and by sex).
6From now on, level refers to kindergarten, primary (regular and new) and secondary.
48
The second section in the kindergarten level has the number of students by grade, age
and sex. For instance, it records female children age four in grade 2. Also, it presents two
variables that are the number of requested places by grade and number of actual available
places by grade. From this section, two key variables are extracted: the total number of male
students (the sum of male students by grade), and the total number of female students. Also,
each record indicates the department, the municipality, the zone, the sector, the schedule
and the DANE code of the school.
The third section presents information about teachers. Colombia has a system of teachers?
seniority that depends on experience, education and speci c courses that the teachers can
take to ascend in the ranking. There are 17 ranks: without rank, rank A, rank B and ranks
1 through rank 14. This section contains data on the number of teachers, by sex and by
ranking. From these data, seven variables are constructed: total number of teachers without
ranking (the sum of female and male teachers without rank), total number of teachers with
rank A and B (the sum of female and male teachers with ranks A or B), total number of
teachers with rank 1 through rank 14 (analogously), the total number of female teachers,
the total number of male teachers and the total number of teachers (sum of female and male
teachers). Again, each entry indicates the department, the municipality, the zone, the sector,
the schedule and the DANE code of the school. Rank gives a proxy for experience of the
teacher, given that higher rank means higher experience.
Finally, the last section at the kindergarten level contains information about education
of teachers. The data distinguish education of teachers by whether the teacher  nished each
49
of four levels; primary, secondary, college, and graduate studies. All variables are available
by sex. From these variables,  ve other are constructed: total number of teachers with only
primary, total number of teachers with secondary studies, total number of teachers with
college studies, total number of teachers with graduate studies and  nally total number of
teachers. These four sections were merged, using the DANE code and the schedule of the
school, in a single Stata dataset.
The  rst section for primary records the total number of students admitted, not passed
and dropouts, each broken down by sex. The second section includes number of students
by sex, grade and age. From these variables, three variables were constructed: total number
of male students, total number of female students and total number of students. The third
and fourth sections give teachers by rank and level of education in an identical way as to the
respective sections for kindergarten. These four sections of primary education were merged,
again using DANE code and schedule, to create a Stata  le for each year.
For new primary, the data are also divided into four sections that are identical to the
sections in the kindergarten and primary categories. From these data, variables were analo-
gously constructed and compiled into a single Stata dataset.
There are two types of secondary education in Colombia: \short" and regular secondary.
In general, short secondary programs are technical schools in which the students prepare for
speci c careers, such as electrician or farmer. These programs can last for four or  ve years.
The duration of regular secondary is six years. The regular secondary school programs have
general courses like mathematics, physics, biology, etc. The aim of this type of secondary
50
school is to prepare individuals for college programs.
As with kindergarten and primary, the  rst section of the data provides information on
the total number of students, the total number of dropouts and the total number of students
who did not pass to the next grade in the short type of secondary. These variables are
available by sex. It is also possible to know the type of short secondary program in which
the students are enrolled: industrial, farming, social promotion, business, pedagogic, arts,
science, and others.
The second section presents data for the regular secondary schools (six grades). It con-
tains information on the total number of registered students, the total number of dropouts
and the total number of students who did not pass to the next grade, distinguished by sex.
The third section presents information on the total number of students by age. Sections three
and four are exactly the same as the respective sections for the other three levels. They give
information about education and rank of teachers. Finally, another section provides number
of teachers by area of teaching and sex. The  fteen areas are mathematics, Spanish, sciences,
religion, aesthetics, physical training, language, philosophy, industrial class, commerce class,
farming, social promotion, systems and others. As in the other three levels, all the sections
in the secondary data were compile in a Stata  le.
Finally, for each year, the school level data were merged into a single data set including
all variables by school and by schedule.
51
1996-1999
In contrast to the 1990-1994 data, the 1996-1999 data are presented in sections by variable.
For example, one archive contains the data for the number of students by school, with an
identi er of the level and the schedule of the school. It is important to stress that the
separation by levels (e.g. kindergarten, primary, secondary) is an important feature of the
data work since it allows the identi cation of changes of inputs by level of school. As in the
previous section, the description of the data will follow the original format in which the data
were given. The following sections are identical for the years 1996-1999.
In the  rst section of the data, each school has a school DANE code (these codes do not
change over the 1990-1999 period), and the schedule of the school. This section of the data
is important because it allows the merger of the 1990-1994 data with the 1996-1999 data
using the DANE code and schedules.
The second section of the data presents, by school-level, data on the department, munic-
ipality, zone (rural or urban), name of the school, address, phone number, calendar (whether
the school runs from February-November (calendar A) or September-June (calendar B)),
sector (private or public), type of provision of energy, and information on the water and
sewer system of the school (no provision, private provision or public provision.) Each entry
has a DANE code.
The third section presents data on administrative personnel and teachers. The data are
presented by school (DANE code) and by schedule. Seven main variables can be extracted
from this section; the number of headmasters that are teachers, the number of teachers,
52
the number of administrative personnel, the number of doctors and dentist, the number of
advisory personnel (including psychologists) and the number of people dedicated to special
education. From these variables the total number of teachers is constructed (the sum of
headmasters and teachers).
The next section of the data presents information on the total number of students in the
school and the number of students who did not pass to the next grade. These two  gures
are discriminated by sex, grade and schedule. Since each entry in the dataset is by school,
by schedule and by grade, in order to get the data by level, the data are collapsed by grade,
schedule, and school.
The information in the  fth section of the data refers to level of education of teachers.
Each line of data has the code of the school, the schedule, the level (kindergarten, primary
or secondary), and the number of teachers with complete or incomplete education in each of
primary, secondary, college, and graduate studies. The same type of data as described for
1990-1994 were extracted from this section.
The next section gives secondary teachers? data by area of education. This section pro-
vides the number of teachers by area of teaching and sex for the secondary level. The  fteen
areas in which the data are presented are mathematics, Spanish, sciences, religion, aesthet-
ics, physical training, language, philosophy, industrial class, commerce class, farming, social
promotion, systems and others. These data are identi ed by school code and schedule.
Finally, the last section of the data provides information about physical characteristics
of schools. The main variables are: furniture in the school (chairs and desks); support
53
materials such as computers and so; number of computers for teaching purposes exclusively;
number of laboratories; laboratories for physics, chemistry and biology; laboratories for
speci c education (construction, farming, etc); number of rooms; number of classrooms;
number of libraries; number of food facilities and dorms; number of sports locations (soccer
and basketball  elds); and number of other rooms.
All the above sections for the years 1996-99 were merged using the DANE code and the
school?s schedule to create a dataset for each year.
Common data and main variables of schools
As we have described, there are several variables that are available for both time periods.
Some variables are available, however, in only one of the periods. In particular, the number
of dropouts is available only for the 1990-94 period, while the physical characteristics of the
schools and the administrative personnel are available only for the 1996-99 period. The rest
of the variables are available for both periods. Given that the data are presented by DANE
code and schedule for both periods 1990-94 and 1996-99, it is possible to have a sub-group
of schools common to the period 1990-99. The number of schools that are observed over the
entire period is 10.481.
Altogether, 10 variables were used in the estimation, following the empirical literature
(Hanushek [33], Hanushek [32]). First, two variables are included to capture students char-
acteristics of the school: the total number of students per school and the fraction of students
that failed to pass the academic year, separated by kindergarten, primary and secondary.
54
Second, three sets of variables that capture teacher characteristics are included: number of
teachers per schools; primary teachers fraction (teachers with only primary/total number
of teachers), secondary teachers fraction (teachers with secondary/total number of teach-
ers), college teacher fraction (teachers with college/total number of teachers), and graduate
teachers fraction (teachers with graduate studies/total number of teachers); average years
of teachers schooling;  nally, math, language, and science teachers fraction (math, language
and science teachers/total number of teachers). Finally, the last variable is the student-
teacher ratio (total number of students/total number of teachers). All of these variables are
calculated within level of education.
3.4.2 Individual test and characteristics
The data on test scores and student characteristics are from the ICFES. The test score
data come from a general test that is administered to all students who are  nishing their
secondary education. Again, following the format of the data, the description will follow the
 les as they were given. It is important to note two points with respect to these data. First,
the data are available for the years 1990, 1991, 1992, 1995 and 1999. Therefore, three years
before decentralization are available to be merged with the data of the schools (1990, 1991
and 1992), and only one year in the post decentralization regime (1999) since school data
for 1995 are not available (though it is possible to merge the 1995 individual data and the
1996 school data). Second, detailed information on the characteristics of individuals is only
available for 1999. However, some individual characteristics are available for the other years.
55
The data for August and March 1989, March 1990, August 1991, March 1992 and August
and March 1995 include several variables: a 7-digit individual identi er; gender; a variable
indicating whether the person is currently enrolled; a school code7; the city where the test
was taken (which is a combination between the Department code and the Municipality code);
a variable indicating whether the person is married or not; the university that the individual
wants to attend; the schedule of the school that the individual attended; the calendar of the
school (A, which is from February to November, or B, which goes from September to June);
the type of secondary school that the individual attended (mainly, academic or technical);
and whether the person has migrated from the original town of birth.
The data for the exam scores are in a separate  le. The data have the individual identi er
(a seven-digit number) and total score in the test and score by areas of the exam (mainly,
mathematics, physics, chemistry, biology, language, and others).
The data for August and March 1999 are divided by date of exam administration. Typ-
ically, students from schools with calendar A (February-November) take the August exam,
and schools with calendar B (September-June) take the March exam. The data provide the
seven-digit individual identi er; a variable indicating whether the students is \validating" or
not (usually, validaters are people who dropped out and later decided to take some classes
to  nish school); gender; whether the person is studying at the moment or not; a National
Test Service school code; the city where the test was taken (which is a combination of the
7This code is not the DANE code, but another code of the National Test Service. However, in another
section of the data this code is matched with the DANE code, allowing the merging between the individual
and school data sources.
56
Department code and the Municipality code); the address and telephone of the person; the
day, month and year of birth; the age; year in which the person  nished or is  nishing high
school; the schedule of the school that the individual attended; the calendar of the school (A
or B); the type of secondary school that the individual attended (regular or short); plans for
next year (work, study or both); number of people in the family; whether the family owns a
house; whether the family has a mortgage; number of people contributing to the income of
the family; range of family income (there are 10 categories given as a ration to the minimum
wage); occupation of father; occupation of mother; number of siblings; number of siblings
with high school; position of individual among siblings; whether the person is married; and
the university that the individual want to attend. The data on exam scores are in a separate
 le. These data have the individual identi er, the total score, and the score by area of the
exam.
Finally, the ICFES data have some information on the school characteristics. The school
data include both school codes - the National Test Service and the DANE code. They allow
the merging of data on individuals and data on schools. Besides these two codes, the data
also include the schedule of the school, the sector of the school (private or public schools),
department of the schools, municipality of the school, calendar of the school, whether the
school is bilingual, and the population attending the school attend by gender.
57
3.5 Changes in characteristics of school inputs before
and after decentralization: Inside the Black Box
For the whole study period (1990-1999) around 40% of the schools are located in urban areas.
However, location varies substantially across public and private schools: 70% of all public
schools are located in rural areas, whereas only 5% of private schools are in rural areas. In
terms of school schedules (morning, afternoon or both), the distribution of schools is the
same in both the public and private groups.
As described in the data section, the focus of this part will be on three sets of school-
level variables. The  rst set describes the student composition of a school, the second set
describes the teacher characteristics at the schools, and the third set compares the number
of students to teachers. Table 3.4 to Table 3.9 present these school level characteristics. In
the tables, the observations are divided into three school levels; kindergarten, primary, and
secondary.
The tables provide averages before and after decentralization, as well as di erences in
the averages. There are three years that are important to consider when constructing the
before-after decentralization periods. The new constitution was enacted in July of 1991,
giving major resources to departments and municipalities from 1992 onwards. The 1993 and
1994 laws delimited the scope of decentralization and organized tasks across the three levels
of government.
Three di erent calculations were performed to derive the pre- and post-decentralization
58
period variables. The simplest one was to take 1991 as the pre-decentralization period
and 1999 as post-decentralization. Second, the average of values in 1991 and 1992 as pre-
decentralization and the average of the values for 1996 and 1999 as post-decentralization were
used. Third, the average of values in 1991 and 1994 was used for the pre-decentralization
period and the same average (1996 and 1999) as in the second method was used for the post-
decentralization period. Sensitivity analyses were carried out to see if the results change
with these three de nitions of the pre- and post-decentralization period. The results are
robust to the years chosen for the calculations.
Several interesting and some surprising results can be seen in Table 3.4 regarding changes
in kindergartens. The top pane of the Table pools public and private schools. The number
of students per school stayed constant during the period of consideration: 55.32 students in
the pre-decentralization period versus 55.09 students in the post-decentralization period. In
terms of the percent of students who did not advance to the next academic year (\non-pass"
students), the average number decreases from 0.17% before decentralization to zero percent
after decentralization.
In general, the proportion of teachers with secondary and graduate studies increased,
whereas the proportion with college education decreased. On average, the number of years of
education of teachers remains stable at around 14.5-15 years. Presumably, a decentralization
process that has a positive impact on education will induce teachers to be more highly
educated. It is, therefore, expected that the proportion of teachers with just primary (or
secondary) school education will decrease and that the other categories will increase. This
59
expectation is not fully borne out, however. The proportion of teachers with secondary
schooling increased slightly (from 0.269 to 0.280), whereas the proportion of teachers with
primary stayed constant (around 0.004); the proportion of teachers with a college decreased
from 0.711 to 0.572, while the proportion of teachers with graduate studies increased from
0.016 to 0.143. Finally, when both types of schools are considered, the teacher-student ratio
increases from 0.054 teachers per student to 0.061.
When the kindergarten sample is divided by sector, several interesting  ndings emerge.
First, on average public schools have fewer students in kindergarten, with the number of
students enrolled in public kindergarten increasing with decentralization (from 40.023 to
51.610); by contrast, the average number of students in private kindergarten decreases over
the decentralization period (67.070 versus 57.756). The proportion of failing students in
both public and private schools decrease dramatically from 0.138 to 0.005 and from 0.186 to
0.004 students, respectively.
Before decentralization, the distribution of teachers? educational attainment was very
similar across sectors. However, college-educated teachers in public schools dropped after
decentralization from 0.693 to 0.503, with the proportion of teachers having done graduate
studies increased from 0.027 to 0.304. By contrast, in private schools the proportion of
teachers with only a secondary degree increased with the number of teachers having a college
education dropping. The average years of education increased in public schools from 14.66
to 15.916 and decreased in privates ones from 14.654 to 14.263.
In terms of teacher-student ratio, another asymmetry between public and private schools
60
is evident: the ratio, which was lower for public schools in the pre-decentralization period
(0.043 in public schools versus 0.062 for private schools) remained constant for public schools
and increased for private ones. Di erences across the decentralization period varied signi -
cantly across sector for 7 out of 8 variables describing kindergartens (Table 3.5). In summary,
the number of students increased more during the pre- to post-decentralization period in the
public schools. The proportion of teachers with graduate studies increased more in the pub-
lic schools than in the private ones, and the average years of education per teacher also
increased more in the public schools. Finally, the student-to-teacher ratio decreased more
for the public schools than for the private ones.
Table 3.6 shows the same set of variables but for primary-level schools. Pooling the pubic
and private schools, the number of primary students decreased from an average of 123 to
117 students per school. The ratio of non-passing students also declined during the pre-post
decentralization period, from 0.205 to 0.018.
The proportion of teachers with secondary education decreased, and the proportion of
teachers with graduate studies increased. Teachers? educational attainment increased by one
year. Finally, during the study period there was an increment in the ratio of teachers to
students, from 0.046 to 0.053.
Looking at primary education by type of school, we see that private schools have more
primary students. However, for both types of schools the number of students per school fell,
with the decline being more pronounced in the private schools (from 180 to 161) than in
public schools (from 112 to 108). The ratio of non-passing students fell signi cantly in both
61
types of schools.
The table shows an increase in teachers? education: the proportion of teachers with
college education falls, while the proportion with a graduate education increases. This trend
is especially pronounced in the public schools. The average year of education increased for
public schools, while decreasing for private ones. Lastly, there is a small increment in the
teacher- student ratio for both types of schools.
In terms of the di erence in di erence (DD) estimates without covariates, 7 of the 8
are statistically signi cantly di erent from zero. In summary, teachers? education levels
increased more for public schools. Class size as measured by the teacher-to-student ratio
rose in both sectors, with greater increments in private schools (Table 3.7).
Table 3.8 reports summary statistics for the same set of variables at the secondary-school
level. Pooling the public and private schools, the number of secondary students increased
across the two periods, from 439 to 478 students per school. The ratio of non-passing students
fell signi cantly, from 0.26 to 0.03. The proportion of teachers with graduate studies and the
average years of teachers? education increased. Lastly, the teacher- student ratio decreased,
slightly, and the proportion of teachers who were physics, chemistry, biology and languages
teachers remained almost the same.
Comparing across sectors, we  nd results similar to those for kindergarten and primary
schools. The number of students in the public schools increased, while the number in private
schools decreased. In both periods, the number of secondary students attending public
schools was greater than the corresponding number in private schools.
62
With respect to teachers? education, the same pattern as in the previous two levels arises:
the proportion of teachers with secondary and college levels of education falls, while the
number of teachers with graduate studies rises and years of teachers? education increased
more in the public sector. In fact, average years of teachers? education were greater in
private schools before decentralization (15.10 versus 15.36), but greater in public schools
after decentralization (16.69 versus 15.95).
The ratio of teachers-to-students decreased for both public schools (from 0.0642 to 0.0529)
and private ones (from 0.0845 to 0.0715). The ratio of \science" teachers (physics, biology,
chemistry and language), which is higher in the public schools, fell for both types of schools
over the two periods.
In conclusion, 8 out of the 9 variables showed a statistically signi cant change for sec-
ondary schools (Table 3.9). In all 8 of these cases, the changes occurred in a direction
suggesting that decentralization improved the relative quality of education in public schools.
The number of students increased more in the public schools; the ratio of non-passing stu-
dents decreased more in public schools; the number of teachers increased more in the public
schools; the proportion of teachers with graduate studies increased more in public schools;
the number of years of education of teachers increased in public schools; and the proportion
of science teachers increased in public schools (relative to private schools).
It is worth noting that, for all levels of education, teachers education in the public sector
preserves  rst-order stochastic dominance:  Pr[ teachers education= E] < 0 for all E <
Emax. In other words, everyone prefers the dominant, after decentralization distribution
63
over the dominated, before decentralization distribution. This property holds for the public
schools, whereas it does not apply for the private ones. This is an important property
that supports the argument of an improvement in the quality of public teachers after the
decentralization.
Across all three levels of education, three principal characteristics emerge concerning
teachers. First, the number of teachers increased more in public than in private schools.
Second, the proportion of teachers with graduate studies increased, with the fraction having
secondary- and college-education falling. Third, the average number of years of education of
teachers increased more in the public sector than in the private one. All in all, a reasonable
case can be made that relative teacher quality improved in public schools.
64
Table 3.1: Main Laws in the Nineties
Year Law Changes
1991 New Constitution Decentralization is implemented:
increment in expenditure functions by local government
1993 Law 60 Development of decentralization;
organization of transfers
1994 Law 115 Empowerment of departments in the education decisions;
more resources to education, specially wages of teachers
Table 3.2: Functions by level of government
Level Activity
National General norms to be implemented by other levels:
curriculum, pedagogic and technical orientation
Departments Administration, coordination and preparation of teachers
Administration of co- nancing funds
Schools infrastructure, investment and maintenance
Municipalities Administration of kindergarten, primary and secondary education
Schools infrastructure, investment and maintenance
Inspection of the provision of education
Adapted from Borjas and Acosta.
65
Table 3.3: Decentralization in numbers
(% of GDP) 1990 1994 1997 1999
Tranfers to localities 2.4 3.5 4.3 5.6
Taxes
Municipalities 0.8 1.2 1.5 1.8
Departments 1.0 0.8 0.9 0.9
National 8.8 10.2 11.1 10.9
Expenditure
Municipalities 2.55 3.85 5.9 6.82
Departments 1.64 1.98 3.82 4.2
Source: Income Mission, several papers
66
Table 3.4: Kindergarten
Pool sample
Number of Observations: 1830
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 55.320 1.318 55.086 1.263
Number of no-pass students 0.165 0.004 0.004 0.001
Prop. of teacher with primary 0.004 0.001 0.004 0.001
Prop. of teacher with secondary 0.269 0.009 0.280 0.009
Prop. of teacher with college 0.711 0.009 0.572 0.010
Prop. of teacher with graduate std. 0.016 0.003 0.143 0.007
Avg. years of education of teachers 14.656 0.049 14.981 0.059
Ratio Student/Teacher 0.054 0.001 0.061 0.001
Public Sector
Number of observations: 795
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 40.023 1.156 51.610 1.441
Number of no-pass students 0.138 0.005 0.005 0.001
Prop. of teacher with primary 0.004 0.002 0.006 0.003
Prop. of teacher with secondary 0.276 0.015 0.187 0.013
Prop. of teacher with college 0.693 0.015 0.503 0.016
Prop. of teacher with graduate std. 0.027 0.005 0.304 0.015
Avg. years of education of teachers 14.660 0.080 15.916 0.093
Ratio Student/Teacher 0.043 0.001 0.044 0.001
67
Table 3.4: Kindergarten, cont.
Private Sector
Number of observations: 1035
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 67.070 2.082 57.756 1.937
Number of no-pass students 0.186 0.006 0.004 0.001
Pro. of teacher with primary 0.005 0.002 0.003 0.002
Prop. of teacher with secondary 0.263 0.012 0.352 0.013
Prop. of teacher with college 0.725 0.012 0.625 0.013
Prop. of teacher with graduate std. 0.007 0.002 0.020 0.003
Avg. years of education of teachers 14.654 0.062 14.263 0.067
Ratio Student/Teacher 0.062 0.002 0.074 0.002
Table 3.5: Kindergarten. DD estimator
Label Dif. in dif.
Mean St.Err t stat
Number of students 20.69 2.05 10.12
Number of no-pass students 0.048 0.008 5.85
Prop. of teacher with primary 0.003 0.004 0.84
Prop. of teacher with secondary -0.18 0.022 -8.15
Prop. of teacher with college -0.085 0.26 -3.36
Prop. of teacher with graduate std. 0.27 0.014 19.10
Avg. years of education of teachers 1.68 0.13 12.88
Ratio Student/Teacher -0.011 0.003 -4.17
68
Table 3.6: Primary
Pool sample
Number of Observations: 7786
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 123.1376 2.5137 116.8260 1.5420
Number of no-pass students 0.2054 0.0111 0.0180 0.0004
Prop. of teacher with primary 0.0060 0.0008 0.0097 0.0010
Prop. of teacher with secondary 0.6047 0.0044 0.4319 0.0046
Prop. of teacher with college 0.3803 0.0043 0.4086 0.0043
Prop. of teacher with graduate std. 0.0090 0.0006 0.1498 0.0030
Avg. years of education of teachers 12.9377 0.0228 14.1832 0.0298
Ratio Student/Teacher 0.0460 0.0003 0.0529 0.0005
Public Sector
Number of observations: 6570
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 112.5610 2.7961 108.6204 1.5939
Number of no-pass students 0.2367 0.0131 0.0203 0.0005
Prop. of teacher with primary 0.0057 0.0008 0.0082 0.0010
Prop. of teacher with secondary 0.6129 0.0048 0.4182 0.0051
Prop. of teacher with college 0.3725 0.0047 0.4024 0.0047
Prop. of teacher with graduate std. 0.0090 0.0007 0.1713 0.0035
Avg. years of education of teachers 12.9002 0.0251 14.3331 0.0329
Ratio Student/Teacher 0.0453 0.0003 0.0513 0.0004
69
Table 3.6: Primary, cont.
Private Sector
Number of observations: 1216
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 180.2821 5.2560 161.1604 4.6275
Number of no-pass students 0.0364 0.0080 0.0054 0.0004
Prop. of teacher with primary 0.0076 0.0021 0.0179 0.0035
Prop. of teacher with secondary 0.5606 0.0102 0.5063 0.0107
Prop. of teacher with college 0.4227 0.0100 0.4420 0.0104
Prop. of teacher with graduate std. 0.0090 0.0013 0.0338 0.0031
Avg. years of education of teachers 13.1401 0.0537 13.3731 0.0634
Ratio Student/Teacher 0.0496 0.0010 0.0618 0.0028
Table 3.7: Primary. DD estimator
Label Dif. in dif.
Mean St.Err t stat
Number of students 15.374 5.604 2.74
Number of no-pass students -0.195 0.031 -6.26
Prop. of teacher with primary -0.008 0.003 -2.25
Prop. of teacher with secondary -0.141 0.015 -9.55
Prop. of teacher with college 0.010 0.015 0.68
Prop. of teacher with graduate std. 0.138 0.008 16.83
Avg. years of education of teachers 1.203 0.086 13.94
Ratio Student/Teacher -0.006 0.001 -4.49
70
Table 3.8: Secondary
Pool sample
Number of Observations: 665
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 439.0842 12.8440 478.4045 12.5026
Number of no-pass students 0.1831 0.0102 0.0318 0.0019
Prop. of teacher with primary 0.0001 0.0001 0.0005 0.0002
Prop. of teacher with secondary 0.2946 0.0071 0.0944 0.0043
Prop. of teacher with college 0.7509 0.0076 0.5987 0.0079
Prop. of teacher with graduate std. 0.0544 0.0039 0.3064 0.0080
Avg. years of education of teachers 15.1897 0.0391 16.4420 0.0364
Ratio Student/Teacher 0.0710 0.0045 0.0591 0.0017
Propor. of science teachers 0.5299 0.0048 0.5095 0.0043
Public Sector
Number of observations: 443
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 460.8939 16.8139 531.0293 16.1686
Number of no-pass students 0.2060 0.0115 0.0366 0.0024
Prop. of teacher with primary 0.0001 0.0001 0.0006 0.0003
Prop. of teacher with secondary 0.2130 0.0091 0.0818 0.0045
Prop. of teacher with college 0.7299 0.0094 0.5698 0.0098
Prop. of teacher with graduate std. 0.0569 0.0044 0.3678 0.0100
Avg. years of education of teachers 15.1046 0.0496 16.6879 0.0419
Ratio Student/Teacher 0.0642 0.0026 0.0529 0.0011
Prop. of science teachers 0.5466 0.0064 0.5256 0.0054
71
Table 3.8: Secondary, cont.
Private Sector
Number of observations: 222
Label Pre-decentr. Post-decentr.
Mean Std.Err Mean Std.Err
Number of students 395.5631 18.5360 373.3919 16.9880
Number of no-pass students 0.1375 0.0199 0.0220 0.0029
Prop. of teacher with primary 0.0000 0.0000 0.0003 0.0003
Prop. of teacher with secondary 0.1578 0.0108 0.1194 0.0088
Prop. of teacher with college 0.7928 0.0123 0.6964 0.0109
Prop. of teacher with graduate std. 0.0495 0.0075 0.1839 0.0090
Avg. years of education of teachers 15.3594 0.0610 15.9512 0.0575
Ratio Student/Teacher 0.0845 0.0123 0.0715 0.0045
Prop. of science teachers 0.4966 0.0065 0.4773 0.0065
Table 3.9: Secondary. Di erence in di erence
Label Dif. in dif.
Mean St.Err t stat
Number of students 77.744 12.140 6.40
Number of no-pass students -0.041 0.019 -2.19
Proport. of teacher with primary 0.003 0.001 2.11
Proport. of teacher with secondary -0.115 0.013 -9.03
Proport. of teacher with college -0.076 0.016 -4.69
Proport. of teacher with graduate std. 0.170 0.015 11.10
Average years of education of teachers 1.114 0.078 14.32
Ratio Student/Teacher 0.005 0.008 0.60
Proportion of science teachers 0.007 0.007 1.08
72
Chapter 4
Empirical Strategy and Estimations
4.1 Introduction
The empirical literature on educational quality can be divided into two categories. The  rst
investigates the determinants of the quality of education (for a comprehensive survey, see
Hanushek [32] and Hanushek [33]), whereas the second one links quality of education to
labor income (Behrman and Birdsall [8], Card and Krueger [18] and Card and Krueger [17].)
In the  rst strand of literature, the quality of education is seen as a product of two inputs,
characteristics of schools and characteristics of individuals. Measurement of quality usually
is done through standardized tests. School inputs include the ratio of teachers to students
or the average experience of teachers. Individual characteristics are signi cant covariates of
education performance, since family background, genetics, etc, play an important role in the
acquisition of education and in the ability to perform on standardized exams.
73
In the second strand, a measure of quality (usually, test scores) is linked to labor income
through human capital theory. Quality of education enhances future wages because the
higher the quality of education, the more valuable are the skills individuals develop. Thus,
there is a theoretical link between quality of education and wages that can be estimated
empirically.
Given the data available, the thesis follows the  rst strand of the literature. The  rst
step in the estimation is to replicate the results of the previous chapter, only for the schools
that are going to be used in the regression analysis.1 The second step is the estimation of a
quality of education equation (Equation 4.1, below), using a treatment-control methodology.
As it is clear from the literature review, an important aspect of the relationship between
di erent degrees of decentralization is the trade o between quality and equity. In the last
part of the chapter we estimate Equation 4.1 for di erent households, according to their
income, and in this manner, shed some light onto the trade o between quality and equity.
The remainder of this chapter consists of three sections. The next section develops the
theoretical foundations for the production function for quality of education. Section 4.3
incorporates this theory within a quasi-experimental estimation framework. Section 4.4
presents the characteristics of the schools used in the estimation and their changes through
time as well as the main results of the estimation. The section also presents the result of the
estimation when the sample is partitioned by household income.
1In Chapter 3 we estimated the changes in inputs using all the schools available in the school level dataset.
When merging these data with the student data, some schools are not represented. Therefore, the  rst step
of the estimation is to replicate the results obtained for all the schools with the sub-sample that remains for
the subsequent estimations.
74
4.2 Quality of education: Production Function
The common model in the literature of quality of education (Q) is a function of observable
individual characteristics (X), school characteristics (Z) and an unobservable component
(U)
Q = f(X;Z;U) (4.1)
The typical approach to constructing a measure of education?s quality is to use a stan-
dardize test (S). In theory these types of exams are designed to measure the capacity to
reason and to solve non-mechanical problems, capacity that is enhanced by high quality edu-
cation. Therefore, such tests provide information on the quality of education that individuals
have.
The characteristics of the school, like the teacher/student ratio, wages of professors, ed-
ucation of professors, etc, as well as other unobservable characteristics like the relationship
between the school and the community, a ect the quality of education that individuals re-
ceive. On the other hand, the characteristics of the individual taking the exam are important.
These characteristics include family background (like income, size of the family, education
of parents, etc) and the demographic characteristics (like sex, race, etc.)
The  rst group of studies, including Coleman et. al. [19], picked a functional form of
the production function and estimated several speci cations of Equation 4.1. One surprising
result of the Coleman et. al. study was that school characteristics seem to play a very
small role in determining test scores, whereas individual characteristic play a decisive one.
75
Later studies also showed this regularity (again, the key references are Hanushek [32] and
Hanushek [33]).2
These results induced two lines of research. In the  rst researchers started to depart
from the notion of perceiving education as analogous to a production of an output in a
 rm, and therefore began to build upon a notion of a more complex production function of
education. Second, others researchers started work more in the quality of data and techniques
of estimation of the \simple" production function.
In production theory, a  rm tries to maximize pro ts through the selection of inputs,
given a production function and a cost structure. In the case of schools, it is not clear that
the objective is quality maximization. For instance, a recent poll in Colombia asked \Of the
next functions, which one is the most important objective of the education in Colombia?"
In response, 34% of the people answered \learning how to live in the community," 26% said
\learning to make the country better," 16% said \the objective of self-realization," 14%
said \a better economic future for people," 9% said \learning of tools to make Colombia
more competitive" and 1% answered \none of the above" (newspaper El Tiempo, April 7
of 2002). Clearly this is the opinion of \individuals in the street", and may not re ect the
opinion of principals or teachers. It is signi cant, however, that a plurality considers civic
values the principal objective of schooling, rather than better preparation or enhanced skills.
Hanushek [32] and [33], Manski [47] and Hoxby [41] forcefully make this point.
2A study by Heynaman and Loxley [37] divides poor countries from rich ones. They  nd that for high-
income countries, individual inputs are more important than school inputs, with the converse true for low-
income countries.
76
On the other hand, it is possible that the production function approach is valid, but
a ected by data, speci cation or estimation problems. Omitted variable bias can be present
in the estimation of Equation 4.1. For instance, the score of an individual may depend on
unobservable characteristics like ability, and these characteristics may be correlated with
socioeconomic factors. For example, higher levels of income are correlated with better nutri-
tion, and nutrition may be correlated with ability. In this case, the coe cients on individual
characteristics X are going to be upwardly biased.
In order to obtain consistent estimates it is possible to use instrumental variables tech-
niques or to implement randomized experiments. Both approaches are quite di cult:  nding
instrumental variables can be as di cult as  nding the unobservable variables that are the
source of the problem, and randomized experiments are expensive and quite uncommon to
 nd.
Another type of problem is measurement. For example, the typical variables used to
capture the relevant inputs of the school may not be the right ones. Hanushek [32] and [33]
identi ed three variables extensively used in the empirical literature: the experience and
education of teachers and the ration of students to teachers. In a comprehensive review
of empirical studies, Hanushek found that none of these three variables plays a statistically
signi cant role in explaining scores on standardized tests.3 Hanushek advance the hypothesis
that the critical variable in explaining scores is, for example, ability in teachers, which in
turn is combined with some characteristic of the school (small classrooms, good facilities,
3In several studies, the sign of the coe cient of the variables is not the expected one and when the sign
is the expected one, very often it is not signi cant.
77
etc) to produce good results in tests. In other words, the typical included variables are
neither necessary nor su cient variables in explaining test results.
In two recent articles, Krueger [42] and Hoxby [38] try to capture the e ect of class size on
standardized tests using exogenous variations in class size. Krueger [42] uses an experiment
implemented in the United States, in which, in a randomized way, children were allocated in
three di erent sizes of classrooms. In this way, and thanks to the fact that neither children
nor parents were allowed to pick the classroom, the e ect of class size on test was in principle
estimable. The empirical results were that size does matter: children in small classrooms had
higher test scores than children in big classrooms. However, the experiments were partially
contaminated, since some individuals changed classrooms from big to small sizes.
Hoxby [38] points out another potential problem in the above experiment. It is possible
that teachers in small classrooms implemented more e ort because they want the evidence
to show that size matters, and therefore, the results may be biased upward in favor of small
class size. Hoxby used another source of external variation. There exist small demographic
variations in the size of birth cohorts. Hoxby uses this variation as an instrument to estimate
the e ect of class size on test scores and  nds no e ect.
78
4.3 A framework for estimation: Quasi-experiments
4.3.1 A general model
One of the most important quests in modern economics is to  nd the e ect of speci c
programs, like job training, on an outcome of interest, like income (Heckman and Robb [36],
Heckman [35], Heckman, LaLonde and Smith [44] and LaLonde [43]). Quasi-experiments
are a way to circumvent the problem of not having a control group randomly chosen by
the researcher: in a quasi-experiment, an external shock a ects some individuals but not
others in a way that is plausibly not subject to self-selection. In the present context, the
institutional change from a centralized system to a decentralized one provides the potentially
exogenous source of variation that separates schools into the treatment and into comparison
groups. Hence, the treatment, in this case, is decentralization. In the comparison group are
students attending schools to which the decentralization does not apply. As discussed in
detail below, three sets of models, based on di erent treatment and control groups, are used
in the estimation.
The institutional change, however, does not mean that the problem of self- selection
into the type of school by an individual is not present: it may be the case that better (or
worse) students may be more likely to attend schools in the control group, for instance.
The literature in program evaluations (which is summarized in Heckman, LaLonde and
Smith [44]) studies the impact of individual outcomes that only depend on the personal
characteristics of \one" individual (or at least, they are modeled in this way). In the case
79
of quality of education, it is necessary to take into account the decisions of the two agents
simultaneously, the schools and the children.
Assuming a linear relationship, Equation 4.1 can be rewritten as
S =   X +  Z +U (4.2)
where S denotes test scores, the proxy for quality of education, Q.
Suppose now that decentralization is undertaken at time k, t = 0 <k<t = 1, and only
certain schools are a ected. Letting Tr denote students in treated schools and Co denote
students in control schools, test scores before the program for both groups can be represented
as4
STr0 =   XTr0 +  ZTr0 +UTr0
SCo0 =   XCo0 +  ZCo0 +UCo0
and the after post-test score can be represented as
STr1 =  +  XTr1 +  ZTr1 +UTr1
SCo1 =   XCo1 +  ZCo1 +UCo1
4The vector X includes a variable equal to 0 in the pre-reform period and 1 in the post-reform period
80
The objective of the researcher is to estimate  ,  and  .  measures the impact of
decentralization, holding schools inputs constant. The proposed estimator is essentially
a di erence-in-di erences estimator, since we have information on school and individual
characteristics both before and after treatment. The estimator is the following. In the  rst
place, the previous equations can be summarized as
S1 = D STr1 + (1 D) SCo1 (4.3)
S0 = D STr0 + (1 D) SCo0 (4.4)
where D = 1 for the treated group and D = 0 for the control one.
These two equations yield the basic equation to be estimated:
Sgt =  0 + 0 Tt + 1 Dg + 2 Dg Tt + 1 Xgt +  Zgt +Ugt (4.5)
where g = Tr;Co, treated (Tr) and control (Co) schools.
One important condition has to hold in order to have consistent estimators for Equa-
tion 4.5: the unobservable characteristics for both groups have to evolve in the same way.
The last condition includes unobservable characteristics in the production function for qual-
ity of education. Clearly, these can be unobservable characteristics of the individuals, such
as ability, or they can be unobservable characteristics of the school, such as motivation of
teachers. In this sense, U can be decomposed in two terms, one re ecting the individual part
81
and the other the school part: U = USchools +UIndividuals.
At least two factors can contaminate the experiment, each of which may induce bias in
the estimation of the e ect of the program. The  rst can be described as general equilibrium
e ects and the second is a problem with the selection of individuals.
The problem of general equilibrium e ects can best be described with an example. Sup-
pose that, based on the idea that decentralization directly a ects public schools and not
private ones, the treated group consists of the public schools and the control includes the
private ones. Further suppose that, on average, the skills of individuals in the private schools
were higher than those in the public schools before decentralization.5 Now, suppose that
after decentralization a signi cant portion of the population decides to move from private
education to public education. In this case, it may be that the quality of education in the
public school increases, not because of decentralization itself, but because of this movement
of better students between the two types of schools. Continuing with the same example,
suppose that the movement of students from the private to public schools is drawn from the
lower, left tail of the ability distribution of students in private schools. In this case, the level
of educational outcomes will increase in the private schools, which implies problems with the
measurement of the direct impact of decentralization on the quality of education.
The second problem can be very severe in the context of education. Presumably public
school students di er from private school students in signi cant ways; income is one example.
In the case in which public education is cheaper than private education, family incomes will
5For instance, if nutrition is better in private schools than in public ones and better health is re ected in
better school performance.
82
di er across school types. Again, a proper comparison group is important in order to avoid
potential contamination of the experiment.
The estimation of Equation 4.5 captures two potential channels of the program?s impact:
the impact through the changes in the characteristics of the schools because of the program,
  [(ZTreatment1  ZTreatment0 ) (ZControl1  ZControl0 )], and a \inputs-constant treatment e ect",
 2.
The economic interpretation of the change in the characteristics of the school is an in-
teresting one. Following with the example in which the treatment group is composed of the
public schools and the control of the private ones, before the decentralization process, public
schools were constrained by the decisions of the central government, and thus potentially
the allocation of resources were not optimal. Once the system was decentralized, public
schools could start changing internal characteristics since decentralization provided more
autonomy to the public schools. The results of the previous chapter indicate that indeed
decentralization induced increases in the mean level of teacher education.
The economic interpretation of the \inputs-constant treatment e ect" of decentralization
is also an interesting one. Decentralization can have some positive unobservable e ects
like a better relationship between the local community and the school. For instance, after
the decentralization the school may have more power to change certain practices to better
accommodate certain parental demand in the community. Presumably, the \inputs-constant
treatment e ect" will capture these e ects.
83
4.3.2 Speci c models for estimation
At this moment, it is important to explain in detail the three di erent sets of treatment and
control groups. Table 4.1 summarizes the models used in the estimations.
Model 1
In the  rst model, the treatment group consists of the public schools and the control group
of the private schools. In this sense, the treated are the students who are in schools a ected
by decentralization, namely in the public ones. A natural option for the control group is the
private schools.
In mathematical terms, the general model of the previous sections is applied to this
speci c one:
S1 = Dg SPu1 + (1 Dg) SPr1 (4.6)
S0 = Dg SPu0 + (1 Dg) SPr0 (4.7)
where D = 1 for public schools (Pu) and D = 0 for private ones (Pr).
Given these two equations, the basic equation to be estimated is
Sgt =  0 + 0 Tt + 1 Dg + 2 Dg Tt + 1 Xgt +  Zgt +Ugt (4.8)
where g = Pu;Pr, public (Pu) and private (Pr) schools.
84
In order to gain an intuition of the model, suppose that the treated group, public schools,
observed two types of shocks: the e ect of decentralization and an external, nationwide
event like a recession. A recession can a ect education outcomes in several ways. The most
direct one is the reduction of resources to schools. If we were analysing test scores only for
public schools, before and after decentralization we will capture not only the e ect of the
decentralization but also the e ect of the negative shock in the economy. In this case, the
control, that is, private schools, will observe the e ect of the nationwide shock, but not the
e ect of decentralization. By taking the di erence in the test scores before and after, for the
public and private schools we can  nd the e ect of decentralization.
In more formal terms, Table 4.2 presents the argument by making use of Equation 4.8.6
In this case,  0 is the common shock that a ects both private and public schools after
decentralization;  1 is the coe cient of the dummy that identi es the public schools;  2
is the coe cient that captures the \inputs-constant treatment e ect," the parameter of
interest. Again, if we take test scores before and after decentralization for public schools, we
will capture both the e ect of decentralization and the nationwide shock,  0 + 2. If we take
the di erence in the test scores for private schools only, before and after decentralization, we
will get the e ect of the nationwide shock,  0. Taking the di erence of these two di erences,
we  nd the parameter of interest:  2 = ( 0 + 2) ( 0).
6The table is a simpler model as it assumes constant the covariates X and Z.
85
Model 2
While institutional change provides a potentially exogenous variation, there may have been
di erences among public schools in their degree of decentralization. For instance, we would
expect that schools in departments with less initial political and administrative autonomy
would have experienced greater e ects of decentralization than schools in departments with
higher initial autonomy. The impact of any decentralization would be greater in the localities
that were highly dependent on the central government prior to decentralization, and which
became more independent after decentralization.
Gonzalez [30] and Vergara and Simpson [61] evaluate the administrative performance of
departments during the nineties by constructing an index that shows their relative position
in the country.7 This index places Bogota, Antionquia and Valle at a higher initial point
of decentralization than the other departments. Administratively these departments were
more autonomous than others in the beginning of decentralization. The major cities of the
country and their respective departments - Bogota (Capital District), Medellin (capital of the
department of Antioquia) and Cali (capital of Valle) - were able to prior to decentralization
make local decisions that did not depend on the central government and had more resources
of their own to implement local policies.
For these reasons, Model 2 places schools in departments with an initially low degree of
decentralization into the treatment group, and as controls schools in the three abovemen-
7Vergara and Simpson [61] rank the autonomy of localities according to four conditions: (1) possibility
of community participation in decisions, (2) a well-organized information system on the administrative
aspects, (3) an organizational structure that can meet the new responsibilities of decentralization, and (4) a
distribution of resources that meets the new responsibilities (Vergara and Simpson [61], page 18).
86
tioned departments. The hypothesis in this case is that the departments that were more
dependent on the central government prior to decentralization had more bene ts from de-
centralization than the departments that were already highly autonomous at the beginning
of decentralization.
In mathematical terms, Model 2 is as follows:
S1 = Dl SHD1 + (1 Dl) SHI1 (4.9)
S0 = Dl SHD0 + (1 Dl) SHI0 (4.10)
where D = 1 for schools in departments initially highly dependent from the central gov-
ernment (HD) and D = 0 for schools in departments initially highly independent from the
central government (HI).
Given these two equations, the basic equation to be estimated is
Slt =  0 + 0 Tt + 3 Dl + 4 Dl Tt + 1 Xlt +  Zlt +Ult (4.11)
where l = HD;HI, treated (HD) and control (HI) schools.
Intuitively, there are some common shocks that a ect all public schools and that are
independent of decentralization. A clear example is a change in the curriculum that a ects
all public schools, which may have an impact on test scores. In theory, schools in initially
highly dependent departments (HD) will have the e ect of decentralization  4 and of the
87
change in curriculum  0 (see Table 4.3) Schools in initially highly independent departments
(HI) will observer only the e ect of the change in the curriculum  0, and therefore, we can
infer the e ect of decentralization by the second di erence, across schools and time.
Model 3
Model 3 is derived from Models 1 and 2: it is a di erence in di erence in di erence (DDD)
model. Retaking the notation of the di erence in di erence estimator (Equations 4.4), we
have now four basic equations:
Sg0 = Dg SPu0 + (1 Dg) SPr0 (4.12)
Sg1 = Dg SPu1 + (1 Dg) SPr1 (4.13)
Sl0 = Dl SHD0 + (1 Dl) SHI0 (4.14)
Sl1 = Dl SHD1 + (1 Dl) SHI1 (4.15)
where l indicates if the school is in a department initially highly dependant on the central
government (HD) or in an initially highly decentralized department (HI) and g indicates
whether the school is public (Pu) or private (Pr). These four equations can be restated as
88
Sg;lt =  0 + 0 Tt + 1 Dg + 2 Dg Tt + 3 Dl + 4 Dl Tt +
 5 Dg Dl + 6 Dg Dl Tt + 1 Xg;lt +  Zg;lt +Ug;lt (4.16)
where g = Pu;Pr and l = HD;HI.
This equation describes the triple di erence estimator, an extension of the di erence in
di erence estimator (Equation 4.5). The triple di erence estimator is a more  exible model
than the previous two. Model 1 does not allow for changes that will a ect all public schools
(e.g. curriculum changes), or in other words, it assumes  3 = 0. Model 2 does not allow for
general shocks in the economy that will a ect both private and public schools; it assumes
that  1 = 0. In contrast, Model 3 allows both types of shocks, events a ecting only public
schools and events a ecting both public and private schools, allowing us to recover the e ect
of decentralization ( 6) after controlling for these types of e ects (see Table 4.4).
4.4 Estimation
All the estimations are based on the same basic equation (Equation 4.5). Two years are used.
1991 represents the pre-decentralization period and 1999 represents the post-decentralization
period. There are three reasons for choosing only these two years for the estimation. First,
for 1990 the number of school observations is signi cantly lower than in any other year.
Second, taking the extreme points ensures that the data truly represents the pre- and post-
89
decentralization periods. As presented in the description of decentralization in Colombia,
the new constitution was enacted at the end of 1991 and Laws of 1993 and 1994 developed
decentralization. The transfers from the central government to the localities started in
1992, and therefore, the impact of decentralization began that year. Third, the e ects
of decentralization may take several years to become apparent. Using the extreme points
ensures that due to the slow nature of the changes, the process of adjustment will have had
enough time to operate. In any case, own sensitivity analysis, found no signi cant di erences
in the results when other years were included.
It is important to emphasize that the estimation employs as data a test taken by prospec-
tive college applicants. Despite the fact that almost everyone graduating from high school
takes the ICFES exam, it is not a requirement to get a high school diploma.
4.4.1 Changes in School Characteristics
The previous Chapter showed evidence on the impact of decentralization on the character-
istics of schools. This section replicates the same procedure but limiting the data to those
schools that are in the estimation of Equation 4.5. Following Hanushek [32] and [33], three
school variables were include in the estimation of Equation 4.5: one variable that captures
the quality of the teachers, or teachers? years of education; a variable that captures the size
of the school, or the total number of teachers; and  nally, one that captures the average size
of the classroom, or the ratio of teachers to students.
Table 4.5 presents the summary statistics for these variables in the public and private
90
schools used in the estimation of Model 1. Consistent with the previous Chapter, the same
basic results hold:  rst, education of teachers increased more in public schools than in private
ones; second, the number of teachers in public schools increased more than in private ones;
 nally, the ratio of teachers to students decreased more in public schools than in privates
ones.
It is important to stress one result from Table 4.5. It shows the same pattern as seen
in the previous Chapter, indicating that the sub-sample of schools used preserve the same
characteristics as the whole sample.
Table 4.6 presents the same evidence as the previous table, but for Model 2. Consis-
tent results emerge from this table. In the treatment group, schools in the initially less
autonomous departments, the quality of the teachers increased more than the quality of
teachers in the control group. The same pattern occurs with the number of teachers. Fi-
nally, the ratio of teachers to students declines more in the treatment group than in the
control group.
Overall, the main argument of the previous Chapter still holds: the quality of teachers
increased more in the treated group.8
4.4.2 Test scores and decentralization: regression analysis
To recapitulate, the objective is the estimation of Equation 4.5. Only schools with informa-
tion in both periods are used. The dependent variable is the test score for a student, and
8However, under the speci cation of Model 3, we do not observe this trend. In Model 3 quality of the
treatment group teachers decreases through time relative to the control group teachers.
91
the right hand side variables consist of school characteristics (Z), individual characteristics
(X), a dummy variable for time (Tt), which is zero for the pre-decentralization period and
one for the post-decentralization period, a dummy variable for the treatment group, which
is one for the treated group and zero for the control group (Dg), and the interaction term
between (Dg) and (Tt), which is the variable that captures the \inputs-constant treatment
e ect" of decentralization. We used as characteristics of the schools the years of education
of teachers, the number of teachers and the ratio teachers / students. As for the variables
capturing characteristics of students taking the exam, three variables were included: income
of the head of the household (in number of minimum wages), and the sex and age of the
student.9
One important aspect of the estimation that needs to be discussed is the fact that it
uses test results and intrinsic individual characteristics from the same year as the school
inputs. Hanushek [32] discuses the issue of timing in this class of models. According to the
author, contemporaneous measures of inputs with contemporaneous test scores can lead to
measurement and speci cation errors. Education of individual i, who went to school from
1989 to 1999, depends on the inputs that he / she received during all the years of schooling.
However, we are matching inputs of a speci c year (1999) with the educational outcome of
individual i of the same year. Therefore, if the set of inputs used in the estimation is not the
relevant set for the education of the individual, we are facing a measurement error problem.
The assumption about timing is that decentralization induces changes in inputs of the
9In this case, other controls, such as the marital status and migration status, were also used but they
were not statistically signi cant nor did the results reported here change upon their inclusion.
92
schools and that those changes are cumulative through time. An individual who started
school in 1988 began to bene t from these changes in inputs from 1991 onwards. In e ect,
the levels seen in 1999 re ect the cumulative changes induced by decentralization from 1990
onwards. As discussed in the previous chapter, the changes in inputs are invariant to the years
taken: the same pattern emerge if we take the average inputs between 1991 and 1992 as pre-
decentralization and 1996 and 1999 as post-decentralization or 1991 as pre-decentralization
and 1999 as post decentralization.
The summary statistics of the individual characteristics used in the estimation of Equa-
tion 4.5 are presented in Table 4.5 and Table 4.6. In the sample, 55% of the population is
female. This percentage is constant across time and the two groups (treatment and control.)
On average, students in public schools are slightly older than the students in private schools.
The mean  uctuates around 18 years. Finally, and probably as a re ection of the 1999
recession in Colombia, the income of the individuals decreased on average of 3% between
the before- and the after-decentralization periods. However, the pattern is quite di erent in
the treatment and control groups; for the treated the average income declined, and for the
control it increased. One potential reason for this pattern can be compositional: if more
people are going into college, then it is possible that the sample re ects people in the left
tail of the distribution for the after-decentralization sample. Barrera and Higuera [7] present
preliminary statistics on college enrolment that are in line with this hypothesis.
As presented in the table, there is an important increase in the sample size between the
two periods. Moreover, the increase occurs in both public and private schools, although it
93
is more pronounced, in relative terms, in the public ones. Given that the span of time is
10 years, and the fact that through time more people are getting primary, secondary and
tertiary education in the country, it is possible that the increase re ects an increase in the
overall educational enrolment.
This relative increase in public schools, however, may raise concerns about selection
problems. In short, if the program is successful in increasing the quality of education, it is
reasonable to expect migration from private to public schools of certain types of individuals
(for instance individuals with a higher ability). A simple and direct way to see if the problem
is a serious one is the generation of CDFs for tests scores for both public and private schools.
Sharp di erences in the DCFs may indicate the presence of selection problems in the two
samples. Graphing the two distributions one observes that the density functions look very
similar for both types of schools, and for both time periods. At least this evidence does not
support the suggestion of selection problems.
The  rst results presented are the di erence in di erence estimates without covariates
(Table 4.5, DD for Mean scores) from Model 1. The DD estimator is -4.72, and it is statis-
tically signi cantly di erent from zero. As the Table shows, test scores for private schools
were, before decentralization, higher than the public ones. For both types of schools, the
mean test score increases through time, but it increases more for the private schools.
The analogous estimation for Model 2 (Table 4.6), shows a di erent picture. The treat-
ment group, which consists of the public schools in departments with initially less autonomy,
had lower test scores than the control group. For both groups the mean test score increased,
94
but more for the treated group. For this reason, the DD estimation in this case is positive
(3.426) and statistically signi cant.
Summing up, these  rst sets of estimators are showing two preliminary results:  rst,
for both Models 1 and 2, treatment mean scores increased between pre-decentralization and
post-decentralization periods; second, mean scores for the control group in Model 1 increased
more and in Model 2 increased less than for their respective treatment groups.
Table 4.7 presents the results from the regressions of test scores against individual char-
acteristics, school characteristics, a time dummy, a treatment dummy, and the interaction
of the last two variables (DD estimation), for the case of Models 1 and 2, and all the com-
binations of dummies of treatment and control groups for Model 3 (2 to the power of 3).
All the results presented are controlled for  xed e ects by departments. It is important to
stress that only schools with observations in both periods were used.
Before presenting Models 1, 2 and 3, Table 4.7 shows the results of a regression taking
only public schools, before and after decentralization. That estimated (label \Model 0")
indicates that test scores in public schools after decentralization are higher than before:
the time-dummy coe cient is 0.91 and statistically signi cant. Clearly, as we discussed in
the previous sections, that result not only has the e ect of decentralization, but also others
e ects such as nationwide shocks and other changes beside decentralization that a ect public
schools. In other words, we cannot take the increment in test scores in public schools as
the sole result of decentralization since other factors are embedded in the estimated e ect.
Precisely, that is the reason of including controls, as Models 1-3 do.
95
In Model 1 the coe cients for teachers educational level and for the number of teachers
in the school are both positive and signi cant. The coe cient for the ratio of teachers
to students is negative and signi cant (95% level of con dence). One explanation for this
negative e ect can be that the relationship between test scores and the ratio of teachers
to students is not linear. It is possible that in small classes students are losing some of the
positive externalities of having other students and in larger sized classrooms the overcrowding
e ect from greater number of students creates a negative externality. The coe cient for the
sex of the individual is positive and signi cant (females tend to have lower scores after
controlling for the other variables). The coe cient for income is positive and signi cant as
expected. The coe cient for the age individual is negative and signi cant such that younger
people tend to perform better in the exam.
The coe cient for the treatment dummy is negative and signi cant (-12.82, with an
standard error of 0.4). That is, on average the public schools have lower test scores. The
time dummy is positive but not statistically signi cant. The coe cient on the di erence
in di erence variable is positive (0.94) and signi cant at the 95% level of con dence. In
other words, after controlling for individual and schools characteristics, there is evidence of
a positive impact of decentralization on test scores.
With respect to Model 2 the results are consistent with the results of Model 1 with some
minor variations. The majority of the coe cients are very similar, not only in terms of the
point estimate but also in terms of their signi cance. The coe cient on the ratio of teachers
to students is, however, now positive but not statistically signi cant. The second di erence
96
in the estimates is the coe cient of the DD variable that changes from 0.94 to 2.38.
This two models yields the same pattern of results: after controlling for individual and
school characteristics, in a model with \inputs-constant treatment e ects" by department,
there is evidence of a positive impact of decentralization on public schools and on schools
located on departments with initially lower autonomy. Moreover, the inclusion of covariates
reverses the sign on the impact of decentralization relative to the DD estimate without
covariates. In other words, covariates matter. The di erence in di erence estimator without
covariates may be collecting some of the intrinsic characteristics of the localities, schools and
individuals. Once the estimation includes controls for localities, schools and individuals, the
\inputs-constant treatment e ect" impact is positive.
The results of Model 3, the di erence in di erence in di erence model, present a clear
contrast to the results of Models 1 and 2 in terms of the \input-constant treatment e ect"
of decentralization. In this case, the coe cient is negative and signi cant (-5.05). Given
that both the public-private e ect and the initially low-high decentralized e ect are positive,
the driving force behind the result is presumably the private schools in highly decentralized
areas. The result, in light of the previous ones, is that despite the positive e ect of decen-
tralization in terms of public-private and low-high treatment-control groups, a more  exible
model that allows for nationwide e ects and public schools-wide e ects gives the result that
decentralization has not incremented the quality of education.
In order to see potential explanations of the results, it is important to analyze the overall
impact of decentralization on the quality of education. The impact of decentralization on
97
the school quality is not only limited to the DD and DDD estimators but also includes the
e ects of decentralization on the inputs of the treated schools. If, as it is argued in the
previous Chapter, all the changes in schools? inputs are causal, then an important part of
the e ect of decentralization is operating via these e ects.10
Table 4.8, Table 4.9, Table 4.10 present the overall e ect of decentralization for each
Model via the two channels, changes in inputs of schools and the DD and DDD estimators.
The impact on inputs is the change in the inputs of the treatment schools times the coe cient
found in the di erent models. The total e ect of decentralization is positive for the  rst two
models (2.87 and 4.02 points respectively) and negative 3.99 points for Model 3.
A main di erence between Table 4.10 and the other two tables (Table 4.8 and Table 4.9)
is the impact of decentralization on the changes in inputs, especially in terms of the quality
of teachers. The di erence across time and across groups (initially low, public schools versus
the rest of schools) for teachers? education is negative (-0.13), representing a clear contrast
with the same data for models 1 (1.69) and 2 (1.80). In plain words, the gap in terms of
the quality of teachers augmented between public schools in initially centralized areas and
the other schools. Clearly, if part of the e ect of decentralization operates via inputs, the
results are not surprising.
Summing up, there is evidence of a positive impact of decentralization on test scores
when the treatment and control groups are public versus private schools and initially low
versus initially high localities. In the last decade, the public schools observed an increase in
10It may be the case that the changes in the inputs are not completely due to decentralization. The results
are, in this case, an upper bound of the e ect of decentralization.
98
test scores. Once we control by individual characteristics, school characteristics and \inputs-
constant treatment e ect" of departments, the di erence in di erence estimator is positive
and signi cant. On top of this e ect, there is evidence that decentralization changed inputs
used in schools, which also induced increments in the mean test score for public schools.
However, in the most general,  exible model, the results are that decentralization did not
induce increments in the quality of education. Potentially this result can be explained by
the fact that the inputs of the schools in the treatment group of Model 3 did not change as
they did in the other two models. In this model we do not observe increments in the teacher
quality in the treatment group.
In order to test the null hypothesis of these estimates equal zero, a general test of the
form R  = 0 was run. In this case, R is the vector of estimates of changes in inputs,  is the
vector of estimated coe cients. The variance of the estimator is Var(R ) = R Var( ) R0,
and the distribution an F(j;n k) = (R )0Var(R )( 1) j(R ) where j is the number of
restrictions. For each Model, the value of the F() was 36.87, 52.25 and 17.60 respectively,
all of them signi cant at the 99%. Therefore, the null hypothesis is rejected.
In any case, the e ect of decentralization, according to the estimates, is small. The mean
of test scores for the public schools in 1999 was 242.56, with an standard deviation of 55.5,
whereas the same data for schools in initially highly dependant localities in 1999 was 242.76,
with and standard deviation of 54.9. Comparing this number with the estimate of the e ect
of decentralization, yield a very small impact of decentralization.
In comparing this result with similar studies done in other countries, it is very di cult
99
given the disparity in the scales of test scores, input measures, and other traits. Follow-
ing Hanushek [32], some ideas on the size of the results can be discussed. First, Hanushek
presents evidence that lead to the conclusion that \there appears to be no strong or system-
atic relationship between school expenditures and students performance". The results of two
models (1 and 2) present evidence that actually changes in inputs of schools, presumably
induced by decentralization, has had an impact on students? performance. However, Model
3 shows that the impact has not been positive. Second, one simple hypothesis that explains
the apparently small impact of decentralization on student performance is that the measure
of the performance is quite parsimonious, and therefore, one cannot expect large changes in
short periods of time.
4.4.3 E ciency versus equity
As it is described in the  rst chapter of the dissertation, there is an apparent trade-o between
e ciency and equity for di erent degrees of decentralization. In order to explore this issue,
we formulate the following question: Are there di erences in the impact of decentralization
across income groups?
Following the same methodology as in the previous section, we run Models 1 to 3 but
for di erent household groups, depending on their income. For individuals who go to public
schools, we split the sample between households that have less than the average income and
those who have more than the average income. Then, we estimate Models 1 to 3 for each
group. In this way, we can explore the e ects of decentralization on di erent income groups.
100
Table 4.11 shows the results for Model 1. The  rst column presents estimates when
the control group consists of the individuals who go to private schools and the treatment
group of the individuals who go public schools and belong to household with less than the
average income. Column two shows the estimates with the same control group but where the
treatment group consists of the individuals who go to public schools and belong to households
with more than the average income.
For the most part, the two regressions are very similar. However, there are two clear
di erences between the results. First, the coe cient on the time dummy for the sample
with public school individuals with lower than mean household income is negative, whereas
for the other sample it is positive. Second, and more importantly, the \inputs-constant
treatment e ect" impact is positive for low-income individuals, whereas it is negative for the
high-income ones. If the treated are low-income individuals in public schools, and the control
are individuals in private schools, the e ect of decentralization is positive, and the contrary
occurs when the treatment group consists of high-income individuals in public schools.
Table 4.12 presents the same type of analysis but for Model 2. The estimation results
for both samples (low- and high-income public school students) are very similar. Moreover,
for both treatment groups the \inputs-constant treatment e ect" impact is positive and
statistically signi cant. However, the point estimate for the low-income treatment group
is lower than for the high-income treatment group implying that decentralization induces
higher gains for high-income individuals, at least through the \inputs-constant treatment
e ect." This result concurs with the theory laid out in Chapter 2.
101
Finally, Table 4.13 presents the results for Model 3. Recall that this model is a nested
model of Models 1 and 2, and therefore, it is a more  exible speci cation of the estimation.
In short, it allows for both nationwide shocks and shocks a ecting only public schools. The
estimations for low- and high-income subsamples in this model preserve the same charac-
teristics as the previous two in which the estimation is very similar for the two subsamples.
In this model, we do not see di erential \inputs-constant treatment e ect" impacts across
low- and high-income samples. For both samples the impact is negative and statistically
signi cant with point estimates that are statistically equal. In other words, the e ect of
decentralization is not di erent across income groups.
In conclusion, the evidence on the impact of decentralization across income groups is
mixed. On one hand, against a priori beliefs, decentralization a ects positively low-income
public students in Model 1. On the other, Models 2 and 3 do not give di erences in the
impact across individuals. In Model 2, the impact is positive, although of slightly di erent
magnitude, for both income groups; in Model 3, the impact is negative for both income
groups.
102
Table 4.1: Models used in the Estimation
Model 1 Treatment: Public Schools
Control: Private Schools
Model 2 Treatment: Public schools in initially less autonomous departments
Control: Public schools in initially more autonomous departments
Model 3 Di erence in di erence in di erence, Models 1 and 2
Table 4.2: Model 1
Before After Di .
Private  0  0 + 0  0
Public  0 + 1  0 + 0 + 1 + 2  0 + 2
Di .  1  1 + 2  2
Table 4.3: Model 2
Before After Di .
H.Indepen.  0  0 + 0  0
H.Depend.  0 + 3  0 + 0 + 3 + 4  0 + 4
Di .  3  3 + 4  4
Table 4.4: Model 3
Before After Di .
H.Indepen.
Private  0  0 + 0  0
Public  0 + 1  0 + 0 + 1 + 2  0 + 2
Di .  1  1 + 2  2
H.Depend.
Private  0 + 3  0 + 0 + 3 + 4  0 + 4
Public  0 + 1 + 3 + 5 ( 0 + 0 + 1 + 2+  0 + 2 + 4 + 6
 3 + 4 + 5 + 6)
Di .  1 + 5  1 + 2 + 5 + 6
Di . in Di .  2 + 6
Third Di .  6
103
Table 4.5: Summary statistics Model 1
Pre-decent. Post-decent. DD
Model 1 Public Private Public Private
schools schools schools schools
A. Schools character.
Teachers education 15.07 15.28 16.41 15.73 0.89
(Standard error) 0.00 0.00 0.00 0.00 0.01
No. of teachers 30.87 23.91 32.96 24.86 1.14
(Standard error) 0.07 0.07 0.05 0.06 0.14
Teacher/students 0.06 0.06 0.05 0.06 -0.00
(Standard error) 0.00 0.00 0.00 0.00 0.00
B. Individ. character.
Sex (% males) 0.45 0.44 0.45 0.44 -0.003
(Standard error) 0.002 0.003 0.001 0.002 0.004
Age 18.57 17.90 18.31 17.66 -0.03
(Standard error) 0.01 0.01 0.01 0.02 0.03
Income 1.80 2.03 1.46 2.63 -0.94
(Standard error) 0.00 0.00 0.00 0.01 0.01
Mean scores 238.31 250.95 242.56 259.92 -4.72
(Standard error) 0.18 0.27 0.15 0.26 0.45
Sample size 74138 39331 136870 59019
104
Table 4.6: Summary statistics Model 2
Pre-decent. Post-decent. DD
Model 2 Highly Less Highly Less
dependent dependent dependent dependent
A. Schools character.
Teachers education 14.90 15.34 16.39 16.45 0.38
(Standard error) 0.00 0.01 0.00 0.01 0.01
No. of teachers 29.64 32.86 33.16 32.51 3.87
(Standard error) 0.09 0.10 0.06 0.07 0.18
Teacher/students 0.06 0.05 0.05 0.04 -0.00
(Standard error) 0.00 0.00 0.00 0.00 0.00
B. Individ. character.
Sex (% males) 0.46 0.44 0.46 0.43 0.002
(Standard error) 0.002 0.003 0.002 0.002 0.005
Age 18.65 18.46 18.34 18.23 -0.072
(Standard error) 0.02 0.01 0.01 0.02 0.04
Income 1.79 1.82 1.42 1.54 -0.087
(Standard error) 0.00 0.00 0.00 0.01 0.01
Mean scores 237.22 240.02 242.76 242.13 3.43
(Standard error) 0.23 0.30 0.18 0.27 0.51
Sample size 45469 28669 93225 43645
105
Table 4.7: Regression analysis
Dependent Var.: Test scores Model 0 Model 1 Model 2 Model 3
Teachers years of education 1.82 1.69 1.80 1.75
(0.12) (0.10) (0.12) (0.10)
Number of teachers 0.25 0.37 0.25 0.36
(0.01) (0.01) (0.01) (0.01)
Ratio teachers / students 2.84 -3.94 3.02 -4.73
(3.10.) (2.13) (3.11) (2.13)
Age of individuals -2.62 -3.11 -2.62 -3.10
(0.03) (0.03) (0.03) (0.03)
Sex of individual 14.03 13.76 14.03 13.77
(0.24) (0.20) (0.24) (0.20)
Family income 4.64 6.30 4.65 6.29
(0.11) (0.08) (0.11) (0.08)
Time dummy (Dt) 0.91 0.48 -0.64 -2.67
(0.31) (0.39) (0.48) (0.53)
Treatment dummy (Dg:Pu.=1) -12.82 -16.61
(0.40) (0.58)
Treatment dummy (Dl:HD.=1) -21.24 -16.29
(6.63) (5.28)
Dg*Dl 8.20
(0.79)
DDg: Dg*Dt 0.94 2.88
(0.48) (0.71)
DDl: Dl*Dt 2.38 6.73
(0.56) (0.77)
DDD: Dg*Dl*Dt -5.05
(0.96)
Constant 230.34 251.08 231.75 254.09
(1.86) (1.62) (1.89) (1.65)
N 190499 276460 190499 276460
F 670 1231 653 1147
Adj R-sq. 0.11 0.15 0.12 0.15
(Standard errors in parenthesis)
106
Table 4.8: Total e ect of decentralization: Model 1
Change in inputs Coe cient Impact (product)
Teachers years of education 0.89 1.69 1.50
Number of teachers 1.14 0.37 0.42
Ratio teachers / students -0.002 -3.94 0.01
Di erence in di erence 0.94 0.94
Total E ect 2.87
Table 4.9: Total e ect of decentralization: Model 2
Change in inputs Coe cient Impact (product)
Teachers years of education 0.38 1.80 0.69
Number of teachers 3.87 0.25 0.97
Ratio teachers / students -0.003 3.02 -0.009
Di erence in di erence 2.38 2.38
Total E ect 4.02
Table 4.10: Total e ect of decentralization: Model 3
Change in inputs Coe cient Impact (product)
Teachers years of education -0.13 1.75 -0.23
Number of teachers 3.35 0.36 1.22
Ratio teachers / students -0.013 -4.73 0.06
Di erence in di erence -5.05 -5.05
Total E ect -3.99
107
Table 4.11: Impact di erences across income groups: Model 1
Dependent Var.: Test scores Below mean income Above mean income
Teachers years of education 1.68 1.69
(0.12) (0.14)
Number of teachers 0.39 0.50
(0.01) (0.01)
Ratio teachers / students -6.73 -4.51
(2.37) (2.52)
Age of individuals -3.64 -3.05
(0.03) (0.03)
Sex of individual 13.99 13.07
(0.23) (0.28)
Family income 7.59 6.49
(0.11) (0.11)
Time dummy (Dt) -0.33 0.48
(0.38) (0.42)
Treatment dummy (Dg:Pu.=1) -13.35 -11.67
(0.44) (0.52)
DDg: Dg*Dt 4.62 -2.69
(0.54) (0.62)
Constant 258.23 248.80
(1.86) (2.22)
N 199744 162677
F 1101 685
Adj R-sq. 0.18 0.14
(Standard errors in parenthesis)
108
Table 4.12: Impact di erences across income groups: Model 2
Dependent Var.: Test scores Below mean income Above mean income
Teachers years of education 1.74 1.90
(0.14) (0.21)
Number of teachers 0.20 0.29
(0.01) (0.01)
Ratio teachers / students 1.91 3.69
(4.30) (4.60)
Age of individuals -3.10 -2.25
(0.04) (0.04)
Sex of individual 14.43 13.24
(0.29) (0.41)
Family income 7.33 2.38
(0.34) (0.23)
Time dummy (Dt) 2.78 -2.18
(0.65) (0.83)
Treatment dummy (Dl:HD.=1) 0.87 8.28
(13.57) (13.98)
DDl: Dl*Dt 1.88 3.59
(0.68) (0.96)
Constant 237.14 230.71
(2.34) (3.29)
N 113783 76716
F 395 205
Adj R-sq. 0.12 0.09
(Standard errors in parenthesis)
109
Table 4.13: Impact di erences across income groups: Model 3
Dependent Var.: Test scores Below mean income Above mean income
Teachers years of education 1.69 1.75
(0.12) (0.14)
Number of teachers 0.39 0.49
(0.01) (0.01)
Ratio teachers / students -7.65 -5.39
(2.37) (2.52)
Age of individuals -3.64 -3.05
(0.03) (0.03)
Sex of individual 14.00 13.07
(0.23) (0.28)
Family income 7.58 6.44
(0.11) (0.11)
Time dummy (Dt) -3.43 -2.52
(0.52) (0.56)
Treatment dummy (Dg:Pu.=1) -16.57 -16.35
(0.65) (0.81)
Treatment dummy (Dl:HD.=1) -12.83 -33.89
(8.43) (16.09)
Dg*Dl 6.92 9.27
(0.87) (1.05)
DDg: Dg*Dt 6.66 -0.30
(0.80) (0.96)
DDl: Dl*Dt 6.64 6.38
(0.75) (0.82)
DDD: Dg*Dl*Dt -5.18 -5.43
(1.03) (1.25)
Constant 261.28 251.29
(1.87) (2.26)
N 199744 162677
F 1026 640
Adj R-sq. 0.18 0.14
(Standard errors in parenthesis)
110
Chapter 5
Conclusions
The main contribution of this thesis is the empirical e ort to test a critical question in
economics: Does decentralization provides better quality of education? This question has
both theoretical and practical implication. On one hand, there is an important strand of
theoretical literature exploring the issue from several points of view. On the other, numerous
countries started to implement decentralization during eighties and nineties.
The second chapter of the dissertation explores the issue from a theoretical point of view.
The leading argument of the discussion is the apparent trade-o between e ciency and
equity of centralized and decentralized systems. There are theoretical reasons for decentral-
ization increasing e ciency. Centralized systems, however, may induce a more homogenous
expenditure of resources across localities, inducing more equity into the system.
Based on theory, decentralized schools might perform better than centralized ones. First,
decentralized schools have better information in terms of the preferences of the individuals
111
in the community and on the idiosyncratic problems that the school faces. Second, local
control might increase community involvement and this can lead to better process of checks
and balances on the local provider of the service. Despite this, there are some reasons
to think that decentralization might imply e ciency losses, based on general equilibrium
models. These models yield results in which decentralization might segregate localities and,
in presence of local externalities that implies losses in e ciency. Also, there may be economies
of scale in the provision of certain services and the size of the government.
The results from the empirical studies, based on court ordered equalizations across com-
munities presented in USA, are mixed at best. Some of them  nd that court orders actually
induce equalization of expenditures and that poor localities increases expenditure. Other
studies  nd that equalization occurs towards the districts with lowest expenditure, that rich
districts are induced to spend less, and that several localities receive educational services
that do not match their preferred consumption bundles of the community.
As a conclusion, the second chapter underlines the importance of empirical tests. Chap-
ters 3 and 4 of the dissertation examine the impact of the Colombia decentralization on the
quality of education.
Chapter 3 describes the main aspects of decentralization in Colombia. In short, Colombia
experienced a partial decentralization in which important resources were transferred from the
central power to the local governments, but with signi cant limitations on the expenditure of
the resources. The chapter presents as well the data used in the estimation. It argues that an
important channel for the impact of decentralization on education is via changes in the inputs
112
of schools. The chapter presents a strong case for a positive impact of decentralization on the
quality of teachers: all the statistics indicate that the relative teacher quality improved in
public schools after decentralization. This conclusion, however, is sensitive to the comparison
groups. As the last chapter shows, decentralization does not imply an increase in a critical
input of the schools in the most  exible model tested.
The last chapter presents result of the impact of decentralization on student test scores
using a quasi-experimental methodology. In terms of comparing public versus private schools,
and public schools in initially highly dependant localities versus public ones in initially
highly independent localities, the results are clear: there is evidence of a positive impact of
decentralization on the schools that we expect to bene t from decentralization. However, the
impact is negative when a more  exible model is used. This last model allows for nationwide
shocks (like recessions) and other changes across public schools (like changes in curriculums),
making possible the estimation with a more complete description of potential events that
have a ected the education system. A possible explanation of the results is that the quality
of teachers, measured as their level of education, has not increased for public schools in the
initially highly dependent departments.
The dissertation concludes with an empirical approximation to the question of the trade-
o between equity and e ciency of decentralization. The results are mixed as well. In two
of the estimated models, decentralization is income-neutral (same directional e ect across
di erent incomes). However, a surprising result emerges from Model 1: decentralization has
a positive impact for low-income individuals and not on high-income students.
113
Bibliography
[1] Olga Lucia Acosta and Richard Bird. The dilemma of decentralization in colombia.
Working Paper, Fedesarrollo and The Income Mission, 2002.
[2] Ehtisham Ahmad and Katherine Baer. Colombia. In Fiscal Federalism in Theory and
Practice. International Monetary Fund, 1997.
[3] Alberto Alesina, Alberto Carrasquilla, and Juan Jose Echavarria. Decentralization in
colombia. Working Paper, Fedesarrollo, Colombia, Sept. 2001.
[4] Julio C. Alonso, Mauricio Olivera, and Israel Fainboim. Del equilibrio al desequilibrio
 scal en colombia: la experiencia se repite (1980-1996). In Como armar el Rompecabezas
Fiscal? Inter-American Development Bank, 2000.
[5] Katherine Baicker. Government decision-making and the incidence of federal mandates.
Dartmounth College, 2000.
[6] Inter-American Development Bank. Economic and Social Progress in Latin America.
Johns Hopkins University Press, November 1996.
114
[7] Felipe Barrera and Lucas Higuera. The supply of training in colombia. Working Paper,
Fedesarrollo, 2003.
[8] J.R. Behrman and N. Birdsall. The quality of schooling. The American Economic
Review, 73(5), December 1983.
[9] Roland Benabou. Workings of a city: Location, education and production. Quarterly
Journal of Economics, 108(3), August 1993.
[10] Roland Benabou. Equity and e ciency in human capital investment: The local con-
nection. Review of Economic Studies, 63, 1996.
[11] Roland Benabou. Heterogeneity, strati cation, and growth: Macroeconomic implica-
tions of community structure and school  nance. American Economic Review, 86(3),
June 1996.
[12] Richard Bird. Finanzas intergubernamentales en colombia. Departamento Nacional de
Planacion, 1982.
[13] Richard M. Bird and Francois Vaillancourt, editors. Fiscal Decentralization in Devel-
oping Countries. Cambridge; New York: Cambridge University Press, 1998.
[14] Sandra E. Black. Do better schools matter? parental valuation of elementary education.
The Quarterly Journal of Economics, 114(2), 1999.
[15] George Borjas and Olga Lucia Acosta. Education and decentralization in colombia.
Working Paper, Fedesarrollo, Colombia, Sept. 2001.
115
[16] Colin D. Campbell and William A. Fischel. Preferences for school  nance systems:
Voters versus judges. National Tax Journal, 49(1), March 1996.
[17] D. Card and A. Krueger. Labor market e ects of school quality: Theory and evidence.
NBER Working Paper Series, Number 5450, 1996.
[18] D. Card and A. Krueger. Does school quality matter? returns to education and the
characteristics of public schools in the united states. Journal of Political Economy,
100(1), February 1992.
[19] J. S. Coleman, J. Campbell, E. Hobson, C. McPartland, J. Mood, F. Wein eld, and
R. York. Equality of Educational Opportunity. Department of Health, Education, and
Welfare, 1966.
[20] J. Duarte-Agudelo. La debilidad del ministerio de educaci on y la politizaci on de la
educaci on: dos problemas a enfrentar en el plan decenal. Coyuntura Econ omica, 14,
1996.
[21] Juan Jose Echavarria, Carolina Renteria, and Roberto Steiner. Bailout of territorial
entities by the central government of colombia. Working Paper, Fedesarrollo, June
2000.
[22] Dennis Epple and Richard E. Romano. Competition between private and public schools,
vouchers, and peer-group e ects. American Economic Review, 88, March 1998.
116
[23] William Evans, Sheila Murray, and Robert Schwab. The impact of court-mandated
school  nance reform. University of Maryland, Working Paper, June 1998.
[24] William Evans, Sheila Murray, and Robert Schwab. The property tax and education
 nance: Uneasy compromises. University of Maryland, Working Paper, April 2000.
[25] Raquel Fernandez and Richard Rogerson. Public education and income distribution:a
dynamic quantitative evaluation of education- nance reform. American Economic Re-
view, 88(4), September 1998.
[26] William A. Fischel. Did serrano cause proposition 13? National Tax Journal, 42(4),
December 1989.
[27] William A. Fischel. Communication. Journal of Economic Literature, XXX, March
1992.
[28] Willian A. Fischel. School  nance litigation and property tax revolts: How undermining
local control turns voters away from public education. Lincoln Institute of Land Policy,
WP98WF1, 1998.
[29] M. Friedman. Capitalism and Freedom. Chicago: University of Chicago Press, 1962.
[30] Edgar A. Gonzalez. Evaluaci on de la decentralizaci on municipal en colombia: Compo-
nente de capacidad institucional. Archivos de Econom ia, Documento 169, DNP Colom-
bia, 2001.
117
[31] Bruce W. Hamilton. Capitalization of intrajurisdictional di erences in local tax prices.
American Economic Review, 66(5), December 1976.
[32] Eric A. Hanushek. The economics of schooling: Production and e ciency in public
schools. Journal of Economic Literature, 24, September 1986.
[33] Eric A. Hanushek. Measuring investment in education. Journal of Economic Perspec-
tives, 10, Fall 1996.
[34] F.A. Hayek. The use of knowledge in society. The American Economic Review, 35(4),
September 1945.
[35] J. Heckman. Randomization and social policy evaluation. In C. Manski and I. Gar nkel,
editors, Evaluating welfare and training programs. Harvard University Press, Cambridge,
1992.
[36] J. Heckman and R. Robb. Alternative methods for evaluating the impact of interven-
tions: an overview. Journal of Econometrics, 30, 1985.
[37] S. Heyneman and W Loxley. The e ect of primary-school quality on academic achieve-
ment across twenty-nine high- and low-income countries. American Journal of Sociology,
88(6), May 1983.
[38] Caroline Hoxby. The e ects of class size on student achivenment: New evidence from
population variation. Quarterly Journal of Economics, 115, November 2000.
118
[39] Caroline Hoxby. Peer e ects in the classroom: Learning from gender and race variation.
NBER Working Paper No 7867, 2000.
[40] Caroline M. Hoxby. Is there an equity-e ciency trade-o in school  nance? tiebout and
a theory of the local public goods producer. National Bureau of Economic Research
Working Paper: 5265, September 1995.
[41] Caroline M. Hoxby. All school  nance equalizations are not created equal. The Quarterly
Journal of Economics, 116(4), 2001.
[42] Alan B. Krueger. Experimental estimates of educational production functions. Quarterly
Journal of Economics, 113, May 1999.
[43] R. LaLonde. Evaluating the econometric evaluations of training programs with experi-
mental data. American Economic Review, 76(4), 1986.
[44] R. LaLonde, J. Heckman, and J. Smith. The economics and econometrics of active
labor market programs. In O. Ashenfelter and D. Card, editors, Handbook of Labor
Economics. Elsevier Press, 2000.
[45] Susanna Loeb. Estimating the e ects of school  nance reform: a framework for a
federalist system. Journal of Public Economics, 80, 2001.
[46] Eduardo Lora and Felipe Barrera. A decade of structural reform in latin america:
Measurement and growth e ects. Working Paper No 350, Inter-American Development
Bank (www.iadb.org/oce/pdf/350.pdf), 1999.
119
[47] C. Manski. Educational choice (vouchers) and social mobility. Economics of Education
Review, 11, 1992.
[48] Thomas J. Nechyba. Public school  nance in a general equilibrium tiebout world:
Equalization programs, peer e ects and private school vouchers. National Bureau of
Economic Research Working Paper: 5642, June 1996.
[49] Thomas J. Nechyba. School  nance induced migration and strati cation patterns: The
impact of private school vouchers. Journal of Public Economic Theory, 1, January 1999.
[50] Thomas J. Nechyba. Mobility, targeting, and private-school vouchers. American Eco-
nomic Review, 90, March 2000.
[51] Douglas North. Institutions, Institutional Change and Economic Performance. Cam-
bridge, University Press, 1990.
[52] Wallace Oates. An essay on  scal federalism. Journal of Economic Literature, XXXVII,
September 1999.
[53] Wallance E. Oates. Fiscal Federalism. New York: Harcourt Brace Jovanovich, 1972.
[54] Paul Peterson. School vouchers and academic performance: Results from three ran-
domised  eld trials. February 2001.
[55] Fabio Silva and Jon Sonstelie. Did serrano cause a decline in school spending? National
Tax Journal, 48, June 1995.
120
[56] Michael Smart, Juan Gonzalo Zapata, and Juan Camilo Chaparro. Integovernmen-
tal tranfers and municipal  nance in colombia. Working Paper, Fedesarrollo and The
Income Mission, 2002.
[57] Jon C. Sonstelie and Paul R. Portney. Pro t maximizing communities and the theory
of local public expenditure. Journal of Urban Economics, 5(2), April 1978.
[58] Teresa Ter-Minassian, editor. Fiscal Federalism in Theory and Practice. International
Monetary Fund, 1997.
[59] C.M. Tiebout. An Economic Theory of Fiscal Decentralization in Public Finance:
Needs, Sources and Utilization. Princeton; New Jersey: Princeton University, 1961.
[60] El Tiempo. Mensaje de urgencia al congreso de colombia. El Tiempo, November 18
2001.
[61] Carmen Helena Vergara and Mary Simpson. Evaluaci on de la decentralizaci on munic-
ipal en colombia: Estudio general sobre los antecedentes, dise~o, avances y resultados
generales del proceso de descentralizacion territorial en el sector educativo. Archivos de
Econom ia, Documento 168, DNP Colombia, 2001.
121