WPS4836
P olicy R eseaRch W oRking P aPeR 4836
i mPact e valuation s eRies n o . 29
The Use and Misuse of Computers
in Education
Evidence from a Randomized Experiment in Colombia
Felipe Barrera-Osorio
Leigh L. Linden
The World Bank
Human Development Network
Education Team
February 2009
Policy ReseaRch WoRking PaPeR 4836
Abstract
This paper presents the evaluation of the program gender. The main reason for these results seems to be the
Computers for Education. The program aims to failure to incorporate the computers into the educational
integrate computers, donated by the private sector, process. Although the program increased the number of
into the teaching of language in public schools. The computers in the treatment schools and provided training
authors conduct a two-year randomized evaluation of to the teachers on how to use the computers in their
the program using a sample of 97 schools and 5,201 classrooms, surveys of both teachers and students suggest
children. Overall, the program seems to have had little that teachers did not incorporate the computers into their
effect on students' test scores and other outcomes. These curriculum.
results are consistent across grade levels, subjects, and
This paper--a product of the Education Team, Human Development Network--is part of a larger effort the network to
estimate the impact of Information and Communication Technologies on education. Policy Research Working Papers are
also posted on the Web at http://econ.worldbank.org. The author may be contacted at fbarrera@worldbank.org.
The Impact Evaluation Series has been established in recognition of the importance of impact evaluation studies for World Bank operations
and for development in general. The series serves as a vehicle for the dissemination of findings of those studies. Papers in this series are part
of the Bank's Policy Research Working Paper Series. The papers carry the names of the authors and should be cited accordingly. The findings,
interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the
International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of
the World Bank or the governments they represent.
Produced by the Research Support Team
The use and misuse of computers in education:
Evidence from a randomized experiment in Colombia 1
Felipe Barrera-Osorio
(World Bank)
Leigh L. Linden
(Columbia University)
JEL: C93, I21, I28
Keywords: education, computer programs, randomization
1
We are grateful to the program Computers for Education, and the Ministry of Communication, for
financial support. We thank past and current staff members of the program for their unconditional
commitment to the evaluation, particularly María Isabel Mejía Jaramillo, Beatriz Eugenia Córdoba,
Francisco Camargo, Martha Patricia Castellanos and Julián Gómez. Fedesarrollo provided personnel and
financial help. We are especially grateful to Mauricio Olivera. Camilo Dominguez and Monica Hernandez
provided outstanding research assistance. Funding from the Knowledge for Change Program at the World
Bank, research grant RF-P101262-TF090460, helped in the final stages of the evaluation. The opinions
expressed in this document do not necessarily represent the views of the World Bank.
Barrera-Osorio: fbarrera@worldbank.org; Linden: leigh.linden@columbia.edu.
I. Introduction
The use of information and communication technologies (ICT) in education is
becoming a major consideration as developing countries focus on improving the quality
of education. Investment in ICT use in education has grown steadily over the past decade
in developing countries, even in the some of the most challenging environments in some
of the least-developed countries. Several countries are determinedly expanding the supply
of computers in their schools in the belief that schools will benefit from the use of the
new technologies and that students need to be exposed early. For instance, several
countries in Africa, Asia, and Latin America are considering a program to procure $100
laptops for schools. Despite the growing adoption of and demand for ICTs in education,
there is very little systematic research and hard data about how ICT is actually used in the
classroom and even less about its impact on educational outcomes, social behavior, or
employment and worker productivity (InfoDev, 2005).
This article aims to increase the available evidence on the use and the impact of
computers in education. We consider the program Computers for Education. The
program is an alliance between the public and private sector to refurbish computers
donated by private organization, install them in public schools, and run a program that
teach teachers to use computer in specific subjects, especially in Spanish. This is an
existing large-scale national program in Colombia.
Unfortunately, while ICT programs are one of the most studied interventions in
the education literature, robust evaluations of ICT programs are still too scarce to provide
general conclusions regarding their effectiveness. The results of the evaluations that do
exist are at best mixed.
-2-
The evaluation literature of such programs is more abundant in developed
countries. Importantly, a large portion of these studies are also correlational analyses for
which there are obvious challenges to causal interpretation of the findings. Two studies
in the U.S. written by the National Center for Educational Statistics (2001a and 2001b)
found a positive relationship between availability of computers in schools and test scores.
For the US, Wenglinsky (1998) measured the amount of computers that were used in
math classes and scores on math tests and found a positive relationship between use of
computers and learning in both 4th and 8th grades. Rouse and Krueger (2004) undertook a
randomized study of a popular instructional computer program, known as Fast ForWord,
which is designed to improve language and reading skills. They concluded that while the
program may have improved some aspects of students' language skills, the gains do not
translate into a broader measure of language acquisition or into actual readings skills.
Similar positive relationships have been found in OECD countries between
computer use and test scores for mathematics (NCES 2001a, Cox et al., 2003), science
(NCES, 2001b, Harrison et. al. 2002) and reading (Harrison et. al. 2002). Kulik (2003)
reviews 75 impact evaluations of technology applications in the US, finding the
following results, among others: (i) students who used computer tutorials in mathematics,
natural science, and social science score significantly higher on tests in these subjects; (ii)
the use of computer-based laboratories did not result in higher scores; and (iii) primary
school students who used tutorial software tutorial in reading scored significantly higher
on reading scores, while very young students who used computers to write their own
stories scored significantly higher on measures of reading skills.
-3-
In developing countries, six existing studies present generally positive but still
mixed conclusions. Linden et. al. (2003) designed an impact evaluation of a computer
assisted learning program in Vadodara, India, on cognitive skills using mathematics and
language tests. The authors find a positive and significant impact on math scores of 0.375
standard deviations. Similarly, Linden (2008) finds positive effects of a remedial math
program when implemented on a supplemental basis and negative effects when
implemented on a pull-out basis as a substitute for the regular classroom teacher's
instruction. Fang, He, and Linden (2008) find strong positive effects on Indian students'
English scores of an electronic English-based curriculum.
However, other evaluations do not find such consistent positive results. Angrist
and Lavy (2002) find no effect in their evaluation of the `Tomorrow-98' program, which
placed 35,000 computers in schools across Israel between 1994 and 1996. They find no
impact on math and Hebrew scores at the fourth or eighth grade level. Finally, the
evaluation of the World Links program found positive impact for both students and
teachers (Kozma, et. al, 2004, Kozma and McGhee, 1999). This program prepares
students and teachers on communication, collaboration and Internet skills in African and
Latin American countries. In Uganda, a special designed performance assessment found
that World Links schools outperformed the non-World Links schools on measures of
communication and reasoning with communication (Quellmalz and Zalles, 2000).
However, unlike the other evaluations, these are based on correlational estimates rather
than rigorous research designs.
While there is still much to be learned, one general result that seems to emerge
from this literature is that positive outcomes of the use of computers in schooling are
-4-
linked to changes in pedagogy, and introducing technology alone will not change the
teaching and learning process. It is not enough to install computers in schools without
training (InfoDev 2005). However despite this general result, very little is known about
just how (and how often) ICTs are used in developing country classrooms when
available. One study has shown, for example, that where ICTs are used for learning, they
are chiefly used to present and disseminate information rather than change the way that
children are taught (InfoDev, 2005).
Our study builds on this existing literature in several ways. First, because our
study is a large scale randomized evaluation of a mature, well developed program, we
add to the existing body of rigorous randomized evaluations of ICT programs in
education. Second, within this evaluation, we implemented modules that are designed to
understand, not just the effect of computers on students' test scores, but also the effect of
the computers and associated training on the teachers teaching methods, including their
use of computers in the classroom.
There are three main conclusions of this evaluation. First, the program
successfully increases the number of computers in the school (by 15 computers) and
increases students' use of the computers. Second, despite this success, the program has
little impact on students' math and Spanish test scores. The program also has little effect
on a host of other academic variables including hours of study, perceptions of school, and
relationships with their peers. The reason seems to be that despite the program's focus on
using the computers for teaching students in a range of subjects (but especially Spanish),
the computers were only used to teach the students computer usage skills. The evidence
suggests that students use of the computers for their intended purpose was limited -- only
-5-
3 to 4 percent of the students in both treatment and control groups reported to use the
computers in the language class for example. Overall the results of this study highlight
the importance of program implementation and measuring the impact of an intervention
on the actual practice of teachers and the learning experiences of students.
The paper is organized as follows. In the next section we describe the program in
more detail. In section three, we discuss the design of the experiment. Section four
contains the results that verify the internal validity of the study and section five contains
the results of the evaluation. Finally, we conclude in section six.
II. The Computadores para Educar Program
The Computadores para Educar (Computer for Education or CPE) was created in
March 2002 by the Minister of Communications, with the objective of refurbishing
computers donated by the private sector to install them in public schools. The program
trains teachers in the use of computers in the classroom, especially in the areas of
language. Since its creation, the program has received 114,541 computers, and
refurbished 73,665 that have been installed in 6,386 public schools in 1,018
municipalities. To date, the program includes 83,092 teachers and more than 2 million
students.
The program creates a partnership between schools and a local university. For
our study, schools were paired with the Universidad de Antioquia. The university then
designs a pedagogical strategy for the school and participates in the implementation of a
20 months training component directly in the schools for teachers. The start of the
training coincides with the school's receipt of a set of refurbished computers. The initial
-6-
phase of the training is provided by CPE program directly and lasts for 8 months and
covers computers installation and adaptation of classroom techniques, which includes
preparation of the rooms where the computers are installed and classroom management
strategies. This phase also includes a first step towards the active use of computers in
education through teachers' workshops.
The second phase lasts one year and is developed by the partner university to take
into account the regional needs of the schools. The objective of this second phase is to
train teachers and coordinators on the relationship between technology and learning.
Among the objectives are: (i) support the education of the children in basic areas
(language, math, natural and social sciences) by integrating the use of ITCs with
pedagogic projects and activities, and (ii) encourage collaborative learning, creativity,
and improve teachers' and students' confidence in the use of technology by integrating
ITCs to their pedagogic processes (CPE, 2008).
The model designed and run by the Universidad de Antioquia focuses on Spanish
education. The program is aimed at training teachers in teaching methodologies using
computers to strengthen students' reading and writing skills through a theoretical socio-
constructivist approach. Its objectives go beyond the impact on standardized tests. In
particular, the program tries to integrate technology in learning pursuing the goal of
"fomenting a socio-cultural vision over the reading-writing teaching and learning
processes" (translated from Henao and Ramirez, no date available, p. 2). Specifically, the
model focuses on different aspects of reading and writing, with a special focus on the
recent developments of e-mail, Internet, and the Hypermedia formats.
-7-
III. The Design of the Experiment
A. Sample
In order to evaluate the program, we randomly assigned 100 interested and
eligible schools into a treatment and a control group. Interested schools were selected as
follows. First, to minimize program implementation and data collection costs, the
schools were chosen to be in close geographic proximity: we chose schools in the north
part of Antioquia, Caldas, Choco, Cordoba, Quindio y Risaralda. For all interested
schools in this area, we further reduced the sample to schools to those with 80 or more
students in order to facilitate the collection of data. From this list of 100 schools a CPE
team visited each school to verify the number of students, the existence of a classroom
that could be refurbish for the computers, and the type of school (public or private).
Once the final sample was created, we conducted a stratified randomization,
stratifying on department and type of school--basic education, basic plus lower
secondary, high secondary. 2 In this process half of the schools received the program and
half were assigned to a control group which did not receive the program. The lottery was
performed at the beginning of August 2006, and the list of schools, with their status, was
given to the Ministry of Communication for the implementation of the program.
B. Data Collection
Data was collected in two phases: a baseline survey conducted immediately after
the randomization but before the start of the treatment and a follow-up survey conducted
two years later. The baseline was conducted between August 14 and September 29,
2
Type refers to the grades covered by each school. Basic includes grades one through five. Lower
secondary includes grades six through eight, and high secondary includes grades nine through eleven.
-8-
2006, with the target of reaching 100 schools. The collection of information was done
directly at the school in an unannounced visit, and questionnaires were administered to
students in grades three through eleven, their math and Spanish teachers, and the school's
principal.
The survey company, SEI (Sistemas Especializados de Informacion), was able to
gather the information from 97 schools: three schools (two treatment and one control)
declined to participate in the data collection activities. Because the follow-up survey was
completed two years later, we focus on students who were in grades through nine at
baseline students in grades ten and eleven at baseline would have either dropped out or
graduated by the time the follow-up survey was administered. The first three columns of
Table 1 contain a tabulation of these schools by research group. In total, the distribution
of schools and students is even. The sample contains forty-nine control schools and
forty-eight treatment schools, and 3,889 control students and 4,327 treatment students in
grades three to nine. Dividing the sample by gender and grade shows a similarly even
distribution of students. Across all of the groups, the largest difference is the 182
additional male students in the treatment group.
During May and June of 2008, the same survey company resurveyed the same 97
schools, focusing on the students who were in grades three through nine in 2006. The
final sample comprises 5,201 students that were found at follow up; 3,015 students were
attritors, yielding an attrition rate of 37 percent. Discussions with local principals suggest
that the high rate of attrition is primarily due to high rates of migration in the rural areas
chosen for the research project. A high attrition rate alone does not violate the internal
validity of a study as long as the types of students who attrit are similar in each research
-9-
groups (Angrist et al., 2002). While we investigate this in more detail in section five
below (and find the attrition rates similar), the distribution of non-attriting students is
presented in the last three columns of Table 1. Like the original baseline tabulation, the
distribution of students is still evenly distributed between the research groups even at
follow-up, despite the high attrition rate.
Both the baseline and follow-up questionnaires followed the same format. Three
different questioners were administered respectively to students, math and Spanish
teachers and the principal of the school. The students' questionnaire was self-
administered by the students with the assistance of the survey team. The variables
included in the students questionnaire were socioeconomic characteristics (age, gender,
family structure, number of siblings, work status, and allowance, among others,); school
outcomes (attendance last week, number of hours of study outside the school, grades in
math and language, repetition, and dropout spells in the past, among others); and attitude
towards the school and the content of classes, including the use of computers. Finally,
the student survey included a shortened version of the national Colombian exam, the
Saber. 3 For the purpose of analysis, the scores on these tests are normalized relative to
the grade-specific control group distribution of scores.
Information was also collected from all math and language teachers present at the
school the day of the visit. We collected background information (age, gender, education,
experience, etc.), information on their knowledge of computers, and their use of
3
Four different tests were used, depending on the student's grade, grouping them as follows depending on
their grade at the time of the survey: 1) third and fourth, 2) fifth and sixth, 3) seventh and eighth, and 4)
ninth, tenth and eleventh. These are national Colombian exams normally administered to students in the
odd grades from grade three to grade nine. As a result, we simply administer the exam normally given to a
particular grade to that grade and the subsequent grade with the exception of grade eleven which receives
the grade nine test (since there is no grade eleven test).
- 10 -
computers in class. Finally, the questionnaire of the school was answered by the principal
or the coordinator of the school. It includes general variables at the school level.
C. Statistical Models
Because the treatment was randomly assigned, we can assess the causal effects of
the program by directly comparing the average outcomes of students in the treatment
group with students in the control group. To do this, we employ three statistical models
estimated through ordinary least squares. First, we use a simple difference model of the
following form:
Yij = 0 + 1T j + ij (1)
The variable Yij in this specification is the variable whose average value is to be
compared between the two groups at the child (i), school (j) level. The variable T j is a
dummy variable for whether or not school j was selected for treatment in the
randomization process. The coefficient 1 then provides the estimated differences in the
variable Yij between the treatment and control group. This specification is primarily used
for two purposes: to compare the treatment and control group using information collected
in the baseline survey to check for comparability and to conduct simple comparisons of
outcome variables (without controlling for baseline characteristics).
Equation (1) can also be used to estimate the treatment effects, but we can
improve the precision with which we estimate the treatment effects by controlling for
- 11 -
baseline characteristics that are correlated with the outcome of interest. This is done with
the following specification which is also estimated through ordinary least squares:
Yij = 0 + 1T j + 2 X ij + ij (2)
The specification is identical to equation (1) with the addition of a vector, X ij , of baseline
and school characteristics.
In Table 3, we compare the relative characteristics for students that attrit from the
baseline at follow-up and students that do not. To make this comparison we use a
difference in differences estimator that compares the difference between attritors and
non-attritors in the treatment group to the same difference in the control group. The
model is estimated through ordinary least squares using the following specification:
Yij = 0 + 1T j + 2 Attrit ij + 3 Attrit ij * T j + ij (3)
The variables Yij , T j , and ij are all defined as in equation (1) and Attrit ij is a dummy
variable set to 1 if student i in school j did not appears in the follow-up. The coefficient
3 in this model is then an estimate of the difference in characteristics between attritors
and non-attritors across the treatment and control groups.
Finally, because treatment assignment occurs at the school level, student behavior
is likely correlated within schools. Not taking this correlation into account, could lead us
to overestimate the precision of the treatment effect estimates and conclude that an effect
- 12 -
exists when, in fact, it does not. To account for this in both models, we cluster the
standard errors ij at the school level (Bertrand, Duflo, and Mullainathan, 2004).
IV. Internal Validity Confirmation
In order to ascribe any observed difference in follow-up characteristics to the
implemented program, the two groups created by the randomization must be sufficiently
similar in all characteristics that such differences could not be responsible for generating
the observed differences in the groups at follow-up. We proceed as follows. First, using
all of the information students provided in the baseline, we verify that the randomization
process did create initially similar treatment and control groups. Some of those students
observed in the baseline, however, then fail to provide information at follow-up, having
attritted form the sample. Because these students are not available at follow-up, the
attrition process could cause an initially balanced sample to become unbalanced if
different types of students attrited from one of the two groups. To check this, we
compare the characteristics of attriting students between the two groups and compare
non-attriting students using the available baseline characteristics.
A. Baseline Comparison
To start, we compare the respective research groups based on the survey
responses at baseline. Table 2 contains the comparison of all students taking a baseline
survey in treatment and control schools. Panel A contains the test score characteristics
and the rows in Panel B contain basic demographic characteristics. Panel C then contains
other relevant academic variables. We present the average of these characteristics for the
- 13 -
treatment group in the first column. Column two contains the control group average.
And the final column contains the estimated difference using equation (1).
The differences between the treatment and control groups are minor. Of nineteen
considered variables three are statistically significant and all of them are practically
small. This includes a three percent difference in the probability that students report
staying after school, an eight percent difference in the probability that a child talks to a
teacher after class, and a 0.11 difference in the number of hours reported studying. The
most significant variables the math and Spanish test scores show no difference
between treatment and control groups. The normalized difference in Spanish scores is
0.07 standard deviations and the difference in math scores is 0.06 standard deviations. 4
B. Attrition
Even if the treatment and control groups are similar at baseline changes in the
composition of the groups between baseline and follow-up could result in large
differences between the treatment and control students who complete a follow-up survey.
And as long as these attrition patterns are similar across treatment and control groups, the
research design would still have internal validity even with extremely high attrition rates
(Angrist et al., 2002). We compare the attrition pattern between the two group using
baseline characteristics in Table 3 and 4. Table 3 contains a direct comparison of attriting
students. Using the same rows as in Table 2, columns one through three contains a
simple comparison of attriting students using equation (1). Columns four and five
4
A more detailed comparison is provided in a baseline report produced after the baseline data collection
(Barrera-Osorio et al, 2006). The report compares the samples in much greater detail than is presented here
concluding the that the samples are sufficiently similar. This report is available from the authors upon
request.
- 14 -
contain the relative difference between attriting and non-attriting students in the treatment
and control groups respectively. Column six then contains the estimated difference in
these differences between treatment and control groups using equation (3). Thus, the first
three columns compare the characteristics of attriting students while the last three
columns compare the relative characteristics of attriting and non-attriting students across
the two research groups.
The results in Table 3 confirm that the same types of children attrited from the
treatment and control groups. Even though on average 35 and 38 percent of baseline
students attrited from the treatment and control groups respectively, the types of attriting
students are exactly the same. First, the attrition rates are very similar, differing by only
3 percentage points. Second, in almost all characteristics, the differences between the
treatment and control groups are minor. The attritors differ in only two characteristics
whether students talk to a teacher outside of class (7.5 percentage points) and the number
of hours students study outside of school (0.13 hours) both of which are small.
Columns four through six then compare the characteristics of attriting and non-
attriting students. In both the treatment and control groups, attriting students are
significantly different from non-attritors. Attritors are about three-quarters of a year
older than non-attritors. They are 11 percentage points more likely to have repeated a
previous grade, and 9 to 11 percentage points more likely to have failed a class. They are
also more likely to work. The patterns, however, are exactly the same in the treatment
and control groups. As shown in column six, the difference in these patterns is only
significant for the number of friends.
- 15 -
Table 4 estimates the ultimate outcome of the attrition process by comparing the
non-attriting students using the baseline characteristics. The table layout is identical to
Table 2. As one would expect from the previous tables, the students remaining in the
study after the follow-up survey are similar in their baseline characteristics. In fact, the
attrition pattern was so similar, that the results in column three of Table 4 are almost
exactly the same as those in column three of Table 2 despite having lost almost a third of
the sample. Focusing on the differences in column three, all of the differences are small,
and only two differences are statistically significant. This includes an 8.5 percentage
point difference in the probability that a child reports seeing a teacher outside of class and
a 3.4 percentage point difference in the probability that a child stays after school.
V. Results
Given that the sample of students completing the survey in both research groups
is comparable, we can estimate the causal effects of the program by directly comparing
the average responses in the treatment and control groups. First, we check to ensure that
the treatment was implemented as planned and that the implementation created the
expected treatment differential between the two groups it did. Second, we estimate the
differences in outcomes. Overall, the program seems to have had little effect on students'
test scores and other outcomes. These results are consistent across grades, subjects, and
different types of students. The main reason for these results may be the implementation
of the program. Surveys of both teachers and students suggest that the program increases
computer use among students and teachers by a surprising small amount, and most of the
use of computers by students is for the purposes of learning to use a computer rather than
- 16 -
studying language. Additionally, the extra computer use reported by teachers is
concentrated in the lower grades with older students' teachers reporting almost no
computer use in both groups.
A. Treatment Differential
As described in section four, the two research groups contained similar students at
the time of the follow-up exam. The research design then depends on the only difference
being the treatment delivered to the schools in the treatment group. Table 5 estimates the
differences in the implementation of the treatment at the school level by research group
using equation (1). As before, columns one and two contain the average characteristics
for the treatment and control groups respectively and we present the estimated difference
in the third column. We use three measures of implementation to compare the schools.
First, using information from CPE staff regarding whether or not the program was
implemented in a specified school, we estimate the average number of schools receiving
the treatment in the first row. The estimates document that the treatment assignment was
very closely followed by CPE. On average 96 percent of the treatment schools were
treated and only 4 percent of the control schools were treated. 5
We also checked for the implementation of the two main components of the CPE
program by CPE and the Universidad de Antioquia the provision of computers and the
training of teachers. Using data from the principals, we estimate in row two the average
number of computers in each school. On average, treatment schools have 13.4 computers
compared to only 5.1 computers in the control schools. This suggests that the program
5
Despite this very high compliance rate, we also estimated all of the estimated differences in outcomes
using a treatment on the treated model to account for the existing non-compliance. As expected, the
estimates are virtually identical to those presented in Tables 6 through 13.
- 17 -
was successful at increasing the computer resources available to students in the treatment
schools. Finally, in row three, we report the average responses of the surveyed teachers
regarding whether or not they received training from the university. In all, 95 percent of
treatment teachers reported receiving training while only 8 percent of the control group
reported receiving training. These results document that the program implementation
complied with the research design and generated the expected treatment differential
between the two research groups.
B. Effects on Test Scores
Given that the groups are on average similar in characteristics except for the
receipt of the treatment by the treatment group, we can estimate the effects of the
treatment by comparing the average characteristics of the research groups at follow-up.
Tables 6 through 9 present results of the impact of the program on test scores.
Table 6 contains the primary outcome measure the average score on the test
administered as part of the follow-up survey. The first four columns of Table 5 presents
percentage of correct answers, and the second four estimates the same treatment effects
using the test scores normalized by the grade-specific control distribution. Within each
group, we first present the average score for the treatment group, the average score for
the control group, and then the simple difference estimated using equation (1). Finally, in
the fourth column we estimate the differences controlling for baseline characteristics
using equation (2).
It is important to note that the estimated respective differences between the
estimates from equation (1) in column three and seven (without controls) and the
- 18 -
estimates from equation (2) in columns four and eight (with controls) are virtually
identical. This underscores the similarity of non-attriting sample. Had there been large
differences in characteristics that were correlated with the follow-up test scores, then the
estimated differences would have differed significantly. The fact that the estimates are
similar is another piece of evidence supporting the comparability of the research groups
at follow-up.
The value of the estimated differences shows that, by all measures, the program
seems to have had little effect on students' test scores. In Spanish, the subject targeted by
the program, students answer on average about 40 percent of the questions correctly, but
the treatment group only answers correctly 1.7 percent more than the control group. This
difference is too small to be statistically significant at conventional level of significance.
Similarly, the program has no ancillary effect on math (neither positive nor negative).
To facilitate comparing the magnitude of these results to those of other studies,
the second four columns make the same comparison using normalized scores. The
average treatment effect in both subjects is less than 0.1 standard deviations. This is
significantly smaller than the estimated effects of other successful programs in the
literature. In addition, the estimated standard errors suggest that the experiment was
powerful enough to detect treatment effects of at least 0.125 standard deviations (at the
ten percent level) which is comparable to the smallest estimated treatment effects for
education programs in developing countries (Kremer, Miguel, and Thornton, 2007). In
addition, we can reject the hypothesis (at the ten percent level) that the true effect of the
program in Spanish is greater than 0.2 standard deviations. Since many viable classroom
interventions have increased test scores by 0.3 to 0.47 standard deviations (for example,
- 19 -
Muralidharan and Sundararaman, 2008; Banerjee, Cole, Duflo, and Linden, 2007; He,
MacLeod, and Linden, 2008; Linden 2008), the test is powerful enough to exclude
treatment effects that would place the program in the range of other programs that
educators would consider when choosing interventions.
Despite the lack of overall average effects, it is possible that the program may
have had significant impacts on individual subjects (as in Glewwe, Kremer, and Moulin,
2007). Table 7 presents the estimated treatment effects by specific components. This
includes encyclopedia, identification, paraphrase, pragmatics and grammar in Spanish 6
and of algebra, arithmetic, geometry and statistics in mathematics. As in Table 5, we
present the average treatment score, the average control score, the difference estimated
with equation (1), and the difference estimated with equation (2). All scores are
normalized to the control group distribution. The results are similar to the overall
average estimates presented in Table 5. For all competencies, we did not detect any
differences across sub-subjects between students in treatment and control schools. The
largest estimated difference is 0.13 for letter identification, a difference that is not
statistically significant at conventional levels.
Table 8 presents differences on the test by gender. Replicating the format of
Table 6, we present the treatment effect measures in normalized test scores. Panel A
contains the results for female students and Panel B contains the results for male students.
For no subject or group of students are the results different than the overall averages
6
The encyclopedia component requires blending previous knowledge and information presented in a
sample text to answer questions. The identification component asks respondents to repeat information
explicitly presented in the text without changing its meaning. The paraphrase questions require recognizing
both explicit and implicit information in the text and being able to convey the information back in written
form. The pragmatics questions ask students to recognize and understand different types of communicative
actions, intentions, aims and purposes of the author and the circumstances under which the text is written.
The grammar component asks students to recognize and understand the semantic function of grammatical
elements in coherence and cohesion of the text (MEN, 2008).
- 20 -
presented in Table 6. For Spanish, girls show a difference in test scores of 0.069 standard
deviations while boys show a difference of 0.087 standard deviations. Neither of these
estimates is statistically significant at conventional levels of significance.
Finally, Table 9 divides the sample by grade. Using the reported grade levels
from the 2006 baseline survey, we report only the estimated treatment effects using
equation (2) for the test scores normalized to the follow-up control distribution for the
indicated subject. The results generally confirm the estimates presented earlier. In
almost all grades, the estimated differences generally small and statistically insignificant
at conventional levels of significance. The exception is students in grade eight and nine.
Students in grade eight seem to perform worse due to the program while students in grade
nine seem to benefit from the program. However, since these are the only statistically
significant results, they are likely due to random variation. The fact that, as we will show
below in Table 12, computers were not used in these grades further supports this
contention.
Table 10 shows the results for outcome variables other than test scores. The table
is organized similarly to Tables 6 and 8 with the differences estimated using equation (1).
For most variables, the program seems to have little effect. There is no change in the
probability that students like school or like the content of school. Similarly, there is no
change in students reported grades or probability of failing a subject. And similarly,
there is no change in students' probability of talking to a teacher outside of class. The
only statistically significant effects are the probability that students did not attend school
in the previous week (reduction of 0.12 percentage points), the probability that a student
- 21 -
stayed after school (reduction of 2.8 percentage points), and the reported number of hours
worked (reduction of 1.735 hours). However, while significant, they are small.
C. Number and Use of Computers
To better understand the lack of estimated treatment effects on student outcomes,
we turn to our measure of computer utilization by teachers and students. A critical
objective of the CPE program is to incorporate computers into actual classroom teaching
at the school. Indeed, the program aims for the use of computers in Spanish classes. The
lack of consistent treatment effects especially on Spanish test scores seems to be at
least partly the results of teachers' and schools' limited use of the program. Teachers
report only a modest increase in computer use an increase that is concentrated in the
lower grades. And students report that their experience of computers due to the treatment
while also modest occurred in computer science and not in Spanish as originally
envisioned by the creators of the program.
Table 11 presents the responses of 426 teachers to our follow-up questionnaire.
As before, we show the average response for the treatment group, the average response
for the control group, and the estimated difference using equation (1). The responses are
disaggregated by the subject that the teacher teaches. The surveys targeted language and
math teachers, and if a teacher taught both subjects, then the teacher's responses are
included in both the language and math teachers' responses.
The data show that the program created a modest increase in computer use
amongst teachers. In response to the question about whether or not they used a computer
last week, 42 percent of the treatment language teachers responded affirmatively
- 22 -
compared to 17 percent of the control group. This is a difference of 25 percentage points
which is statistically significant at the one percent level. The fact that some schools in
the control group use computers without the program is hardly surprising, but it does
seem surprising that only 44 percent of teachers in the treatment group report using a
computer. Interestingly the results were the same for math teachers, suggesting that the
computers provided by the program were not used exclusively for the Spanish program,
but rather, were used by teachers in general.
The next set of questions attempt to identify what the teachers used the computers
to do. For each topic, teachers were asked how many times they used a computer for the
identified task in the previous week. Again, teachers responded that they used the
computers in the treatment group (for both language and math) about a half a day more a
week than the control group, differences that are both statistically significant at the five
percent level. However, when asked how many times the computers were used for in
class activities, the teachers responses indicate that the computers were not used more
often in the treatment group for these purposes. Teachers, for example, may have been
using the computer for preparatory activities rather than for in class activities.
Using the information on which teachers did and did not use computers, we can in
turn ask for each school which grades had a teacher that reported using a computer in the
previous week. These results are presented in Table 12. The format is the same as for
Table 10, except that the analysis is conducted at the school level, difference are again
estimated using equation (1). The results suggest that for both language and math, the
program generated changes in schools' use of computers only in the early grades. In
Spanish, for example, the program seems to increase computer in only 20 and 16 percent
- 23 -
of the schools respectively in grades three and four, both differences are statistically
significant at the five percent level. However, in higher grades, the differences are much
smaller. There is a difference of 10 percentage points in the fifth grade that is not
statistically significant at conventional levels, and then the difference for all other grades
is statistically insignificant and less than 10 percent in magnitude.
We also included questions for students regarding computer use, and their results
corroborate the data from the teachers. This data are presented in Table 13. In the first
three columns we present the average responses for the treatment group, the control
group, and the difference (estimated using equation (1)) for all 5,201 students responding
to both the baseline and follow-up surveys. In the second three columns, we present the
same three statistics, but only for students who report having used a computer in the last
week. Thus the first three columns show the overall average rates of use for each
individual purpose across all students, and the second column identifies differences in
computer use between the treatment and control students for students using a computer.
Similar to the responses for the teachers, about 25 percent more students report
using a computer in the last week, an increase to 66 percent for the treatment group
versus 41 percent for the control group. Focusing on where students used a computer, the
increase in the treatment group seems to be largely due to increased computer use at
school (a 30 percent increase) which is largely consistent with the idea that the use is
generated by the provision of computers through the CPE program. There is no
difference in this respect between students in general and only students using computers.
Given that students use the computers more often at school, which subjects do
they use them for? Panel B presents estimated differences in the subjects in which
- 24 -
students' report using computers. It seems that the largest (and only statistically or
practically significant change) is in the use of computers for computer science class. The
magnitude of this result, 29 percent, is consistent with the overall reported use of
computers for both groups of students. This is also consistent with the data from teachers
suggesting that they did not report an increase in the use of computers in class. Both
confirm that despite the intent of the program to increase computer use in the teaching of
Spanish, the program did not have that effect.
VI. Conclusion
The results of this evaluation provide a sobering example of the potential limits of
ICT interventions aimed at improving the methods that teachers use in the classroom.
Our estimates suggest that this widely implemented national program has no effect on
students' academic performance. The primary reason seems to be a failure of the
implementation of the program. Despite receiving computers, training, and technical
assistance, the teachers in the program simply failed to incorporate the new technology
into their classroom teaching.
This example provides an important lesson both for researchers and for policy
makers. For policy makers it emphasizes the importance of program implementation and
monitoring. In this case, the program simply assumed that once equipped and trained,
teachers would voluntarily incorporate the provided technology into their classrooms.
Mere training and equipment does not seem to be sufficient.
For researchers, it suggests two important research questions. First, the obvious
question is once trained and equipped, how can we get teachers to use the resources that
- 25 -
they are given? There are examples of teacher using new techniques and technologies
(Duflo, Hanna, Ryan, 2007; He, MacLeod, and Linden, 2008), but why these efforts were
successful at changing teacher behavior and other were not remains an open question.
Second, while all too often process evaluations are incorrectly interpreted as measures of
impact (Duflo, 2004), process evaluations remain an important component of impact
evaluations. Particularly in a literature with results that are as mixed as those that
estimate the effects of ICT on education, it is important to be able to distinguish between
programs that have little effect because of implementation issues and those that fail
because of poor pedagogical design.
- 26 -
References
Angrist, J. and Lavy, V. (2002) "New Evidence on Classroom Computers and Pupil
Learning". Economic Journal, 112, pp. 735-765.
Banerjee, A., Cole, S., Duflo, E. and Linden, L (2007) "Remedying Education: Evidence
from Two Randomized Experiments in India," Quarterly Journal of Economics
122(3), pp. 1235-1264.
Barrera-Osorio, F., C. Dominguez, L. Linden, and M. Olivera. 2006. "Computers for
Education: Evidence from a Randomized Experiment. Baseline report", processed,
Fedesarrollo, Bogota, Colombia.
Bertrand, M., E. Duflow, and S. Mullainathan (2004) "How Much Should We Trust
Differences-in-Differences Estimates?" Quarterly Journal of Economics 119 pp 249-
275.
Computadores para Educar (2008) www.computadoresparaeducar.gov.co. Accessed
December 9, 2008.
Cox, M., Abbott, C., Webb, M., Blakely, B., Beauchamp, T. and Rhodes, V. (2003) ICT
and Attainment--A Review of the Research Literature, Coventry: Becta (ICT in
Schools Research and Evaluation Series)
Duflo, E. (2004). "Scaling Up and Evaluation," Annual World Bank Conference on
Development Economics 2004. pp 341-369.
Duflo, E., Hanna, R. and Ryan, S. (2007) "Monitoring Works: Getting Teachers to Come to
School," NBER Working Paper No. 11880, 2005; BREAD Working Paper No. 103,
November 2007.
- 27 -
Glewwe, P., Kremer, M. and Moulin, S. (2007) "Many Children Left Behind? Textbooks
and Test Scores in Kenya," NBER Working Papers 13300, National Bureau of
Economic Research, Inc.
Harrison, C., C. Comber, T. Fisher, K. Haw, C. Lewin, E. Lunzer, A. McFarlane, D.
Mavers, P. Scrimshaw, B. Somekh, and R. Watling, (2002) "ImpCT2: The Impact of
Information and Communication Technilogies on Pupil Learning and Attainment,"
mimeo Department for Education and Skills (DfES)/Becta, London.
He, F., Linden, L. and MacLeod, M. (2008) "How to Teach English in India: Testing the
Relative Productivity of Instruction Methods within the Pratham English Language
Education Program," mimeo Columbia University.
Henao, O. and Ramirez, D. (date not available) "Impacto de una propuesta de
capacitación en didáctica de la lecto-escritura con un enfoque socio-constructivista y
el uso de TICs, en las concepciones teóricas y metodológicas de los docentes,"
mimeo Universidad de Antioquia, Medellin, Colombia.
InfoDev (2005), "Knowledge Maps: ICTs in Education", November, www.infoDev.org.
Kozma, R. and McGhee, R. (1999). World Links for Development: Accomplishments and
challenges. Monitoring and evaluation annual report: 1998-1999. Menlo Park, CA:
SRI International.
Kozma, R., McGhee, R., Quellmalz, E. and Zalles, D. (2004), "Closing the digital divide:
Evaluation of the World Links program". International Journal of Educational
Development, 24(4), pp. 361-381.
Kremer, M., Miguel, E. and Thornton, R. (2007). "Incentives to Learn," mimeo MIT
University.
- 28 -
Kulik, J. (2003), "Effects of Using Instructional Technology in Elementary and
Secondary Schools: What Controlled Evaluation Studies Say", SRI Inernational,
Arlington, VA, May.
Linden, L., Banerjee, A. and Duflo, E. (2003), "Computer-Assisted Learning: Evidence
from a Randomized Experiment", Poverty Action Lab Paper No. 5, October.
Linden, L. (2008) "Complement or Substitute? The Effect of Technology on Student
Achievement in India," mimeo Columbia University.
Ministerio de Educación Nacional de Colombia (MEN) (2008).
http://menweb.mineducacion.gov.co/saber/componentes.php. Accessed December 10,
2008.
Muralidharan, K. and Sundararaman, V. (2008) "Teacher Performance Pay: Experiment
Evidence from India," mimeo University of California at San Diego.
National Center for Educational Statistics (NCES) (2001a) "The Nation's Report Card:
Mathematics 2000", Washington DC.
National Center for Educational Statistics (NCES) (2001b) "The Nation's Report Card:
Science 2000", Washington DC.
Quellmalz, E. and D. Zalles (2000) World Links for Development: Student assessment
Uganda field test. Menlo Park, CA: SRI International.
Rouse and Krueger (2004) "Putting computerized instruction to the test: a randomized
evaluation of a `scientifically based' reading program" Journal Economics of
Education Review, 23(4), pp. 323-338.
Wenglinsky, H. (1998) "Does it compute? The relationship between educational
technology and student achievement in mathematics", Princeton, NJ: ETS.
- 29 -
Table 1: Distrubtion of Surveyed Schools
Baseline Survey FollowUp Survey
Control Treatment Total Control Treatment Total
Schools 49 48 97 49 48 97
Students 3889 4327 8216 2403 2798 5201
Students by Gender
Male 1851 2133 3984 1146 1410 2556
Female 2038 2194 4232 1257 1388 2645
Students by Grade
Third 1230 1209 2439 764 804 1568
Fourth 375 365 740 242 251 493
Fifth 370 441 811 235 279 514
Sixth 767 833 1600 455 536 991
Seventh 576 689 1265 345 427 772
Eighth 321 469 790 198 276 474
Ninth 250 321 571 164 225 389
Note: This table contains a tabulation of the schools and students that were surveyed in the baseline survey and then both the
baseline and followup surveys. One hundred schools were originally randomized, but three schools (2 treatment and 1 control)
refused to participate in data collection.
- 30 -
Table 2: Average Characteristics of Students Completing Baseline Survey
Treatment Control Difference
Characteristics Average Average
Panel A: Test Scores
Language score 0.06 0.01 0.07
(0.09) (0.08) (0.12)
Math score 0.04 0.02 0.06
(0.08) (0.08) (0.11)
Total score 0.07 0.02 0.08
(0.10) (0.09) (0.13)
Panel B: Demographic Characteristics
Gender 0.51 0.52 0.02
(0.01) (0.02) (0.02)
Age 12.05 11.85 0.20
(0.26) (0.35) (0.43)
N. parents in the household 1.55 1.59 0.04
(0.02) (0.02) (0.03)
N. siblings 3.77 4.03 0.27
(0.20) (0.18) (0.27)
Receives allowance 0.76 0.72 0.04
(0.02) (0.03) (0.03)
N. friends 17.88 15.52 2.36
(1.79) (1.16) (2.12)
Hours of work 6.50 7.58 1.08
(0.37) (0.76) (0.84)
Panel C: Academic Variables
Transport time to school 2.43 2.41 0.02
(0.05) (0.07) (0.08)
Attended school last year 0.97 0.97 0.00
(0.00) (0.01) (0.01)
Repeated a grade in the past 0.35 0.36 0.01
(0.02) (0.02) (0.02)
N. days absent this year 2.07 1.89 0.18
(0.10) (0.13) (0.16)
Failing subject 0.37 0.35 0.03
(0.02) (0.02) (0.03)
Stays at school after classes 0.09 0.12 0.032*
(0.01) (0.02) (0.02)
Talks to teacher outside class 0.62 0.70 0.082**
(0.02) (0.03) (0.03)
N. hours study outside school 1.45 1.33 0.112*
(0.05) (0.04) (0.06)
Note: This table contains the average characteristics of all students completing the baselind survey in
grades three through nine. This includes 8,216 students 4,327 in the treatment group and 3,889 in
the control group. Differences are generated by estimating equation (1) using ordinary least squares
with standard errors clustered by school. Significance at the one, five, and ten percent levels is indicated
by ***, **, and * respectively.
- 31 -
Table 3: Comparison of Attriting Students by Baseline Characteristics
Attritors Attritors less NonAttritors
Treatment Control Difference Treatment Control Difference
Characteristics Average Average Average Average
Percentage of Baseline Students 0.353 0.382 0.029
(0.031)
Panel A: Test Scores
Language normalized score 0.012 0.018 0.03 0.071 0.018 0.053
(0.064) (0.082) (0.104) (0.057) (0.044) (0.071)
Math normalized score 0.049 0.047 0.095 0.011 0.047 0.058
(0.095) (0.075) (0.120) (0.050) (0.042) (0.065)
Total normalized score 0.039 0.039 0.078 0.042 0.039 0.002
(0.089) (0.091) (0.126) (0.065) (0.043) (0.077)
Panel B: Demographic Characteristics
Gender 0.527 0.526 0.002 0.031 0.002 0.029
(0.016) (0.021) (0.026) (0.023) (0.021) (0.031)
Age 12.527 12.328 0.199 0.738*** 0.777*** 0.04
(0.260) (0.357) (0.439) (0.123) (0.131) (0.179)
N. parents in the household 1.458 1.516 0.058 0.137*** 0.112*** 0.025
(0.028) (0.030) (0.041) (0.024) (0.027) (0.036)
N. siblings 3.88 4.101 0.222 0.174* 0.11 0.064
(0.175) (0.181) (0.251) (0.092) (0.121) (0.151)
Receives allowance 0.762 0.702 0.060* 0.001 0.027 0.026
(0.015) (0.030) (0.033) (0.017) (0.018) (0.025)
N. friends 17.823 14.533 3.29 0.083 1.582** 1.499*
(1.629) (1.215) (2.022) (0.650) (0.623) (0.895)
Hours of work 7.343 8.286 0.944 1.391** 1.199** 0.192
(0.544) (0.912) (1.056) (0.572) (0.594) (0.820)
Panel C: Academic Variables
Transport time to school 2.384 2.379 0.004 0.075** 0.052* 0.023
(0.055) (0.078) (0.095) (0.030) (0.031) (0.043)
Attended school last year 0.957 0.954 0.003 0.020*** 0.025*** 0.005
(0.007) (0.006) (0.009) (0.006) (0.007) (0.009)
Repeated a grade in the past 0.422 0.43 0.008 0.111*** 0.107*** 0.003
(0.014) (0.018) (0.022) (0.014) (0.017) (0.022)
N. days absent this year 2.196 1.906 0.29 0.226 0.032 0.194
(0.162) (0.123) (0.202) (0.155) (0.145) (0.211)
Failing subject 0.444 0.403 0.041 0.113*** 0.094*** 0.019
(0.022) (0.033) (0.039) (0.018) (0.030) (0.034)
Stays at school after classes 0.103 0.131 0.028 0.016 0.011 0.005
(0.011) (0.019) (0.022) (0.011) (0.014) (0.018)
Talks to teacher outside class 0.627 0.702 0.075** 0.018 0.007 0.011
(0.022) (0.024) (0.032) (0.014) (0.018) (0.023)
N. hours study outside school 1.433 1.302 0.131** 0.019 0.05 0.032
(0.035) (0.046) (0.057) (0.045) (0.045) (0.063)
Note: This table describes the characteristics of students attriting at followup. The first three columns contain the averge characteristics of attriting
students. Differences are estimated using equation (1). The last three columns compare the relative characteristics of attriting and nonattriting
students using equation (3). Standard errors in all models are clustered by school. Significance at the one, five, and ten percent levels is indicated by
***, **, and * respectively.
- 32 -
Table 4: Comparison of NonAttriting Students Using Baseline Characteristics
Treatment Control Difference
Characteristics Average Average
Panel A: Test Scores
Language normalized score 0.083 0 0.083
(0.104) (0.083) (0.133)
Math normalized score 0.037 0 0.037
(0.081) (0.081) (0.114)
Total normalized score 0.08 0 0.08
(0.111) (0.097) (0.147)
Panel B: Demographic Characteristics
Gender 0.496 0.523 0.027
(0.016) (0.024) (0.029)
Age 11.79 11.551 0.239
(0.268) (0.361) (0.447)
N. parents in the household 1.595 1.628 0.033
(0.019) (0.023) (0.030)
N. siblings 3.705 3.991 0.286
(0.222) (0.201) (0.298)
Receives allowance 0.763 0.729 0.034
(0.019) (0.029) (0.034)
N. friends 17.906 16.115 1.791
(1.914) (1.154) (2.223)
Hours of work 5.951 7.087 1.136
(0.398) (0.729) (0.826)
Panel C: Academic Variables
Transport means to school 1.607 1.444 0.163
(0.150) (0.066) (0.163)
Attended school last year 0.976 0.979 0.003
(0.004) (0.005) (0.007)
Repeated a grade in the past 0.311 0.322 0.011
(0.020) (0.018) (0.027)
N. days absent this year 1.969 1.873 0.096
(0.096) (0.167) (0.191)
Failing subject 0.331 0.309 0.022
(0.016) (0.020) (0.026)
Stays at school after classes 0.087 0.12 0.034*
(0.010) (0.015) (0.018)
Talks to teacher outside class 0.61 0.695 0.085**
(0.021) (0.031) (0.037)
N. hours study outside school 1.452 1.353 0.099
(0.057) (0.043) (0.071)
Note: This table contains a comparison of nonattriting students by baseline characteristics using 5,201 non
attriting students (2,798 tretment and 2,403 control). Difference are generated by estimating equation (1)
using ordinary least squares with standard errors clustered by school. Significance at the one, five, and ten
percent levels is indicated by ***, **, and * respectively.
- 33 -
Table 5: First Stage, Distribution of Treatment by Research Group
Treatment Control Difference
Variable Average Average
CPE Reported School Treated 0.958 0.041 0.918***
(0.029) (0.029) (0.041)
Number of Computers at School 13.383 5.102 8.281***
(1.279) (0.753) (1.485)
Percentage of Teachers Trained 0.947 0.082 0.865***
(0.031) (0.04) (0.05)
Note: This table compares characteristics of the 97 schools in our sample. All regressions are
estimated using equation (1) at the school level. Significance at the one, five, and ten percent levels
are indicated by ***, **, and * respectively.
Table 6: FollowUp Test Scores
Percentage Correct Normalized Score
Treatment Control Difference Treatment Control Difference
Difference Difference
Test Sections Average Average w/ Cntrls Average Average w/ Cntrls
Spanish Section 0.42 0.402 0.017 0.015 0.099 0 0.099 0.077
(0.014) (0.013) (0.019) (0.015) (0.071) (0.059) (0.092) (0.076)
Math Section 0.238 0.23 0.008 0.014 0.014 0.07 0.07 0.088
(0.018) (0.011) (0.021) (0.019) (0.019) (0.098) (0.110) (0.109)
Total Score 0.334 0.321 0.013 0.015 0.015 0.111 0.111 0.109
(0.014) (0.011) (0.018) (0.015) (0.015) (0.096) (0.116) (0.104)
Note: This table contains a comparison of the treatment and control groups using the tests administered in the followup survey. Results are presented first for the percentage of correct
answers and then using normalized scores. The first column contains the average scores in the treatment group. The second column contains the average scores in the control group.
Column three contains the simple difference using equation (1), and column four contains the difference estimate controling for baseline characteristics using equation (2). All standard
errors are clustered by school. Sample includes 5,201 students who completed followup survey. Significance at the one, five, and ten percent levels is indicated by ***, **, and *
respectively.
- 34 -
Table 7: Treatment Estimates by Subject
Treatment Control Difference Difference
Subject Average Average w/ Controls
Panel A: Spanish Subjects
Encyclopedia 0.08 0.00 0.08 0.08
(0.04) (0.04) (0.05) (0.05)
Identification 0.11 0.00 0.11 0.13
(0.06) (0.04) (0.07) (0.10)
Paraphrase 0.07 0.00 0.07 0.07
(0.06) (0.05) (0.08) (0.06)
Pragmatics 0.04 0.00 0.04 0.06
(0.05) (0.04) (0.06) (0.06)
Panel B: Math Subjects
Algebra 0.04 0.00 0.04 0.09
(0.08) (0.05) (0.09) (0.14)
Aritmetics 0.01 0.00 0.01 0.01
(0.06) (0.03) (0.07) (0.07)
Geometry 0.07 0.00 0.07 0.10
(0.09) (0.05) (0.10) (0.10)
Statistics 0.11 0.00 0.11 0.12
(0.08) (0.04) (0.09) (0.09)
Note: This table contains a comparison of the treatment and control groups using the tests administered in the
followup survey. Scores are normalized relative to the control group followup survey distribution. The first
column for each subject area contains the average scores in the treatment group. The second column contains
the average scores in the control group. Column three contains a simple difference estimates using equation
(1) and column four contains the difference estimate controling for baseline characteristics using equation (2).
All standard errors are clustered by school. Sample includes 5,201 students who completed followup survey.
Significance at the one, five, and ten percent levels is indicated by ***, **, and * respectively.
- 35 -
Table 8: Difference in Test Scores by Gender at FollowUp
Treatment Control Difference Difference
Test by sections Average Average w/ Controls
Panel A: Female Students
Language score 0.146 0.051 0.095 0.069
(0.077) (0.066) (0.101) (0.087)
Math score 0.088 0.039 0.127 0.143
(0.117) (0.047) (0.125) (0.122)
Total score 0.155 0.017 0.138 0.136
(0.110) (0.065) (0.127) (0.115)
Panel B: Male Students
Language score 0.055 0.047 0.102 0.087
(0.072) (0.065) (0.096) (0.079)
Math score 0.051 0.035 0.016 0.027
(0.087) (0.063) (0.107) (0.107)
Total score 0.068 0.015 0.083 0.079
(0.091) (0.076) (0.118) (0.102)
Note: This table contains the difference between treatment and control groups disaggregated by gender. Panel A presents
the results for female students while Panel B presents the results for male students. The first column presents the
average score on the specified subject for the treatment group. The second column contains the average score for the
control group. The third column presents the simple difference in averages estimated using equation (1) and the final
column presents the average difference controling for baseline characteristics using equation (2). All standard errors are
clustered by school. Sample includes the 5,201 students who completed the followup survey. Significance at the one,
five, and ten percent levels is indicated by ***, **, and * respectively.
- 36 -
Table 9: Difference in Test Scores by Grade at FollowUp
Grade in 2006
Test by Sections 3 4 5 6 7 8 9
Spanish Section 0.139 0.115 0.047 0.092 0.020 0.257*** 0.302**
(0.157) (0.140) (0.160) (0.115) (0.158) (0.098) (0.134)
Math Section 0.170 0.225 0.125 0.042 0.029 0.186 0.355**
(0.226) (0.213) (0.237) (0.158) (0.164) (0.151) (0.160)
Total Score 0.189 0.215 0.133 0.032 0.007 0.298*** 0.426***
(0.225) (0.168) (0.241) (0.145) (0.175) (0.111) (0.138)
Note: This table reports the estimated difference in normalized test scores between the treatment and control groups using equation (2). Estimates are
disaggregated by grade and subject as indicated. Standard errors are clustered by school. The sample size for each grade is given in Table 1, but all 5,201
students who completed the followup survey are included in this table. Significance at the one, five, and ten percent levels is indicated by ***, **, and *
respectively.
- 37 -
Table 10: Difference in Other Academic Outcomes at FollowUp
Treatment Control Difference
Characteristics Average Average
Attended school last year 0.96 0.94 0.02
(0.007) (0.014) (0.016)
Did not attend school last week 0.246 0.37 0.124*
(0.024) (0.067) (0.070)
N. days she did not attended 1.793 2.064 0.271
(0.114) (0.212) (0.239)
Likes school 0.977 0.973 0.004
(0.004) (0.003) (0.005)
Likes contents at school 0.987 0.986 0.001
(0.002) (0.004) (0.004)
Grade report 4.023 4.095 0.072
(0.047) (0.037) (0.060)
Failing subject 0.396 0.365 0.03
(0.028) (0.032) (0.042)
Stays at school after classes 0.064 0.093 0.028*
(0.010) (0.011) (0.015)
Talks to teacher outside class 0.647 0.657 0.009
(0.021) (0.029) (0.035)
Hours of work 8.185 9.92 1.735**
(0.493) (0.613) (0.782)
Note: This table contains the average responses of students in the respective groups. Differences are
estimated using equation (1). Standard errors are clustered at the school level. Sample includes all 5,201
students who completed a followup survey. Significance at the one, five, and ten percent levels is
indicated by ***, **, and * respectively.
- 38 -
Table 11: Teachers' Reported Utilizaton of Computers at FollowUp
Language teachers Math teachers
Treatment Control Difference Treatment Control Difference
Variables Average Average Average Average
Did you use computer last week? 0.416 0.170 0.246*** 0.438 0.182 0.256***
(0.075) (0.041) (0.085) (0.074) (0.043) (0.085)
How many times last week did you?
Use a computer in relationship to your classes 0.876 0.388 0.488** 0.932 0.436 0.496**
(0.188) (0.110) (0.217) (0.173) (0.116) (0.207)
Use a computer for group work in class 1.478 1.444 0.033 1.486 1.621 0.135
(0.176) (0.257) (0.309) (0.197) (0.404) (0.445)
Use a computer for lectures 0.866 0.679 0.187 0.789 0.903 0.114
(0.275) (0.137) (0.306) (0.263) (0.182) (0.318)
Use a computer for class excercises 1.552 1.036 0.517 1.414 1.226 0.188
(0.234) (0.235) (0.329) (0.206) (0.275) (0.341)
Use a computer in class for Internet research 0.746 0.321 0.425 0.761 0.419 0.341
(0.326) (0.124) (0.347) (0.305) (0.198) (0.362)
Allowed class free use of the computeres 1.299 0.893 0.406 1.324 1.387 0.063
(0.320) (0.245) (0.400) (0.303) (0.408) (0.505)
Requires use of a computer for homework 2.313 3.000 0.687 2.569 3.133 0.564
(0.302) (0.507) (0.584) (0.323) (0.504) (0.594)
Note: This table presents the results of teachers reported use of computers for the indicated activities. Statistics are provided for teachers teaching Spanish and teacher teaching
math. Teachers who teach both subject are included in both sets of statistics. Differences were estimated using equation (1) with standard errors clustered at the school level.
Sample includes 426 teachers. Significance at the one, five, and ten percent levels is indicated by ***, **, and * respectively.
- 39 -
Table 12: Fraction of Schools Reporting Computer Use at FollowUp by Grade
Language teachers Math teachers
Treatment Control Difference Treatment Control Difference
Grade Average Average Average Average
Third 0.45 0.25 0.202** 0.45 0.22 0.222**
(0.07) (0.06) (0.10) (0.07) (0.06) (0.10)
Fourth 0.43 0.27 0.160* 0.45 0.25 0.202**
(0.07) (0.06) (0.10) (0.07) (0.06) (0.10)
Fifth 0.36 0.27 0.10 0.38 0.27 0.12
(0.07) (0.06) (0.10) (0.07) (0.06) (0.10)
Sixth 0.09 0.02 0.07 0.13 0.06 0.07
(0.04) (0.02) (0.05) (0.05) (0.04) (0.06)
Sevent 0.06 0.06 0.00 0.13 0.10 0.03
(0.04) (0.04) (0.05) (0.05) (0.04) (0.07)
Eight 0.04 0.06 0.02 0.11 0.12 0.02
(0.03) (0.04) (0.05) (0.05) (0.05) (0.07)
Ninth 0.04 0.04 0.00 0.09 0.08 0.00
(0.03) (0.03) (0.04) (0.04) (0.04) (0.06)
Tenth 0.02 0.02 0.00 0.06 0.08 0.02
(0.02) (0.02) (0.03) (0.04) (0.04) (0.05)
Eleventh 0.02 0.02 0.00 0.06 0.06 0.00
(0.02) (0.02) (0.03) (0.04) (0.04) (0.05)
Note: This table contains the fraction of schools, in each research group, that have at least one teacher teaching in the indicated
grade that reports using a computer. Differences were estimated using equation (1) at the school level. Significance at the one,
five, and ten percent levels is indicated by ***, **, and * respectively.
- 40 -
Table 13: Student Reported Computer Utilization in Previous Week at FollowUp
All Students Students using Computers
Treatment Control Difference Treatment Control Difference
Average Average Average Average
Did you use a computer last week? 0.66 0.41 0.252***
(0.04) (0.05) (0.06)
Where did you use a computer last week?
at school 0.57 0.27 0.300*** 0.85 0.64 0.214*
(0.04) (0.06) (0.07) (0.04) (0.10) (0.11)
at home 0.04 0.05 0.01 0.06 0.11 0.05
(0.01) (0.02) (0.02) (0.02) (0.04) (0.04)
at an internet café 0.08 0.07 0.00 0.12 0.17 0.06
(0.04) (0.03) (0.05) (0.05) (0.08) (0.09)
at a friends home 0.02 0.03 0.01 0.04 0.07 0.038**
(0.00) (0.01) (0.01) (0.01) (0.02) (0.02)
In which subjects did you use a computer?
Mathematics 0.03 0.02 0.01 0.03 0.02 0.00
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Spanish 0.04 0.03 0.02 0.05 0.03 0.01
(0.01) (0.01) (0.02) (0.01) (0.01) (0.02)
Natural Sciences 0.03 0.03 0.00 0.03 0.04 0.01
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Social Sciences 0.02 0.02 0.00 0.02 0.02 0.00
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Computer Science 0.51 0.21 0.294*** 0.52 0.22 0.296***
(0.05) (0.05) (0.07) (0.05) (0.05) (0.07)
Art 0.02 0.02 0.01 0.02 0.02 0.01
(0.00) (0.01) (0.01) (0.00) (0.01) (0.01)
Note: This table contains estimates of differences in students' reported use of computers. The first three columns contain estimates for all 5,201 children
completing a followup survey. The last three columns contain estimates for just those students who report having used a computer in the last week. All
estimated differences are generated by estimating equation (1) with standard errors clustered at the school level. Significance at the one, five, and ten percent
levels is indicated by ***, **, and * respectively.
- 41 -