Learning to Teach by Learning to Learn∗

Vesall Nourani Nava Ashraf Abhijit Banerjee

UChicago LSE MIT

November 10, 2023

Click here for latest version

Abstract

Massive learning gaps persist in most developing countries, undermining rapid

gains in school attendance. While a pedagogy based on rote learning of facts is often
cited as a factor for limited school effectiveness, evidence of success of introducing
new pedagogies to teachers is scarce. Indeed, most in-service teacher interventions
evaluated in the literature have been ineffective. In this paper, we report on the
randomized evaluation in Uganda of an intervention in which teachers were trained
in a “learning how to learn” approach. The curriculum, called Preparation for Social
Action, invited teachers to learn how to learn like scientists: posing sharp questions,
framing specific hypotheses, using evidence and data gathered from everyday life
whenever possible. We find evidence of change in pedagogy, with dramatic effects on
learning: The intervention raised the pass rate in the national exam that determines
progression from elementary to secondary school from 51% to 75%, which places the
program in the top five percentile of all rigorously evaluated education interventions
in terms of learning-adjusted years of schooling per USD. Effects persist four years
after the first cohort of teachers was trained, including in the learning outcomes of
students who continue to secondary school.

∗ We would like to thank Kimanya Ngeyo Foundation for Science and Education for providing us the opportunity to study
their teacher training program in Uganda. Many seminar and conference presentations have contributed to the refinement
of ideas in this paper. Their commitment to systematically learning about their own activities has benefited this study
immensely. This study is funded by the Abdul Latif Jameel Poverty Action Lab’s Post-Primary Education Initiative and,
in its early stages, by the International Growth Centre’s Small Project Facility in Uganda. The corresponding author
acknowledges support from the National Science Foundation under grants numbers DGE-1144153 and 1809427. The RCT
in this paper was pre-registered as AEARCTR-0002647. IRB approval was provided by LSE IRB and UNCST #REC REF
0510-2017. Numerous research and field assistants have provided excellent support along the way. Special thanks goes
to Azizi Buyinza, Andrew Chemonges, Ed Davenport, Moustafa El-Kashlan, Stefan Faridani, Joshua Obaya Humphreys,
Timo Kapelari, David Liedig, Beatrice Montano, Kim Sarnoff, Misha Seong and Nick Swanson.
1 Introduction
While the last few decades have witnessed an unprecedented increase in access to school
worldwide, this access has not been accompanied by a commensurate increase in student
achievement. In Uganda, the country where our study takes place, close to 100% of
primary age students attend primary school - but only 35.5% of students who started
primary earlier in the late 2000’s qualified for the last grade of primary education by 2016,
and of these only 59% of primary-school leavers transitioned to lower secondary school
(UNESCO, 2020). Even after reaching grade 7 (the last year of primary), 30% cannot
read and understand a story in English, the primary medium of education (UWEZO,
2016). The results from many other developing countries such as India, Pakistan and
Nigeria, are similarly concerning.1
Dealing with this learning crisis is widely recognized to be a key challenge of the
day (World Bank Group, 2018). The public discussion of the strategies to deal with it
emphasize teacher education. The main strategy across diverse contexts is to revamp the
teacher education framework.2 However the Global Education Evidence Advisory Panel’s
recent (2020) report on “Best Buys in Education,” lists in-service teacher training as an
area where lots of resources are spent “but evidence on how to do it effectively is low”,
and goes on to say

“However, there is little evidence showing that the typical stand-alone general-
skills in-service training is cost-effective. Indeed, much of the rigorous ev-
idence that is available suggests that it does not improve student learning
outcomes.”

The Report however contrasts this conclusion about general-skills in-service trainings
with trainings that are dedicated to help teachers implement specific pedagogical practices
strongly supported by evidence from multiple studies.
This paper stakes out a claim for a specific category of general-skills in-service training.
We report evidence from an RCT in Uganda of very large improvements in student out-
1
In India, for example, only 50% of fifth graders can do second grade level maths. Only 16% of children
in rural India attending Grade 1 of primary school were able to read a Grade 1 level text. While every
child in Grade 1 in India is expected to recognize 2-digit-numbers (according to the National Council
of Educational Research and Training’s specification of learning outcomes), only 41.1% of children in
Grade 1 were able to do so (ASER Centre, 2019). In rural Pakistan, only 59% of children in Grade 5
were able to read a story in their local language in 2019, while 57% were able to do division. In urban
Pakistan, the numbers were 70% and 66% respectively (ASER Pakistan, 2019). After having received
over three years of language training, 80 percent of students in Nigeria were not able to read a simple
word of English. In Uganda, around half of the students who had received three years of mathematics
teaching were not able to order numbers between 0 and 100 (Bold et al., 2019).
2
For example, India’s recent New Education Policy (Ministry of Human Resource Development, 2020)
and Uganda’s Effective Teacher Policy (MoES, 2014).

1
comes resulting from a one-year-long in-service teacher training intervention, combining
three periods of intense study during term breaks with monthly in-service visits from tu-
tors. Teachers study a subset of a curriculum titled Preparation for Social Action (PSA).3
The teachers are trained to teach in a way that is inspired by the scientific method, em-
phasizing the use of precise language, posing sharp questions, framing precise hypotheses
and using evidence to decide. To make this approach more effective, the teacher is en-
couraged to use examples from everyday life wherever possible, so that the questions, the
hypotheses and the relevant evidence are likely to be concrete. The training takes place
through demonstrations, where the trainer takes the teachers through multiple instances
of the method, applied to specific contexts.
However this is still very far from a scenario where the trainer sets very specific teaching
goals and lays out the exact steps to get there. The teachers in this program still have
to figure out how to apply these ideas in the context of what they are teaching, which
can be English, Social Studies or some branch of Science, and do so for multiple different
topics within each field. In this sense, the intervention we study is very different from
Teaching at the Right Level (TaRL) and Structured Lesson Plans, which are the two
pedagogy/teacher training interventions that have been widely and successfully replicated
in multiple developing countries, and much closer to general-skills training. Both of
those interventions narrow down the pedagogy to a relatively small number of steps
combined with a clear set of proximate goals — for example, basic skills development
for young children. The current intervention, on the other hand, enables a broad-based
shift in the capabilities of the teachers, so that the teacher becomes less the source of the
answer and more able to assist students to learn to find the answers themselves across
diverse applications and contexts. This is important, for example, when teaching young
adolescents, aged 11 to 15, who study multiple different fields and are expected to cover
much more material, which makes it harder to provide teachers with a minute-by-minute
lesson plan.
In 2017, we randomly selected 18 out of 35 primary schools in a subcounty of Jinja,
Uganda using a pairwise matching procedure (Bruhn & McKenzie, 2009) to participate in
the training.4 The District Education Office (DEO) invited these schools to send two to
three primary school teachers (emphasizing upper primary — grades 4 through 7 — over
lower primary) to a training opportunity offered by KN. Therefore, our analysis measures
treatment effects for schools where two cohorts of teachers participated in the training
3
The curriculum was originally developed by Fundación para la Aplicación y Enseñanza de las Ciencias
(FUNDAEC) in Colombia in the 1970s and implemented in Uganda by an Ugandan NGO called Kimanya
Ngeyo Foundation for Science and Education (KN) and the Jinja District (Government) Education Office.
The translation of FUNDAEC into English is “The Foundation for the Application and Teaching of the
Sciences.”
4
Between 2017 and mid-line data collection in 2019, three schools became insolvent, causing them to
close. Due to pairwise matching, we proceed with the analysis unhindered by dropping each school in
the pair from the study, resulting in outcomes from 29 schools analyzed in the present study.

2
— one in 2018 and the other in 2019.
The intervention had large and positive effects on almost every outcome where we
expected one, as recorded in the pre-analysis plan.5 Student performance in treatment
schools increases by 0.5 standard deviations in high stakes (primary leaving) national
exams that determine whether or not a student can proceed to secondary school — the
pass rate in treated schools increases by 24 percentage points relative to a control school
mean of 51%. This result reflects improvements across the entire distribution of student
achievement and for boys and girls alike. There were large improvements in performance
in all the fields that were featured in the training — English, Social Studies and Science
— and these are largely robust to using Randomization Inference to deal with the small
number of clusters and Benjamini-Hochberg adjustments for multiple hypothesis testing.
The one area where there was no clear evidence of improvement was mathematics, which
was explicitly left out of the teacher training due to time constraints.6
To understand better the mechanisms at the student level, we were interested in
whether the students have a deeper understanding of the conceptual material or whether
the main source of improvement was more effective rote memorization, resulting from
say greater effort or higher self-confidence. To do this we administered an exam using
questions from a commonly utilized assessment booklet in Uganda (SIPRO, 2019) and
classified questions in this booklet that reflect higher-order learning using classification
categories described in (Burdett, 2017).7 Student performance in these tests improves by
between 0.5 and 0.8 standard deviations.
To test the idea that the students actually absorbed the scientific method, we imple-
mented a novel field test of scientific competencies. We coordinated with the district
education office to institute a series of science shows, where student-led experiments
tested methods of soil conservation and clean water preparation in practice. We present
results of these experiments to assess practical scientific competencies.8 Specifically, an
independent panel of judges hired and trained by IPA-Uganda assessed the performance
of students by attending the shows held by each of the schools in the study. The treat-
ment students demonstrate a higher aptitude for scientific thinking as measured by the
judges’ assessments of the school-based science shows. Students proficiency in designing
5
The experiment is registered in the AEA RCT Registry under AEARCTR-0002647, where the pre-
analysis plan was published on October 2, 2020.
6
As a consequence of these results, KN is revisiting approaches to including PSA materials focused
on mathematical capabilities.
7
According to this classification, there are, roughly, three broad categories of assessment questions
that measure student recall, application of understanding, and critical thinking and reasoning. As
Burdett (2017) notes, there is a paucity of questions assessing critical thinking and reasoning in most
developing country examinations. We, therefore, augmented the common assessment with additional
questions testing critical thinking and reasoning using the framework described in Liu et al. (2014).
8
The National Curriculum Development Centre (NCDC) strongly encourages the use of science shows,
though they are scarcely used in practice. We drew on this encouragement in our conversations with
schools and the district education office and tied our measurement of scientific competencies almost
exactly to those requested by the NCDC.

3
experiments, articulating and analyzing hypotheses, measuring observations, and com-
municating results increase by 31%, 80%, 61% and 60% respectively over control school
peers, all significant at conventional levels.
We use several other measures of teacher outcomes to link improvements in student
learning to pedagogy. Classroom observations test whether there are observable changes
in the conduct of the classroom and we find that students in treated schools are 39
percent more likely to be engaged in the activity conducted by a teacher relative to their
peers in control schools. Students in treated schools are 6 percentage points more likely
to have asked teachers questions in class when they don’t understand a concept that is
being presented relative to a control school mean of 22%. We develop a novel measure
indicative of the degree of attention they give to their students by quizzing them on
details of their students’ lives and past attendance. Teachers in treated schools are 10
percentage points more likely to know these details.
In 2022, four years after the initial cohort of teachers were trained, we administered
assessments to all primary school students, secondary students who we can trace to the
primary schools in our study and primary school teachers. We also carried out a household
census with recall data on the previous five years of school enrollment for each household
member to develop measures of persistent school enrollment and grade progression. We
find that all treatment effects associated with student learning persist in the long run,
including in the learning outcomes of secondary students in 2022 who were assigned to
treated schools in 2018 but are no longer taught by trained teachers. Furthermore, we do
not find differences in teacher test scores when we test them on questions relevant to the
subjects they teach. However, we do find that large effects in a series of questions that test
whether teachers are more precise and consistent in their “use of language” — exercises
where they choose between words such as “science” and “technology” or “understand” and
“know” in sentences where an untrained eye might not be able to differentiate word choice
precisely.
The student learning results in particular, measured in Learning Adjusted Years of
Schooling (LAYS), place the program in the top five percentile of all education interven-
tions across the current list of 150 studies analyzed by Angrist et al. (2020). A conser-
vative estimate suggests that an additional $100 of funding increases learning adjusted
years of schooling by 9.62 years.
While we do not have direct observations of what happens during the training itself,
there is a large body of writing that can help us understand why the PSA approach has
the potential to promote improved learning (Freire, 1970, Ranciére, 1991, Duckworth,
2006). Teachers in these settings, as in most of the world, are accustomed to “chalk and
talk,” or lecture and blackboard-centered pedagogy (Glewwe & Muralidharan, 2016). To
quote Paolo Friere, the Brazilian educationist who partly inspired FUNDAEC’s work,
Narration (with the teacher as narrator) leads the students to memorize me-

4
 chanically the narrated account. Worse yet, it turns them into “containers,”
into “receptacles” to be “filled” by the teachers. The more completely she fills
the receptacles, the better a teacher she is. The more meekly the receptacles
permit themselves to be filled, the better students they are. Education thus
becomes an act of depositing, in which the students are the depositories and
the teacher is the depositor.

By contrast, the PSA training adopts a premise that, in part, suggests that knowledge
is to be discovered by the learner (Freire, 1970) and that one of the teacher’s primary
roles is to help the student discover strategies for learning. FUNDAEC describes its
tutors which would include the trainer who trained the teachers as well as the trained
teachers themselves thus:

Tutors guide the students through the textbooks, raise questions, clarify ob-
scure matters, encourage reflection on real-life experience and supervise ex-
periments and social action. They do not lecture or dictate but nor are they
mere facilitators of group discussion (Arbab & Lample, 2005).

In practice what the teacher does differently can be quite simple. For example, science
teachers typically introduce plant biology by drawing a plant on the blackboard and ask-
ing students to memorize the location of the “roots,” “stem,” “leaves,” “buds,” “flowers,”
etc. Instead, teachers might now bring actual plants into the classroom and ask students
to differentiate the parts of a plant according to what they observe. The teacher might
ask the students to reflect on the process of a growth of a plant by planting, observing,
and measuring key aspects of a plant at distinct stages, thereby facilitating the use of
scientific methods. Once the learner has exerted sufficient effort to understand plants
through observation, the teacher provides clarity by linking the learner’s efforts to what
we know about the core characteristics of plants, but the source of knowledge has shifted
in a fundamental manner through this exercise. The pedagogical act now disassociates
teachers from the origin of knowledge, without compromising the teacher’s role as an ex-
perienced guide, and students slowly begin to see how their own critical thinking abilities
can be a source of insights into the reality around them.
Moreover this process of discovery, experimentation and research spills over to problem-
solving within the school as well. Whenever teachers complain about a challenge they
face in the school environment, such as student absenteeism, or a confusing directive from
school managers, KN tutors ask them “Do you know what the source of this challenge
is? Have you done enough research to find out?” 9 If the trained teachers take this lesson
9
For example, one teacher shared that a pupil of theirs was consistently late to school. Prior to
participating in the training, the teacher’s main strategy was to berate the pupil and send him back
home so that he would learn his lesson and arrive on time the following day. This did not happen and
the pupil was consistently sent home day after day. When the teacher brought this challenge up during

5
to heart and take the lead in solving problems, the culture of the entire school, including
teachers not included in the program, may change.
The emphasis on how to learn to learn is closely related to the increasing emphasis
on metacognition (thinking about thinking) in the literature on how to improve teaching
practices (Perkins, 1992, Kolencik & Hillwig, 2011, Tanner, 2012, Lang, 2012, Swartz
& Perkins, 2017). This literature argues that giving students more control over the
learning process helps them reach a deeper understanding of the material and become
more engaged, but also gain confidence, which helps them make better use of what they
know.
To the best of our knowledge, this is the first rigorous impact evaluation of a “learning
to learn” intervention. Evidence from multiple RCTs have established the idea that the
traditional top-down pedagogy, where the teacher teaches a fixed curriculum and is largely
unresponsive to specific gaps in what the students know, is a major source of the learning
gap in many developing countries. When the pedagogy is altered to focus on specific child
level gaps, they make rapid progress (Piper et al., 2014, Banerjee et al., 2017). While
most of these studies focus on basic skills for younger children, Muralidharan et al. (2019)
show that a tablet-based TaRL program that coaches children based on where they are
making mistakes, can lead to large gains in learning levels for children in Upper Primary
or Middle school. Bando et al. (2019) show that another technique, inquiry- and problem-
based pedagogy, improve student performance in math and science, primarily for boys,
in a way that increases performance over time. We offer that a general-skills training
in which the teacher learns how to learn may enable her to draw on the appropriate
pedagogy depending on the learning objective on hand.
Moreover, this insight helps us think more broadly about the relationship between
human capital accumulation, technology and economic growth (Becker, 2009, Goldin &
Katz, 2009). A growing literature argues that the type of human capital required of
modern economic growth matters (Kremer, 1993, Acemoglu et al., 2012, Hanushek &
Woessmann, 2012, Flabbi & Gatti, 2018). More often than not, labor needs to be flexible
to the demands of modern technologies. We propose that metacognitive, “learning to
learn,” approaches to human capital development, rather than measuring simple years
of schooling, may capture vital skills for economic growth and development. We explore
this further in our conclusion.
The rest of the paper proceeds in the following manner. In section 2 we provide details
of KN’s Intervention, articulating our theory of change and its relationship to our research
questions and hypotheses in our pre-analysis plan. In section 3, we provide more details
a KN teacher-training, the tutor asked her to investigate why the student was consistently late. After
some research, the teacher learned that the student spent his morning hours selling donuts at the local
market in order to pay school fees. Feeling sympathy towards this student, the teacher began to allow
the student to arrive late to school whenever necessary and engaged more deeply with his family to see
if there were other ways of supporting the child so that he did not miss important class time.

6
on our experimental design and data. In section 4, we describe our identification strategy
and share and interpret results before providing a cost-benefit analysis of KN’s teacher
training program. We conclude with thoughts for next steps in section 7. All appendix
materials are available online.

2 The Intervention and Conceptual Framework

2.1 Preparation for Social Action (PSA) Teacher Training
The PSA curriculum has been in development since the mid-1970s and is continuously
reviewed and updated through an ongoing process of action-research, often taking up
to ten years to produce texts before distribution.10 KN coordinates the study of PSA
materials among groups of tutors in Eastern Uganda, including in Jinja district where
our study is based. This section weaves together a description of the content of the
training with a conceptual framework that explains how the approach cultivates a general
approach to developing teacher capabilities.

2.1.1 Learning How to Learn

The PSA materials are studied by the teachers together with the tutor in a highly par-
ticipatory manner. The initial experience is intended to enable the teacher to experience
a new type of learning environment and is, therefore, jarring for most Ugandan teachers
accustomed to training where they are passive recipients of lectures. Instead, the tutor
starts by handing out a PSA text that all teachers take turns reading out loud. The text
poses questions that the teachers are invited to answer. Initially, teachers tend to be
hesitant to share their thoughts lest they embarrass themselves by providing a “wrong”
answer.
As an example, here is the first page of the first text in the PSA training, on “Prop-
erties,”:

When we look at the world around us, we immediately notice that objects
have different shapes and sizes. This enables us to differentiate things and
describe them to others. For example, in everyday conversation we speak of
round things, square things, long, wide, big, and small things. Let us think
about these concepts to find out how much we really understand them.
At one time or another, we have all played in the mud or the sand building
things of different shapes. Perhaps we are now too scrupulous to play with
10
In the early 2000s, FUNDAEC, the organization that develops the PSA materials, began partnering
with collaborating institutions around the world which now includes a growing network of organizations
across Africa, including KN, which was established in 2007.

7
 mud, so let us use a bit of clay or play-dough to make some interesting shapes
and try to describe them to one another.
[Participants spend 15 minutes making shapes out of clay and then attempt
to describe the shapes to one another for another 30 minutes with questions
interspersed by the tutor. Next, the text asks.]
Was it easy to describe all the shapes?

Teachers initially provide common responses to a spherical object someone may have
crafted such as “round,” “circular,” or it “looks like a ball.” The tutor introduces questions
that challenge these descriptions such as “What do you mean by circular? I thought circles
were flat?” or “But that is not entirely round. I notice a few bumps on that object. How
would you describe it in its entirety?” The conversation around the shapes teachers have
crafted soon transitions into a conversation about dimensionality as well as one about
regular and irregular shape names — all through an activity where they attempt to
describe the shapes they themselves made. As teachers struggle to find words that make
their description more precise, the tutor helps them realize that their prior descriptions
required an underlying understanding of the words “round” or “circle.”
While seemingly simple, the novelty of the exercise exposes teachers to new learning
experiences that the cohort of teachers can reflect on collectively. In section ?? we de-
scribe how later parts of the training help teachers systematically analyze the pedagogical
environment that led to these learning experiences. For now, it is worth pointing out that
as teachers analyze the exercise of describing shapes they recognize that prior knowledge
is a pre-requisite of precise language and is required to generate new knowledge. This
underscores the necessity of being precise with one’s language, and for a shared under-
standing of prior beliefs/concepts. If this can matter with something as simple as creating
a sphere out of clay, how much more important can it be with more complex shapes and
concepts?11
In later lessons, for example as teachers carry out activities to learn about the growth
of a plant, the participant recognizes the values of distinguishing and describing different
parts of a plant: its root system, its branches, its leaves, and so on. As the participant
learns to describe each part in greater detail, her description of the process of the growth
of a plant is rendered more accurate.
11
The tutor might subsequently ask, “what other shapes and words are important to know to describe
these objects more precisely?” The text then reads “as we deal with more and more complex shapes,
the task of describing them becomes increasingly difficult.” Teachers consider how they can describe
increasingly complex objects by combining regular shapes in their descriptions. For example, a house
can be described as having a long rectangular cuboid concrete slab as a foundation with a cuboid structure
on top of the foundation. On top of the cuboid is a pyramid or prism-like shape that we call a roof.

8
2.1.2 Developing Scientific Capabilities

The above exercise illustrates a non-hierarchical mode of teaching related to one aim of
science: to describe reality with increasing degrees of precision. However, it builds on this
illustration with a series of additional experiences that foster the development of scientific
capabilities in teachers.12 These dimensions include, but are not limited to: 1) the use
of precise language, 2) observing and describing the world around oneself with increasing
clarity, 3) designing experiments and testing hypotheses, 4) distinguishing objective from
subjective statements, 5) identifying sources of data and processes worth studying, and
so on. We list below concrete examples of exercises done in the teacher training.

Using precise language. In “Transition to Agriculture,” one exercise reads “Depending

on how it is used, a word can carry one or another meaning, which is easily seen by
the explanations given in a dictionary.” Teachers are then asked to consider how the
context in the reading determines the meaning of a pre-selected set of words used
in the materials. They recognize that precision of language does not only depend
on the definition of a word, but also on the context in which it is placed.

Observing and describing the world with increasing clarity. When agriculture is
introduced in the module “Planting Crops” teachers are asked to observe the leaf
structure of the plants in their environment and classify them according to their
biological properties.

Designing experiments and testing hypotheses. Participants set up an experiment

that allows them to measure the boiling point of water, testing the hypothesis that
the boiling point is 100◦ C.

Distinguishing between objective and subjective statements. When articulating

hypotheses, teachers are careful not to use subjective terms such as “improved soil
health,” but rather objective measurements of soil health, potentially observable
to all, such as soil pH, moisture absorption, precise proportions of clay, sand and
loam, and so on.

Identifying processes worth studying and sources of data. When learning about
“urban gardens” they measure the amount of labor and financial input as well as
both the nutritional and economic value of the crops produced. In a text titled
“Nurturing Young Minds,” participants carry out interviews with village elders to
collect information on past and present forms of child rearing and education, often
focusing on the moral, social and intellectual dimensions of child development. This
12
There are multiple ways in which the PSA materials foster these capabilities; We provide coarse
summaries of each course in appendix A.

9
 allows for further and deeper analysis of the forces that enable or impede develop-
ment of their learners. In ‘environmental issues,” teachers visit local markets and
investigate waste management practices among small and large scale manufacturers.

These are but a few examples of experiences teachers have as they experience various
modalities of learning. Beyond providing them with an illustration of a non-hierarchical
mode of teaching the deepening of awareness of scientific approaches to learning generates
an orientation of humility, an openness to data and refinement of ideas.

2.1.3 Understanding How Knowledge is Produced

The above examples of learning activities are not experienced in isolation from each
other during training. From the very first exercise of describing a shape, teachers are
asked to reflect on how any descriptive language or shared understanding relies on as-
sumptions/prior concepts. This is reinforced by the scaffolding of the learning approach
within the training — the learning experience in PSA is cumulative and makes clear that
knowledge can grow along a path of discovery that starts from initial building blocks.
As learners encounter gaps in their understanding, new observations lead to the need to
develop more precise language to learn about, and describe, any object of reality. The
experience of the training thus sheds light onto how science — and knowledge — ad-
vance, by incrementally adding basic shared concepts to describe increasingly complex
phenomena.
A passage in “The Heating and Cooling of Matter” helps illustrate this point. In this
course, participants develop an appreciation for the atomic theory of matter and how the
evolution of this theory helped humanity understand the relationship between heat and
matter through an iterative process of observation, refinement of models, and further tests
over time. Toward the end of the course, participants reflect on the following questions

The way we have chosen to introduce the subject of heat may lead you to
believe that a theory is somehow the conclusion one reaches after making
many observations. But pause a little and ask yourselves if this is really
true. Would people actually begin to observe any process in nature if they
did not already know something about it? what kind of questions would you
ask about the process of heating and cooling if you did not have some notion,
right or wrong, of what heat is? Think of very small children. Does what they
learn about heat and temperature arise directly from their sense of touch and
sight, or does somebody have to teach them about these concepts?

Teachers may spend over an hour discussing these question in their group. The result
of this exercise is a deeper appreciation that (scientific) knowledge can emanate from

10
observations made by anyone, including young children. In order for knowledge to ad-
vance, people build on the accumulated insights of past experiences as well as the forward
movement of new ideas — including new observations made as language of description
becomes more refined. The ability of each individual to contribute meaningful insights
and observations beckons teachers towards an orientation of humility to the perspectives
of others, including children — in gathering and incorporating new data. Knowledge
of the process of knowledge production gives rise to a sense of agency in the ability to
generate knowledge oneself.

2.1.4 Creating Conceptual Connections

Learning comes with understanding new concepts, including the meaning of the term
concept itself, and learning how concepts generalize to broader contexts. For example,
continuing in “Properties,” teachers begin discussing how the concept of “properties,” can
be applied to their own understandings of their students. Towards the end of “Properties,”
teachers read:

Up to now we have been studying the properties of matter. We have examined

certain general properties common to all matter and a few specific ones that
make each substance different from others. We would now like to ask whether
the concept of property also applies to human beings. Do they have general
and specific properties that we can use to describe them?...
It so happens that we do not normally use the word property to describe
human beings. We use words such as characteristic, quality, and attribute.
The underlying concept, however is not so different. In the same way that it
is the property of water to flow, certain virtues such as honesty, courage, and
generosity are properties of human beings. So, too, is the capacity to think
and to express thoughts through language.

The material, with the help of the tutor, carefully draws a connection between how we
observe and describe the properties of nature and our observations and descriptions of hu-
man beings and human behavior. Creating this conceptual connection can have profound
effects. For example, on several occasions the PSA materials explicitly or implicitly ask
teachers to examine what it means for a material substance to change phases — say, from
a liquid to a gaseous state. Following the sections that analyze “properties” of human
being, the tutor also helps the teachers to reflect on their own students. They might ask:
“What are the qualities you desire in your students? How can you facilitate a process
that helps the students acquire these qualities? How can we apply our understanding of
the “change of phase” of substances to this process?” 13
13
This extends to practical exercises as well. In “The Heating and Cooling of Matter,” teachers ex-

11
 The fact that material substance can change phase becomes an analogy for how human
“properties” (qualities) may also not be fixed and static, but rather subject to change
and transformation. This can change how teachers view their students, from having
fixed qualities (bad at math, for example) to capable of change - just like other material
substances.
A second example from “Properties” further illustrates how scientific concepts can
affect teachers’ perceptions of their students. In one lesson, teachers explore the concept
of “objective” and “subjective” descriptions.14 For example, as they discuss the scientific
imperative of striving to make and test objective statements, they gain insights into the
concept by reading a series of statements and discussing whether they are objective or
subjective descriptions, e.g., “the school is the largest building in the village” or “he has a
big heart.” The tutor then asks them to reflect on the various statements they might make
about their students and whether it describes an objective or subjective characteristic
of their students. Teachers soon realize that statements that describe students may
differ from teacher to teacher. They begin to recognize pitfalls in their thinking —
that subjective terms such as “smart,” “slow,” and “difficult” may reflect their own biased
perspective more than anything else. They begin to realize that their colleagues might see
the very same student in a different light, propelling them to seek the advice of others to
reduce the bias in their opinion. The recognition of subjective bias even propels changes
in the teacher’s own attitudes and behaviors towards students, a process we describe in
more detail when describing our conceptual framework and predictions in section 2.2.
The concepts “change in phase,” “subjective” and “objective” descriptions are only a
few concepts that the teachers unpack in depth. Additional concepts include the dif-
ference between “information” and “concept” and the associated ideas of “assimilating
information” and “understanding concepts.” This exploration of concepts accompanies
the other features of the rich new learning experiences teachers have over the course of
the year. Together, these experiences help teachers appreciate how anyone can learn
to produce new knowledge, or “learn how to learn.” This process can begin with very
basic building blocks where each block gives rise to subsequent questions that require
more investigation. Teachers recognize that when knowledge is explored and organized
in this manner, the educational experience allows the learner to appreciate the role they
can play in advancing knowledge generation, granting them with a sense of agency for
periment with measuring the boiling point of water — the point at which molecules break apart and
water changes phase from liquid to gas. The concrete nature of this exercise aids tutors to initiate a
conversation in which teachers think more concretely about the “transformation” they desire to see in
their students. Tutors are not satisfied with vague statements and point out how the teachers’ description
of this transformation should strive towards being as concrete and precise as how they can now describe
the change of phase of a substance.
14
In a similar vein, teachers discuss the difference between “general” and “specific” properties of sub-
stances. A dialogue between a KN tutor and participating teachers, highlighting aspects of KN tutors’
pedagogy, is available to read in Appendix A.7.

12
contributing meaningfully to processes of discovery. Equipped with this experience and
realization, the final step of the training involves assisting the teacher to create similar
learning experiences for their students, that is, “learning to teach.”

2.1.5 Learning How to Teach

The second part of the training pivots to teaching, using the same insights used to learn
how to learn. Teachers begin by carrying out an explicit analysis of the PSA curriculum
using “Discourse for Social Action (DSA): Unit 2 Education,” part of another sequence
of courses developed by FUNDAEC. This material is used in the teacher training to help
the teacher think about how the curriculum and pedagogy in PSA influenced her own
thinking — a meta-cognitive analysis of the pedagogical act. By unpacking the path
between pedagogy and learning, this process helps the teacher generalize beyond her own
experience and to think how her pedagogy influences her pupil’s thinking.
It does so by walking the reader through FUNDAEC’s description of the aim and
purpose of its own educational programs — “to assist our students in their spiritual and
intellectual growth as individuals and in the development of their capacity to contribute
to the transformation of society.” Teachers reflect more broadly on the aim of education,
for example by understanding the difference between “assimilating information” and “un-
derstanding concepts,” by recognizing the importance of integrating theory with practice,
being with doing, moral and academic achievement, all while avoiding the tendency to
simplify reality by espousing extreme views and false dichotomies such as “the aim of
education is to help students understand concepts rather than assimilating information.”
It is important to note that what distinguishes this approach from other approaches
to teacher training is that, prior to carrying out the meta-cognitive analysis, teachers
in KN’s training first experience the PSA curriculum as learners. Thus, their analysis
is not of the curriculum or pedagogy alone — they also analyze their own experiences
together with an analysis of the PSA curriculum and pedagogy, rendering the act of
generalizing pedagogy more concrete and practicable. Rather than copying and pasting
the PSA curriculum for use in their own classroom, teachers see the value of thinking
clearly about the purpose of any curriculum and ensuring that the pedagogy and content
are coherent with this purpose.
What does this mean in practice? When reflecting on the difference between assimilat-
ing information and understanding concepts, for example, teachers recognize that much
of their (former) pedagogy is aimed at helping their students memorize definitions, for-
mulas or procedures depending on the subject. From their own experience with the PSA
material as well as reflecting on their aim as an educator, they realize that memorization
may be insufficient as an objective of an educational endeavor. They recognize that a
different pedagogy can engender forms of learning that allows one to apply understand-

13
ing to ever-broadening contexts.15 They analyze the approach taken by the KN tutor to
guide participation in the PSA courses and reflect on the underlying aim, purpose, and
structure of Uganda’s National Curriculum and how it correlates with their own evolving
notions of the aim and purpose of education. Crucially, rather than mirroring approaches
to teaching offered by their own teachers, they are able to think more critically about
pedagogical choices that advance student learning more holistically, beyond the mere
assimilation of information.
Each of the pieces of KN’s training are brought together through this exercise, and
teachers see how the experience of “learning how to learn” will lead them to learn how to
teach: they use their newfound scientific capabilities to observe and describe their own
practices with increasing clarity; they consider more carefully the concepts underlying
the curriculum they are told to teach; and they reflect on how their own pedagogy can
fit the learning task at hand, articulating concrete plans and approaches that advance
student learning towards diverse learning objectives. This newfound approach to learning
has profound implications for how teachers teach. We summarize the implications in the
next subsection.

2.2 Summary and testable hypotheses

The traditional approach to teaching in these schools is best described as “explication”
— an approach that works under the assumption that knowledge is a static and well-
defined object, and the main problem is that students lack this knowledge which teachers
possess Ranciére (1991). Under this set of presumptions, even a teacher with the best of
intentions is unlikely to explore solutions or answers together with the student.
By contrast, the pedagogical approach teachers experience in KN’s training conceives
of knowledge as an evolving construct and emphasizes that everyone can participate, to
varying degrees, in its generation and application. The most important function of the
training may be to make teachers more aware of the implications of this proposition with
the goal of helping them develop an appreciation for the freedom of thought required of
students who learn through a process of discovery.
The theory behind this pedagogy is that this shift in their perspective on teaching has
many implications for the teaching-learning process. We summarize the main potential
changes here:

1. The training leads teachers to internalize and adopt a pedagogical approach that
is more exploratory in nature, and to develop a taste for active learning that they
then try to inculcate in their students.
15
It is during this moment in the training that one representative teacher commented: “I have been
teaching for 30 years and am only now realizing that all I was ever doing was assisting my pupils to
assimilate information.”

14
 2. It also makes the teachers more inquisitive about the world around them and the
material they are teaching, which makes their own learning, and consequently their
teaching, less rote-based.

3. Taking a more inquiry-based approach to teaching makes the teachers more inter-
ested in the thought-processes of the students, which in turn makes them more
sympathetic to the students and less adversarial.

4. Both because of the closer connection with the students and because of the overall
emphasis on more open learning, the teachers are more likely to encourage students
to explore ways to answer a question on their own (and perhaps explore it with
them) than handing them a textbook answer.

5. Being more open about the sources of knowledge also enables teachers to learn more
from their colleagues, but also share their own knowledge and insights with their
fellow-teachers.

These changes are all multidimensional, and there are surely many other potential
changes beyond those that we have listed here. In particular, we expect that these
changes will have knock-on effects on the families of these students and their village. Our
plan is to pursue many of these questions in our ongoing work on this project. Here, we
focus on the changes in the main educational outcomes, two years and four years after
the beginning of the intervention. We listed many of these as a set of testable hypotheses
for the outcomes we measure in our pre-analysis plan, which we list below:

(Q1) Does the teacher training change teacher pedagogy, attitudes, and effort?

(H1.1) The teacher training will increase the following classrooms outcomes: stu-
dent engagement,the number of students asking questions in class, pedagogical
techniques that engage critical thinking and practical exploration, pedagogical
techniques that facilitate understanding of concepts and deeper learning. It
will decrease the use of corporal punishment and other harsh responses to poor
student behavior.
(H1.2) Trained teachers exhibit higher effort and knowledge of their students.
(H1.3) Teachers will be more willing to learn from others and collaborate with their
peers following the teacher training.

(Q2) Does exposure to trained teachers improve student outcomes?

(H2.1) Traditional learning outcomes are higher for students taught by trained teach-
ers via: Pass-through Rates and standardized test scores(Primary Leaving
Exam — PLE)

15
 (a) Jinja District in Uganda (b) Sub-county in the Study

Figure 1: Study Area

(H2.2) Higher order learning outcomes are higher for students taught by trained teach-
ers.
(H2.3) Scientific capabilities are higher for students taught by trained teachers.
(H2.4) Students taught by trained teachers find more creative uses of local resources.

There are implications of the theory that are not directly addressed in the hypotheses
outlined above such as the degree of active learning pedagogy, teacher inquisitiveness,
and an increase in teacher sympathy towards students. We provide additional results
that explore these aspects of the theory in our discussion in section 4.

3 Experimental Design and Data

Jinja District is located in Uganda’s Eastern region and contains the source of the river
Nile (Figure 1a). The study takes place in one sub-county in Jinja District (now Jinja
City).16
16
We had initially planned to augment this sample with schools from four additional sub-counties and
divisions in Jinja District. We collected baseline data from 53 additional schools to this effect in 2019.
However we did not anticipate at the time that Covid-19 would effectively cause schools in Uganda to be
closed for two full academic years, 2020 and 2021. We released our initial manuscript at the end of 2020
when it seemed schools would be re-opened in 2021 prior to the global concern with the Delta variant of
Covid-19. We managed to extend the final data collected as part of the study to the end of 2022, but
decided not to collect data in the augmented sample in lieu of collecting more long-run outcomes in the
original set of schools in Budondo sub-county, those selected into the study in 2017.

16
 Feb 5 May 4 May 28 Aug 24 Sep 17 Dec 7
Intense T1 Intense T2 Intense T3

T1 visits T2 visits T3 visits

Figure 2: Timing of Teacher Training in 2018

3.1 Experimental Implementation & Timeline

We focus on KN’s use of the PSA materials in its approach to train school teachers, an
initiative that began in late 2014. KN collaborates with the district education officials
(DEO) to invite schools to select 2-5 teachers per school (depending on the size of the
school) to participate in the in-service Teacher Training Program over the course of the
school year.
The training combines three two-week periods of full-time programming during school
holidays with in-service periods and classroom visits during the school year. There are
three school terms during the course of a typical school year in Uganda. In 2018, for
example, term 1 took place from February 5 to May 4, term 2 from May 28 to August
24 and term 3 from September 17 to December 7, the gaps between terms are utilized by
KN for “intense” study of PSA materials as depicted in Figure 2. During these periods,
teachers study PSA materials for roughly 11 days together with a KN trained tutor.
During school terms, PSA tutors try to visit teachers in their classrooms at least once
each month to observe how teachers engage with students and reflect on how to improve
pedagogy and other dimensions of practice.
KN worked with the DEO to invite 50 teachers to participate in the training in each
year of the intervention that we analyze in this paper, inviting teachers from both primary
and upper primary. They expected that some would not be able to attend and were aiming
for roughly 40 teachers participating in each training. In total, 42 teachers attended a
portion of the intense sections of the training in 2018 and 46 attended a portion of the
training in 2019. Each cohort of teachers is dubbed a “unit,” and KN employs two PSA
tutors to engage with each unit of teachers at any given time.17 Intense study periods
were held at a central location near the schools so that practical exercises could link the
PSA materials to local contexts in which lessons can be applied. Notwithstanding that
teachers are not paid to participate (only a transport subsidy is provided, which schools
are requested to refund), training participation rates are high — 33 and 32 teachers
attended the majority of the training in 2018 and 2019 respectively.
17
We also note that, in conversations with KN, it became clear that the tutors they hire do not
necessarily have experience as educators in formal educational settings — indeed, one of the two tutors
of the Budondo unit had not completed secondary school before he started engaging with the PSA
program in the late 2000s.

17
3.2 Sampling
3.2.1 School Sampling Frame

To identify the sample in Budondo subcounty, we took a census of all primary schools
in the summer of 2017 and collected information on the ownership status of each school,
enrollment numbers, the highest class taught, and the number of teachers (among other
measures). Given the study’s focus on upper primary teachers, we limited schools to those
that teach up to the highest primary school level, P7. This left us with 37 remaining
schools of which 2 closed doors before the end of the 2017 school year. We included
the remaining 35 schools in our sample - 15 of which are government-owned. KN began
working with the randomly selected treatment schools in December 2017 after baseline
data was collected.

3.2.2 School Assignment to Treatment

We randomly selected 18 out of 35 Budondo primary schools and offered training op-
portunities only to these schools. Randomization was conducted in November 2017 by
pairwise matching (Bruhn & McKenzie, 2009) on 7 school characteristics drawn from
administrative data.18 First, one school was chosen uniformly at random for treatment.
Then, the remaining 34 schools were greedily matched into pairs according to minimum
Mahalanobis distance among the 7 matching characteristics. Finally, one school from
each of the 17 pairs was chosen uniformly at random for treatment. Importantly, this
procedure allows us to treat each of the 17 pairs, as well as the odd school, as strata in
which one side of the pair was randomly selected into the treatment. Thus, if schools drop
out of the study for any reason, we can drop the pair of schools and proceed unhindered
without losing identification. Three private schools were no longer solvent in 2019, which
caused us to drop them and their paired schools from the study at that point. Thus, the
proceeding analysis takes into account the remaining 29 schools in Budondo sub-county.
The top panel in Table 1 presents summary statistics for school characteristics as
measured in 2017 for the sample of 29 schools whose outcomes we analyze in this paper.
Out of the 29 schools, 15 are government schools and, in 2017, each school employed 13.2
teachers on average for a total of 383 teachers with 248 of them teaching at some level
of upper primary. Of these, 43% teach P6 and P7 in addition to other classes, which
shows that teachers have responsibilities for many classes across the school. Only 24%
of teachers have a post-secondary diploma, as is now mandated of all teachers by MoES
in the newly implemented National Teacher Policy, and 20% participated in in-service
18
Government ownership, student enrollment, number of teachers, fraction of teachers who teach 6th
and 7th year primary school, mean population under 15 years old in feeder villages, mean number of
years that teachers have worked at the school, and fraction of teachers who completed upper secondary
school.

18
trainings prior to 2017.19 Roughly 520 students were enrolled in each school on average in
2017 with an average pupil-teacher ratio of 38.2. This table further breaks down averages
by control and treatment status.
Teachers often transfer in and out of schools in Uganda, especially in private school
settings. Due to poor record keeping around such transfers, we cannot know with precision
the percent of trained teachers in a given school at a given time. However, we do know
that no trained teacher transferred to a control school from a treated school. We also
know that, of the 233 upper primary teachers in the sample in 2019, 127 teach in treated
schools and 49 (38.6%) participated in the training at some point during the two years
of the intervention.

3.2.3 Teacher, Classroom and Student Sample

Outcomes will be measured at the teacher-, student- and classroom-level to address the
research questions above. Due to resource constraints, we limit the analysis specified in
the pre-analysis plan and in this paper to measures from the following three sampling
frames within our sample schools: upper primary teachers, upper primary classrooms,
students in primary six (survey and researcher-administered assessments) and students
sitting for primary leaving exams. The measures we use are best thought of as mea-
sures collected of multiple cross-sections across time, though in cases where teachers are
retained, we have teacher-level panel data. For teacher outcomes, we interview the uni-
verse of upper primary teachers in our school sample. Furthermore, we conduct classroom
observations of all upper primary classes during the morning hours of operation at schools
in the sample (typically between 8:00 AM and 1:00 PM, though this can vary by school
and school day). In 2019, we interviewed 230 teachers in this way and observe 238 P4 to
P6 class periods and 69 P7 class periods across the 29 schools.
In the student survey, we also asked P6 students to identify which of the upper elemen-
tary teachers regularly teach them at their school. When considering teacher behavior
measures constructed using student responses, we restrict the teacher sample to those
teachers who teach P6 students as indicated by at least 90% of surveyed students in a
school — there are 95 such teachers in our sample as indicated by student responses.
Our sampling frame for students who participate in the student survey and assess-
ment is the universe of P6 students in each school.20 We select 10 to 15 such students
(depending on the size of the school) in each school to our sample, balancing across
school-specified performance quintiles and student gender.21 We provide alternative stu-
19
MoES is now establishing a credit-based continuous professional development (CPD) training system
that requires each registered teach to obtain a certain number of CPD credits every two years.
20
We began collecting student data in 2019, therefore we do not have baseline student survey data for
Budondo schools.
21
We use Term 1 records from 2019, provided to us by the schools in the study, to determine the rank
and gender of students in each school. This list serves as our sampling frame as well.

19
dents within quintile-gender strata in case a student is unavailable for interviews and
assessments on the day our team of enumerators visits the school. We select 329 students
to participate in the student survey and researcher assessment in this way across the 29
schools.
We had the opportunity to collect long-run data on a set of indicators in 2022 after the
nearly two-year-long Covid-19 induced school lockdowns in Uganda. Given the unique
circumstances of the intermediary period, we will present and analyze this data in section
5 as long-run follow-up data.

3.3 Instruments
Our study will make use of a data-set compiled of survey data, classroom observations,
school administrative data (including exam results), a researcher-administered assessment
and a series of science fairs organized by the DEO in collaboration with the co-authors.
These data are collected as follows:
Survey data. Survey data is collected by trained enumerators hired by IPA-Uganda.
We work with two types of survey data in this paper: teacher and student. The teacher
and student surveys are conducted at the schools in our sample during the third term
on the same day of a given academic year.22 Student surveys were collected in 2019 and
teacher surveys were collected in both 2017 and 2019.
Classroom observations. Classroom observation data is collected by trained enu-
merators leveraging the Stallings observation tool. This tool provides information on
teacher time-use in class. Observations are in the form of a “snapshot” in which the
enumerator observes what the teacher is doing, the degree of student engagement in the
activity facilitated by the teacher, what non-engaged students are doing, and the mate-
rials used by teachers and students. We calibrate the Stallings tool so that enumerators
collected ten such snapshots for each scheduled class in the school’s timetable.
School administrative data. We attempted to collect school administrative data
from all schools that agreed to participate in the study. These data include information
on teacher and student attendance, student registration, and student exam scores in
each term. In practice, this data proved difficult to collect and its use was limited to
constructing sample frames for student surveys.23
Researcher Administered Student Assessment. In 2019, primary six students
who participated in student interviews also responded to questions on an assessment we
provided them. In 2022, we carried out assessments with all students in the school. We
22
We intended to collect household and community-leader surveys in March-April 2020 at the home-
steads of respondents. However, during the ongoing Covid-19 pandemic, we were forced to pause these
activities.
23
While they are required by law to do so, many schools did not systematically collect data on teacher
and student attendance.

20
describe both assessments in greater detail, including the variables we construct out of
it, in appendix sections C and D.
Science Shows Held in Collaboration with DEO. Each school in Budondo sub-
county was requested to hold a science show in October 2019. During these shows,
student groups in P6 classes, and organized by teachers at the school, presented results
of an experiment they conducted on themes related to the P6 curriculum. Schools formed
79 groups in this way, depending on school size. Two independent judges assessed student
performance during the course of these science shows. We describe these shows in greater
detail in section appendix sections C and E.
Uganda National Examination Bureau Data. Primary 7 “candidates” register
for the Primary Leaving Examination (PLE) in June of each school year. They sit for
the examinations in November and results are provided in January to assess whether a
candidate can attend secondary school. Students from different schools often sit at a
centralized school for exam-taking purposes. Therefore, we collect student registration
numbers from each school after students have registered but prior to students having taken
the PLE. We merge these registration numbers with official records from the Uganda
National Examination Bureau (UNEB) which provide student-level results for Math,
English, Science and Social Studies. We purchase these results in February of each year
relevant to our study.24 We collected 964 records from schools in this manner in 2019,
which were merged with 899 records in the official UNEB data. The remainder (65
students) did not take the PLE for reasons we do not observe. We categorize these
students as failing the PLEwhen analyzing pass rates.
Budondo census data.: Finally, we collect detailed census data in 2018 and 2022 in
Budondo sub-county which includes household rosters with information on age, gender,
highest education attainment, schooling status in 2018, schooling status in 2017 among
other variables. We use this information to provide a richer understanding of the context
around schooling in Budondo in 3.4 and appendix section B.2.
Table 2 provides a timeline for all data used in this study. We have post-treatment
data for instruments and school terms highlighted in blue in this table. These instruments
were used to make observations and measurements in Term 3 of 2019. These outcomes are
analyzed in section 4. The treatment effects in this section are conditional on two cohorts
of teachers participating in the teacher-training from Budondo subcounty. Additionally,
we collected some long-run outcomes on students and teachers in 2022 after schools were
re-opened. We describe these in section 5.
24
Given poor practices around record-keeping in a subset of schools, we were unable to compile the
list of pre-registered PLE takers at baseline for the full set of schools in our study. However, the records
we collect in 2019 contain information for the 29 schools in this paper.

21
3.3.1 Balance Checks and Summary Statistics

Table 1 further provides balance tests for a selection of variables across the set of instru-
ments we use at [Link] balance tests are broken up according to the intersection
of the dataset used to generate the variable and the analyzed unit of observation. For
example, we compare school characteristics at the school-level and teacher characteristics
at the teacher-level, clustering by school. We conduct five specifications in this way: 1)
school characteristics, 2) teacher survey, 3) teacher dyad outcomes, 4) classroom observa-
tion outcomes, 5) student PLE results.25 For each specification, we estimate the following
equation (with indices changing according to the unit of observation):

Treatedis = βxis + ϵis , (1)

where Treatedis indicates the treatment status of individual/classroom snapshot i (or

dyad ij) in school s, xis is the vector of covariates tested for the given unit of observation
and ϵis is the error term clustered at the school level. We wish to understand whether
covariates are significantly different across treatment and control schools, the joint distri-
bution (according to each specification) of these differences is reflected in β. In addition
to presenting p values of each covariate, Table 1 presents joint tests of orthogonality of
these covariates against treatment status, which is reflected in the F score (and the p
value associated with the F score). None of the p values associated with the F scores are
significant at conventional levels.

3.4 Schooling in Budondo Subcounty

The learning crisis is evident in Budondo subcounty. Grade repeat rates are between 12
and 18 percent across every level of the schooling pipeline (as measured between 2017
and 2018 in the household census and summarized in Table B.1). Appendix Figure B.2
shows the distribution of ages enrolled in different levels of schooling in primary school.
Here, we see a very large age variance in schooling level and this variation widens as
students advance in school levels (as expected with a high repeat rate). For example, in
Primary 1, there is significant density throughout ages 6 and 10, suggesting that either
students start Primary 1 quite late, or there is a significant tendency of repeating classes
— most likely, both factors play a role. The age-variation in Primary 4 (the transition
year between primary and middle school) and Primary 7 increases relative to Primary 1.
In Primary 7, there is high density throughout ages 13-18. A likely cause of the patterns
25
For 2017 PLE results, we needed to collect PLE registration numbers ex post. Some schools had lost
these numbers, thus they are only available in 2017 for 26 schools. As we were collecting PLE registration
numbers for schools we added to the study in early 2020, COVID-19 induced lockdowns impeded our
access to schools which prevented us from collecting 2019 PLE registration numbers from each school.
We intend to retrieve these as soon as conditions allow.

22
in these figures is poor teaching quality which we explore further in our analysis below.26
[Vesall: Removed Section Measurement of Outcome Variables and placed it in the
appendix.]

4 Identification and Results

We estimate the intent to treat effect (ITT ) by comparing outcomes across schools invited
to participate in the training (treatment) and other (control) schools. We regress endline
outcome yisp for teacher or student i (or dyad ij) in school s and matched pair p surveyed
by enumerator e on a dummy Tsp , indicating treatment status of school s within its
pair. We include a vector of fixed effects depending on the characteristics of the variable
selected but will always include pair fixed effects, γp , and generally include enumerator
fixed effects when available, ζe .27 The coefficient of interest β1 is the intent to treat effect.
Since treatment is assigned at the school level, we cluster standard errors at the school
level. Formally, the estimating equation for the ITT is equation 2:

yispe = β0 + β1 Tsp + γp + ζe + ϵispe (2)

We pre-committed to using estimators that cohere with the distribution of outcome

variables of interest. The table of all specifications and estimators, linking variables
described in section C to a variant of equation 2, including the unit of observation and
each specification’s associated fixed effects is available in Table G.3.
To test the robustness of our results, we report two sets of p values associated with
the ITT in each specification. First, we report the conventional p value derived from the
use of the estimator specified in Table G.3. That is, we will present p values that test
the null that the ITT is significantly different from zero for the average school. Second,
following Young (2018), we report the p value derived from randomization inference with
at least 1,000 permutations of the treatment variable. That is, we will present p values
that test the sharp null that the ITT is zero for all schools in the sample. These p values
have the added benefit that they allow for tests that are exact — with a distribution that
is known no matter the sample size or characteristics of the error terms. Nevertheless,
when drawing inference from our results we will take results of all tests into consideration.
Since we are testing multiple outcomes for a number of our hypotheses, we control the
false discovery rate (FDR) across outcome measures within each hypothesis using the
Benjamini-Hochberg method detailed in Benjamini & Hochberg (1995). The set of pre-
26
A more thorough description of the schooling and socio-economic context in Budondo sub-county is
available in appendix section B.2
27
Enumerator fixed effects are not available, for example, in administrative data. Nor do they make
sense to use when we construct teacher outcomes using student responses. Variables stemming from
classroom observations of teachers will also include “class” or “grade” fixed effects.

23
specified results, containing both specification and randomization inference p values are
presented in Tables 3 through 5. These tables also present the Benjamini-Hochberg
corrected critical p values at the 5% level and furthermore highlight, using the ± symbol,
whether a given test can be rejected at either the 5% or 10% level after accounting for
the FDR under the critical p values at the 5% and 10% levels respectively. Given the
sensitivity of test statistics to the small sample size, we separately windsorize our data by
chopping off the top and bottom 5% of observations within treatment and control schools
separately. We replicate analysis with this sample and present results in appendix Tables
G.4 through G.6. We note that the statistical significance of all of the results we discuss
below are stronger after windsorizing our data.

4.1 Tests of (H2.1): Traditional Learning Outcomes.

Table 3 presents results associated with our analysis of high-stakes primary leaving exam
scores and pass-through rates to secondary school and from P6 to P7. In the first four
columns, we find that students who attend a school whose teachers were invited to partic-
ipate in KN’s training scored 0.64σ higher in English, 0.56σ higher in Science, and 0.44σ
higher in social studies than students in control schools — all statistically significant at
the 1% level and robust to multiple hypothesis testing (controlling the FDR across all
variables in the table) and randomization inference. Students perform higher in math by
0.12σ, but this result is not statistically significant.28 The Ugandan National Examination
Bureau aggregates scores across all subjects to produce an “aggregate” score. Students
in treated schools perform 0.51σ higher than students in control schools. Additionally,
student performance along this dimension in treated schools has first-order stochastic
dominance over the distribution of performance in control schools as is inferred from Ap-
pendix Figure H.1b, suggesting that the program increases outcomes for every student
across the distribution of student performance.29
The fifth column shows that students in treated schools are 24 percentage points more
likely to pass the PLE examination at division three or above relative to a control school
mean of 51% — this is in the same range as the number reported by UNICEF of 59% of
primary-school leavers transitioning to lower secondary school.30 In the sixth column, we
show that students in P6 in 2018 are 12 percentage points more likely to transition to P7
28
The result is, however, significant upon windsorizing the data as demonstrated in Table G.4, which
shows a treatment effect of 0.19σ after chopping off the top and bottom 5% of the sample within
treatment and control schools separately. Nevertheless, as we discuss in the introduction, of the six PSA
materials KN studies in the teacher training, they have yet to find a way to integrate a book specific to
mathematical capabilities. These books exist in the PSA curriculum and as a result of these findings KN
is now considering what it can do to integrate more of the lessons from these materials into its training.
29
We provide distributions of non-normalized performance across all measures in Table 3 in Figure
H.2, distinguishing between the distribution of performance in control and treatment schools.
30
Students scoring below division 3 are almost never admitted into secondary schools — even students
scoring in division 3 may struggle to gain admittance.

24
than students in control schools. This is relative to a control school mean of 0.81, which is
similar to the range reflected in responses to the household census in 2017 and reported in
appendix Table B.1. Notice that in this final column we include enumerator fixed effects
into the specification because the variable is sourced from the student survey, whereas
PLE data are sourced from UNEB. These are large effects. We use the same table to
extrapolate total increases in expected years of schooling for a P6 student whose chances
of progressing to P7 increase by 12 percentage points and whose chances of progressing
from P7 to secondary increase by 25 percentage points. Such a student participates in
0.59 extra years of schooling, a 5.5% increase over the average student in control schools.
If we extrapolate pass rates to all grades along the pipeline (except for secondary grades),
years of schooling for an average student in treated schools increases by 0.86 years, or
9.6% more than the average student in control schools.

4.2 Tests of (H2.2) through (H2.4): Higher-Order Learning, Sci-

ence Shows and Creativity Outcomes.
In section 2.2, we describe how the pedagogy applied by teachers will also lead students
to seek ways of engaging with knowledge in novel and creative ways — here, we test
those hypotheses. The first two columns of Table 4 measure treatment effects associated
with higher-order learning outcomes. Column one shows that P6 students who attend a
treated school scored 0.73σ higher on questions that measure applied understanding and
0.45σ higher on questions that measure critical thinking, both results of which are robust
to multiple-hypothesis testing and randomization inference. In third and fourth columns
we explore whether students are better able to their critical thinking skills to concrete
practices that solve locally relevant problems. The third column does this by measuring
the treatment effect associated with our aggregate measure of performance in science
shows. We find that students in treated schools score 0.87 points higher relative to a
control school mean of 2.63 — a 33% increase.31 Finally, students score 0.44 points higher
in the creativity measure we constructed relative to their control school counterparts,
suggesting that they are better able to consider ways in which common, every-day, objects
can be used to serve creative and unique ends. While these two tests are robust to multiple
hypothesis testing, they are not, however, robust to randomization inference. Looking at
appendix figure H.3, it appears possible that this may be driven more by the influence
of outliers than any meaningful pattern in the treatment effect that may stray from our
interpretation of the data. We discuss this possibility in further detail in section 4.4
below.
31
We decompose this measure in section 4.4 below to gain further insights.

25
4.3 Tests of (H1.1) to (H1.3): Teacher Pedagogy and Effort Out-
comes
We attribute these improvements in student outcomes to the improved teacher pedagogy.
The first evidence of this is demonstrated in the first column of Table 5. Here, we analyze
the classroom observer’s assessment of the proportion of students who were engaged
in the activity led by the teacher during observed class sessions. Recall, we do not
ask the observer to record a precise percentage of engaged students, which calls for an
ordered logit estimator. We show that for every snapshot observed across P4 through
P6, accounting for class fixed effects, classrooms in treated schools are 39% more likely to
have a larger proportion of engaged students in class than control schools — the control
school average of 4.41 reflects that enumerators indicated that close to half of the students
were engaged in an activity with the teacher on average.
Next, we use teacher measures from the student survey to analyze teacher outcomes.
Each measure in column 2 and 3 should be interpreted as reflecting the percent of a
given teacher’s students who stated that they ask questions in class when they don’t
understand a concept (column 2) and who stated that the teacher has used a form of
corporal punishment (column 3), both reflecting actual incidences spanning the two most
recent school terms. To facilitate descriptions of results, we call the variable in column
2 “student inquisitiveness.” On average, student inquisitiveness is 22% in control schools
as depicted in column two of Table 5. This same number is 6 percentage points higher in
treated schools as measured by a tobit estimator that treats zero as a lower bound and
1 as an upper bound, suggesting that students are more inquisitive in treated schools.
However, they are no less likely to be be subject to corporal punishment in treated
schools. The baseline level of corporal punishment is large — the average teacher was
mentioned by 53% of her students as a teacher who has used a cane to beat the students
in the past. This outcome does not change on average across treatment and control
schools, which we suspect is due in part to how it is measured. Recall, we asked each
student to mention teachers who applied corporal punishment at any point in the last
two school terms. Thus, teachers in treated school who were in the second cohort may
have applied corporal punishment prior to the second training which invites them to
think more carefully about their relationship with their students. Thus, questions that
measure a reduction of an unwanted behavior may not yet manifest in the data at this
time.32
Despite this result, column four suggests that teachers put more effort into learning
intimate details about their students’ lives. We quiz teachers on the knowledge they
32
There is also a strongly entrenched belief, often reinforced by parents, that suggests learning, dis-
cipline and upright conduct is developed through corporal punishment. This belief may be difficult for
teachers to shake off in the short-run, however the evolution of this belief will be a subject of future
research.

26
have regarding the school attendance and living circumstances (relationship with the
guardian they stay with) of a randomly selected subset of students. When matching
student and teacher responses, we find that the average teacher in control schools provides
correct responses 62% of the time. Teachers in treatment schools are 10 percentage
points more likely to provide a correct response using a Tobit estimator that censors
observations below 0 and above 1. The distribution of outcomes along these lines is also
worth mentioning. Appendix Figure H.4 provides a distribution of all outcome measures.
Column five reports treatment effects associated with teachers’ desire to learn, as
measured through a change in interactions with colleagues with the objective of learning
from one another. The measure reflects the degree of intensity of teacher learning within
dyad with 3 being the most intense relationship and 0 indicating no such relationship. We
can divide the mean and coefficient by 3 to interpret results in percentage point terms.
Teachers in treatment schools are 9 percentage points more likely to form a learning
relationship with their colleague relative to a control school mean of 60%.33
Results in columns four and five support the view that the emphasis on metacognition
in the training facilitates a desire to learn how to overcome a given challenge rather than
take it as given. This is reflected in their willingness to learn more about about their
students’ lives and their willingness to learn with their colleagues about how to improve
their pedagogy, consistent with results in columns 4 and 5.

4.4 Additional Discussion

The previous set of results contain pre-specified analysis outlined in our pre-analysis plan.
Below, we explore the data further to provide additional insights into the nature of change
created by the intervention.
Science Shows. We go into further detail and analyze each of the outcomes asso-
ciated with the science shows. Science show judges assessed student groups based on
12 outcome variables that we place into five categories: 1) measures that assess pupils’
ability to frame the question they are investigating (Framing); 2) measures that assess
the quality of the pupils’ experiment (Experiment); 3) Measures that assess whether the
experiment and hypotheses are well-aligned (Hypothesis); 4) student groups’ ability to
systematically measure and log outcome variables associated with the experiments and
hypotheses (Measurement); and 5) students’ ability to independently and accurately ar-
ticulate the conclusions they drew from the experiment (Articulating). Each of the twelve
measures are outlined in detail in Table E.1.
Appendix Figure H.5 displays the distribution of each of the 12 outcome variables and
shows a large degree of bunching of outcome variables on the lowest measure — 1 —
33
Recall that this is unlikely to be due to experimenter demand effects because we verify the existence
of a collaborative relationship between two teachers through the response of both teachers within the
pair.

27
across many of the outcomes. As a result, we use Tobit estimators to analyze treatment
effects, setting one as the lower bound and 10 as the upper bound across all specifications.
Table 6 estimates the treatment effect of each of theses 12 measures.
Apart from outcome measures reflecting the students’ ability to frame their research
question, we find significant impacts across all categories of outcome measures.34 In other
words, there is reason to believe that we observe positive treatment effects in students’
abilities to design an experiment, describe and analyze hypotheses, measure outcome
variables related to the hypothesis, and clearly articulate their arguments. The most
robust outcomes are variables in the latter two categories, whose randomization inference
p values suggest significant impact at the 10% level, accounting for the false discovery
rate within outcome variable category.
Furthermore, these effects are large. Take, for instance, treatment effects of 2.12 and
2.37 in columns (11) and (12) respectively. The control school means for these outcomes
is 3.20 and 2.65. Therefore, on average, treated schools perform 66% and 89% higher in
these two categories respectively.
We also display a sample of the text summaries articulated by the science show judges
for the 10 highest performing student groups and 10 median performing student groups
in Appendix Tables E.2 and E.3 to provide qualitative insights demonstrating the clear
difference between a science project receiving a high score and low score. It is clear from
reading these summaries that our elicited measures had their intended effect of measuring
an important dimension of student learning — applied critical thinking and reasoning —
through science shows.
Experimenter Demand and Hawthorne Effects. One natural concern of a
teacher-training intervention is that teachers in trained schools have been conditioned by
the training — therefore, their subjective responses to survey questions will reflect that
they know what the researchers are looking for. Notice however that none of the vari-
ables constructed in our analysis rely on a single teacher’s subjective response. Even our
main indicators of teacher behavior are provided primarily an outside observer (column
1 of Table 5) or their own (columns 2 and 3 of Table 5). The two measures constructed
using teacher responses in columns 4 and 5 of Table 5 are also conditioned by either their
students (column 4) or their colleagues (column 5). As a result, we do not believe that
there should be any issue concerning experimenter demand effects.
Hawthorne effects are similarly unlikely since Table 3 provides results from a stan-
dardized test score administered by a government institution. Students across treatment
and control schools were unlikely to change their effort because their results would be
used in the study — passing the PLE is already a high-stakes exercise requiring most of
34
In retrospect, the letter we crafted together with the Jinja DEO framed the research exercise for
the students and teachers, so it is not surprising that there is no treatment effect along the “framing”
dimension.

28
a child’s attention.
Active Learning Pedagogy. Among the arguments we make in our theory in Section
2.2, we claim that teachers who go through the training are better able to implement
active learning pedagogy. Active learning pedagogy involves students in an experiential
learning process similar to the ones that teachers in KN’s training participate in. We use
teacher survey responses and classroom observations to show that teachers are, in fact,
utilizing pedagogical principles that suggest a more active learning environment. We do
this in two ways.
First, in the teacher survey we ask teachers whose schools have a garden plot how
many days they spent either in the clearing, planting, weeding, or harvesting stages of
growing crops. There is a positive, but mostly insignificant difference between treated and
control schools along this measure. However, when we further ask teachers to indicate
how many days they spend on the garden plot working with their students, we see large
and significant differences. Using a Tobit estimator censored below by zero, Table G.8
shows that teachers in treated schools spend 1.93, 2.80, 2.52, and 2.14 more days engaged
in each of the four activities, respectively, than their control school counterparts.
Second, a teacher adept at implementing active learning pedagogy can engage learn-
ers in learning activities even when she is not around. This is often the case in P7,
where students are tasked with “revising” course content from previous years of school
in preparation for the PLE. We observed classrooms of PLE takers two months prior to
the exam in 2019 and noted that 20.1% of the time, teachers were “out of the classroom”
during our observations.35 However, just because teachers are not around does not mean
that students are not engaged in a learning activity. To check whether this is the case,
We further observe what activity the students are engaged in when the teacher is not
conducting an activity with pupils. There are 168 instances of classroom observations
where this is the case in P7. In these cases, the classroom observer can indicate whether
the students are distracted by indicating that learners are “not engaged” or “socializing,”
or that learners are engaging in the following learning activity categories “assignment
or classwork,” “copying,” or other.36 Figure H.6 shows the distribution of activities in
P7 classrooms when teachers are not around, where we see that learners are engaged in
learning activities 75.2% percent of the time in treatment schools while only engaged in
learning activities 52.7% of the time in control schools.
Teachers are more inquisitive. We also argue that the teacher training makes
teachers more inquisitive and their teaching becomes less rote-based. The latter is re-
flected in the improvements in student learning in treatment schools along dimensions
of higher-order learning. We present additional evidence here that teachers are more
35
This as compared to 12.7% of teachers in P4 to P6 who were “out of the classroom.”
36
We compile “other learning activities” using responses that suggest the learners are “Reading out
loud,” engaged in “Explanation/lecture,” “Interactive demonstration,” “Question and answer/ discussion,”
or Practice and drill.

29
inquisitive. In the teacher survey, we ask teachers whether they have ever been asked a
question by their pupils that they don’t know the answer to (1 if yes, 0 if no). Table G.9
shows that there is no significant difference across treatment and control in response to
this question. However, conditional on responding yes, we further ask teachers to share
the most common way they respond to such a question. Teachers respond freely and we
record their responses according whether they carry out further research (inquisitive) or
if they ignore the question (ignore).37 We find that teachers in treatment schools are 10
percentage points more likely to state that they are inquisitive and 6 percentage points
less likely to ignore the student.
We do not believe these answers to be driven by experimenter demand. In order to
suggest this, we interact school-level average teacher inquisitiveness with treatment and
regress the interaction effect against science show outcomes in Table G.10. Teacher in-
quisitiveness is negatively correlated with performance in control schools, but positively
correlated in treatment schools across all categories of outcomes, a result we depict graph-
ically in Figure H.7. What this result shows is that teachers who present themselves as
inquisitive are more able to improve science show performance in treated schools as op-
posed to control schools. This suggests that teachers in treated schools know how to use
their inquisitiveness to promote student exploration. Inquisitiveness may be a correlated
feature of a broader capability that enables teachers to facilitate a process of exploration
that leads to learning.
Teachers are more sympathetic and less adverserial. Our pre-specified result
on corporal punishment does not show much change across treatment and control school
outcomes. However, there is still reason to believe that teachers become more sympathetic
and less adverserial towards their students. In the teacher survey, we ask a series of yes
and no responses in this regard. First, we ask whether students are free to disagree with
their teachers about concepts they are being taught. Second, we ask if students are free to
decide how to spend time during a lesson. Third, we ask teachers whether they agree with
the statement that they are “learn as much from my pupils as they learn from me” (on a
five-point likert scale, we code a response of “Strongly Agree” — the strongest agreement
response — as 1 and 0 otherwise). We show that teachers are 50% more likely to agree
with the first statement and 76% more likely to agree with the third, both significant at
five percent and one percent levels, respectively. An index that averages across all three
variables is also significinat and positive at the 1% confidence level.
We also ask teachers how they respond to pupils who mis-behave in class. Teachers
can select from a menu of strategies and we code their responses as “gentle” or “harsh.”
Table G.12 shows that teachers in treatment schools are 56% more likely to provide gentle
37
For the former, teachers response is either “Tell the student you do not know the answer and will do
research to find the right answer” or “suggest to the student how to investigate the answer on his or her
own.” For the latter, teachers response is either “Ignore the student and do not respond” or “Provide the
best response you can provide.”

30
responses and 43% less likely to provide harsh responses.
Again, the above results could be consistent with experimenter demand effects. Anec-
dotal evidence suggests that as teachers learn more about their students’ lives, they often
find that student misbehavior is a result of student hunger. To address this, they often
ramp up school feeding programs. If this is the case, then students should be less hungry
at school in treatment schools, a hypothesis we test in appendix Table G.13. We ask
students whether they ate a meal at school yesterday and whether they agree with the
statement “I am usually hungry during the school day” (five point likert scale). We show
that students in treated schools are 17 percentage points more likely to eat a meal at
school the previous school day and are more likely to disagree with the statement about
hunger at school.

5 Long-Run Outcomes Measured in 2022

As described earlier, we had originally hoped to augment the sample by 53 additional
schools in 2020 and study end-line outcomes in 2021, where we would also study long-
term follow up results in the original set of schools in our study. This plan was thwarted
by the Covid-19 pandemic, which resulted in a two-year school lockdown in Uganda.
Schools were not re-opened until February 2022 (only “candidate classes” were allowed to
meet in 2021), one additional school closed entirely (causing us to drop two schools from
the study), the approach to training was constrained by an inability to carry out tutor
visits to schools, and many teachers chose not to return to their previous posts. Indeed,
of the 388 teachers from these schools in 2019, only 236 (61%) remained in 2022. The
total number of teachers reduced from 388 to 340 — consistent with anecdotes around
the country that seemed to suggest a general drop in the teacher population due to post-
covid exit from the profession. This also affected the teachers trained prior to 2020 in
similar proportion — such teachers made up 23% of all teachers in treated schools in
2022. Kimanya Ngeyo continued with the study portion of their training, studying in
small groups of teachers to ensure adequate social distancing. However, since schools
were closed, tutors were unable to follow up with teachers in classrooms, reflecting on
teacher practices for the two years of government-imposed school closures. With this
caveat in mind, we still note that by the time data was collected in 2022 58% of teachers
found in treated schools in 2022 (112 of the 192) had participated in training of some kind
— though only 55 of these participated in the training prior to 2020 under the original
model.

31
5.1 Persistence and Enrollment in Secondary Higher in Treated
Schools
In 2022, we collected data on all school-going children in our study sample using a house-
hold census. We briefly interviewed the most senior member of the household, requested
that they list all household members and their school-going status for each year since
2018. This allows us to assign a treatment status to children who attended schools in
our sample in 2018 in order to estimate measures on long-run enrollment in primary or
secondary school as well as persistence through school. We can study two questions: did
the intervention affect student enrollment in any school after four years? did the inter-
vention increase secondary enrollment rates after four years of the intervention? did the
intervention increase persistence, the observation of a student passing through successive
classes, after four years?
Specifically, we define an outcome variable related to each of these variables in 2022
and assign students to treatment based on the school they were enrolled in in 2018.
For the first question, we define an outcome variable using a dummy equal to one if a
student is enrolled in any kind of school in 2022. For the second question, we define an
outcome variable equal to one if a student is enrolled in secondary school in 2022 (we
only define this for students in grades 5 or higher in 2018). For the third question, we
take the difference between the student’s grade level in 2022 and 2018. If a student was
not enrolled in school in 2022, we replace missing values with the most recent grade in
which the student was enrolled.
Table 7 presents outcomes in three panels, one for each question. Column one ag-
gregates all observations of students enrolled in sample primary schools in 2018 and the
remaining seven collumns disaggregate by the student’s 2018 grade-level. We find strong
and large positive effects across all relevant specifications. The only meaningful null re-
sults are in Panel A, which analyzes enrollment in any school, for students enrolled in
P1 or P2 in 2018. Notice that the control-school average of whether a student continues
to be enrolled in 2022 for this subset is at least 0.97, consistent with a more general ob-
servation that almost all primary-school aged children enroll in universal primary school.
Consistent with our PLE results from 2019, we find that students enrolled in treated
primary schools are 10 percentage points more likely to be enrolled in a secondary school
in 2022. Finally, we find that treated students across all grades, including lower primary,
have passed through more grades relative to their control school counter-parts. The effect
seems to increase for students enrolled in higher grades in 2018 — from 0.09 additional
grades for treated P1 students to 0.25 additional grades for treated P7 students.
It is worth noting that this measure of treatment assignment does not reflect any type
of treatment intensity at the student-level. For example, students assigned to treated
schools in 2018 could have transferred to control schools or vice versa. Thus, the treatment

32
effects described above are likely lower bounds for the effect of the treatment on the
specified outcomes.

5.2 Secondary School Performance when Treated in Primary

We are interested in understanding whether the ability to “learn how to learn” transfers
to the students — one indicator of this is whether treatment effects persist to learning
outcomes even after students’ exposure to treated teachers ends. In 2022, one of our co-
authors initiated a RCT within secondary schools, where he conducted a student learning
assessment in lower secondary (S1 to S4, corresponding to grades 8 through 11 in the
U.S. context). We use these data to analyze a long-run effect on student learning.
Following a fuzzy match procedure, we identified 528 students in the secondary as-
sessment data who we could identify in the 2022 household census with a high degree
of confidence (following manual checks). Of these, 320 attended primary schools from
our sample at some point between 2018 and 2021 — 189 students attended a treated
school in 2018 and 131 did not (only 9 of the 131 attended a treated school after 2018
before entering secondary school). Furthermore, 165 of the 320 were in S1, 89 in S2, 63
in S3 and 3 in S4. Consistent with our effects on persistence, a larger share of treated
students identified in the assessment data are in later grades of secondary (53% in S2 and
above) relative to control students (42%), which increases our confidence in the matches
we found. Questions for the assessment were compiled from past Trends in International
Mathematics and Science Study (TIMSS) assessments and select reading comprehension
questions tailored to the local context. Additionally, we tested students on the precise
use of words that teachers engage with during the course of the teacher training — activ-
ities include juxtaposing “information” with “concept,” “science” with “technology,” and
“knowing” with “understanding.” Greater consistency in differentiating these words across
sentences leads to higher performance.
The results suggest a fairly large and persistent effect on student learning even after
student exposure to treated primary teachers ends. On average, students who attended
primary schools in 2018 outperformed students who attended control schools by 0.33
standard deviations on an index that aggregates all assessment questions into a single
outcome. Effects are positive across all subjects (ranging from 0.08 to 0.24 SD), including
Biology, Chemistry, Physics, Math, Reading and Language. Effects are statistically sig-
nificant for Math (0.14 SD) and Reading (0.24 SD) at the 5% and 10% level respectively.

5.3 Teacher and Student Assessments in 2022

Finally, in 2022 we were able to administer and digitize written student assessments to
all upper primary students in our student sample to better understand the distribution
of learning outcomes across grade levels. We also took this opportunity to administer

33
subject-specific assessments to teachers to see whether the training affected teacher con-
tent knowledge. We were unfortunately not able to administer assessments to teachers in
all schools since many teachers were too busy to take our exams. We nevertheless present
results for all teachers who were available.
We present student results in table 9. We continue to provide results on critical think-
ing (in Panel D, similar to Table D.1) but break down additional test questions into their
respective subject categories of math (Panel A), English (Panel B) and Science (Panel
C). Student learning outcomes remain positive, significant and large for all categories of
outcomes, including mathematics in grades four through six in Panel A. Treatment effects
on math outcomes are largest in P4 (0.59 standard deviations) and equal to 0.43 standard
deviations in P5 and P6. We suspect that these effects are driven by improvements in
pedagogy in earlier grades of schooling. Consistent with our earlier result in column 4
of Table 5, it is possible that teachers are more aware of their students’ mathematical
abilities at younger ages. Given the cumulative nature of math skills, and the importance
of having a strong mathematical foundation for later math outcomes, we suspect that
improved pedagogy in earlier grade levels results in improved outcomes in later grades as
well.38
Teacher learning outcomes are presented in table 10. We note that columns 1 and 2
reflect teacher responses to questions related to subject proficiency and critical thinking —
we include teacher-subject fixed effects for column 1 since the questions differed depending
on the subject that the teacher taught. The critical thinking questions are similar to those
we give to primary students.
We do not see a significant effect in either teacher content knowledge or critical think-
ing. This subjects that teachers across treated and control schools have roughly similar
critical thinking skills and knowledge of the content they are teaching — at the very least,
they demonstrate that they know the content they should be teaching in both types of
schools.
When it comes to teachers’ nuanced use of language, we see large and significant
treatment effects between 0.34 and 0.36 standard deviations relative to control school
teachers. Treatment school teachers are able to more precisely and consistently use
the word “understand” or “know” when presented with sentences in which the difference
between the use of one or the other word is a bit nuanced. A similar effect is observed
in teachers’ use of the words “science” and “technology.” While these are concepts that
are directly explored in the teacher training material, the word “tender” is a more neutral
term. Here, teachers are presented with the statement “the principal treated children
with tenderness.” and were asked to determine whether an additional set of sentences
38
These results extend to foundational learning outcomes in lower primary as well. We administered
oral assessments in literacy and numeracy to ten lower primary students per class-school pair using a
common assessment used in Uganda by UWEZO TWAWEZA (UWEZO, 2016). Here, again we find
positive effects that are large and significant at P1. Results available upon request.

34
used the term “tender” in similar fashion (e.g., “the meat was very tender” is not a similar
use of this term). Again, we see large treatment effects with regard to this measure on
the order of 0.25 standard deviations.

6 Cost Effectiveness Analysis

To gauge the cost-effectiveness of our learning outcomes, we use Learning-Adjusted Years
of Schooling (LAYS) by following the procedure outlined by Angrist et al. (2020). LAYS
effectively controls for differences in education quality across countries when measuring
cost-effectiveness — it seeks to render an absolute measure of cost-effectiveness for edu-
cation interventions. For example, 10 LAYS per $100 reflects an increase of 10 years of
high quality education per $100 spent. Angrist et al. (2020) use 3 LAYS per $100 as a
threshold for “the most cost-effective programmes”; the deworming intervention in Miguel
& Kremer (2004), for example, is estimated to have 5.68 LAYS per $100.
To calculate LAYS, we divide the treatment effect on aggregate PLE exam scores for
2019 by the benchmark high-performance learning rate of 0.8σ per year.39 This gives the
improvement in learning outcomes expressed in terms of equivalent high-quality schooling
years, which Angrist et al. (2020) define as learning-adjusted years of schooling (LAYS).
We multiply this number by the number of students exposed to trained teachers. As we
discuss in section 2 above, KN trains teachers across all levels of primary schooling, though
we have only measured treatment effects on learning outcomes for P6 and P7 students.
The treatment effect on higher order learning outcomes using questions similar to the
PLE questions (SIPRO, 2019), and without any preparation or prior warning, suggests
larger treatment effects for P6 students relative to their P7 counterparts, perhaps because
they were exposed to trained teachers at an earlier age. Thus, there is no reason to believe
that students in lower levels somehow benefit less from the teacher training, so we do
not differentiate between different grade levels and implicitly assume that the effect on
lower-level students is the same as that on the students taking PLE exams in 2019.40
Dividing this number by per-student costs, we obtain 9.62 LAYS per $100.41 This
number is ranked in the top 5% of all interventions currently measured using the LAYS
method in Angrist et al. (2020). Furthermore, the four interventions ranked higher are
39
We use 0.51σ, which measures the treatment effect for aggregate PLE exam scores (summing over
all subjects). Results available upon request, but also note the number is similar to the one presented
in Figure H.1b.
40
We could, alternatively, divide costs by the number of teachers in the training who teach P6 and P7
and re-calculate LAYS. Doing this does not change our LAYS figure dramatically.
41
Overall annual cost of the intervention for a single cohort of teachers in Budondo is $27,682. This
sum covers the original 18 treated schools with 8,541 students as measured in 2017: omitting closed
schools, it decreases to 8,361 which is the number we use in our analysis. The treatment effect measured
in this paper assumed 1 year of training for each of the 2 cohorts of teachers, so we multiply the annual
cost by 2 to obtain our final per-student cost.

35
largely driven by their low costs — learning gains are likely to plateau in these settings
without an improvement in the quality of teaching.42
Finally, we note that this estimate is likely a lower bound. We only use improvements
along a single dimension of learning measures — specifically, aggregate standardized test
score results — however, it is clear from our analysis that there are many dimensions
through which learning is improved in this setting. It does not appear that LAYS, or
any other cost-effectiveness measure to date, has adopted an approach that captures the
possible multi-dimensionality of learning, so we refrain from introducing one ourselves at
this time. This is particularly relevant when cost-effectiveness analysis is implemented
on programs that only focus on a single learning outcome, such as reading, measured via
literacy assessments. The treatment effect on aggregate PLE scores already reduces the
dimensionality of learning across four PLE subjects while also leaving out improvements
in critical and scientific thinking which are likely to contain learning dimensions that are
not correlated with PLE results.

7 Conclusions
A novel intervention aimed at helping teachers learn how to learn has large treatment
effects on the quality of student learning along many dimensions — standardized test
scores improve as do our measures of higher order learning, scientific capabilities and stu-
dent creativity. We outline a conceptual framework that proposes a way of understanding
these changes in terms of changes in the teacher. Our evidence shows that teachers adjust
their pedagogy and relationships with students by teaching in a more engaging manner,
by creating an environment in class in which students ask more questions, by being more
aware of details of their students’ lives and by increasing the degree to which they learn
from their colleagues.
In the paper, we contrast the teacher’s new pedagogy, the one promoted by the PSA
materials, to her old one, what is sometimes described as “banking pedagogy”. This new
pedagogy is also different from the kind of highly structured pedagogy which has shown
some success in developing countries (Piper et al., 2014, Banerjee et al., 2017, Piper et
al., 2018, Muralidharan et al., 2019). Structured pedagogy is based on the idea that
for learning certain specific skills, it is important that the student goes through a set of
fixed steps that are universally applicable. While this is likely to be an important feature
of a high-quality educational experience, it is clear that it cannot be a substitute for
promoting a deeper understanding and that this part of learning probably becomes more
important as students move to higher grades. To get to this, teachers need to get students
42
Angrist et al. (2020) Describe these four interventions as “providing information on the returns to
schooling in Madagascar (Nguyen, 2008), creating school links to village councils in Indonesia (Pradhan
et al., 2014), and grouping students by ability level in Kenya (Duflo et al., 2011).”

36
to reflect on questions that do not, or sometimes can not, have a textbook answer. To
guide a learner through such an experience requires skill and human interaction through
a mode that is unlikely to be replaced by a script. Our evidence suggests that it requires
that the teacher herself possesses the capabilities required of someone willing to learn
about the world around her.
Anecdotal reports from KN tutors as well as our own evidence also suggests that
this transition to a new pedagogy moves hand in hand with changes in teacher-student
relationships. As teachers begin to ask more open-ended questions, they start to realize
that students are reluctant to speak up. After all, the students were accustomed to having
a teacher feed them the information they absorbed in hierarchical fashion. Over time, the
teacher engages the pupil in new ways, often by engaging them outside of class in village
settings and becoming more aware of the challenges pupils face at home. This results in
a softening of their relationship, which allows the pupils to be more open in class and
facilitates the teacher’s application of the new pedagogy.
Our conceptual framework suggests that a key difference between treated and con-
trol schools is that students in treated schools were taught in ways that increasingly
exposed them to conceptual understandings of content as opposed to learning that pro-
motes memorization and rote learning. Given the emphasis on measuring rote learning
in standardized tests in Uganda (Burdett, 2017), we expected that the increase in PLE
scores would be quite modest and were surprised to learn of the magnitude of the effect
after two years of the training — one that increased the effective (division 3) pass rate
by 25 percentage points (50% above control school equivalent).
The size of the impact hints at the potentially important relationship between con-
ceptual mappings and absorption of information. There is growing research suggesting
that people are not always attentive to information they are exposed to (e.g., Hanna et
al. (2014)), or they ascribe incorrect meanings to information signals they receive (e.g.,
Michelson et al. (2021)). The proposed policy response is often to improve on the informa-
tion signals (e.g., by “nudging” or labeling) to prevent mistakes in judgement. However,
this presupposes that proper judgement cannot be taught and learned. Our results sug-
gest that a better understanding of underlying concepts within a field of inquiry might
promote higher absorption of information by an individual.43
We hope this paper will promote further research into “learning to learn” teacher
training models beyond primary school settings in Uganda. Indeed, we intend to study
a modified teacher training program offered by KN in secondary schools as soon as con-
ditions allow.44 Moreover we imagine taking this set of ideas beyond the school setting.
For example we want to ask whether something similar could be applied to the relation-
43
For example– as explored in one of the PSA materials– would an increase in a farmer’s conceptual
understanding of the ways in which genetic and environmental factors influence crop growth affect their
ability to absorb information accurately?
44
This paper was written in 2020 during the height of the Covid-19 pandemic.

37
ship between agricultural extension workers and smallholder farmers. These field officers
apparently often operate under the assumption that smallholders lack knowledge of new
technologies and approaches and it is their job to “explicate” this knowledge to them.45
What would happen if instead the extension workers learned how to learn and extended
this approach to learning to smallholder partners through processes of action-research?
More broadly, we note that this orientation to learning contributes to a growing con-
versation on how our conception of social reality shapes behaviors. Kranton (2019) argues
that the assumptions laden in the policy-environment we act in often shapes the behaviors
of human beings.46 In this paper, we provide evidence that when teachers’ assumption
of knowledge changes in a way that is reflected in their attitudes and in their practice,
then behaviors also change — this is absent of any change in incentives for the teacher.
We note, too, that this paper has left out an important feature of KN’s teacher training:
participants in the PSA program are encouraged to view themselves as “promoters of
community well-being” and to seek ways of applying their approaches to learning to
advance processes of community development. We plan to study these community level
impacts in the future.

45
We discuss the meaning of this term in section 2.2.
46
There are many examples in Bowles (2017), the book Kranton (2019) is reviewing, that show, for
example, that monetary incentives crowd out embedded norms and moral codes when they do not
consider the context people may have previously used to make decisions.

38
Statistics Summary Statistics Balance Tests
All Control Treated
Mean Sd Mean Sd N Mean Sd N β p value

School Characteristics
Gov vs. Private 0.48 0.51 0.43 0.51 14 0.53 0.52 15 0.15 0.67
N Teachers 13.21 4.05 11.79 2.42 14 14.53 4.84 15 0.02 0.83
Percent P6 and P7 Teachers 0.43 0.08 0.45 0.07 14 0.41 0.08 15 -0.45 0.79
Percent Teachers Have Diploma 0.24 0.23 0.27 0.24 14 0.20 0.21 15 -0.10 0.91
Percent Any In-Service Training < 2017 0.21 0.18 0.22 0.18 14 0.20 0.19 15 -0.17 0.78
Enrolment 520.28 289.42 487.64 268.35 14 550.73 314.01 15 0.00 0.73
Pupil-Teacher Ratio 38.20 14.78 40.12 17.21 14 36.41 12.43 15 -0.01 0.71
Specification Statistics F Score (p value) 1.74 (0.14)
Clusters 29
Teacher Survey Outcomes
Teacher Gender 1.35 0.48 1.35 0.48 113 1.34 0.48 132 0.00 0.99
Attended Training Last Year 0.26 0.44 0.28 0.45 113 0.23 0.43 132 -0.06 0.47
Farm Land 0.68 0.47 0.73 0.45 113 0.68 0.47 132 -0.08 0.60
Boys Better Pupils 3.19 1.52 3.32 1.51 113 3.11 1.55 132 -0.02 0.46
HT Listens 1.31 0.69 1.40 0.83 100 1.24 0.55 119 -0.07 0.30
How Satisfied 2.06 0.86 2.07 0.89 113 2.07 0.85 132 -0.01 0.83
Connect Family 1.86 0.85 1.94 0.97 113 1.84 0.74 132 -0.02 0.62
Specification Statistics F Score (p value) 0.51 (0.82)
Clusters 29
Teacher Dyad Outcomes
ij Speak About Classroom Management 0.84 0.37 0.84 0.36 1,280 0.83 0.38 2,057 0.03 0.45
ij Visit Each Others Classrooms to Learn 0.79 0.41 0.81 0.39 1,280 0.77 0.42 2,057 -0.07 0.30
ij Plan Classroom Activities Together 0.65 0.48 0.68 0.47 1,280 0.63 0.48 2,057 -0.04 0.55
Specification Statistics F Score (p value) 0.77 (0.52)
Clusters 29
Classroom Observation Outcomes
Share of Engaged Pupils 4.04 2.18 3.96 2.17 1,611 4.16 2.16 1,549 0.02 0.15
Activity: Q and A 0.16 0.36 0.16 0.37 1,612 0.15 0.35 1,551 -0.05 0.30
Activity: Practice and Drill 0.03 0.17 0.03 0.17 1,612 0.03 0.17 1,551 -0.03 0.78
Teacher Out of Class 0.24 0.43 0.24 0.42 1,612 0.24 0.43 1,551 0.07 0.44
Materials: None 0.50 0.50 0.50 0.50 1,598 0.49 0.50 1,536 0.05 0.48
Materials: Textbooks 0.04 0.20 0.03 0.17 1,598 0.05 0.22 1,536 0.19 0.11
Materials: Blackboard 0.36 0.48 0.35 0.48 1,598 0.36 0.48 1,536 0.05 0.48
Specification Statistics F Score (p value) 1.27 (0.30)
Clusters 29
Student PLE Outcomes
PLE English 7.02 1.69 7.39 1.55 364 6.74 1.73 486 -0.08 0.05
PLE Science 6.55 1.79 6.88 1.80 364 6.31 1.75 486 -0.04 0.47
PLE Maths 7.08 1.51 7.25 1.50 364 6.96 1.50 486 0.02 0.24
PLE Social Studies 5.59 1.75 5.80 1.69 364 5.43 1.79 486 0.05 0.31
Specification Statistics F Score (p value) 1.86 (0.15)
Clusters 27

Notes: This table reflects balance tests of teacher, teacher dyad, and classroom outcomes and covariates. We present summary statistics of each
measure, displaying means and standard deviations for the whole sample “All,” the sample of teachers/teacher dyads/classroom observations in control
schools “Control,” and the sample of teachers/teacher dyads/classroom observations in treatment schools “Treated” (the latter two also include number
of observations). We have baseline data for all schools for variables that do not use the Student Survey in their construction. For school admin data,
including teacher attendance, we were in the process of collecting baseline data outside of Budondo sub-county in March 2020 when the Covid-19
related lockdowns started — we exclude these variables form balance tests here. Balance tests reflect an OLS regression with the specification
Treatedis = βX + ϵis where i represents student, s represents school, X represents the vector of covariates in the rows of this table, β is the vector of
coefficients associated with each covariate and ϵis is the error term clustered at the school level. Each specification is run for each data set separately,
datasets are separated by the horizontal lines in the table. The F Score and number of clusters is reported for each specification. The F Score’s p
value (in parentheses) reports results of the null hypothesis test that coefficients are jointly orthogonal within a given specification.

Table 1: Summary Statistics and Balance Table

39
 Year of Data Collection: 2017 2018 2019 2020 2020-21 2022
School Term: T3 T1 T2 T3 T1 T2 T3 T1 Remaining Terms T1 T2 T3
School Characteristics
Budondo Sub-county
Teacher Survey
Teacher Assessment
Classroom Observations
Student Survey
Student Assessment
Science Show COVID-19
Administrative Data Subset of Schools Subset of Schools Closings
PLE Registration
UNEB Data
Household Census
Intervention Cohort 1 Cohort 2 Cohort 3a Cohort 4
Used to measure treatment effect after two cohorts of teachers participated in the training.
Cohorts of teachers participating in KN training
40

Baseline Data Collected and analyzed in balance table (except for household census).
Long run outcome data.
Notes: a Cohort 3 did not participate in the training as designed because groups studied in small groups to promote social distancing and
tutors were not able to accompany teachers in schools to advance teacher “reflections” based on classroom observations (schools were closed
and did not accept visitors).

Table 2: Timeline of Data Collection in the Study

 Hypothesis: (H2.1): Traditional Learning Outcomes
Primary Leaving Exam Results 2019 Passthrough 2018-2019
Outcome Variable: English Science Mathematics Social Studies PLE Pass P6 to P7
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
Treatment (ITT ) 0.64 0.56 0.12 0.44 0.24 0.12∗∗∗
(0.13) (0.15) (0.15) (0.12) (0.07) (0.04)
H0 : ITT = 0
p value [0.00]±± [0.00]±± [0.43] [0.00]±± [0.00]±± [0.01]±±
RI p value [0.01]±± [0.03]±± [0.55] [0.02]± [0.04]± [0.06]±
BH Critical p value (5%) [0.01] [0.03] [0.05] [0.02] [0.03] [0.04]
Pair FE Yes Yes Yes Yes Yes Yes
Enum FE No No No No No Yes
Source of Data UNEB UNEB UNEB UNEB UNEB Student Survey
Unit of Observation Student Student Student Student Student Student
Standardized Variable Yes Yes Yes Yes No No
Range of Outcome Variable [-1.07, 3.47] [-1.29, 2.96] [-1.24, 3.49] [-1.54, 2.68] {0,1} {0,1}
Control School Mean 0.00 0.00 0.00 0.00 0.51 0.81
Clusters 29 29 29 29 29 29
41

Observations 899 899 899 899 964 328

Estimator OLS OLS OLS OLS OLS OLS

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Randomization Inference using 1,000 permutations of
school-level treatment indicator within matched-pair strata.
Variable Descriptions: Units of observation in all columns are students, though the data sets and, hence, number of observations differ. The first
through fifth columns utilize official test scores for the primary leaving exams (PLE) purchased from Uganda’s National Examination Bureau in
2020, relating results from tests taken in November 2019. Students can receive scores ranging from 1 (best) to 9 (worst) for each exam. We have
transformed this measure to facilitate interpretation such that 9 is the top score and 1 is the worst score (reflected in figures H.2a through H.2d).
Furthermore, all measures in the first four columns have been standardized using the pooled subject-specific mean and standard deviation of control
schools. Therefore, treatment effect measures reflect standard deviation differences between treatment schools and control schools
in columns one through four. We analyze the PLE pass rate and P6 to P7 pass-through rate in columns 5 and 6 respectively. Both are
binary measures equaling 1 if a student passed the exam/grade level and 0 otherwise. The number of observations in column five is greater than that
in columns one through four because students who registered but were not present for the PLE exam are classified as students who must repeat the
PLE exam in the following year. Students are divided into four “divisions” based on PLE scores, division 1 reflects high performance and division 4 is
the lowest. Students receiving division 4 marks and below are considered to have the PLE (typically with an aggregate score above 28). The variable
analyzed in column 6 is constructed using the student survey. For more details, see pages 20-21 of the pre-analysis plan. We standardize variables in
columns one through four according to the control school mean and standard deviation.

Table 3: Preliminary Student Outcomes for Hypothesis (H2.1): Traditional Learning Outcomes Increase
 Hypotheses: (H2.2): Higher Order Learning (H2.3): Science Show (H2.4): Creativity
Outcome Variables Apply/Understand Critical Thinking Index (Mean) Index
Treatment (ITT ) 0.73∗∗∗ 0.45∗∗ 0.87∗∗ 0.44∗∗
(0.14) (0.16) (0.40) (0.20)
H0 : ITT = 0
p value [0.00]±± [0.01]±± [0.03]±± [0.03]±±
RI p value [0.00]±± [0.02]±± [0.13] [0.35]
BH Critical p value (5%) [0.03] [0.05] [0.05] [0.05]
Pair FE Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes
Source of Data Student Assessment Student Assessment Science Show Student Survey
Unit of Observation Student Student Student Group Student
Standardized Variable Yes Yes No No
Range of Outcome Variable [-2.00, 1.68] [-1.75, 1.95] [1, 9.16] [0, 20]
Control School Mean 0.00 0.00 2.63 6.21
42

Clusters 29 29 29 29
Observations 329 329 158 329
Estimator OLS OLS Tobit Tobit

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Randomization Inference using 1,000 permutations
of school-level treatment indicator within matched-pair strata. Tobit estimator in the third column treats 1 as the lower bound and 10 as the upper
bound; in the fourth column, 0 is the lower bound with no upper bound.
Variable Descriptions: The unit of observation in the first, second and fourth columns is individual students, and student groups in the third column.
Each group was judged by two separate judges, thus there are 79 total groups across 29 schools in the analysis. The first and second columns report
results from the researcher administered student assessments, measuring applied understanding and critical thinking using the procedure outlined on
pages 21 and 22 of the pre-analysis plan. The science show outcome in the third column averages across the 12 outcomes measured by science show
judges and described on page 25 and table 1 of the pre-analysis plan. The creativity score follows the procedure described on pages 26 and 27 of the
pre-analysis plan. We standardize variables in columns one and two according to the control school mean and standard deviation.

Table 4: Preliminary Student Outcomes for Hypotheses (H2.2) through ((H2.4)): Higher Order Learning, Science Shows and Creativity
 Hypothesis: (H1.1): Pedagogy (H1.2): Effort (H1.3): Learning
Outcome Variable: Share of Engaged Pupils Student Inquisitiveness Corporal Punishment Knowledge of Student Teacher Network
Treatment (ITT ) 0.39∗∗∗ 0.06∗∗ −0.01 0.10∗∗∗ 0.27∗∗
(0.15) (0.03) (0.04) (0.03) (0.11)
H0 : ITT = 0
p value [0.01]±± [0.02]±± [0.81] [0.00]±± [0.02]±±
RI p value [0.01]±± [0.28] [0.94] [0.06]± [0.14]
BH Critical p value (5%) [0.02] [0.03] [0.05] [0.05] [0.05]
Pair FE Yes Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes No
Grade FE Yes No No No No
Source of Data Classroom Observations Student Survey Student Survey Stud. + Teach. Survey Teacher Network
Unit of Observation Classroom Snapshots P6 Teachers P6 Teachers P6 Teachers Teacher Dyads
Range of Outcome Variable {1,2,. . . ,6} [0,1] [0,1] [0,1] {0,1,2,3}
Control School Mean 4.41 0.22 0.53 0.62 1.80
Clusters 29 29 29 29 29
Observations 2,380 95 95 95 1,466
43

Estimator Ologit Tobit Tobit Tobit OLS

Analysis Notes: Standard errors are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01.
Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”) as well as the Benjamini Hochberg (BH) “BH Critical p value” at the 5% level
within hypothesis. Hypothesis categories are delineated in the top row of the table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level; ±± suggests a
significant discovery at the 5% level. Randomization Inference using 1,000 permutations of school-level treatment indicator within matched-pair strata. Tobit estimator in the second through
fourth columns treats zero as the lower bound and one as the upper bound.
Variable Descriptions: Units of observation in the first column are classroom snapshots observed using the Stallings instrument — 10 “snapshots” are taking during each class; therefore, 238
classes (P4 to P6) were observed across 29 schools. “Share of Engaged Pupils” indicates the observer’s assessment of the share of pupils engaged in an activity with a teacher: 1 = No pupil,
2 = One pupil, 3 = A few pupils, 4 = Half of the pupils, 5 = Most of the pupils, and 6 = All pupils (details on page 16 of the pre-analysis plan). Teachers in the second to fourth columns
are restricted to those who at least 90% of P6 students listed as one of their teachers. The second column analyzes the percent of students who indicate that they have asked questions in
the previous two school terms to the specified teacher when they do not understand a topic. The third column analyzes the percent of students who indicate that the specified teacher has
“caned” or beaten them at some point in the previous two school terms (details on page 17 of the pre-analysis plan). The fourth column combines student and teacher responses regarding
student attendance and the students relationship with their guardians for each student. We analyze the treatment effect on the percent of correct teacher responses (details available on pages
18-19 of the pre-analysis plan). Unit of observation in the fifth column is within-school teacher dyads. We sum across binary measures of teacher interactions along three dimensions where
both teachers indicate the existence of a link reflecting teacher collaboration (details available on pages 19-20 of the pre-analysis plan).

Table 5: Preliminary Teacher Outcomes for Hypotheses (H1.1) through (H1.3): Teacher Pedagogy, Teacher Effort, Teacher Learning
 Source of Data Science Show
Category Framing Experiment Hypothesis Measurement Articulating
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Treatment (ITT ) −0.79 0.39 1.34∗∗ 0.51 0.22 2.29∗ 2.84∗∗ 0.99 2.55∗∗ 1.83∗∗∗ 2.12∗∗∗ 2.37∗∗
(0.81) (0.94) (0.62) (0.78) (0.92) (1.18) (1.25) (0.87) (0.98) (0.62) (0.79) (0.96)
H0 : ITT = 0
p value [0.33] [0.68] [0.03]± [0.52] [0.81] [0.05]± [0.02]±±[0.26] [0.01]±±[0.00]±±[0.01]±± [0.01]±±
RI p value [0.50] [0.76] [0.14] [0.63] [0.89] [0.20] [0.14] [0.41] [0.06]± [0.05] [0.04]± [0.08]±
BH Critical p value (5%) [0.03] [0.05] [0.02] [0.03] [0.05] [0.05] [0.03] [0.05] [0.03] [0.02] [0.03] [0.05]
Pair FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Judge FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Standardized Variable No No No No No No No No No No No No
44

Range of Outcome Variable [1,10] [1,10] [1,10] [1,10] [1,8] [1,10] [1,10] [1,10] [1,10] [1,10] [1,10] [1,10]
Control School Mean 3.24 2.84 3.81 2.19 1.8 2.71 2.4 1.65 2.92 2.16 3.2 2.65
Clusters 29 29 29 29 29 29 29 29 29 29 29 29
Observations 158 158 158 158 158 158 158 158 158 158 158 158
Estimator Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Hypotheses are grouped according to the categories
outlined in the second row of the table. Randomization Inference using 1,000 permutations of school-level treatment indicator within matched-pair
strata. Tobit estimator treats 1 as the lower bound and 10 as the upper bound across all columns.

Table 6: Granular Science Show Outcomes

 Class Attended in 2018
Observation Set (Class) All P1 P2 P3 P4 P5 P6 P7
Panel A: Outcome Variable = Is Student Attending School in 2022?
Treatment 2018 (ITT ) 0.03∗∗ 0.01 0.01 0.02∗ 0.06∗∗∗ 0.07∗ 0.05 0.09∗∗
(0.01) (0.00) (0.01) (0.01) (0.02) (0.03) (0.04) (0.04)
H0 : ITT = 0
p value [0.02] [0.21] [0.31] [0.1] [0.00] [0.06] [0.15] [0.03]

Pair FE Yes Yes Yes Yes Yes Yes Yes Yes

Range of Outcome Variable {0, 1} {0, 1} {0, 1} {0, 1} {0, 1} {0, 1} {0, 1} {0, 1}
Control School Mean 0.89 0.99 0.97 0.93 0.84 0.8 0.76 0.69
Panel B: Outcome Variable = Is Student Attending Secondary School in 2022?
Treatment 2018 (ITT ) 0.1∗∗ 0.12∗∗ 0.07∗ 0.08∗
(0.04) (0.06) (0.03) (0.05)
H0 : ITT = 0
p value [0.02] [0.04] [0.07] [0.08]
45

Pair FE Yes Yes Yes Yes

Range of Outcome Variable {0, 1} {0, 1} {0, 1} {0, 1}
Control School Mean 0.62 0.52 0.72 0.68
Panel C: Outcome Variable = Class 2022 - Class 2018 (Persistence)
Treatment 2018 (ITT ) 0.14∗∗∗ 0.09∗ 0.11∗∗ 0.17∗∗∗ 0.19∗∗∗ 0.17∗ 0.21∗∗∗ 0.25∗∗
(0.04) (0.05) (0.05) (0.06) (0.05) (0.09) (0.08) (0.1)
H0 : ITT = 0
p value [0.00] [0.06] [0.04] [0.01] [0.00] [0.07] [0.01] [0.01]

Pair FE Yes Yes Yes Yes Yes Yes Yes Yes

Range of Outcome Variable {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4} {0, 1, .., 4}
Control School Mean 2.55 2.66 2.71 2.65 2.51 2.39 2.24 2.15
Unit of Observation School-going child between grades 1 and 7 in 2018
Source of Data 2022 Household Census (All)
Clusters 29 29 29 29 29 29 29 29
Observations 6,056 1,264 1,077 1,081 949 712 565 408
Estimator OLS (All)
Analysis Notes: Standard errors are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,” less than 0.1, ∗∗
less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to Treat effect in which a student is assigned to treatment based on their 2018 school
assignment only. All observations in column 1 reflect number of observations in Panel A and C. Total observations in Panel C sums observations across
columns representing P5 to P7.

Table 7: Treatment Effects on Long-Run School Enrollment and Persistence

 Subject
Dependent Variable: All Bio Chem Physics Math Reading Language
Panel A: All Secondary Students Matched with Census
Treatment 2018 (ITT ) 0.33∗∗∗ 0.14 0.18 0.08 0.14∗∗ 0.24∗ 0.20
(0.13) (0.14) (0.15) (0.14) (0.06) (0.13) (0.14)
H0 : ITT = 0
p value [0.01] [0.32] [0.22] [0.56] [0.04] [0.07] [0.16]

Pair FE Yes Yes Yes Yes Yes Yes Yes

Source of Data 2022 Secondary Student Assessment
Unit of Observation Secondary School Students Matched with HH Census in 2022
Standardized Variable Yes Yes Yes Yes Yes Yes Yes
Range of Outcome Variable [-2.69, 3.63] [-2.12, 1.59] [-1.98, 1.57] [-1.66, 1.93] [-.42, 5.60] [-2.11, 2.08] [-2.67, 1.06]
Control School Mean −0.32 −0.18 −0.22 −0.07 −0.15 −0.25 −0.12
Observations 306 306 306 306 306 306 308
46

Estimator OLS OLS OLS OLS OLS OLS OLS

Table 8: Long-Term Learning Outcomes for Students Enrolled in Secondary School in 2022
 Table 9: Student Learning Effects in 2022

Grade P4 P5 P6 P7

Panel A: Student Math Scores

Treatment (ITT ) 0.59∗∗∗ 0.43∗∗∗ 0.43∗∗∗ 0.04
(0.15) (0.08) (0.15) (0.08)

Panel B: Student English Scores

Treatment (ITT ) 0.45∗∗∗ 0.41∗∗∗ 0.36∗∗∗ 0.25∗∗∗
(0.09) (0.13) (0.09) (0.06)

Panel C: Student Science Scores

Treatment (ITT ) 0.60∗∗∗ 0.33∗∗ 0.51∗∗∗ 0.15
(0.13) (0.13) (0.12) (0.13)

Panel D: Student Critical Thinking Scores

Treatment (ITT ) 0.29∗∗∗ 0.12 0.21∗∗∗ 0.26∗∗∗
(0.06) (0.14) (0.05) (0.07)

Pair FE Yes Yes Yes Yes

Standardized Yes Yes Yes Yes
Observations 1,105 1,030 973 842
Clusters 27 27 27 27
Analysis Notes: Standard errors are clustered at the school
level. ∗ , reflects a coefficient p value from the original spec-
ification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗
less than 0.01. Coefficients represent the Intent to Treat
effect based the school’s assignment to treatment. All out-
comes are standardized within class and effects reflect %
change in standard deviation units in treated schools.

47
 Table 10: Teacher Learning Effects in 2022

Subject Critical-Thinking Understand/Know Science/Technology Tender

Treatment (ITT ) 0.07 0.10 0.36∗∗∗ 0.34∗∗ 0.25∗∗
(0.14) (0.11) (0.12) (0.17) (0.11)

English Teacher −0.06

(0.17)

Math Teacher −0.06

(0.17)

Pair FE Yes Yes Yes Yes Yes

Standardized Yes Yes Yes Yes Yes
Observations 147 251 248 248 248
Clusters 23 24 24 24 24
48

Analysis Notes: Standard errors are clustered at the school level. ∗ , reflects a coefficient p value from the original
specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to
Treat effect based the school’s assignment to treatment. All outcomes are standardized and effects reflect % change
in standard deviation units in treated schools. Column 1 results only consist of responses from science, math or
english teachers’ subject assessments. Science teachers is the omitted category.
References
Acemoglu, D., et al. (2012). What Does Human Capital Do? A Review of Goldin and
Katz’s The Race Between Education and Technology. Journal of Economic Literature,
50 (2), 426–63.

Angrist, N., Evans, D., Filmer, D. P., Glennerster, R., Rogers, F. H., & Sabarwal, S.
(2020, October). How to Improve Education Outcomes Most Efficiently ? A Compar-
ison of 150 Interventions Using the New Learning-Adjusted Years of Schooling Metric
(Policy Research Working Paper Series No. 9450). The World Bank. Retrieved from
[Link]

Arbab, F., & Lample, P. (2005). Educational Concepts.

ASER Centre. (2019). Annual Status of Education Report (Rural) [Com-

puter software manual]. New Delhi — last accessed November 3, 2020. Re-
trieved from [Link]
20/[Link]

ASER Pakistan. (2019). Annual status of education report (provisional) [Com-

puter software manual]. Lahore — last accessed November 3, 2020. Re-
trieved from [Link]
20/[Link]

Bando, R., Näslund-Hadley, E., & Gertler, P. (2019, September). Effect of Inquiry
and Problem Based Pedagogy on Learning: Evidence from 10 Field Experiments in
Four Countries (Working Paper No. 26280). National Bureau of Economic Research.
Retrieved from [Link] doi: 10.3386/w26280

Banerjee, A., Banerji, R., Berry, J., Duflo, E., Kannan, H., Mukerji, S., . . . Walton,
M. (2017, November). From Proof of Concept to Scalable Policies: Challenges and
Solutions, with an Application. Journal of Economic Perspectives, 31 (4), 73-102.
Retrieved from [Link] doi:
10.1257/jep.31.4.73

Baten, J., de Haas, M., Kempter, E., & Selhausen, F. M. (2020). Educational Gender
Inequality in Sub-Saharan Africa: A Long-term Perspective. African Economic History
Working Paper Series, 54 .

Becker, G. S. (2009). Human Capital: A Theoretical and Empirical Analysis with Special
Reference to Education. University of Chicago Press, 3rd Edition.

Benjamini, Y., & Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical
and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society.
Series B (Methodological), 57 (1), 289–300. Retrieved from [Link]
stable/2346101

Bold, T., Filmer, D. P., Molina, E., & Svensson, J. (2019). The Lost Human Capital:
Teacher Knowledge and Student Achievement in Africa. Policy Research Working
Papers, 8849 .

Bowles, S. (2017). The Moral Economy. Yale University Press.

49
Bradler, C., Neckermann, S., & Warnke, A. J. (2019). Incentivizing Creativity: A Large-
Scale Experiment with Performance Bonuses and Gifts. Journal of Labor Economics,
37 (3), 793–851.
Brock-Utne, B. (2002). Whose Education for All?: The Recolonization of the African
Mind (Vol. 1445). Routledge.
Bruhn, M., & McKenzie, D. (2009, October). In Pursuit of Balance: Randomization
in Practice in Development Field Experiments. American Economic Journal: Applied
Economics, 1 (4), 200-232. Retrieved from [Link]
.1257/app.1.4.200 doi: 10.1257/app.1.4.200
Bude, U. (1983). The Adaptation Concept in British Colonial Education. Comparative
Education, 19 (3), 341–355.
Burdett, N. (2017). Review of High Stakes Examination Instruments in Primary and
Secondary School in Developing Countries. Research on Improving Systems of Educa-
tion (RISE) Working Paper 17/018 . Retrieved from [Link]
sites/default/files/publications/RISE_WP-018_Burdett_0.pdf
De La Warr, L. (1937). Higher Education in East Africa: Notes and Correspondence
of the De La Warr Commission. Manuscripts in the Africana Collection, Makerere
University Library.
Duckworth, E. (2006). The Having of Wonderful Ideas and Other Essays on Teaching
and Learning. Teachers College Press.
Duflo, E., Dupas, P., & Kremer, M. (2011, August). Peer Effects, Teacher Incentives,
and the Impact of Tracking: Evidence from a Randomized Evaluation in Kenya. Amer-
ican Economic Review , 101 (5), 1739-74. Retrieved from [Link]
articles?id=10.1257/aer.101.5.1739 doi: 10.1257/aer.101.5.1739
Feierman, S. (1985). Struggles for Control: The Social Roots of Health and Healing in
Modern Africa. African studies review , 28 (2/3), 73–147.
Flabbi, L., & Gatti, R. (2018). A Primer on Human Capital. Policy Research Work-
ing Paper (8309). Retrieved from [Link]
514331516372468005/pdf/[Link]
Freire, P. (1970). Pedagogy of the Oppressed. New York: Continuum.
Gilligan, D. O., Karachiwalla, N., Kasirye, I., Lucas, A., & Neal, D. (2018). Educator
Incentives and Educational Triage in Rural Primary Schools. IZA Discussion Paper
Series, 11516 .
Glewwe, P., & Muralidharan, K. (2016). Chapter 10 - Improving Education Out-
comes in Developing Countries: Evidence, Knowledge Gaps, and Policy Implications.
In E. A. Hanushek, S. Machin, & L. Woessmann (Eds.), (Vol. 5, p. 653 - 743).
Elsevier. Retrieved from [Link]
B9780444634597000105 doi: [Link]
Goldin, C., & Katz, L. F. (2009). The Race Between Technology and Education. Harvard
University Press.

50
Guilford, J. P. (1967). The Nature of Human Intelligence. McGraw-Hill.
Hanna, R., Mullainathan, S., & Schwartzstein, J. (2014). Learning Through Noticing:
Theory and Evidence From a Field Experiment. The Quarterly Journal of Economics,
129 (3), 1311–1353.
Hanson, H. E. (2010). Indigenous Adaptation: Uganda’s Village Schools, ca. 1880–1937.
Comparative Education Review , 54 (2), 155–174.
Hanushek, E. A., & Woessmann, L. (2012). Do Better Schools Lead to More Growth?
Cognitive Skills, Economic Outcomes, and Causation. Journal of Economic Growth,
17 (4), 267–321.
Jansen, J. (1989). Curriculum Reconstruction in Post-colonial Africa: A Review of the
Literature. International Journal of Educational Development, 9 (3), 219–231.
Kolencik, P. L., & Hillwig, S. A. (2011).
In Encouraging Metacognition: Supporting Learners through Metacognitive Teaching
Strategies. (p. 183). Peter Lang Inc., International Academic Publishers.
Kranton, R. (2019, March). The Devil Is in the Details: Implications of Samuel Bowles’s
The Moral Economy for Economics and Policy Research. Journal of Economic Lit-
erature, 57 (1), 147-60. Retrieved from [Link]
.1257/jel.20171463 doi: 10.1257/jel.20171463
Kremer, M. (1993). The O-ring Theory of Economic Development. The Quarterly Journal
of Economics, 108 (3), 551–575.
Lang, J. M. (2012). Metacognition and Student Learning. The Chronicle of Higher
Education.
Liu, O. L., Frankel, L., & Roohr, K. C. (2014). Assessing Critical Thinking in Higher Ed-
ucation: Current State and Directions for Next-generation Assessment. ETS Research
Report Series, 2014 (1), 1–23.
Michelson, H., Fairbairn, A., Ellison, B., Maertens, A., & Manyong, V. (2021). Mis-
perceived Quality: Fertilizer in Tanzania. Journal of Development Economics, 148 ,
102579. Retrieved from [Link]
S0304387820301541 doi: [Link]
Miguel, E., & Kremer, M. (2004, January). Worms: Identifying Impacts on Education and
Health in the Presence of Treatment Externalities. Econometrica, 72 (1), 159-217. Re-
trieved from [Link]
Ministry of Human Resource Development. (2020). National Education Policy 2020.
Government of India.
Moulton, J. (2002). Uganda: External and domestic efforts to revive a derelict primary
school system. Contributions to the Study of Education, 82 , 53–86.
Muralidharan, K., Singh, A., & Ganimian, A. J. (2019, April). Disrupting Education?
Experimental Evidence on Technology-Aided Instruction in India. American Economic
Review , 109 (4), 1426-60. Retrieved from [Link]
.1257/aer.20171112 doi: 10.1257/aer.20171112

51
National Curriculum Development Centre (NCDC). (2016). Primary 6 Science Curricu-
lum. National Curriculum Development Centre.
Nguyen, T. (2008). Information, Role Models and Perceived Returns to Education:
Experimental Evidence from Madagascar. Unpublished manuscript.
Perkins, D. N. (1992). Smart Schools: Better Thinking and Learning for Every Child.
The Free Press.
Piper, B., Zuilkowski, S. S., Dubeck, M., Jepkemei, E., & King, S. J. (2018). Identify-
ing the Essential Ingredients to Literacy and Numeracy Improvement: Teacher Pro-
fessional Development and Coaching, Student Textbooks, and Structured Teachers’
Guides. World Development, 106 , 324–336.
Piper, B., Zuilkowski, S. S., & Mugenda, A. (2014). Improving Reading Outcomes in
Kenya: First-year Effects of the PRIMR Initiative. International Journal of Educa-
tional Development, 37 , 11–21.
Pradhan, M., Suryadarma, D., Beatty, A., Wong, M., Gaduh, A., Alisjahbana, A., &
Artha, R. P. (2014, April). Improving Educational Quality through Enhancing Commu-
nity Participation: Results from a Randomized Field Experiment in Indonesia. Amer-
ican Economic Journal: Applied Economics, 6 (2), 105-26. Retrieved from https://
[Link]/articles?id=10.1257/app.6.2.105 doi: 10.1257/app.6.2.105
Ranciére, J. (1991). The Ignorant Schoolmaster. Stanford University Press.
SIPRO. (2019). Sipro Mid-term II Examination 2019: Primary Six. The SIPRO Educa-
tional Services Ltd.
Ssekamwa, J. C. (1997). History and Development of Education in Uganda. Fountain
Pub Limited.
Swartz, R. J., & Perkins, D. (2017). Psychology Library Editions: Cognitive Science. In
R. J. Swartz & D. Perkins (Eds.), Teaching thinking: Issues and approaches (Vol. 24).
Routledge.
Tanner, K. D. (2012). Promoting Student Metacognition. CBE—Life Sciences Education,
11 (2), 113–120.
The World Bank. (2020, October). Cost-Effective Approaches to Improve Global Learning:
What Does Recent Evidence Tell us Are “Smart Buys” for Improving Learning in Low-
and Middle-Income Countries? Retrieved from [Link]
curated/en/719211603835247448/pdf/Cost-Effective-Approaches-to-Improve
-Global-Learning-What-Does-Recent-Evidence-Tell-Us-Are-Smart-Buys-for
-[Link]
Uganda Ministry of Education and Sports (MoES). (2014, Jun). Teachers Initiative
in Sub-Saharan Africa (TISSA) Uganda. Ministry of Education and Sports, Uganda.
Retrieved from [Link]
_files/tissa_uganda_-_2014.pdf
Uganda Ministry of Education and Sports (MoES). (2017, May). A Teacher Incentive
Framework (TIF) For Uganda. Ministry of Education and Sports, Uganda.

52
United Nations Educational Scientific and Cultural Organization (UNESCO). (2020).
Uganda: Education and literacy (statistics) [Computer software manual]. — last ac-
cessed November 3, 2020. Retrieved from [Link]

UWEZO. (2016). Are Our Children Learning? Uwezo Uganda Sixth Learning Assessment
Report [Computer software manual]. Kampala: Twaweza East Africa — last accessed
November 3, 2020. Retrieved from [Link]
2016/12/[Link]

World Bank Group. (2015). User Guide: Conducting Classroom Observations. World
Bank Group.

World Bank Group. (2018). World Development Report 2018: Learning to Realize
Education’s Promise. Washington, DC: The World Bank . Retrieved from https://
[Link]/en/publication/wdr2018

Young, A. (2018). Channeling Fisher: Randomization Tests and the Statistical In-
significance of Seemingly Significant Experimental Results. The Quarterly Journal of
Economics, 134 (2), 557-598. Retrieved from [Link]
doi: 10.1093/qje/qjy029

53
Online Appendix for Learning to
Teach by Learning to Learn

Vesall Nourani Nava Ashraf Abhijit Banerjee

November 11, 2023

i
Appendix
Table of Contents
A Summaries of Each PSA Course in the Teacher Training iii
A.1 Primary Elements of Description: Properties . . . . . . . . . . . . . . v
A.2 Matter: The Heating and Cooling of Matter . . . . . . . . . . . . . . . vi
A.3 Dawn of Civilization: Transition to Agriculture . . . . . . . . . . . . . vi
A.4 Nurturing Young Minds: To Describe the World . . . . . . . . . . . . . vii
A.5 Promoting a Healthy Environment: Environmental Issues . . . . . . . . ix
A.6 Food Production on Small Farms: Planting Crops . . . . . . . . . . . . x
A.7 Example of Facilitation Style in Teacher Training . . . . . . . . . . . . xi

B Ugandan Educational Context xiii

B.1 History of Schooling and Education in Uganda . . . . . . . . . . . . . xiii
B.2 Study Context — Jinja District, Uganda . . . . . . . . . . . . . . . . . xiv
B.3 Additional Socio-Economic Context in Budondo . . . . . . . . . . . . . xviii

C Measurement of Outcome Variables xxiii

C.1 (Q1) Does the teacher training change teacher pedagogy? . . . . . . . xxiii
C.2 (Q2) Does the teacher training change student outcomes? . . . . . . . xxv

D Researcher Provided Student Assessment xxviii

E Judging the Science Shows xxxv

F Categories of Creativity score xlv

G Additional Appendix Tables xlix

H Additional Appendix Figures lxiv

ii
A Summaries of Each PSA Course in the Teacher Train-
ing
The entire PSA program is organized around the concept of a capability, conceived as
developed capacity to think and to act purposefully in a particular sphere of activity.
Each unit of instruction has been designed to contribute to the strengthening of one or
more such capabilities. The capabilities can be broadly categorized into four areas: math-
ematics, science, language, and processes of community life. The spiritual dimension of
human reality is explicitly discussed given its importance to the lives of most populations
served by the FUNDAEC materials, and its role in endowing participant reflections with
deeper purpose and meaning. The approach taken by FUNDAEC materials and tutors
can be coarsely described as inquisitive and exploratory — allowing participants to be
the protagonists of their own learning and not imposing one or another world-view. The
materials help participants consider the validity of all perspectives by assisting them to
express and verify their own thoughts on a particular matter without blindly adopting
any point of view.

Table A.1: PSA Courses Studied During Intense Training Sessions

Term# Book Unit Capacities Developed

1 Primary Ele- Properties To think about the nature of
ments of Descrip- language and the way it is
tion used to describe reality.
1 Matter The Heating and Observing and describing
Cooling of Matter the transformation of mate-
rial substances.
1 DSA Basic Concepts —
2 Dawn of Civiliza- Transition to Reading Comprehension.
tion Agriculture Understanding human ex-
perience in the context of
the evolution of civilization
2 Nurturing Young To Describe the Teach young children (ages
Minds World 4-5) to describe the world
with increasing levels of clar-
ity.
2 DSA Education —
3 Promoting a Environmental Capacity to create aware-
Healthy Environ- Issues ness and identify solutions
ment of community environmen-
tal issues.
3 Food Production Planting Crops Effective participation in the
on Small Farms development, selection, and
use of technology.
3 DSA Science —

Each of the courses in the training, along with the term during which they are studied
and a coarse summary of their associated capacities, are listed and summarized in Table

iii
A.1. There are four aspects of the courses that are worth pointing out. First, the
courses chosen are designed to develop participants’ facility with the scientific method by
discussing and implementing applications relevant to the local context. For example, in
“Planting Crops” (T3) teachers learn how to apply experimental methods to planting, use
local flora to learn how to describe plant systems, etc. This allows teachers an opportunity
to reflect on the ways in which exploratory methods can be introduced into the classroom
using processes students themselves are familiar with.
Second, it is worth noting that the FUNDAEC courses have an explicit aim of lead-
ing the participants to think of themselves as “Promoters of Community Well-being.”
Thus, the focus on local application of the scientific method is aligned with this aim.
The courses are also explicit about exploring both the “spiritual qualities” a promoter
of community well-being might possess in addition to the relationships and qualities a
promoter of community well-being seeks to engender in others. Thus, the participants in
the course are attentive to how their new-found scientific capabilities can be applied to
learning about the material, social and moral dimensions of community well-being. This
theme is explored in some depth in the courses covered in the third term of the train-
ing. In the unit titled “Planting Crops” (T3) teachers develop an understanding of how
experimental methods can be used to learn about the productivity and social impacts of
diverse types of agricultural technologies. In the unit titled “Environmental Issues” (T3)
teachers explore the relationship between human activity and pollutants stemming from
systems of transportation, small- and large-scale manufacturing, and waste-management
processes. In both courses, care is taken to allow participants to reflect on the moral
implications of various approaches to agricultural production and use of resources that
have potential negative spillovers.
Third, and relatedly, many of the concepts teachers learn — and capacities they ac-
quire — are reinforced in subsequent lessons through actions teachers undertake in the
communities they live in. For example, in the unit titled “Properties” (T1), teachers are
exposed to the primary elements of descriptions. Specifically, they reflect on the words
one might use to describe substances that exist in the world, their positions in space and
their relationships with one another. Then, in the unit titled “To Describe the World”
(T2), teachers are asked to organize groups of pre-school children through hands-on ac-
tivities that enable them to describe the world around them with increasing levels of
clarity. Using resources from their village environment — such as leaves, flowers, sticks,
and pebbles — they help children understand the concept of sets, size of sets, characteris-
tics of elements of sets, numbers, shapes, etc. many of which were previously explored by
teachers in “Properties.” This act of organizing village children to carry out these activi-
ties allows teachers a concrete setting in the community to apply the linguistic capacity
they gained in earlier courses. Similarly, while studying “Transition to Agriculture” (T2),
teachers develop the ability to describe the emergence of culture and civilization in a
manner that contextualizes modern experiences by an analysis of the past. This concept
is reinforced by asking teachers to interview elders in the community to learn how pro-
cesses around child-rearing and agricultural production have evolved over generations in
the local community.
Fourth, there are natural opportunities embedded in each course for teachers to con-
template the ways in which these capabilities can be used to improve their understanding
of and responses to classroom processes. We share one such example when describing
the change of phase of a substance through the texts “Properties” and the “Heating and
Cooling of Matter” (T1).

iv
 Across all of the terms teachers study units in a course titled Discourse for Social
Action (DSA). Many of the courses in DSA cover concepts that allow participants spaces
for meta-cognitive reflection of the reasons for covering various topics the PSA curricu-
lum. Specifically, participants consider aspects of epistemology associated with scientific
methods of knowledge generation. In this vein, participants reflect on the use of assump-
tions in developing scientific theory, methods used by various scientific disciplines, the
connection between knowledge and reality, and purposes tied to the progress of scientifi-
cally generated knowledge. Teachers engage themes that range from avoiding a tendency
to seek simplistic answers, on the one hand, to the value of scientific investigation for
processes of social and economic development on the other. The spaces for reflection
offered during the study of DSA allow them to correlate the capacities gained in studying
PSA materials to social processes relevant to the education of children.
The below sections describe the content of courses in the PSA sequence utilized by
Kimanya-Ngeyo in their primary-school teacher training as of November 2018. Each
course is often accompanied by a section titled “To the Tutor,” that FUNDAEC addresses
to the tutor who is guiding a group of participants through the course materials. The
below summaries often copy text from these “To the tutor” sections as well as some select
sections of the course that are more relevant to the teacher training.
These courses are continuously refined by FUNDAEC, so the summaries described
below are subject to change. These summaries do not capture the fullness of the content
and especially to not capture the value of the mode in which these materials are studied
— by mutually exploring their implications with other peers with the aid of a tutor who
facilitates the learning process. They also leave out many of the exercises and workbook
activities that aim to enhance the participant’s understanding of course content.

A.1 Primary Elements of Description: Properties

This unit was created to address a shortcoming perceived by FUNDAEC when it started
to offer its courses. This was that young people often lacked facility with words and
concepts needed to describe the world around them. While they could recognize various
shapes and speak vaguely about the size and position of objects, they had not reflected
on the meanings the words associated with these ideas convey.
The first few lessons of the unit acquaint participants with a number of concepts and
words used in making descriptions at the most basic level: shape, size, and position.
The difference between objective and subjective descriptions is then briefly examined,
followed by a review of the general properties of matter in its solid, liquid, and gaseous
phases. Some of the properties of specific substances, such as melting and boiling point,
color, density, and hardness, are examined next. The concept of property is used to enter
into a discussion of the qualities of human beings; of these, the qualities of truthfulness,
justice, and love are treated in detail. In a similar vein, participants are asked in the
final lesson to think about the characteristics of the kind of communities that, as future
agents of change, they will be called upon to build.
Many ideas employed in describing the world necessarily come from various branches
of knowledge, and so the lessons touch on a number of subject matters. It should be
remembered, however, that the unit is primarily concerned with developing capabilities
in the area of language. The words and concepts emphasized in the lessons are intended
to help the student think about the nature of language. The lesson dedicated to shape,
to take one example, should not be treated like an exercise in geometry; its purpose is to

v
encourage participants to reflect on the concept of shape and its role in expression.
The capabilities used in describing reality are not developed through knowledge of
physical existence alone. FUNDAEC conceives of reality as having a social and spiri-
tual dimension as well; the unit therefore is not content to undertake an exploration of
physical properties only. By introducing higher entities — the human being and human
communities — it is able to extend the discussion into the social and spiritual realms, and
students are compelled from the outset of their exploration of language to think about
different facets of existence and the means for describing it. It is hoped that, in this way,
they will begin to form habits of the mind that avoid the tendency towards fragmentation
which invariably gives rise to unnecessary contradictions and imaginary dichotomies in
life.
Sections: 1) Shape 2) Size 3) Space 4) Position 5) Objective and Subjective Descrip-
tions 6) Matter and Its General Properties 7) Phases of Matter 8) Specific Properties
of Matter 9) Melting Point and Boiling Point 10) Color 11) Density and Hardness 12)
Qualities of Human Beings 13) Truthfulness 14) Justice 15) Love 16) Characteristics of
a Community.

A.2 Matter: The Heating and Cooling of Matter

This book aims at fostering in participants some of the capabilities that scientists draw
upon to study and understand the world around them. It enables participants to execute
a chief task of science: that of building models of the way the real world behaves. It does
so by focusing analysis around a phenomenon that is easy to understand and is present
in every day life: basic thermodynamics. This approach takes the participant through
basic concepts required to understand simple models of thermodynamics while placing
them on the threshold of understanding atomic theory. The final lessons opens the way
for a general consideration of models and theories. These sections allow participants to
reflect on several important, yet abstract, ideas: how are scientists conceived as members
of “scientific communities” who consume and generate bodies of knowledge, not all of
which are valid, that accumulate over long periods of time. They further reflect on the
range of validity of scientific models and the relationship between beliefs of scientists and
the models they create. Participants reflect on the ways in which scientific models evolve
when presented with observations outside of the range of validity of previously accepted
models.
Sections in the course: 1) What is Science? 2) Observation 3) Temperature 4) Heat 5)
Force 6) Pressure 7) Variables That Determine the State of a Gas. 8) Change of Phase
9) Particles 10) Models and Theories

A.3 Dawn of Civilization: Transition to Agriculture

The primary purpose of the unit is to contribute to the capabilities of reading with good
comprehension and expressing ideas with clarity. The story of humanity’s transition from
hunting and gathering to agriculture serves as the means through which this purpose is
achieved. As such, the unit also seeks to enhance a range of capabilities associated with
the process of civilization building. Not least among these is the capability of placing
events and developments in historical context. Each of the twelve lessons in this unit
(apart from the first) follow the same format. Each begins with a reading intended to help
participants think about the historical developments that eventually led to the emergence

vi
of civilization. In their entirety, the readings present a view of human existence that has
two dimensions, material and spiritual. Human progress, it is suggested, is driven by
knowledge of both these aspects of reality.
To convey concepts that might otherwise be too challenging, the unit introduces an
imaginary study group learning about early human history and uses the dialogue between
its members as a means of addressing theoretical and abstract subjects. The sections of
each lesson that deal with capacities of language are “Comprehension,” “Building Vocabu-
lary,” “About Language.” The final four sections are “Extension” (an idea is amplified and
considered in the light of new, additional information), “Reflections” (think about how a
concept relates in concrete terms to the world around them), “Writing” (participants in-
vest thought into a relevant question and formulate a cogent, well-written response). The
final section is called “Investigation” and encourages them to carry out modest research
based on what they have learned and to articulate their findings in writing, helping them
to see early in their education that knowledge has highly practical dimensions.
The sections of the unit (in order) are: 1) Introduction 2) Human Evolution 3) Hunter-
Gatherers 4) The Process of Domestication 5) What Led to Domestication 6) The Man
with a Red Feather on His Head 7) Genetics 8) Material Development 9) Archaeology
10) Spiritual Development 11) Science 12) Religion.

A.4 Nurturing Young Minds: To Describe the World

This book focuses on developing capacity in contributing to processes of community
life while simultaneously developing mathematical and language capabilities by enlisting
participants in processes around children’s education. In the first lesson, emphasis is
placed on nurturing an attitude in which one does not view oneself as superior to another.
This implies that our mode of engaging with others is one of a mode of service — readiness
to learn with others, acquiring and utilizing knowledge for the upliftment and progress of
the community. This course will focus on service a “promoter of community well-being”
can render with children and asks participants to identify a group of 5-6 children of
pre-school age that a participant can meet with frequently and take through educational
activities. In order to maintain a proper posture of service, the course focuses its first
lesson on the non-cognitive qualities, conceived of as “spiritual qualities,” required by
someone adopting a posture of service. Participants have the opportunity to reflect on the
relationship between the following qualities and a mode of service: love, faith, kindness,
sensitivity to others’ needs, respect, patience, honesty and trustworthiness, humility and
diligence.
The educational activities the participant undertakes with children is that of describing
sets. They are encouraged to learn the concept of a set, the concepts of belonging to a
set, elements of a set, and of the number of elements in a set (more vs. fewer). Along the
way, participants think of ways they can introduce this activity with objects from their
immediate environments. For example, a participant will collect stones, seeds or leaves.
Then, the participant will group stones together with stones, seeds together with seeds
and so on. He or she will show these groupings to the children and say “this is a set of
stones,” “this is a set of seeds,” etc. until they have done this with a sufficient number of
sets. Then, the participant will help the children do the exact same thing and to verify
the presence of the various sets. The next activity invites children to form their own sets.
The participant will also introduce games to the children that enable them to engage in
physical activity. The teacher can refer to children as a “set of children squatting,” which

vii
will prompt the children to squat and so on.
Following these activities, it should be possible to reinforce the concept of a set by
helping children understand what it means to belong to a set — i.e., to possess a certain
characteristic. The teachers are asked not to use the word characteristic, but to use
examples from the children’s immediate environment to differentiate sets using their
characteristics. E.g., a set of flowers can be divided into a set of all flowers or a set of red
flowers (flowers possessing a red characteristic). Next, with the examples in front of the
participant, they are encouraged to point to two sets with the same characteristic (e.g.
flowers, sticks, etc.) and ask the children “which of these sets has more elements?” Repeat
this step and verify the children’s understanding. Many other activities are suggested
including coloring sheets, games, songs and other activities that will assist the children
learn the concepts associated with sets. Some activities encourage the children to think
of the characteristics that describe themselves to see if they would put themselves in a
set of, say, kind children. The unit emphasizes that one does not need a classroom, only
a corner of a living room or place in a yard.
The next educational activity is centered around exploring “God’s creation.” Teachers
first reflect on what it means to possess consciousness. They are asked to imagine a human
being singing a beautiful tune and a bird singing a tune. Is the bird conscious of what it
is doing? Is the human being? This is not meant to be a subject of discussion with the
children. It is intended for the teachers so that they reflect meaningfully on what it means
to claim that one “knows” something. Then, moving forward, activities are described that
one could conduct with the children so they can explore the world around them using
their senses. First their sight, then hearing and touch. The participant is asked to think
of activities for the sense of smell and taste. As an example of one of the activities under
the sense of touch the teacher is encouraged to give children an object that they will be
tasked with describing according to its roughness, smoothness, softness, etc. Following
this activity children are asked whether there are things that one can know without one’s
senses. This will allow children to reflect on the reality of non-material things.
Next, given the list of objects identified by the children, teachers will introduce them
to the classification. Again, teachers need not use the term “classification,” and should
instead focus on helping children group objects into useful categories. A natural first step
is to identify whether one of the listed items falls under the animal kingdom, mineral
kingdom, plant, etc. The teachers might then encourage the children to think about
differences between human beings and animals. Specifically, teachers are encouraged to
talk to children about the human soul and the non-material realms of reality it inhabits.
This is important because it develops children’s abstract thinking ability while building
this ability atop of activities that use physical and motor skills to understand the world.
Counting is also introduced in a similar vein. Reflections allow teachers to think about
the value of the concrete measurement of reality that counting allows while being aware
of uncountable aspects of reality that the rigidity of counting prohibits. Next, teachers
are introduced to approaches of introducing addition and subtraction, space, and time
and change.
The text is explicit in its exploration of spiritual reality by framing discussions around
“spiritual qualities.” In the context of a lesson on addition and subtraction, for example,
teachers are able to introduce the concept of “more and less.” The text suggests that
teachers can ask children to connect the concepts of more and less to the attitudes of
giving and sharing. If you share your food with your friend, do you have more or less
food? Does this make you richer or poorer? Does the act of giving one child love

viii
subtract a mothers amount of love she can give to another child? In discussing concepts
associated with space, teachers reflect with children about what it means to have an
“orderly” space. They are asked to explore this concept by discussing the relationship
between children’s belongings and their immediate environment. Do your belongings have
a particular place in your room? What about the way you appear to others? Care is taken
that teachers explore the benefits of orderliness that emerges from inner discipline and
not from externally imposed discipline. In discussing time and change, teachers explore
whether all indicators of change can be measured. Is it desirable to reduce the intellectual
and spiritual growth of a person to a group of indicators that can be cleverly measured?
The text states that clearly measurement might help, but it is important not to take this
to an extreme. How might a teacher assist students to develop their capacity to evaluate
their own process of learning?

A.5 Promoting a Healthy Environment: Environmental Issues

The participants of PSA courses see themselves as “promoters of community well-being.”
As such, they need to develop an understanding of the kinds of challenges faced by
communities. These can be categorized into three types: those a community can meet
using its own resources, those that a community requires help from the outside to solve
and those challenges that are well beyond the possibilities of an average village to solve.
Despite the scope of challenges in the third category, promoters of community well-
being acknowledge that the problem cannot be ignored at the level of a microregion, the
geographic unit participants are most concerned about.
This course develops a fundamental capacity required by promoters of community
well-being is one in which they discuss such challenges with local populations and create
awareness of the issues involved. After all, can people do anything to address their
problems if they are not conscious of them and remain unaware of possible solutions? This
course focuses on the question of the physical environment and the challenge of raising
awareness regarding the environmental issues in a microregion. It begins by stating a
simple truth: that every human activity affects the physical environment. Participants
identify activities in their microregion and begin to develop an awareness of how these
activities influence the physical environment. Two sections are dedicated to an analysis
of pollutants that emerge out of systems of transportation. Participants research changes
in transportation systems in their microregions in the last twenty years and engage in
conversations with other members of the community with the purpose of raising awareness
about the relationship between transportation systems and pollution. They are asked
specifically to take care “to ensure that the [conversations] you create take the form of a
conversation, not a sermon.”
Participants also gain an understanding of possible solutions to the problem of air
pollutants generated by transportation systems. These fall into three broad categories:
1) technological remedies, 2) those that relate to individual behavior, 3) actions and
decisions of our governments. Promoters of community well-being can likely contribute
to solutions in categories 2 and 3 and participants reflect on ways in which they would
assist their peers to reflect on these types of solutions.
After exploring this theme, the course switches its attention to small-scale manufac-
turing and large-scale manufacturing and again encourages participants to reflect on how
these activities negatively effect the physical environment and the solutions to these neg-
ative effects. These sections model a conversation among a few local manufacturers to

ix
aid participants to think of their own roles as promoters of community well-being. In this
way, participants are exposed to more complex forms of waste management and think
about how wasteful substances are discharged into the air and water while others are
thrown away as refuse.
A final human activity that participants analyze is that of daily individual and fam-
ily activities: namely the waste generated from the consumption behavior we engage in.
Participants are asked to keep a log of the types of waste they produce in their daily
activities in order to recognize the relationship between this activity and the production
of waste. There are three questions that are the subject of study and conversation among
the participants: Can we reduce the amount of waste material generated? Can we find a
beneficial use for the waste material? Can we find a safe method of disposal? These ques-
tions are discussed while also encouraging participants to be moderate and to consider
the desirable goal of prosperity in their communities. Similarly to the challenges associ-
ated with transportation and small-and-large-scale manufacturing, participants consider
solutions associated with reducing consumption patterns, finding beneficial uses for waste
materials, and how to safely dispose of parts that cannot be re-used. In the final sec-
tion, participants learn the distinction between biodegradeable and nonbiodegradeable
waste. This forms the basis to understanding the production of compost in the unit
titled “Planting Crops.”
Sequence of sections in this unit: 1) Awareness 2) Human Activity 3) Transportation 4)
Factors that Affect Air Polution 5) Solutions to Air Pollution 6) Small-Scale Manufactur-
ing 7) Waste Management in Small-Scale Manufacturing 8) Large-Scale Manufacturing I
9) Large-Scale Manufacturing II 10) Consumption and Solid Waste 11) Solutions to Solid
Waste

A.6 Food Production on Small Farms: Planting Crops

Technology extends the fruits of science to humanity. It is a fundamental theme through-
out all courses, but is treated extensively in this unit. The capabilities explored in this
unit broadly fall in the category of “science” and “processes of community life.” Given the
central role of agriculture, its processes and activities represent an excellent means for
helping people advance in the capabilities they will need to participate effectively in the
development, selection, and use of technology. Participants will do this by exploring how
farming a particular rotation of crops on a small plot of land can generate high yields
year after year.
The first lesson of “Planting Crops” lays a foundation for the unit as a whole. Here
the concept of primary production is introduced, and through a series of short exercises
participants are asked to reflect on the importance of agriculture and animal husbandry
for human societies. This paves the way for a brief discussion on the historical evolution of
agriculture production and on the two sources of knowledge available about this essential
area of activity: modern science and technology, on the one hand, and the experience of
the farmers of the world, on the other. The concept of appropriate technology follows
naturally from this idea, and participants are helped to think about the implications of
technological choice for the way of life of a region. It is not intended for them to reach
any conclusions in this respect. They should simply gain a sense of the scope of the issues
involved.
The sections in this book follow a conversation among four individuals who represent
different ways of thinking about social change. A young man who wishes to help his

x
grandmother improve production on her small farm is seeking advice from others on how
to proceed. His grandmother provides a perspective informed by traditional knowledge. A
friend studying at a nearby university brings the view of specialized science of agriculture.
His cousin shares ideas as a participant in the FUNDAEC courses. They seek a balanced
approach to change — characterized by a healthy relationship between cultural heritage
of a local population and modern science.
The sequence of ideas in the course are as follows: 1) Primary Production 2) Factors
That Determine Crop Production 3) Social, Economic, and Cultural Factors 4) Intro-
duction to Soils 5) Physical Properties of Soil 6) The Chemical Composition of Soil 7)
Fertilizers 8) Soil Biology 9) Preparing the Land 10) City Gardens 11) Seeds. Unique
to this unit are exercises dubbed “Practice and Experimentation,” which provides par-
ticipants with the opportunity to apply what they are learning and develop their skills
and abilities, and “Investigation”, which encourages them to broaden the scope of their
practical knowledge by, for example, interviewing farmers.
The central concern of this unit is on technology. The lessons seek to raise discussion
above the formula of “either-or” or even the harmonious co-existence of both. The unit
attempts to show students that in every region, including their own, a learning process
can be set in motion by which new knowledge is generated from the interaction between
the traditional knowledge system and modern science and technology, which can then be
applied to the problems of everyday life and used to promote the sound progress of the
region. The hope is that students will come away from the unit with the understanding
that such a learning process is one in which they, along with many others, can participate.

A.7 Example of Facilitation Style in Teacher Training

In “Properties,” participants gain insights into the importance of precision of language
in scientific thinking. After studying different parameters used to describe reality (e.g.,
temperature, color, size, etc.), teachers are asked to reflect on the concept of shape, one
such parameter. In an exercise, the tutor guides a conversation aimed at responding to
the veracity of the following statement: “Shape is a specific (not general) property of
substances.” Here is a snippet of the ensuing conversation recorded by one of the PIs:

Participant A It’s false.

Tutor Can you go ahead and explain a little more why you think it’s false?

Participant A As we were listing the specific properties of matter, we were listing shape
and size. The general properties are common. Shape is one that is common to
many substances. It seems every type of matter has shape. However, those specific
properties that allow us to distinguish one substance from another, shape is not a
type of property.

Tutor Hmm. Ok. How about others, what do you say?

Participant B For me, I say it’s specific. If you want to distinguish a table and some-
thing else then you use its shape.

Participant C Let’s say there’s a table that has triangular shape. Does it mean that
tables only have triangular shapes? What about oranges and apples? They have
the same shape. But they’re different substances.

xi
Tutor It’s quite challenging, isn’t it? If you think about it, it’s challenging to state
the reason why you can’t use shape or size to distinguish one thing or another.
However, the other participant is saying I can use shape to distinguish one object
from another. I can say that an orange is spherical and it helps me distinguish it
from something that is not spherical.

Participant C That’s true, but we’re examining general vs. specific properties. Her
statement is right, but we’re trying to talk about those properties that are specific.
When you talk about specific properties, they need to be such that they make the
substance different from others.

Tutor [sensing that participant C has grasped the concept, but others have not] Is shape
a general property of matter?

Participant D How do you distinguish a bus from the lorry? Don’t you use shape?

Participant E We’re not saying we can’t describe objects using general properties. But
there are those properties that further describe matter but they create a sense of
difference.

At this point, the tutor points the participants back to the use of the term spe-
cific and general properties in the text. It seems many in the group have reached
the understanding of participants E and C, but a few are still holding out.

Tutor Shape helps us describe substances, but is it a specific property of substances?

We’re not here for debating with anyone. We’re here to explore concepts for the
purposes of understanding. If you have reservations about what you are discussing
and it appears that the majority is on the other side which you are not yet it’s best
to put a star so that you come back and try to understand. You shouldn’t just
accept that the majority is right. You have to understand on your own. Let’s move
on.

xii
B Ugandan Educational Context
B.1 History of Schooling and Education in Uganda
One way of characterizing the history of schooling in Uganda is through the cyclical
manner in which education access and education quality have been prioritized across
different points in time. Consistent with global trends, the call of the moment is to
enhance education quality through the primary and secondary schooling pipeline. The
manner in which this increase in quality is to take shape is subject to intense public
discussion in Uganda.
To situate the context of the current moment in a broader perspective, it is help-
ful to understand the landscape of schooling in Uganda across time — we provide a
more detailed summary in appendix section B.1. Missionary groups in the late 19th cen-
tury spurred a nation-wide literacy movement in Uganda spanning many tribes (Hanson,
2010). During this period, Ugandans used informal networks to self-organize into com-
munity schools, allowing children, youth, adults and elders to study literacy side-by-side,
often before, during and after completing farming-related chores for the day. According
to British accounts, by the early 1930s, Ugandans had organized 5,673 community schools
in this manner, promoting the participation of scores of villages along the way.
Seeing a strength in Ugandans’ initiative to implement formal education, multiple
parties sought to move beyond basic literacy by increasing education quality: namely,
British colonialists sought to bring a standard of education that prioritized observational
science that was increasingly desired by the younger generations of Ugandans (Feierman,
1985, Ssekamwa, 1997). Education began prioritizing “practical” forms of knowledge such
as agricultural and other technical disciplines (Bude, 1983). To implement a policy of
curricular change, the Governor of the Colony began taxing the teacher “profession” and
used the revenues to transition 218 of the “best” community schools to this new form.
This tradeoff intended to enhance education quality at the expense of education access.47
In the lead-up to independence, it was clear that Ugandans again began to emphasize
increased schooling access to the broader population. This required increasing the number
of well-trained public servants and increasing government revenue to support trainings
and schools (Ssekamwa, 1997). Numerous conferences were held across Africa during
the same time to identify pathways forward that would combine indigenous schooling
approaches with the best of imported versions of schooling (Jansen, 1989). Ultimately,
however, the rapid pace of modernization led to the Ugandan government choosing to
expand the existing (colonial) educational systems rather than experimenting with new
forms of schooling (Brock-Utne, 2002).
The 1960s through the late 1980s was characterized by political instability in Uganda,
which sidetracked many of the planned educational initiatives. Towards the beginning
of the 1990s, Uganda began to re-prioritize educational access. They were among the
47
This change in the curriculum also had an impact on girl participation in schooling. For example, in
conversations recorded by the De La Warr commission 1937, Archdeacon Bowers, expressing the Church
Mission Society’s position stated “We feel very strongly that the education of girls should have a strong
bias towards the building of homes. The time has not come to give them an academic education,” to
which Martin Kayamba, a highly ranked Tanzanian in the colonial service, responded “In our tribal life
women are very important; we have women chiefs ruling people; women doctors; women agriculturalists
who support their families, and their influence is very strong not only in the homes but outside. I think
myself the education of women is just as important if not more than that of men.” The rise in gender
inequality in education during this period in Africa is also documented rigorously using econometric and
statistical methods in (Baten et al., 2020).

xiii
first countries to implement a policy of universal primary education (UPE) in 1997,
rapidly increasing the number of learners attending primary school — a 300% increase in
enrollment was observed between 1996 and 2004. However, the familiar tension between
access and quality re-emerged which was perhaps hindered by the separation of efforts to
increase access and quality of education in Uganda (Moulton, 2002).
Today, close to 100% of the population enrolls in primary school. The Ministry of
Education and Sports (MoES) is now intensifying its focus on increasing education qual-
ity, which is also an important factor in increasing access to secondary education.48 The
main avenue through which this focus is exercised is by considering the training and
incentive framework for teachers. This is reflected in the MoES’s publication on the
Teacher Initiative in sub-Saharan Africa (TISSA) (MoES, 2014) and the teacher incen-
tive framework (MoES, 2017), both of which emphasize an important role for recruiting
teachers with higher credentials and also systematic incorporation of teacher trainings
that can make teachers more effective, better compensated, connected to community
processes, and driven by a greater sense of purpose. Importantly, these initiatives are
taking place alongside a curricular reform of lower secondary school which emphasizes
“competency-based” education and teacher policies that are calling for programs that
promote continuous professional development. Thus, teacher trainings similar to the one
offered by KN are very timely.

B.2 Study Context — Jinja District, Uganda

We zoom in on Budondo sub-county in Jinja district to provide a more detailed picture
of the schooling landscape. In March-April 2018, we administered a community and
school census across 38 villages in Budondo, aimed at providing a rich perspective on the
patterns of schooling and economic activities of the households in this region. According
to this census, there are 8,982 households, split across 38 villages. The mean number
of members per household is 5.1 (median 5) and the mean number of children aged 16
or under, per household is 2.7 (median 2). The head of household is male in 75% of
cases and has an average age of 44.1 years (median 42). Additionally, 10% of households
have only one member, and both the head and spouse of the household residing in the
sub-county in 68% of household cases. Each household has lived in the same village for
19.0 years on average (median 14). A vast majority of the households have homes with
an iron roof (99.5%). The region is ethnically composed of Basoga, with a share of 73.5%
of households claiming Basoga heritage with clusters of Banyole (6.3%), Bagisu (5.1%),
Baganda (4.3%), Iteso (2.3%), and others. Most of the diversity is concentrated in the
trading center of Buyala, which contains three villages 3,230 households.
We turn our attention to the nature of schooling in Budondo. First, it is worth
noting that district education offices (DEO) are largely autonomous across Uganda in
administering primary education. They are tasked with delivering education using the
National Curriculum, and the district-level administrators have final say over the staffing
of schools in their district.49
Of the 44,445 individuals for whom we have age-data in Budondo sub-county, 75%
are below the age of 30 and 84% of children and youth between the ages of 5 and 20
attended some form of school in 2017 — be it nursery, primary, secondary or a form of
higher education. Figure B.1a shows the distribution of schooling by age and category
48
It should be noted that Uganda also implemented a universal secondary education in the early 2000s.
49
The district administrators themselves are accountable to the Primary Education Office in MoES.

xiv
 (a)

Figure B.1: Schooling in Budondo

within Budondo. A majority of individuals attend primary school by the age of 7 and
the percent of individuals attending “no school” is decreasing until roughly the age of 13
when it begins to increase again. Of the individuals between the ages of 13 and 20 who
are not in primary school, 33% no longer attend any form of school.
Figure B.2 shows the distribution of ages enrolled in different levels of schooling in
primary school.50 We can clearly see that there is a very large age variance in level
schooling and this variation widens as students advance in school levels. For example,
in Primary 1, there is significant density throughout ages 6 and 10, suggesting that
either students start Primary 1 quite late, or those who start around age 5 or 6 have a
tendency to repeat classes — most likely, both factors play a role. The distributions of
ages in Primary 4 (the transition year between primary and middle school) and 7 Primary
flattens relative to Primary 1. In Primary 7, there is high high density throughout ages
13-18. Below, we use features of the household census to show that a factor driving this
process is the very large repeat rate — households indicate that students in P1 through P6
repeat the previous year’s class between 14 and 18 percent of the time, a figure consistent
with UWEZO (2016) and our own student survey data we describe in further detail later
in the paper.51 There is no meaningful difference in the dropout, repeat, and transition
50
Similar distributions for secondary school are available upon request.
51
The repeat rate in P7 is slightly lower, however, it is worth noting that Gilligan et al. (2018) argue
that “Educational Triage” plays a role in the repeat rate at all levels of the primary schooling pipeline.
They argue that educators encourage their weaker students to repeat or drop out classes so that the
“performance” of the school in high-stakes primary leaving examinations improves. The indicator of

xv
 Figure B.2: Distribution of Ages Enrolled in Different Primary School Levels.

rates across girls and boys in primary school.

For each individual not in school in 2018, we also request information on the highest
level of schooling that individual reached. Thus, we can analyze the distribution of the
highest year of schooling achieved across four different cohorts (age 30-39; 40-49; 50-59;
60+) in Figure B.3 to understand how progression through school has changed across
generations. The distribution is very bi-modal for men, showing spikes at S4 and S6 as
we expect from the dropout rates discussed above. The distribution is becoming more
bi-modal over time for younger cohorts of women (in older cohorts, few women attended
secondary school). In summary, that data suggest an important role for enhancing educa-
tion quality in the early years that could address both the ends of participation and access
to schooling (especially at the higher levels of schooling) as well as quality of learning.

B.2.1 Dropout, Repetition, and Transition Rates

Table B.1 computes the dropout rate, repetition rate, and transition rate at each year
of school. To do this, we take advantage of the fact that the household roster in the
community census requests each child’s school history in the previous year as well as the
attendance at the beginning of 2018. We can define “dropout” as someone who was in
school in 2017 but said they were not attending school in 2018. The “dropout rate” for a
given year is then the numbers of dropouts from that year divided by the total number of
students in that year. The repetition rate is defined as the number of people who report
expecting to be in the same school level in 2018 as they were in 2017, divided by the total
number of people in that school level in 2017. Finally, the transition rate is defined as
achievement in PLE examinations is valued by district education offices and It is believed that these
measures of performance are key drivers of school choice.

xvi
Figure B.3: Distribution of highest year of schooling achieved by generation/cohort. Generations
are split into age groups of 10 years. Nearly everyone is finished with school by age 30.

the number of people who expect to be in a higher year than they were last year divided
by the total number of people in that year last year.52
Dropout rates are relatively low — ranging between 3 and 8 percent in primary school.
Dropout rates spike at S4 and S6 levels of schooling, reflective of the O-level and A-
level requirements of these school levels. They are also surprisingly balanced by gender.
Primary 7 has a higher dropout rate than most, but not overwhelmingly so.
Table B.2 gives a sense of the cumulative impact of these dropout rates: each row
produces the estimated share of an initial Primary 1 cohort that reach the row-specified
level of education. These numbers are sensitive to assumptions about how repeaters
behave: if we assume they never transition then only 20% reach to S1; if we assume they
always transition then 73% reach S1 (eventually). This table presents a somewhat middle-
road but we hope to eventually improve these estimates by using multi-year data where
we can actually track a cohort across time. With that said, the table shows that just
over half of the original cohort reach secondary school (the passthrough rate is actually
higher for women than men).
Finally, given our interest in parent-teacher interactions, it is also worth noting that we
asked household heads to answer the following question regarding each of their children
who attend school: “Have you interacted with [child’s] teacher?” Just over half (53%)
of the (non-missing) responses “Yes.” These “Yes” respondents are equally split between
52
It is worth nothing that the number of people who expect to be in school this year but were not in
school last year is very small, thus we ignore them in this analysis.

xvii
 Table B.1: School Continuation Table

Dropout Repeat Transition Number

Male Female Male Female Male Female Male Female
P1 .03 .03 .19 .16 .77 .80 1,071 965
P2 .02 .02 .14 .08 .84 .89 1,017 888
P3 .03 .03 .18 .15 .78 .81 966 911
P4 .03 .03 .18 .17 .78 .80 986 936
P5 .05 .03 .18 .19 .75 .78 849 865
P6 .08 .05 .18 .15 .73 .79 638 705
P7 .09 .09 .08 .11 .82 .78 480 565
S1 .04 .03 .02 .03 .93 .94 354 384
S2 .03 .04 .03 .03 .92 .92 289 350
S3 .04 .05 .05 .03 .90 .92 319 317
S4 .32 .37 .07 .09 .48 .42 283 314
S5 .04 .06 .00 .02 .96 .91 68 47
S6 .38 .38 .07 .03 .37 .42 86 64
Notes: The numbers reflect estimated probabilities of dropping out, repeating, or
transitioning for each school level according to survey responses from the household
census. Details for how the calculations are made are described in the text. We
distinguish each category of variables by male and female students. N refers to the
number of individuals in that school-level in the previous year. The rows do not
always sum to 1, especially in Secondary 6, because individuals sometimes report
that they had yet to decide their schooling status in 2018 and also because of a
small number of people who state that they transition to a lower class.

male and female students.

B.3 Additional Socio-Economic Context in Budondo

Table B.3 shows the distribution of occupation for each member of the household, as
specified by the household head, at a series of different age categories. Budondo is
commonly referred to as the “bread-basket” of Jinja. This is evident from the distribution
of individuals who state that their profession is “Farmer.” However, these categories are
not mutually exclusive — an individual can select more than one category. The large
share of people reporting "Other" in the 20-29 category are largely women.
Figure B.4 shows the main composition of ethnicities for each of the 38 villages. Basoga
clearly dominates in general, though not in every single village. Figure ?? shows a similar
picture, but for religions rather than for ethnicities. The distribution varies a little bit
across villages, but all villages have some representation of the major religions. Just
over one-third of people are Muslim, and most of the rest are Christians split across a
few denominations. Figure B.6 shows the same picture for a household’s main source of
power. There is a fair split across most villages, and it is clear that several villages are
not electrified whatsoever.

xviii
 Table B.2: Cumulative School Continuation Table

Leftover Share (M) Leftover Share (F)

Primary 1 1 1
Primary 2 .95 .96
Primary 3 .93 .93
Primary 4 .87 .89
Primary 5 .81 .84
Primary 6 .73 .78
Primary 7 .62 .71
Secondary 1 .55 .62
Secondary 2 .52 .59
Secondary 3 .5 .56
Secondary 4 .47 .53
Secondary 5 .27 .27
Secondary 6 .26 .25
Post Sec. 6 .12 .13
Notes: Rates are sensitive to assumptions about
repeaters as described in the text.

Figure B.4: Distribution of ethnicity by village. The ethnicities are grouped into the four biggest
ones and ’Other’ and the villages are sorted by the size of the village. Note that this is done at the
individual, rather than household, level.

xix
 Table B.3: Occupations by Age

Age Category: 15-19 20-29 30-39 40-49 50-59 60+

Farmer .12 .39 .58 .66 .7 .61
Teacher 0 .02 .03 .03 .02 .01
Taxi/Boda-boda driver .01 .05 .07 .06 .03 .01
Truck driver 0 .01 .01 .02 .01 0
Artisan 0 0 0 0 0 0
Construction worker/Mason .01 .05 .05 .03 .02 .01
Factory worker 0 .01 .01 .01 .01 0
Carpenter/Carver 0 .01 .01 .01 .01 0
Herbalist 0 0 0 .01 0 0
Hairdresser .01 .03 .01 0 0 0
Health worker 0 .01 .01 .01 .01 .01
Office worker 0 .01 .01 .01 .01 0
Shop attendant/Trader .01 .08 .11 .11 .09 .04
Baker/Cook 0 .02 .03 .02 .01 0
Priest 0 0 0 0 0 0
Elder 0 0 0 0 .01 .06
Student .66 .11 0 0 0 0
Works in household .02 .02 .02 .01 .01 0
Broker/Middleman/Property Master 0 0 0 0 0 0
Mechanic .01 .03 .02 .02 .03 .01
Fishing 0 .02 .02 .01 0 0
Tailoring 0 .02 .02 .02 .02 .01
Sugar Cane Loading .01 .02 0 0 0 0
None .14 .17 .08 .07 .08 .23
Other(specify) .01 .03 .03 .03 .03 .03
N 5,646 6,744 3,914 3,177 2,085 2,309
Notes: Rows need not sum to 1 as multiple occupation selections are possible.

xx
Figure B.5: Distribution of religion by village. The religions are grouped into the four biggest ones
and ’Other’ and the villages are sorted by the size of the village. Note that this is done at the
individual, rather than household, level.

xxi
Figure B.6: Distribution of main source of power by village. The power sources are grouped into
the four biggest ones and ’Other’ and the villages are sorted by the size of the village. Note that this
is done at the individual, rather than household, level.

xxii
C Measurement of Outcome Variables
We will organize the description of measurements according to the research questions
articulated in section 2.2. Each of these measures will be constructed using data from
the above-mentioned sources. All variables described below are also described in similar
fashion in our pre-analysis plan. We refrain from providing summary statistics of each
variable in a separate table since we present the number of observations and control school
means for each variable in our results table.

C.1 (Q1) Does the teacher training change teacher pedagogy?

(H1.1) Classroom Pedagogy Improves.
Using the Stallings classroom observation tool and student survey responses, we con-
struct measures of student engagement and participatory pedagogy in school.53 We con-
struct the variables under each outcome measure for hypothesis (H1.1) in the following
manner.
• Student Engagement (Stallings): After identifying the activity the teacher is
engaging students in, the enumerator needs to specify a response to the question
“How many pupils are engaged in the activity with the teachers?” The enumerator
can respond with 1) “No pupil,” 2) “One pupil,” 3) “A few pupils,” 4) “Half of the
pupils,” 5) “Most of the pupils,” or 6) “All pupils.” 54
• Students ask questions in class (Teacher Outcome from Student Survey):
We ask students to indicate whether they asked teachers any questions during
class (in terms 2 and 3) to help them with a problem they didn’t understand.
Respondents selecting “yes” continue to indicate for which teachers this statement
is true. We sum student responses by teacher and divide by the number of students
surveyed. Formally, and for each school k, let yijk be equal to one if teacher i was
mentioned Pby student j and zero otherwise. The variable we construct is equal
to yik = j yijk /Nk where Nk reflects the number of students surveyed in school
k. We restrict analysis of this variable to those teachers who teach P6 students as
indicated by at least 90% of surveyed students mentioning a teacher in this way.
• Corporal Punishment (Teacher Outcome from Student Survey): We gently
ask students to indicate whether a teacher has “caned” or beaten him or her. We
stress to the students that they do not have to respond to the question if they do
not wish to, but that their responses will remain confidential no matter how they
respond.55 If the students say “yes,” we ask them to indicate which teachers caned
or beat him or her. We construct teacher-level outcomes in the same manner as the
“Students ask questions in class” variable above.
(H1.2) Teacher Motivation and Effort:
53
We will include TIPPS measures related to critical thinking pedagogy and pedagogy facilitating
deeper learning and understanding of concepts when the data become available, likely in mid-2021.
54
We follow the instructions in World Bank Group (2015) and emphasize that even if only one student
is involved in an activity, students are coded as “engaged” if they are paying attention to the activity.
For example, if a single student is asked by the teacher to read a passage out loud, all students are
considered “engage” if they are at least listening attentively to the passage being read.
55
“Caning” is a common term used for using something like a thin bamboo stem to hit a child on his
or her head as a form of corporal discipline.

xxiii
 • Teacher Attendance (School Admin Data): Schools are required to collect
daily teacher sign-in sheets, however not all schools do so. For the schools that
collect this information, we sum and then digitize the days of teacher attendance
per term at the teacher level.
• Teacher Knowledge of Student (Teacher Outcome from Student Survey):
In the student survey we ask students in P6 to indicate their attendance at the
school during the prior two school days. We additionally ask these students to indi-
cate their relationship with the people they live with. Then, in the teacher survey,
we randomly selected six of the students we surveyed and added two additional
names of students who do not attend the teacher’s school (hereafter dubbed “fake
students”). We first ask the teacher to identify whether the named student attends
the school (question 1). If the teacher responds “yes” to this question, we proceed
to ask, questions 2, 3 — whether the named student was present in school during
the previous 1-2 school days (one question for each school day) — and question 4
— what is the relationship that the student has with the people he or she stays
with. We match teacher and student responses and create an index reflecting the
accuracy of the teacher’s response

Formally, and for each school, let qijk reflect teacher i’s response regarding student
j for question k = {1, 2, 3, 4} — student j’s response for question k is reflected by
sjk . The index we construct amounts to the percent of teacher i’s correct responses
across all students and questions:

k 1(qijk = sjk )
P P
j
Teacher Knowledge of Studenti = P P , (3)
j k1

where 1(qijk = sjk ) is equal to one when the teacher’s response matches the student’s
response and zero otherwise.

(H1.3) Teacher Collaboration:

• Teachers learn from and collaborate with colleagues ( Teacher Survey):

Recall that to compile our sample for teacher interviews, we acquire a complete
list of teachers in each school in the second term of a given school year and inter-
view upper primary teachers in the third term of that school year. We pre-load
teacher names in each school into our teacher survey and ask each interviewed
teacher to respond to the following three questions regarding classroom and school
learning/collaboration with other teachers:
1. Have you and [colleague] spoken about how to better manage your classrooms?
2. Have you or [colleague] visited any of each others classroom to help improve
teaching practices?
3. Have you planned classroom activities together with [colleague]?
Thus, for a given teacher i and colleague j we observe i’s description of the sym-
metric relationship she has with j — yijnet,k for questions k ∈ {1, 2, 3}. Importantly,
if school s has a sample of Ns teachers, then we observe both i’s responses (yijs net,k
)
and j’s response (yjis ) for each ij or ji pair in {1, 2, · · · Ns }. To add precision to
net,k

xxiv
 our measures, we only consider affirmative responses in which ijs and jis agree.56
In other words,
net∗,k
yijs net∗,k
= yijs net,k
= (yjis = Y ∩ yijs
net,k
= Y)
(4)
For all k in {1, 2, 3} and i, j in Ns

where (yjis
net,k
= Y ∩ yijs
net,k
= Y) indicates that both i and j answered “Yes” to
social network question number k.
The final variable we construct for the pre-analysis plan is the dyad-specific average
response for each question yijs
net,k
. In other words, we analyze

1 X
net
yijs = net∗,k
yijs (5)
3
k∈{1,2,3}

C.2 (Q2) Does the teacher training change student outcomes?

(H2.1) Traditional learning outcomes increase (PLE scores, Student Assess-
ment and Student Survey):

• Primary Leaving Examination (PLE) at the end of P7 (UNEB and School

Admin Data): As described above, we collect school records on PLE registration
and combine it with official data from UNEB. Lower scores indicate higher perfor-
mance and range from 1 to 9 for each subject. To ease interpretation, we flip this
relationship and make 9 the highest performance and 1 the lowest performance.
Furthermore, we aggregate scores from each of the four subjects in a manner that
reflects the official approach taken by UNEB to determine student performance.
Therefore, our aggregate measure ranges from 4 (lowest performance) to 36 (high-
est performance) which we standardize according to the control school mean and
standard deviation.

• Student Pass-through Rates (Student Survey): In the student survey, we

ask students to indicate the grade they are currently in, the grade they were in
one year prior and the grade they were in two years prior. We indicate that a
student has had a successful pass-through rate if the most recent year’s grade is
one (or more) grade ahead of the previous year’s grade. Formally, and for example,
a successful pass-through for student i in school s in time t (t reflects the school
year) is characterized by
pass
yist = 1(Gradet − 1 >= Gradet−1 ) (6)

where 1 reflects an indicator variable equal to one if (Gradet − 1 >= Gradet−1 ) is

true and zero otherwise.57
56
A subset of sampled teachers are not included in our pre-loaded survey program because they were
not listed as school teachers by school administrators prior to term 3. Therefore, for these teachers
net,k net,k
we only have observations for either yijs or yjis but not both. We exclude these teachers (whether
present in i or j) from our analysis which preserves the precision of our measure and imposes symmetry
on dyadic responses.
57
We furthermore ask whether students attended the same school in the prior year. In the pre-analysis
plan we will pre-commit to analyzing this variable for all students in our sample, though exploratory

xxv
 • Researcher-provided Student Assessment (Student Assessment):

(H2.2) Higher order learning outcomes increase

• Measuring Recall, Understanding and Critical Thinking (Student As-

sessment): Close to 90% of assessment questions in developing countries, as ana-
lyzed by Burdett (2017), test students’ ability to recall and memorize information.
Thus, when analyzing student learning using official examinations such as PLE,
researchers are most likely correlating student recall of facts and information with
treatment status. These are important features of student learning, but not nearly
enough to be prioritized to the extent that they are. We analyzed a common assess-
ment booklet in Uganda, and tried to identify the questions that follow elements of
Liu et al.’s (2014) conceptual framework for critical thinking assessments. Specifi-
cally, questions that: 1) evaluate evidence and its use, 2) analyze and evaluate ar-
guments, 3) understand implications and consequences, 4) develop sound and valid
arguments, 5) understand causation and explanation. These elements cohere with
Burdett’s (2017) description of reasoning/critical thinking. Burdett (2017) classi-
fies questions into three categories of learning according to whether they measure
recall, understanding (or application of knowledge), or reasoning/critical thinking.
We found very few questions that address the last category of questions in SIPRO’s
(2019) standard assessment. This required us to construct our own questions mea-
suring reasoning and critical thinking, which we did by following principles of critical
thinking assessment design discussed in Liu et al. (2014). For further details on our
approach, see appendix section D.

(H2.3) Field-based scientific competency increases (Science Shows):

• Science Shows. The Ugandan National Curriculum Development Centre (NCDC)

is pushing a curriculum transition towards a “competency-based” framework for edu-
cation, starting with curriculum reform in secondary education. Recently, however,
it began specifying expected competencies in students in primary training as well.
Given the national emphasis on scientific competencies, and together with the Jinja
District Education Office (DEO), we develop a field-based measure of testing sci-
entific competencies in students across all of the schools in our study area. This
activity is designed to measure the following competencies articulated by NCDC’s
(2016) curriculum guide for Primary 6 students:

Pupils apply correct scientific processes in investigations of various phe-

nomena by: 1) identifying problems, 2) designing and practicing scientific
investigation processes, 3) examining the evidence useful in inferences, 4)
demonstrating the skills of observation, classification, accurate measure-
ment and recording, 5) making predictions and formulating hypothesis
for evidence, 6) communicating findings accurately and honestly, 7) an-
alyzing causes and effects, 8) using a variety of sources for acquiring
information, and 9) recording information with reasonable accuracy.

We develop a 12-category rubric that measures each of these competencies on a

scale of 1 to 10, with 10 reflecting better performance. Extensive details of the
analysis will analyze whether this variable responds differently when the analysis only includes students
who attended the same school year after year.

xxvi
 approach taken by the DEO to invite the schools to participate in science shows,
preparation of judges, and the events themselves are provided in appendix section
E.
After the science show. The judges provide scores to the research team and the
DEO. The research team aggregates the scores of each group by taking the mean
across all twelve quantitative variables in the judges rubric represented in Table
E.1.

(H2.4) Student creativity increases (Student Survey)

• Creativity Score Student Survey : We utilize a well-known tool designed by

psychologists to measure student creativity. We follow the procedure implemented
by Bradler et al. (2019), who utilize this measure in an experimental economic
setting. The method effectively involves a one minute exercise in which students
generate a list of creative uses of a common object in their environment. We repeat
this exercise for three objects: plastic bottles, maize plant residues, and bamboo
stems. Once data are collected, researchers construct a creativity score by giving
a point for each feasible use of the object and multiplying points according to the
degree of uniqueness in the gathered data. Details of this procedure are outlined in
appendix section F.

Additional Long-Run Student Learning Outcomes

In order to match students in secondary schools to a primary-school treatment status,
we use a fuzzy merge procedure to match student identifiers in the secondary assessment
with details collected in the census. We know, for example, that 3.428 students from
census households attended lower secondary school in 2022. Of these, 1,206 students
attend a secondary school in our new sample. Of these, 655 students attended a primary
school in our sample. We are able to match

xxvii
D Researcher Provided Student Assessment
The following pages display the student assessment as received by the students following
enumerator administration of the student survey. Each individual students participating
in the survey sits independently on a bench nearby the school. Enumerators were en-
couraged to tell the students to “try their best” if they had any questions regarding the
survey, though such questions were minimized during the piloting phase of the student
assessment.
Using SIPRO (2019), and recognizing that our sample only includes P6 students at
the beginning of term 3, we borrow questions from a subset of the term 2 middle-of-term
assessment questions sold to schools for the researcher-administered assessment.
These assessments rarely contain questions relevant for measuring higher order learn-
ing. To address this shortcoming, we designed several questions that intend to measure
higher-order learning principles described in Liu et al. (2014). Specifically, we tried to
identify one question for each of the following elements of Liu et al.’s (2014) conceptual
framework for critical thinking assessments: 1) evaluate evidence and its use, 2) analyze
and evaluate arguments, 3) understand implications and consequences, 4) develop sound
and valid arguments, 5) understand causation and explanation. These elements cohere
with Burdett’s (2017) description of reasoning/critical thinking.
Appendix section D displays the format we used for the student assessments. Ques-
tions 1 through 5 were borrowed directly from SIPRO (2019) while questions 6 through
9 were generated by the research team. Students fill out the assessment on paper and
enumerators enter student answers into our pre-programmed tablet-based form. Enumer-
ators were also trained to assess the completeness of sentences and grammar structure
of responses in question 4. Appendix Table D.1 demonstrates how we classified each
assessment question according to Burdett (2017).
We utilize factor analysis to predict three coarse learning outcomes based on our
approach to classifying assessment questions. The three outcomes reflect different dimen-
sions of student ability: 1) recall, 2) understanding and 3) reasoning or critical thinking
ability. Out of the 23 opportunities to receive points from the SIPRO (2019) exam, 52%
fall under recall, 39% fall under understanding and 9% fall under reasoning.
The questions we add to the student assessment all fall under the reasoning category.
However, since such questions are rare in the Ugandan context, we only use them to
predict reasoning ability if they achieve a benchmark threshold of validity.58 Specifically,
since 2 of the questions from SIPRO (2019) represent the reasoning criteria, we analyze
correlations between these two questions and each of the seven questions we construct
ourselves. If the average correlation across the two questions that provide benchmark for
our “reasoning” criteria is less than 0.05, we exclude the question from our construction
of the reasoning variable.
Table D.1 reflects how each of the questions in the assessment, and the modes of
student response, were classified to reflect three categories of learning: 1) student recall
and recognition of information (recall), 2) student’s ability to understand and apply a
concept (understand), 3) students ability to analyze, and evaluate an argument (reason).
Throughout, we roughly follow Burdett (2017) when analyzing each question type and
classifying them for our own analysis.

58
Indeed, Burdett (2017) shows that such questions are very rare throughout the African continent,
though Uganda utilizes such questions with greater frequency than its sister countries across Africa.

xxviii
 Broadest Classification of Assessment Questions: Level 1 - Recall Level 2 - Understand/Apply Level 3 - Reason
Second level of classification: Remember Understand Apply Analyze Evaluate Create
Third level of classification: Recognizing Recall Interpreting Exemplifying Classifying Summarizing Inferring Comparing Explaining Executing Implementing Differentiating Organizing Attributing Checking Critiquing Generating Planning Producing
Question Text SIPRO (2019)? Subject Classification
1.a) Give any two examples of leguminous crops. Science Recall
1.b) Mention any two ways of caring for crops in the Science Recall
garden
2.a) Name two examples of vitamin deficiency diseases Science Recall
3.a) Which part of a plant carries out reproduction? Science Recall
3.b) Which part of a plant provides attachment to Science Recall
branches?
4.a) At which school was the fire outbreak? English Understand
4.a) ↑Proper Sentence Structure English Recall
4.b) On which day did the fire break out? English Understand
4.b) ↑Proper Sentence Structure English Recall
4.c) Whose dormitory got burnt? English Understand
4.c) ↑Proper Sentence Structure English Recall
4.d) Where were the children when the fire broke out? English Understand
4.d) ↑Proper Sentence Structure English Recall
4.e) Why was the police not able to save the property? English Understand
4.e) ↑Proper Sentence Structure English Recall
4.f) What did the fire brigade officers tell the school English Understand
xxix

management?
4.f) ↑Proper Sentence Structure English Recall
4.g) Where was the school advised to put smoke de- English Understand
tectors?
4.g) ↑Proper Sentence Structure English Recall
5.a) How many patients were admitted on Sunday? Math Understand
5.b) Which day had the least number of patients ad- Math Understand
mitted?
5.c) How many more patients were admitted on Math Reason
Wednesday than Thursday?
5.d) Find the total number of patients admitted in the Math Reason
whole week.
6.a) Which statement best supports Sarah’s thoughts? Critical Thinking Reason
6.b) Which statement best supports Aminah’s Critical Thinking Reason
thoughts?
6.c) Which statement best supports both Aminah and Critical Thinking Reason
Sarah’s thoughts at the same time?
6.d) Which statement does not support either Aminah Critical Thinking Reason
or Sarah’s thoughts?
7) Sarah’s Evaluation of a statement Critical Thinking Reason
8) Aminah’s Causal Reasoning Critical Thinking Reason
9) Correlation vs. Causation Critical Thinking Reason

Table D.1: Classification of Assessment Questions

School:______________________________________________________

Name:_______________________________________________________

Class:_______________________________________ Stream:___________

Age: ___________________ Sex: (Male OR Female – Circle one)

Village Where You live: ____________________________________________

Name of parent/guardian: _______________________________________

Your position in class at the end of term 2: ___________________________

1. A. Give any two examples of leguminous crops.

i. _____________________________________________________________________________

ii. ____________________________________________________________________________

B. Mention any two ways of caring for crops in the gardens.

i. _____________________________________________________________________________

ii. ____________________________________________________________________________

2. A. Name two examples of vitamin deficiency diseases.

i. _____________________________________________________________________________

ii. ____________________________________________________________________________

3. A. Which part of a plant carries out the following activities?

i) reproduction. _________________________________________________________________

ii) provides attachment to branches._________________________________________________

B. State any two uses of leaves to plants.

i. _____________________________________________________________________________

ii. ____________________________________________________________________________
 4. Read the passage below carefully and then answer, in FULL SENTENCES, the questions that follow.

It was a sad Saturday morning at Mogga Primary School last term. That morning shocked everyone with the
sudden burning of one of the dormitories for Primary Four children. Fortunately, by the time of the fire
outbreak, all of them were in class writing weekly tests, so no children were injured but a lot of property was
burnt. The Police Fire Brigade rushed to the scene to save property but failed because by the time it arrived,
most properties had burnt to ashes.

The fire brigade officers advised the school management to put up enough fire fighting equipment to use in
case of a fire outbreak. The burnt dormitory had only one fire extinguisher which made it hard for people to
stop the fire. The school did not have smoke detectors so they were advised to put smoke detectors on every
building in the school.

a) At which school was the fire outbreak?

______________________________________________________________________________

b) On which day did the fire break out?

______________________________________________________________________________

c) Whose dormitory got burnt?

______________________________________________________________________________

d) Where were the children when the fire broke out?

______________________________________________________________________________

e) Why was the police not able to save the property?

______________________________________________________________________________

f) What did the fire brigade officers tell the school management?

______________________________________________________________________________

g) Where was the school advised to put smoke detectors?

______________________________________________________________________________
 5. The bar graph below shows the number of patients admitted to the life care medical centre in a week. Study and
use it to answer the questions that follow.

Number of Patients Admitted to Medical Centre

50
45
40
Number of patients

35
30
25
20
15
10
5
0
MON TUE WED THUR FRI SAT SUN
Days in a week

a) How many patients were admitted on Sunday?

______________________________________________________________________________

b) Which day had the least number of patients admitted?

______________________________________________________________________________

c) How many more patients were admitted on Wednesday than Thursday?

______________________________________________________________________________

d) Find the total number of patients admitted in the whole week.

______________________________________________________________________________
Question 6: In the survey, you shared your thoughts about Sarah and Aminah and how they think about
education. Remember,

• SARAH BELIEVES: the purpose of education is to help people get jobs.

• AMINAH BELIEVES that the purpose of education is to help people make positive changes in their
communities.
Think about these sentences.
1. 1 out of 5 (20%) of students repeat a class each year.
2. It is easier for someone with education to find a job.
3. Without education, it is difficult to improve your life.
4. In every village, the most helpful people are those with the most education.
According to you, which sentences belong in each of the blank spaces below?
(Choose one sentence for each blank space and write the sentence number in the blank)
a.) Which statement best supports Sarah’s thoughts? _______________________________________

b.) Which statement best supports Aminah’s thoughts? ______________________________________

c.) Which statement best supports both Aminah and Sarah’s thoughts at the same time? ____________

d.) Which statement does not support either Aminah or Sarah’s thoughts? _______________________

Question 7: SARAH heard that the following statement is TRUE.

1. People who are educated have more skills BECAUSE of their education.
SARAH decides that this also means:
2. People with skills are more educated BECAUSE of their skills. (CIRCLE ONE)
Is SARAH correct in making this decision?
CORRECT NOT CORRECT

Question 8: Aminah believes the following two statements are TRUE.

1. People who are educated serve their community BECAUSE of their education.

2. People who serve their community create peaceful communities BECAUSE of their service.
Thinking only about AMINAH’s beliefs, would AMINAH say the following:
3. People create peaceful communities BECAUSE of their education. (CIRCLE ONE)

YES NO
Question 9:
In many parts of the world, the following statement is true:
1. Schools that have poor performance receive more teachers to help students perform.
Is statement number 2 bellow CORRECT or NOT CORRECT?
2. If the number of teachers in a school decreases, the school’s performance will improve.

CORRECT NOT CORRECT

E Judging the Science Shows
Prior to the science show. In order to measure these competencies we worked with
the DEO to organize science shows at each of the schools in our study and trained judges
to utilize a rubric that measures student competencies along each of these criteria. We
describe each of these activities in turn below.
First, NCDC (2016) articulates a set of activities that P6 students could implement
in the course of a science show. In term two and three science includes lessons on natural
resource conservation (with an emphasis on soil) and clean water preparation (for drink-
ing) at home. Together with the DEO, we drafted a letter to each school that, in part,
reads:

“Each stream of primary 6 class will organize 3 or 4 groups of students who

will prepare a science-oriented activity on one of two subject competencies:

1. “Experiments that compare at least two different approaches of conserv-

ing resources in the community such as soil, minerals, fuel, water, and
air.
2. “Experiments that compare at least two different approaches of preparing
clean and safe water for drinking and washing.”

This letter is sent to schools towards the end of term 2 (typically July) and schools
are informed that science shows will take place in the middle of term 3 (October), giving
them three months to prepare. The letter states that each group of students will present
the results of their experiments in front of pre-appointed science show judges and each
presentation should take 10-15 minutes with an expected 5 additional minutes of questions
and answers with the judges.
For schools participating in the science shows, the student surveys ask students to
describe 1) the research question they are exploring in their science show project, 2)
the hypothesis around the question, 3) why they think the hypothesis might be true,
4) which two approaches they will be comparing in their experiment and 5), what they
should observe from these approaches according to their hypothesis. The purpose for these
questions was to ensure that schools were preparing adequately for the science show and
also to provide the team of researchers material that could be adapted to utilize during
the training of science show judges.
The rubric provided to the judges is displayed in appendix E and is shared with each
school as an attachment to the letter. We measure each of NCDC’s desired competencies
on a ten point scale using twelve questions. Table E.1 summarizes the rubric item used
to assess each of these competencies. Our aim is to make each measure as objective a
description of each competency as possible. Schools are also told that the DEO will invite
a selection of community members to attend each show on the specified day and time
and the highest ranking school will receive a certificate from the DEO that will also be
displayed at the district office headquarters.
One week prior to the training, IPA Uganda helped identify a pool of six highly
qualified enumerators to act as science show judges who participated in a training. The
researchers modified student survey responses that emulated possible science shows the
team would be expected to judge in the following month and asked each of the judges to
provide scores for each of the items on the rubric. For three days, judges spent one hour
assessing a handful of mock science shows using the form containing the rubric. Each of

xxxv
 Framing
1. Identifying the problem: The pupils clearly stated the 2. Relevance of the problem: The pupils clearly stated why
problem they are hoping to address with their project. the problem is an important one to address.
1 = Did not state the problem at all. 1 = Did not indicate why the problem is important.
5 = State a problem, but not clearly. 5 = Indicated why the problem is important, but not convincingly.
10 = Clearly stated the problem. 10 = Convincingly indicated why the problem is important.
Experiment
3. Designing an experiment: The pupils conducted an experi- 4. Designing an experiment: The experiment was very cre-
ment that clearly outlines how they would study the problem ative — it tries to learn about the problem in a way that
they identified. others have tried before.
1 = Did not conduct an experiment at all. 1 = Others have tried to address the problem in a similar manner.
5 = Conducted an experiment that was not clear. 5 = The approach taken was creative, but not original.
10 = Conducted a very clear experiment. 10 = The approach was creative and original (never seen before).
5. Designing an experiment: The pupils tested a technology
not seen before.
1 = The technologies tested are known by all.
5 = The technologies are known by some but not all.
10 = The technologies were completely original and creative.
Hypothesis
6. Describing a hypothesis: The students had a clearly artic- 7. Analyzing a hypothesis: The students clearly linked their
ulated hypothesis. hypothesis to the design of the experiment.
1 = I had no idea what hypothesis the students were testing. 1 = There was no link between the hypothesis and the experiment.
5 = The students mentioned a hypothesis, but it was not clear. 5 = The link between the hypothesis and experiment was there but not
clear.
10 = The students mentioned a very clear hypothesis, comparing at 10 = It was very clear how the experiment would address the hypoth-
least two groups. esis.
Measurement
8. Making observations: The students captured observations 9. Making observations: The students made accurate mea-
in a systematic manner (for example, by using a logbook). surements of their observations using relevant instruments.
1 = There is no evidence that the students recorded observations. 1 = The students did not make measurements.
5 = Students recorded observations, but not in a structured way. 5 = The students made measurements that were not accurate.
10 = The observations were recorded in a very structured and system- 10 = The students made very accurate measurements.
atic way.
10. Sources of information: The students used a sufficient
number of information sources to make their conclusions (for
examples, they ran the experiment more than once or they
measured more than one outcome.
1 = The students did not use any sources of information to support
conclusions.
5 = There were at least two sources of information in support of con-
clusions.
10 = The students made every effort to support conclusions with all
available evidence.
Articulating
11. Communicating information: The students communicated 12. Communicating information: The students communicated
their findings independently. their findings accurately.
1 = The students could not describe what they did without help. 1 = The students looked like they did not understand what they were
saying.
5 = The students described their work sometimes on their own, some- 5 = The students understood what they were saying, but it wasn’t
times with help. always connected to what they did and learned from their experiments.
10 = The students were the only ones describing their work. 10 = The students understood their work and connected it to their
experiment.

Table E.1: Science Show: Judge’s Rubric

the judge’s scores was entered into an excel spreadsheet and compared with one another.
A PI for the project then identified questions where there was disagreement around the
responses provided by each judge. Though imperfect, this process allowed the PIs to
identify judges who could articulate the reasons behind the scores they provided in a
clear and sound manner. Through this process, two of the judges were selected to work
as judges for the science show on behalf of the PI and the DEO.
The day of the science show. Science shows were held in two to three schools

xxxvi
each day from middle of October 2019 until the beginning of November 2019 (9:00 AM,
11:30 AM, and 2:30). The team of science show judges and two assistants (who also
serve as backup judges) arrived at each school one hour prior to each show to assist
the headteacher with logistics. In a typical science show, the group of students tasked
with presenting their experiment lead the judges and community members to the site
of their experiment — agricultural experiments were often implemented outdoors near
the school’s garden plot. Groups differed in the modality of their presentation — some
selected a group leader who was tasked with presenting the work of the group, others
presented their experiment as a team. Regardless of the mode of presentation, the judges
were trained to ask a set of questions that they would direct to the entire team and that
they would use in scoring each presentation. These question types are also articulated in
the judge rubric in appendix E and are summarized by the following four items:

1. When did you start preparing for this activity? How did you come up with the
question that you studied?

2. Describe how you decided to investigate the question?

3. How did you measure outcome [specify the outcome from the experiment context]?

4. [Instruction: Provide an alternative explanation for the observations made by the

students. Ask the students how they would respond to this explanation.] 59

After engaging with each of the presentations, the judges invite the teachers and
head teachers to address the students and community members. They then close by
expressing their appreciation for the student and school’s efforts and stating that the
DEO will communicate the results of the shows at a later date.
For the sake of transparency, we present qualitative comments provided by judges in
Tables E.2 and E.3 for the top 10 and median 10 group performances in the science shows
alongside their overall ranking. The following pages display the rubric provided to science
show judges for the purpose of assessing student performance in the science shows.

59
This exercise was incorporated into the training of the science show judges prior to the science shows.
The training was thorough enough that the judges developed a standard set of alternative explanations
for most student experiments.

xxxvii
 Table E.2: Top 10 Student Performance in Science Shows

Rank Score Research Hypothesis Research Experiment

Top 10 ranked Science Groups
Filtered water is clean compared to un- The pupils collected muddy water and a water bottle and cut off its bottom and turned it upside down. They put cotton wool at the
1 110
filtered water bottom [of the bottle], added gravel, coarse sand, fine sand, and charcoal pest then poured muddy water through the bottle.
Filtered water is clean as compared to They got cotton placed them in a container, added gravel stones, coast sand, fine sand,charcoal paste. Added a container of muddy water
2 109
unfiltered water and placed on the Stand with another container below to receive the clean water.
Mulched soils are not carried away by The pupils obtained two soil samples and put them into two boxes. They covered one soil sample with green grass. After, they poured
3 100 running water compared to the soils water onto the covered soil sample to determine how much soil will be lost and put a basin below the soil sample where water was dropping.
that are left bare They then poured water on the soil which was not covered with grass.
Mulched soils are not carried away by The pupils obtained two soil samples and placed them in two different containers. They left one of the container bare and the other covered
4 100 running water while compared to soils the soil with grass. They then poured water in both containers and placed basins below to help capture the soil that was lost in both
that are left bare containers
Experiment 1: Decanting- They collected muddy water in a container, prepared 3 containers and poured muddy water into the first container
and left it to settle for some time. They then transferred the water into the second container and left it to settle again. They then poured
Boiled water is safe for drinking com- the water into the third container and later poured it into the kettle. They put the water on fire and left it to boil and later poured the
5 93
pared to filtered water water into a clean container. Experiment 2: Filtration collected muddy water. Students got a funnel, a clean container, and a filter funnel.
Poured the water into clean container through the funnel to remove the residue. They then poured the water into the kettle to boil it.
After boiling they put the water into a clean container and covered it.
xxxviii

Collect muddy water in a container and leave it to settle as you prepare 3 clean containers. Pour the muddy water in the first container
leaving the dirt below and leave it for sometime to settle. Transfer the water carefully into another basin leaving the dirt within and wait
Boiled water is clean and safe for drink-
6 88 for it to settle, put the water in the third basin and then put it in the kettle then boil it. Let it to cool for sometime and then put it in Ina
ing compared to filtered water
covered container ready for drinking Get two containers and put muddy water in one of the basins then use a sieve and pour the water into
the other container and then boil the water and put it in the container covered ready for drinking.
Soils that are covered have more mois-
7 86 ture and nutrients compared to soils They got maize stocks, banana leaves, and dry grass and covered the soil.
that are not covered
Filtered water is clean compared to un- [The students] got muddy water and put a piece of cloth through the funnel. They then poured dirty water into the container through the
8 85
filtered water funnel. They repeated the process a second time.
Wastes like banana peelings, cassava
peelings and potato peelings decay to The pupils collected different items like banana peelings, potato peelings, ash, cassava peels, egg shells, kaveras, broken bottles plastics ,
9 83 make soil fertile other than broken bot- old clothes and shoes and a container having soil. They got banana peelings which they put in the soil which decayed to make manure and
tles and kaveera that instead damage made the soil fertile.
the soil
Filtered water is clean compared to wa- The pupils got a cup and placed in it mudy water , placed a clean cloth on the cup and poured water though Into another container. He
10 82
ter that is not filtered removed the dirt from the cloth and then poured the water through again.
Notes: Responses were collected from science show judges as described previously in the paper. Science group participants were graded out of 120 points on their ability to conduct scientific experiments and present their findings
with clarity. This table shows the research questions, processes, and hypotheses of the groups that ranked in the top 10 out of 182 according to judges. We also collected further qualitative insights on the nature of the experiments;
their measurement and observation; and the students’ perceptions of the importance of this research to their community, among others.
 Table E.3: Median 10 Student Performance in Science Shows

Rank Score Research Hypothesis Research Experiment

101-110th Ranked Science Groups
The pupils put soils in 4 pieces of [a] jerrican, planted beans in beans in the first jerrican, planted maize in the second jerrican, tomatoes
101 23 Not defined
in the third, and Sukuma in the fourth.
The pupils collected plastic bottles and sticks. They then put holes in the bottles and arranged bottles onto the sticks. They then got nails
102 23 Not defined
and used them to make a dustbin out of the bottles.
Covered soil is fertile because it stops
103 23 soil erosion compared to soil that is not The [students] covered one plot of soil with banana leaves and dug some lines in another plot to create terraces.
covered
The pupils got two samples of soil, one was covered with dry grass and the other one terraces was made showing that once water comes
104 23 Not defined
with a high speed it is reduced to control soil erosion.
The pupils collected soil and divided it into 3 plots, the first plot had maize and the others had no plant[s] on them. They then applied
manure around the maize plant on the first plot. They then spread [manure] onto the soil of the second plot and mixed it with the soil and
xxxix

105 22 Not defined

afterwards planted seeds and covered. In the third plot, the students dug holes and added manure into the holes, mixed with the soils and
planted 4 maize grains in each hole.
Covered soil is fertile because soil ero-
106 22 sion doesn’t occur compared to the one Get banana leaves and grass and cover the soil. It will control soil erosion. Draw lines in The soil to make terraces
that is not covered
107 22 Not defined Pupils got soil, planted egg plants and covered the soil with dry plant leaves. They also put soil in a jerrican and planted trees.
Pupils divided the land into 3 plots.
Planted sweet potatoes on the first
108 22
plot, banana on the second plot, maize
on the third plot
The pupils got poles and used nails to make a tiptap stand. After they got a 3 litre jerrican and put holes into it and passed a [goose] wire
109 21 Not defined through the holes on the jerrican and tied it onto the tiptap stand. They tied another goose wire on the handle of the jerrican and tied it
to a pole. They then poured water into the jerrican and covered it.
110 20 Not defined The pupils obtained dirty water into a jerrican, they put it into a glass and let it settle. They then poured the water into another glass.
Notes: Responses were collected from science show judges as described previously in the paper. Science group participants were graded out of 120 points on their ability to conduct scientific experiments and present their findings
with clarity. This table shows the research questions, processes, and hypotheses of the groups that ranked in the between 70th and 79th out of 182 according to judges. We used this range of observations as those that were ranked
lower had difficulty providing research questions and hypotheses and thus wouldn’t have been provided greater clarity and the qualitative nature of science show scores. We also collected further qualitative insights on the nature of
the experiments; their measurement and observation; and the students’ perceptions of the importance of this research to their community, among others.
School:___________________________________Group Number:__________________

Date:____________________ Time: ________________ Name of Judge: _____________________

JUDGE SUMMARIZE THE GROUP’S ACTIVITY:

WHAT ARE THE NAMES OF THE PUPILS WHO PRESENTED FOR THIS GROUP:

WHICH PUPIL SEEMS TO BE THE LEADER OF THE GROUP:

_________________________________________________________________________________

WHAT WAS THEIR RESEARCH QUESTION?

WHAT PROCESS ARE THEY STUDYING?

DESCRIBE THE EXPERIMENT THAT THE PUPILS ARE UNDERTAKING:

WHAT WAS THE GROUP’S HYPOTHESIS:

WHAT DID THE PUPILS OBSERVE:

2
School:___________________________________Group Number:__________________

Date:____________________ Time: ________________ Name of Judge: _____________________

HOW DID THEY MEASURE THESE OBSERVATIONS:

WHAT CONCLUSIONS DID THE PUPILS REACH:

JUDGE TALLIES:

1. Identifying the problem: The pupils clearly stated the problem they are hoping to address
with their project. (Choose a number between 1 and 10)

SCORE:

2. Relevance of the problem: The pupils clearly stated why the problem is an important one to
address.
SCORE:

3. Designing an experiment: The pupils conducted an experiment that clearly outlines how they
would study the problem they identified.
SCORE:

4. Designing an experiment: The experiment was very creative – it tries to learn about the
problem in a way that others have tried before.
SCORE:

5. Designing an experiment: The pupils tested a technology not seen before.

SCORE:

6. Describing a hypothesis: The students had a clearly articulated hypothesis. SCORE:

3
School:___________________________________Group Number:__________________

Date:____________________ Time: ________________ Name of Judge: _____________________

7. Analyzing a hypothesis: The students clearly linked their hypothesis to the design of the
experiment.
SCORE:

8. Making observations: The students captured observations in a systematic manner (for

example, by using a logbook).
SCORE:

9. Making observations: The students made accurate measurements of their observations

using relevant instruments.
SCORE:

10. Sources of information: The students used a sufficient number of information sources to
make their conclusions (for example, they ran the experiment more than once or they
measured more than one outcome).

SCORE:

11. Communicating information: The students communicated their findings independently.

SCORE:

12. Communicating information: The students communicated their findings accurately.

SCORE:

13. Did the pupils demonstrate their activity in a practical way? (YES ------------ NO)

IF YES, DESCRIBE ___________________________________________________

4
School:___________________________________Group Number:__________________

Date:____________________ Time: ________________ Name of Judge: _____________________

Q&A Session

INSTRUCTIONS TO JUDGES: DURING THE Q&A SESSION, ENCOURAGE RESPONSES FROM ALL OF
THE STUDENTS. IN OTHER WORDS, IF ONLY ONE STUDENT RESPONDS TO THE QUESTION, PLEASE
CALL ON OTHER STUDENTS AT RANDOM FROM WITHIN THE GROUP.

1. JUDGE ASK THE STUDENTS: When did you start preparing for this activity? How did you
come up with the question that you studied?

Summarize Response:

INSTRUCTION TO JUDGE: Allow the students to freely respond, and then indicate which of the
following best resembles the student response. [CHOOSE ONE]

A. The learners formulated their own questions or hypothesis to be tested.

B. Teacher suggests topic areas or provides samples to help learners formulate own
questions or hypothesis.
C. Teacher offers learners lists of questions or hypotheses from which to select.
D. Teacher provides learners with specific stated questions or hypotheses to be
investigated.
E. Other ___________________________________________________________________
F. Cannot respond because no clear choice observed.

2. JUDGE ASK THE STUDENTS: Describe how you decided to investigate the question?

Summarize Response:

INSTRUCTION TO JUDGE: The students shared in their presentation how they answered the question.
What we want to know here is how they decided whether this approach to answering the question
would be a good one.

A. Learners developed the procedures and protocols to independently plan and conduct a
full investigation.
B. Teachers encouraged the learners to plan and conduct a full investigation, providing
support along the way.
C. Teacher provided the guidelines for learners to plan and conduct part of an
investigation. Some choices are made by learners.
D. Teacher provides the procedures and protocols for the students to conduct the
investigation.
E. Other__________________________________________________________________
F. Cannot respond because no clear choice observed.

5
School:___________________________________Group Number:__________________

Date:____________________ Time: ________________ Name of Judge: _____________________

3. JUDGE ASK THE STUDENTS: How did you measure outcome XXX?

Judge’s Outcome identified:

Pupil’s Response:

4. JUDGE ASK THE STUDENTS: Provide an alternative explanation for the observations made
by the students. Ask the students how they would respond to this observation.

Judge’s Alternative Explanation:

Pupils’ Response:

INSTRUCTION TO JUDGE: Select one of the following.

A. The pupils were very convincing in their response and responded to this observation
adequately.
B. The pupils provided a response to this observation and it was somewhat convincing.
C. The pupils provided a response to this observation but it was not convincing.
D. The pupils did not respond at all to the alternative explanation.
E. Other_______________________________________________________

Other Questions:

• What is the importance of this activity ( you could add if they are confused about this
question…. to the community)?
• ___________________________________________________________________________
• ___________________________________________________________________________

6
F Categories of Creativity score
Bradler et al. (2019) measure creative performance with the unusual uses task originally
developed as Guilford’s alternative uses task (Guilford, 1967). In the unusual uses task,
subjects are asked to name as many unique and unusual uses as they can in a limited
amount of time for an item common to their subject context. In our setting, we asked
students to identify unusual uses for three different items: a plastic bottle, maize plant
remains, and bamboo stems. The enumerator first demonstrated how the activity works
by providing some creative uses of a plastic bag and then asking the student to share
additional ideas for creative uses of plastic bags. Once the students understand the
nature of the activity, they are told that they have one minute to list as many creative
and uncommon uses of the next item that they possibly can. Once they are told the item,
the enumerator starts a timer on a tablet that sounds an alarm when 60 seconds have
passed. In the interim, the enumerator writes each creative use of an object mentioned
by a student to later record into the tablet.
We evaluate students’ answers using two measures of the unusual uses task: “valid-
ity” and “originality”.60 Each valid use is valued at one point. We then categorize the
answers of students according to their primary functions.61 The originality of responses
is measured by the statistical infrequency of answers according to the categories they
corresponded to. Specifically, we give one additional point to a valid answer if less than
10% of participants gave the same answer and allotted two additional points to a valid
answer as very original if less than 1% of subjects gave that answer.62
Answers that were unclear were counted as invalid. Table F.1 shows the categories
and percent responses for data collected in 2019for the task involving a plastic bottle.
The respective tables for the tasks involving the remains of maize plants and bamboo
stems can be found in the appendix (Tables F.2 and F.3 respectively).

60
We omit “flexibility,” and “elaboration,” measurement items in the original creative uses task (Guil-
ford, 1967). Flexibility reflects the variety of a student’s response and is determined by counting the
number of different categories into which responses fall. Elaboration reflects the level of detail in their
response. Since we utilized enumerators in drawing out responses from students, we feel that these two
measures are more compromised by enumerator effects than “validity” and “originality,” which complicate
the use of these concepts in empirical analysis even with enumerator fixed effects.
61
Unlike Bradler et al. (2019) we do not delineate response categories ex ante. Rather, we anonymize
responses and ask two members of our research team to categorize the universe of responses. In case of
disagreements a PI intervenes and makes the final decision.
62
Bradler et al. (2019) use thresholds of 8% and 1% respectively.

xlv
Category Percent
Category Percent
Storage 40.1 .. ..
Drinking 26.2 . .
Toy 22.7
Irrigation 17.1 Fishing 0.6
Planting 14.4 Light Protector 0.5
Income 11.4 Science Experiments 0.5
Charcoal Lighter 11.1 Shoes 0.4
Recycling 8.6 Life Jacket 0.3
Construction 8.5 Plastic Wire 0.3
Funnel 4.7 Workout Equipment 0.3
Decoration 3.6 Weapon 0.3
Plates 2.3 Control Soil Erosion 0.3
Musical Instrument 1.5 Gutter 0.2
Stove Stand 1.4 Pipes 0.1
Animal Feeders 1.3 Make Dice 0.1
Washing 1.1 Making Bricks 0.1
Measurement 0.9 Tyre 0.1
Scare Crow 0.8 Communication Lines 0.1
Tool Handle 0.8 Roof 0.1
Jerrycan Repair 0.8 Note: This table shows the categories
Rubbish Container 0.7 used for the student creativity scores
Filter 0.7 for the exercise involving plastic bot-
Curtains 0.6 tles. Each mention receives one point
Spray 0.6 and objects mentioned less than 10%
.. .. and 1% frequency receive one and two
. .
extra points respectively.

Table F.1: Categories used for creativity scores for unusual uses of plastic bottles

xlvi
Category Percent
Fire 41.8 Category Percent
Mulching 38.3 .. ..
. .
Fertilizer 30.3
Cooking 29.4 Baskets 0.2
Animal Feed 28.9 Traditional Practices 0.2
Construction 10.2 First Aid 0.2
Toy 6.1 Wind Breaker 0.2
Fencing 3.9 Building Shades 0.2
Roofing 3.8 Creating Salt 0.2
Disciplinary Tool 2.3 Creating Smoke 0.2
Income 1.5 Sugar Cane 0.1
Cover 1.2 Goal Posts 0.1
Scare Crow 1.1 Riding Aid 0.1
Plant Support 1.0 Counters 0.1
Cleaning 0.8 Hat 0.1
Decoration 0.8 Harvest Beans 0.1
Hygiene 0.7 Regerminate 0.1
Making Ashes 0.7 Maize Packaging 0.1
Mats 0.7 Saw Dust 0.1
Medicine 0.7
Note: This table shows the categories
Straw Bed for Animals 0.6
used for the student creativity scores
Charcoal 0.6
for the exercise involving maize resid-
Making Tools 0.5
ual. Each mention receives one point
Furniture 0.3
and objects mentioned less than 10%
Weapon 0.3
.. .. and 1% frequency receive one and two
. . extra points respectively.

Table F.2: Categories used for creativity scores for unusual uses of maize plant residues

xlvii
Category Percent
Construction 60.4 Category Percent
Disciplinary Tool 35.4 .. ..
. .
Fire Wood 20.2
Fencing 11.7 Fertilizer 0.6
Cooking 11.3 Charcoal 0.6
Furniture 6.4 Plumbing 0.6
Income 6.1 Counting 0.6
Music Instrument 4.9 Rain System 0.5
Pole 4.9 Straws 0.5
Animal Feed 4.6 Control Soil Erosion 0.4
Decoration 4.2 Wiring 0.3
Roofing 3.8 Ladder 0.3
Fishing 2.8 Bikes 0.2
Plant Support 2.7 Power Generation 0.1
Crafts 2.5 Charcoal Lighter 0.1
Weapons 2.3 Toothbrush/Floss 0.1
Medicine 2.2 Smoking Pipe 0.1
Dishes 1.9 Floaters 0.1
Boats 1.7 Pounding Machines 0.1
Toy 1.3
Note: This table shows the cate-
Art 1.3
gories used for the student creativity
Praying 0.9
scores for the exercise involving bam-
Mulching 0.9
boo stems. Each mention receives one
Gates 0.8
point and objects mentioned less than
Measurement 0.6
.. .. 10% and 1% frequency receive one and
. . two extra points respectively.

Table F.3: Categories used for creativity scores for unusual uses of bamboo stems

xlviii
G Additional Appendix Tables

Statistics Summary Statistics Balance Tests

All Control Treated
Mean Sd Mean Sd N Mean Sd N β p value

Teacher Survey Outcomes (Student Variables — 2019 Only)

Student Asks Questions 0.24 0.16 0.22 0.16 124 0.25 0.15 119 0.11 0.73
Corporeal Punishment 0.43 0.29 0.42 0.27 162 0.44 0.30 169 0.12 0.54
Knowledge of Student 0.45 0.23 0.43 0.21 247 0.46 0.24 238 0.48 0.12
Specification Statistics F Score (p value) 1.34 (0.27)
Clusters 54
Teacher Survey Outcomes
Teacher Gender 1.39 0.49 1.40 0.49 415 1.42 0.49 422 0.03 0.50
Attended Training Last Year 0.33 0.47 0.36 0.48 415 0.32 0.47 422 -0.05 0.32
School Has Farm Land 0.65 0.48 0.67 0.47 415 0.68 0.47 422 0.01 0.94
Boys Better Pupils Than Girls? 3.33 1.52 3.33 1.52 415 3.37 1.51 422 0.01 0.72
Head Teacher Listens to Me 1.30 0.64 1.35 0.71 382 1.25 0.56 384 -0.06 0.08
How Satisfied with Job? 2.13 0.97 2.16 1.00 415 2.13 0.95 422 -0.01 0.61
Students Connect School to Family Life 1.69 0.84 1.72 0.90 415 1.68 0.78 422 -0.01 0.72
Specification Statistics F Score (p value) 0.96 (0.47)
Clusters 89
Teacher Dyad Outcomes
ij Speak About Classroom Management 0.82 0.38 0.83 0.37 3,108 0.81 0.39 3,908 -0.02 0.60
ij Visit Each Others Classrooms to Learn 0.73 0.44 0.74 0.44 3,108 0.72 0.45 3,908 0.01 0.89
ij Plan Classroom Activities Together 0.62 0.49 0.64 0.48 3,108 0.59 0.49 3,908 -0.05 0.21
Specification Statistics F Score (p value) 0.87 (0.46)
Clusters 88
Classroom Observation Outcomes
Share of Engaged Pupils 4.55 1.97 4.54 1.97 4,806 4.52 2.01 4,833 0.01 0.47
Activity: Q and A 0.19 0.39 0.20 0.40 4,807 0.18 0.39 4,834 -0.04 0.15
Activity: Practice and Drill 0.05 0.22 0.06 0.23 4,807 0.05 0.22 4,834 -0.04 0.34
Activity: Assignment 0.09 0.29 0.10 0.30 4,807 0.08 0.28 4,834 -0.06 0.10
Activity: Copying 0.10 0.30 0.10 0.30 4,807 0.09 0.29 4,834 -0.04 0.24
Teacher Out of Class 0.13 0.34 0.13 0.34 4,807 0.15 0.35 4,834 0.03 0.60
Materials: None 0.42 0.49 0.42 0.49 4,792 0.43 0.50 4,813 -0.03 0.65
Materials: Textbooks 0.04 0.19 0.04 0.18 4,792 0.04 0.20 4,813 0.02 0.77
Materials: Notebooks 0.10 0.30 0.11 0.31 4,792 0.10 0.30 4,813 -0.03 0.68
Materials: Blackboard 0.40 0.49 0.40 0.49 4,792 0.38 0.49 4,813 -0.03 0.63
Specification Statistics F Score (p value) 0.74 (0.69)
Clusters 89

Notes: This table reflects balance tests of teacher, teacher dyad, and classroom outcomes and covariates. We present summary statistics
of each measure, displaying means and standard deviations for the whole sample “All,” the sample of teachers/teacher dyads/classroom
observations in control schools “Control,” and the sample of teachers/teacher dyads/classroom observations in treatment schools “Treated”
(the latter two also include number of observations). At the time of writing, we have baseline data for all schools for variables that do not
use the Student Survey in their construction. For school admin data, including teacher attendance, we were in the process of collecting
baseline data outside of Budondo sub-county in March 2020 when the Covid-19 related lockdowns started — we exclude these variables
form balance tests here. Balance tests reflect an OLS regression with the specification Treatedis = βX + ϵis where i represents student,
s represents school, X represents the vector of covariates in the rows of this table, β is the vector of coefficients associated with each
covariate and ϵis is the error term clustered at the school level. Each specification is run for each data set separately, datasets are
separated by the horizontal lines in the table. The F Score and number of clusters is reported for each specification. The F Score’s p
value (in parentheses) reports results of the null hypothesis test that coefficients are jointly orthogonal within a given specification.

Table G.1: Balance Test for Teacher Outcomes and Covariates

xlix
Statistics Summary Statistics Balance Tests
All Control Treated
Mean Sd Mean Sd N Mean Sd N β p value
Student Survey Outcomes
Passthrough Rate 2018-2019 0.88 0.33 0.87 0.33 283 0.90 0.30 314 0.06 0.34
Passthrough Rate 2017-2018 0.91 0.29 0.91 0.29 282 0.92 0.28 314 0.03 0.68
Creativity Index 7.46 3.91 7.82 4.05 283 7.88 4.22 315 0.00 0.88
Specification Statistics F Score (p value) 0.33 (0.80)
Clusters 54
Student Assessment Outcomes
Recall and Recognition -0.10 2.32 0.46 2.04 280 0.68 2.19 317 0.01 0.65
Apply Understanding -0.04 1.73 0.43 1.42 280 0.50 1.29 317 -0.01 0.83
Critical Thinking -0.04 1.50 0.16 1.48 280 0.29 1.45 317 0.01 0.59
Specification Statistics F Score (p value) 0.22 (0.88)
Clusters 54
Student PLE Outcomes
English 2.98 1.69 2.61 1.55 364 3.26 1.73 486 0.06 0.15
Science 3.45 1.79 3.12 1.80 364 3.69 1.75 486 0.05 0.40
Mathematics 2.92 1.51 2.75 1.50 364 3.04 1.50 486 -0.03 0.16
Social Studies 4.41 1.75 4.20 1.69 364 4.57 1.79 486 -0.06 0.33
Female 0.54 0.50 0.54 0.50 376 0.55 0.50 494 0.00 0.99
Specification Statistics F Score (p value) 1.66 (0.18)
Clusters 26

Notes: This table reflects balance tests of student outcomes and covariates. We present summary statistics of each
measure, displaying means and standard deviations for the whole sample “All,” the sample of students in control schools
“Control,” and the sample of students in treatment schools “Treated” (the latter two also include number of observations).
At the time of writing, we do not have baseline data for all schools. For Student Survey Outcomes and Student Assessment
Outcomes, we did not collect student data until 2019, which was two years into the intervention in Budondo sub-county.
For PLE outcomes, we were in the process of collecting baseline data outside of Budondo sub-county in March 2020 when
the Covid-19 related lockdowns started. Balance tests reflect a regression with the specification Treatedis = βX + ϵis
where i represents student, s represents school, X represents the vector of covariates in the rows of this table, β is
the vector of coefficients associated with each covariate and ϵis is the error term clustered at the school level. Each
specification is run for each data set separately, datasets are separated by the horizontal lines in the table. The F Score
and number of clusters is reported for each specification. The F Score’s p value (in parentheses) reports results of the
null hypothesis test that coefficients are jointly orthogonal within a given specification.

Table G.2: Balance Test for Student Outcomes and Covariates

l
 Table G.3: Summary of All Pre-analysis Specifications

Dependent Variables: Fixed Effects

Estimator Datasets Unit of Obs. Notes
Pair Enum Grade
(H1.1) Classroom Pedagogy Improves
Student Engagement Ologit Stallings Classroom Ordered proportion of engaged
students.
Critical thinking and OLS TIPPS Classroom Factor analysis using classroom
explorationa
observations
Understanding concepts and OLS TIPPS Classroom Factor analysis using classroom
deep learninga observations
Students ask questions Tobit Student Survey Teacher Transform student response
into teacher-level variable.
li

Censored below by 0 and

above by 1.
Corporal punishment Tobit Student Survey Teacher Transform student response
into teacher-level variable.
Censored below by 0 and
above by 1.
(H1.2) Teacher motivation and effort increases
Teacher attendance OLS School Admin Data Teacher
Teacher knowledge of stu- Tobit Teacher and Student Teacher Combine teacher and student
dent Survey responses. Censored below by
0 and above by 1.
(H1.3) Teacher collaboration increases
a
indicates that we were not able to measure this outcome. Continued on next page. . .
 . . . continued from previous page
Dependent Variables: Fixed Effects
Estimator Datasets Unit of Obs. Notes
Pair Enum Grade
Teacher learning and collab- OLS Teacher Survey Teacher Dyad Mean of responses for ij pair.
oration
(H2.1) Traditional learning outcomes increase
English OLS UNEB P7 Student Standardized score
Science OLS UNEB P7 Student Standardized score
Math OLS UNEB P7 Student Standardized score
Social Studies OLS UNEB P7 Student Standardized score
Pass PLE OLS UNEB P7 Student Standardized score
P6 Pass-through Rate OLS Student Survey P6 Student Most recent year
lii

(H2.2) Higher order learning outcomes increase

Understanding OLS Student Assessment P6 Student Standardized score of factor
analysis
Critical Thinking OLS Student Assessment P6 Student Standardized score of factor
analysis
(H2.3) Field-based scientific competency increases
Science Show Result Tobit Science Show Student Average across all scores in
Group rubric; censored below by 1 and
above by 10
(H2.4) Student creativity increases
a
indicates that we were not able to measure this outcome. Continued on next page. . .
 . . . continued from previous page
Dependent Variables: Fixed Effects
Estimator Datasets Unit of Obs. Notes
Pair Enum Grade
Creativity score Tobit Student Survey P6 Student Score measuring number and
uniqueness of student ideas;
censored below by 0

Note: All specifications will measure the Treatment Effect using the Intent to Treat (ITT) coefficient. All standard errors across all
specifications are clustered at the school level. Acronyms/Abbreviations: UNEB = “Ugandan National Examinations Bureau”; TIPPS
= “Teacher Instructional Practices and Processes System”; OLS = “Ordinary Least Squares”; P7 = “Primary Seven”; P6 = “Primary
Six”; Enum = “Enumerator” or “Judge”; Obs. = “Observations”. Dependent variables grouped by hypotheses described in section ??.
liii
 Hypothesis: (H2.1): Traditional Learning Outcomes
Primary Leaving Exam Results 2019 Passthrough 2018-2019
Outcome Variable: English Science Mathematics Social Studies PLE Pass P6 to P7
∗∗∗ ∗∗∗ ∗ ∗∗∗ ∗∗∗
Treatment (ITT ) 0.60 0.53 0.19 0.49 0.27 0.14∗∗∗
(0.09) (0.12) (0.09) (0.10) (0.07) (0.04)
H0 : ITT = 0
p value [0.00]±± [0.00]±± [0.05]± [0.00]±± [0.00]±± [0.00]±±
RI p value [0.00]±± [0.01]±± [0.13] [0.00]±± [0.01]±± [0.02]±±
BH Critical p value (5%) [0.01] [0.03] [0.05] [0.02] [0.03] [0.04]
Pair FE Yes Yes Yes Yes Yes Yes
Enum FE No No No No No Yes
Source of Data UNEB UNEB UNEB UNEB UNEB Student Survey
Unit of Observation Student Student Student Student Student Student
Standardized Variable Yes Yes Yes Yes No No
Range of Outcome Variable [-1.28, 2.44] [-1.56, 2.48] [-1.29, 3.15] [-1.82, 2.35] {0,1} {0,1}
Control School Mean −0.57 −0.51 −0.34 −0.48 0.46 0.78
Clusters 29 29 29 29 29 29
Observations 808 808 808 808 866 294
liv

Estimator OLS OLS OLS OLS OLS OLS

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Randomization Inference using 1,000 permutations
of school-level treatment indicator within matched-pair strata. These results should be treated as preliminary since estimation took place in October
2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Units of observation in all columns are students, though the data sets and, hence, number of observations differ. The first
through fifth columns utilize official test scores for the primary leaving exams (PLE) purchased from Uganda’s National Examination Bureau in
2020, relating results from tests taken in November 2019. Students can receive scores ranging from 1 (best) to 9 (worst) for each exam. We have
transformed this measure to facilitate interpretation such that 9 is the top score and 1 is the worst score (reflected in figures H.2a through H.2d).
Furthermore, all measures in the first four columns have been standardized using the pooled subject-specific mean and standard deviation of control
schools. Therefore, treatment effect measures reflect standard deviation differences between treatment schools and control schools
in columns one through four. We analyze the PLE pass rate and P6 to P7 pass-through rate in columns 5 and 6 respectively. Both are
binary measures equaling 1 if a student passed the exam/grade level and 0 otherwise. The number of observations in column five is greater than that
in columns one through four because students who registered but were not present for the PLE exam are classified as students who must repeat the
PLE exam in the following year. Students are divided into four “divisions” based on PLE scores, division 1 reflects high performance and division 4 is
the lowest. Students receiving division 4 marks and below are considered to have the PLE (typically with an aggregate score above 28). The variable
analyzed in column 6 is constructed using the student survey. For more details, see pages 20-21 of the pre-analysis plan.

Table G.4: Windsorized Data: Test of Hypothesis (H2.1): Traditional Learning Outcomes Increase
 Hypotheses: (H2.2): Higher Order Learning (H2.3): Science Show (H2.4):Creativity
Outcome Variables Apply/Understand Critical Thinking Index (Mean) Index
Treatment (ITT ) 0.80∗∗∗ 0.44∗∗∗ 0.78∗∗∗ 0.79∗∗∗
(0.13) (0.13) (0.29) (0.17)
H0 : ITT = 0
p value [0.00]±± [0.00]±± [0.01]±± [0.00]±±
RI p value [0.00]±± [0.00]±± [0.04]±± [0.00]±±
BH Critical p value (5%) [0.03] [0.05] [0.05] [0.05]
Pair FE Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes
Source of Data Student Assessment Student Assessment Science Show Student Survey
Unit of Observation Student Student Student Group Student
Standardized Variable Yes Yes No No
Range of Outcome Variable [-2.45, 1.40] [-1.84, 1.89] [1, 9.16] [0, 20]
Control School Mean −0.53 −0.27 2.16 5.47
lv

Clusters 29 29 28 29
Observations 295 295 141 295
Estimator OLS OLS Tobit Tobit

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Randomization Inference using 1,000 permutations
of school-level treatment indicator within matched-pair strata. Tobit estimator in the third column treats 1 as the lower bound and 10 as the upper
bound; in the fourth column, 0 is the lower bound with no upper bound. These results should be treated as preliminary since estimation took place
in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: The unit of observation in the first, second and fourth columns is individual students, and student groups in the third column.
Each group was judged by two separate judges, thus there are 79 total groups across 29 schools in the analysis. The first and second columns report
results from the researcher administered student assessments, measuring applied understanding and critical thinking using the procedure outlined on
pages 21 and 22 of the pre-analysis plan. The science show outcome in the third column averages across the 12 outcomes measured by science show
judges and described on page 25 and table 1 of the pre-analysis plan. The creativity score follows the procedure described on pages 26 and 27 of the
pre-analysis plan.

Table G.5: Windsorized Data: Tests of Hypotheses (H2.2) through (H2.4): Higher Order Learning, Science Shows and Creativity
 Hypothesis: (H1.1): Pedagogy (H1.2): Effort (H1.3): Learning
Outcome Variable: Share of Engaged Pupils Student Inquisitiveness Corporal Punishment Knowledge of Student Teacher Network
∗∗ ∗∗∗ ∗∗∗
Treatment (ITT ) 0.38 0.07 −0.00 0.14 0.32∗∗∗
(0.15) (0.02) (0.05) (0.02) (0.10)
H0 : ITT = 0
p value [0.01]±± [0.00]±± [0.94] [0.00]±± [0.00]±±
RI p value [0.00]±± [0.19] [0.99] [0.00]±± [0.08]±
BH Critical p value (5%) [0.02] [0.03] [0.05] [0.05] [0.05]
Pair FE Yes Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes No
Grade FE Yes No No No No
Source of Data Classroom Observations Student Survey Student Survey Stud. + Teach. Survey Teacher Network
Unit of Observation Classroom Snapshots P6 Teachers P6 Teachers P6 Teachers Teacher Dyads
Range of Outcome Variable {1,2,. . . ,6} [0,1] [0,1] [0,1] {0,1,2,3}
Control School Mean 4.23 0.18 0.48 0.57 1.66
Clusters 29 29 29 29 29
Observations 2,140 84 83 83 1,318
lvi

Estimator Ologit Tobit Tobit Tobit OLS

Analysis Notes: Standard errors are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than
0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”) as well as the Benjamini Hochberg (BH) “BH Critical p value” at the
5% level within hypothesis. Hypothesis categories are delineated in the top row of the table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level;
±±
suggests a significant discovery at the 5% level. Randomization Inference using 1,000 permutations of school-level treatment indicator within matched-pair strata. Tobit estimator in the
second through fourth columns treats zero as the lower bound and one as the upper bound. These results should be treated as preliminary since estimation took place in October 2020 and
only includes schools from Budondo sub-county.
Variable Descriptions: Units of observation in the first column are classroom snapshots observed using the Stallings instrument — 10 “snapshots” are taking during each class; therefore, 238
classes (P4 to P6) were observed across 29 schools. “Share of Engaged Pupils” indicates the observer’s assessment of the share of pupils engaged in an activity with a teacher: 1 = No pupil,
2 = One pupil, 3 = A few pupils, 4 = Half of the pupils, 5 = Most of the pupils, and 6 = All pupils (details on page 16 of the pre-analysis plan). Teachers in the second to fourth columns
are restricted to those who at least 90% of P6 students listed as one of their teachers. The second column analyzes the percent of students who indicate that they have asked questions in
the previous two school terms to the specified teacher when they do not understand a topic. The third column analyzes the percent of students who indicate that the specified teacher has
“caned” or beaten them at some point in the previous two school terms (details on page 17 of the pre-analysis plan). The fourth column combines student and teacher responses regarding
student attendance and the students relationship with their guardians for each student. We analyze the treatment effect on the percent of correct teacher responses (details available on pages
18-19 of the pre-analysis plan). Unit of observation in the fifth column is within-school teacher dyads. We sum across binary measures of teacher interactions along three dimensions where
both teachers indicate the existence of a link reflecting teacher collaboration (details available on pages 19-20 of the pre-analysis plan).

Table G.6: Windsorized Data: Tests of Hypotheses (H1.1) through (H1.3): Teacher Pedagogy, Teacher Effort, Teacher Learning
 Source of Data Science Show
Category Framing Experiment Hypothesis Measurement Articulating
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Treatment (ITT ) −0.46 0.46 1.03∗∗ 0.36 0.06 1.11∗∗ 1.41∗∗∗ 0.20∗∗ 1.41∗∗∗ 0.83∗∗∗ 1.19∗∗∗ 1.28∗∗∗
(0.34) (0.34) (0.38) (0.29) (0.13) (0.41) (0.42) (0.08) (0.44) (0.21) (0.31) (0.45)
H0 : ITT = 0
p value [0.19] [0.18] [0.01]±±[0.23] [0.66] [0.01]±±[0.00]±±[0.01]±±[0.00]±±[0.00]±±[0.00]±± [0.01]±±
RI p value [0.42] [0.36] [0.06] [0.29] [0.71] [0.05]±±[0.03]± [0.04]±±[0.02]± [0.02]±±[0.01]±± [0.03]±±
BH Critical p value (5%) [0.05] [0.03] [0.02] [0.03] [0.05] [0.05] [0.03] [0.05] [0.02] [0.03] [0.03] [0.05]
Pair FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Judge FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Standardized Variable No No No No No No No No No No No No
Range of Outcome Variable [1,10] [1,10] [1,10] [1,10] [1,8] [1,10] [1,10] [1,10] [1,10] [1,10] [1,10] [1,10]
Control School Mean
lvii

2.58 2.13 3.31 1.55 1.24 2.15 1.78 1.19 2.31 1.6 2.66 2.15
Clusters 28 29 29 29 29 29 29 29 28 28 28 29
Observations 141 141 141 141 141 141 141 141 141 141 141 141
Estimator OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS

Notes: Standard errors are clustered at the school level. *, reflects a coefficient p value from the original specification, “p value,” less than 0.1, ** less
than 0.05 and *** less than 0.01. Coefficients represent the Intent to Treat effect. We report p values using randomization inference (“RI p value”)
as well as the Benjamini-Hochberg (BH) “BH Critical p value” at the 5% level within hypothesis. ± suggests a significant discovery, accounting for
multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. Hypotheses are grouped according to the categories
outlined in the second row of the table. Randomization Inference using 100 permutations of school-level treatment indicator within matched-pair
strata. Tobit estimator in the third column treats 1 as the lower bound and 10 as the upper bound; in the fourth column, 0 is the lower bound with
no upper bound. These results should be treated as preliminary since estimation took place in October 2020 and only includes schools from Budondo
sub-county.

Table G.7: Windsorized Data: Granular Science Show Outcomes

 Category: Days Teacher Spent on School Farm Days Teacher Spent with Pupils
Outcome Variable: Clearing Planting Weeding Harvesting Clearing Planting Weeding Harvesting
Treatment (ITT ) 1.21 0.90∗ 1.22 0.35 1.93∗ 2.80∗∗∗ 2.52∗∗ 2.14∗∗∗
(1.20) (0.48) (0.85) (0.53) (1.07) (0.80) (1.27) (0.67)
H0 : ITT = 0
p value [0.31] [0.07]± [0.15] [0.51] [0.07] [0.00]±± [0.05]± [0.00]±±
BH Critical p value (5%) [0.04] [0.05] [0.01] [0.03] [0.01] [0.03] [0.04] [0.05]
Pair FE Yes Yes Yes Yes Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes Yes Yes Yes Yes
Source of Data Teacher Survey Teacher Survey Teacher Survey Teacher Survey Teacher Survey Teacher Survey Teacher Survey Teacher Survey
Unit of Observation Teachers Teachers Teachers Teachers Teachers Teachers Teachers Teachers
Range of Outcome Variable [0, 30] [0, 14] [0, 20] [0, 8] [0, 30] [0, 13] [0, 15] [0, 10]
Control School Mean 2.75 1.59 1.93 1.12 2.31 1.04 1.63 0.59
lviii

Clusters 24 24 24 24 24 24 24 24
Observations 168 168 168 168 168 168 168 168
Estimator Tobit Tobit Tobit Tobit Tobit Tobit Tobit Tobit

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗
less than 0.01. Coefficients represent the Intent to Treat effect. We report the Benjamini Hochberg (BH) at the 5% level within hypothesis. Hypothesis categories are delineated in the top row of the
table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. These results should be treated as preliminary
since estimation took place in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables constructed using teacher survey. In the case that a teacher indicated that the school has a school plot, the teacher was then asked “In the the current term (term 3)
or previous term (term 2) how many days total have you worked on this land for: clearing, planting, weeding, and harvesting (or other)”. The enumerator proceeded in step-wise fashion to ask about
the number of days spent for each activity. The response for this exercise is analyzed in the first four columns of this table for each activity, respectively. Then, the same question was repeated but
the teacher was asked “how many days have you worked on this land WITH YOUR PUPILS.” The response for this latter exercise is analyzed in the last four columns of the table. We display this
result graphically in Figure ??.

Table G.8: Teacher Survey: Working on Land at School

Category: Conditional on Not Knowing
Outcome Variable: Question Inquisitive Ignore
Treatment (ITT ) −0.06 0.10∗∗ −0.06∗∗
(0.06) (0.04) (0.03)
H0 : ITT = 0
p value [0.31] [0.01]±± [0.03]±±
BH Critical p value (5%) [0.05] [0.02] [0.03]
Pair FE Yes Yes Yes
Enum FE Yes Yes Yes
Source of Data Teacher Survey Teacher Survey Teacher Survey
Unit of Observation Teachers Teachers Teachers
Range of Outcome Variable {0,1} {0,1} {0,1}
Control School Mean 0.61 0.82 0.08
Clusters 29 29 29
Observations 230 144 144
Estimator OLS OLS OLS

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a
coefficient p value from the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗
less than 0.01. Coefficients represent the Intent to Treat effect. We report the Benjamini Hochberg
(BH) at the 5% level within hypothesis. Hypothesis categories are delineated in the top row of the
table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level;
±±
suggests a significant discovery at the 5% level. These results should be treated as preliminary
since estimation took place in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables constructed using teacher survey. “Question” indicates that a
teacher responded yes (equal to 1) if they were asked a question they did not know the answer to (0
otherwise). Conditional on being asked such a question, the teacher is given an opportunity to state
the most common way they respond. The enumerator codes responses according to whether the
teacher says she 1) “Ignores the student and do not respond,” 2) “Tell him that you do not know and
will do research to find the right answer,” 3) “Suggest to him or her how to investigate the answer
on his or her own,” 4) “Provide the best response to the student you can provide,” or 5) “Other.”
We code the second and third responses as “Inquisitive” and set the dependent variable equal to one
if the teacher provided either one of these as a response and zero otherwise. We code the first and
fourth responses as “Ignore” suggesting that the teacher effectively ignores the students question.
Again, the dependent variable in this column is equal to one if the teacher mentions response number
one or four and zero otherwise.”

Table G.9: Teacher Survey: Teacher Inquisitiveness

lix
 Source of Data: Science Show
Category: Experiment Hypothesis Measurement Articulating Average
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Treatment 1.72 ∗∗∗
0.73 0.10 1.99 ∗∗∗
2.09 ∗∗∗
0.87 ∗∗∗
1.41 ∗∗∗
2.12 ∗∗∗
2.16 ∗∗∗
2.06 ∗∗∗
1.36∗∗∗
(0.50) (0.44) (0.28) (0.65) (0.52) (0.18) (0.34) (0.55) (0.72) (0.63) (0.37)
Teacher Outcome
Inquisitive −1.59∗∗ −1.36∗∗ −0.78∗∗ −2.11∗∗ −1.92∗∗ −1.47∗∗∗ −1.46∗∗ −1.74∗∗ −1.69∗ −1.97∗∗ −1.68∗∗∗
(0.68) (0.60) (0.36) (0.87) (0.75) (0.37) (0.53) (0.77) (0.89) (0.80) (0.52)
Treatment × Inquisitive 1.59∗∗ 1.24∗ 0.86∗∗ 2.29∗∗ 2.20∗∗∗ 1.15∗∗∗ 1.27∗∗ 1.13 1.30 2.09∗∗ 1.64∗∗∗
(0.75) (0.66) (0.39) (0.89) (0.70) (0.39) (0.51) (0.79) (0.90) (0.85) (0.54)
Enum FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Pair FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Source of Data Science Show Science Show Science Show Science Show Science Show Science Show Science Show Science Show Science Show Science Show Science Show
Unit of Observation Student Group Student Group Student Group Student Group Student Group Student Group Student Group Student Group Student Group Student Group Student Group
Standardized Variable No No No No No No No No No No No
Range of Outcome Variable {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,8} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,10} {1,2,. . . ,9}
Control School Mean 3.81 2.18 1.8 2.70 2.4 1.65 2.16 2.92 3.2 2.65 2.63
lx

Clusters 29 29 29 29 29 29 29 29 29 29 29
Observations 158 158 158 158 158 158 158 158 158 158 158
Estimator OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to
Treat effect. Hypothesis categories are delineated in the top row of the table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. These results should be
treated as preliminary since estimation took place in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables in this table reflect scores provided by judges to student groups during the course of the science show. Column (1) reflects the clarity of an experiment; (2) the creativity of an experiment; (3) the uniqueness of an
experiment. Column (4) measures the degree to which students described their hypotheses clearly and (5) measures the degree to which the analysis of their hypothesis is congruent with their experiment. Column (6) measures the degree to which
students kept records of their observations (7) the degree of accurate measurement and (8) sufficient use of resources. Column (9) measures whether students articulated their thoughts independently (without a teacher) and (10) measures whether
they articulated their thoughts accurately (they understood what they were saying). Column (11) averages across all other columns.
Teacher Outcome: We present teacher outcome interacted with treatment. “Further Research” is our measure of teacher inquisitiveness. We ask teachers whether students have ever asked them a question they don’t know the answer to. If they
respond “Yes,” we further ask them how they respond when a student asks such a question. We code teacher responses according to whether they 1) Ignore the student and do not respond; 2) Tell him that you do not know and will do research to
find the right answer; 3) Suggest to him or her how to investigate the answer on his or her own; or 4) Provide the best response to the student you can provide. “Further Research” averages across all teachers in a school who responded using category
2) or 3) in the follow up to the initial question. We interact each student’s response by the within-school teacher average for this variable.

Table G.10: Science Show: Heterogeneity by Teacher Inquisitiveness

 Category: Horizontal Treatment of & Attitudes Towards Students Average
Outcome Variable: Free to Disagree Decide Learn from Pupils Horizontal
Treatment (ITT ) 0.50∗∗ 0.20 0.76∗∗∗ 0.12∗∗∗
(0.23) (0.27) (0.27) (0.04)
H0 : ITT = 0
p value [0.03]± [0.45] [0.01]±± [0.01]±±
BH Critical p value (5%) [0.02] [0.05] [0.03] [0.05]
Pair FE Yes Yes Yes Yes
Enum FE No No Yes Yes
Source of Data Teacher Survey Teacher Survey Teacher Survey Teacher Survey
Unit of Observation Teachers Teachers Teachers Teachers
Range of Outcome Variable {0,1} {0,1} {0,1} [0,1]
Control School Mean 0.68 0.24 0.59 0.50
Clusters 29 29 29 29
Observations 230 230 230 230
lxi

Estimator Logit Logit Logit OLS

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a coefficient p value from
the original specification, “p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to
Treat effect. We report the Benjamini Hochberg (BH) at the 5% level within hypothesis. Hypothesis categories are delineated
in the top row of the table. ± suggests a significant discovery, accounting for multiple hypothesis tests, at the 10% level; ±±
suggests a significant discovery at the 5% level. These results should be treated as preliminary since estimation took place
in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables constructed using teacher survey. All are binary variables equal to one if the teacher indicates
“yes” and zero otherwise. For each column variable, the questions we ask teachers are as follows: Free to Disagree → Are
pupils free to disagree with the concepts you are teaching them? Decide → Are pupils free to decide how to spend time during
a lesson? “Learn from pupils” indicates whether a teacher is more likely to strongly agree with the statement “I learn as
much from my pupils as they learn from me.” Equal to 1 if the teacher chooses “Strongly Agree” and zero otherwise (choices
include: Strongly Disagree, Somewhat Disagree, Neither Agree nor Disagree, and Somewhat Agree). We omit enumerator
fixed effects in the first two columns due to insufficient variation.

Table G.11: Teacher Survey: Classroom Practices Demonstrate More Horizontal Relationships with Students
 Category: Gentle Response Harsh Response Ordinal Difference
Outcome Variable: Ordinal Binary Ordinal Binary Gentle-Harsh
Treatment (ITT ) 0.51∗∗ 0.56∗∗∗ −0.37∗∗ −0.43∗ 0.53∗∗∗
(0.24) (0.20) (0.16) (0.22) (0.18)
H0 : ITT = 0
p value [0.04] [0.01]±± [0.02]±± [0.06]± [0.00]±±
BH Critical p value (5%) [0.01] [0.02] [0.03] [0.04] [0.05]
Pair FE Yes Yes Yes Yes Yes
Enum FE No No No No No
Source of Data Teacher Survey Teacher Survey Teacher Survey Teacher Survey Teacher Survey
Unit of Observation Teachers Teachers Teachers Teachers Teachers
Range of Outcome Variable {0,1,2} {0,1} {0,1,2,3} {0,1} {-3,-2,. . . ,2}
Control School Mean 0.60 0.43 0.46 0.28 0.13
lxii

Clusters 29 29 29 29 29
Observations 230 230 230 230 230
Estimator Ologit Logit Ologit Logit Ologit

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a coefficient p value from the original specification,
“p value,” less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to Treat effect. We report the Benjamini Hochberg
(BH) at the 5% level within hypothesis. Hypothesis categories are delineated in the top row of the table. ± suggests a significant discovery, accounting
for multiple hypothesis tests, at the 10% level; ±± suggests a significant discovery at the 5% level. These results should be treated as preliminary
since estimation took place in October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables constructed using teacher survey. We ask the teacher “When pupils misbehave in the classroom, what are the different
techniques you use to help make sure pupils listen to you and improve their behavior?” The enumerator allows the teacher to respond unmprompted
and then classifies responses according to whether the teacher 1) makes efforts to understand why the poor behavior is taking place, 2) distract the
pupils in a manner that will bring attention back, 3) remind them to pay attention or become serious, 4) separate them from the rest of the class, 5)
hit or beat them, 6) Shout at them, 7) punishment, but not beating, or 8) other. We code responses of 1) or 2) as “gentle” responses and 5, 6, and 7 as
“harsh” responses. Ordinal measures sum sums over the total number of “gentle” or “harsh” responses provided by the teacher. Binary measures equal
1 if only gentle or harsh responses were provided by a teacher and zero otherwise. Differences between gentle and harsh responses reflect differences
in the ordinal measures. We omit enumerator fixed effects from this table due to insuficient variation within enumerator for some outcomes.

Table G.12: Teacher Survey: Teacher Response to Poor Pupil Behavior

 Category: Hunger at School
Outcome Variable: Ate yesterday? I am usually hungry during school I am usually NOT hungry during school Difference
Treatment (ITT ) 0.17∗∗∗ −0.05 0.05 0.10∗∗
(0.05) (0.03) (0.04) (0.05)
H0 : ITT = 0
p value [0.00]±± [0.11] [0.16] [0.04]±±
BH Critical p value (5%) [0.01] [0.03] [0.04] [0.05]
Pair FE Yes Yes Yes Yes
Enum FE Yes Yes Yes Yes
Source of Data Student Survey Student Survey Student Survey Student Survey
Unit of Observation Students Students Students Students
Range of Outcome Variable {0, 1} {0, 1} {0, 1} {-1, 0, 1}
Control School Mean 0.69 0.16 0.07 −0.08
Clusters 29 29 29 29
lxiii

Observations 329 329 329 329

Estimator OLS OLS OLS OLS

Analysis Notes: Standard errors in open parentheses are clustered at the school level. ∗ , reflects a coefficient p value from the original specification, “p value,”
less than 0.1, ∗∗ less than 0.05 and ∗∗∗ less than 0.01. Coefficients represent the Intent to Treat effect. We report the Benjamini Hochberg (BH) at the 5%
level within hypothesis. Hypothesis categories are delineated in the top row of the table. ± suggests a significant discovery, accounting for multiple hypothesis
tests, at the 10% level; ±± suggests a significant discovery at the 5% level. These results should be treated as preliminary since estimation took place in
October 2020 and only includes schools from Budondo sub-county.
Variable Descriptions: Variables constructed using student survey. For the first variable, the student was asked “Did you eat anything at school yesterday (or
Friday if today is Monday)?” A response of yes was coded as 1 and a response of no was coded as 0. For the second and third variables, the student was asked
“How strongly do you agree with the following: I am usually hungry during the school day.” Students who said “Strongly Agree” are coded as 1 in column two
and zero otherwise. Students who said “Strongly Disagree” are coded as 1 in column three and zero otherwise. We take the difference of these two variables
in column four: column three minus column two.

Table G.13: Student Survey: Eating at School

H Additional Appendix Figures

Control Schools: Division 4 = 22.2%

.4
Treated Schools: Division 4 = 17.3%

Kernel Density (Bandwidth = 0.3)

.1 .2 .3

µC=0 µT=.18
0

-2 0 2 4
PLE 2017 Result

Control Treated

(a) 2017

Control Schools: Division 4 = 27.2%

.4

Treated Schools: Division 4 = 12.7%

Kernel Density (Bandwidth = 0.3)
.1 .2 .3

µC=0 µT=.53
0

-2 0 2 4
PLE 2019 Result

Control Treated

(b) 2019
Notes: These graphs present kernel densities of the aggregate PLE distribution across 24
schools only. We were unable to procure student registration numbers for the PLE exam in
2017 from all of the schools in our study. Therefore, this graph shows PLE score distributions
for the 24 schools in which every pair of schools from our pair-wise matching procedure pro-
vided PLE registration numbers for all P7 candidates in both 2017 and 2019. Distributions are
standardized according to the control school mean and standard deviation within year. Larger
values reflect higher performance. The shaded regions reflect the portion of the distribution
that did not pass the division 4 threshold, which is equivalent to failing the PLE. The percent-
ages associated with these shaded areas are printed on the body of the graph. Symbols µC and
µT correspond to the year-specific mean of the control and treatment school PLE outcomes,
respectively. Notice that µC = 0 due to the standardization of PLE scores. Therefore, µT
reflects the difference between control school and treatment school averages according to the
standard deviation of control school PLE score distributions in any given year.

Figure H.1: Distribution of PLE Outcomes Across Treatment and Control Schools

lxiv
 30
Percent of Sample

Percent of Sample
30

20
20

10
10
0

0
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9

Worse Control Treated Better Worse

Control Treated
Better
(a) PLE English (N = 899) (b) PLE Science (N = 899)

25
Percent of Sample

Percent of Sample
30

20
15
20

10
10

5
0

1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Worse Better
Worse Control Treated Better Control Treated

(c) PLE Maths (N = 899) (d) PLE Social Studies (N = 899)

80

80
Percent of Sample

Percent of Sample
60

60
40

40
20

20
0

Repeat Pass Repeat Pass

Control Treated Control Treated

(e) PLE Pass (N = 964) (f) P6 Passthrough 2018-19 (N = 329)

Notes: All variables in this figure have discrete ranges. Variables in sub-figures H.2a through
H.2e are sourced from official results of the 2019 P7 Primary Leaving Examination for schools
in the study sample, procurred from Uganda’s National Bureau of Education (UNEB). We
transform variables to ease interpetation such that larger scores reflect better performance.
The variable represented by sub-figure H.2f reflects P6 students’ responses when asked to
confirm the class they were in during the 2018 school year. If students state that they were in
P6 in 2018 then we code them as “repeat” students in 2019.

Figure H.2: Histograms of Outcome Variables in Table 3 Prior to Standardizing

lxv
 25

25
Percent of Sample

Percent of Sample
20

20
15

15
10
10

5
5

0
0

-2 -1 0 1 2 -2 -1 0 1 2

Worse Better Worse Better

Control Treated Control Treated

(a) Applied Understanding (N = 329) (b) Critical Thinking (N = 329)

22
50
Percent of Sample

Percent of Sample
16.5
40
30

11
20

5.5
10
0

1 2 3 4 5 6 7 8 9 10 0 5 10 15
Worse Better Less More
Control Treated Creative Control Treated Creative

(c) Science Show (Mean) (N = 158) (d) Creativity Index (N = 329)

Notes: Variables represented by sub-figures H.3a through H.3c have, in principle, continuous
ranges. Variables in sub-figures H.3a and H.3b were constructed using principal components
analysis as detailed on pages 21 and 22 of the pre-analysis plan. Thereafter, the variables are
standardized according to the control school mean and standard deviation. Sub-figure H.3c
averages across all measures captured by science show judges and documented in table 1 of
the pre-analysis plan. The variable in sub-figure H.3d has a discrete range and the procedure
we use for constructing it is described on pages 26-27 of the pre-analysis plan.

Figure H.3: Histograms of Outcome Variables in Table 4

lxvi
 Percent of Sample

30
60

Percent of Sample
40

20
20
0

10
l

ls

ls

ls

ils
pi

pi

pi

pi

pi

up
pu

pu

pu

pu

pu

lp
o

ne

he

he
N

Al
O

fe

ft

ft
A

fo

to

0
al

os
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
H

Control Treated Control Treated

(a) Share of Engaged Pupils (N = 2,380) (b) Student Asks Questions (N = 104)
30

40
Percent of Sample

Percent of Sample
30
20

20
10

10
0

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1

Control Treated Control Treated

(c) Corporal Punishment (N = 104) (d) Knowledge of Student (N = 104)

50
Percent of Sample
40
30
20
10
0

0 1 2 3

Control Treated

(e) Teacher Network (N = 1,466)

Notes: Variables in figures H.4a and H.4e have discrete ranges. Variables in remaining sub-
figures are, in principle, continuous and are constructed using the student survey (and the
teacher survey in the case of figure H.4d). Students respond to questions regarding specific
teachers for variables reflected by figures H.4b and H.4c — we then construct a teacher-level
measure for each question by summing over the students who responded in the affirmative for
each teacher and dividing by the total number of students surveyed. The variable reflected
in figure H.4d measures the percent of correct responses provided by P6 teachers’ regarding
different aspects of specific students’ lives. The variable represented by H.4e sums over three
dimensions of teacher interactions reflecting the degree of within-school teacher collaboration
and communication.

Figure H.4: Histograms of Outcome Variables in Table 5

lxvii
 20
40
40
Percent of Sample

Percent of Sample

Percent of Sample
15
30
30

10
20
20

10
10

5
0

0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Worse Better Worse Better Worse Better

(a) 1-Statement of Problem (b) 2-Importance of Problem (c) 3-Clear Experiment

80

50
50
Percent of Sample

Percent of Sample

Percent of Sample
40
40

60

30
30

40

20
20

20

10
10
0

0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Worse Better Worse Better Worse Better

(d) 4-Creative Experiment (e) 5-Unique Technology (f) 6-Clear Hypothesis

60

80

40
Percent of Sample

Percent of Sample

Percent of Sample
60

30
40

40

20
20

20

10
0

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Worse Better Worse Better Worse Better

(g) 7-Linked Hypothesis (h) 8-Systematic Observations (i) 9-Accurate Measurements

50

25

40
Percent of Sample

Percent of Sample

Percent of Sample
40

20

30
30

15

20
20

10

10
10

5
0

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Worse Better Worse Better Worse Better

(j) 10-Information Sources (k) 11-Indep. Communication (l) 12-Accurate Communication

Figure H.5: Histograms of Granular Science Show Outcome Variables in Table 6

lxviii
Notes: We observe classrooms in P7 when no students are engaged in an activity with the
teacher. The observer categorizes the activity the students are engaged in according to whether
they are engaged in learning activities (blue) or they distracted (red).

Figure H.6: Activities Observed in P7 Classrooms When Teacher is Not Around

lxix
 2
Science Show (Mean)
1
0
-1
-2
-3

0 .2 .4 .6 .8 1
Further Research

Control CI 90% Treated CI 90%

(a) Science Show (Mean)

Y Axis Notes: Variable on y axis is a “residual” school-level average science show outcome.
It is “residual” in the sense that the y axis averages the residual term after regressing the
student outcome by school-pair and enumerator fixed effects.
X Axis Notes: The variable on the x axis averages teacher outcomes within school. “Further
Research” is our measure of teacher inquisitiveness within school. We ask teachers whether
students have ever asked them a question they don’t know the answer to. If they respond
“Yes,” we further ask them how they respond when a student asks such a question. We code
teacher responses according to whether they 1) Ignore the student and do not respond; 2)
Tell him that you do not know and will do research to find the right answer; 3) Suggest to
him or her how to investigate the answer on his or her own; or 4) Provide the best response
to the student you can provide. “Further Research” averages across all teachers in a school
who responded using category 2) or 3) in the follow up to the initial question. We interact
each student’s response by the within-school teacher average for this variable, using only
the school-level average as an observation.
Graph Notes: We graph the scatter plot of science show group outcomes against teacher
outcomes, analyzing only one observation per school by taking the average across group
outcomes. We further graph the line of linear best fit within treatment status with a
confidence interval of 90%.

Figure H.7: Science Show Heterogeneity by Teacher Inquisitiveness in Treatment vs. Control Schools

lxx