The Impact of Drought on Educational Attainment: An Empirical Cross-Country Study of Sub-Saharan Africa

Report 2022/5

Mina Skille Mariussen

The Impact of Drought on Educational Attainment: An Empirical Cross-Country Study of Sub-Saharan Africa

Category	Reports
Sub-Categories	EnvironmentStatistics and Empirical AnalysisClimate change and the green transformation
Year	2022
Report Number	5
Author(s)	Mina Skille Mariussen
	https://vista-analyse.no/no/publikasjoner/the-impact-of-drought-on-educational-attainment/
Download	file_download (1.6 MB)
Read in browser	PDF

Report 2022/05	 | 	For 	the 	University of Oslo and Vista Analyse AS

The Impact of Drought on Educational Attainment
An Empirical Cross	-Country Study of Sub	-Saharan Africa

Master of Philosophy in Economics
Mina Skille Mariussen

The Impact of Drought on Educational Attainment

Vista Analyse	 | 2022/05	 	2

D	ocument details
Title	 	The Impact of Drought on Educational Attainment
Report Number	 	2022/05
Author	 	Mina Skille Mariussen 	Mina Skille Mariussen
Date of completion 	 	11.11.2021
Source front page photo 	 	Arc GIS Pro
Keywords	 	Educational attainment, drought, climate change, 	empirical analysis, master
thesis

About us
Vista Analysis is a social science consultancy with its main emphasis on economic research, policy anal-
ysis and advice, and evaluations. We carry out proje	cts to the highest professional standards, with in-
dependence and integrity. Our key thematic areas include climate change, energy, transport, urban
planning and welfare issues.
Our employees have high academic credentials and broad experience within 	consulting. When needed
we utilise an extensive network of companies and resource persons nationally and internationally. The
company is fully employee	-owned.

The Impact of Drought on Educational Attainment

Vista Analyse	 | 2022/05	 	3

Foreword
Climate change will probably intensify drought in areas that already are dry. It is impo	rtant for adapta-
tation and mitigation to assess and understand the impact of drought. Using econometric methods this
thesis estimates the impact of drought on educational attainment in Sub	-Saharan Africa. The thesis 	is
co	-financed by Vista 	Analyse as part 	of our research program in environment and development.

March 10, 2022
Haakon Vennemo
Managing partner
Vista Analyse

The Impact of Drought on Educational Attainment

Vista 	Analyse	 | 2022/05	 	4

The 	Impact of 	D	rought on Educational Attainment
An Empirical Cross	-Country Study of Sub	-Saharan Africa

Mina Skille Mariussen
Master of Philosophy in Economics
Department of Economics
Submitted: 	November 2021

©Mina Skille Mariussen

2021

The Impact of Drought on Educational Attainment: An Empirical Cross	-Country Study of Sub	-
Saharan Africa

Mina Skille Mariussen

http://www.duo.uio.no/

Trykk: Reprosentralen, Universitet	et i Oslo

Abstract

This thesis investigates the effect of drought on education	al attainment	 in ten countries south
of 	Sahara	, by 	using	 geocoded data on climate and households across a time span of 	seventy
years	. Most	 previous studies	 have 	considered	 smaller geographical areas 	and periods 	in
isolation and	 remain 	inconclusive when it comes to the effect of weather shocks on
educational outcomes	 (Randell & Gray, 2016	; Shah, 2017)	.

Combining 	climate data sourced from the Peace Research Institute of Oslo 	with	 micro	-level
household data sourced from the Demographic and Health Survey Program	 was	 made
possible by using 	the software 	ArcGIS	 Pro	.

Results	 show that in the presence of drought, individuals attain around a quarter of a year less
education, 	and the probability of having any education is reduced by roughly 	three	 percent.
This effect is ma	inly driven by an adverse effect for females. Females see	m to be kept from
education in times of drought, while males seem to increase their attendance. This
heterogeneous effect on genders seems to be reversed when comparing subsamples of those
born in the 1950s and 1990s, indicating a shift in gender norms	 wit	h regards to	 education
across time	.

The results	 also 	show	 that having any education increases the probability of being literate by
sixty	 percent, 	highlighting	 the importance of ensuring 	universal	 access to education. Having
experienced a drought reduces	 the probability of literacy by around 	two	 percent. 	There seem
to be no apparent differences in 	educational 	adaptation patterns between rural and urban
residents	, in the presence of a drought.	 An event study of the famine in Ethiopia 	in the 1980s
shows that 	persistent	 drought incidences can increase 	the 	length	 of education	, possibly
through altering long	-term investment patterns	 in education	.

iii

Preface

I would like to thank my 	supervisor Jo Thori Lind for valuable guidance throughout the work
with this thesis. He has provided invaluable	 feedback	 and 	encouragement thorough the whole
process. I also would like to thank Vista Analyse for the opportunity to write the thesis 	in an
inclusive and strongly	 competent 	environment. 	 Particularly I thank 	co	-supervisor 	Haakon
Vennemo for useful discussions and feedback on the thesis.
I also thank my previous colleagues at the Development Co	-Operation Directorate 	of the
OECD for invaluable discussions on relevant trends and topics within the field of
development co	-operation, which ultimately led to the topic of this thesis.
Any errors 	in th	is thesis are solely my own responsibility.

Table of Contents
1 	Introduction	 ................................	................................	................................	..............................	1
2 	Background	 ................................	................................	................................	...............................	3
2.1 Climate change, ag	ricultural dependence, and drought	 ................................	...............................	3
2.2 Educational attainment	 ................................	................................	................................	..............	5
2.3 The 	gender aspect of weather shocks and educational attainment	 ................................	...............	5
3 	Literature review	................................	................................	................................	.......................	7
3.1 Drought and mechanisms	 ................................	................................	................................	..........	7
3.2 Linkages between droughts and educational attainment	 ................................	.............................	8
4 	Data	 ................................	................................	................................	................................	.........	 11
4.1 Demographic and Health Surveys and data management	 ................................	.........................	 11
4.1.1	 	The DHS data	 ................................	................................	................................	............	 11
4.1.2 Data restrictions	................................	................................	................................	................	 12
4.2	 	PRIO	-GRID and SPEI	 ................................	................................	................................	.......	 13
4.2.1	 	PRIO	-GRID	 ................................	................................	................................	...............	 13
4.2.2 SPEI global dataset	 ................................	................................	................................	...........	 14
4.3 Humanitarian Data Exchange	 ................................	................................	................................	.. 16
4.4 Combining the datasets using ArcGIS Pro	 ................................	................................	...............	 16
4.5 Construction of two variables for drought experiences	 ................................	.............................	 17
4.6 Descriptive statistics of constructed dataset	 ................................	................................	.............	 18
4.6.1 Drought experiences	 ................................	................................	................................	.........	 19
4.6.2 Rural and urban areas	 ................................	................................	................................	.......	 21
4.6.3 Educational attainment	 ................................	................................	................................	.....	 22
4.6.4 Literacy	 ................................	................................	................................	............................	 23
5 	Empirical strategies	................................	................................	................................	.................	 26
5.1 Ordinary Least Squares estimation with fixed effects	 ................................	...............................	 26
5.2 Linear probability model on the level of education attained	 ................................	.....................	 29
5.3 Control variables included in the 	analysis	 ................................	................................	................	 30
5.4 Clustered standard errors to resolve spatial correlation in the error terms	 ................................	. 33
5.5 Synthetic control method to evaluate impact of persistent drought	 ................................	...........	 34
6 	Results	 ................................	................................	................................	................................	.....	 35
6.1 Years of education and drought	 ................................	................................	...............................	 35
6.1.1 Effect of drought on years of education attained for the full sample	................................	... 36
6.1.2 Effects of drought	 on subsets of countries	 ................................	................................	.........	 37
6.1.3 Heterogeneous effects on gender	................................	................................	.......................	 40
6.1.4 Heterogeneous effects on place of residence	 ................................	................................	.....	 43
6.1.4 Adaptability over time	 ................................	................................	................................	......	 45

6.2 	Effect of drought on different levels of education	 ................................	................................	....	 48
6.3 Effect of drought on literacy	 ................................	................................	................................	....	 50
6.4 An event study of the Ethiopian famine 1983	-1985	 ................................	................................	.. 54
6.5 Robustness evaluation	 ................................	................................	................................	.............	 57
7 	Weaknesses and possible extensions	 ................................	................................	.......................	 60
8 	Discussion and conclusion	 ................................	................................	................................	.......	 62
9 	Bibliography	 ................................	................................	................................	............................	 65
10	 	Appendix	................................	................................	................................	..............................	 69
10.1	 	Additional tables	................................	................................	................................	................	 69
10.2	 	Further explanations	 ................................	................................	................................	..........	 76

List of figures
Figure 2.1: Land used for agriculture over time in selected countries	 ................................	......	4
Figure 2.2: Gender parity index for primary school attendance over time	 ................................	6
Figure 4.1: The ten countries selected for analysis	 ................................	................................	 12
Figure 4.2: SPEI drought values and 	their possible implications	 ................................	...........	 15
Figure 4.3: Combining ADM1 with DHS clusters using ArcGIS Pro	 ................................	....	 17
 Figure 4.4: Distribution of respondents by country residence and decade born	 .....................	 18
Figure 4.5: Frequency of drought during eligibility to attend school by country	 ....................	 20
Figure 4.6: Percent of individuals who experienced drought while eligible to attend school	 .. 21
Figure 4.7: Mean years of education reported across time	 ................................	.....................	 22
Figure 4.8: Trends in the highest level of education attained over time	 ................................	. 23
Figure 4.9: Evolution of lit	eracy by gender	 ................................	................................	...........	 24
Figure 6.1: Mean years of education in Ethiopia and the synthetic control group	 ..................	 55
Figure 6.2: The gap in the predicted error	 ................................	................................	.............	 56

List of tables
Table 4.1: Gender frequency	 ................................	................................	................................	. 19
Table 4.2: Frequency of extreme and moderate drought during eligibility to attend school	 ....	 19
Table 4.3: Distribution of urban and rural residence among respondents	 ...............................	 21
Table 4.4: Percentage distribution of educational attainment	 ................................	.................	 22
Table 4.5: Percentage of literacy by gender	................................	................................	...........	 24
Table 4.6: Correlation of educational levels and literacy	 ................................	.......................	 25
Table 6.1: The effect of drought on years 	of education attained	 ................................	............	 36
Table 6.2: The effect of drought in dry and humid countries	 ................................	.................	 38
Table 6.3: Heterogeneous effects on gender by sample	 ................................	.........................	 41
Table 6.4: Heterogeneous effects on place of residence	 ................................	........................	 44
Table 6.5: Heterogeneous effects on gender across time	 ................................	.......................	 47
Table 6.6: Percentage of sample according to levels of education	 ................................	.........	 48
Table 6.7: Any education and drought	 ................................	................................	..................	 49
Table 6.8: Literacy and drought	 ................................	................................	............................	 51
Table 6.9: The effect of education on literacy	 ................................	................................	.......	 52
Table 6.10: Comparing restricted and full sample	 ................................	................................	. 58

vii

List of tables and figures in appendix
Table A 1: All gender interaction terms	 ................................	................................	.......................	 69
Table A 2: Effect of drought on separate samples	 ................................	................................	. 70
Table A 3: Heterogeneous effects of drought on any education	 ................................	.............	 70
Table A 4: Probability of having completed primary and secondary	................................	......	 71
Table A 5: Effect of drought on 	literacy when controlling for education	 ...............................	 72
Table A 6: Analysis replicated for Ethiopia	 ................................	................................	..........	 73
Table A7: Clustering by country	 ................................	................................	...........................	 74
Table A 8: Correlation matrix between drought and education	 ................................	..............	 74
Table A 9: Variables used in regression analysis	 ................................	................................	... 75
Table A 10: Comparing the linear model to the orde	red logistic and probit models	 ...............	 77
Figure A 11: Mean years of education over time in Ethiopia, Uganda and Tanzania	 .............	 78
Table A 12: Difference	-in-difference results	 ................................	................................	.........	 79

1 	Introduction

Climate change 	is often referred to as the 	most severe	 challenge 	of our time	. Rising
temperatures 	increase the frequency of	 extreme weather events	 such as floods, 	droughts,
wildfires,	 and 	heatwaves across the globe	 (Masson	-Delmotte, et al., 2021)	.
Agricultural dependent populations in 	sub	-Saharan Africa are especially vulnerable 	when
faced with	 extreme 	climatic events, and their adaptability 	is limited	 by lack of outside
opportunities and	 relatively	 low	 levels of educatio	n (Agamile, Dimova, & Golan, 2021	; The
World Bank, 2021)	.
Education is one of the most important tools in combating poverty and achieving the
Sustainable Development Goals	 ahead of 2030	 (United Nations, 2021)	. As extreme weather
events become more frequent it 	is therefore	 increasingly important to 	better understand the
linkages between 	extreme weather	 events and educational outcomes. 	Several previous studies
have investigated the effect of extreme weather events on educational attainment	, but there
exists	 no consensus 	on t	he magnitude and direction of effect	 (Randell & Gray, 2016	; Shah,
2017)	.
This thesis 	combines	 local climatic data	 on drought 	occurrences	 from the 	Standardized
Precipitation	-Evapotranspiration Index	, with household data across ten countries	, sourced
from the Demographic and Health Surveys	. This is done	 with the goal of contributing to a
better understanding of the effect of 	extreme weather events on educational attainment	. The
data is combined using the software 	ArcGIS Pro	, connecting the	 geocoded data on individual
school attendance 	with	 data on drought occurrence	s from 1946 through 2014. 	The 	main 	study
includes 	244	 000	 individuals, distributed over ten countries located 	south of Sahara. 	This
thesis has a particular focus on	 heterogeneous effect	s on gender in the presence of drought.
Linkages between drought and educational attainment	 are	 investigated 	using	 ordinary least
squares	 estimation	, and a linear 	probability model	, both 	with fixed effects and clustered
standard errors. 	The method	s employ	 fixe	d effects both on small geographical areas, and
across	 time,	 with the aim of 	investigating whether a causal relationship between drought and
educational attainment exists	, and how this has developed across time	. Clustered standard
errors are 	employed to obtain heteroscedasticity robust standard errors.

First, the effect of drought on years of education attained is investigated, 	both on the 	full
sample and on two	 separate	 regions 	based on climatic conditions. 	Results show that having
experienced moderate drought 	while in 	school 	yields 	a reduction in years of attended
education by around a quarter of a year. 	This effect is mainly driven by an adverse effect for
the females in the 	sample and	 can indicate that	 the	 parent	s´ valuation of	 expected future
payoff from education is differentiate	d between the genders	. The effect is however reversed
when studying 	in isolation 	the 	individuals born in the 1990s, indicating a shift in gender
norms	 with regards to	 education	.

Second	ly, the effect of drought on the level of attained education 	is investigated	, finding that
individuals who 	have 	experience	d drought are 	three	 percent less likely to ever have attended
education. 	Third	ly, the effect	s of 	both 	education 	and drought 	on literac	y, 	are	 studied
separately	, showing that education is important to increase literacy, and that individuals are
two	 percent less likely to be literate having experienced a drought.

Lastly, the effect of persistent drought is 	briefly considered,	 through an event study of the
Ethiopian famine in the 1980s. This is done 	using synthetic control methodology	 to construct
an artificial control group	. Th	e latter analysis 	indicates that the l	ong	-term effects on
educational attainment from persistent and severe shocks to 	household	 income can increase
investments in education.

The s	tatistical software S	tata	 is used in 	generation of 	all 	estimations	 and calculations t	hat
follow.

Th	e thesis is structured in the following way	: section 2 presents a background on the 	topics of
the thesis	 and section 3 presents a brief overview of the current literature on the topic. 	Section
4 describes the data used to investigate the topic,	 section 5 des	cribes the methods that 	are
applied,	 and section 6 presents the findings. 	Lastly, section 7 	discusses possible weaknesses
and extensions, before section 	8 presents a brief conclusion.	 The references are found in
section 9 and an appendix is included 	in section 10.

2 	B	ackground

This chapter will p	rovide further insight into why the topic of 	drought 	occurrences	 and
educational 	attainment 	is an interesting and relevant field of 	study.	 The chapter	 starts with
background on climate change, agricultural dependence, and drought prevalence, before it
briefly 	elaborates on the 	importance of education and determinants of educational 	attainment
in developing countries. 	The chapter ends b	y introducing 	the gender aspect of educational
attainment 	and weather shocks.
2.1 	Climate change	, agricultural 	dependence	, and drought

Rising global	 surface t	emperatures 	are 	caus	ing	 extreme weather events	, such as droughts,	 to
become more frequent	 (Masson	-Delmotte, et al., 2021)	. In this thesis the focus will lie on the
possible impacts on educational 	attainment 	through 	weather induced 	shifts in 	household
resources, possibly through changes in 	agricultural production	 and 	food prices	. Droughts are
one 	of the 	extreme weather 	types	 which ha	ve the most unambiguous 	implication for
agricultural production, and 	for which 	there 	exist	 available measures across space and time
(Begue	ría, Latorre, Reig, & Vicente	-Serrano, 2021)	. This thesis will therefore focus on 	the
effect of droughts on educational outcomes.
Climate change	 is likely to yield both more frequent and severe droughts towards the end of
the 21	st century 	(Haile, et al., 2020)	.1 The study by 	Haile et	 al. find that drought	s are	 likely to
increase in 	frequency in 	Somalia, South 	Sudan,	 and Tanzania, while 	a decrease in drought
frequency	 is expected to occur in Uganda, Kenya and in the highlands of 	Ethiopia. 	The
different 	expected 	patterns of drought prevalence across regions 	in coming decades 	makes
understanding its 	possible 	implications 	critical	 to increase adaptability and achieve the 	SDGs
ahead of 2030	 (United Nations, 2021)	.
One 	of the 	way	s droughts can affect 	educational 	attainment 	is through 	changes in agricultural
prod	uction	, spurring adaptations in the 	choice of education	. In sub	-Saharan Africa the percent
employed in agriculture was measured to 53 	percent	 in 2019, making it the largest industry in
the region 	(International Labour Organization, ILOSTAT database, 2021)	. Given the high
 	1 Climate change refers to long	-term changes in weather and temperatures. Climate change induced by humans commenced as the industrial revolution yielded an
increasing need for energy, which largely has been acquired by burning fossil fuels, generating green	house gas emissions. At an aggregate level these greenhouse
gas emissions trap heat from the sun, causing a long	-term shift towards higher temperatures 	(United Nations, 2021)	.

proportion of	 the working population employed in this industry 	the region is	 vulnerable 	to
climatic changes affecting the industry.
Figure 2.1 shows the development in land used for agriculture in the ten selected countries
from 1960 until 2010. 	The total average 	of land used for agriculture 	in these countries ha	s
risen with about ten percent	age points	 during these decades	, indicating 	high dependence on
agricultural production	.
Figure 	2.1: Land used for agriculture	 over time in selected countries

Note: 	Development	 in land used for agriculture in the ten selected countries from 1960 until 2010	. The
figure is developed using Stata.	 Figure s	ource: 	(UNESCO Institute for Statistics, 2021)	.
Studies have	 observed	 that v	ariations in rainfall and temperatures 	have 	historically	 impacted
agricultural production 	more 	adversely	 in 	sub	-Saharan Africa	 since the 1950s	, relative to
other regions	 of the world	 (Barrios, 	Bazoumana, & Strobl, 2008)	. Th	e study 	by Barrios	,
Bazoumana and Eric	 find that 	even small climatic changes 	is 	impacting agricultur	al
production negatively	 in this region due to a lack of 	endowments and technology to
adequately adapt. 	This thesis will therefore focus 	primarily	 on the effect of smaller drought
incidences	 on educational attainment	.

2.2 	Educational attainment

Sustainable development goal number 4 	states	 “Ensure inclusive and equitable quality
education and promote lifelong learning opportunities for all”	 (United Nations, 2021)	. Within
this lies the 	goal	 of universal access to education, ensur	ing	 opportunities for all individuals.
Reaching this 	goal	 would	 yield value for individuals, but also on the aggregate level a country
can 	benefit 	by	 reaching 	higher long	-term economic growth rates if higher educational	 levels
are 	attained	 (Crespo Cuaresma & Lutz, 2007)	. Researchers have found	 that education is
necessary for long	-term economic growth	, by increasing	 human capital development, 	and that
the highest level of growth is obtained by ensuring 	universal 	access to primary and secondary
education 	(Crespo Cuaresma & Lutz, 2007)	.
There are 	several 	factors influencing	 educational 	attainment	 in developing countries	. One
important factor is t	he dependency and employment in	 agricultural production	. In economies
where agricultural production is the main source of income	, the 	true 	cost of education is non	-
zero	 despite 	school enrolment 	being	 free	 of charge	 (Glory & Nsikak	-Abasi, 2013)	. This is due
to the 	high 	alternative 	value children may produce partaking in agricultural production. 	Glory
and Nsikak	-Abasi find poverty 	and family size 	to be the most important factor	s influencing
child labor rates in agriculture	.
Another factor	 influenc	ing	 educational enrollment is the di	stance to the nearest school, which
is largely predicted by rural and urban 	residency	 (Alesina, Hohmann, Michalopoulos, &
Papaioannou, 2021)	. Further, cultural norms towards education plays an important role. The
latter 	is especially important when evaluating gender aspects of educational 	attainment	. The	se
factors will all be 	considered as determinants of educational 	attainment 	throughout this thesis.

2.3 The gender aspect of weather shocks and educational 	attainment

In year 2000 world leaders gathered around the Millennium Development Goals (MDS). The
audacious goals set out to achieve universal primary education, promote gen	der equality and
empower women, among other goals	 (United Nations, 2021)	. With the increased focus on
equal access to opportunities the world has seen a rise in education for 	females	. For 	sub	-
Saharan Africa the percentage of female pupils in primary education has risen from 45
percent in year 2000 to 48 percent in 2019 	(UNESCO Institute for Statistics, 2021)	. This is in
line with the overall trend for the globe.

Figure 	2.2: Gender parity index for primary school attendance over time

Note: 	The gender parity inde	x (GPI)	 of primary school attendance in the ten selected countries	 from
1970 to 2020	. The figure is developed using Stata.	 Figure s	ource	: (UNESCO Institute for 	Statistics,
2021)
From figure 	2.2 we observe that i	n 19	70	 the GPI indicated that for every tenth boy 	enrolled 	in
school there w	ere	 only 	4.5	 girls	 in 	Ethiopia and 	5.5 girls in	 Niger. The ratio has however
tended towards one for most of the 	studied	 countries during the last decade	s (UNESCO
Institute for Statistics, 2021)	.
Access to education	 and paid employment outside the household are on average better in
urban 	areas 	than rural ones. This implies a 	relaxation of social values and norms 	that have
historically	 been	 imposed on females	, enabling them to easier achieve greater independence
(Tacoli & Satterthwaite, 2013)	. Therefore, a	dvocating for 	females´	 school attendance has 	the
possibility of inducing 	great long	-term effects. Education helps limit poverty	 and	 lower
fertility rates,	 it increases labor force participation and empowerment, which in turn increases
the likelihood that future children are sent to school	 (Murray, Paola, Demeranville, & Hurst,
2010	; Tacoli & Satterthwaite, 2013)	. Given the importance of universal access 	to 	edu	cation
and t	he 	changing patterns	 of drought	 occurrences it	 seems 	increasingly 	important to
understand the mechanisms 	it might induce with regards to gender disparity in education.

3 	Literature	 review

This chapter will provide an overview of 	some of 	the available literature on 	relevant
economic mechanisms induced by 	drought	s and the	 linkages to educational 	attainment.
3.1 	Drought	 and 	mechanisms

Extreme climatic events such as floods and droughts 	usually 	affect 	crops 	and thereby
livelihoods 	across large regions	 simultaneously	. Consequences of such	 events are 	diverse	,
ranging 	from increased poverty	 and	 forced 	migration, to changing 	employment	 and
educational	 patterns 	(Agamile, 	et al.	, 2021	; Barrios, 	et al.	, 2008	; Bjorkman	-Nyqvist, 2013	).
A recent	ly published report by several multilateral 	organizations	 showed that 60 percent of
Africa	’s population currently suffer from moderate to severe food insecurity, and 	that 	this
share has been rising since 2014	 (FAO, IFAD, UNICEF, WFP and WHO, 2021)	. Climatic
changes, such as 	the	 increase in drought	 frequency	, affecting agricultural productivity	, and
thereby increasing food prices	 is identified as 	one	 of the largest drivers of this development in
recent years	.
The adverse effects of climate change 	in the less developed regions are often s	tudied	 through
the 	lens	 of crop failures	, leading to 	loss of livelihoods and 	higher food prices	, reducing real
income	 (Hertel & Rosch, 2010)	. This effect is particularly strong for the very poor, who
already live on marginalized plots of land, wit	h li	mited possibilities of	 adapting to new
conditions	 (Randell & Gray, 2016)	. However, 	higher food prices can lead to an increase in
income for producers in regions that are	 not as	 adversely	 affected, generating heterogeneous
effects between those that are directly affected and those that are not	 (Hertel & Rosch, 2010)	.
The effects 	of weather induced shifts in agricultural productivity 	can also translate into 	shifts
in labor markets	 (Agamile	 et al.	, 2021)	.
A study from Uganda 	find that women are allocated more productive land following drought	s
(Agamile	 et al	., 2021).	 Authors argue that the pattern is likely driven by a difference in
outside options for the ge	nders. M	ales have	 on average 	higher education 	and access to 	more
resources 	than females and 	are 	therefore 	more likely to be more mobile and able to transition
to non	-agricultural employment, leaving women with more productive plots of land after the
drought has subsided.

Another possible effect of extreme weather events is that there is a short	-term 	increase 	in the
need for unskilled labor, 	such as child 	labor	, for instance in order to replant fields and salvage
existing crops 	(Hertel & Rosch, 2010)	. Extreme weather events can also 	reduce the need for
unskilled labor, through the loss of crops to cultivate. 	Therefore, the short	-term effects 	on
demand for 	unskilled 	labor 	depend largely on the nature of the weather events and available
adaptation mechanisms. 	A study from Zambia in 2012	 show	 however	 that the impact of
climatic variability on crops alone 	can 	reduce the growth rate of GDP by 0	.4 percentage
points yearly	 (Thurlow, Zhu, & Xinshen, 2012)	. T	hurlow et al.	 demonstrate that the impacts
on the growth rate of GDP are even larger in years with extreme drought or other extreme
weather events.
Long	-term effects of such events	 on 	labor markets	 are 	also	 ambiguous, but some 	observations
point in the direction of 	less use of child labor, and more 	investment in education, as 	work
outside the agricultural sector 	becomes more attractive 	when the vulnerability in the sector
becomes more apparent	 (Opiyo, Wasonga, Nyangito, & Schilling, 2015)	. In 	conclusion,	 the
nature of the 	mechanisms 	induced by 	extreme weather events	 depend on the length and
severity of the	 episodes	 and can therefore have 	diverse effects on the economy.

3.2 	Linkages between droughts and educational 	attainment

Extreme weather events	 ha	ve	 direct negative effects on development, leaving countries to
rebuild rather than 	improve 	on 	existing structures. If climate change also 	impairs	 educational
attainment 	the 	adverse 	effect 	on development might be amplified over time. 	Despite this
research on linkages between weather shocks and educational 	attainment 	in developing
countries is limited and to a large extent diverging	 in conclusions	 (Randell & Gray, 2016)	.
When considering linkages between drought frequency	 and educational attainment the notion
of human capital investments and 	alternative value of 	time	 becomes important 	(Shah, 2017)	.
A study of nearly two million children in rural India find 	evidence that children are more
likely to attend school during years of drought	 (Shah, 2017)	. The study asserts that a viable
alternative to	 education	 is in many cases a	gricultural work or work in the home.
Given that extreme weather events yield	 less 	crops to cultivate 	and less need for child labor,
the value of time spent in agricultural production	 is reduced. This can	 induce more education
for children	. This finding is corroborated by 	a study performed using data from rural

Zimbabwe	 (Nordstrom & Cotton, 2020)	. In addition to exposing the above mechani	sm	, the
study	 from Zimbabwe	 also	 find 	that	 the	 learning output 	of children 	is 	reduced	 in the presence
of drought,	 regardless of more time spent in school	. This	 sugges	ts that the 	adverse effect of
excess stress put 	on families 	during droughts 	is offsetting the effect of more time spent in
school on learning outcomes.
Contrary to findings in Zimbabwe and India a study of 	educational choices in	 rural	 Ethiopia
shows	 that rainfall above mean long	-term levels results in 	an increased likelihood of 	more
education	 (Randell & Gray, 2016)	. The	 study find that higher levels of rainfall is associated
with higher crop productivity, 	and higher household food consumption, 	which 	the authors
argue reduces the need for children participating in income generating activities. 	It seems that
less droughts, and 	thereby 	stress o	n family resources increases the 	likelihood of attaining
more education for rural households in Ethiopia.
The	 literature is also 	concerned with understanding 	how effects of weather shocks yield
heterogeneous effects on 	the 	gender	s. A study from Uganda 	finds	 that 	drier	 seasons lower the
school enrolment of older	 female	s (Bjorkman	-Nyqvist, 2013)	. This study 	finds	 no significant
effect on	 school enrollment of	 younger 	females	, nor on 	males	 of any age. 	The study	 implies
that when stress is asserted on the family, the value of home producti	on for older 	females
increases, while 	males	 and young 	females	 are kept in school due to the l	ower	 need for their
labor in agricultur	al production	.
A study of Ethiopia concludes that	 more humid	 seasons	 reduce schooling for all pupils, 	and
that the effect is more negative 	for 	males	 than it is for 	females	 (Mani, Hoddinott, & Strauss,
2013)	. The two studies 	from Uganda and Ethiopia thus conclude similarly in terms of 	the
relative change in schooling between the genders in response to a drier environment. 	Th	e
studies concludes 	that 	more 	rainfall increases crops, and this is when the value of male
production in	creases relative to the value of traditional female production. 	Based on the two
studies of 	Mani 	et al. 	and Bjorkman	-Nyq	vist 	more humid	 seasons	 reduces 	male schooling,
while 	drier	 seasons	 reduces female schooling	, and the explanation of 	such 	effects lie in a
change in relative value of their labor	 in home or agricultural production	 compared 	to 	the
value of remaining in school.
The difference in results 	regarding the overall effect of drought on educational outcomes
could be due to the different approaches adopted by 	researchers but	 is also likely to be a
product of different samples. 	Zimbabwe, Ethiopia, Uganda and India are different countries,

with diff	erent a	daptation	 mechanism and 	differing	 cultural aspects. 	Therefore, studying each
country in isolation	, as is primarily done in the existing literature	, is likely to yield a
difference of results.
Furthermore, it is difficult in 	such	 studies 	to know if all possible	 underlying	 effects
influencing educational attainment 	are controlled for	. The problem of omitted variable bias,
endogeneity	 and outliers often limit the validity of results.	 Therefore, st	udying the effects
across several countries	, with many individuals 	across time, and with climate data 	for small
areas	, as 	done in this thesis,	 is a	n interesting approach to 	contribute to	- and 	expand on the
existing literature within 	this 	research	 field.

4 	Data

This chapter will provide a description of the 	data sources	 used in this th	esi	s, some of	 the
processing of the data, and descriptive statistics of the data	set that is constructed	. The main
data sources	 used are the Demographic and Health Surveys	 (DHS)	 from ten countries in 	sub	-
Saharan Africa, and climatic data sourced from 	SPEI, made availabl	e through the PRIO	-
GRID structure. 	Data on first administrative borders is sourced from Humanitarian Data
Exchange.
4.1 D	emographic and Health Survey	s and data management

4.1.1	 	The DHS	 data

Data on individual educational attainment and other 	individual 	characteristics	 was sourced
from 	the Demographic and Health Survey	 Program	. The DHS 	conduct	 surveys 	by 	count	ry
usually at 	five	-year	 intervals.	 The sample size of the 	surveys range 	between five and thirty
thousand individuals during each	 survey	 round. 	The DHS 	records households into clusters
based on their location. A small village will 	usually	 be 	recorded as 	one cluster, while in a
larger city several clu	sters will be 	recorded. The DHS clusters are given GPS coordinates	,
yielding	 an accuracy of 	about 	fifteen	 meters	 of clusters	 (Burgert, Zachary, & Colston, 2013)	.
The 	primary purpose of 	the 	DHS is 	to record	 health outcomes, but individuals 	are also 	asked
about living conditions, 	education	, employment,	 and 	status within the household, among other
characteristics	 (The DHS Program, 2021)	.
For use in this thesis a	ccess	 to surveys from 	ten countries w	as applied for	 based on locations
of countries 	and 	the 	availab	ility of	 surveys	 with geocod	ed	 clusters	. The countries 	selected for
analysis 	are	 Burkina Faso, 	the Democratic Republic of	 Congo, Ethiopia, Kenya, Malawi,
Mali	, Niger	, Nigeria	, Tanzania, and 	Uganda	. Surveys 	in the selected countries 	are conducted
in 	the 	timespan from 1990 until 2018.

Figure 	4.1: The ten 	countries 	selected 	for analysis

Note: 	The ten 	selected countries	 are 	represented	 in different color	s, and with first administrative
levels	 outlined within each country. The figure is 	developed 	using the software	 ArcGIS	 Pro	.
By merging 	the cross	-sectional 	data	sets from the DHS	 across both countries and time a	 large
dataset is constructed	 containing 	around 750	 000 individual observations. 	Individuals in the
dataset 	are	 interviewed 	once	, they are born in the year span from 1932 to 	2003 	and	 are
eligible to attend school 	between 1937 and 2013	. This span in time means we can study
droughts and educational 	attainment 	for around 	seventy	 years.

4.1.	2 Data restrictions

Due to the structure of the data several restrictions had to be made	. Firstly, 	as 	this thesis
considers linkages between educational outcomes and experiences of drought during
childhood, 	the sample had to be limited to individuals who reported having 	always 	lived 	in
the same place	. This is done as the respondents 	are recorde	d in the geocoded clusters 	at the
time of the survey	, and do not give location of their childhood residency	. The 	possible
restrictions this impose 	on validity of results 	are 	investigated	 in section 	6.5.
Secondly, individuals who had missing 	values 	or 	non	-logical values 	for	 central variables, such
as educational attainment, were excluded.	 Values which 	are not likely to be correct, such as

an 	individual having attended school for 	forty years	, while 	being thirty	-five years of age	 are
considered as non	-logical values	.
Thirdly, as the PRIO	-GRID only 	is operationalized until year 2014 individuals who were
eligible to 	end schooling past this year were excluded	 from the sample	. These restrictions
limit t	he sample size to around 2	44	 000 individuals across the ten countries.
After	 imposing restrictions	, two of the selected countries only have one survey to draw from.
As individuals of different ages are asked to partake in each survey round	, it is possible to
study 	developments across time despite only having one survey	 for these two countries	. The
remaining eight countries have from two to four available surveys	 from which data is sourced.

4.2	 	PRIO	-GRID 	and SPEI

The climatic data	 from SPEI 	and 	percentage of land used for agriculture	, used in this thesis	 is
sourced through	 the PRIO	-GRID 	map 	structure	.

4.2.1	 	PRIO	-GRID

Data on drought	 frequency	 was 	accessed	 using a	 spatial	 grid structure developed by 	the Peace
Research Institute Oslo 	(PRIO)	.2 The grid structure developed by 	PRIO 	consists of 	a 0	.5x0.5	-
degree 	grid 	resolution map	 covering the globe	 (Tollefsen, Strand, & Buhaug, 2012)	.3
PRIO	-GRID	 draws on multiple 	data 	sources to generate various variables describing drought
and other extreme weather events within each grid cell. As the PRIO	-GRID draws on several
data sources	, the 	availability of 	variables 	varies	. Some variables are 	given as 	snapshots 	in
given years, such as land within the grid cell that is used for agriculture, 	which is 	given every
tenth year, while other variables, such as drought measures are given every year, for every
grid cell	 (Tollefsen, Bahgat, Nordkvelle, 	& Buhaug, 2015)	. PRIO draws on three different
sources of drought measurements; the SPI index, the S	PEI	 index	 and the S	PEI Global
Drought Monitor	, allowing users to identify the measures 	that 	is best suited for their analysis	.
 	2 The purpose of PRIO	-GRID is to facilitate easie	r use of spatial and comparable data worldwide, by drawing on multiple different spatial data sources. The 	structure compiles data from a range of other spatial data sources and enables the usage of those within the given grid struc	ture	 (Tollefsen, Strand, & Buhaug, 	PRIO	-GRID: A unified spatial data structure, 2012)	. 	3 This implies that the cell area covered by 	grid	s around equator has an area of roughly 55 x 55 kilometers.

The PRIO	-GRID cell structure and data pertaining to the grids are 	freely	 available and
downloadable from their online portal	 (Peace Research Institute Oslo, 2021)	. For use in this
thesis variables 	describing	 droughts	 and 	land used for agriculture within g	rid	 cells 	was
downloaded from the online portal and combined with 	micro	-level	 household 	data 	from DHS
using ArcGIS Pro.

4.2.2 SPEI	 global 	dataset

The Standardized 	Precipitation	-Evapotranspiration Index	 (SPEI)	 is one of 	several	 indexes
measur	ing	 droughts across the globe	 (Tollefsen, 	et al.	, 2012)	.4 The SPEI contains monthly
values 	for drought measurements 	across the globe on a 0	.5x0	.5-degree 	resolution	. The values
of the SPEI are given as deviations from normal, long	-term conditions in each locality, which
makes it is well suited to compare 	adaptations to 	droughts across regions	 (Begueria, Vicente	-
Serrano, & Angulo	-Marti	nez, 2010)	.
This 	implies	 that the same value of the index would present the same deviation from the mean
in two different 	localities	, reflecting relative conditions rather than absolute ones	 (Begueria, 	et
al.	, 2010)	. For use in this thesis, it is relevant that the SPEI reports deviation from mean
conditions, as 	what is studied 	are	 mechanisms of adaptation to shocks. Therefore,	 other
index	es, reporting absolute measures 	are inferior to the SPEI for the given purpose	.5
The values 	of the SPEI typically range between 2,5 and 	-2,5, although 	the index 	in theory
could reach infinity in either direction. 	The 	drought 	variable captures the severity of drought
within the rain season	 of the given grid cell.	6

4 The SPEI is made available to the public through a global dataset named the SPEIbase	 (Begueria, Vicente	-Serrano, & Angulo	-Martinez, 2010)	. The 	SPEI	 uses 	monthly climat	ic water balance as measurement, effectively measuring the difference between monthly rainfall and monthly potential evapotra	nspiration.	 	5 Note that	 the similar SPI differ from the SPEI by not including temperatures in its measures. This makes the SPEI bet	ter suited to capture drought incidences as 	consequence of climate changes, considering temperatures are changing in response to climate change	 (Begueria, Vicente	-Serrano, & Angulo	-Martinez, 2010)	 	(Masson	-Delmotte, et al., 2021)	 . 	6 The rain season is defined as the three consecutive months where it on average rained the most during a year within a grid ce	ll. Data on rain seasons within 	each cell used in the calculation 	of the SPEIbase is obtained from Schneider et.al 	(Schneider, et al., 2015)	.

Figure 	4.2: SPEI drought values	 and their possible implications
Dryness	 	Possible Impacts	 	Standardized
Precipitation	-
Evapotranspiration
Index	 (SPEI)
Moderate
drought
dummy
Extreme
drought
dummy
Moderate
Drought
Some damage to crops	 	-0.8 to 	-1.2	 	1

0
Severe
Drought
Crop or pasture losses likely
Water shortages common
-1.3 to 	-1.5	 	1

0
Extreme
Drought
Major crop losses
Widespread water shortages or
restrictions
-1.6 to 	-1.9	 	1 	1
Exceptional
Drought
Exceptional and widespread crop
losses
Shortages of water creating water
emergencies
-2.0 or less	 	1 	1
Note: 	The values of the SPEI that captures dry conditions 	and their possible impacts	 on crops and
livelihoods	, along with 	a representation of the two drought dummy variables. 	Figure	 source:
(National Drought Mitigation Center, 2021)
In th	is the	sis two threshold	 values 	of drought prevalence	 are	 used	 to study 	educational
adaptations to 	droughts	. One set of dumm	y variables	 capturing the existence of 	at least a
moderate drought (	yielding 	1 for 	SPEI 	values below 	-0,8) within each 	grid	 cell 	in each year	 is
generated	, while a second set of dumm	y variables	 capture the existence of	 at least an	 extreme
drought (	yielding 	1 for 	SPEI 	values below 	-1,6)	 is also generated	.7
This construction implies that 	if 	the 	dummy variable for m	oderate drought 	is one, while the
extreme drought dummy is zero, the drought 	experienced 	ranges 	from moderate to severe,
while if both dummy variables are one 	the drought ranges from moderate to 	exceptional
drought	. Consequently,	 the 	extreme drought dummy variable cannot be one	 unless the
moderate drought dummy 	variable 	is also one	.
A dataset	 with t	he two drought dummy variables is merged with the DHS clusters, where the
common variable is the grid cell of each observation.	8 Th	e process	 of combining the two
geocoded datasets is performed	 using ArcGIS 	and	 is described in more detail below.
 	7 The two threshold values are generated using the variable from the SPEIbase named “	droughtend_speibase	” in the PRIO	-GRID dataset 	(Tollefsen, Bahgat, 	Nordkvelle, & Buhaug, 2015)	 (Begueria, Vicente	-Serrano, & Angulo	-Martinez, 2010)	. The SPEIbase and “	droughtend_speibase	” was selected to generate the 	threshold values based on several criteria. The variable is recorded for the longest period of time within the PRIO	-GRID structure, 1949 to 2014, and 	it includes 	temperatures 	and measurements a	re deviations from long	-term mean values.
8 To be able to combine	 data from the SPEI with DHS data the data is reshaped separately for each dummy threshold such that dummies for each year 194	6 to 	2014 	are kept 	on the x	-axis	 while keeping the grid numbering on the y	-axis of the dataset. The two 	resulting datasets are then merged yielding one dataset with dummies 	showing whether there was existence of moderate and extreme drought for each year within each grid cell in the PRIO	-GRID structure.

4.3 Humanitarian Data Exchange

In addition to the use of data from the DHS and the SPEI, data on first administrative levels
(ADM1) 	is sourced from the Humanitarian Data Exchange	 (Humanitarian Data Exchange ,
2021)	. The inclusion of data on administrative levels is done t	o facilitate clustering of the
standard errors at 	this 	level	 to obtain heteroscedasticity robust standard errors. This 	met	hod is
further explained in chapter 5. 	Shapefiles with administrative borders 	for each country 	are
downloaded from the Humanitarian Data Exchange	 website	 and 	merged with the other data
sources using Arc	GIS	 Pro	. For each of the ten countries t	he 	selected	 sha	pefile	 is the 	most
recent version	 available.
Th	is is done	 despite the DHS also recording corresponding ADM1 levels for each interviewed
cluster. 	This is done to ensure grid cells pertain to one ADM1 level, rather than evolving with
name changes of 	ADM1 levels between	 DHS	 rounds.

4.4 Combining 	the datasets	 using ArcG	IS	 Pro

The software 	ArcGIS Pro is used in c	ombining the 	DHS clusters with 	PRIO	-GRID 	and
ADM1 levels	. All three 	data sources	 are available in shapefiles, which are representable	 using
the 	software	. The DHS c	lusters 	with longitude and latitude coordinates 	from each survey were
uploaded and merged into one shapefile, before uploading the PRIO	-GRID structure	 onto the
software. The	 polygons of PRIO	-GRID were joined with the poi	nts of the DHS clusters using
a spatial join tool	, capturing the points lying within each 	grid	 polygon	, and generating a new
layer of data with the combined information	 (ArcGIS Pro, 2021)	. The clusters that lie on
polygon borders are excluded in this procedure	, as including them in an arbitrary 	grid	 cell	s
could 	potentially 	lead to biased estimates	.
The new data layer	 containing 	numbers for	 identification	, cluster	s a	nd 	grid	 cells	, among other
variables,	 is then exportable to Excel format	, which in turn is imported 	into	 the statistical
software 	Stata.	 This is then merged with 	the combined DHS dataset, using 	cluster number,
year of 	interview and country of residence	 as the common id	entifiers. The merge is performed
using a 	one	-to-many	 merge in Stata, as one 	grid	 cell contains several individual records in the
DHS dataset.

ArcGIS Pro is also used in the same way to combine DHS clust	ers with first administrative
levels (ADM1). 	 Figure 	4.3 illustrates how this 	looks like for Burkina Faso and Malawi	.
 Figure 	4.3: Combining ADM1 with DHS clusters using ArcGIS Pro

Note: 	The 	DHS clusters from different survey round	s are	 represented by different 	colored	 points	,
while the within country borders are those of 	the ADM1 levels	. Burkina Faso is represented to the left,
while Malawi is represented to the right. The	 figure	s are	 developed using ArcG	IS Pro.

4.5 Construction of 	two 	variable	s for	 drought experiences

In this thesis 	individuals’	 experienc	es of	 drought 	occurrences are	 recorded in the period in
which 	individuals	 are eligible to attend school. 	The eligibility to attend school is defined as
the years in which 	individuals	 are five to 	fifteen	 years of age	. The selection of this span of
years 	is based on the m	ean and standard deviation of the 	main 	dependent variable “Years of
education”. 	The variable has a mean of 4,5 years, and a standard deviation of 4,6 years	 across
the whole sample	. Therefore, 	the selected year span of eligibility to attend school 	is deemed
reasonable.
For each individual 	two variables describing the years they were five and fifteen years of age
are constructed. 	This is then merged with the two sets 	of 	threshold 	values	 for drou	ght	, for
which 	two 	dummy 	variables 	for the years 	from 	1946 t	o 2014 exists, describing whether there
was 	moderate and extreme drought in each 	of the 	year	s. Based on these yearly dummy
variables 	for drought occurrence 	and 	the range of years where 	individual	s are	 eli	gible	 to

attend school, 	individuals are either affected by a drought during these years, or they are not	.
The result are two dummy variables	 named	 “Moderate drought while in school”	 and
“Extreme drought while in 	school” for each individual observation	 in the dataset	.
Using a year span of eligibility of school attendance, rather than a record of actual school
attendance has the 	benefit of capturing individuals who might have left school	 short	-term	 due
to drought	 experiences	. Using 	the	 year span of when individuals attended school would not
have captured this 	short	-term 	effect	.

4.6 Descriptive statistics 	of 	constructed 	dataset

The 	resulting 	dataset on which analysis is based 	consists of around 2	44	 000 individual
respondents, 	interviewed	 in 	years between 1990 and 2018	 and	 attending school between 1937
and 2003.
 Figure 	4.4: Distribution of respondents by country residence and decade born

Note: T	he percent of individual observations by country of 	residence	 and decade 	in which respondents
were 	born	. The figures are developed using Stata.
From figure 	4.4 we observe that 	Ethiopia, Malawi and Nigeria are the three largest countries
in the sample, while 	Niger, 	the 	Democratic 	Republic of	 Congo and Uganda are the three
countries	 with fewest observations. 	The	 majority of respondents, almost 	thirty	 percent	, were

born in the 1980	s. Only 236 individual respondents were born in the 1930s and due to the
impos	ing	 restrictions of completion of school within 2014	, which is the last recording by
PRIO	-GRID	, no one is born in the 21	st century. 	The sample size of those born in the earliest
decades is much smaller than the samples 	of those born in more recent decades. This fact,
together with the 	acknowledgement that some of the	 ones born in the 1930s and 1940s	 can
have difficulties remembering th	e exact years they attended school, mean the accuracy of
estimates based on these samples will be l	ower than the estimates for the more recent
samples.
The DHS primarily ask women about health and family planning, and thus females are
overrepresented in t	he surveys. In the constructed dataset	, sixty	-five	 percent 	of respondents
are	 female,	 and 	thirty	-five	 percent are 	male. 	Therefore,	 it is expected that the 	results are more
accurate for females than they are for males due t	o the	 larger sample size.	 Also	, g	iven that the
effect on males and females of experiencing drought is 	heterogenous, the overall effect will be
driven mostly by the effect experienced by females.
Table 	4.1: Gender frequency

4.6.1 Drought 	experiences

Due to the way the drought dummy variables are constructed, 	more individuals have
experienced moderate drought than extreme drought. Table 	4.2	 shows 	these percentages in the
constructed dataset.

Table 	4.2: Frequency of extreme and moderate drought during eligibility to attend
school
Extreme drought 	 	 Moderate drought
 	Frequency	 	Percent	 	Frequency	 	Percent
No	 	133 734	 	54.84	 	27 096	 	11.11
Yes	 	110 126	 	45.16	 	216 764	 	88.89
Total	 	243 860	 	100.00	 	243 860	 	100.00

Gender	 	Frequency	 	Percent
Male	 	84 064	 	34.47
Female	 	159 796	 	65.53
Total	 	243	 860	 	100.00

From table 	4.2 we observe that around 	fifty	 percent of individuals have experienced extreme
drought during their years when eligible to attend school. In contrast, almost 	ninety	 percent of
individuals have experienced moderate drought in the years	 when	 eligible to attend school.
Naturally, 	there are differences in these percentages across the ten countries in the sample	, as	.
represented in figure 	4.5.

Figure 	4.5: Frequency of drought during eligibility to attend school by country

Note: T	he percent of individual	s who have experienced 	extreme and moderate drought during
eligibility to attend school, 	within each of the selected countries. 	The figures are developed using
Stata.
From figure 	4.5 we observe t	hat the highest percentage has experienced extreme drought in
Mali and Niger, while in Nigeria, Tanzania and Uganda fewest 	individuals have	 experienced
drought while in school.
The percentage 	of those that have experienced moder	ate drought 	is on average higher and
more even 	across the ten countries 	than that of extreme drought. 	W	e also	 observe that
Tanzania and Uganda are two of the countries where there 	have	 been relatively few droughts
while individuals were eligible to attend 	school	.
Across time there are differences in the percentage of the sample who experienced drought
while eligible to attend school	. Figure 	4.6 show how this percentage has developed over time.

Figure 	4.6: Percent of individuals who experienced drought while eligible to attend
school

Note: 	The pe	rcent of individuals 	who have experienced mo	derate and 	extreme drought	 while eligible
to attend school	, over tim	e in the sample	. Calculated by the mean level 	for those born in the same
de	cade.	 The figure	 is developed using Stata.
From figure 4.	6 we observe that 	more individuals have experienced moderate drought, than
extreme drought	 while eligible to attend school	. For those born around 1970, almost 	sixty
percent experienced extreme drought while eligible to attend school	. The estimated
percentage for the earliest decades is less reliable than the 	more recent	 decades	, due to 	the
larg	er sample size.
4.6.2 Rural and urban areas

Acknowledging	 the	 exist	ence of	 a rural	-urban divide in educational attainment, it is
interesting to consider how this is represented in the constructed dataset	 (Sumida & Kawata,
2021)	. From table 	4.3 we observe the distribution of the sample that reside in urban and rural
areas.
Table 	4.3: Distribution of u	rban and rural residence among respondents
Respondents 	by residenc	e
Residenc	e 	Frequency	 	Percent
Urban	 	57 993	 	23.78
Rural	 	185 867	 	76.22
Total	 	243 860	 	100.00

From table 	4.3 we observe that seventy	-six percent reside in rural residences, while twenty	-
four percent live in urban areas. Considering the divide in educational attainment	 by

residen	ce	, table 	4.4	 represents the distribution of educational attainment according to type	 of
residence.
Table 	4.4: Percentage distribution of educational attainment
Education	 	Urban percent	 	Rural percent
Any education	 	79.01	 	54.65
Completed primary	 	53.57	 	24.67
Secondary or higher	 	41.87	 	14.53

From table 	4.4 we observe that the educational attainment is much lower in rural residences,
compared to urban ones. In urban areas 	seventy	-nine	 percent report having any education,
compared to 	only fifty	-five	 percent in rural areas. The difference	 is even larger when
considering secondary or higher education, with urban and rural residents reporting
respectively 	forty	-two	 and 	fifteen	 percent.
4.6.3 Educational attainment

The DHS record	s several variables related to educational attainment. 	Individuals	 report their
highest year of education, they report how many years they have attended 	education	 and their
highest level of 	completed education. 	As this thesis 	investigates the	 possibility that
individuals alter their 	educational patterns in the short term, the main variable of interest is the
variable that records 	how many years individuals have attended education. Figure 	4.7 shows
the development in this 	variable over time.
Figure 	4.7: Mean years of education reported across time

Note: 	Mean	 years of education over time	, c	alculated by the mean 	reported by	 those born in the same
year	 for each gender	 separately. 	The figure is developed using Stata.

From figure 	4.7 we observe that the mean years of education has been trending upwards since
1930.	 The mean years of education has risen remarkably from 1930s until the 1990s, from a	n
overall	 mean of 0	.3 to almost 6	.5 years of education. 	We also observe that there is a ge	nder
gap in the mean years of reported education but that this gap is narrowing.
From the dataset constructed we note changes across time 	also 	with respect to the highest
level of education attained. 	Figure 	4.8 displays this	 overall trend	, with the 	highest level of
educational 	attainment 	reported	 divided into three categories	.
Figure 	4.8: Trends in the highest level of education attained 	over time

Note: 	Developments in e	duc	ational attainment over time, calculated by the 	percentages	 of those born
in the same de	cade. 	The figure is developed using Stata.

From figure 	4.8 we observe that 	the share reporting having no education declines over the
decades	, while those reporting having 	any 	primary education and 	secondary or higher
education increase	, indicating that most individuals now start education	.

4.6.4 Literacy

Another 	variable that will be considered in the analysis is	 the	 level of literacy, as an indicator
of the quality of education. 	Table 	4.5 shows	 the level of literacy	 in the whole sample, by
gender	. A more comprehensive explanation o	n the construction of 	this variable will be given
in section 6.3.

Table 	4.5: Percentage	 of literacy by gender
Literacy	 	Female 	 Male
Not literal	 	54.66	 	35.79
Literal	 	45.34	 	64.21
Total	 	100.00	 	100.00

From table 	4.5 we observe that 	more males 	than women 	in the 	total 	sample are literal.
However, there has been changes to this 	percentage 	across time.	 Figure 	4.9 shows	 this
evolution.
Figure 	4.9: Evolution of literacy by gender

Note: 	Percent of literal individuals	 over time	, for 	the	 gender	s separately	. Calculated by the
percent	age of literacy	 for those born in the same	 decade	. The figure is developed using Stata.
From figure 	4.9 we observe that	 that the	 percentage of females who are literate	 lie
consistently below the one for males, but that the gap is narrowing.
As this thesis considers the 	effect on drought on educational attainment, it is interesting to
note how educational attainment correlates with literacy in the sample. 	In table 4.6 f	our
variables, representing the highest level of education attained by individuals	, and the
correlation to literacy is presented.

Table 	4.6: Correlation of educational levels and 	literacy
 Level of 	education	 	 Literal
 No education	 	-0.291
 Primary education	 	0.195
 Secondary education	 	0.354
 Higher education	 	0.391

From table 4.6 we observe 	that	 all levels of education h	ave	 a positive correlation with
literacy, 	while having no education correlates 	inversely	, indicating that education 	is 	an
important t	ool t	o increase literacy.

5 	Empirical strateg	ies

The following section will 	describe the empirical strateg	ies	 that 	are 	used to study linkages
between 	drought prevalence and 	educational attainment. 	In a first step	 ordinary least squares
estimation is used to evaluate the 	effect of drought on years of education obtained. Secondly,
a linear probability model is used to 	evaluate the effect on the level of education attained	 and
to evaluate the effect of drought on literacy. 	Both 	strategies involve fixed effects and
clustering	 to obtain 	heteroskedasticity	 robust standard errors. 	Lastly, the method of synthetic
control	s is applied to consider the effect of persistent drought on educational attainment.
5.1 	Ordinary L	east 	Squares 	estimation with fixed effects

The met	hod of 	ordinary least squares (	OLS	) is applied to study the effect of drought on years
of attained education. 	The dataset that is created consists of individuals who either have
experienced drought during years eligible to attend school, or they have not. 	Having this data
structure,	 it is theoretically possible to estimate the causal effect of drought experiences on
educational attainment, conditional on several criteria.
For OLS estimation to yield unbiased and consistent estimators, 	the Gauss	-Markov
assumptions 	must	 be fulfilled. 	If 	the assumptions are fulfilled 	the resulting least squares
estimators, are those that have the lowest possible sampling variance	 (Stock & Watson, 2015)	.
The following section mentions 	assumptions	 that are central to OLS estimation	 and how they
relate to the data	set	 and study design 	used	, before elaborating on the use of fixed effects in the
estimation.
A central 	underlying assumption for OLS to yield unbiased estimates is 	that the conditional
mean of the error term should be zero, meaning that there is no relationship between the
explan	atory variables and the error term. 	If this assumption 	holds,	 there should exist no
omitted variable bias influencing the estimated coefficients of the model.
If 	there is no presence of omitted variable bias	, the results of the estimation	s are close to th	e
true population parameters, and not influenced by other factors that are not controlled for.
When studying the effect of drought on educational attainment, it is important to note that
there exist several factors which affects educational attainment, tha	t could potentially bias the
results. In the analysis, some control variables are included, but the majority of omitted

variable bias is excluded due to the use of entity fixed effects on small geographical areas, as
well as time fixed effects.
There are 	several factors that can influence educational attainment and be correlated with
drought prevalence. Cultural aspects, norms, 	conflict	, and attitudes towards 	female education
are only some of the factors that can play an important role in determining the level of
educational attainment reached by the individual. These are factors which are difficult to
control for in the analysis and may result in omitted variable bias affecti	ng the validity of the
estimate	 if not controlled for 	using alternative methods	.
As 	observations had to be restricted to certain individuals who have not moved residence
since childhood	, the constructed dataset 	violates the 	OLS 	assumption concerning the
randomness of 	the sample	. However, an estimation of the magnitude of th	is potential bias is
performed in section 6.5, 	indicating	 that 	this	 does not introduce major bias.
Another assumption states	 that the model should be linear in its parameters	 (Stock & Watson,
2015)	. In the specification used no coefficients enter multiplicatively or quadratically, but
explanatory variables are interacted.	 For the constructed dataset, the presence of perfect mul	ti-
collinearity is tested for by 	considering	 the correlation of explanatory variables.	 This	 test
indicates that 	the explanatory variables do not perfectly predict one another, and 	the presence
of 	perfect multi	-collinearity 	is excluded	.
The assumption of	 no heteroscedasticity and	 no 	serial 	autocorrelation 	in the sample is
important for OLS to yield precise estimates. 	Heteroscedasticity is observed when t	he
variance of the 	error terms vary systematically with the predictor	 and serial correlation means
that the covariance of the error terms is non	-zero, meaning that the error terms can be
predicted by one another	 (Stock & Watson, 2015)	. To test for heteroscedasticity in the dataset
the White test	 is applied	 to a regression without fixed effects or clustered standard errors	. This
test indicates that 	the data is 	heteroscedastic	 with high significance	.
The presence of spatial correlation between treated and untreated groups in 	the constructed
dat	aset	 and heteroscedastic 	error terms 	presents a possible 	issue for the estimation, if not
addressed	. The 	adverse 	effects from this 	are	 limited 	using	 clustered standard errors	 in
estimation, producing heteroscedasticity robust standard errors	. T	his 	method 	is described
more in depth 	in section 5.4	.

As omitted variable bias 	is one of the most common 	issues 	when trying to expose causal
relationships, several methods 	have been	 developed to limit its adverse effects	 (Stock &
Watson, 2015)	. One way of addressing 	issues caused by omitted variable bias is to use fixed
effects on entity and/or time. 	An alternative to fixed effects estimation is the usage 	of random
effects estimation. To	 rule out the usage of random effects in the estimation, the Hausman test
is performed using Stata. Prior to running the Hausman test two estimations are performed,
one with fixed effects and one with random effects. The Hau	sman test then compares the
estimated coefficients from both models and rejects the usage of random effects if the
estimated coefficients are sufficiently different from one another. The Hausman test
confirmed that fixed effects perform better than random 	effect for estimation with the
constructed dataset.
In 	this thesis	 effects are fixed at 	relatively 	small geographical area	s, as well as across time,
implicitly studying variation within each selected area for individuals that are born in the
same 	decade. Employing both entity	- and time fixed effects omits much of the potential
omitted variable bias from variables that vary across areas and across time	 (Stock & Watson,
2015)	. Using fixed effects on small geographical area	s, within each country,	 mean cross	-
country differences are also accounted for.
Employing both entity and time fixed effects is appropriate when there are omitted variables
that are constant over time but vary across areas, and omitted variables that are c	onstant
across areas but vary over time. In this analysis it can be argued that both types of omitted
variables exist. 	Two e	xample	s of such variables 	are	 that cultural aspects vary between areas,
and attitudes towards universal education has varied 	across	 time.
When employing entity fixed effects one is subtracting the mean values for each area from
individual observations in the given area, and therefore studying deviations from mean values.
Fixed effects estimation 	on area 	assumes that everyone in the sa	me area has been subject to
the same incidences and trends, allowing us to detect the effect of someone being affected by
drought.
Similarly, to the intuition of the entity fixed model, employing time fixed effects allows each
time period	, or decade,	 to h	ave its own intercept. This allows to study effects that vary over
time, but not across 	areas	. The intercept of different decades means we subtract the mean of
each decade, leaving us to study the effect for those born in the same decade.

The functional f	orm 	with both entity	- and time fixed effects i	s illustrated in equation 5.	1
where	 αi is the intercept for each area studied, and	 λt is the intercept for each time period
studied	.
Years	 of education	it = β0+β1Drought	1,it +⋯	+βkXk,it+αi+ λt+ µit 	 	(5.	1)
In	 equation 5.1 β	1 is the main coefficient of interest in this thesis, 	while 	βk represents the
coefficients of the control variables. 	This is the equation that is used to study the effect of
drought on years of attained education	.
The statistical software 	Stata is used to run the model with fixed effects 	using	 the constructed
dataset. Employing this method means that there needs to 	exist	 variation in experiences of
drought within the f	ixed 	areas for those born in the same decade, fo	r meaningful results to
emerge. If everyone in an area, born in the same decade, experience drought, the model will
not provide meaningful results.
Omitted variable bias can still occur due to area specific characteristics that vary over time,
for which t	he fixed effect method does not control. For this to be an issue, there must be such
omitted variables that are correlated to both drought occurrence and educational attainment.

5.2 Linear probability model on the level of education 	attained

In addition to 	recording 	years of attained education, the DHS also records the 	highest level of
education attained by 	the 	individual.	 To study the effect of drought on different levels of
attained education a linear probability model with OLS	, fixed effect	s and clustering 	is
applied.
The DHS records 	the highest level of educational 	attainment in five 	categories	: no education,
incomplete primary education, complete primary, 	incomplete secondary, complete secondary
and higher educatio	n.
W	e have reason to believe that 	a drought occurrence affects children of different ages
differently 	(Bjorkman	-Nyqvist, 2013	; Mani, 	et al.	, 2013)	. This	 indicat	es that there is 	a non	-
linearity in 	the effect of drought	 on different levels of education	. This 	implies that 	regression
analysis with a categorical	 dependent variable can yield biased results. 	To resolve this issue 	a
linear probability model	, with a binary dependent variable, is applied	 on three different level	s
of education	 separately	. The linear probability model is chosen over the ordered logistic and

ordered probit models due to its 	clear	 interpretation	 and 	the 	simple	 inclusion of fixed effects	.
A 	brief 	demonstration of the performance of 	the ordered logistic and ordered probit models 	is
presented 	in 	appendix 10.1	 table 	A10	.
The linear probability model 	entails	 that the dependent variable 	takes	 the value of either 	one
or 	zero	, and the 	coefficients	 of the explanatory variables are interpreted as 	the change in
probability that the dependent variable takes the value 	one	 if the explanatory variable increase
by one unit	. The three dependent variables used to study the effect of drought on ed	ucational
level attained are	: having	 any education, having completed primary 	education,	 and having
completed secondary or higher education. 	The 	functional form 	of this model 	is presented by
equation 5.2.
P(Educational	 level	it =	1)= β0+ β1Drought	 it + β2X1,it +⋯	+βkXk,it+ αi + λt+ µit 	(5.2)

5.3 Control variables 	included in the analysis

Control variables are added to the estimation, 	to reduce the omitted variable bias	 that is not
eliminated by fixed effects,	 as much as possible. This 	is done to resolve	 issues related to
endogeneity and identification issues of the model. 	However, including more control
variables 	can increase the variance of the 	main 	variable of interes	t, leaving us with a less
accurate estimate	 (Stock & Watson, 2015)	. Therefore, 	the control variables that are added 	to
the estimation are 	thoroughly assessed 	before inclusion, 	with respect to the	 possibility that
they cause 	omitted variable bias.
An e	xample of 	a variabl	e that can be correlated to both drought occurrence and 	educational
obtainment, while 	varying over time 	and 	within locations,	 is the wealth 	quintile	 to which an
individual belongs.	 This also varies for individuals born in the same decade.
Wealth of a family is 	a strong predictor of educational 	attainment 	of its children	 (Knight &
Shi, 1996)	. At the same time wealth could be strongly correlated with drought occurrence in
agricultural dependent areas. 	Th	e wealth of families	 can vary over time within the	 areas 	and is
therefore	 included as a control variable in the specification. 	However, it is important to note
that controlling for wealth can pose a potential issue, as some of the educational adaptation
patterns to drought 	are expected to happen due to	 a shift in family income, 	which potentially

also 	affects wealth. This means income could	 have been 	the dependent variable studied, and
therefore should not be used as an explanatory variable.
However, t	he DHS does not record income directly 	but calculate	s a wealth index 	that is
relative 	for each survey based on the 	household’s	 ownership of assets 	such as 	refrigerator	, TV
and	 access to 	water sources	, that are 	indicative of	 their living conditions	 (Rutstein & Johnson,
2004)	. As living conditions	 can be an important predictor of educational attainment, the
wealth 	quintile 	is included as a control variable despite 	being related to income and the
potential 	issues	 this might cause	 (Filmer & Pritchett, 2001)	.
Other control	 variables are added as they pose great predictive power on 	educational
attainment. One of these variables is the type of residence of individuals	, whether they live in
urban or rural areas. As demonstrated in 	section 4.	6.2 educational attainment vary greatly
between rural and urban areas.	 There exist several 	explanations	 of	 this difference in
educational attainment between rural and urban areas. 	In	 rural areas t	he distance	 to school	 is
on average	 longer, infrastructure is 	generally	 less developed	, and c	ultural norms towards
especially 	female education tends to be more conservative for rural residen	ts 	(Knight & Shi,
1996	; Linard, et al., 2012)	.
Furthermore, child labor and agricultural production is highly correlated 	in many countries. 	In
2010 it was estimated that 60 percent of all child 	laborers	 work in agriculture	, and the
majority of these are unpaid family members	 (International Labour Organization, 2010)	.
With this being 	an issue mainly in rural 	areas, where agricultural 	production	 is higher	, it adds
to the	 divide in educational attainment between rural and urban areas.
Another control variable that is included in the specification is 	the 	gender of individuals.	 Due
to 	the gender disparity in education	 gender 	proves	 an important control variable. 	This
becomes e	specially 	important 	when 	studying 	heterogeneous effects o	f drou	ght on	 gender	.
Religion is included as another control variable, as this can be indicative of cultural norms,
which can vary 	both 	across and within areas. The composition of religion	 in an area	 could
also vary from decade to decade, although families tend to keep to the same religion over
time.
The land used for agriculture within the area is another control variable	, sourced from PRIO	-
GRID and the European Space Agency Glob	al 	Cover Portal	 (ESA Globcover Project, 2021)	.
It is possible to infer that the higher this variable is, the more dependent on agriculture

individuals within this area are. This control variable is recorded in the year the individuals
were eligibl	e to attend school, given that this was during or after the 1960s. For individuals
attending school prior to 1960 the value recorded in 1960 is inferred. This is done as the first
recording of the variable 	is in 1960.
The age of respondents is also include	d as a control variable. 	This is important as wealth and
religion	 are	 variables which are	 recorded at the time of the survey. 	Individuals who are in
their 30s might differ from individuals who are in their 40s, with respect to for instance
accumulated wealth. 	In 	addition,	 the decade in which individuals were born 	is included as a
time fixed effect, as we know there has been 	great	 development with regards to educational
attainment over time.
Information is given in the year individuals partook in the survey. To limit bias from this,
years in which individuals partook is divided into ranges of five years and included as a time
fixe	d effect. However, as information of variables 	such as religion and wealth 	are given at the
time of the survey in which individuals partook, assumptions must be made regarding the
relevance of those variables in predicting conditions during childhood. 	It i	s possible to argue
that low social mobility makes relative wealth quintile today predictive of conditions in the
past	 (Alesina, 	et al.	, 2021)	. The choice has therefore been made to include it in the analysis
regardless of its shortcomings. Furthermore, in	dividuals seldomly move between religions.
Therefore, religion today is used to control for the outcome in the past.
Controlling for age, survey	 period and decade in which individuals are born mean we have a
strong correlation between these 	three 	variables	, and the signs and magnitudes of these
estimated coefficients should therefore not be interpreted	. The decision is made to include
these three v	ariables 	as controls as it is believed that 	they together control for much of
potential omitted variable bias.
The main coefficients of interest are th	e coefficients	 of the 	two 	drought 	threshold 	variables.
Therefore, the 	coefficients 	of several of the control variables 	are	 not displayed in the
regression tables. 	The full regression tables and dataset can be obtained at request. 	 A
summary of the variables and their functions in the 	regressions is found in appendix 1	0.2 table
A9.

5.4 Clustered standard errors 	to resolve spatial correlation in the error terms

As climatic shocks are determined by large weather systems, they are likely to affect large
areas simultaneously. This means that 	the 	grid	 cells experiencing	 droughts are not at random
within an area	, implying	 existence of 	spatial 	correlation in the 	error terms	.
This issue can be solved 	using	 clustered standard errors. 	By employing clustering 	on the
standard 	errors,	 we 	allow the error terms to be arbitrarily correlated withing the clusters. 	If the
appropriate clustering level is employed the method	 yield 	heteroscedasticity 	robust standard
error	s. This can matter greatly for the statistical significance of the independent variables, as
the size 	of	 standard errors usually 	increase in size	, compared with estimations without
clustering	.
Clustering on first administrative level 	(ADM1) 	is likely to be the best option to resolve the
spatial correlation in the error terms, as th	ese are	 larger areas than the g	rid cells and might
resolve the issue of drought prevalence 	over 	larger areas. 	Cluste	ring could also be done on
country level, but 	for the larger 	countr	ies	 in the sample 	drought 	is less likely to affect the
entire country 	at the same time.
To be able to cluster 	the grid cells	 within the ADM1	 levels	 sourced from Humanitarian Data
Exchange	 some 	grid cells 	need to	 be split	. PRIO	-GRID cells do not remain within country,
nor within administrative borders	, generating an overlap between grid cells and other borders.
The overlap means grid cells are not nes	ted within administrative levels. 	To 	solve	 this overlap
a new set of 	grid	 cells are generated, by combining a 	unique 	country code, the code for
ADM1	 levels	 and the 	PRIO	-GRID	 cell code. Th	e resulting, 	new g	rid cells pertain to one
country and one ADM1 	level	 only	. The procedure of generating new grid cells is explained in
more detail in appendix 10.	2. Depending on	 analysis	, the resulting clusters	 based on ADM1
levels	 range between 160 	and	 190	 clusters	.

5.5 Synthetic 	contro	l met	hod to evaluate 	impact of persistent drought

The main focus of this thesis 	is 	on 	evaluating the	 effect	s of short	-term dro	ught experiences	 on
educational attainment	. However, as droughts 	can 	span across several years	, the	 possible
effects	 this	 can cause	 is also investigated briefly	 using a famine in Ethiopia as an event stu	dy	.
To evaluate 	whether	 mean years of education developed differently in Ethiopia compared to
other countries, 	who did not experience this famine, the method of synthetic controls is
applied	 (Abadie & Gardeazabal, 2003)	.
The method extends on the methodology of difference	-in-difference estimation, where	 the
trend	s of	 two groups	, one affected and one unaffected,	 are compared 	before and after a shock
or an 	intervention to determine	 its	 effects. 	A central assumption of the difference	-in-
difference method	ology	 is that the treated and control groups have the same trend in the
outcome variable prior to the shock	 (Stock & Watson, 2015)	. Otherwise, the observed
difference	-in-difference estimator can be biased by other underlying trends, and not be a
product of the shock alone.
As 	it is 	often 	challenging to 	find	 two groups where the trend is e	qual prior to the shock, the
method of synthetic controls 	poses an advantage	. The method 	derived by Abadie and
Gardeazabal entails 	creat	ing	 an artificial	 control group by combining selected 	untreated
groups, in a weighted average	, so that the artificial co	ntrol group 	resemble the trend of the
treated group prior to the shock	 as much as possible	 (Abadie & Gardeazabal, 2003)	.
For the dataset at hand, 	the 	synthetic control method is used to study the effect of the
infamous famine in Ethiopia during the 1980s. The other nine countries 	in the constructed
dataset 	are used to generate 	the	 synthetic control group	, in a weighted average	, to lay as close
to the 	pr	e-famine Ethiopian trend. 	This entails that some of the nine 	countries are used more
than others in the construction of the weighted average, but some information is sourced from
all nine countries.

6 	Results

This section will give an overview of the results 	of the analysis	. Several 	linkages between
drought and educational outcomes are investigated	, including the effect on years of education,
the level of education, the 	effect on literacy and heterogeneous effe	cts	 of drought	. An analysis
is also 	conducted to assess the effects 	of droughts 	across time	.
Lastly the section provides a	 brief event study of the Ethiopian famine, and a 	robustness
analysis of the model	s that 	are	 used to investigate the relations between education and
drought.
In all the 	sections	 of this chapter	 both threshold values of drought are included. D	ue to the
way the	se variables are	 constructed, anyone who experiences extreme drought ha	s also
experience	d moderate drought. Therefore, in all instances where the extreme drought dummy
variable 	is 	one	, the moderate drought dummy 	variable 	is also 	one	.
When interpreting the coefficients of the drought dummies it is important to consider the
change in 	control group between the two dumm	y variables	. The coefficient of the moderate
drought dummy is interpreted as the effect of going from no drought, to experiencing a
moderate drought or worse, while the coefficient of the extreme drought dummy is the effec	t
of going from a situation of moderate drought to even worse drought. Therefore, the effect of
going from no drought to a situation with extreme drought is	 interpreted as	 the sum of the two
estimate	d drought 	coefficients	.
Using fixed effects, the 	interpretation of the coefficients is the 	estimated 	effect within each
grid cell area for all specifications, and within each decade when decade born is included as a
fixed effect	.

6.1 	Years of e	ducation	 and drought

A correlation matrix of the two 	threshold values	 of drought with years of education is
generated and	 can be found in appendix	 10.1	 table A8	. The matrix shows negative and
significant correlation of both types of drought and education. 	However, this effect 	might be
driven by other factors, and do not indicate a causal relationship	 between the variables	. To
investigate whether a significant relationship exists the method	s described in chapter 5 are
applied	.

The following four subsections all investigate the relationship between drought and years of
attained education.
6.1.1 Effect of drought 	on 	years of 	education	 attained	 for the	 full sample

In table 6.1	 five specifications are used to study the links between drought and years of
attained education. All specifications in this, and the following regression tables, include
fixed effects on the constructed grid cell level, and clustering on first administrati	ve level to
obtain 	heteroscedasticity	 robust standard errors. 	Significant results 	at minimum 95 percent
confidence intervals 	for 	key variables of interest 	are 	outlined in bold in all 	tables to follow.

Table 	6.1: The effect of drought on years of education attained
Years of education attained	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate Drought	 	-0.124	 	-0.0993	 	-0.252	*** 	-0.249	*** 	-0.232	*** 	 	 	(0.130)	 	(0.125)	 	(0.0590)	 	(0.0593)	 	(0.0571)	 	 	 	 	 	 	 	 	 	Extreme Drought	 	-0.0265	 	-0.00866	 	0.0327	 	0.0371	 	0.0481	 	 	 	(0.177)	 	(0.164)	 	(0.0591)	 	(0.0626)	 	(0.0582)	 	 	 	 	 	 	 	 	 	Female	 	 	-1.443	*** 	-1.440	*** 	-1.451	*** 	-1.420	*** 	 	 	 	(0.0884)	 	(0.0880)	 	(0.0888)	 	(0.0933)	 	 	 	 	 	 	 	 	 	Urban	 	 	2.848	*** 	2.677	*** 	2.687	*** 	1.604	*** 	 	 	 	(0.211)	 	(0.206)	 	(0.191)	 	(0.163)	 	 	 	 	 	 	 	 	 	Constant	 	4.600	*** 	4.839	*** 	1.941	*** 	1.199	** 	-0.214	 	 	 	(0.0909)	 	(0.111)	 	(0.201)	 	(0.374)	 	(0.444)	 	 	N 	243718	 	243718	 	243718	 	233690	 	228872	 	 	 R2 	0.000	 	0.091	 	0.162	 	0.180	 	0.238
F (1, 	188)	 	1.16	 	0.68	 	9.77 	8.65	 	7.28	 	 	P-value	 	0.2834	 	0.4101	 	0.0021	** 	0.00	37*** 	0.00	77**
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed
No No No No No No
No No No No No No
No No No Yes	 No No
Yes	 No No Yes	 Yes	 No
Yes	 Yes	 Yes	 Yes	 Yes	 Yes

Note: 	Standard errors are clustered at the ADM1 	level and	 are 	heteroscedasticity robust. They are given in parentheses for all specifications	. All specifications 	have fixed effects on grid cell level. 	 	The 	F-test is testing the significance of the sum of the two drought dummy variables	. Degrees of freedom are given in parenthesis, and the p	-value in the 	following row. 	For specification 5 the F	-statistic has d	egrees of freedom F (1, 157). 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.

From regression table 6.1	 we observe that experiencing moderate drought has a negative and
significant effect on years of education attained in specification 3	-5. Depending on selected

specification, individuals who have experienced moderate drought spend from a fifth of a
year, to	 a quarter of a year, less in school than those who have not experienced drought.
The added effect of going from a moderate drought to an extreme drought is small, and not
statistically different from zero. 	The 	F-test is testing the 	significance of the su	m of the two
drought dummies	. From this we observe that the effect of an extreme drought compared to a
situation of no drought is significantly different from zero in specification 3	-5. 	This indicates
that 	some of the effect on 	education	 detected in the coefficient of the moderate drought
variable	 is driven by the effect of an extreme drought.

The 	coefficient	s of the drought variables	 change	 as religion, wealth	, agriculture, age and
period interviewed are included as control	 variables	. This indicate that the introduced control
variables explain some of the variation in years of attained education. 	The 	included 	R2 shows
how much of the variation within each fixed area and time period in educational attainment is
explained 	by the model. 	However, including more 	explanatory variables will always increase
this R2 value	 (Stock & Watson, 2015)	. Therefore,	 not much emphasis will be put on
interpreting this	 value with respect to the 	goodness of fit of the model	.
6.1.2 	Effects of drought on subsets of countries

Several studies have found that 	in 	areas that 	frequently experience drought, people have
adopted 	both short	- and long	-term adaption strategies, to mitigate the adverse effects of crop
loss 	(Opiyo, 	et al.	, 2015)	. To investigate whether this 	implies that effects of droughts on
educational outcomes are different 	depending on long	-term climatic conditions	, the dataset is
divided	 into two subsets.
In 	table 6.2, 	panel A the relative	ly humid countries	 of Nigeria and the Democratic Republic of
Congo are investigated 	and in	 panel 	B, the dr	ier	 countries 	of Burkina Faso, Mali, Niger,
Ethiopia,	 and Kenya 	are	 studied	 (Ellis & Galvin, 1994)	.

Table 	6.2: The effect of drought in dry and humid countries
Years of education 	attained
Panel A: 	humid	 countries	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate Drought	 	-0.0735	 	-0.0636	 	-0.0589	 	-0.0599	 	-0.0408	 	 	(0.187)	 	(0.187)	 	(0.0929)	 	(0.0908)	 	(0.0880)	 	 	 	 	 	 	 	Extreme Drought	 	-0.957***	 	-0.952***	 	-0.202*	 	-0.201*	 	-0.200*	 	 	(0.166)	 	(0.155)	 	(0.0990)	 	(0.0989)	 	(0.0903)	 	 	 	 	 	 	 	Female	 	 	-2.241***	 	-2.366***	 	-2.345***	 	-2.259***	 	 	 	(0.141)	 	(0.140)	 	(0.145)	 	(0.155)	 	 	 	 	 	 	 	Urban	 	 	2.588***	 	2.386***	 	2.414***	 	0.820***	 	 	 	(0.243)	 	(0.242)	 	(0.254)	 	(0.111)	 	 	 	 	 	 	 	Constant	 	6.208***	 	6.862***	 	2.345***	 	2.526***	 	0.289	 	 	(0.162)	 	(0.191)	 	(0.339)	 	(0.671)	 	(0.473)	 	N 	62714	 	62714	 	62714	 	62314	 	61611	 	R2 	0.008	 	0.108	 	0.176	 	0.196	 	0.291
F (1, 62)	 	18.20	 	20.14	 	4.40	 	4.38	 	4.01	 	P-value	 	0.0001	*** 	0.0000	*** 	0.0401*	 	0.0404	* 	0.0497	*
Panel B: Dry countries	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate Drought	 	0.130	 	0.139	 	-0.335	*** 	-0.311	*** 	-0.301	*** 	 	(0.147)	 	(0.138)	 	(0.0840)	 	(0.0843)	 	(0.0811)	 	 	 	 	 	 	 	Extreme Drought	 	-0.115	 	-0.0811	 	0.0307	 	0.0199	 	0.0255	 	 	(0.0790)	 	(0.0727)	 	(0.0456)	 	(0.0433)	 	(0.0434)	 	 	 	 	 	 	 	Female	 	 	-1.262	*** 	-1.184	*** 	-1.142	*** 	-1.154	*** 	 	 	(0.135)	 	(0.121)	 	(0.109)	 	(0.114)	 	 	 	 	 	 	 	Urban	 	 	3.360	*** 	3.216	*** 	3.044	*** 	2.150	*** 	 	 	(0.425)	 	(0.411)	 	(0.338)	 	(0.295)	 	 	 	 	 	 	 	Constant	 	3.279	*** 	3.178	*** 	0.932	*** 	0.0719	 	-0.622	 	 	(0.137)	 	(0.144)	 	(0.232)	 	(0.537)	 	(0.476)	 	N 	115625	 	115625	 	115625	 	115158	 	114682	 	R2 	0.000	 	0.117	 	0.175	 	0.191	 	0.225
F (1, 87)	 	0.01	 	0.14	 	15.13	 	13.04	 	12.25	 	P-value	 	0.9249	 	0.7057	 	0.000	2*** 	0.0005	*** 	0.0007	***
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed 	 	 Gender interactions
No No No No No No No
No No No No No No No
No No No Yes	 No No No
Yes	 No No Yes	 Yes	 No No
Yes	 Yes	 Yes	 Yes	 Yes	 Yes	 No
Note: 	Standard errors are clustered at the ADM1 level, and 	heteroscedasticity robust. They are given in parentheses for all specifications	. All specifications have 	fixed effects on grid cell level. 	 	The 	F-test is testing the significance of the sum of the two drought dummy variables	. Degrees of freedom are given in parenthesis, and the p	-value in the 	following row. 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.

In the 	relatively humid countries	 of DR Congo and 	Nigeria 	going from experiencing a
moderate drought to 	experiencing an 	extreme drought has a negative and significant effect on
years of education.	 Specification	 3-5 estimate this effect to about a quarter 	of a year 	less time
spent in school.
For the 	drier	 countries	 it is the moderate drought dummy that is 	negative and 	significant	.
Going from a situation with no drought, to a situation with drought 	implies	 individuals attain
about a third 	of a year 	less schooling. 	The effect of going from a moderate drought 	to an
extreme drought is not significant in explaining years of education.
The F	-test of 	significance of the sum of the two drought 	coefficients	 is included in 	both
panels	, showing that the effect of an extreme drought is 	different from zero in the majority of
specifications	.
We observe that for the dr	ier	 countries, the effect of a drought is 	pronounced	 enough to adapt
educational patterns if the drought is 	moderate or wor	se, while for the more humid 	countries
educational patterns are only adapted if a drought is categorized as extreme or worse. 	Th	e
observed 	difference between the dry and 	more humid	 countries	 can be explained by the
construction of the SPEI	 as a deviation 	from mean long	-term conditions in each locality	, as
described more in depth in section 4.2.2. 	Therefore, it is natural that 	the more humid 	countries
only feel the impact of the drought enough to adapt 	education	al patterns 	when the drought is
categorized	 as extreme. 	For the 	drier	 countries	, a drought	 of any kind	 is enough of a deviation
from their already dry long	-term mean 	to adapt.
To investigate whether 	a drought 	has fewer consequences 	in already 	drier	 areas	, due to 	better
adaptation mechanism	s, an absolute measure of droughts should be applied. 	The results do
however indicate that 	the variables used for analysis work appropriately for the purpose of
this analysis	, as they confirm th	at the	 construction of the SPEI 	values	 make sense	.
Choosing selected countries for analysis	 is 	interesting but	 can 	risk threatening the validity of
results	 by 	testing	 multiple hypothesis.	 The more 	hypothesizes	 are tested, the more likely it is
to observe a significant result	 by 	random	, as the chance that the result lie outside of the 95
percent confidence interval increases	 (Stock & Watson, 2015)	. Stated differently, if	 twenty
hypotheses	 are tested, one is likely to yield significan	t results due to the acceptance of a
significance level of 	95 percent	. This is worth noting as we assess the significance of	 repeated
testing	.

6.1.3 Heterogeneous effects on gender

The results	 from regression table 6.1 and 6.2	 show that a drought does affect the years of
education attained negatively 	from a third of a year to a quarter of a year, depending on
sample and specification. 	From the descriptive statistics we 	noted 	that there are 	differences in
educational attainment 	between the genders, and 	previous research has shown that there 	exists
a gen	dered aspect of adaptation to weather shocks 	(Bjorkman	-Nyqvist, 2013	; Hertel & Rosch,
2010)	. Therefore, it	 is interesting investigate whether there are heterogeneous effects on years
of education attained 	with respect to	 gender	 in our sample	.
This is done by 	introducing 	an 	interaction term describing whether an individual 	is 	male and
has 	experienced drought	 while eligible to attend school	. In regression table 	6.3 all
specifications	 include control variables for 	age, wealth, agricultural land, type of residence,
religion	 and 	year interviewed	. All 	specifications	 have 	entity 	fixed 	effects on grid 	cell 	level
and time fixed effects on decade in which individuals were born.
The tree panels 	in table 6.3 	contain the same samples as presented in section 	6.1	 and 6.2	, and
the interaction terms are included for male	. This allows us also to study the effect 	of d	rought
for females	, presented by the	 estimated	 coefficient of the drought variables	. For males, the
total effect of a drought 	on education 	is the sum of the 	estimated 	coefficient of the 	interaction
variable and the corresponding 	estimated 	coefficient of the drought variable. 	A table with all
four interaction terms is included in 	appendix 10.1	 table A	1.

Table 	6.3: Heterogeneous effects on 	gender	 by sample
Years of education attained

Panel B: 	Humid	 countries	 	(1) 	(2) 	(3) 	 	Moderate	 	Extreme	 	Moderate & Extreme
Extreme Drought	 	 	-0.454	*** 	-0.427	** 	 	 	(0.125)	 	(0.124)	 	 	 	 	 	Male x Extreme Drought	 	 	0.764	*** 	0.697	*** 	 	 	(0.197)	 	(0.200)	 	 	 	 	 	Moderate Drought	 	-0.2	64*	 	 	-0.143	 	 	(0.115)	 	 	(0.115)	 	 	 	 	 	Male x Moderate Drought	 	0.56	3** 	 	0.330	 	 	(0.172)	 	 	(0.174)	 	 	 	 	 	Female	 	-1.774	*** 	-2.041	*** 	-1.778	*** 	 	(0.215)	 	(0.174)	 	(0.215)	 	 	 	 	 	Urban	 	0.822	*** 	0.823	*** 	0.823	*** 	 	(0.111)	 	(0.110)	 	(0.110)	 	 	 	 	 	Constant	 	-0.0569	 	0.118	 	-0.0456	 	 	(0.466)	 	(0.454)	 	(0.459)	 	N 	61611	 	61611	 	61611	 	R2 	0.291	 	0.292	 	0.292
F (1, 62)	 	 	 	18.90	 	P-value	 	 	 	0.0001	***

Panel C: 	Dry 	countries	 	(1) 	(2) 	(3) 	 	Moderate	 	Extreme	 	Moderate & Extreme
Extreme Drought	 	 	-0.0456	 	0.0167	 	 	 	(0.0485)	 	(0.0480)	 	 	 	 	 	Male x Extreme Drought	 	 	0.0551	 	0.0210	 	 	 	(0.0980)	 	(0.108)	 	 	 	 	 	Moderate Drought	 	-0.4	03***	 	 	-0.408***	 	 	(0.0933)	 	 	(0.0896)	 	 	 	 	 	Male x Moderate Drought	 	0.234	 	 	0.221	 	 	(0.126)	 	 	(0.142)
Panel A: Full sample	 	(1) 	(2) 	(3) 	 	Moderate	 	Extreme	 	Moderate & Extreme
Extreme Drought	 	 	0.0315	 	0.0759	 	 	 	(0.0800)	 	(0.0768)	 	 	 	 	 	Male x Extreme Drought	 	 	-0.0490	 	-0.0820	 	 	 	(0.100)	 	(0.101)	 	 	 	 	 	Moderate Drought	 	-0.278	*** 	 	-0.311	*** 	 	(0.0678)	 	 	(0.0677)	 	 	 	 	 	Male x Moderate Drought	 	0.168	 	 	0.208	* 	 	(0.103)	 	 	(0.104)	 	 	 	 	 	Female	 	-1.271	*** 	-1.443	*** 	-1.271	*** 	 	(0.126)	 	(0.0973)	 	(0.126)	 	 	 	 	 	Urban	 	1.604	*** 	1.604	*** 	1.604	*** 	 	(0.162)	 	(0.163)	 	(0.162)	 	 	 	 	 	Constant	 	-0.277	 	-0.341	 	-0.284	 	 	(0.454)	 	(0.440)	 	(0.452)	 	N 	228872	 	228872	 	228872	 	R2 	0.238	 	0.238	 	0.238
F (1, 157)	 	 	 	1.03	 	P-value	 	 	 	0.3115

The regression design in table 6.3 implies the following
with respect to the effect of 	an extreme 	drought:

The effect on females:
βModerateDrought 	+ β	ExtremeDrought 	 	 	(1)

The e	ffect on males:
βModerateDrought 	+ β	ExtremeDrought	 + β	Male x ModerateDrought 	+
βMale x ExtremeDrought	 	 	 	(2)

By combining (1) and (2) we note that if heterogeneous
effects of extreme drought exist:
βMale x ModerateDrought 	+ β	Male x ExtremeDrought	 ≠ (3)

The F	-test in column 3 is testing equation (3)

Text box 6.1.

Female	 	-0.940***	 	-1.126***	 	-0.940***	 	 	(0.143)	 	(0.130)	 	(0.143)	 	 	 	 	 	Urban	 	2.150***	 	2.150***	 	2.150***	 	 	(0.295)	 	(0.295)	 	(0.295)	 	 	 	 	 	Constant	 	-0.698	 	-0.801	 	-0.697	 	 	(0.478)	 	(0.464)	 	(0.476)	 	N 	114682	 	114682	 	114682	 	R2 	0.225	 	0.224	 	0.225
F (1, 87)	 	 	 	3.34	 	P-value	 	 	 	0.0710	 	Note: 	Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all specifi	cations	. All specifications 	include controls for gender, age, wealth, agricultural land, type of residence and period in whic	h they were interviewed. All specifications have fixed effects 	grid	 	level and clustered at ADM1.	 	The 	F-test is testing the significance of the sum of the two drought dummy variables	. Degrees of freedom are given in parenthesis, and 	the p	-value in the 	following row. 	 	Significance levels: * p<0.05, ** p<0.01, *** p<0.001

In specification 1 we observe 	the two types of drought, and the interaction term for 	male and
moderate drought. 	In specification 2 we observe the same, only for 	extreme drought, and in
specification number 3 we see all four terms describing drought and the interaction	s with
male. 	Below each panel is the 	F-statistic and its 	p-value displayed	 in column 3	. This	 F-test
indicates that the effect of extreme drought has heterogeneous effects 	on gender in the
relatively	 humid	 countries	, but not in the two other samples.

From 	panel A, 	the full sample	, specification numbe	r 3 we observe that there are clear
heterogene	ous effects of moderate drought between the genders. 	The effect for a 	male 	of
experiencing moderate drought is interpreted as the sum of the coefficients of 	moderate
drought and the interaction term. 	Following the reasoning from 	text box 6.1. we know that
heterogeneous effects between the genders exist in specification 1 and 2 if the coefficient on
the interaction term with male is significant. 	From specification 3, we note that relative to
females, ma	les attend a quarter of a year more education in the presence of drought.

Running the model with interaction terms for females	, as shown in 	appendix 10.1	, confirm
that there is a	n overall 	small, negative effect for males, but this effect is not significant. 	For
females the negative effect is significant and negative, indicating that they attain around a
third of a year less education	 in the full sample.

From panel B, we observe that the 	heterogeneous effects 	between males and females are
larger in the more humid 	countries	, with 	males	 experiencing a moderate drought attaining
more than	 half a year 	more	 education	 relative to 	females	. The heterogeneous effec	t of
experiencing an extreme drought is even larger. 	 Given that the effect 	for males of

experiencing a drought is likely to be positive, the negative effect of drought on years of
education detected in the previous sections is likely to be driven by an ad	verse effect for
females.

From panel C, the 	drier	 countries	, we o	bserve that the effect of experiencing a	 moderate
drought, relative to no drought is about the same as it was for the humid 	countries	 when the
drought went from moderate to extreme. 	Although the 	estimated 	coefficient of the male
interaction term is positive, while the drought coefficient is negative, t	he	 heterogenous effect
on gender	, represented by the F	-statistic	 is 	not 	statistically significant.

Th	e difference 	in adaptation 	to the drought levels 	between humid and dry 	countries	 is likely
to be explained by the same reasoning as was applied in section 6.1.2 regarding the
construction of the SPEI values. It is however interesting to note that 	all three	 sample	s seem
to confirm that 	droughts 	can contribute to widen the gender gap in	 education	, and that the
heterogeneous effect seems to be largest in the more humid 	countries	.

The	 larger heterogenous effect in the humid 	countries	 can possibly be explained by the role of
males in agricultural prod	uction. The humid 	countries	 have possibly not adopted as many
adaptation mechanisms as the dry 	countries	, meaning that crops are more affected, and male
labor less needed in times of drought. 	Detecting the true 	drivers behind this	 pattern	 is 	an
avenue of further research.

6.1.4 Het	erogeneous effects 	on place of residence

From the descriptive statistics	 we observed that there is a large difference in educational
attainment between 	rural and urban areas, with 	seventy	-nine	 percent in urban 	areas	 having
any education, compared to 	fifty	-four	 percent in the rural areas.
Knowing that the type of residence is a	n important predictor of educational attainment	 it is
interesting to study whe	ther there 	exists	 any difference in 	educational 	adaptation to droughts
between the 	types of 	residences. 	Rural residents	 are likely to be 	more affected by droughts,
due to a higher dependence on agriculture 	(Thurlow, 	et al.	, 2012)	. However, 	Thurlow	 et al.
showed that also residen	ts in urban area	s can fee	l the impact 	of droughts 	through 	both 	a rise
in food prices	 and because many rely on	 small	-scale plots of land for subsistence	 farming	.

Table 6.4 show	s the 	analysis from section 6.	1.3 with interaction terms for urban residence
instead of for male	.

Table 	6.4: Heterogeneous effects on place of residence
Years of education attained	 	(1) 	(2) 	(3)
Extreme Drought	 	 	0.0216	 	0.0566
 	 	(0.0719)	 	(0.0685)
Urban x Extreme Drought	 	 	-0.0321	 	-0.0373
 	 	(0.145)	 	(0.135)
Moderate Drought	 	-0.210	* 	 	-0.235	**
 	(0.0878)	 	 	(0.0838)
Urban x Moderate Drought	 	0.00868	 	 	0.0125
 	(0.272)	 	 	(0.262)
Female	 	1.421	*** 	-1.421	*** 	-1.420	***
 	(0.0934)	 	(0.0932)	 	(0.0935)
Urban	 	1.611	*** 	1.617	*** 	1.608	***
 	(0.345)	 	(0.175)	 	(0.345)
Constant	 	-0.213	 	-0.349	 	-0.214
 	(0.461)	 	(0.433)	 	(0.458)
N 	228872	 	228872	 	228872
R2 	0.238	 	0.238	 	0.238
Note	: Standard errors are clustered at the ADM1 level, and 	heteroscedasticity robust. They are given in parentheses for all specifications	. All specifications 	include controls for gender, age, wealth, agricultural land, type of residence and period in which they were interviewed. All	 specifications have fixed effects 	grid	 	level and clustered at ADM1.	 	Significance levels: * p<0.05, ** p<0.01,	 *** p<0.001

From table 6.4 we note t	hat the interaction terms between urban residence and drought are not
significant, when controlling for the 	effect of residence on educational attainment. 	From
specification 3 we cannot conclude that 	the sum of the two interaction terms is non	-zero.
Considering the signs of the 	coefficients	 of the interaction terms, it is p	ossible	 that a moderate
drought has slightly less adverse effects for 	residents living in urban areas	 than for rural
residents	, while an extreme drought 	has slightly more adverse effects	. The 	coefficients	 are
very low, and the robust standard errors 	relatively hi	gh, indicating that we cannot 	detect a
causal relationship, nor be confident that the sign	s of the coefficients 	represent an indication
of relationship.
The results show that there are no apparent differences in how residents of urban and rural
areas adap	t educational patterns in the presence of a drought. 	This can indicate that the
economic shoc	k induced	 by lower crops affect 	urban residents most likely both directly

through lower 	yields	 from	 subsistence	 farming	, as well as	 indirectly 	through	 higher food
prices.

6.1.	4 Adaptability over time

Employing	 time fixed effects 	on	 the 	decade in which individual were born mean we are 	only
studying variation between 	those born in the same decade. This 	is beneficial as it 	eliminates
much of the 	omitted variable bias arising from individuals from different decades 	being
influenced by different 	trends.
Consequently, the approach does	 not allow to study trends across time	. Therefore, 	in 	the
following analysis 	specification number 5 from 	table	 6.1 is applied to individuals born 	in
different 	decades separately. 	The table	 that summarizes the findings from this is found in
appendix 10.1	 table 	A2.
The results indicate that 	experiencing drought 	had a negative effect on years of attained
education for those born in the 1940s and 	in 	the 1970s. For the remaining decades the
coefficients	 on the drought variables are 	not significantly different from zero. 	 However, the
magnitude	s of the coefficients 	indicates	 that the effect of drought on attained education 	seem
to 	have been 	larger for those born 	before and during the 1980s, compared to 	the 	individuals
born in the 1990s. 	This 	indicate	s that there has been changes in adaptation to droughts across
the decades, with 	experiences of 	droughts having a less negative effect on education	 in recent
decades 	compared to the earlier 	decades	.
Figure 4.5 showed the development in percentage of individ	uals 	who	 had experienced
drought. From this we observed two 	local 	maximums	 in the curve of extreme drought, one for
those born in 	the 	1940	s and one for those born in	 the	 1970	s. Noting how this coincides with
the decades in which the effect of 	drought is significant and negative in predicting years of
attained education	 we can infer two different 	possible 	pathways in which drought experiences
influence educational patterns.
For those born in the 1940s, around 35 percent 	experienced extreme dro	ught 	while eligible to
attend school. 	Compared to 	the level for those born in the 1930s, this is an increase of almost
50 percent. 	The strong effect of drought for those born in the 1940s	 can in the light of this, be
explained as 	a consequence of the	 large	 relative	 change from the previous level	. Adaptation
mechanisms 	to drought	 experiences might have been inadequate in accommodating the large

change, and 	consequently	 the 	educational patte	rns were affected. 	 The lack of significant
results of drought for those born in 	the 1960s can thus be due to 	a lower relative change from
the previous decade	 and 	increased	 adaptability	.
The peak	 in 	percent of drought experiences	 for those born in 	the 	1970	s show 	that 	60 percent
of individuals experienced extreme drought while they were eligible to attend school. 	This
high percentage	 implies that drought was widespread, a	ffecting many individuals 	at the same
time	. The data also con	firms this	, as extreme drought was experienced in all	 ten	 countries for
those born in the 1970s	. Almost 30 percent of these experiences were recorded in Ethiopia.
Observing that e	xperiences of drought for those born in the 1970s 	induce the largest
educational adaptation, it is possible to 	infer that widespread 	drought has	 a larger impact than
drought which affects less individuals	 simultaneously	. W	idespread drought affect	ing	 larger
areas, causing 	loss of crops and higher food prices across large areas, a	ffect household
resources negatively, possibly making families prioritize who to send to school in the short
term.
The lack of significant results of drought experiences in th	e decades	 following 1970	 can be
due to 	several mechanisms. Among these are 	an increase in access to education, and possibly
increased adaptation mechanisms	 in response to the many drought experiences for those born
in the 1970s.
Since 	studying each decade	 in isolation produced interesting results, an analysis is also
applied to consider how heterogeneous effects on genders in response to drought has changes
across the decades. 	This is done 	by introducing gender interaction terms to 	samples of 	those
born in the 1950s and those born in the 19	90	s.
Ta	ble 	6.5 shows the results from th	e study of heterogeneous effect	 between the genders in
the	se two decades	.

Table 	6.5: Heterogeneous effects on 	gender across time
 	Panel A	 	1950	-1960	 	 	Panel B	 	1990	-2000	 	 	Years of education attained	 	(1) 	(2) 	(3) 	(1) 	(2) 	(3) 	 	Education	 	Education	 	Education	 	Education	 	Education	 	Education
Extreme	 Drought	 	 	-0.184	 	-0.142	 	 	0.272	 	0.275	 	 	 	(0.113)	 	(0.115)	 	 	(0.142)	 	(0.143)	 	 	 	 	 	 	 	 	Male x Extreme Drought	 	 	0.559	** 	0.419	* 	 	-0.531	** 	-0.542	** 	 	 	(0.182)	 	(0.170)	 	 	(0.174)	 	(0.186)	 	 	 	 	 	 	 	 	Moderate	 Drought	 	-0.184	* 	 	-0.129	 	0.0708	 	 	-0.0146	 	 	(0.0917)	 	 	(0.0907)	 	(0.135)	 	 	(0.134)	 	 	 	 	 	 	 	 	Male x Moderate Drought	 	0.588	** 	 	0.417	* 	-0.151	 	 	0.0556	 	 	(0.198)	 	 	(0.182)	 	(0.187)	 	 	(0.192)	 	 	 	 	 	 	 	 	Female	 	-0.956	*** 	-1.239	*** 	-0.957	*** 	-0.733	*** 	-0.781	*** 	-0.737	*** 	 	(0.201)	 	(0.154)	 	(0.200)	 	(0.200)	 	(0.168)	 	(0.201)	 	 	 	 	 	 	 	 	Urban	 	1.743	*** 	1.741	*** 	1.743	*** 	1.278	*** 	1.283	*** 	1.283	*** 	 	(0.369)	 	(0.370)	 	(0.369)	 	(0.130)	 	(0.130)	 	(0.130)	 	 	 	 	 	 	 	 	Constant	 	1.566	*** 	1.814	*** 	1.589	*** 	0.903	*** 	0.889	*** 	0.848	*** 	 	(0.377)	 	(0.402)	 	(0.388)	 	(0.242)	 	(0.220)	 	(0.246)
N 	16489	 	16489	 	16489	 	41970	 	41970	 	41970	 	R2 	0.173	 	0.173	 	0.174	 	0.163	 	0.164	 	0.164
F – statistic 	 	 	F (1, 151) = 	11.17	 	 	F (1, 137) = 	5.43	 	P-value	 	 	 	0.001**	 	 	 	0.0212*
 Note: 	Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all specifi	cations	. All specifications 	includ	e controls for gender, age, wealth, agricultural land, type of residence and period in which they were interviewed. 	All specifications have fixed effects 	grid	 	level 	and clustered at ADM1.	 	The 	F-test is testing the difference 	between the effect on males and females	, by testing significance of the sum of the interaction terms on male. 	The degrees of 	freedom are given in parenthesis, and the p	-value in the following row. 	 	Significance levels: * p<	0.05, ** p<0.01, *** p<0.001

From table 6.	5 we observe 	that 	the 	coefficients	 indicat	e the heterogeneous effect on the
genders is reversed 	when comparing	 those born in the	 1950	s to 1990	s. From panel A we
observe tha	t for those born in the 1950s 	and experience a 	moderate 	drought, 	males attend
almost half a year more 	education	, relative to the females	 in this sample	. This difference in
educational attainment is increased 	further	 if 	an 	extreme drought	 is experienced	.
From panel B,	 those born between 1990 and 2000	, we observe a different pattern. 	We observe
that 	in the presence of drought 	males attain 	about a half of a year less education	, relative to
the females in the sample	.
Th	e observed	 pattern	 of heterogeneity of e	ffect between genders and decades studied	 can 	have
multiple 	explanations	, and knowing the exact causes requires more in	-depth study. However,
it can indicate that 	the value of 	male child	-labor in agriculture decreased	 during 	periods 	of
drought	 in the first sample	. The results from the second sample can indicate that the increased
focus on 	access to education for females have 	been influential.

From table 6.	5 we	 also note 	two other trends. Being 	female 	had a more negative impact on
years of attained education 	for those born in	 1950 than 	in 	1990. 	The negative effect has been
reduced by about 	a fifth of a	 year. This indicates a positive development	 in gender equality	 in
educational opportunities	, but also 	shows that there are still 	improvements to be made	. We
also note that the 	negative 	impact of living in a rural area on years of education has decreased.
This can indicate that 	infrastructure 	is improving and access to 	education 	is increasing.

6.2 	Effect of drought 	on	 different levels of 	education

Investigating the effect of drought on years of attained education is beneficial, as it allows to
study the possibility that individuals are kept out of 	education	 for limited amount of time,
before retur	ning.
However, as previous studies have indicated children of different ages 	are	 affected differently
it is interesting to consider whether the experience of drought has any implication for the
probability of attaining a certain level of education	 (Bjorkman	-Nyqvist, 2013)	. From section
4.6. we know that the level of educational attainment has been changing over time at the
aggregated level in our sample.
Investigating th	e relation between drought and 	educational level	 is done by redefining the five
levels of education recorded by the DHS 	int	o three binary variables	. These binary variables
capture whether individuals have any education, whether they have completed primary
education 	and whether they have attended secondary or higher education.
Table 	6.6: Percentage of sample according to levels of education
 	Some education	 	Completed primary 	 	Secondary 	or higher
No	 	36.19	 	 	67.33	 	 	77.67
Yes	 	63.81	 	 	32.67	 	 	22.33
Total	 	100.00	 	 	100.00	 	 	100.00

From table 6.	6 we observe that 	sixty	-four	 percent of the sample has some education, while
only 	twenty	-two	 percent has secondary or higher education. 	The three binary variables are
used as the dependent variable	 to assess the impact of drought on the probability of attending
the different levels o	f education.

Table 	6.7 shows the estimation results from 	the first set of regressions, where 	having 	any
education is used as the binary variable	, leaving only those with no education in the control
group.
Table 	6.7: Any education and drought
Any Education	 	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate Drought	 	-0.0104	 	-0.00740	 	-0.0308	*** 	-0.0292	*** 	-0.0281	***
 	(0.0150)	 	(0.0145)	 	(0.00709)	 	(0.00728)	 	(0.00738)	 	 	 	 	 	 	 	Extreme Drought	 	0.000816	 	0.00106	 	0.00193	 	0.00208	 	0.00244
 	(0.0185)	 	(0.0173)	 	(0.00616)	 	(0.00594)	 	(0.00564)	 	 	 	 	 	 	 	Female	 	 	-0.160	*** 	-0.160	*** 	-0.169	*** 	-0.168	*** 	 	 	(0.0117)	 	(0.0111)	 	(0.0115)	 	(0.0122)	 	 	 	 	 	 	 	Urban	 	 	0.211	*** 	0.191	*** 	0.194	*** 	0.121	*** 	 	 	(0.0206)	 	(0.0203)	 	(0.0196)	 	(0.0144)	 	N 	243841	 	243841	 	243841	 	233811	 	228993	 	R2 	0.000	 	0.063	 	0.145	 	0.158	 	0.182
F (1, 188)	 	0.53	 	0.19	 	12.79	 	10.79	 	9.13	 	P-value	 	0.4692	 	0.6660	 	0.0004	*** 	0.0012	** 	0.0029	**
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed
No No No No No No
No No No No No No
No No No Yes	 No No
Yes	 No No Yes	 Yes	 No
Yes	 Yes	 Yes	 Yes	 Yes	 Yes
Note: 	Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are 	given in parentheses for all specifications	. All specifications have 	fixed effects on grid cell level. 	 	The 	F-test is testing the significance of the sum of the two drought dummy variables	. Degrees of freedom are given in parenthesis, and the p	-value in th	e 	following row. 	For specification 5 the F	-statistic has degrees of freedom F (1, 157).	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

Table 6.	7 show	s that 	experienc	ing	 of at least a moderate drought decreases the probability of
having any education by roughly three percent	 in specification 3	-5. The F	-test indicates that
experiencing a	n extreme drought, relative to a situation of no drought reduces the probability
of hav	ing any education by almost 	three	 percent.

The analysis is also performed with interaction terms for male, to investigate whether there
exist heterogeneous effects also here. 	The results from this analysis indicate that the negative
effect on the probability of having any education is partly driven by a more adverse effect for
females 	relative to that of	 males. 	The table 	with these results 	can be found in 	appendix 10.1
table 	A3	.

Experiences of drought has no significant effect on the probability of completing primary
education	 or secondary and higher education. 	No heterogeneous effects on gender are
observed for 	the	 impact of drought on the probability of these	 two	 levels of 	education	. The
table 	with results for	 these two levels of education is found in 	appendix 10.1	 table A4.

The results from this section can indicate that 	either drought 	experiences hinder	 individuals
from ever attending school, by 	reducing the probability 	by 	three	 percent	, or it 	only	 keeps
individuals	 out of school for a limited amount of time, before returning. Th	is reasoning can
explain	 why 	there is no 	observed 	effect 	of drought 	on the 	probability of attending the 	higher
levels of education.	 The result 	from this section 	justifies why the main approach in this thesis
is to investigate how droughts affect the 	years of education attained	, rather than level attained	.

To be better suited to evaluate how 	droughts affect the level of attained education 	data on
when during their time as eligible to attend school individuals experienced droughts should be
used.

6.3 Effect of drought on literacy

Investigating the effect of droughts on 	educational level and 	years of education obtained is
important but	 might undermine the importance 	of 	the 	quality 	of 	education	 for 	long	-term
development	 (Hanushek, 2013	; Nordstrom & Cotton, 2020)	. Therefore	, investigating	 the
effect of droughts on literacy can give a	n indication of 	the possible impacts 	on th	e quality of
education	 in the presence of droughts.
The DHS defines literacy as individuals who attended 	secondary	 level schooling or who can
read a whole sentence or parts of a sentence	 (Demographic and Health Survey Program,
2021)	. As 	the aim is	 to investigate the relationship between schooling, literacy and drought	,
the	 literacy 	variable 	is 	for the purpose 	defined solely based on whether an individual can read
parts of a sentence	.9
 	9 Including also the part of DHS´ definition of liter	acy of having attended secondary schooling in construction of the variable would mean we could not study the
effect of secondary schooling on literacy, 	as literacy would be in both the dependent variable and the explanatory variable, introducing mechanical	 correlation	.

In this part of the 	analysis,	 surveys conducted 	prior to year 2000 	are excluded 	as 	the	se do not
ask the literacy question in the same way as the more recent surveys	 do	. The 	resulting d	ataset
is used 	to define 	a binary variable 	of whether an individual can read or not.	 This yields a
data	set 	of 2	16	 000 individuals, 	where	 fifty	-five	 percent	 of respondents 	are defined as literate	.
A simple 	correlation analysis	 show	s that there is a significant, negative correlation between
having experienced drought in the years between 	five	 and 	fifteen	 years of age and 	being
literate, corroborating the findings of	 Nordstrom and Cotton who found that drought can yield
less learning	 in the presence of drought 	(Nordstrom & Cotton, 2020)	.
To investigate the relationship between drought prevalence and literacy a linear probability
model is employed	. The equation used to study the effect of drought on 	literacy is the
following:
P(Literacy	it=	1)=	βo+ β1Drought	+ βiXi……+ αi+ λt+ui 	 	(6.	1)
As previously, 	Xi are the control variables, indicated in the specifications, and α	i and λ	t are the
entity and time fixed intercepts. 	Table	 6.8 shows the results 	from	 this estimation.
Table 	6.8: Literacy and drought
Literacy 	 	(1) 	(2) 	(3) 	(4)
Moderate Drought	 	-0.0257	* 	-0.0249	** 	-0.0208	** 	-0.0204	** 	 	(0.0128)	 	(0.00757)	 	(0.00641)	 	(0.00630)	 	 	 	 	 	 	Extreme Drought	 	0.00477	 	0.00669	 	0.0153	** 	0.0148	** 	 	(0.0182)	 	(0.00906)	 	(0.00525)	 	(0.00496)	 	 	 	 	 	 	Female	 	-0.205	*** 	-0.224	*** 	-0.221	*** 	-0.218	*** 	 	(0.0138)	 	(0.0149)	 	(0.0142)	 	(0.0140)	 	 	 	 	 	 	Urban	 	0.229	*** 	0.116	*** 	0.112	*** 	0.113	*** 	 	(0.0207)	 	(0.0161)	 	(0.0149)	 	(0.0148)	 	 	 	 	 	 	Constant	 	0.647	*** 	0.680	*** 	0.307	*** 	0.433	*** 	 	(0.0129)	 	(0.0221)	 	(0.0571)	 	(0.0578)	 	N 	215884	 	213486	 	204812	 	204812	 	R2 	0.073	 	0.151	 	0.171	 	0.172
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed
No No No No No No

Yes	 Yes	 Yes	 No Yes	 No

Yes	 Yes	 Yes	 Yes	 Yes	 No

Yes	 Yes	 Yes	 Yes	 Yes	 Yes

Note: 	Standard errors are clustered at the ADM1 level, and 	heteroscedasticity robust. They are given in parentheses for all specifications	. All specifications have 	fixed effects on grid cell level. Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

From table 6.	8 we note that experiencing a drought lower	s the probability of being literate by
around 	two	 percent. This can be due to lower quality of education during times of drought,
but it can also be a direct result of 	individuals attending less 	education	, as observed in section
6.1. 	To investigate these link	ages an analysis of the 	role of education on t	he probability of
literacy is 	performed.
Education is 	one of 	the most important 	tools 	to achieve literacy	. Controlling for attained
education in 	table 	6.8 	would however not be desired	 as we have already observed a direct
relationship between drought and education, making 	education an outcome of the treatment
variable	. Therefore, 	the link from attained education to literacy is studied in isolation in 	table
6.9.
The e	quation	 used 	to stud	y the 	effect 	of 	education	 on literacy is the following:

P(Literacy	������������=	1)=	βo+ β1EducationalLevel	������+ βiXi……+ αi+ λt+ui 	(6.2)
Table 6.	9 shows the results of this estimation. 	All f	our	 specifications	 in table 6.	9 include
control variables for gender, residence, 	religion, wealth, agricultural land, age	, and survey
period. They all 	have fixed effects on the constructed	 grid 	cell 	level, and time	 fixed effect	 on
decade born.
Table 	6.9: The effect of education on literacy
Literacy	 	(1) 	(2) 	(3) 	(4)
Started Education	 	0.672	*** 	 	 	0.605	*** 	 	(0.00952)	 	 	 	(0.0116)	 	 	 	 	 	 	Completed Primary	 	 	0.375	*** 	 	0.0933	** 	 	 	(0.0186)	 	 	(0.0283)	 	 	 	 	 	 	Secondary Higher	 	 	 	0.344	*** 	0.135	*** 	 	 	 	(0.0218)	 	(0.0265)	 	 	 	 	 	 	Female	 	-0.103	*** 	-0.183	*** 	-0.193	*** 	-0.0954	*** 	 	(0.00702)	 	(0.0125)	 	(0.0128)	 	(0.00736)	 	 	 	 	 	 	Urban	 	0.0407	*** 	0.0723	*** 	0.0742	*** 	0.0229	*** 	 	(0.00614)	 	(0.0103)	 	(0.0109)	 	(0.00516)	 	 	 	 	 	 	Constant	 	0.167	*** 	0.275	*** 	0.300	*** 	0.106	** 	 	(0.0416)	 	(0.0597)	 	(0.0617)	 	(0.0395)	 	N 	204812	 	204812	 	204812	 	204812	 	R2 	0.474	 	0.288	 	0.254	 	0.505	 	Note: 	Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all 	specifications	. All specifications have 	fixed effects on grid cell level 	and decade born. The control variables included are gender, residence, religion, wealth, agricultural land, age, and survey p	eriod	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001
In table 6.	9 we observe how the different levels of education 	all 	contribute to a higher
probability of being literate. 	By exploiting the way the three binary variables of education are

constructed, yielding an increase in the control 	group for each level	 of education	, specification
1-3 s	how the effect of the binary variables 	separately.
In specification number 4 all three educational levels are included. Th	e estimated 	coefficients
from this specification 	allows us to observe the margi	nal gain in terms of probability of
literacy of reaching a higher educational level.
We note that the marginal effect o	n the probability of literacy 	of 	having started 	any 	education
is much higher than the marginal effect of attaining the higher levels of education on the
probability of being literate. 	This indicate that	 some individuals do learn to read without
formal education, and for those attending 	education	, most learn	 to read during the first years
of education	.
From 	regression table 6.	8 and 6.	9 we have observed that experiencing a drought 	reduces the
probability of being literate of 	around 	two percent	, and that 	having started a	ny education is
the greatest contributor	 to 	increasing the probability of 	literacy.
Running a	n analysis 	on 	the 	relationship between drought and literacy	, while controlling for
educational attainment would 	give an indication of 	how well the 	educational	 institutions
perform 	during	 drought events. 	The results from this 	are	 found	 in table	 A5 in 	the 	appendix
10.1	. The	se results indicate that 	when controlling for educational attainment, 	going from a
situation of moderate drought to a situation of	 extreme drought has a positive and highly
significant effect on the probability of being literate of around 	one	 percent. 	The coefficient of
the moderate drought variable is negative, but not significant.
This 	can indicate that 	the negative impact of less	 education is induced 	when a 	moderate
drought	 occurs	, but if the drought becomes even worse 	communities 	could have	 had time to
adapt	 to new conditions	, and the added effect 	on learning outcomes 	can be 	slightly 	positive.
Results from this section imply that increasing the likelihood of individuals attending school
is important, as this leads to a higher probability of being literate. This is an important notion,
considering that results from regression table 6.2 indicated 	that a drought could decrease the
likelihood of having any education by 	three	 percent. Preventing the adverse effects of
droughts, by adopting measures for better adaptability, could therefore increase school
attendance, and thereby increase literacy. This	 could be an important part of reaching a higher
long	-term growth rate for developing countries.

There are however several shortcomings of the linear probability model. The fitted prediction
line of OLS can predict less than zero and above one as the chan	ge in probability of
occurrence, when 	this is not feasible	. Also, 	as 	the linear probability model with OLS 	fits a
straight line to a curve that is likely to be 	non	-linear, it usually 	overstates the 	predicted
changes in the dependent 	variable when	 the predicted change is close to zero or one	 (Stock &
Watson, 2015)	. This means that 	the predicted reduction of 	two	 percent in probability of
literacy 	due to drought occurrences 	is likely to be overestimated.

6.4 A	n event study 	of 	the 	Ethiopia	n famine	 1983	-198	6

The analysis conducted 	until this point	 consider whether individuals have experienced
moderate or extreme drought during the years they were eligible to attend school. Some of the
results show a 	significant 	and negative 	effect of drought on education	al attainment	. The
analysis has not 	considere	d in isolation 	the effects of 	long	-term drought occurrences, that can
cause 	famine	s.
In the years from 1984 to 198	6 Ethiopia 	experienced one of the worst famines in recent
history	, affecting nearly	 eight million people, and causing 	up to 	one million fatalities 	(Dercon
& Porter, 2014	; Vestal, 1985)	. The reasons behind famines experienced in Ethiopia are
several, ranging from poorly led governments, to 	conflicts	 and 	severe and long	-lasting
drough	ts (Dercon & Porter, 2014)	. This makes it difficult to consider	 the effect of the 	long	-
lasting 	drought in isolation. Therefore, 	this event study considers the long	-term impacts on
education from a 	severe famine, causing among other effects a 	sharp rise in food prices,
which 	is at least 	partially 	attributable	 to 	a long	-lasting 	drought.
First, 	some of the methods from this chapter are applied to Ethiopia in isolation	, to evaluate
whether 	droughts aff	ect Ethiopia differently than observed using the full sample	. The tables
with results from this 	can be	 found in appendix 10.1	 table A6.
Experiencing moderate drought 	reduces the years of attended education by 	almost 	0.4 	in
Ethiopia, 	depending on 	specification	. This is larger than the overall trend observed in section
6.1.1.	 Experiencing a moderate drought reduces the probability of being literate with 	almost
four	 percent, and the added effect of experiencing an extreme drought is 	incr	easing the
probability of being literate with almost 	two	 percent. 	The probability of having any education
is reduced by 	around 	six	 percent if 	a moderate drought is experienced. 	It seems that 	the effects

of 	experiencing a	 drought 	while eligible to attend school 	are larger in Ethiopia, 	relative to the
results from	 the full sample.
To 	investigate the 	effect	s of 	the 	persistent drought 	in the 1980s in Ethiopia 	the method of
synthetic controls	, developed by Abadie et al. 	is applied	 (Abadie & Gardeazabal, 2003)	. This
method uses information from 	untreated 	countries to generate a control group that has 	as
similar trend to that of Ethiopia prior to the famine	 as possib	le. Applying	 a difference	-in-
difference approach is 	inferior to the method of synthetic controls, as no countries in the
sample have a 	similar trend to that of Ethiopia prior to the famine. A difference	-in-difference
approach is however used to evaluate the validity of results from this 	method	. This analysis
can be found in the appendix	 10.1	 figure	 A11 and 	table 	A12. The 	difference	-in-difference
analysis 	confirms the trend that is found	 using the method of s	ynthetic controls	.
Figure 	6.1: Mean years of education in Ethiopia and the synthetic control group

Note: Mean years of education in Ethiopia and the constructed synthetic control group	, calculated 	as
the mean	 years of 	educational atta	inment of those who started education 	in the same year. 	The figure
is d	eveloped using Stata.
From figure 6.4 we observe th	e trend across time in mean years of education reported in
Ethiopia, relative to that of the generated synthetic control group. All the other nine countries
in the sample 	are used	 in a weighted average	 to generate a synthetic control group that 	nearly
follows the trend of Ethiopia	 prior to 1982	. In order to replicate the trend of Ethiopia prior to
the famine m	ost weight is put on the observations from 	Mali, 	Niger and Burkina Faso in
generation of the synthetic control group	.
From figure 6.4 w	e observe t	hat 	the trend of the synthetic control group and 	the trend of
Ethiopia follows one another relatively closely prior to 1982, while 	for those who started

school after 1982 the trend in 	mean years of education in Ethiopia is 	consistently 	larger than
for the 	synthetic control group. 	There are however several shortcomings of this method	,
making it difficult to determine the true causes of this observed trend.
Most 	notably, we know that several of the other countries have also experienced famines
during the period, making the synthetic control group biased. 	Also, Ethiopia has experienced
several famines across time, making this famine 	not a shock in isolation. 	Theref	ore,	 the
magnitude of the effect	 is not	 described	. However, it is possible to conclude	 from figure 6.4
that 	relative to the 	synthetic control group constructed	, Ethiopia has a larger increase in 	mean
length of education, and that this increase 	started with	 the	 commence of the	 infamous famine
in the 19	82	.
Figure 	6.2: The gap in the predicted error

Note: The ev	olution in the gap 	between the mean years of education in Ethiopia and 	the mean years of
education in the constructed synthetic 	cont	rol group. 	The figure is d	eveloped using Stata.
Figure 6.2 shows the evolution in the 	gap between the mean years of education in Ethiopia
and the 	synthetic control group	, visualizing better the differen	ce in	 trends 	between the
synthetic control and 	Ethiopia as 	observed in figure 6.1. 	The better the fit of the synthetic
control group, the smaller the gap in the prediction should be prior to the 	famine	. This gap is
higher than desired for the years 	from	 1950 to around 1972, indicating that the synthetic
control group does not follow the trend of Ethiopia adequately. 	However, the larger gap in the
prediction 	post famine, 	can 	indicate that 	Ethiopia 	benefited from the famine in terms of
educational outcomes.	 Note	 that 	other possible elements that could influence educational
attainment at country level	, such as investments in 	the educational sector	, infrastructure
improvements, aid inflows, 	and	 conflicts are 	not controlled for	.

Considering th	e shortcomings of this method the findings should be interpreted with caution.
However, 	the results 	from this thesis 	indicate opposite directions of effects of short	-term and
long	-term drought occurrences. 	It is possible that droughts have a negative impact on
households once it commences due to a strain on household resources, but that the long	-term
effec	t is that paren	ts prioritize education as a long	-term investment, as they have observed the
vulnerability in agricultural production. This pattern was observed in a study of long	-term
adaptation mechanisms to drought in Kenya 	(Opiyo, 	et al.	, 2015)	. From figure 2.1 we
obs	erved how 	land used for agriculture in Ethiopia declined sharply during the 1990s	,
possibly motivated by 	long	-lasting drought episodes and 	higher levels of education.
This part of the analysis should be 	extended 	using	 a more appropriate 	control group	, including
important 	control variables	, and	 considering shocks that are more isolated,	 to be able to
conclude on the direction and 	magnitude of effect	 with more certainty	.

6.5 Robustness 	evaluation

This section evaluates the 	performance of the 	empirical strategies by 	testing several of the	ir
underlying 	assumptions.
As mentioned	 in section 5.4	, the Hausman test is run to rule out the use of random effects. All
Hausman tests performed on specification	s reject the hypothesis of random effects being
superior to fixed effects with the constructed dataset.
The issues of heteroscedasticity and autocorrelation are addressed by using clustered standard
errors in all specifications, generating 	heteroscedastic	ity 	robust standard errors.
Restricting the sample to individuals who have lived in 	the 	place of 	current 	residence always
violates the central assumption of random sampling for OLS to yield unbiased 	and consistent
estimates. To estimate the effect of this possible source of bias the main specifications
employed in part 6.1.1 are imposed also on the sample pri	or to this restriction. 	The results
from this	 are found in table 6.	10	.

Table 	6.10	: Comparing restricted and full sample
Years of education	 	Panel A: prior to restriction	 	Panel B: after the 	restriction	 	(1) 	(2) 	(3) 	(4) 	(5) 	(1) 	(2) 	(3) 	(4) 	(5) 	Moderate Drought	 	-0.268	** 	-0.204	** 	-0.175	*** 	-0.182	*** 	-0.151	*** 	-0.124	 	-0.0993	 	-0.252	*** 	-0.249	*** 	-0.232	***
 	(0.0813)	 	(0.0751)	 	(0.0412)	 	(0.0392)	 	(0.0359)	 	(0.130)	 	(0.125)	 	(0.0590)	 	(0.0593)	 	(0.0571)	 	 	 	 	 	 	 	 	 	 	 	 	Extreme Drought	 	-0.224	* 	-0.209	* 	-0.142	** 	-0.109	** 	-0.0819	* 	-0.0265	 	-0.00866	 	0.0327	 	0.0371	 	0.0481
 	(0.0965)	 	(0.0907)	 	(0.0433)	 	(0.0367)	 	(0.0363)	 	(0.177)	 	(0.164)	 	(0.0591)	 	(0.0626)	 	(0.0582)	 	 	 	 	 	 	 	 	 	 	 	 	Female	 	 	-1.271	*** 	-1.438	*** 	-1.385	*** 	-1.315	*** 	 	-1.443	*** 	-1.440	*** 	-1.451	*** 	-1.420	*** 	 	 	(0.0695)	 	(0.0705)	 	(0.0689)	 	(0.0629)	 	 	(0.0884)	 	(0.0880)	 	(0.0888)	 	(0.0933)	 	 	 	 	 	 	 	 	 	 	 	 	Urban	 	 	3.180	*** 	3.100	*** 	3.106	*** 	1.349	*** 	 	2.848	*** 	2.677	*** 	2.687	*** 	1.604	*** 	 	 	(0.130)	 	(0.126)	 	(0.127)	 	(0.0967)	 	 	(0.211)	 	(0.206)	 	(0.191)	 	(0.163)	 	 	 	 	 	 	 	 	 	 	 	 	Constant	 	5.445	*** 	5.302	*** 	1.200	*** 	-2.988	*** 	-4.743	*** 	4.600	*** 	4.839	*** 	1.941	*** 	1.199	** 	-0.214	 	 	(0.0559)	 	(0.0707)	 	(0.258)	 	(0.617)	 	(0.766)	 	(0.0909)	 	(0.111)	 	(0.201)	 	(0.374)	 	(0.444)	 	N 	591907	 	591907	 	591907	 	591907	 	584181	 	243718	 	243718	 	243718	 	233690	 	228872	 	R2 	0.001	 	0.107	 	0.157	 	0.162	 	0.244	 	0.000	 	0.091	 	0.162	 	0.180	 	0.238	 	Note: 	Standard errors	 are clustered at the ADM1 level, and 	heteroscedasticity robust. They are 	given	 in parentheses for all specifications	. All specifications have 	fixed effects on grid cell level and decade born. The control variables included are gender, residence, religion, wealth, agr	icultural land, age, and s	urvey period	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

From table 6.	10	, panel A	 we observe 	a larger	 estimated	 coefficient of extreme drought
compared to the 	effect observed in the 	restricted sample	, presented in panel B	. In 	specification
5, panel A, the sum of the effect on the two drought dumm	y variables	 almost equal the
coefficient of moderate drought in specification 5, panel B.
This indicates that by restricting the sample to those who have not moved since childhood we
are not introducing large bias on the overall effect of drought on educational outcomes. This
can be due to individuals reporting having moved residence 	while	 having moved short enough
to still be within an area that was or was not affected by drought. In t	he restricted sample we
seem to be detecting a slightly lower overall effect than the one detected in the full sample
which strengthens the confidence in the results	 from this thesis	.
Analysis	 is run on 	subsets of countries	 in section 6.1.2	, without large 	differences in the overall
adaptability of education to drought being detected. This also strengthens the confidence in
the validity of estimates.
Clustering on country level rather than ADM1 level is 	done as an 	alternativ	e. The 	results from
this are	 found in appendix 10.1	 table A7	. Clustering on country rather than ADM1 level
yields	 the same estimated 	coefficients	 of the	 three	 specifications which yield significant
results	 in section 6.1.1	. However, t	he heteroscedasticity robust standard errors are 	marginally
smaller in 	specification number 1, 3 and 4, indicating 	marginally 	more precise results for
these 	three 	specifications	 when clustering on country level	. Despite this,	 clustering on ADM1
level is 	preferred	 as ADM1 level is believed to be better suited to capture the	 geographical
spread	 of drought than countries are	, especially 	for	 the larger countries	 in the sample	.

Appendix 	10.	2 also shows a robustness check of the 	linear probability model by comparing it
to the use of ordered logistic and ordered probit models	. This can be	 found	 in 	table 	A10	.

7 	Weaknesses 	and 	possible 	extensions

This thesis has only used one drought measure, the SPEI, 	to evaluate the effect of droughts on
educational attainment. Alternative measures, such as the SPI or the SPEI global drought
monitor could also have been used, despite being calculated slightly dif	ferently, as stated in
section 	4.2.2. To extend on th	e results from this thesis different 	drought 	measures	 could 	be
included.
With the use of a more frequent drought measure it would be possible to detect whether there
exist heterogeneous effects of droughts depending on the time during the year when 	a drought
commences. This could be used to predict the importance of alternative	 value of labor in
educational choices, as it would be possible to detect whether droughts have larger effect on
educational outcomes in seasons where child labor is valued differently to other seasons	, for
instance during the main harvest season	.
Two thr	eshold values of drought were generated based on the available descriptions of
consequences for crops and livelihoods of the different 	SPEI 	values. Other threshold values
could be generated to investigate 	how this	 choice	 affects 	results.	 Also,	 other types 	of extreme
weather events, such as floods could be in	cluded and investigated	 to supplement the results
presented 	in this thesis	.
As 	described in section 4 and 6, restrictions had to be 	im	posed on the sample 	to yield 	an
appropriate	 dataset for analysis. Omitting all individuals who have moved residency since
childhood is in violation of one of the central 	assumptions underlying the empirical method
applied.	 Those who choose to not move might differ in other aspects from the individu	als who
choose to move. Furthermore, 	living in a place with devastating droughts over time could
induce forced migration, leaving out the individuals who have been 	most 	exposed to droughts
in the sample. 	Although section 6.5 demonstrated that this likely h	ad little implication for the
direction and magnitude of 	the results	, a possible	 extension	 is to 	use more randomized data to
rep	licate	 the 	analysis.
Another source of potential bias	 in the analysis	 is the geographical displacement procedure
used by the DHS to ensure confidentiality of individuals´ responses. 	The DHS regards
confidentiality	 highly, and t	o be granted access to the geocoded datasets, a separate
application must be submitted. These data	sets provide a longitude and latitude coordinate
associated with each interviewed cluster. 	These coordinates are geographically 	displaced by

up to 	two	 kilometers in urban areas and up to 	five	 kilometers in 	rural	 areas, with a further 	one
percent of 	rural	 clusters being displaced even further, and up to	 ten	 kilometers from its true
location	 (Burgert, Colston, Roy, & Blake, 2013)	.
Using ArcGIS clusters were placed within the 	0.5	x0.5	-degree grid cell	s and based on this
procedure 	individuals have either experienced 	drought,	 or they have not	 during childhood	.
Considering that the DHS clusters are somewhat displaced it is possible that some of the
clusters identified within one grid	 cell	, truly pertain to	 a neighboring grid. Further study could
consider	 the possibility of omitting the clusters which fall within a proximity to the border of
a grid by 	two	 and 	five	 kilometers, respectively in rural and urban areas to ensure 	that this does
not influence result	s. However, as droughts usually affect larger areas 	simultaneously	 the
implications for the validity of results 	are	 assumed to be small.
The findings of this thesis do not necessarily 	extend to explain how extreme weather events
will affect educational att	ainment in the future in th	is region, nor in other regions. 	Firstly, 	the
frequency, spatial coverage and severity of droughts seem to be changing	 around the world
(Masson	-Delmotte, et al., 2021)	. This	, together with 	rapidly changing technologies,	 might
induce 	different 	adaptation mechanisms	 to such events	. Secondly, as observed 	by considering
the most recent subsample of individuals, 	the 	adverse 	effect	s of 	droughts 	on educational
attainment 	seem	 to be 	subsiding	.
Sub-Saharan Africa has 	different characteristics from the rest of the world, both in terms of
population, weather, availability of technology 	and dependence on agricultural production
(The World Bank, 2021)	. Therefore, the results uncovered in this thesis 	will most likely not
be replicated in other regions of the world. 	Another	 possible 	extension	 is to use 	DHS
household data 	to investigate whether 	other effects of 	drought	 on educational outcomes	 exist
in Lat	in-America or Asia, where the DHS also conducts surveys.

8 	Discussion and c	onclu	sion

This thesis has used geocoded data to investigate linkages between drought episodes and
educational attainment. Education is one of the most powerful tools to combat poverty, reach
the SDGs, and enable higher long	-term economic growth. Climate change, brin	ging with it
more extreme weather poses a critical threat to agricultural dependent populations in sub	-
Saharan Africa, possibly lowering their prospects for income generating activities. This thesis
has investigated the possible implications this has for e	ducational attainment. The findings
from this thesis contributes to the somewhat limited existing literature on the topic of extreme
weather events and education.
By combining data from the DHS with climatic data from the SPEI, significant results were
fo	und, indicating that episodes of drought during childhood decreases the years of attained
education with about a quarter of a year for the full sample. Applying the same analysis to
subsamples of countries based on differing climatic conditions show that t	he effect is
replicated in more humid 	countries	, while slightly more adverse in 	drier	 countries	. These
results are largely explained by the construction of the SPEI as deviations from long	-term
means. 	This means that 	according to 	values	 recorded by the SPE	I a	 drought in an already dry
region is therefore more severe, in absolute terms, than the same drought detected in a humid
region.
Experiencing a drought decreases the probability of having any education by 	three	 percent.
Both this effect, and the negative, significant effect of droughts on years of attained education
are strongly heterogeneous for the genders, implying that females w	here most adversely
affected by droughts. 	However, w	hen only considering a more recent sample, this effect
seems to have been reversed.	 As r	esults	 show that the probability of being literate increases
by 	sixty	 percent if the individual has attended any edu	cation, 	it indicat	es that individuals 	do
learn something in school, 	and	 highlights the importance of 	ensuring 	universal	 access to
education.
The broader question of this thesis considers how households adapt to exogeneous shocks to
household 	income. The heterogeneous effects on gender can be explained in the context of the
substitution and income effects. When deciding on school enrollment for their children,
parents trade off the value the children can produce in income generating activities,	 with the
expected future value generated by education. A drought usually yields less crops to cultivate,

and therefore lower household income. This can lead to prioritizations of whom to send to
school, and consequently less education.
On the other hand, less crops to cultivate 	imply that	 the alternative value of child labor in
agriculture is reduced, meaning families have less to lose by sending their children to school.
It is possible to imagine that these two effects, along with sever	al others, are at play when
families decided on school attendance for their children.
Observing how females attained less education in the presence of drought, while males
attained more it is reasonable to conclude that the size of the two effects differe	d between the
genders. At least in earlier 	decades	 it seems households prioritized ensuring education for
males in the presence of household income shocks. Furthermore, the relative value of male
labor in agriculture relative to education shrinks when ther	e are less crops to cultivate. The
same cannot be claimed for the value of female labor if primarily used for home production.
These 	results	 are in line with the findings on heterogeneous effects 	of weather shocks on
educational outcomes 	between the gender	s found in both Ethiopia and Uganda	 (Bjorkman	-
Nyqvist, 2013	; Mani, 	et al.	, 2013	). A possible avenue for further research is studying these
mechanisms more in depth.
Results from this thesis also show that the experience of drought can decrease the likelihood
of literacy by 	two	 percent, indicating that learning outcomes are lower during times of stress
on household income. There exist several possible explanations behi	nd this effect. A recent
study of the effects of the COVID	-19 pandemic on rural household in Uganda conclude	s that
households lowered their expenditures on food consumption by as much as 	forty	-four	 percent
and exhausted their savings in response to the 	imp	osed 	lockdown	 (Mahmus & Riley, 2020)	.
These findings indicate that many households in rural Africa have scarce resources, and
therefore changes to their disposable income have large implications. Mitigating the adverse
effects 	of extreme weather on crops is therefore important to ensure adequate nutrition and
resources to prioritize education. The adverse effects of drought	s on crops can at least
partially explain why results show that individuals experiencing droughts have a lo	wer
probability of having any education and being literate.
As it is o	bserv	ed	 that	 adaptations to drought seems to be stronger during earlier decades, than
for the more recent 	decades	, it can indicate that the increased focus on equal and adequate
access to education has been influential. Briefly studying the effects of a severe famine in
Ethiopia seem to indicate that persistent drought episodes can increase educational attainment,

possibly through increasing the expected payoff from education, rel	ative to that of a career	 in
a vulnerable agricultural	 sector	. Therefore, the long	-term effects of persistent droughts can
have positive implications for	 the level of	 investment in education	. This is	 contra	ry	 to the
findings 	from earlier decades 	on	 negativ	e short	-term effects 	of 	experiencing 	drought 	on 	the
years of 	attained 	education	.

9 	Bibliography

Abadie, A., & Gardeazabal, J. (2003). The Economic Costs of Conflict: A Case Study of the Basque
Country. 	The American Economic Review	, Volum 93(1), p 113	-131.
Abel, G., Brottrager, M., Cuaresma, J., & Muttarak, R. (2019). Climate, conflict and forced migr	ation.
Global Environmental Change	.
Agamile, P., Dimova, R., & Golan, J. (2021). Crop Choice, Drought and Gender: New Insights from
Smallholders’ Response to Weather Shocks in Rural Uganda. 	Journal of Agricultural
Economics	, Volum 72(3), p 829	-856.
Alesina	, A., Hohmann, S., Michalopoulos, S., & Papaioannou, E. (2021). Intergenerational mobility in
Africa. 	Econometrica	, Volum 89(1), p 1	-35.
ArcGIS Pro. (2021, September 20). 	Spatial Join (Analysis)	. Retrieved from esri:
https://pro.arcgis.com/en/pro	-app/latest/tool	-reference/analysis/spatial	-join.htm
Barrios, S., Bazoumana, O., & Strobl, E. (2008). 	The impact of climatic change on agricultural
production: Is it different for Africa? 	Food policy	, Volum 33(4), p. 287	-298.
Beguería, S., Latorre, B., Reig, F., & Vicente	-Serrano, S. M. (2021, May 19). 	Global SPEI database 	.
Retrieved from Global SPEI database : https://spei.csic.es/database.html
Begueria, S., Vicente	-Serrano, S. M., & Angulo	-Martinez, M. (2010). 	A mulitscalar g	lobal drought
dataset: The SPEIbase: a new gridded product for the analysis of drought variability and
drought impacts. 	American Meteorological Society	, Volum 91(10), p.1351	-1354.
Bjorkman	-Nyqvist, M. (2013). Income shocks and gender gaps in Education: Evi	dence from Uganda.
Journal of Development Economics	, Volum 105, p. 237	-253.
Burgert, C. R., Colston, J., Roy, T., & Blake, Z. (2013). 	Geographic Displacement Procedure and
Georeferenced Data Release Policy for the Demographic and Health Surveys. DHS Spatia	l
Analysis Reports.	 Calverton, Maryland: USAID.
Burgert, R. C., Zachary, B., & Colston, J. (2013). 	Incorporating geographic information into
demographic and health surveys: A field guide to GPS data collection.	 Calverton, Maryland :
The United States Agenc	y for International Development.
Crespo Cuaresma, J., & Lutz, W. (2007). 	Human Capital, Age Structure and Economic Growth:
Evidence from a New Dataset. IIASA Interim Report.	 Laxenburg, Austria: IIASA.
Demographic and Health Survey Program. (2021, August 29	). Literacy	. Retrieved from Guide to DHS
Statistics DHS	-7: https://dhsprogram.com/data/Guide	-to-DHS	-Statistics/Literacy.htm
Dercon, S., & Porter, C. (2014). Live Aid Revisited. 	Oxford University Press	, Volum 12(4), p. 927	-948.
Ellis, J., & Galvin, K. (1994). Climate Patterns and Land	-use Practices in the Dry Zones of Africa.
BioScience 	, Volume 44 (5), p. 340	-349.
ESA Globcover Project. (2021, September 24). 	ESA	. Retrieved from GlobCover:
http://due.esrin.esa.int/page_globcover.	php

FAO, IFAD, UNICEF, WFP and WHO. (2021). 	The State of Food Security and Nutrition in the World
2021.	 Rome, Italy: FAO, IFAD, UNICEF, WFP and WHO.
Filmer, D., & Pritchett, L. (2001). Estimating wealth effects without expentidure data	-or tears: An
applica	tion to educational enrollments in states of India. 	Demography	, Volum 38(1), p. 115	-
132.
Glory, E., & Nsikak	-Abasi, E. (2013). Child labour in agriculture among poor rural households: some
issues and facts. 	European journal of Physical and Agricultural Sci	ences	, Volum 2 (1), p. 13	-19.
Haile, G. G., Tang, Q., Hosseini‐Moghari, S.‐M., Liu, X., Gebremicael, T. G., Leng, G., . . . Yun, X. (2020).
Projected Impacts of Climate Change on Drought Patterns over East Africa. 	Earth´s Future	,
Volum 8 (7), p. n/a.
Hanus	hek, E. (2013). Economic growth in developing countries: The role of human capital. 	Economics
of education review	, Volume 37, p. 204	-212.
Hertel, T. W., & Rosch, S. D. (2010). Climate Change, Agriculture and Poverty. 	Applied Economic
Perspectives and Polic	y, Volum 32(3), p. 355	-385.
Hongbin, L., Junsen, Z., & And, Z. (2008). The quantity	-quality trade	-off of children in a developing
country: Identification using Chinese twons. 	Demography	.
Humanitarian Data Exchange . (2021, September 23). 	Humanitarian Data 	Exchange 	. Retrieved from
Dataset, Administrative Divisions:
https://data.humdata.org/search?ext_administrative_divisions=1&q=&ext_page_size=25
International Labour Organization. (2010). 	Accelerating action against child labour 	– Global Report
under the fo	llow	-up to the ILO Declaration on Fundamental Principles and Rights at Work.
Geneva: International Labour Organization.
International Labour Organization, ILOSTAT database. (2021, January 29). 	The World Bank	. Retrieved
from Employment in 	agriculture (% of total employment) (modeled ILO estimate) 	- Sub	-
Saharan Africa: https://data.worldbank.org/indicator/SL.AGR.EMPL.ZS?locations=ZG
Knight, J., & Shi, L. (1996). Educational attainment and the rural	-urban divide in China. 	Oxford Bulletin
of E	conomics and Statistics	, Volume 58(1), p. 83	-117.
Kotsdam, A., & Flatø, M. (2020). Rainfall shocks and intimate partner violence in sub	-Saharan Africa.
Journal of Peace Research 	.
Linard, C., Gilbert, M., Snow, R., Noor, A., Tatem, A. J., & Schumann, G. J.	-P. (2012). Population
Distribution, Settlement Patterns and Accessibility across Africa in 2010. 	PLoS One	, Volum
7(2), p.e31743	-e31743.
Mahmus, M., & Riley, E. (2020). Household response to an extreme shock: Evidence on the
immediate impact of the Covid	-19 lockdown on economic outcomes and well	-being in rural
Uganda. 	World Development	.
Mani, S., Hoddinott, J., & Strauss, J. (2013). Determinants of Schooling: Empirical Evidence from Rural
Ethiopia. 	Journal of African Economies,	, Volum 22 (5), p. 693	-731.
Ma	sson	-Delmotte, V., Zhai, A., Pirani, S. C., Péan, S., Berger, N., Caud, Y., . . . B, Y. (2021). 	Climate
Change 2021 The Physical Scence Basis. Contribution of Working Group I to the Sixth

Assessment Report of the Intergovernmental Panel on Climate Change.	 Cambridge University
Press.
Murray, U., Paola, T., Demeranville, J., & Hurst, P. (2010). Breaking the rural poverty cycle: Getting
girls and boys out of work and into school. Washington DC.
National Drought Mitigation Center. (2021, August 2). 	U.S. Drought	 Monitor	. Retrieved from Drought
Classification:
https://droughtmonitor.unl.edu/About/AbouttheData/DroughtClassification.aspx
Nordstrom, A., & Cotton, C. (2020). Impact of a Severe Drought on Education: More Schooling but
Less Learning. 	Environment and dev	elopment economics	, Volum 24(4), p. 413	-436.
Opiyo, F., Wasonga, O., Nyangito, M., & Schilling, J. (2015). Drought Adaptation and Coping Strategies
Among the Turkana Pastoralists of Northern Kenya. 	International Journal of Disaster Risk
Science	, Volum 6(3)	, p. 295	-309.
Peace Research Institute Oslo. (2021, August 5). 	Download the PRIO	-GRID data	. Retrieved from PRIO	-
GRID 2.0: https://grid.prio.org/#/download
Randell, H., & Gray, C. (2016). Climate variability and educational obtainment: Evidence from rural
Ethiopia. 	Global Environmental Change	, Volum 41, p. 111	-123.
Rutstein, S. O., & Johnson, K. (2004). 	The DHS Wealth Index.	 Maryland: DHS comparative reports.
Schneider, U., Becker, A., Finger, P., Meyer	-Christoffer, A., Rudolf, B., & Ziese, M. (2015). 	GPCC F	ull
Data Reanalysis Version 7.0 at 0.5°: Monthly Land	-Surface Precipitation from Rain	-Gauges
built on GTS	-based and Historic Data.	 Putgarten, Germany: Global Precipitation Climatology
Centre at Deutscher Wetterdienst.
Shah, M. (2017). Drought of Opportunit	ies: Contemporaneous and Long Term Impacts of Rainfall
Shocks on Human Capital. 	Journal of Political Economy	, Volum 125(2), p. 527	-561.
Stock, J. H., & Watson, M. W. (2015). 	Introduction to Econometrics.	 (Third Edition). Harlow, England:
Pearson Education Limited.
Sumida, S., & Kawata, K. (2021). An analysis of the learning performance gap between urban and
rural areas in sub	-Saharan Africa. 	South African Journal of Education	.
Tacoli, C., & Satterthwaite,	 D. (2013). Gender and urban change. 	Environment & Urbanization	, Volum
25(1), p. 3	-8.
The DHS Program. (2021, August 2). 	DHS Model Questionnaires 	. Retrieved from The DHS Program:
https://dhsprogram.com/Methodology/Survey	-Types/DHS	-Questionnaires.cfm
The W	orld Bank. (2021, September 3). 	Education, School enrollment, primary	. Retrieved from The
World Bank: https://data.worldbank.org/topic/education?locations=ZG	-1W
The World Bank. (2021, September 21). 	Sub	-Sarahan Africa	. Retrieved from The World Bank Data:
https://data.worldbank.org/country/ZG
Thurlow, J., Zhu, T., & Xinshen, D. (2012). Current Climate Variability and Future Climate Change:
Estimated Growth and Poverty Impacts for Zambia. 	Review of Development Economics	,
Volum 16. p. 394	–411.

Tollefsen, A. F.	, Bahgat, K., Nordkvelle, J., & Buhaug, H. (2015). 	PRIO	-GRID v.2.0 Codebook.	 Oslo,
Norway: Peace Research Institute Oslo.
Tollefsen, A. F., Strand, H., & Buhaug, H. (2012). PRIO	-GRID: A unified spatial data structure. 	Journal
of Peace Research	, Volum 49(2), p. 363	-374.
UNESCO Institute for Statistics. (2021, August 10). 	Agricultural land (% of land area) 	- Sub	-Saharan
Africa	. Retrieved from The World Bank:
https://data.worldbank.org/indicator/AG.LND.AGRI.ZS?locations=ZG
UNESCO Institute for Statistics. (2021, September 15). 	School enrollment, primary (gross), gender
parity index (GPI) 	- Sub	-Saharan Africa	. Retrieved from The W	orld Bank:
https://data.worldbank.org/indicator/SE.ENR.PRIM.FM.ZS?locations=ZG
United Nations. (2021, September 21). 	United Nations	. Retrieved from United Nations Sustainable
Development : https://sdgs.un.org/goals/goal4
United Nations. (2021, September 20	). United Nations Climate Action	. Retrieved from What is
climate change?: https://www.un.org/en/climatechange/what	-is-climate	-change
Vestal, T. M. (1985). Famine in Ethiopia: Crisis of Many Dimensions. 	Africa Today	, Volume 32 (4), p. 7	-
28.
World Vision. (2	021, August 11). 	1980s Ethiopia famine: Facts, FAQs, and how to help	. Retrieved from
World Vision: https://www.worldvision.org/disaster	-relief	-news	-stories/1980s	-ethiopia	-
famine	-facts

10	 Appendix

10.1	 Additional tables

Table 	A 	1: All gender interaction terms
Full sample	 	(1) 	(2) 	(3) 	 	Years of Education	 	Female x Moderate Drought	 	-0.278	*** 	 	-0.311	*** 	 	(0.0716)	 	 	(0.0677)	 	 	 	 	 	Male x Moderate Drought	 	-0.110	 	 	-0.102	 	 	(0.0791)	 	 	(0.0876)	 	 	 	 	 	Female x 	Extreme Drought	 	 	0.0315	 	0.0759	 	 	 	(0.0800)	 	(0.0768)	 	 	 	 	 	Male x Extreme Drought	 	 	-0.0175	 	-0.00618	 	 	 	(0.0645)	 	(0.0700)	 	 	 	 	 	Female	 	-1.270	*** 	-1.443	*** 	-1.271	*** 	 	(0.126)	 	(0.0973)	 	(0.126)	 	 	 	 	 	Urban	 	1.604	*** 	1.604	*** 	1.604	*** 	 	(0.162)	 	(0.163)	 	(0.162)	 	 	 	 	 	Constant	 	-0.274	 	-0.341	 	-0.284	 	 	(0.458)	 	(0.440)	 	(0.452)	 	N 	228872	 	228872	 	228872	 	R2 	0.238	 	0.238	 	0.238	 	 	 	 	 	Wet region	 	(1) 	(2) 	(3) 	 	Years of Education	 	Female x Moderate Drought	 	-0.264*	 	 	-0.143	 	 	(0.113)	 	 	(0.115)	 	 	 	 	 	Male x 	Moderate Drought	 	0.298*	 	 	0.187	 	 	(0.122)	 	 	(0.127)	 	 	 	 	 	Female x Extreme Drought	 	 	-0.454***	 	-0.427**	 	 	 	(0.125)	 	(0.124)	 	 	 	 	 	Male x Extreme Drought	 	 	0.309*	 	0.270	 	 	 	(0.135)	 	(0.141)	 	 	 	 	 	Female	 	-1.778***	 	-2.041***	 	-1.778***	 	 	(0.214)	 	(0.174)	 	(0.215)	 	 	 	 	 	Urban	 	0.824***	 	0.823***	 	0.823***	 	 	(0.111)	 	(0.110)	 	(0.110)	 	 	 	 	 	Constant	 	-0.0946	 	0.118	 	-0.0456	 	 	(0.480)	 	(0.454)	 	(0.459)	 	N 	61611	 	61611	 	61611	 	R2 	0.291	 	0.292	 	0.292	 	 	 	 	 	Dry region 	 	(1) 	(2) 	(3) 	 	Years of Education	 	Female x Moderate 	Drought	 	-0.401***	 	 	-0.408***	 	 	(0.0901)	 	 	(0.0896)	 	 	 	 	 	Male x Moderate Drought	 	-0.167	 	 	-0.186	 	 	(0.110)	 	 	(0.127)	 	 	 	 	 	Female x Extreme Drought	 	 	-0.0456	 	0.0167	 	 	 	(0.0485)	 	(0.0480)	 	 	 	 	 	Male x Extreme Drought	 	 	0.00945	 	0.0376	 	 	 	(0.0798)	 	(0.0893)	 	 	 	 	 	Female	 	-0.940***	 	-1.126***	 	-0.940***	 	 	(0.143)	 	(0.130)	 	(0.143)	 	 	 	 	 	Urban	 	2.150***	 	2.150***	 	2.150***	 	 	(0.295)	 	(0.295)	 	(0.295)	 	 	 	 	 	Constant	 	-0.698	 	-0.801	 	-0.697

(0.478)	 	(0.464)	 	(0.476)	 	N 	114682	 	114682	 	114682	 	R2 	0.225	 	0.224	 	0.225	 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All specifications have 	fixed effects on grid cell level and include controls for gender, age, wealth, agricultural 	land, residency, religion and period interviewed. 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.

Table 	A 	2: Effect of drought on 	separate samples
 	Decade born: Effect of experiencing a drought
Years of
education
1940	-1950	 	1950	-1960	 	1960	-1970	 	1970	-1980	 	1980	-1990	 	1990	-2000

Moderate
Drought
-0.256	** 	0.0399	 	0.261	 	-0.426	** 	-0.167	 	-0.0000651
 	(0.0921)	 	(0.0802)	 	(0.147)	 	(0.132)	 	(0.0860)	 	(0.104)

Extreme
Drought
0.236	 	0.0110	 	0.102	 	0.189	 	-0.161	 	0.0666
 	(0.147)	 	(0.101)	 	(0.0717)	 	(0.105)	 	(0.0886)	 	(0.113)

Female	 	-0.561	*** 	-1.415	*** 	-1.798	*** 	-1.803	*** 	-1.464	*** 	-0.603	***
 	(0.166)	 	(0.148)	 	(0.134)	 	(0.122)	 	(0.125)	 	(0.138)

Urban	 	1.306	*** 	1.739	*** 	1.765	*** 	1.802	*** 	1.425	*** 	1.279	***
 	(0.251)	 	(0.372)	 	(0.305)	 	(0.247)	 	(0.131)	 	(0.130)

Constant	 	-0.931	 	1.851	*** 	1.876	*** 	2.700	*** 	3.614	*** 	0.859	**
 	(0.851)	 	(0.385)	 	(0.278)	 	(0.272)	 	(0.258)	 	(0.256)
N 	4330	 	16489	 	37962	 	60481	 	67406	 	41970
R2 	0.140	 	0.172	 	0.186	 	0.193	 	0.180	 	0.162
P-value 	 	0.9018	 	0.6718	 	0.0238	* 	0.1411	 	0.0150	* 	0.6190
Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All specifications have 	fixed effects on grid cell level and include controls for gender, age, wealth, agricultural 	land, residency, religion and period interviewed. 	 	The 	given p	-value is that of the 	F-statistic	 testing the joint significance of the two drought dummy variables. 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.
Table 	A 	3: Heterogeneous effects of drought on any education
 	Any Education	 	 	 	 	(1) 	(2) 	(3)
Extreme Drought	 	 	-0.00325	 	0.00186	 	 	 	(0.00862)	 	(0.00827)	 	 	 	 	 	Male x Extreme Drought	 	 	0.00458	 	0.00146	 	 	 	(0.0122)	 	(0.0119)	 	 	 	 	 	Moderate 	Drought	 	-0.0354	*** 	 	-0.0361	*** 	 	(0.00824)	 	 	(0.00823)	 	 	 	 	 	Male x Moderate Drought	 	0.0213	 	 	0.0205	 	 	(0.0116)	 	 	(0.0105)	 	 	 	 	 	Female	 	-0.148	*** 	-0.166	*** 	-0.148	*** 	 	(0.0126)	 	(0.0110)	 	(0.0126)	 	 	 	 	 	Urban	 	0.121	*** 	0.121	*** 	0.121	*** 	 	(0.0144)	 	(0.0144)	 	(0.0144)	 	 	 	 	 	Constant	 	0.362	*** 	0.353	*** 	0.362	*** 	 	(0.0532)	 	(0.0501)	 	(0.0524)
N 	228993	 	228993	 	228993	 	R2 	0.182	 	0.182	 	0.182
P-value	 (extreme)	 	 	 	0.1496	 	P-value	 (moderate)	 	 	 	0.0013	** 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All specifications have 	fixed effects on grid cell level and include controls for gender, age, wealth, agricultural 	land, residency, religion and period interviewed. 	 	The 	given p	-value is that of the 	F-statistic	 testing the joint significance of the drought dummy variable and the interaction term. 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.

Table 	A 	4: Probability of having completed primary and secondary
 	Completed primary	 	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate Drought	 	0.00337	 	0.00478	 	-0.00733	 	-0.00615	 	-0.00578	 	 	(0.00987)	 	(0.00965)	 	(0.00677)	 	(0.00672)	 	(0.00631)	 	 	 	 	 	 	 	Extreme Drought	 	-0.00267	 	-0.00160	 	0.000739	 	-0.000693	 	0.000864	 	 	(0.0125)	 	(0.0119)	 	(0.00501)	 	(0.00534)	 	(0.00506)	 	 	 	 	 	 	 	Female	 	 	-0.0862	*** 	-0.0864	*** 	-0.0978	*** 	-0.0931	*** 	 	 	(0.00767)	 	(0.00773)	 	(0.00698)	 	(0.00697)	 	 	 	 	 	 	 	Urban	 	 	0.194	*** 	0.182	*** 	0.184	*** 	0.115	*** 	 	 	(0.0182)	 	(0.0181)	 	(0.0170)	 	(0.0143)	 	 	 	 	 	 	 	Constant	 	0.314	*** 	0.322	*** 	0.136	*** 	0.341	*** 	0.286	*** 	 	(0.00708)	 	(0.00956)	 	(0.0154)	 	(0.0428)	 	(0.0434)	 	N 	243841	 	243841	 	243841	 	233811	 	228993	 	R2 	0.000	 	0.031	 	0.060	 	0.069	 	0.091
 	Secondary or higher education	 	 	(1) 	(2) 	(3) 	(4) 	(5)
Moderate 	Drought	 	-0.00800	 	-0.00705	 	-0.0115	 	-0.0125	 	-0.0116	 	 	(0.0102)	 	(0.0100)	 	(0.00767)	 	(0.00749)	 	(0.00710)	 	 	 	 	 	 	 	Extreme Drought	 	-0.00871	 	-0.00786	 	-0.000963	 	-0.00343	 	-0.00235	 	 	(0.0116)	 	(0.0111)	 	(0.00393)	 	(0.00378)	 	(0.00377)	 	 	 	 	 	 	 	Female	 	 	-0.0648	*** 	-0.0641	*** 	-0.0756	*** 	-0.0717	*** 	 	 	(0.00641)	 	(0.00652)	 	(0.00570)	 	(0.00585)	 	 	 	 	 	 	 	Urban	 	 	0.190	*** 	0.178	*** 	0.178	*** 	0.112	*** 	 	 	(0.0159)	 	(0.0160)	 	(0.0152)	 	(0.0135)	 	 	 	 	 	 	 	Constant	 	0.221	*** 	0.217	*** 	0.0621	*** 	0.305	*** 	0.247	*** 	 	(0.00816)	 	(0.0103)	 	(0.0135)	 	(0.0475)	 	(0.0477)	 	N 	243841	 	243841	 	243841	 	233811	 	228993	 	R2 	0.000	 	0.031	 	0.071	 	0.076	 	0.100
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed
No No No No No No
No No No No No No
No No No Yes	 No No
Yes	 No No Yes	 Yes	 No
Yes	 Yes	 Yes	 Yes	 Yes	 Yes	 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All specifications have 	fixed effects on grid cell 	level.	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

Table 	A 	5: Effect of drought on literacy when controlling for education
Literacy	 	(1) 	(2) 	(3) 	(4) 	 	Literal	 	Literal	 	Literal	 	Literal
Started Education	 	0.672	*** 	 	 	0.604	*** 	 	(0.00956)	 	 	 	(0.0117)	 	 	 	 	 	 	Completed Primary	 	 	0.375	*** 	 	0.0933	** 	 	 	(0.0186)	 	 	(0.0282)	 	 	 	 	 	 	Secondary Higher	 	 	 	0.344	*** 	0.135	*** 	 	 	 	(0.0218)	 	(0.0265)	 	 	 	 	 	 	Moderate Drought	 	-0.00355	 	-0.0194	*** 	-0.0179	** 	-0.00400	 	 	(0.00365)	 	(0.00527)	 	(0.00587)	 	(0.00315)	 	 	 	 	 	 	Extreme Drought	 	0.00956	*** 	0.0125	** 	0.0142	** 	0.00928	*** 	 	(0.00270)	 	(0.00401)	 	(0.00512)	 	(0.00257)	 	 	 	 	 	 	Female	 	-0.103	*** 	-0.183	*** 	-0.193	*** 	-0.0954	*** 	 	(0.00702)	 	(0.0125)	 	(0.0128)	 	(0.00737)	 	 	 	 	 	 	Urban	 	0.0407	*** 	0.0723	*** 	0.0742	*** 	0.0229	*** 	 	(0.00614)	 	(0.0103)	 	(0.0109)	 	(0.00516)	 	 	 	 	 	 	Constant	 	0.167	*** 	0.281	*** 	0.305	*** 	0.106	** 	 	(0.0414)	 	(0.0601)	 	(0.0620)	 	(0.0394)
N 	204812	 	204812	 	204812	 	204812	 	R2 	0.475	 	0.288	 	0.254	 	0.505
Note: Standard errors are 	clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all specifications. All 	specifications have fixed effects on grid cell level and decade born, in addition to and controls for gender, age, wealth, ag	ricultural la	nd, 	residency, religion and period interviewed	. 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

Table 	A 	6: Analysis	 replicated for Ethiopia
Years of education	 	(1) 	(2) 	(3) 	(4) 	(5) 	Ethiopia
Moderate 	Drought	 	-0.0187	 	-0.0625	 	-0.0625	 	-0.287	 	-0.339	* 	 	(0.253)	 	(0.221)	 	(0.221)	 	(0.146)	 	(0.127)	 	 	 	 	 	 	 	Extreme Drought	 	-0.201	 	-0.114	 	-0.114	 	-0.119	 	-0.0683	 	 	(0.123)	 	(0.100)	 	(0.100)	 	(0.0909)	 	(0.0861)	 	 	 	 	 	 	 	Female	 	 	-1.830	*** 	-1.830	*** 	-2.031	*** 	-1.685	*** 	 	 	(0.203)	 	(0.203)	 	(0.204)	 	(0.209)	 	 	 	 	 	 	 	Urban	 	 	5.804	*** 	5.804	*** 	5.289	*** 	4.093	*** 	 	 	(0.368)	 	(0.368)	 	(0.273)	 	(0.337)	 	 	 	 	 	 	 	Constant	 	3.698	*** 	3.476	*** 	3.476	*** 	6.451	*** 	3.641	*** 	 	(0.230)	 	(0.205)	 	(0.205)	 	(0.470)	 	(0.498)
N 	49511	 	49511	 	49511	 	49299	 	49299	 	R2 	0.001	 	0.236	 	0.236	 	0.296	 	0.339	 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All specifications have 	fixed effects on grid cell level and include 	the same control variables 	as in regression tab	le 6.1. 	 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.
Literacy	 	(1) 	(2) 	(3) 	(4) 	Ethiopia
Moderate Drought	 	-0.0144	 	-0.0437	*** 	-0.0365	** 	-0.0359	** 	 	(0.0226)	 	(0.00894)	 	(0.0109)	 	(0.0105)	 	 	 	 	 	 	Extreme Drought	 	-0.00605	 	-0.00173	 	0.0183	* 	0.0182	* 	 	(0.00950)	 	(0.0115)	 	(0.00789)	 	(0.00778)	 	 	 	 	 	 	Female	 	-0.294	*** 	-0.319	*** 	-0.280	*** 	-0.277	*** 	 	(0.0285)	 	(0.0300)	 	(0.0265)	 	(0.0261)	 	 	 	 	 	 	Urban	 	0.474	*** 	0.315	*** 	0.275	*** 	0.273	*** 	 	(0.0336)	 	(0.0362)	 	(0.0303)	 	(0.0310)	 	 	 	 	 	 	Constant	 	0.554	*** 	0.654	*** 	0.228	*** 	0.328	*** 	 	(0.0298)	 	(0.0526)	 	(0.0512)	 	(0.0632)
N 	49485	 	49485	 	49273	 	49273	 	R2 	0.179	 	0.250	 	0.274	 	0.275	 	 	 	 	 	 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in 	parentheses for all specifications. All specifications have 	fixed effects on grid cell level and include the same control variables as in regression table 6.	8. 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.
Any education	 	(1) 	(2) 	(3) 	(4) 	(5) 	Ethiopia
Moderate Drought	 	-0.0169	 	-0.0192	 	-0.0698	** 	-0.0648	** 	-0.0621	** 	 	(0.0389)	 	(0.0336)	 	(0.0157)	 	(0.0151)	 	(0.0152)	 	 	 	 	 	 	 	Extreme Drought	 	-0.0217	 	-0.0115	 	0.0219	 	0.0184	 	0.0192	 	 	(0.0160)	 	(0.0142)	 	(0.0120)	 	(0.0105)	 	(0.0107)	 	 	 	 	 	 	 	Female	 	 	-0.249	*** 	-0.210	*** 	-0.226	*** 	-0.220	*** 	 	 	(0.0296)	 	(0.0301)	 	(0.0289)	 	(0.0292)	 	 	 	 	 	 	 	Urban	 	 	0.454	*** 	0.398	*** 	0.367	*** 	0.244	*** 	 	 	(0.0280)	 	(0.0393)	 	(0.0273)	 	(0.0303)
N 	49511	 	49511	 	49511	 	49299	 	49299	 	R2 	0.001	 	0.147	 	0.298	 	0.311	 	0.333
 Note: 	Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all specifi	cations. All specifications have 	fixed effects on grid cell level and include the same control variables as in regression table 6	.7. 	Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001.

Table 	A7: Clustering by country
Clustering on country	 	 	(1) 	(2) 	(3) 	(4) 	(5) 	 	Education	 	Education	 	Education	 	Education	 	Education
Moderate Drought	 	-0.124	 	-0.0993	 	-0.252	** 	-0.249	*** 	-0.232	** 	 	(0.152)	 	(0.148)	 	(0.0541)	 	(0.0515)	 	(0.0581)	 	 	 	 	 	 	 	Extreme Drought	 	-0.0265	 	-0.00866	 	0.0327	 	0.0371	 	0.0481	 	 	(0.311)	 	(0.296)	 	(0.0965)	 	(0.0969)	 	(0.0896)	 	 	 	 	 	 	 	Female	 	 	-1.443	*** 	-1.440	*** 	-1.451	*** 	-1.420	*** 	 	 	(0.245)	 	(0.253)	 	(0.245)	 	(0.234)	 	 	 	 	 	 	 	Urban	 	 	2.848	*** 	2.677	*** 	2.687	*** 	1.604	** 	 	 	(0.536)	 	(0.506)	 	(0.465)	 	(0.467)	 	 	 	 	 	 	 	Constant	 	4.600	*** 	4.839	*** 	1.941	*** 	1.199	 	-0.214	 	 	(0.126)	 	(0.270)	 	(0.277)	 	(0.554)	 	(0.737)
N 	243718	 	243718	 	243718	 	233690	 	228872	 	R2 	0.000	 	0.091	 	0.162	 	0.180	 	0.238
Religion	 	 Wealth	 	 Agriculture 	 	 Decade born	 	 Age	 	 Year interviewed
No No No No No No
No No No No No No
No No No Yes	 No No
Yes	 No No Yes	 Yes	 No
Yes	 Yes	 Yes	 Yes	 Yes	 Yes
Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications. All 	specifications have fixed effects on grid cell. Significance levels: 	 * p < 0.05, 	** p < 0.01, 	*** p < 0.001

Table 	A 	8: Correlation matrix between drought and education
Variables	 	(1) 	(2) 	(3)
(1) 	Years of e	ducation	 	1.000	 	 	 	 	 	 	 	(2) Moderate Drought	 	-0.079*	 	1.000	 	 	 	(0.000)	 	 	 	(3) Extreme Drought	 	-0.136*	 	0.310*	 	1.000	 	 	(0.000)	 	(0.000)	 	 	*** 	p<0.01, ** p<0.05, * p<0.1

Table 	A 	9: Variables used in regression	 analysis
Variable 	 	Usage 	 	Description 	 	Source
Years of education	 	Dependent variable	 	How many years the
individual has attended any
education
DHS
Highest level of
education
Dependent variable	 	The highest level of attained
education
DHS
Literacy 	 	Dependent variable	 	Defined as 	the ability to
read parts of a sentence
DHS
Extreme drought 	while
in school
Main v	ariable of
interest
Dummy variable 	capturing
incidences of SPEI below 	-
1,6	 during the years
individuals are eligible to
attend school
SPEI through PRIO	-
GRID
Moderate drought 	while
in school
Main v	ariable of
interest
Dummy variable capturing
incidences of SPEI below 	-
0,8	 during the years
individuals are eligible to
attend school
SPEI through PRIO	-
GRID
Gender	 	Explanatory variable	 	Recoded to female, where
female = 1 is female
DHS
Residency	 	Explanatory variable	 	Recoded to urban where
urban = 1 is urban residence
DHS
Wealth index	 	Explanatory variable	 	Index ranging from 1	-5,
where 1 is poorest and 5 is
richest. Computed as a
relative measure for each
survey round.
DHS
Religion	 	Explanatory variable	 	Recording religion of
individuals.
DHS
Age	 	Explanatory variable	 	Age of respondent 	at time
of survey	.
DHS
Agricultural land 	 	Explanatory variable	 	Gives the percentage of land
used for agriculture within
each grid cell. Variable is
recoded into ten ranges,
from 0	-9, where means
the cell has 	- 10 percent
agricultural land, and 9
indic	ates that the cell has
between 90 and 100 percent
land used for agriculture.
European Space
Agency Global Cover
Portal through PRIO	-
GRID
First a	dministrative
levels
Clustering variable	 	 	Humanitarian Data
Exchange

10.2	 Further explanations

Construction of 	new grid cells
The drought variables are recorded in grid cells, sourced from SPEI. These 	cells to not stay
within country or administrative borders. 	Individual country residence is recorded by the
DHS. However, as grid cells cross borders, two individuals living on either side of a country
border will be placed in the same grid, regardless of their country residence. 	Th	e procedure 	of
split	ting	 grid	 cel	ls between countries is important when employing fixed effects on grid level	,
as otherwise one would be assuming	 that an individual on either side of a country border
experiences the same kind of local development	. Hence fixing on th	e original grid cell le	vel
provided by PRIO	 could yield biased estimates.
To	 ensure that 	grid cells do not overlap 	borders used for analysis, new grid cells were
constructed.	 This was done by	 generating a new set of grid cells, which are a combination of a
country code and the grid code it is ensured that grids are split between countries. Using this
grid code to repeat the process on ADM1 levels as well, ensures that each grid pertains to one
country, and one ADM1 level.	 After the process, some grid cell levels still pertained to more
than one ADM1 border, due to 	codes used for generation adding up to the same numbe	rs. In
these cases, codes were manually 	recoded 	to ensure every new grid cell h	ad a unique code.
This process ensured grid cells were nested within only one country and one ADM1 level
over time.

Explanatory	 variables and 	the role 	of agricultural dependence
The signs and magnitude of the majority of control variables are not included in the
regression tables. 	Therefore, a	 short	 description of their role is given here.
Increasing wealth predicts higher education, higher age predicts less education, and
conseq	uently t	he more recent an individual was born the more education it attains. The level
of agricultural land in the area is not significant in explaining the years of education in any of
the specifications	.
All types of religion are significant in explainin	g educational attainment. Relative to being
catholic, being a protestant yields almost half a year more in school, while being Muslim,
animist, having no religion or reporting “other religion” all yield around a year less schooling
compared to 	individuals 	with 	catholic 	believes	.

Also, the period in which individuals where interviewed is important in the specifications.
Relative to being interviewed in the period between 1990 and 1995 individuals who are
interviewed between 2015 and 2020 are likely to attend	 school for almost 1,5 years longer.
The decade in which individuals are born is an important time fixed effect, proving highly
significant. We observe that relative to being born in the 1930ies being born in each preceding
decade yields higher years of e	ducation.
In an alternative specification interaction terms between levels of agricultural land and
drought occurrence are included to check whether the hypothesis that drought shocks affect
education in agricultural heavy areas more than in non	-agricultu	ral areas hold. These
interaction terms are not significant in explaining years of education attained. This might
indicate that the potential effect of drought on education goes through other channels than
changes in agricultural produce.

Robustness of t	he linear model
The linear model is selected 	for analysis in sections	 6.2 and 6.3 	mainly 	as it is easier to
include fixed effects to this model, compared to the alternatives of an ordered logistic or
ordered probit model. The ordered models also assume som	e form of linearity between the
dependent variable and the effect of the treatment variable 	(Stock & Watson, 2015)	. Below is
a table showing the results from employing an ordered logistic and ordered probit model to
the same analysis as performed in section 6.2.
Table 	A 	10	: Comparing t	he linear model 	to the	 ordered logistic and probit models
Educa	tional attainment	 	Linear model	 	Ordered logistic model	 	Ordered probit model
 	 	 	 	 	 	 	Moderate Drought	 	-0.0183	 	-0.0103	 	-0.124	 	-0.0653	 	-0.0672	 	-0.0380	 	 	(0.0378)	 	(0.0363)	 	(0.100)	 	(0.0790)	 	(0.0563)	 	(0.0446)	 	 	 	 	 	 	 	 	Extreme Drought	 	-0.0285	 	-0.0230	 	-0.500	*** 	-0.455	*** 	-0.300	*** 	-0.279	*** 	 	(0.0475)	 	(0.0437)	 	(0.117)	 	(0.123)	 	(0.0670)	 	(0.0717)	 	 	 	 	 	 	 	 	Female	 	 	-0.455	*** 	 	-0.582	*** 	 	-0.358	*** 	 	 	(0.0270)	 	 	(0.0586)	 	 	(0.0335)	 	 	 	 	 	 	 	 	Urban	 	 	0.950	*** 	 	1.559	*** 	 	0.896	*** 	 	 	(0.0677)	 	 	(0.134)	 	 	(0.0759)	 	N 	243840	 	243840	 	243840	 	243840	 	243840	 	243840
 	 	 	 	 	 	 	Note: Standard errors are clustered at the ADM1 level, and heteroscedasticity robust. They are given in parentheses for all s	pecifications.	 	Significance levels	:* p < 0.05, 	** p < 0.01, 	*** p < 0.001

Although the selected specification for the linear model yields non	-significant results of
drought on the levels of educational attainment, the sign of the results of the ordered probit
and logit model are in line with the general f	indings. Being negative, they indicate that if
individuals experience drought they are more likely to be in the lower levels of educational
attainment. This indicates that the linear model behaves well. Since the results of the linear
model is without sign	ificance, 	it is possible to	 conclude that 	the 	selected 	model is 	more
conservativ	e, and hence	 the significant results that do emerge can be trusted more.

Difference	-in	-difference approach to evaluate the effect of famine in Ethiopia
One of the central underpinning of the difference in difference approach is that the treated and
untreated groups have the same trend prior to the shock. This is the reason why the main
meth	od applied to consider the effects of the famine is 	the synthetic control method	.
However, a difference in difference analysis is performed to 	evaluate the validity of	 the
results from the synthetic control analysis. 	This	 approach compares the trend in educational
outcomes for people in Ethiopia before and after the famine, with this trend in other countries,
that were less affected by famines. Tanzania and Uganda are selected as the comparisons with
Ethiopia, as the obser	vations in the dataset indicate that these two countries did not
experience a persistent drought. The observations are limited to those that attended school
after 1960.
Figure 	A 	11	: Mean years of education	 over time	 in Ethiopia, Ugan	da and Tanzania

Note: 	Mean years of education in Ethiopia, Uganda and 	Tanzania, calculated by the mean
years 	of education of those born in the same year	. The figure is developed using Stata.

Table 	A 	12	: Diff	erence	-in-diff	erence	 results
Number of observations: 66589
 	Before famine After famine
 Control: 9853 11588
 Treated: 31781 13367

Outcome variable	 	Years of education
Before	 	 	Control	 	5.206	 	Treated	 	3.078	 	Difference 	 	-2.127	 	 	 	After	 	 	Control 	 	6.711	 	Treated	 	5.612	 	Difference	 	-1.099
Diff	-in-diff	 	1.028

The results from the difference in difference analysis, in table 	A12	, indicate that the mean
level of education in Ethiopia was more than 	two	 years behind that of Uganda and Tanzania
prior to 1982. After 1982 the difference in means between the two countri	es is reduced to
around 	one	 year. All differences are significant at 	one	 percent level.
From figure 	A11	 we observe that those born around 1975, indicated by the vertical line,
experience a small reduction in the mean years of education reported, 	possibly	 due to the
short	-term effect of the	 famine. However, the mean years of education is trending sharply
upwards for those that start education after the famine. This trend is not observed for Uganda
and Tanzania.
This can indicate that the effect of the fami	ne in Ethiopia had a positive effect on years of
education attained, relative to that of Uganda and Tanzania. There exist however many other
possible reasons for the observed trends in mean educational attainment, both in Ethiopia and
the countries used as	 controls.	 The trend observed using the difference	-in-difference approach
supports t	he findings 	from section 6.4.