-
Essays in Natural Resources and Development Economics
by
Danamona Holinirina Andrianarimanana
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Agricultural and Resource Economics
in the
Graduate Division
of the
University of California, Berkeley
Committee in charge:
Professor Maximilian Au↵hammer, Chair
Professor Alain de Janvry
Associate Professor Solomon Hsiang
Spring 2018
-
Essays in Natural Resources and Development Economics
Copyright 2018
by
Danamona Holinirina Andrianarimanana
-
Abstract
Essays in Natural Resources and Development Economics
by
Danamona Holinirina Andrianarimanana
Doctor of Philosophy in Agricultural and Resource Economics
University of California, Berkeley
Professor Maximilian Au↵hammer, Chair
There is great geographical overlap between key areas of natural
resources, global biodi-versity and regions of acute poverty. The
world’s poorest people, including 59% of thepopulation of Asia,
Africa and Latin America live in rural areas alongside great
naturalresources on which they heavily rely for food source and
income generation. However,proximity does not imply free unlimited
access and often involves a great deal of trade-o↵s and risk
ranging from natural weather and catastrophic shocks a↵ecting
resourceavailability, productivity and even human lives, to changes
in governance and resourceuse regulations. In this dissertation, I
study the linkages between natural resources use,livelihoods,
governance and the environment, using the case study of Madagascar,
a low-income country with great biodiversity and natural resources
endowment. In particular, Istudy how di↵erent types of regulations
and restrictions a↵ect household resource use andwell-being. In
Chapter 1, I evaluate the health and wealth trade-o↵s of the widely
prac-ticed fire use in agriculture in Madagascar, using
high-frequency satellite data to modelpollution exposure taking
advantage of random variation in wind direction. In Chapter2, I
study how poor households cope with natural disasters using the
quasi-experimentsetting of high frequency cyclones in Madagascar.
In Chapter 3, I take advantage of aunique dataset coupled with the
staggered rollout of a biodiversity conservation policy tostudy the
impacts of community-based conservation on bushmeat hunting in
northeasternMadagascar.
In the first chapter, I study the impacts of agricultural fires
on local health and on agri-cultural productivity in Madagascar.
Every year, despite agricultural fires being illegal,25% to 50% of
grasslands and 7% to 10% of forests are set on fire due to
slash-and-burnagriculture and livestock farming. This leads to
great pollution throughout the island,yet there is limited
empirical evidence on the health impacts of fires in the island. I
firstestimate the health impacts of fires by using high frequency
and high resolution satel-lite data on fire location and wind speed
on the day of fire to model pollution exposurearound population
centers. Identification comes from the random variation in wind
di-rection and the frequent change in pollution source. I find that
agricultural fires greatlyimpact birth outcomes and respiratory
health of infants and that fires are responsible forover 4,000
“missing infants”, or 0.7% of all births across the island every
year. To identifythe agricultural impacts of fires, I use an
instrumental variable strategy taking advantageof a rapid expansion
of protected areas in Madagascar that led to tripling of
protectedareas and delimitation of numerous potential parks. I use
proposed parks, areas thatwere physically delimited as potential
o�cial protected areas, as an instrument for fires.Delimitation of
proposed parks led to reduced fire activity, however, since parks
were notactually implemented, surrounding populations were
una↵ected by potential economicreturns or changes in behavior that
would raise concerns regarding the validity of the
1
-
exclusion restriction. Grassland fires led to increased
livestock production and yields forcassava and corn, whereas forest
fires increased corn farming land and harvest, leadingto decreased
food prices. These quantity and price e↵ects increased consumer
surplus byUSD1.884 billion per year, implying that, for the output
gains to outweigh the mortalityimpacts, one would have to assume a
value of statistical life of less than USD440,000,whereas typical
values for VSL range from 4 million to 9 million USD. Therefore
themortality costs of fires alone, excluding hospitalization costs
and morbidity, exceed thebenefits from increased agricultural
production. Given that land use rights are ambigu-ous and
government resources in regulating forest fires are limited, a more
cooperativeand integrative approach such as payments for ecosystem
services might be e↵ective inincentivizing farmers to engage in
less frequent more sustainable fire activity. In the sec-ond
chapter, I use cyclone track data and hourly wind direction data to
model cycloneexposure and study the impact of tropical storms in
Madagascar. Madagascar is thesecond most exposed country to
multi-disaster risks in Africa, and experiences multipleepisodes of
droughts, floods, locust invasions and cyclones every year. On
average, theisland yearly experiences three to five cyclones that
claim 10% to 30% of annual GDPin post-disaster losses and damages.
Indeed, 74% of total labor is employed in agricul-ture,
furthermore, agricultural products including exports amount to 45%
of GDP. Yet,there is little government e↵ort in terms of risk
mitigation, resilience building and evendisaster relief. Looking at
the impact of cyclones on household well-being along
multipledimensions, I find that both rural and urban households are
negatively impacted by cy-clones in Madagascar despite better
infrastructure and less reliance on natural resourcesin urban
areas. While rural areas experience more physical losses than urban
areas asmeasured by cyclone e↵ects on housing and access to
electricity, rural households are ableto smooth consumption and are
less prone to cyclone-driven poverty compared to theirurban
counterparts. In this latter group, average cyclones have no
significant impacton physical assets, but lead to lower consumption
and higher rates of transient poverty.I show that this is the
result of a strong informal safety net between rural and
urbanfamilies through informal insurance and relief in the form
inter-household transfers. Toprovide relief to rural families,
urban households reduce expenditure in non-food expen-diture
including education. This suggests that, while partially e↵ective
in managing riskand achieving consumption smoothing along some key
dimensions, lack of formal insur-ance diverts resources away from
potentially productive investments such as educationand towards
unequivocally necessary informal relief. In the third chapter, I
use a uniquehousehold-level panel data to evaluate how
community-based conservation impacts bush-meat or wildlife hunting
and consumption in the northeastern rainforests of Madagascar,where
lemurs, bats, carnivores, tenrecs and bush pigs are commonly
consumed to sat-isfy nutritional needs. Taking advantage of the
staggered rollout of the policy, I findthat community-based
conservation has decreased overall hunting in the study area
byreducing opportunistic hunting and hunting by less reliant,
richer households. This e↵ectwas larger among relatively more
educated households. Furthermore, community-basedconservation
successfully modified consumption patterns among poorer households
suchthat illegal hunting (hunting of lemurs and bats) was reduced
and substituted by huntingpractices conforming with conservation
practices (seasonal hunting of sustainable prey).While these
results are encouraging given the increasing shift towards
decentralization,it is important to note that, in my study setting,
community-based conservation wasfound to have some limitations.
First, e↵ects did not persist and faded over time. Sec-ond, not all
types of hunting were successfully reduced and the policy led to
increasedactive hunting through weapons and traps as households
respond by retaliating and over-extracting resources in fear of
completely losing access in the future. The e↵ectivenessof
community-based conservation on opportunistic hunting and bushmeat
purchase was
2
-
found to be heterogeneous based on income and education. Better
community integrationand dissemination of community conservation
design principles is therefore recommendedas it has proven to
e↵ectively reduce illegal hunting and also has the potential of
solvingthe retaliation and fear-based extraction behavior.
Furthermore, given that biodiversityis a global public good, local
users should not be the only bearers of conservation costsand
alternative livelihood strategies need to be introduced for the
long-run success ofconservation e↵orts.
3
-
Contents
Contents i
List of Figures ii
List of Tables ii
Acknowledgments v
1 Fire as an agricultural input: are the output gains worth the
mortality
impacts? A case study of Madagascar 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 1
1.2 Background: Madagascar’s fire problem . . . . . . . . . . .
. . . . . . . . . . 3
1.3 Health analysis data and methods . . . . . . . . . . . . . .
. . . . . . . . . . 6
1.4 Health results and discussion . . . . . . . . . . . . . . .
. . . . . . . . . . . . 9
1.5 Agricultural outcomes analysis data and methods . . . . . .
. . . . . . . . . 16
1.6 Agricultural outcomes results and discussion . . . . . . . .
. . . . . . . . . . 18
1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 21
1.8 Figures and tables . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 24
2 The role of inter-household transfers in coping with
post-disaster losses in
Madagascar 46
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 46
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 48
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 49
2.4 Estimation strategy . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 53
2.5 Results and discussion . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 54
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 58
2.7 Figures and tables . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 60
3 To hunt or not to hunt: evaluation of community-based
biodiversity con-
servation in northeastern Madagascar 70
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 70
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 72
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 75
3.4 Estimation strategy . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 77
3.5 Results and discussion . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 78
i
-
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . 83
3.7 Figures and tables . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 85
References 95
List of Figures
Figure 1.1: Study Area: Madagascar . . . . . . . . . . . . . . .
. . . . . . . . . . . 24
Figure 1.2: Average number of daily fires (January to March) . .
. . . . . . . . . . 24
Figure 1.3: Average number of daily fires (April to December) .
. . . . . . . . . . . 25
Figure 1.4: Area with observed fire (km2) . . . . . . . . . . .
. . . . . . . . . . . . 26
Figure 1.5: Monthly average of area with observed fire (km2) . .
. . . . . . . . . . 26
Figure 1.6: Fire pixels against yearly agricultural land . . . .
. . . . . . . . . . . . 26
Figure 1.7: Fire pixels against yearly livestock . . . . . . . .
. . . . . . . . . . . . . 26
Figure 1.8: Decomposition of fires by wind direction . . . . . .
. . . . . . . . . . . 26
Figure 1.9: Protected areas and parks in Madagascar . . . . . .
. . . . . . . . . . . 27
Figure 2.1: Madagascar’s full cyclone history . . . . . . . . .
. . . . . . . . . . . . 60
Figure 2.2: Madagascar’s cyclones during study period (1995 -
2010) . . . . . . . . 61
Figure 2.3: Number of cyclones observed per month since 1851 . .
. . . . . . . . . 62
Figure 2.4: Commune wind speed variation during cyclone Gafilo .
. . . . . . . . . 62
Figure 2.5: Cyclone exposure of Malagasy communes during the
period 1970 - 2010 63
Figure 3.1: Study area and villages . . . . . . . . . . . . . .
. . . . . . . . . . . . . 85
Figure 3.2: Reported obtention method for bushmeat consumed . .
. . . . . . . . . 86
Figure 3.3: Bushmeat consumption during study period . . . . . .
. . . . . . . . . 86
List of Tables
Table 1.1: Aerosol and fire summary statistics . . . . . . . . .
. . . . . . . . . . . 28
Table 1.2: Health summary statistics . . . . . . . . . . . . . .
. . . . . . . . . . . 28
Table 1.3: Regression of pollution on monthly count of fires . .
. . . . . . . . . . . 29
Table 1.4: Regression of pollution on monthly count of
confidence-weighted fires . 30
Table 1.5: Regression of pollution on monthly count of fires
weighted by confidence,
angle and distance . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . 30
Table 1.6: Regression of pollution on monthly count of fires,
quadratic specification 31
Table 1.7: Decomposing fires by di↵erent angles . . . . . . . .
. . . . . . . . . . . 31
Table 1.8: Regression of birth outcomes on monthly count of
confidence-weighted
fires . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 32
Table 1.9: Regression of birth outcomes on forest fires and
grassland fires . . . . . 33
Table 1.10: E↵ects of fires on cases of respiratory and
diarrheal diseases (All ages) . 34
ii
-
Table 1.11: E↵ects of fires on cases of respiratory and
diarrheal diseases (Children 0
- 5years) . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 34
Table 1.12: E↵ect of fires on cases of respiratory and diarrheal
diseases (Infants 0 -
12 months) . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . 35
Table 1.13: E↵ect of fires on cases of respiratory and diarrheal
diseases by type of
fire (All ages) . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 35
Table 1.14: E↵ect of fires on cases of respiratory and diarrheal
diseases by type of
fire (Children 0 - 5 years) . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 36
Table 1.15: E↵ect of fires on cases of respiratory and diarrheal
diseases by type of
fire (0 to 12 months) . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 37
Table 1.16: Agricultural production summary statistics . . . . .
. . . . . . . . . . . 38
Table 1.17: Reduced form impacts of all fires on area cropped .
. . . . . . . . . . . 38
Table 1.18: Reduced form impacts of all fires on quantity
harvested . . . . . . . . . 39
Table 1.19: Reduced form impacts of all fires on crop yield . .
. . . . . . . . . . . . 39
Table 1.20: Reduced form impacts of grassland and forest fires
on area cropped . . 40
Table 1.21: Reduced form impacts of grassland and forest fires
on quantity harvested 40
Table 1.22: Reduced form impacts of grassland fires and forest
fires on crop yield . 41
Table 1.23: Reduced form impacts of grassland fires by trimester
on yield and livestock 41
Table 1.24: First stage . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . 42
Table 1.25: IV impacts on cropped area (All fires) . . . . . . .
. . . . . . . . . . . 42
Table 1.26: IV impacts on cropped quantity (All fires) . . . . .
. . . . . . . . . . . 43
Table 1.27: IV impacts on crop yield (All fires) . . . . . . . .
. . . . . . . . . . . . 43
Table 1.28: IV impacts on cropped area (Grassland and forest
fires) . . . . . . . . . 44
Table 1.29: IV impacts on cropped quantity (Grassland and forest
fires) . . . . . . 44
Table 1.30: IV impacts on crop yield (Grassland and forest
fires) . . . . . . . . . . 45
Table 1.31: IV impacts on food prices (All fires) . . . . . . .
. . . . . . . . . . . . . 45
Table 2.1: Cyclone summary statistics (IBTrACS) . . . . . . . .
. . . . . . . . . . 63
Table 2.2: Cyclone summary statistics (constructed exposure) . .
. . . . . . . . . 64
Table 2.3: Household summary statistics . . . . . . . . . . . .
. . . . . . . . . . . 64
Table 2.4: Transfer summary statistics . . . . . . . . . . . . .
. . . . . . . . . . . 65
Table 2.5: Balance of household characteristics based on cyclone
exposure . . . . . 65
Table 2.6: Impact of cyclones on household well-being . . . . .
. . . . . . . . . . . 66
Table 2.7: Impact of cyclones on per capita transfers and
consumption . . . . . . 66
Table 2.8: Impact of cyclone dummy on household well-being . . .
. . . . . . . . . 67
Table 2.9: Impact of cyclone dummy on per capita transfers and
consumption (log) 67
Table 2.10: Impact of 90th percentile cyclones on household
well-being . . . . . . . 68
iii
-
Table 2.11: Impact of 90th percentile cyclones on transfers and
consumption per
capita (log) . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 68
Table 2.12: Impact of cyclones on transfers . . . . . . . . . .
. . . . . . . . . . . . 69
Table 2.13: Indirect impact of cyclones on household well-being
. . . . . . . . . . . 69
Table 3.1: Household summary statistics . . . . . . . . . . . .
. . . . . . . . . . . 87
Table 3.2: Hunting behavior summary statistics . . . . . . . . .
. . . . . . . . . . 87
Table 3.3: Relationship between GCF and ZOC with pre-policy
hunting trends . . 88
Table 3.4: Impact of GCF and ZOC policies on hunting behavior .
. . . . . . . . 88
Table 3.5: Household characteristics and bushmeat consumption .
. . . . . . . . . 89
Table 3.6: Impact of GCF and ZOC policies on quantity of
bushmeat obtained
opportunistically . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 89
Table 3.7: Impact of GCF and ZOC policies on quantity of
bushmeat purchased . 90
Table 3.8: Impact of GCF and ZOC policies on quantity of
bushmeat consumed at
a friend’s house . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . 90
Table 3.9: Impact of GCF and ZOC policies on hunting with traps
. . . . . . . . . 91
Table 3.10: Impact of GCF and ZOC policies on active hunting . .
. . . . . . . . . 91
Table 3.11: Heterogeneity of impact of GCF on all hunting . . .
. . . . . . . . . . . 92
Table 3.12: Heterogeneity of impact of GCF on opportunistic
hunting . . . . . . . . 92
Table 3.13: Heterogeneity of impact of GCF on purchased bushmeat
. . . . . . . . 93
Table 3.14: Heterogeneity of impact of GCF on hunting using trap
. . . . . . . . . 93
Table 3.15: Heterogeneity of impact of GCF on active hunting . .
. . . . . . . . . . 94
iv
-
Acknowledgments
This dissertation would not have been possible without the
invaluable support and guid-ance of many people whom I would like
to thank and recognize.
My three advisors, Maximilian Au↵hammer, Alain de Janvry and
Solomon Hsiang, haveprovided invaluable support, both
professionally and personally. They generously sharedtheir thoughts
regarding my academic work, and continuously supported me through
theentire process. I am also grateful to Elisabeth Sadoulet, David
Zilberman and Je↵rey Perlo↵for their suggestions and feedback
earlier in the process.
I am indebted to the participants in the Environmental and
Resources Economics seminar(Department of Agricultural and Resource
Economics) for their thoughtful feedback on mypresentations.
Specifically, I am deeply grateful to Brian Wright and Peter Berck
for theiruseful comments. Tamma Carleton, John Loeser, Edward Rubin
and Deirdre Sutula alsoprovided useful suggestions at various
stages.
The local data used in the first and second chapters were
generated by the Ministry ofHealth in Madagascar and by the
Malagasy National Institute of Statistics (INSTAT). Iam deeply
thankful for their cooperation and exceptional communication. The
local datain the third chapter was collected and generously shared
by Christopher Golden and theMadagascar Health and Environmental
Research organization (MAHERY). I am extremelygrateful for his
support and mentorship throughout the years and for his helpful
commentson the third chapter.
Financial support for the first chapter of this dissertation
came from the University ofCalifornia’s Graduate Division, the
Agricultural and Resources Economics Department andthe Ethan Ligon
Family Foundation. I acknowledge their generous support.
Finally, I am deeply grateful to my family for their unwavering
support and constantencouragement. This journey would not have been
possible without their endless generosity,support and love. They
have been and remain a source of inspiration and motivation for
me.
v
-
Chapter 1
Fire as an agricultural input: are
the output gains worth the mortality
impacts? A case study of Madagascar
1.1 Introduction
Fire is regularly used in agricultural practices around the
world and is a significant com-ponent of global fires (Korontzi et
al., 2006). While present in all continents (Stohl et al.,2007;
Chen et al., 2013; Zha et al., 2013), the highest volume of
agricultural fires isfound in developing countries. India and China
alone are responsible for forty percent ofglobal agricultural fires
and for twelve percent of global greenhouse gas (GHG)
emissions.This widely used agricultural practice has been sparsely
researched in economics, partlydue to the traditional lack of
reliable data on fires and partly due to the challenges
inidentifying the e↵ects of this anthropogenic activity. The
majority of existing studiesuse national-level data and focus on
national and global ecological, climatic and envi-ronmental
consequences of fires. Hence, to date, there is a paucity of
evidence on thesubnational and individual-level consequences of
fires as an agricultural input, an issuethat is important for at
least two reasons.
First, it is well-known that the agricultural sector is
predominant in developing coun-tries, in terms of national share of
labor, generation of income and livelihood reliance(Johnston and
Mellor, 1961). In low-income countries, forty to sixty percent of
nationalincome is typically produced in agriculture and fifty to
eighty percent of labor force isengaged in agricultural production.
However, it has also been shown that the large quan-tities of land
and labor committed to agriculture are used at low levels of
productivity.Maximizing profitability and increasing productivity
in agriculture is therefore essen-tial because it is a key sector
for development, thanks to its ability to enhance growththrough
product, factor and market contributions, to reduce poverty through
increasedfood security and better health and to provide
environmental services (Hirschman, 1958;Mellor et al., 1966;
Cipolla, 1976; Conway, 1998; Christiaensen et al., 2011; De
Janvryand Sadoulet, 2009). As a result, there is a great deal of
research tackling the issues ofimproving input e�ciency and
technology adoption in agriculture (Besley et al., 1994;Foster and
Rosenzweig, 1995; Duflo et al., 2008; Suri, 2011; Conley and Udry,
2010; Fos-ter and Rosenzweig, 2010; Cai et al., 2015). Yet, limited
work has been conducted ineconomics on rigorously understanding and
measuring the profitability and productivityof fire as agricultural
input.
Second, as a controversial farming technique, a discussion of
the use of agriculturalfires does not come without the recognition
of its negative impacts. To date, the discus-sion is largely
dominated by the important and relevant themes of loss of forest
cover,biodiversity and biomass; as well as global climate change
impacts (DeBano, 2000; Nasiet al., 2002; Pettus, 2009). There is
limited research and evidence on the local impacts of
1
-
agricultural pollution, a significant and unavoidable
consequences of agricultural fires. Itis naturally expected that
agricultural fires will lead to increased pollution, which in
turnis expected to have significant health impacts on surrounding
localities. General pollutionresearch shows that pollution leads to
increased in-utero mortality, lower birth rates andincreased infant
mortality rates with larger impacts in poorer communities (Chay
andGreenstone, 2003; Currie and Neidell, 2005; Arceo et al., 2016).
However, the externalvalidity of these results with regards to
developing countries is questionable, given thatlow-income
countries typically have di↵erent baseline levels of pollution and
health thanthe rich countries, where most studies are
conducted.
Evidence on the impacts of general pollution on health in
developing countries is stillsparse and results are mixed. Studies
show that, in India, the most successful pollu-tion regulation has
not led to any significant decline in infant mortality (Greenstone
andHanna, 2014). As for agricultural pollution studies, the few
existing studies seem to con-form with the general wisdom. A study
of wildfires in the Australian monsoon tropicsshows that one unit
increase in PM10 per cubic meter of air per 24-hour period due
towildfires leads to a 26% increase in daily asthma presentations
to the emergency depart-ment of the Royal Darwin Hospital, with a
threshold at 40 mg/m3 PM10 (Bowman andJohnston, 2005). Two projects
that are closest to my study use satellite data to investi-gate the
impact of fires on infant health. The first takes advantage of the
1997 massivewildfires in Indonesia and concludes that fires
contributed to 16,400 fewer survivals or a1% reduction of cohort
size especially in poor areas as inferred to missing children in
2000census (Jayachandran, 2009). The second evaluates the impacts
of smoke from sugarcaneharvest fires in Brazil on infant health and
finds that late-pregnancy exposure to upwindfires decreases birth
weight, gestational length and in utero survival (Rangel and
Vogl,2016). Furthermore, they find that non-upwind fires are
correlated with better infanthealth, highlighting the role of
agricultural fires in driving both pollution and
economicactivities.
Using the case study of Madagascar, a low-income country that
heavily relies on firein agriculture, my paper contributes to the
sparse literature on agricultural fires in devel-oping countries
and adds to the literature by extending the analysis on health
impacts tothe general population and by explicitly identifying the
economic gains from agriculturalfires in order to evaluate the
health versus wealth tradeo↵ of fires. While these e↵ects
areimportant in and of themselves, assessing the trade-o↵ is
essential in advising regulatorsin the appropriate direction and
costs of future fire regulations. Using high-resolutiondaily fire
and weather satellite data combined with monthly hospitalization
data andyearly agricultural data, I evaluate the health
consequences of agricultural fires on sur-rounding localities and
identify the e↵ects of fires land area used for agriculture,
harvestof main crops and yields. In order to identify the health
e↵ects of fire, I decompose firesinto downwind and upwind fires and
use wind direction on the day of fire as source ofexogenous
variation. The underlying assumption is that while both fires can
potentiallya↵ect health through income and liquidity e↵ects, upwind
fires will di↵erentially impacta↵ected populations through
pollution. To identify the impacts of agricultural fires
onagricultural outcomes, I take advantage of an international
conservation movement thatled to rapidly tripling the area of
protected parks in Madagascar. In particular, I use “pro-posed
parks”, areas that were physically delimited as potential o�cial
protected areas,as an instrument for fires. The assumption is that
the rapid creation of proposed parksled to a decrease in fires and
that the choice of proposed park is not otherwise relatedto
agricultural outcomes since they were driven by international
conservation trends and
2
-
given that no other changes were implemented besides the
physical delimitation of parkboundaries. To proceed with the
analysis, the remainder of this document is organizedas follows.
Section 1.2 describes the fire situation in my study area,
Madagascar. Section1.3 describes the data and methods for the
health analysis. Section 1.4 presents and dis-cusses the health
results. Section 1.5 describes the data and methods for the
agriculturaloutcomes analysis followed by a discussion of the
results in Section 1.6. Finally, Section1.7 concludes.
1.2 Background: Madagascar’s fire problem
1.2.1 Fire use in agriculture
Agricultural fires account for around ten percent of all global
fire activity (Korontzi et al.,2006) and can be used during the
pre-planting, harvesting and/or post-harvesting peri-ods. These
uses include clearing crop residue, fertilizing the soil,
eliminating pests andweed, wildfire prevention and pastoral
management. In Madagascar, one of many devel-oping countries with
high seasonal fire activity (Figure 1.1a), slash-and-burn
agricultureand pastoral management are the two main reasons for
agricultural fire use. Madagas-car, located in the Indian Ocean
east of Mozambique, is the fourth largest island in theworld, with
a land mass of 587,000km2 and 24.24 million inhabitants. It is a
renownedbiodiversity hotspot and a critical priority for
development and conservation e↵orts dueto a large overlap of
chronic poverty and unparalleled levels of endemism and species
di-versity. Madagascar has large but declining forest cover spread
across the island (Figure1.1b): every year, 25% to 50% of the
island’s grasslands and 7% to 10% of rainforestsare set aflame, yet
bushfires do not figure among the list of natural hazards as
thosefires are usually set by farmers and cattle herders for
agricultural purposes. Slash-and-burn agriculture is widely used in
developing countries, and refers to the process of firstpartially
clearing forests then burning the cut trees for the establishment
of plantations.This second phase of burning the remaining plants
leaves vegetable ashes that providenatural fertilizers leading to
three or four years of abundant crops followed by decliningsoil
fertility and yield. After this short period of high fertility,
nutrients leach out ofthe soil so that it becomes too poor for
agriculture. The cleared land is then left tofallow and farmers
typically need to repeat the process of slash-and-burn on new
forests(Gay-des Combes et al., 2017). While it is clear that
slash-and-burn agro-ecosystemsare central to livelihoods in poor
countries, there are mixed views on its sustainability.The most
optimistic view regarding slash-and-burn claims that it is
sustainable if used inan ecologically sound fashion (with
appropriate burning frequencies and fallow lengths)because it does
not require outside inputs but rather is based on natural elements
forfertilizers, pesticides and irrigation (Kleinman et al., 1995).
On the other end of thespectrum, some studies have suggested that
soil fertility decreases rapidly and remark-ably after the initial
release of plant nutrients and requires treatment with
herbicides,insecticides and fungicides to reduce root competition
and plant diseases. Furthermore,short-term success requires set up
of crop specific shelter-wood systems (Brinkmann andDo Nascimento,
1973). That is, while both sides agree that repeated use is harmful
inthe long run there is some disagreement in terms of how
successful slash-and-burn is inthe short run and recommendations
for appropriate cropping and fallow cycle lengthsvary from five to
ten years (Juo and Manu, 1996). Other studies have shown that there
is
3
-
significant variation in farmer e�ciency in slash-and-burn fires
and that this variation canbe attributed to human and social
capital, access to information and technology, credit,soil
fertility and environmental policies that might lead to under or
over exploitation ofa given plot (Binam et al., 2004).
Another common use of fire is in maintaining pasture for
livestock. This is referredto as pastoral fires and consists of
burning pastures just after (early dry season, EDSfires) or just
before (late dry season, LDS fires) the rainy season for greener
pasture andto prevent bush encroachment. Using fire for pasture
management is also complex asthere are tradeo↵s in choosing how
frequently to burn as well as when to burn. Frequentburning (yearly
or two-yearly) can drastically increase pastoral area by rapidly
depletingwoody cover but leads to poor pasture condition in the
long run. As for the timing offires, livestock farmers face a
di�cult trade-o↵ between short-term use of grass biomassfor grazing
and longer term use as fuel to manage tree–grass balances with
fire. Whileearly fires are recommended to reduce the chances of
late season wildfires, EDS wereassociated with declining pasture
condition due to the longer exposure to post fire grazingon early
burnt sites. Late season fires were most e↵ective for managing
woody cover whilstmaintaining higher pasture production and
perennial grass composition (Dyer and Smith,2003; Cowley et al.,
2014). Overall, it is clear that optimal use of fire depends on a
set oflocality-specific characteristics which include vegetation
type, land use, grazing intensityand the prevailing seasonal timing
and frequency of fire.
1.2.2 Fire regulations and politics
The century-old practice of burning has been a constant source
of conflict and tensionin Madagascar. On the one hand, a broad
group of conservationists and governmentalinstitutions believe that
the island sees too much fire and that fires should be stoppedas
they lead to deforestation, desertification, rangeland
impoverishment, soil degrada-tion and accelerated erosion. Wary of
these negative consequences, past colonial rulers,post-independence
leaders, the Malagasy Forest Service and international
environmentalagencies have used laws and information campaigns to
ban fire throughout the island,with a few exceptions. Prior to
colonization of Madagascar by the French in 1896, farm-ers and
herders managed fires locally through mutual understanding,
evolving traditions,and conflict resolution mechanisms overseen by
elders or royalty. In the 19th century,after several wars of
unification, Madagascar had its first recorded state-level
regulationof fires, but only regarding forests, which were seen as
a valuable source of timber (Dez,1968). These regulations became
much stricter and larger in extent during colonizationas the French
saw fires as a threat to natural resources (to protect productive
assets andto avoid soil loss and deforestation) as well as to the
stability of the colonial authority.Post-colonial bureaucratic
perceptions of fire in Madagascar were then deeply influencedby
these views so that strict anti-fire regulations remained. From
independence in 1960until today, anti-fire regulations have evolved
and hardened leading to more “theoretical”fire repression and
criminalization of fire. In practice, however, there are still many
moreillegal than authorized fires. According to the current fire
legislation, preventive fires andcounter-season fires are
authorized while burning is strictly prohibited during the
drymonths (April to November). As per Ordinance 60-127, infractions
would be punishedby a fine and/or six months to three years in
prison (Kull, 2016).
On the other hand, Malagasy farmers and herders rely on fire to
meet their livelihoodneeds, to manage resources, and are therefore
trapped by anti-fire laws. Thanks to the
4
-
nature of fires (fires can be set from a distance and can be
blamed on natural phenomena)and because of the weakness of the
state and its inability to enforce anti-fire laws, a veryhigh
volume of fire still persists across the island while pressures and
tensions betweenboth sides remain.
1.2.3 Fire patterns
As shown in Figures 1.2 and 1.3, which represent the daily
average number of fires percommune per month over the study period
(2000 - 2015), there is high fire activity acrossthe island with
considerable heterogeneity across space and time. The fire season
typicallystarts in April, which also marks the end of the rainy
season, in the south western regionsof Madagascar. These are early
dry season fires that are used for pasture management.Fire activity
increases as the year goes on and moves from south west to the
north east.Late dry season pastoral fires typically happen during
August and September and slash-and-burn fires are practiced around
October and November. Other and less commoncauses of fire in
Madagascar include charcoal production, hunting, customary
causes,criminal causes and natural causes (Kull, 2012; Styger et
al., 2007; Randriambelo et al.,1998). Fires in Madagascar are
seasonal and are closely linked to agricultural seasons.This
seasonality of fires is further illustrated in Figure 1.4 and
Figure 1.5. The blue linerepresents slash-and-burn fires which are
defined as a fire that occurs in a forested area.The orange line
represents pastoral fires which are fires that occur in the
grasslands. It canbe seen that both types of fires are seasonal but
there are more pastoral fires than forestfires in terms of area
burnt per month, which is consistent with Madagascar’s
vegetationcover as forests only account for ten percent of total
land area. Figure 1.5 shows thatboth early dry fires and late dry
fires are practiced but the latter is more widely used.Finally,
Figure 1.6 and Figure 1.7 plot fires along with the area of
agricultural land usedfor production and livestock production and
suggest that fires track economic activity,especially while looking
at agricultural land area. This is at most a correlation and
theexistence of any causal relationship between fires and
agricultural outcomes is yet to beestablished in this paper.
Fire is evidently important for the Malagasy economy, both in
terms of potential agri-cultural benefits, as well as in actual
uses. The subject has been and still is extensivelyresearched in
Madagascar, but through the lens of ecological and environmental
studies(Kull, 2012, 2004), biophysical analyses (Styger et al.,
2007; Clark, 2012) as well as geopo-litical essays (Randriamalala
and Liu, 2010). These studies show that anthropologic firesin
Madagascar are essential to farmers’ livelihoods but come at a high
environmentaland biodiversity costs (Kull, 2004). They further
argue that taking farmers’ reliance onfire into account while
drafting fire regulations would lead to drastic improvements inboth
farmers’ well-being and government’s resource use and that
criminalizing fires willonly lead to suboptimal and unsustainable
uses of fire. However, these policy recom-mendations are not
evidence-based as traditional studies of fire in Madagascar rely
onself-reported bureaucratic accounts that are plagued with missing
data and measurementerror problems. This paper provides some of the
first robust evidence assessing this claimfor Madagascar and even
adds to the local literature given that the discussion of
healthe↵ects of fires are virtually absent from the main debate.
The next section highlights mydata and methodology for the first
part of the analysis, which focuses on health.
5
-
1.3 Health analysis data and methods
1.3.1 Fire data
Data on wildfires and smoke are obtained from National
Aeronautic and Space Ad-ministration (NASA)’s Moderate Resolution
Imaging Spectro-radiometer (MODIS) At-mosphere Science Team.
According to NASA, this fire data from MODIS is the mostadvanced
global data product for fire monitoring but still builds on
heritage algorithmsfor operational fire monitoring from NOAA’s
Geostationary Operational EnvironmentalSatellites (GOES) and
Advanced very-high-resolution radiometer (AVHRR) sensors.
Inparticular, it consists of geo-spatial maps containing
information on the location andcount of fires per one
square-kilometer grid available at the daily level. Throughout
thehealth analysis section, the administrative unit of analysis is
the commune, which is thefourth administrative division in
Madagascar. There are 1,433 communes in Madagascar.On any given day
and for any given commune, I draw a 50km radius circle around
thecommune’s population centroid and match every observed fire
pixel to the correspondingwind data in order to di↵erentiate
between upwind and downwind fires. Fire summarystatistics are
reported in Table 1.1. On average, 12.61 fires per day and 72 fires
per monthare observed within a 50km radius of communes.
1.3.2 Pollution data
Smoke is measured using the aerosol optical depth (AOD)
variable, which comparesreflectance intensity in a particular band
against a reference value and attribute thediscrepancy to
particulates in the air. That is, it measures the degree to which
aerosolsor airborne particles such as windblown dust, sea salts,
volcanic ash, smoke from firesand pollution from factories, prevent
the transmission of light by absorption or scatteringof light. AOD,
as represented by MODIS’s maps takes on values between 0 and 1
wherea value of less than 0.1 indicates a crystal clear sky with
maximum visibility, while themaximum value of 1 indicates the
presence of aerosols so dense that the Sun would notbe visible even
at noon. Despite caveats, AOD is the next best when ground
measuresare not available (Jayachandran, 2009; Foster et al., 2009;
Gendron-Carrier et al., 2018).This paper constructs monthly
measures of pollution from daily observations of 3-kmgridded AOD
data. Aerosol summary statistics are reported in Table1.1. The
averagemean AOD is 0.09 compared to 0.20 for the whole continent of
Africa (Gendron-Carrieret. al, 2016) and the average maximum AOD is
0.34.
1.3.3 Wind data and other weather controls
Wind data consist of daily-averaged 23⇥12 - degree wind vectors
from NASA’s Modern-Era
Retrospective analysis for Research and Applications (MERRA)
database. Northwardand eastward wind vectors are used to compute
wind direction as angles in degrees andwind speed.
In all of my estimation, I control for temperature, rain and
their interaction. I usedaily gridded (1km2 grids) daytime and
nighttime temperature data from MODIS and14 ⇥
14 - degree rainfall data from NASA’s Tropical Rainfall
Measuring Mission (TRMM)
satellite.
6
-
1.3.4 Decomposing fire by wind direction
To construct the treatment variable, I first count the daily
number of 1km-by-1km gridcells containing a fire within a 50km
radius of each commune’s population centroid.Population centroid
data are obtained from the Gridded Population of the World
(GPW,version 4) database. To be conservative, I omit fires that are
within 5km as they arelikely to be within the urbanized areas and
might not be related to agriculture. Instead,I include them as a
control variable in all of my specifications. For each fire pixel
withinthe 50km radius, I compute the di↵erence between the bearing
of the pixel relative to thepopulation centroid and the wind
direction itself. A fire is classified as “Upwind fire” ifthe
absolute value of the di↵erence described above is less than 90
degrees, “Downwindfire” if the di↵erence is greater than 270
degrees and as “Other fire” otherwise. Figure1.8 illustrates this.
Alternatively, I also group the two latter categories such that
firesthat are not within the 90 degree quadrant are classified as a
“Non-upwind fire.
1.3.5 Birth outcomes data and hospital records
Data on birth outcomes and pulmonary diseases are obtained from
commune-level monthlyhospital records from the Malagasy Ministry of
Health. This dataset include all communesin Madagascar and contains
information on the number of consultations, diagnosis
andhospitalizations per month, broken down by diseases and by age
group. My primaryoutcomes of interest are birth outcomes (prenatal
consultations, total number of births,number of live births, birth
weight, and maternal deaths), respiratory diseases
(asthma,respiratory infection, pneumonia, respiratory diseases,
cough and cold), diarrheal diseasesand malnutrition. Unlike most
studies on infant health, I do not have data on gestationalage at
birth. Table 1.2 presents summary statistics for the health
dataset. The averagerate of live births per month is 972.57 per
1,000 births. Among live births, 86 per 1,000infants are born with
a low birth weight (
-
1.3.6.a Impact of fires on pollution
To test the key assumption that upwind fires raise pollution
more than downwind fires,I run the following regression using a
Spatial Fixed E↵ect model:
AODit = ↵Uupwindfires
it+↵Ddownwindfiresit +↵
Ootherfiresit +W0it�+ ⌧i + �t + uit (1)
The dependent variable AODit is the pollution concentration in
commune i in month tas measured by Aerosol Optical Depth. I look at
both average and maximum AOD sinceboth the average and the extreme
values of pollution are known to significantly impacthealth. In
this health specification, the main independent variables of
interest are theupwindfires
itvariable, the count of upwind fires, downwindfiresit the count
of downwind
and otherfiresit the count of other fires all measured during
month t in commune i.Results from this estimation are reported in
Table 1.3. The vector of covariates Witconsists of weather controls
that include monthly average temperature, humidity andtheir
interactions, minimum and maximum temperature and wind speed. These
weathercovariates are associated with both fire incidence and
pollutant concentrations but arenot di↵erentially associated with
fires by wind direction. ⌧i and �t are commune andyear-month fixed
e↵ects, respectively, so that this estimation captures
within-commune,within-time variation. As robustness checks, I also
use confidence-weighted, missingvalues adjusted and non-linear
expressions of the fire count variables.
Spatial Fixed E↵ects models are appropriate and necessary for
this paper’s analysisgiven the importance of spatial patterns of
fires, weather and health. Furthermore, dueto the seasonality of
fires and weather, it is also important to account for
temporalcorrelations in the model as well as in the standard
errors. To account for both spatialand temporal correlations, all
specifications use Conley standard errors, computed usingcode from
Hsiang (2010).
1.3.6.b Impact of fires on birth outcomes
My birth outcomes regression specification is similar to
equation (1) in terms of indepen-dent variables and controls, but
it uses a distributed lag-model since pollution exposurethroughout
pregnancy is expected to impact birth outcomes. Therefore, I
include ninemonth lags and run the following regression:
Birth outcomeit =9X
m=0
↵Umupwindfires
i,t�m +9X
m=0
↵Dmdownwindfiresi,t�m
+9X
m=0
↵Omotherfiresi,t�m +
9X
m=0
W 0i,t�m�m + ⌧i + �t + uit
(2)
where Birth outcomeit is the average of a specific birth outcome
reported in com-mune i during the month t, ⌧i and �t are commune
and year-month fixed e↵ects, and↵Um,↵D
mand ↵O
mare the distributed-lag versions of the coe�cients from
equation (1). To
follow the literature on in utero exposure, I focus on the three
trimesters of gestation andreport three-month coe�cients in Table
1.8. These coe�cients represent the impact ofincreasing fires by
one occurrence per month for three months. Given that I do not
havethe exact date of birth and that I am counting backward from
the month of birth, these
8
-
coe�cients do not exactly correspond to the three gestation
periods. An implication ofthis research design is that I implicitly
assume that all born babies are born at the ninthmonth of
pregnancy, that is, that no babies are prematurely born. This
assumption leadsto selection biases on the first and second
trimester coe�cients. The coe�cient on thelast trimester is more
reliable and can be assumed to be not impacted by this bias since
itis reasonable to expect that all babies were in utero for at
least three months. A centralassumption for identification in this
specification is that pregnant mothers living upwindfrom high-burn
areas are selected on characteristics relevant to infant health.
Given thefrequent change of fire source location and exogeneity of
wind direction, and the absenceof within-commune and between
commune migration, I argue that this is not a concernin this
analysis.
1.3.6.c Impact of fires on birth outcomes
Finally, I also look at the impact of fires on pulmonary
diseases incidence among thegeneral population as well as among
infants. My general health outcomes regressionspecification is
similar to equation (2) in terms of independent variables and
controls,but more flexible in terms of lags as di↵erent diseases
might have di↵erent incubationperiods and expected e↵ect:
Health outcomeait =MX
m=0
↵Umupwindfires
i,t�m +MX
m=0
↵Dmdownwindfiresi,t�m
+MX
m=0
↵Omotherfiresi,t�m +
MX
m=0
W 0i,t�m�m + ⌧i + �t + uit
(3)
where Health outcomeit is the average of a specific health
outcome for the age groupa reported in commune i during the month
t, ⌧i and �t are commune and year-monthfixed e↵ects, and ↵U
m,↵D
mand ↵O
mare the distributed-lag versions of the coe�cients from
equation (1).In both specifications from equations (2) and (3),
I also use an independent variable
that captures the di↵erential impact of upwind fires relative to
non-upwind fires. In this
setup, the reported coe�cient is ↵ =t̄P
m=t↵Um� ↵D
m� ↵O
mwhere t is the start month and
t̄ is the end month per aggregated period or trimester.
Therefore, I run the followingestimation:
Health outcomeait = ↵Firesit +W0i,t�m�m + ⌧i + �t + uit
(4)
1.4 Health results and discussion
1.4.1 Impact of fires on air pollution
Table 1.3 reports estimations of equation (1), showing that
upwind fires di↵erentiallyincrease pollution as measured by mean
and maximum AOD. The results from this table
9
-
use the simplest specification of fire, count of fires within a
month within a 50km radiusof the commune centroid, and the model is
augmented in later estimations by weightingby confidence, number of
missing pixels, angle and distance to the centroid. Columns(1) and
(2) report results for unweighted counts of fires, disregarding
their orientation tothe wind and both specifications exclude
weather controls. Columns (3) and (4) reportresults for unweighted
counts of fires, disregarding their orientation to the wind
andinclude weather controls. From column (1), one additional fire
in the past month raisesmean AOD by 0.00022 and raises maximum AOD
by 0.00104. In order to provide amore intuitive way of interpreting
the results, I interpret all further results in terms ofstandard
deviations rather than in terms of the regression coe�cients. That
is, a onestandard deviation increase in fires within 5 to 50km of a
commune centroid leads toa 0.45 standard deviation increase in
average aerosol optical depth and a 0.59 standarddeviation increase
in maximum aerosol optical depth. Surprisingly, fires within 5km
ofthe population centroid have no significant impact on average AOD
but significantlyincreases maximum AOD. From column (4), I find
that a one standard deviation increasein fires within 5km of the
population centroid raises maximum AOD by 0.011 standarddeviations.
When including weather covariates in columns (3) and (4), I find
that non-decomposed fires are no longer significant predictors of
average aerosol optical depth,however the coe�cients in the
regression of maximum aerosol optical depth are bothhighly
significant. A one standard deviation increase in fires within 5 to
50km of acommune centroid increases maximum aerosol optical depth
by 0.17 standard deviationsand a one standard deviation increase in
fires within 5km increases maximum aerosoloptical depth by 0.013
standard deviations.
In columns (5) through (8), I decompose the sum of fires into
upwind fires, downwindfires, and other fires within a 5 to 50km
radius of population centroids in a given month.In columns (5) and
(6), I also decompose the sum of fires within 5km of the commune
intoupwind fires, downwind fires, and other fires, whereas in
columns (7) and (8) they aresummed together since I am not treating
them as agricultural fires but include them ascontrols. This
di↵erence in specification does not influence the magnitude and
significanceof the main coe�cients of interest therefore remaining
results will use the specification incolumns (7) and (8). When
looking at the impact of fires on average aerosol optical depth,in
columns (5) and (7), I find that both upwind and downwind fires
significantly predictpollution while other fires do not.
Furthermore, the e↵ect of upwind fires on averageon aerosol optical
depth is approximately four to five times larger than the e↵ects
ofdownwind fires. A one standard deviation increase in upwind fire
raises average aerosoloptical depth by 0.08 standard deviations or
by 7.3 percent of the mean, whereas a onestandard deviation
increase in downwind fires raises average aerosol optical depth
by0.018 standard deviations or by 1.6 percent of the mean. The
coe�cient of upwind firesis significant at the 1% level, whereas
the coe�cient on downwind fires is significant onlyat the 10%
level. I find a similar trend while analyzing the impact of
downwind fires onmaximum aerosol optical index. A one standard
deviation increase in upwind fires raisesmaximum aerosol optical
depth by 0.11 standard deviations or by 9.0 percent of the
mean,whereas a one standard deviation increase in downwind fires
raises maximum aerosoloptical depth by 0.02 standard deviations or
by 1.9 percent of the mean. Other fires alsosignificantly impact
maximum aerosol optical depth. A one standard deviation increasein
other fires increases maximum aerosol optical depth by 0.025
standard deviations. Allthree coe�cients of interest in columns (6)
and (8) are statistically significant. Comparedto previous findings
in the literature, Rangel and Vogl (2016) also find that upwind
10
-
fires are associated with a four to five times larger increase
in PM10 concentration thandownwind fires, however the magnitude of
impact they find is much larger (an upwindfire raises PM10 by 0.16
standard deviations).
In the next set of regressions, I use confidence-weighted fire
counts. That is, each firepixel is weighted by the probability or
confidence that it is an actual fire as is documentedin the MODIS
dataset. Table 1.4 reports the results from this new fire
specification.Results are generally similar to what I have found in
the first set of regressions. Fromcolumn (1), I find that a one
standard deviation increase in fires within 5 to 50km ofthe
population centroid raises average aerosol optical depth by 0.01
standard deviationsand raises maximum aerosol optical depth by 0.18
standard deviations. A one standarddeviation increase in fires
within 5km of the population centroid raises maximum AOD by0.043
standard deviations but has no significant impact on average
aerosol optical depth.In columns (3) through (6), I decompose the
sum of fires within 5 to 50km of the communeinto upwind fires,
downwind fires, and other fires and I again find that decomposing
thefires within 5km or leaving them together does not change the
magnitude and significanceof my independent variables of
interest.
While my results on maximum aerosol optical depth are unchanged
in this second setof regressions, I find that downwind fires are no
longer a significant predictor of averageaerosol optical depth,
unlike other fires. From columns (3) and (5), I find that a
onestandard deviation increase in upwind fires raises average
aerosol optical depth by 0.07standard deviations and raises maximum
aerosol optical depth by 0.10 standard devia-tions, compared to
0.08 standard deviations and 0.11 standard deviations respectively
inTable 1.3. A one standard deviation increase in downwind fires
raises maximum aerosoloptical depth by 0.025 standard deviations
compared to 0.02 standard deviations in Table1.3. Finally, a one
standard deviation increase in other fires raises average aerosol
opticaldepth by 0.025 standard deviations and raises maximum
aerosol optical depth by 0.08standard deviations.
After weighting by confidence, I also weight the count of fires
by angle and distance ofthe fire to the population centroid. That
is, before summing up individual fires in orderto compute the
monthly count of fires, I first multiply each fire by the cosine of
the angleof its bearing to the population centroid and the wind
direction at the fire location (a firethat has the exact same
bearing to the commune as wind direction will be multiplied by
1,whereas a fire whose bearing is in the opposite direction of wind
direction will be appliedan angle weight of -1) and then divide by
distance squared. Results are reported in Table1.5 and are
generally consistent with earlier findings. From column (1), I find
that a onestandard deviation increase in fires within 5 to 50km of
the population centroid raisesaverage aerosol optical depth by 0.02
standard deviations and raises maximum aerosoloptical depth by 0.03
standard deviations. In columns (3) and (4), I decompose the sum
offires within 5 to 50km of the commune into upwind fires, downwind
fires, and other fires.The interpretation of coe�cients here is
slightly tricky, since now, all weighted upwindfires have a
positive sign (the smaller the angle between bearing and wind
direction, thecloser to 1 the weight is), all weighted downwind
fires have a negative sign (the smaller theangle between bearing
and the opposite of wind direction, the more negative the weightis)
and other fires can be either negative or positive. In columns (5)
and (6), I recodeangles to be positive and I again find that
results are consistent with earlier estimates. Ifind that a one
standard deviation increase in upwind fires raises average aerosol
opticaldepth by 0.065 standard deviations and raises maximum
aerosol optical depth by 0.11standard deviations, compared to 0.08
standard deviations and 0.11 standard deviations
11
-
respectively in Table 1.3. A one standard deviation increase in
downwind fires raisesaverage maximum optical depth by 0.04 standard
deviations. A one standard deviationincrease in other fires raises
average aerosol optical depth by 0.027 standard deviationsand
raises maximum aerosol optical depth by 0.05 standard deviations
compared to 0.025standard deviations and 0.08 standard deviations
respectively in Table 1.4.
Finally, in Table 1.6, I include quadratic terms to test for a
quadratic relationshipbetween fires and pollution. I find that the
magnitude and coe�cients of the linear termsare unchanged and
quadratic terms are significant but very close to zero, suggestinga
linear relationship between fires and pollution. A one standard
deviation increasein upwind fires raises average aerosol optical
depth by 0.06 standard deviations andraises maximum aerosol optical
depth by 0.08 standard deviations. A one standarddeviation increase
in other fires raises maximum aerosol optical depth by 0.02
standarddeviations and has no impact on average aerosol optical
depth. Downwind fires do notsignificantly predict average and
maximum aerosol optical depth. As a final robustnesscheck, I
decompose fires by di↵erent central angles and I find that my
results are generallyunchanged (Table 1.7).
These sets of estimation help me conclude that my results are
robust to di↵erentspecifications of fire and wind. Going forward, I
use the specification in Table 1.4, whichuses confidence-weighted
count of fires, as my preferred specification.
1.4.2 Impact of fires on birth outcomes
I now investigate the impact of fires on birth outcomes by
estimating equation (2). Resultsare reported in Table 1.8. In
particular, I look at the impact of exposure to upwindfires,
downwind fires and other fires during the three semesters of
pregnancy on cohortsize, fetal deaths, stillbirths, birthweight and
maternal deaths. I find that upwind fireslead to smaller birth
cohorts, fewer livebirths, more fetal deaths and more
maternaldeaths. Upwind fires in all trimesters significantly reduce
cohort size, and, surprisingly,the earlier the fire, the larger the
coe�cient. One additional upwind fire in the lasttrimester of
pregnancy decreases cohort size by 0.023%, whereas an upwind fire
duringthe first trimester of pregnancy reduces cohort size by
0.049%, with the largest and mostconsistent e↵ect being that on
cohort size. That is, a one standard deviation increase infirst
trimester upwind fires leads to a 0.86% decrease in monthly births
per commune, aone standard deviation increase in second trimester
upwind fires leads to a 1.57% decreasein monthly births per commune
and a one standard deviation increase in third trimesterupwind
fires leads to a 1.83% decrease in monthly births per commune.
Other fires alsoreduce cohort size, whereas downwind fires lead to
the opposite e↵ect. A one standarddeviation increase in other fires
during the last trimester leads to a 1.78% smaller birthcohort,
whereas a one standard deviation increase in downwind fires during
the lasttrimester leads to a 0.94% larger birth cohort. A one
standard deviation increase in“other” fires during the first
trimester leads to a 1.17% smaller birth cohort, whereas aone
standard deviation increase in downwind fires during the first
trimester leads to a1.18% larger birth cohort. When applying these
coe�cients to the a↵ected cohorts in mydataset, I find that, on
average, agricultural fires (all fires within 5 to 50km) lead to
12.62fewer births per commune per month. The e↵ect of fires on
birth size was negative for94.5% of commune-months and positive for
the remaining 5.5%. For commune-monthsin the latter category, these
occurrences correspond to months when there are manymore downwind
fires than upwind fires and other fires. On average, aggregating
for
12
-
a↵ected cohorts in the entire island, agricultural fires lead to
4,290 fewer births everyyear (CI = [3, 870 � 4, 710]). The
underlying assumption here is that pollution leadsto fewer
pregnancies and more fetal deaths. Column (2) reports the impact of
fires onfetal deaths and, while it shows that upwind fires increase
fetal deaths, all coe�cientsare extremely small in magnitude. This
is due to the very small mean value of fetaldeaths recorded at
clinics (µfetal
m= 0.0002, sdfetal
m= 0.16). Anecdotal evidence suggests
that most fetal deaths are likely to be unreported or even
unknown if they happen earlyenough in the pregnancy.
In column (3), I investigate the impact of fires on livebirths
per 1,000 births. Ifind that exposure during only the last
trimester significantly impacts livebirths. A onestandard deviation
increase in last trimester upwind fires leads to 1.11 fewer
livebirthsper 1,000 births per month and a one standard deviation
increase in other fires duringthe last trimester reduces livebirths
per 1,000 births by 0.93. Applying this to a↵ectedcohorts across
the nation, I find that, across the island, agricultural fires lead
to 98.87fewer livebirths per year (CI = [47.96 � 149.78]).
Therefore, on average, agriculturalfires in Madagascar lead to
4,389 “missing infants” per year or 1.95% of a↵ected cohortsand
0.7% of all births in Madagascar. These results are in line with
Jayachandran (2009)who finds that wildfires in Indonesia lead to a
1.1% reduction in birth cohort size ofa↵ected populations. I find
no significant impact on newborns’ probability of having abirth
weight lower than 2,500g, of being hospitalized after birth and on
maternal deaths.
Next, I look at the impact of fires on birth outcomes based on
the type of vegetationthat is being burnt (forest or grassland).
Results are reported in Table 1.9 and I alsofind that fires impact
cohort size and livebirths but conditional on surviving, there are
noimpacts on birthweight, hospitalization after birth and maternal
deaths. However, forestfires and grassland fires di↵erentially
impact cohort size. Forest fires during the lasttrimester of
pregnancy negatively impact birth outcomes, whereas grassland fires
in thefirst trimester of pregnancy positively predict health. A one
standard deviation increase inupwind forest fires relative to
non-upwind fires during the last trimester leads to a 0.54%decrease
in monthly births per commune, to a 0.006 standard deviation
increase in fetaldeaths and to 1.35 fewer livebirths per 1,000
births. A one standard deviation increasein upwind forest fires
relative to non-upwind fires during the first trimester leads to
1.17fewer livebirths per 1,000 births. Grassland fires in the two
last trimesters of pregnancyhave no significant impact on birth
outcomes, whereas a one standard deviation increasein upwind
grassland fires relative to non-upwind fires lead to 0.52% larger
cohort sizeand 1.15 more livebirths per 1,000 births. These results
suggest that forest fires are morepolluting than grassland fires,
as would be expected due to the higher mass of vegetationburnt and
the longer duration of forest fires, and that the income e↵ect from
grasslandfires might outweigh the pollution e↵ects as shown by
their positive impact on birthoutcomes. In the next section, I
investigate the relationship between fires and an arrayof diseases
related to respiratory health and nutrition.
1.4.3 Impact of fires on pulmonary disease
1.4.3.a Impact on disease prevalence
Results from the estimation of equation (4) are reported in
Tables 10 through 15. In Table1.10, I investigate the di↵erential
e↵ect of current and lagged fires on the prevalenceof several
diseases that might be associated with fires and smoke among the
generalpopulation. For this, I construct a variable defined as the
di↵erence between upwind fires
13
-
and non-upwind fires. Since both upwind and non-upwind fires can
be associated witheconomic activity, which might also impact
health, taking the di↵erence between the twowill isolate the impact
of pollution. I find that fires lead to higher prevalence of
respiratoryinfections, diarrheal diseases and malnutrition. Note
that I include diarrheal diseases asfires might also be associated
with water pollution if particles from fires are brought towater
sources by wind. I look at the impact on malnutrition to test for
nutrition e↵ects offires through increased production of food. I
find that a one standard deviation increasein upwind fires relative
to downwind fires in the current month leads to a 1.19% increasein
cases of respiratory diseases (asthma, pneumonia, respiratory
infections and cough)diagnosed per month per commune which is
mostly driven by the impact of fires onrespiratory infections
(columns (1) and (4)). Indeed, a one standard deviation increasein
upwind fires relative to downwind fires in the current month leads
to a 1.00% increasein cases of respiratory infections diagnosed per
month per commune. Column (3) showsthat a one standard deviation
increase in upwind fires relative to downwind fires in theprevious
month leads to a 0.85% increase in cases of pneumonia diagnosed in
the currentmonth. There does not seem to be any di↵erential impacts
of upwind fires on asthmaand cough. From column (6), I find that a
one standard deviation increase in upwindfires relative to downwind
fires in the current month (in the previous month) leads to a1.33%
(0.68%) increase in diarrheal cases diagnosed. This suggests that
fires also leadto water pollution as particles carried from fires
might end up in water sources. Finally,I also find that a one
standard deviation increase in upwind fires relative to
downwindfires in the current month leads to a 0.55% increase in
malnutrition cases diagnosed.
While these e↵ects are significant both statistically and
economically speaking, I ex-pect an even larger magnitude among
vulnerable populations as supported by the lit-erature on pollution
and health. To test this, I look at the impact of fires on
diseaseprevalence among children that are under five years of age
and report the results in Table1.11. Results are generally similar
to what was previously discussed, and as expected,children are
found to be more sensitive to pollution and e↵ects are more
persistent. I findthat a one standard deviation increase in upwind
fires relative to downwind fires in thecurrent (previous) month
leads to a 1.33% (0.85%) increase in cases of respiratory dis-eases
diagnosed. A one standard deviation increase in upwind fires
relative to downwindfires in the current (previous) month leads to
a 0.26% (0.36%) increase in cases of asthmadiagnosed. A one
standard deviation increase in upwind fires relative to downwind
firesin the current (previous) month leads to a 1.10% (0.85%)
increase in cases of respiratoryinfections diagnosed. A one
standard deviation increase in upwind fires relative to down-wind
fires in the current month leads to a 0.71% increase in cases of
common coughsbut fires during the previous month do not seem to
matter. In terms of non-respiratorydiseases, I find that a one
standard deviation increase in upwind fires relative to down-wind
fires in the current (previous) month leads to a 1.36% (0.65%)
increase in cases ofdiarrhea diagnosed. Finally, a one standard
deviation increase in upwind fires relativeto downwind fires in the
current month leads to a 0.65% increase in cases of malnutri-tion.
The magnitudes of impact here are slightly larger than those found
for the entirepopulation. I find no impact of fires on
pneumonia.
Finally, I look at the impact on infant health as reported in
Table 1.12. Resultshere are somewhere in between the two cases
discussed above. A one standard deviationincrease in upwind fires
relative to downwind fires in the current (previous) month leadsto
a 0.78% (0.59%) increase in cases of respiratory diseases
diagnosed. Fires are no longera significant predictor of asthma
prevalence. A one standard deviation increase in upwind
14
-
fires relative to downwind fires in the current month leads to a
0.52% increase in casesof pneumonia diagnosed. A one standard
deviation increase in upwind fires relative todownwind fires in the
current (previous) month leads to a 0.97% (0.59%) increase incases
of respiratory infections diagnosed. A one standard deviation
increase in upwindfires relative to downwind fires in the current
month leads to a 0.71% increase in casesof common coughs, a 0.84%
increase in diarrheal cases and a 0.39% increase in cases
ofmalnutrition diagnosed. E↵ects on infants seem to be smaller in
magnitude to those onchildren under five, this might be due to
lower smoke exposure of infants compared to olderchildren who can
wander near pollution sources. I no longer find a significant
impact onasthma, which might be in part due to the di�culty of
diagnosing asthma among infants.Indeed, mothers might keep infants
indoors during high smoke season, whereas childrenare more likely
to be outside even during high smoke season. Furthermore, compared
toan infant, a five year old would have a longer exposure to
pollution due to age, hencee↵ects might be larger in magnitude.
1.4.3.b Impact on disease prevalence by type of fire
In this section, I replicate the estimation above while
distinguishing between forest firesand grassland fires. Table 1.13
reports the impact of fires among populations of allages and finds
that most of the e↵ects found in the previous section were driven
byforest fires. A one standard deviation increase in forest fires
in the current (previous)month increases prevalence of pneumonia by
0.94% (1.30%), respiratory infections by0.84% (0.88%), diarrhea by
1.19 % and malnutrition 1.13% (1.14%). E↵ects of fire
onmalnutrition are large in magnitude and persistent. I find no
impact of grassland fires onany of the diseases. These results are
not surprising given that, all else equal in terms ofduration and
distance of fires, a larger mass of vegetation is burnt during
forest fires. Theimpact on diarrhea is expected to occur through
pollution of surrounding water sources.
In Table 1.14, I report the results on the impact of fires on
children that are un-der five years of age. I find that both forest
fires and grassland fires impact childrenhealth, however, impacts
of forest fires are larger in magnitude and more persistent,
es-pecially on asthma. This is not surprising given that forest
fires are more polluting andis consistent with several studies in
the health and pollution literature documenting in-creases in
asthma incidence as a result of increased pollution(Lee et al.,
2002; Trasandeand Thurston, 2005; Halonen et al., 2008;
Pénard-Morand et al., 2010). A one standarddeviation increase in
forest fires in the current (previous) month increases prevalence
ofasthma by 0.55% (0.81%), respiratory infections by 1.94% (1.20%),
cough by 0.97%, diar-rhea by 1.16 % and malnutrition 0.90% (1.31%).
E↵ects of fire on malnutrition are large inmagnitude and
persistent. As for grassland fires, they only impact respiratory
infectionsand malnutrition. Furthermore, negative e↵ects are only
found for the contemporaneousmonth. A one standard deviation
increase in grassland fires in the current month in-creases
prevalence of respiratory infections by 0.84% and malnutrition by
0.55%. A onestandard deviation increase in grassland fires three
months earlier decreases prevalenceof malnutrition by 0.62%.
In Table 1.15, I report the results on the impact of fires on
infants. I find significant andpersistent impact of forest fires on
the diseases of interest except for asthma. Grasslandfires only
negatively and significantly impact diarrheal diseases and
malnutrition. Aone standard deviation increase in forest fires in
the current (previous) month increasesprevalence of pneumonia by
0.74% (1.43%), respiratory infections by 0.84% (1.14%),
15
-
cough by 0.71%. Lagged forest fires from three months earlier to
the previous month leadto increases in diarrhea diagnoses. Finally,
a one standard deviation increase in forestfires in the previous
month leads to a 0.65% increase in malnutrition among
infants.Overall, these results show that both forest fires and
grassland fires lead to increasedprevalence of pollution-related
diseases but forest fires have a larger e↵ect. Adults seemto be
impacted by forest fires only, whereas children and infants are
also sensitive tograssland fires, which is expected to lead to less
pollution, consistent with the healthliterature regarding the
vulnerability of children and infants.
Now that it has been established that agricultural fires
negatively predict healththrough their e↵ects on fetal and infant
deaths as well as on pulmonary and nutritionrelated diseases, I
will investigate the flip side which predicts that agricultural
fires willincrease output in the long run and yield in the short
run.
1.5 Agricultural outcomes analysis data and methods
1.5.1 Agricultural production data
Data on agricultural production is obtained from the Ministry of
Agriculture and fromthe National Institute of Statistics.
Agricultural data mainly consists of yearly cropproduction data and
monthly food prices which are available at the district level,
thethird administrative division in Madagascar. Summary statistics
for the main food cropsas well as for sugar, vanilla and cattle are
reported in Table 1.16. The most commoncrops, by hectares planted
are rice, cassava and corn and are planted in all districts.Rice is
the staple food and cassava and corn are typically seen as inferior
substitutes.On average, a given district uses twelve thousands
hectares for rice, four thousands forcassava, three thousands
hectares for corn every year and harvests forty thousands tonsof
rice, thirty-one thousands tons of cassava and five thousands tons
of corn. Livestockproduction is also a significant part of
agriculture in Madagascar where the average yearlycount of cattle
is 2.5 million with a maximum of eleven millions in a given
district in ayear, or half of the Malagasy population.
1.5.2 Estimation strategy
In the absence of biases and causality challenges, estimating
the reduced form in equation(5) below would provide the causal
impact of fires on agricultural outcome.
Yict = ↵0 + ↵Fireit +Wit� + ⌧i + �t + uit (5)
The dependent variable Yit is a crop-specific outcome (area
planted, quantity har-vested and yearly yield) for commune i in
year t. The independent variable of interest isFireit, which is
defined as the sum of fires inside commune i during year t. The
vectorof covariates Wit consists of weather controls that include
yearly average temperature,humidity and their interactions, minimum
and maximum temperature and wind speed. Ialso control for pollution
as pollution might impact agricultural outcome either directlyor
through its health impacts on morbidity. ⌧i and �t are commune and
year fixed e↵ects,respectively, so that this estimation captures
within-commune, within-time variation.
However, there are many identification concerns with the above
specification. Thefirst and obvious issue is reverse causality,
given that these agricultural fires are moti-
16
-
vated by expected agricultural output gains. Second, atmospheric
and weather conditionsplay an important role in both agricultural
productivity and ignition and spread of fire.Finally, seasonality
in harvest, temperature and humidity may create spurious
relation-ships between agricultural fires and output. To overcome
these concerns, I use exogenousvariation from the process of rapid
tripling the areas of protected parks in Madagascar,a commitment
made during the 2003 Durban Action Plan, as part of an
internationalconservation trend.
During the Fifth International Union for Conservation of Nature
World Parks Congressheld in Durban South Africa in September 2003,
the then president Marc Ravalomananaannounced that Madagascar was
going to triple the amount of land under o�cial pro-tected status,
from the existing 1.7 million hectares to a total of 6 million
hectares or 10%of the country’s surface area by the year 2010. This
benchmark of protecting 10 percentof a country’s biome is Target 11
of the Convention on Biological Diversity, a treaty thatwas signed
by 150 government leaders at the 1992 Rio Earth Summit and is
currentlyachieved in approximately 55 percent of all terrestrial
eco-regions. While this ambitiousgoal was not achieved by 2010,
Madagascar successfully secured 10% of its land areaby 2011. Figure
1.9 illustrates this expansive creation of new parks between 2003
and2011. In the year following the announcement in 2004, local and
international conser-vation organizations physically mapped out and
delimited potential new parks based onan ecosystem approach taking
into account the overall and spatial distribution of biodi-versity
throughout the island and their current conservation status. Out of
the proposedlist, new parks were rolled out every year from 2005 to
2011 and not all proposed parkswere turned into o�cial protected
areas. I use proposed parks as an instrument for firesas specified
in equations (7) and (6) below. Actual parks are a good candidate
for arelevant instrument, however the exclusion restriction might
be violated as new parksmight impact agricultural output through
both reduced fire activity and tourism-inducedincreased economic
activity and access to information and technology.
Qit= �0 + �1 ˆFireit + �2Wit + ↵i + �t + uit (6)
Fireit = �0 + �1Proposed Parkit + �3Wit + ↵i + �t + vit (7)
The dependent variable in the second stage in equation (6) is
Yit, a crop-specificoutcome (area planted, quantity harvested and
yearly yield) for commune i in year tand Wit is the same vector of
weather covariates defined in equation (5). ⌧i and �t arecommune
and year fixed e↵ects, respectively. In the first stage described
in equation (7),the instrumented variable is Fireit, which is the
sum of fires inside commune i during yeart. The variable Proposed
Park
itis the area of potential parks that were proposed but not
implemented as o�cial parks in commune i in year t. In such
instances, park limits werephysically drawn with the intention of
creating a new park leading to de facto protectionand reduction in
fires but since there was no o�cial implementations, proposed parks
areexpected to a↵ect agricultural outcomes only through fires and
not through other factorsrelated to park creation such as better
road infrastructure, higher volume of informationthrough tourists
and higher economic activity.
17
-
1.6 Agricultural outcomes results and discussion
1.6.1 Reduced form
First, I report the results from the reduced form specification
in equation (5). Despitethe causality concerns discussed above,
working with the reduced form is useful becauseit allows to
understand crop-specific patterns and uses of fire. Tables 1.17,
1.18 and 1.19report the impact of all fires during the current and
three previous years on area planted,quantity harvested and yield
respectively for di↵erent crops. From the fire literature, Iexpect
that more fires would lead to more land area available for planting
hence moreagricultural output and a higher yield to due a more
fertile land post-burning (Gay-des Combes et al., 2017; Kleinman et
al., 1995). The coe�cients for the current yearand the year before
are either negative (cassava and vanilla) or not significant,
whereasthe two and three-year lag coe�cients are positive and
significant for rice only (Table1.17). Unsurprisingly, this
suggests a lag in the process of turning a newly cleared landinto a
field, which is especially true for rice as it requires irrigation.
Table 1.18 reportsthe impact of fires on quantity harvested and
show no significant impact except for anegative impact of
contemporaneous fire on quantity of vanilla harvested and a
positiveimpact of previous year’s fire on livestock production. The
results on vanilla are notsurprising given that vanilla is planted
in the forest, therefore there is a tradeo↵ betweenclearing forest
for land and planting vanilla. Contemporaneous fires might also
decreasethe quantity of vanilla harvested by burning of the plant
and crop destruction before itcan be harvested. Table 1.19 reports
the impact of fires on yield and is consistent with thetwo previous
tables. I find a negative association between recent fires and
yield of vanillaand no significant impact of recent fires on other
crops’ yield. Older fires are associatedwith increased corn yield.
From looking at these first results, it is clear that timing
offires matter and that impacts are di↵erent based on the nature of
the crop. To furtherinvestigate this, I split fires into grassland
fires and forest fires as I expect the earlier tomainly a↵ect yield
and livestock production and the latter to a↵ect area cropped.
Resultsare reported in Tables 1.20 through 1.22. From Table 1.20
column (1), which looks atthe impact of fires on total area
cropped, I indeed find no significant impact of grasslandfires and
a positive and persistent e↵ect of lagged forest fires. A forest
fire in the currentyear does not lead to increased cropped area in
the same year. Table 1.20 column (2),looks at the impact of fires
on area used for rice plantation, and results are reversedthere. A
grassland fire occurring two years before the year reported is
associated withincreased area used for rice. None of the forest
fire coe�cients are significant and thisis not surprising given
that grasslands are much more likely to be turned into rice
fieldthan forested areas. Table 1.20 column (3) looks at the impact
of fires on land area usedfor corn production and shows that
grassland fires are not associated with any increasedland planted
unlike forest fires. Contemporaneous forest fires are associated
with lessarea used for corn production, whereas forest fires
occurring two years before the yearreported is associated with more
land used for corn. Finally, contemporaneous forest firesdecrease
the area used for cassava and vanilla production, whereas lagged
grassland firesdecrease the area used for sugar and vanilla
production.
Next, I look at the relationship between each type of fire and
crop harvest, resultsare reported in Table 1.21. I find that lagged
grassland fires are associated with a higherproduction of cattle
and rice, and less vanilla harvested, which is consistent with
theimpacts on area cropped discussed earlier. On the other hand, I
find that lagged forest
18
-
fires are associated with higher quantity of corn and cassava
harvested but fewer cattleand vanilla. The negative impact of
forest fires on cattle production is surprising, butwould be
consistent with switching activities from herding to farming.
Finally, I look atthe impact of fires on yield and find that forest
fires have no significant impact on yield,whereas lagged grassland
fires are associated with higher corn and bean yield (Table1.22).
However, e↵ects are very small in magnitude. To further investigate
the impactof grassland fires on yield, I split grassland fires into
di↵erent trimesters of the year.Results are reported in Table 1.23.
I find that, for livestock production, as predicted bythe pastoral
fire literature, impacts are driven by late dry season burning
(July to Augustbefore the rainy season) and that there is no
significant impact of early dry season burning(April to June, which
is immediately after the rainy season). For rice and corn,
burningduring July to August and September to December both seem
e↵ective.
To sum up this section, it is clear that agricultural outcomes
are highly correlatedwith fires. Recent fires negatively impact
crops and yields, whereas longer term firesincrease area cropped,
quantity and yield. Area e↵ects are driven by forest fires,
whereasyield e↵ects are more associated with grassland fires. In
the next section, I report anddiscuss the results from the
instrumental variable approach to see whether these
observedpatterns are causal.
1.6.2 Instrumental variable
Table 1.24 reports the first stage described in equation (7).
Both o�cial and proposedparks lead to fewer fires and
unsurprisingly, the magnitude of e↵ects is larger for o�cialparks.
On average, the creation of a new national park is associated with
a 29% decreasein the number of all fires in the district, whereas
having new land proposed for parkcreation is associated with a 17%
decrease in all fire activity. For both o�cial andproposed parks,
the e↵ects are larger for grassland fires than for forest fires.
The creationof an o�cial park is associated with a 32% decrease in
grassland fires but only a 16%decrease in forest fires. Proposing
new land as a potential park is associated with a 23%reduction in
grassland fires and 8.5% decrease in forest fires. There are two
possibleexplanations that could be consistent with this. First,
forest fires might be harder tomonitor and regulate than grassland
fires if they are not located within park boundaries.Second, under
the assumption that forest fires are primarily used for clearing
land andthat grassland fires are used for livestock consumption and
the occasional pest controlthen the smaller reduction in forest
fires would imply that households have more inelasticdemand for