Can Workfare Programs Moderate Violence? Evidence
from India
Thiemo Fetzer
∗
October 18, 2013
Abstract
This paper seeks to answer the question, whether the Indian National Rural Employment Guarantee scheme, designed as social insurance, can mitigate the effect of
income shocks on insurgency violence in India and thus affect the dynamics of conflict. The paper proceeds by establishing that the violence processes are functions
of income or proxies of income. Exploiting variation within-district over time from
a newly created conflict dataset covering the whole of India, I show that the introduction of NREGA has reduced the rainfall dependency of conflict. This finding
is robust to many ways of studying the data. I then explore whether this finding
can be reconciled with the dynamics of NREGA take-up and the agricultural labour
market. I show that NREGA take-up is highly rainfall-dependant; furthermore, I
find some evidence that NREGA functions as a stabiliser to agricultural wages, in
light of rainfall shocks.
Keywords: social insurance, conflict, India, insurgency
JEL Codes: D74, Q34, J65
∗
London School of Economics and STICERD, Houghton Street, WCA2 2AE London. Support from
the Kondard-Adenauer Foundation, the Bagri Fellowship of the Asia Research Centre and the LSE-TISSTATA grant is gratefully acknowledged. I would like to thank Gerard Padro-i-Miquel, Maitreesh Ghatak,
Tim Besley, Claudia Eger, Jon de Quidt, Sam Marden, Jon Colmer, Mirko Draca, Christian Fons-Rosen,
Kelly Zhang, Benjamin Guin, Erlend Berg, Oliver Vanden Eynde and seminar participants at LSE and
CEP for helpful comments. Special thanks goes to Clement Imbert for kindly sharing some data.
1
1
Introduction
Can public work programs have a significant effect on the dynamics of violence in
Indian intra-state conflicts? This paper tries to find an answer to this question. It
is very challenging to find an environment in which the relationship between conflict
and social insurance can be studied, as typically, the two do not go hand in hand.
Either, conflict is too intense for the state to function properly and provide social
insurance or there is no conflict to study, while social security systems are aptly
functioning.
India serves as a unique testing ground to study this question. Firstly, the country has suffered from many low-intensity intra-state conflicts throughout its history.
All these conflicts are endemic, but have all in all a relatively low intensity so that
the state still functions on many dimensions. Secondly, India introduced from 2006
onwards a social insurance scheme through the National Rural Employment Guarantee Act (NREGA), that effectively serves as an insurance, providing employment
opportunities in times of need.
There are some appealing features of this program. It is the biggest public
employment scheme in mankind’s history, currently reaching up to 47.9 million rural
households annually, generating 210 million person-days of employment.
1
Hence,
it is reasonable to assume that the program, due to its scale, may have an impact
on the dynamics of conflict.
I make three contributions. This is the first paper to study the relationship
between insurgency violence and social insurance. In particular, the focus of this
paper is not on the levels of violence, but rather on the elasticity of violence with
respect to income and how this relationship changes after the introduction of the
workfare program.
The second contribution lies in studying the dynamics of rural labour markets in
India and how these are affected by social security systems. In particular, I am able
to comment on the pass-through of productivity shocks on real agricultural wages
and how this relationship is cushioned once a stable outside option offers itself to
workers. I thus focus on the insurance value of public employment that serves as a
income smoothing device and thus, may be a substitute to other forms of insurance,
without the underlying moral hazard problems.
The third contribution is a methodological one. This is the first paper to use a
novel violence dataset that covers the whole of South Asia and has been constructed
using scalable Natural Language Processing Tools (presented in Fetzer (2013)). This
1
See http://nrega.nic.in/netnrega/mpr_ht/nregampr.aspx, accessed on 14.02.2013.
2
paper highlights the possibility to use semi-automated machine-learning routines for
data cleaning and preparation in a field of economics research, where data availability
is always a severe constraint.
The key findings of this paper are as follows. First, violence in India is indeed
strongly income dependent. This is not a new finding, as it has been observed
previously by Vanden Eynde (2011) or Kapur et al. (2012). However, I am able to
show that this result holds up using novel remote-sensed rainfall data and a broader
conflict dataset.
Secondly, I show that the introduction of the workfare program has almost completely removed the income dependency of violence. This does not necessarily mean
that India has become a more peaceful place. In fact, Khanna and Zimmermann
(2013) provide some evidence from a regression discontinuity design that suggests
that immediately after the introduction, there has been an increase in the levels of
violence. This does not square with my results, since I provide evidence that suggests that one of the key drivers of violence has lost its bite. This finding is very
robust to different ways of studying the data and thus represents the core finding. It
ties in with a long literature that has studied the income dependence of conflict (see
e.g. Sarsons (2011); Miguel et al. (2004); Jia (2013); Burke et al. (2010); Vanden
Eynde (2011); Kapur et al. (2012); Dube and Vargas (2013)).
In the third step, I try to reconcile the stark finding on the impact of NREGA with
possible mechanisms that suggest why violence is income dependent. In particular, I
first show that NREGA take-up is highly rainfall dependent, suggesting that public
employment serves indeed as a form of insurance. Secondly, I show that agricultural
wages have become less dependent on agricultural productivity once NREGA is
introduced.
This suggests that NREGA did have a strong impact on rural labour markets in a
very distinct way. While the existing literature has focused on the programs impact
on wage levels (Imbert and Papp (2012); Berg et al. (2012); Zimmermann (2012)), I
show that NREGA stabilises real wages in light of productivity shocks. This suggests
two things: first, NREGA provides direct benefits when facing productivity shocks.
This may directly reduce incentives to join rebel movements. Secondly, the program
has indirect effects on labour markets when productivity shocks occur, moderating
wage volatility even for agents who do not participate in the program.
Both effects may help understand why violence has become less rainfall dependent. Social insurance can increase the opportunity cost of joining and supporting
rebel groups in times of adverse economic conditions and thus, they may help to
“win the hearts and minds”.
3
The paper is organised as follows. The second section provides some background
on the workfare program as well as on the literature. Section 3 gives details on the
data used in this study, while section 4 provides the empirical strategy and the main
results. Section 5 explores the extent to which the results can be reconciled with
NREGA participation as well as wage data. The last section concludes.
2
Background
2.1
Insurgencies and Social insurance
There are many small-scale insurgencies plighting India.2 The three main conflicts
are in the North East of India, mainly comprising the so-called Seven Sister States,
the Naxalite Insurgency that stretches through the Red Corridor across the East of
India, and thirdly, the conflict in Kashmir. The conflicts can be grouped roughly
into movements for political rights (e.g. Assam, Kashmir, Tamil’s and Punjab), for
social justice (the Naxalite conflict and the conflicts in the North East) and conflict
on religious grounds (such in Ladakh [Kashmir] or the various religious conflicts
between Muslim, Hindu and Christian groups all over India).
The intensity of these conflicts varies significantly over time. The Kashmir conflict has reduced in intensity significantly, while conflicts in the centre and the North
East continue unabatedly. The conflict between the Assamese separatists and the
Indian state has been on-going for more than forty years and has lead to a death-toll
in excess of 30,000.3 Concerning Naxalism, there exist no widely acknowledged data
on the number of casualties, but the conflict has intensified in recent years with 2010
being considered as one of the bloodiest years ever.4 It is difficult to study each conflict in isolation, as especially the conflicts in the East and North-East of India are
indeed related. The Indian Home Minister e.g. suggests that the northeastern state
of Assam has been emerging “as the new theater of Maoist groups”, with collaboration between United Liberation Front of Assam rebel groups collaborating with
the Naxalites. Hence, studying a conflict in isolation may be insightful, but fails to
capture possible broader underlying relationships.
The paper attempts to contribute to the literature on economic shocks and conflict. Since Miguel et al. (2004) a vast literature has emerged that seeks to identify
2
Nilakantan and Singhal (2012) provides some estimates of the economic cost of the Naxalite conflict
in the state of Andhra Pradesh.
3
See http://www.globalsecurity.org/military/world/war/assam.htm, accessed on 14.02.2013.
4
See http://www.google.com/hostednews/afp/article/ALeqM5hjfxXhNgGjp9JIyCj2ubaSKI6wIA?
docId=CNG.134eae01c393f94f33516bafd808dfc9.371, accessed on 02.04.2013.
4
the impact of income shocks on violence. This research has shown that incidence and
intensity of violence depend on incomes. Adverse income shocks increase violence,
whereas positive income shocks reduce it.
Underlying these studies are theoretical models that link income shocks with
violence through an opportunity cost channel. The theoretical frameworks typically
study a labour market consisting of two sectors, legal (wage) employment and illegal
insurgent activity. Such activities may have a broad range: from providing shelter
to insurgents, to supporting their activities by providing information or by actually
joining the insurgent groups.
This suggests that, if the opportunity cost of joining or support insurgent movements is relatively high, this should inhibit the capability of insurgent groups to
affect violence (and potentially also the type of targets insurgents chose to attack).
On the ground, this “opportunity cost” argument is widely acknowledged. The US
Army provides best-practice instructions for troops in the field on how to effectively
design counter-insurgency efforts.
Unemployed males of military age may join the insurgency to provide
for their families. Hiring these people for public works projects [...] can
remove the economic incentive to join the insurgency.
Social insurance can serve as a means to stabilise the outside option of workers,
in particular, in times when the outside option is particularly low, and thus, mediate
the relationship between income and violence. This question lies at the heart of this
paper.5
There is little other evidence on the interaction between social insurance and
insurgency. This paper attempts to fill this gap and tries to link up and expand
results from previous studies on the National Rural Employment Guarantee Act
(NREGA) in India, primarily the work by Johnson (2009), Berg et al. (2012) and
Imbert and Papp (2012).
5
Historically, the pacifying role of social insurance has been widely acknowledged. The social reforms
that Bismarck introduced in Germany were the carrot part of his agenda of “Zuckerbrot und Peitsche”.
It was introduced in the early 1880’s to weaken growing popular support for radical socialist movements
that threatened to overthrow the prevailing class system, while at the same time recognising that
The real grievance of the worker is the insecurity of his existence; he is not sure that he will
always have work, he is not sure that he will always be healthy, and he foresees that he will
one day be old and unfit to work.
There is some evidence by economic historians who found that the Bismarckian social welfare reforms significantly reduced German migration to the US (Khoudour-Castéras (2008)) as it weakened the
economic incentives to move. Using micro data from Sweden, Persson (2012) studies welfare-to-work programs and finds that they appear to have increased general crime, but not economically motivated crime,
in Stockholm.
5
2.2
The NREGA workfare program
The NREGA is a country-wide workfare program, which was passed as an act in
2005. It is the largest known workfare program, generating 2.76 billion person-days
of employment during the financial year 2010/2011. It stipulates that Indians in
rural areas are entitled to work 100 days per year on public projects. The program is
demand-led, so that the Gram Panchayat has to provide NREGA work if inhabitants
require such employment. This implies that - if participating in the program is only
attractive, when facing depressed outside options - we would not expect the program
to have a sudden impact on violence, but only through the insurance channel. Thus,
only when facing a negative income shock, should there be an effect of the program
on violence by increasing the opportunity cost of joining or supporting rebel forces.
The implementation of the program is very decentralised. The Gram Panchayat
sets up a list of projects. These typically can range anywhere between road construction, well digging, forestry or other forms of micro-irrigation. These projects could
thus provide two benefits: first, it could offer the workers a direct benefit through
the wage payments and secondly, it could have longer lasting impacts on agricultural
productivity.
The NREGA act further requires that 60 percent of the budget for a project be
allocated to wages. Also, the use of machines or contractors is prohibited. Workers
have to apply for work in NREGA projects by filling out a written application form,
after receipt of the application, the Gram Panchayat has to provide work within
two weeks after receipt of the application. If the panchayat fails to provide work,
a daily unemployment allowance (which is below minimum wage) is to be paid.
The projects on which workers are employed have to be in close proximity to the
home of the worker (at most 5 km distance) and there is additional renumeration
for transportation costs or living expenses, while on the work site.
The wages are to be paid by piece rate (depending on the nature of the project)
or through daily wage rate. These have to be above the state-level minimum wage.
The NREGA wages must be paid by cheque or by transfer to post-office or bank
accounts. In the financial year 2010-2011, expenditure on NREGA reached $ 7.88
billion, thus representing 0.5 per cent of Indian GDP.
This paper contributes to the growing literature that evaluates the NREGA workfare program. Several papers have found that NREGA lead to increases in agricultural wages (Zimmermann (2012), Berg et al. (2012),Imbert and Papp (2012)). I will
show that the stabilisation in agricultural wages takes place in case of a negative
rainfall shock - which corresponds to times when demand for NREGA employment
6
is found to be particularly high.
As with any other public works programme, NREGA has been criticised by many
stakeholders for its inherent inefficiency and susceptibility to corruption. Indeed,
Niehaus (2011) find evidence of widespread corruption in the system. They see
however, that the extent of corruption is moderated by increases in minimum wages.
There are a few papers that have studied NREGA take-up behaviour. Johnson
(2009) finds that take-up is highly seasonal and concentrated in the off-season. I
confirm his findings, but find evidence that this take-up is driven by rainfall shocks
in the preceding growing season, suggesting that agents do use NREGA to smooth
consumption when being faced by an adverse shock. This highlights the potential
consumption smoothing benefits from public employment (Gruber (1997)). I contribute to the growing literature on NREGA by combining these three observations
and linking them to the nature and path of insurgency violence in India. I next
discuss the data used in this study, before presenting the specification and the main
results.
3
Data
3.1
Violence
Empirical research on the economics of conflict almost always suffer from severe data
limitations. This lies in the nature of the subject of study, that typically places that
exhibit conflict are only weakly institutionalised with little official report of violence
and little press and media coverage. Blattman and Miguel (2009)’s review cites that
the correlation across different civil war datasets ranges from 0.42 to 0.96, which
may be the reason why empirical results are often not reproducible using similar
identification strategies, but different datasets or variable definitions (e.g. Ciccone
(2011)).
There exists no broad conflict dataset that covers India or South East Asia as a
whole. This gap was filled through the violence dataset introduced in Fetzer (2013).
This paper documents the process through which in the Indian context 28,638 newspaper reports were transformed into a workable conflict dataset using both machinelearning, semi-automated coding techniques and scalable manual hand-coding methods.6 This section sketches the semi-automated process through which the daily
6
The raw material was a set of 28,638 newspaper clippings collected by the Institute for Conflict
Management in New Delhi through the South Asian Panel on Terrorism (SATP) since 2001, see http:
//www.satp.org, accessed in October 2012.
7
newspaper clippings are transformed (more details are provided in Fetzer (2013)).
A typical sample may look as follows:
Two unidentified terrorists massacred six members of a family and left
a seventh injured at Mangnar Top, Poonch district, on December 31,
2001. Local residents refused to cremate the bodies of the slain victims,
insisting that a Union Minister should visit the area and take notice of
the increasing terrorist violence there.
The semi-automated routine defines a terrorist-incident as an Event-tuple, E =
{L, T, V, S, O} defined by a location L, a date or time of the event T , a verb V that
indicates the type of violent act, and the verb’s associated subject S, the perpetrator
of the act and the object O that was subjected to the act V . The semi-automated
routine tries to fill all these elements of the tuple for each sentence using common
machine-learning algorithms implemented in natural language processing packages.
In the above text-snippet, only one sentence satisfies the requirement of all elements being present, yielding:
object
verb
z }| { z
}|
{
massacred
six
members
of
a
family
Two
unidentified
terrorists
{z
}
|
subject
verb
object
verb
z}|{ z }| { z }| {
and left a seventh injured
time & location
z
}|
{
at Mangnar Top, Poonch district, on December 31, 2001.
which is transformed into:
E1 = {0 Mangar Top Poonch0 ,0 December 31 20010 ,
0
massacre0 ,0 two unidentified terrorists0 ,
0
six members of a family at Mangnar Top, Poonch district0 }
An incident is counted as long as all pieces of information can be deduced from
the underlying sentence. This is essentially mimicking the process through which
humans would code this data manually. An exhaustive list of verbs is used to spot
events and a sentence is normalised to contain at most one event. The individual
elements of the tuple E are then transformed by assigning labels to the snippets
indicating whether the actor was a terrorist, security force or a civilian and similarly
8
a
a
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
ong
ong
Figure 1: Spatial Dimension of Terrorist Attacks before 2006 (left) and after 2006 (right)
for who subjected to the act V .7
The data has been evaluated and correlates very well with hand-coded data.
The correlation between this automatically retrieved data and the hand-coded data
used by Vanden Eynde (2011) is at least 93%; however, the automatically retrieved
dataset covers the whole of India and most of South Asia.
The key dependent variable is the number of violent incidences per district and
quarter. This includes all incidences with at least one fatality, but also accounts
for general attacks on infrastructure, such as the destruction of telecommunication
masts, or attacks without fatalities. These are informative of the insurgents fighting
capacity, but are typically not included in other conflict datasets.
Figure 1 shows the log of the total number of incidences between 2001 and 2006
across India. It clearly highlights the three major conflict areas: first, the Naxalite
conflict, which stretches across India, in the so-called ”red corridor”.8 The second
source of major conflicts occur in the north-east, in the so-called ”7 Sister States”.
There, various insurgency outfits seek to obtain independence from the Indian Union.
7
Note
that
n
the
sentence
there
ex sts
a
further
event
E2
=
{ Mangar Top Poonch December 31 2001 eft
two un dent fied terror sts a seventh njured at Mangnar Top Poonch d str ct } As descr bed n Fetzer
(2013) a sentence w be counted as conta n ng nformat on of at most one nc dent
8
The states affected nc ude Andhra Pradesh Maharashtra Karnataka Or ssa B har Jharkhand
Chatt schargh and West Benga
9
300
Number of Incidents per Quarter
100
150
200
250
50
2000q1
2002q1
2004q1
2006q1
2008q1
2010q1
2012q1
Figure 2: Number of Terrorist Incidents over Time
The third major conflict is in the Kashmir region in the north west.
Figure 1 plots the spatial distribution and intensity of violence after 2006. It
becomes clear that violence seems to have intensified mainly in the ”red corridor”,
while it stayed the same in the ”7 sister states”, but was decreasing in Kashmir. The
intensification of violence, in particular in the Naxalite conflict and in the North East
has also been noted in anecdotal accounts, with 2010 being considered one of the
bloodiest years ever.9
For the main exercises of the paper, I will study India as a whole, but leave
out Kashmir, as this conflict has very strong inter-state dimensions. The empirical
specifications will include flexible conflict-specific time fixed effects in order to take
into account the specific dynamics these conflicts have.
The map suggested that violence levels were increasing over time, in particular
in the East and North East of India. This is confirmed in Figure 2, which plots the
time-series of recorded incidents in the area of study.
Aside from the novel violence dataset, I also invoke a new high resolution observational weather data obtained through remote sensing techniques techniques from
a novel precipitation radar, this is described in the next section.
3.2
Weather Data
This paper uses data from the Tropical Rainfall Measuring Mission (TRMM) satellite, which is jointly operated by the National Aeronautics and Space Administration
9
See http://www.google.com/hostednews/afp/article/ALeqM5hjfxXhNgGjp9JIyCj2ubaSKI6wIA?
docId=CNG.134eae01c393f94f33516bafd808dfc9.371, accessed on 02.04.2013.
10
(NASA) and the Japan Aerospace and Exploration Agency (JAXA). The satellite
carries a set of five instruments to construct gridded rainfall rates at very high spatial
and temporal resolution. It is the only available data source that subjects itself to
a systematic global verification program to validate the rainfall estimates using observational data from the ground. Due to the high spatial and temporal resolution
it is providing more consistent rainfall estimates than any other available ground
based observations and is considered the highest quality rainfall dataset with global
coverage that is currently available (Li et al. (2012)). Its adequacy to pick up the
spatial heterogeneity in precipitation has been highlighted and verified in the Indian
context by Rahman and Sengupta (2007), who have shown that it outperforms e.g.
the Global Precipitation Climatology Centre (GPCC) rain gauge analysis data that
has been used extensively in economics research.10
This is particularly relevant in the case of India, where observational weather
data is scarce and of potentially systematically varying quality. There are three main
drawbacks. First, most observations come from rain gauges, where measurements
are taken once a day. Climatologist are concerned about rain gauges in particular
in tropical- or subtropical areas, since most rainfall is convective. Such convective
rainfalls are highly local, generating intermittent and scattered rainfall, which may
not be picked up using rain gauges, if the network is not spatially fine enough.
The TRMM satellite orbits the earth every 90 minutes, thus providing multiple
observations each day. An alternative is to consider data from weather radars.
Rainfall radar may provide estimates for rainfall in a radius of 200 km around the
station, however it is unreliable for distances in excess of 200 km. In the Indian
case, rainfall radar data is not made available and would be problematic, since most
reporting radar stations are clustered along the coast. The third general concern
regarding observational weather data is the fact that reporting may be endogenous
e.g. to violence or other variables that are correlated with the dynamics of violence.
This has been found to be the case in hilly regions. This has been highlighted recently
by Smith et al. (2011), who show that Somalian piracy has generated a ”black hole”
in the Indian ocean, where observational weather data from merchant vessels is not
available anymore, as vessels take routes avoiding piracy infested areas.11
10
For example by Miguel et al. (2004), Ferrara and Harari (2012) and Kudamatsu et al. (2012). As this
data product comes at a coarse spatial resolution, researchers typically apply inverse distance weighting
methods to interpolate between grid-points. This typically underestimate the spatial variability or induce
variability where there is actually none (see Haberlandt (2007)). This will affect the estimation of extreme
values that these papers typically rely on for identification (Skaugen and Andersen (2010)).
11
See Besley et al. (2012) for a discussion. Another example is the case of Vanden Eynde (2011), who
had to merge several districts together in order to obtain consistent rainfall estimates, since many stations
simply fail to report rainfall estimates. Most of these stations are located in places with conflict or in
11
The TRMM Multi-Satellite Precipitation Analysis provides daily rainfall from
1998 to 2012 at a fine spatial resolution of 0.25 by 0.25 degree grid-cell size. The
data from the various instruments aboard the satellite are cleaned and calibrated
using additional data from the accumulated Climate Assessment and Monitoring
System (CAMS) and the Global Precipitation Climatology Centre (GPCC) rain
gauge analysis data. The output of the algorithm are 3-hourly rainfall rates for that
time-period. This is then scaled up to obtain monthly mean precipitation rates,
which in turn are transformed into overall monthly rainfall.
Any data product derived from remote sensing suffer from a problem known as
“error propagation” (see Leung et al. (2005)). As the raw data is transformed in the
analytical process, simple small measurement errors may be propagated due to the
mathematical and numerical transformations. As remote sensing products are using
some machine-learning algorithms to classify observations, the error propagation is
specific to the algorithm used for classification (see Burnicki et al. (2007)). In order
to take this problem into account, I obtain the PERSIANN precipitation dataset
that is based on the TRMM Multi-Satellite Data as well. This dataset is used to
confirm the main results of this paper and can address the potential issues regarding
error-propagation since the type of error propagation is specific to which algorithm
is used.12 Due to the fine spatial resolution, it gives on average 6 grid-cells per
districts. From this, I compute the monthly rainfall for this district.
In the next section, I detail the NREGA data used in this paper.
3.3
NREGA Data
The main specifications I will study are reduced form regressions highlighting how
the relationship between rainfall- and violence changes after the introduction of
NREGA.
In order do corroborate this reduced form approach, I establish that take-up for
NREGA is responsive to rainfall shocks as well. I use the NREGA participation
data derived from the so-called Monthly Progress Reports, which provide monthly
cumulative participation figures.13
Figure 3 shows a map of when the different districts across India were phased in.
It becomes clear that a lot of districts in the east of India were receiving NREGA
early on, as these are among the poorest districts.
newly created districts or states.
12
The TRMM data is from the algorithm 3B-43 (version 7), while the latest PERSIANN precipitation
data was obtained from http://chrs.web.uci.edu/persiann/ on 22.04.2013.
13
Refer to Appendix A.2 for further discussion of the available NREGA participation data.
12
The key concern for identification will consist of the endogeneity of the roll out
to violence. However, the relevant identifying assumption in this paper is that the
roll out was not endogenous to the previously existing relationship between rainfall
and conflict. As I do not attempt to identify a level effect of NREGA, this means
that the roll-out may well be endogenous to the levels of violence, without affecting
my identification. Furthermore, in the estimation I will specify the time fixed-effects
by phase of implementation, such that I do not estimate the effect of NREGA off of
variation across different phases.
30
3.0
lat
2.5
2.0
1.5
20
1.0
10
70
80
long
90
Figure 3: Phases of the NREGA Roll-out across India
3.4
Agricultural Wages
I will claim that the effect of NREGA operates through two main channels: first of
all, the direct provision of public employment reduces hardships caused by negative
income shocks. However, there is also a general equilibrium effect. This has been
observed in previous work by Berg et al. (2012),Imbert and Papp (2012) and Zimmermann (2012), who find that the introduction of NREGA had a level effect on
real wages. I expand this analysis to exploit the within-year weather variation.
13
In particular, I test whether NREGA stabilises wages using Agricultural Wage
Data from the Agricultural Wages in India (AWI) series which has been published
by the Indian Ministry of Agriculture since 1951. It is unique in offering monthly
wage rates by district (sometimes even containing multiple locations per district),
and separate wage series for several categories of labour and by gender. It remains
the most widely used source for analysis of Indian rural and agricultural wages even
though the quality of the data is questionable. I detail some of the issues with this
dataset in appendix A.1. I now proceed with the empirical specification and the
main results.
4
Empirical Specification and Main Results
4.1
Weather Shocks and Insurgency Violence
In the first step, I establish the reduced form relationship between rainfall and violence. I show that the estimated effects of weather variation and violence is significant
and appears to function mainly through an income channel.
I estimate the following specification:
0
E(Apcdt ) = δd exp (bpct + ηRdpt−1 + Xpdt π + cpdt )
(1)
where ad are district fixed-effects which capture time-invariant district features
that affect conflict, such as e.g. terrain ruggedness. This is particularly relevant as
there are some districts that had persistently high insurgency presence. The variable
bpct is a set of phase-and region-specific time fixed effects. These capture common
time-shocks that affect the districts belonging to each phase within a geographic
region. Making these time fixed-effects region specific is important, since, given the
size of India, there are likely going to be region-time specific shocks.14 Making the
time-fixed effects specific to a NREGA phase is especially important for the later
exercises. These fixed effects ensure that the variation I use to estimate η is coming
from within the different phases. This improves the robustness of the results to
potential omitted variables.15 The matrix Xdt contains district-level covariates, such
as measures of terrain ruggedness, elevation and other socio-economic characteristics
14
The geographic regions I consider are the states in the Red Corridor (Andhra Pradesh, Orissa, Bihar, West Bengal, Chhattisgarh, Jharkhand, Karnataka and Maharashtra). The states in the Northeast
(Assam, Meghalaya, Sikkim, Tripura, Mizoram, Nagaland and Manipur). The remaining states, mainly
in the west of India are contained in its own group.
15
I also study specifications with region-specific time fixed effects, the results do not change. These are
available from the author upon request.
14
taken from the 2001 census interacted with a full set of time-dummies or rainfall
controls that are included in some specifications.
The variable Rdpt−1 measures the log of rainfall in district d in the preceding
growing season. I focus on rainfall during the monsoon season, also known as the
“kharif season” as this rainfall is most relevant to agricultural production. The
income realisations of a Monsoon occur during the Monsoon but particularly after
the offset of the Monsoon, when agricultural produce is harvested. This motivates
studying the impact of the preceding agricultural season.
16
I define the Monsoon-period in two ways which are based on the principal crops
grown using the state specific Indian crop calendar.17 I use a narrow- and a broad
definition of the Monsoon period. This varies from state to state as the typical onset
dates are early May for the north east of India, while the onset may be as late as late
June for central India. For most states, the narrow Monsoon-period ranges from June
to September, while the broad ranges from May to November. The narrow Monsoon
period accounts for 80% of the annual rainfall, while the broad Monsoon period
covers roughly 90% of annual precipitation. For the main analysis I will use the
narrower definition that focuses on the key principal crops. The broader measure
is used for robustness checks. Using the lag of the rainfall in the growing season
alleviates some of the concerns that contemporaneous rainfall has a direct effect
on, e.g. the ability of insurgents to fight. From the perspective of an agricultural
economist, using the lags makes sense as well, as the harvesting season is mostly
in November and December for the most important Monsoon crops. Hence, income
realisations are at the end of a year. An additional argument in favour of using lags is
that it takes time for recruited insurgents to become actively supporting insurgency
activities.
As the dependent variable is a count variable, equation 2 is estimated using
conditional-fixed effects Poisson regressions (Cameron and Trivedi (1999)).18 The
above specification is estimated using a Poisson model to account for the count
nature of the dependent variable. I use a Pseudo Maximum Likelihood Poisson
16
A lot of other measures for rainfall have been proposed, and sometimes they render different results
in empirical analysis. Ciccone (2011) analysis of Miguel et al. (2004) highlights that serial correlation
induced through transformation of the dependent variable may drive some of their key results. I abstain
from performing crude transformations on the rainfall data due to the nature of the agricultural production
function described in the next section - however, I will show that my results are robust to using a set of
common transformations.
17
In particular the key reference is the crop specific calendar maintained by the Indian Food Security Mission, available via http://nfsm.gov.in/nfsmmis/RPT/CalenderReport.aspx, accessed on
12.05.2013.
18
The results are similar to using OLS. These are presented in the appendix in Table 8 and Table 9.
15
(PPML) estimator. I use the estimator implemented by Santos Silva and Tenreyro
(2006) as it overcomes some of the numerical problems in STATA’s implementation
of poisson regression Silva (2011).
A common concern for Poisson model is that of over dispersion. A negativebinomial estimator is a viable alternative, however, this estimator is not really a true
fixed effect estimator and suffers from the incidental parameters problem (Greene
(2004)). The PPML estimator does not require the data to have equi-dispersion. It
is consistent, so long as the conditional mean is correctly specified. The estimator
is even optimal if the conditional variance is proportional to the mean, hence over
dispersion is not an issue. Note further that conditional and unconditional likelihood
yield identical estimates, but typically the former is chosen as the computation is
quicker (Cameron and Trivedi (1999)). Silva and Tenreyro (2011) have shown that
the Poisson estimator performs very well, even in case of excess zeroes. The results
for this exercise are presented in Table 1.
Column (1) studies the baseline whereby the total of the preceding year’s rainfall is regressed on violence in the succeeding four quarters. It becomes clear that
rainfall matters strongly for conflict. The coefficients are interpreted as elasticity,
suggesting that a 1% increase in rainfall reduces conflict by 0.597%. The second
column separates out the effect of Monsoon-season rainfall and non-Monsoon season
rainfall. The effect of Monsoon-season rainfall is highly significant while the rainfall
outside the Monsoon appears not to matter for violence. In column (3) I estimate
the effect by NREGA implementation phase. There is some degree of heterogeneity
and it appears that the elasticity is strongest for districts in the second phase.
Columns (4) to (6) perform essentially the same exercises where now the dependent variable is the incidence of violence – that is a dummy that is equal to 1 in case
there was an incident in a given quarter in a district. This specification is estimated
using plain OLS. A very similar pattern emerges. A 1% increase in rainfall decreases
the incidence of violence by 0.026%. This effect is entirely driven by rainfall during the Monsoon season. The last column finds that the incidence of violence is
responding strongest to rainfall for districts that received NREGA in the first phase.
These regressions taken together suggest that there is a strong link between
rainfall and violence. However, it is unclear whether an income channel is actually at
work. The approach also highlights the differential effect of Monsoon season rainfall
and non-Monsoon rainfall. In order to substantiate these reduced form regressions,
in the next step, I will establish the relevance of these variables for agricultural
production and agricultural incomes.
I estimate an agricultural production function to support my reduced form ap-
16
Table 1: How does Rainfall Affect Violence?
Intensity
(1)
Base
log(Annual Rain)
(2)
Monsoon
Incidence
(3)
by Phase
-0.597***
(0.220)
log(Monsoon)
log(Outside Monsoon)
(4)
Base
(5)
Monsoon
(6)
by Phase
-0.026***
(0.007)
-0.713***
(0.181)
-0.024***
(0.006)
0.025
(0.129)
0.001
(0.002)
Phase 1 x log(Monsoon)
-0.836***
(0.257)
-0.049***
(0.014)
Phase 2 x log(Monsoon)
-2.278***
(0.371)
-0.036***
(0.013)
Phase 3 x log(Monsoon)
-1.375***
(0.492)
-0.006
(0.005)
District FE
Region x Phase x Time FE
Mean Attacks
Observations
Number of Groups
Yes
Yes
.768
9805
222
Yes
Yes
.768
9805
222
Yes
Yes
.754
9660
222
Yes
Yes
.0969
27693
543
Yes
Yes
.0969
27693
543
Yes
Yes
.0969
27693
543
Notes: The dependent variable is the number of terrorist incidences per district and quarter. Significance levels
are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in the parentheses are clustered at the district level.
Monsoon measures rainfall during the previous main agricultural growing season.
17
proach. In particular, I obtained data on agricultural production at district level for
all major crops, as well as farm-harvest prices at state-level. This data is used to
compute the agricultural GDP per capita at the district level.19
I estimate the following specification:
log(ydt ) = δd + bpct + ηRdt + cpdt
(2)
where Rdt is measuring monsoon rainfall. The results from these regressions are
presented in Table 2. The sample is constrained to those districts that are used for
the conflict estimation in order to establish the relevance for this area. Note that the
data is unbalanced. Columns (1) - (3) essentially do the same exercises as before.
It becomes evident that rainfall during the Monsoon season matters very strongly
for agricultural output. A 1% increase in annual rain increases agricultural GDP by
0.669%. The effect is driven mostly by rainfall during the Monsoon season, as can be
seen in column (2). Rainfall outside the Monsoon season matters as well, however,
the marginal effect is significantly smaller by a factor of four. This supports that
using the rainfall outside the monsoon season may serve as a reasonably well placebo
test. Column (3) explores the effect of Monsoon rainfall by NREGA implementation
phase. It becomes evident that agricultural GDP in districts that receive NREGA
first is most responsive to rainfall, as is expected, as these are the places that are
most underdeveloped.
The last two columns present results when just studying particular crops. Column (4) studies grain production, while column (5) considers the most important
cash crop. The effect of rainfall on cash crop production is a lot weaker, in particular in districts in the first phase. A comparison of column (3) and (4) suggests that
focussing on just grain crops (which has been done e.g. in Vanden Eynde (2011)) is
reasonable in particular for districts in the first NREGA implementation phase.
As mentioned before, many papers on the relationship between agricultural incomes and violence use different transformations on the rainfall variable. A common
form that these take is to considering only rainfall below or above a certain threshold
as constituting a negative productivity shock. In the case of India, the relationship
between rainfall and output is however fairly monotone. To highlight this, I estimated the production function using local-linear regression method developed by
Fan (1992). The results are depicted in Figure 4. This graph highlights that the
19
Note that the dataset is unbalanced as some states fail to report crop-production data throughout. I
focus on the main grain- and cash-crops grown in India. The major grain crops are: rice, wheat, bajra,
gram, jowar, maize, small-millets, pules, barley. The major cash crops are: cotton, jute, sugarcane and
tobacco.
18
relationship is fairly monotone. It becomes clear that a lack of rainfall is associated
with significantly lower agricultural output. One can not rule out that abundant
rainfall is correlated with lower agricultural production as well. However, it is unlikely that the existing rainfall data is able to pick up local flash floods sufficiently
-1
-.5
0
.5
1
well, as the spatial resolution is simply too coarse.
-1
-.5
0
xgrid
.5
1
Figure 4: Relationship between Rainfall and Agricultural Output
The reduced form regressions suggest that a low Monsoon rainfall is correlated
with significantly higher levels of violence. Furthermore, the analysis of the production function suggests that agricultural income is likely to be a driver behind the
reduced form results. I will now explore the extent to which the introduction of
NREGA has moderated the rainfall- and violence relationship.
4.2
Moderation of Violence through NREGA
In the next step, I explore whether the introduction of NREGA has moderated the
relationship between rainfall shocks and violence. I construct a dummy variable
Tdt

1 if NREGA available in district d at time t,
=
0 else.
that captures whether NREGA is available in a district d at a given time-point t.
I estimate the following specification:
0
E(Apcdt ) = δd exp (bpct + αTdpt + ηRdt−1 + γTdt Rdt−1 + Xdt π + cpdt )
19
(3)
Table 2: How does Seasonal Rainfall Affect Agricultural GDP
Agricultural GDP
(1)
Base
(3)
by Phase
(4)
Grain
(5)
Cash Crop
NREGA x Phase 1 x
log(Monsoon)
0.755***
(0.127)
0.777***
(0.083)
0.176
(0.112)
NREGA x Phase 2 x
log(Monsoon)
0.233***
(0.080)
0.139
(0.089)
0.053
(0.122)
NREGA x Phase 3 x
log(Monsoon)
0.470***
(0.117)
0.355***
(0.085)
0.277*
(0.146)
Yes
Yes
1812
188
Yes
Yes
2141
216
Yes
Yes
2141
216
log(Annualrain)
(2)
Monsoon
Crops
0.669***
(0.087)
log(Monsoon)
0.454***
(0.070)
log(Outside Monsoon)
0.117***
(0.042)
District FE
Region x Phase x Year FE
Number of Groups
Observations
Yes
Yes
1812
188
Yes
Yes
1812
188
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors are clustered
at the district level and presented in the parentheses. The sample is restricted to places where
agricultural GDP data could be constructed and are used for the main conflict specification, i.e
have experienced some conflict. Note that not for all districts, GDP data is available as some
states do not report price data.
20
The coefficient γ captures here the change in the functional relationship between
lagged rainfall and violence as NREGA is introduced. Note the timing in the regression specification. The effect of NREGA on violence is happening in the subsequent
four quarters, moderating rainfall shocks in the preceding agricultural season t − 1.
The argument for this choice of timing is very simple. As argued before, the full
extent of the income realisations occur towards the end of the year as the main harvesting season is in November and December. Furthermore, as will be shown further
below, take-up is particularly strong following a deficient Monsoon.
The inclusion of phase and region-specific fixed-effects implies that the variation
used to identify the effect comes from within phase-regions over time and thus, I can
not estimate the level effect as the dummy is perfectly collinear with the fixed effects.
The above specification identifies the effect of NREGA γ as long as the introduction
of NREGA was not correlated with the way that rainfall translates into violence
before the introduction of NREGA. This identifying assumption implies that the
above specification is identified even if the roll-out of NREGA was correlated with
levels of violence, as has been argued in Zimmermann (2012). The endogeneity of
the roll-out to the levels of violence however, means that identifying a level effect of
NREGA is very difficult.
The results from specification 4 are presented in table 3. Column (1) performs
the baseline studying rainfall throughout the year, while in the second column I
focus on Monsoon period rainfall. It is evident that the moderation of the rainfall
dependency is mainly operating through the Monsoon season rainfall. This will be
explored further in the robustness checks. The third column estimates the effect by
NREGA implementation phase. The effect is weakest for districts that were phasedin last and is strongest in magnitude for districts in the first NREGA implementation
phase.
The same pattern of results holds up when the dependent variable is the incidence
of violence. Here it becomes even more evident that the results are mainly driven
by districts in the first two NREGA implementation phases.
21
Table 3: The Effect of NREGA on the Relationship between Rainfall and Violence
Intensity
(1)
Base
log(Annual Rain)
NREGA x log(Annual Rain)
(2)
Monsoon
Incidence
(3)
by Phase
(4)
Base
-1.370***
(0.404)
-0.033***
(0.008)
1.136**
(0.489)
0.032***
(0.011)
(5)
Monsoon
log(Monsoon)
-1.316***
(0.285)
-0.033***
(0.007)
NREGA x log(Monsoon)
1.005***
(0.372)
0.032***
(0.010)
(6)
by Phase
22
Phase 1 x log(Monsoon)
-1.469***
(0.431)
-0.087***
(0.024)
NREGA x Phase 1 x log(Monsoon)
1.801***
(0.535)
0.090***
(0.034)
Phase 2 x log(Monsoon)
-2.048***
(0.669)
-0.048***
(0.018)
1.736**
(0.794)
0.046**
(0.022)
-1.177***
(0.449)
-0.007
(0.005)
-0.672*
(0.369)
0.004
(0.006)
NREGA x Phase 2 x log(Monsoon)
Phase 3 x log(Monsoon)
NREGA x Phase 3 x log(Monsoon)
District FE
Region x Phase x Time FE
Mean Attacks
Observations
Number of Groups
Yes
Yes
.768
9805
222
Yes
Yes
.768
9805
222
Yes
Yes
.768
9805
222
Yes
Yes
.0969
27693
543
Yes
Yes
.0969
27693
543
Yes
Yes
.0969
27693
543
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in the parentheses are clustered at
the district level. All rainfall measures are for the preceding year or preceding growing season.
All in all, these results suggest that the relationship between rainfall and violence
changes after the introduction of NREGA. This suggests that there is some effect of
the NREGA on the dynamics of violence. A key concern with the above specification however is, that the relationship between rainfall and violence may have been
changing over time, independently from the introduction of NREGA - i.e. there
could be time-specific changes to the way that rainfall translates into violence, that
are independent of NREGA, but may be picked up by the interaction term. In order
to address this concern, I estimate a very flexible specification, where I allow the
effect of rainfall on violence to be a different for each quarter of each year.20
The specification I estimate is
E(Apcdt ) = δd exp(bpct + αTdpt +
X
ηt Rdt−1
t
+
3
X
p=1
ηp Rdt−1 +
3
X
γp Tdpt Pp Rdt−1 + X0dt π + cpdt )
p=1
This specification still allows for the estimation of a phase-specific rainfall-effect
ηp and also a phase-specific NREGA effect γp , as the way that rainfall translates
into violence may be different across phases, which is not picked up by the simple
time specific effects ηt which are homogeneous across the three phases. The results
from this specification are best presented graphically.
Figure 7 plots the overall effect per district, which is the linear constraint:
ηˆt + ηˆp + γˆp Tpt . The blue dashed lines indicate the NREGA effect obtained from the
specification that does not control for common time specific rainfall effects. It becomes evident that this more demanding specification confirms the previous findings
that suggest that the relationship between rainfall and violence has changed systematically after the introduction of NREGA. The graphs for districts in the second and
third phase are presented in Appendix. They confirm that the effect is mainly driven
by districts in the first two phases of the NREGA implementation. I now present
a set of robustness checks that suggest that the NREGA effect is indeed robust to
many ways of studying the data. This is a prelude to the study of mechanisms
through which NREGA may be operating as moderating the income dependency of
conflict.
20
That is to say since the sample period is 2000q2 to 2012q4, I estimate 50 individual rainfall effect
coefficients.
23
-4
-2
0
2
4
Phase 1 Districts - Effect of Rainfall on Violence over Time
2000q3
2003q3
2006q3
Time
2009q3
2012q3
Figure 5: Plotting Coefficients of the Effect of Rainfall on Violence over Time
4.3
Robustness Checks
There are some natural robustness checks that need to be performed. They can be
crudely classified into three groups. First, a check of the robustness of the results to
adding more weather variables and interactions of the relevant rainfall variable with
respect to district characteristics, as these characteristics could be driving some of
the effects.
The second are some placebo-checks, i.e. shifting the reform a few years ahead
or studying rainfall that should not matter for agricultural incomes. The last set
of checks are robustness exercises regarding the measure of the “shock”. This is
important as has been highlighted by Ciccone (2011).
I perform these three types of checks. The results are presented in table 4. The
first column provides the base-results for reference.
Column (2) adds the rainfall characteristics interacted with some district controls, such as elevation, household size, scheduled-caste and scheduled tribe share of
the overall population and the literacy rate. The overall results remain the same. In
column (3) I add contemporaneous rainfall. There appears to be a negative coefficient of the rainfall dependency of contemporaneous rainfall as well. The coefficient
is however, significantly smaller and the results on lagged rainfall remain as they
were.
Columns (4) and (5) perform two placebo tests. In column (4) I move the reform three years ahead of time. The effect of rainfall on violence is still significant,
but there is no statistically significant “NREGA Effect” from this placebo reform.
24
Column (5) is the placebo test that essentially gives strong indication regarding the
underlying income channel as mechanism. The effect of rainfall outside the Monsoon season has only a marginal effect on agricultural incomes, and hence, we would
expect any NREGA effect operating through this rainfall to be marginal. This is
indeed the case. The signs of the coefficient on rainfall outside the Monsoon season
is consistent with the pattern on the baseline results, but the coefficients are far
away from gaining statistical significant and are significantly smaller in magnitude.
The last three columns study alternative measures of rainfall shocks. Column
(6) uses the deficiency from district mean. This yields very similar coefficients - as
expected - but is problematic to some extent as the district-specific means are not
long term means, since the rainfall data is only available from 1998 onwards. Hence,
these means are estimated with some error. Column (7) uses deviations from district
mean, where the deviation is normalised by the standard deviation. This essentially
transforms the deficiency measure used in column (6) into a z-score. The coefficient
pattern remains basically unchanged, though the interpretation of the coefficients is
less straightforward.
The last column performs an “IV” type analysis using the second remote sensing
rainfall data as an instrument for the TRMM rainfall data in order to address the
issues of error propagation discussed in the data section. The estimated coefficients
are very similar, though the NREGA effect is now just significant at 10%, though
the p-value is closer to the 5% significance level.
These results render me confident that I am capturing a genuine NREGA effect
that is mainly operating through an income channel. NREGA moderates the income
dependency of conflict by providing a stable and reliable outside-option in times of
need.
In the next section, I show that the observed NREGA effect can be reconciled
with NREGA take-up behaviour, which is itself, very responsive to Monsoon rainfall.
I also show that there is some evidence that NREGA moderates the pass-through
of rainfall shocks on agricultural wages, suggesting that there could be a general
equilibrium effect as well, through which NREGA moderates the income dependency
of violence.
25
Table 4: Robustness of the Effect of NREGA on the Relationship between Rainfall and Violence
Base
Interactions and Controls
Placebo tests
(1)
(2)
Rainfall
(3)
Current
(4)
Reform
log(Monsoon)
-1.316***
(0.285)
-1.845***
(0.254)
-1.627***
(0.250)
NREGA x log(Monsoon)
1.005***
(0.372)
1.412***
(0.331)
1.520***
(0.293)
26
log(Monsoon)
(contemporaneuous)
0.098
(0.226)
NREGA x log(Monsoon)
(contemporaneuous)
-0.654**
(0.277)
(5)
Rain
(6)
Deficiency
(7)
Norm
(8)
IV
-0.951***
(0.348)
1.586***
(0.313)
-0.192***
(0.046)
1.653***
(0.481)
0.302
(0.367)
-1.202***
(0.348)
0.117**
(0.058)
-1.063*
(0.544)
Yes
Yes
No
.768
9805
222
Yes
Yes
No
.768
9805
222
Yes
Yes
No
.779
9232
219
log(Outside Monsoon)
-0.206
(0.233)
NREGA x log(Outside
Monsoon)
0.189
(0.295)
District FE
Region x Phase x Time FE
Rainfall x Controls
Mean Attacks
Observations
Number of Groups
Yes
Yes
No
.768
9805
222
Yes
Yes
Yes
.754
9660
219
Yes
Yes
No
.768
9805
222
Robustness to Measures
Yes
Yes
No
.768
9805
222
Yes
Yes
No
.768
9805
222
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in the parentheses are clustered at the district level.
5
NREGA and Economic Outcomes
The following sections try to reconcile the finding that NREGA reduced the income
dependency of insurgent violence by studying two margins. First, if NREGA moderates the income-dependency of violence by reducing the amount of labour that is
available to support insurgent activities, then this should be reflected in the take-up
behaviour of the employment guarantee.
In the second step, I study agricultural wages in general and their responsiveness
to seasonal rainfall variations. If NREGA serves as a stabiliser, it should moderate
the dependency of agricultural wages on agricultural productivity - in particular of
negative productivity shocks.
Both of these affect the local labour market, which may reduce the availability
of labour for insurgency activities.
5.1
NREGA Take-up Behavior
I proceed by studying NREGA take-up behaviour. It has been argued that NREGA
may serve as a substitute for weather insurance, as it provides a stable outside option
in times when agricultural incomes are otherwise depressed due to low rainfall. This
has been observed using village-level NREGA participation data for a few years in
Andhra Pradesh by Johnson (2009). This section thus is an extension of their work,
as it studies broader areas of India.
I obtained district-level NREGA participation data from the so-called Monthly
Progress Reports.21 I study two margins of adjustment – the overall response that
captures the number of days demanded by households, as well as the extensive
margin response, which captures the share of households in a district participating
in the program.
I estimate the following specification:
Ppcdt = δd + bpct + ηηRdt−1 + pcdt
(4)
where Ppcdt measures the participation in the program in district d of phase p
in region c at time t. I use annual NREGA participation data, which is based on
financial calendar year commencing each April. The results from this specification
is presented in table 5.22 It becomes evident that high rainfall realisations in the
preceding growing season have an adverse effect on NREGA participation, both on
21
22
A discussion of the available data sources and some limitations is presented in Appendix A.2.
Appendix Table 10 present very similar results studying the whole of India.
27
the intensive and the extensive margin. A 1% increase in Monsoon rainfall reduces
participation by 0.142%. The elasticity of rainfall with respect to employment days
is 0.486. The extensive and intensive margin response are strongest for districts that
have received NREGA in the first two phases. This is consistent with the findings
from the previous sections.
The rainfall outside the Monsoon has a moderately positive and statistically
significant effect on NREGA participation. This effect is only contemporaneous,
which is unlikely to capture anything related to incomes.
In order to corrobate these findings on take-up, I estimate the above specification
using the monthly participation data. This allows me to plot the effect month-bymonth. The results are depicted in figure 6.
Figure 6: NREGA Participation and Monsoon Rain
The shaded area depicts the Monsoon season from June to September. The
effect of Monsoon rainfall on NREGA participation is only statistically significant
and negative from August onwards - as the Monsoon begins to withdraw. The
coefficients are negative for May already. This is due to the fact that the Monsoon
onset may be earlier from year to year. The key feature is that the effect stays
negative for the month following the Monsoon, but becomes gradually weaker over
time.
These results together suggest strongly that NREGA participation may be a key
driver behind the observed weakening elasticity between rainfall and violence across
28
India. I now study whether NREGA also serves as a stabiliser to agricultural wages
and thus, may have an indirect effect on violence by reducing the rainfall dependence
of agricultural wages and thus, incomes.
Table 5: How does Rainfall Affect NREGA Take-Up?
Extensive
(1)
overall
log(Monsoon) (lagged)
(2)
by phase
-0.142***
(0.043)
Intensive
(3)
overall
(4)
by phase
-0.486***
(0.113)
Phase 1 x log(Monsoon) (lagged)
-0.212***
(0.060)
-0.531***
(0.130)
Phase 2 x log(Monsoon) (lagged)
-0.071
(0.099)
-0.374*
(0.205)
Phase 3 x log(Monsoon) (lagged)
0.006
(0.087)
-0.472
(0.352)
log(Monsoon) (contemporaneuous)
log(Outside Monsoon) (contemporaneuous)
log(Outside Monsoon) (lagged)
-0.044
(0.043)
-0.051
(0.043)
-0.206
(0.132)
-0.206
(0.129)
0.042***
(0.015)
0.045***
(0.016)
0.112*
(0.059)
0.111*
(0.058)
-0.007
(0.016)
-0.004
(0.015)
-0.003
(0.033)
-0.002
(0.033)
Yes
Yes
.415
1007
216
Yes
Yes
.415
1007
216
Yes
Yes
15
1006
216
Yes
Yes
15
1006
216
District FE
Region x Phase x Year FE
hline Mean of the DV
Observations
Number of Groups
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in the parentheses
are clustered at the district level. The samples are restricted to those districts that were used in the
violence regressions. The panel is naturally unbalanced due to the phase-wise roll-out.
5.2
Impact of NREGA on Agricultural Wages
Previous studies have focused on the impact of NREGA on the level of agricultural
wages. The most prominent contributions are Imbert and Papp (2012), Berg et al.
(2012), Azam (2011) and Zimmermann (2012). All of these papers try to identify the
causal effect of the workfare program on wage levels. This paper argues that NREGA
29
essentially works as a social insurance, providing an outside option in times when
agricultural productivity is depressed. It is this employment channel, through which
insurgencies are recruiting new rebels to expand its fighting capacity. If NREGA
works as a social insurance, one should see that it stabilises agricultural wages, when
these would otherwise be low - i.e. in light of productivity shocks.
In this section I will show that this is indeed the case - the effect of NREGA
on agricultural wages is particularly pronounced in times, when agricultural productivity would otherwise be depressed. This effect is on contemporaneous wages,
i.e. NREGA stabilises contemporaneous wages. At first sight, this may be seen as
being at odds with the results obtained in the previous section on NREGA participation, which depended strongly on the Monsoon seasonal rainfall in the preceding
growing season. However, as NREGA merely provides a binding outside option,
agricultural wages may adjust upwards in light of low rainfall realisation, without
increasing NREGA participation directly. This suggests that landlords hiring agricultural labour bear the financial burden of providing insurance, and not necessarily
the state.
The specification I study are very similar to the ones in section 4.1 and 4.2. In
particular I estimate:
log(wpcdt ) = δd + bpct + ηRdpt−1 + Xpdt 0 π + tpst + cpdt
(5)
where wpcdt measures the log of real-daily agricultural wages. The nominal wages
are adjusted using state-level price indices for agricultural workers. I also control
for a state-phase-specific time-trends. This becomes necessary as wages are growing
very rapidly, but the growth differs from one state to the other and within NREGA
implementation phase. This has also been observed by Berg et al. (2012).
I present results for the whole dataset for which there is wage-data.23
The exercise proceeds very much in the fashion of the previous steps. Column (1)
in table 6 suggests that contemporaneous rainfall is a driver of agricultural wages.
All coefficients are interpreted as elasticities. Since the coefficient on on contemporaneous rainfall is 0.044, this implies that the pass-through of productivity shocks
to agricultural wages is limited.
The second column depicts the NREGA effect; the introduction of NREGA has
removed the rainfall-dependency of agricultural wages. This thus suggests that
NREGA serves as a stabiliser to agricultural wages, as expected. Columns (3) and
23
Constraining the sample to only include districts in which there is conflict does not yield any results.
This is most likely due to very poor data quality for these districts. The issues with the data are detailed
in A.1.
30
(4) are placebo-checks. In column (3) I study the role of rainfall outside the growing
season; NREGA appears to not have an effect coming through rainfall falling outside
the Monsoon season. Column (4) is a simple placebo-check, moving the introduction
date of NREGA back by three years. The last two columns can be seen as additional
robustness checks.
In particular, in column (5) I constrain the labour activities to be included in
the calculation of the annual real wage to be only activities that are likely relevant
for the harvesting and planting. These are the ”reaping and harvesting” wages
and the ”weeding” wages. The latter activity only becomes relevant if there is
sufficient rainfall and the fields need to be regularly cleared of any other weeds
growing aside the agricultural crops. The planting activities include the ”sowing”
and ”ploughman” wage categories. These activities are relevant at the onset or even
before the onset of the Monsoon, when the intensity of the Monsoon rainfall is not
yet known.
These results taken together suggest that there appears to be a general equilibrium effect of NREGA on agricultural wages that is correlated with rainfall. This
supports the notion of this paper that NREGA may serve as effective insurance
against adverse shocks and thus, reduce volatility in incomes, which may have induced agents to join or indirectly support rebel movements.
31
Table 6: How does NREGA Affect the Rainfall Dependency of Agricultural Wages?
32
(1)
Baseline
(2)
NREGA Effect
(3)
Placebo (1)
(4)
Placebo (2)
(5)
Harvesting
(6)
Planting
log(Monsoon)
(contemporaneuous)
0.044**
(0.017)
0.067***
(0.019)
0.044**
(0.017)
0.065**
(0.028)
0.055**
(0.021)
0.031
(0.024)
log(Monsoon)
(lagged)
0.010
(0.016)
0.003
(0.016)
0.012
(0.016)
0.010
(0.015)
0.012
(0.017)
-0.006
(0.018)
log(Outside Monsoon)
(contemporaneuous)
-0.001
(0.008)
-0.000
(0.008)
0.003
(0.010)
-0.002
(0.008)
0.003
(0.010)
-0.017*
(0.010)
log(Outside Monsoon)
(lagged)
0.012
(0.008)
0.009
(0.008)
0.012
(0.008)
0.011
(0.008)
0.017*
(0.010)
0.002
(0.008)
-0.065***
(0.024)
-0.039
(0.028)
Yes
Yes
Yes
4.1
1987
280
Yes
Yes
Yes
4.15
1987
280
NREGA * log(Monsoon)
(contemporaneuous)
-0.082***
(0.024)
NREGA (Placebo) *
log(Outside Monsoon) (contemporaneuous)
-0.009
(0.011)
NREGA (Placebo) *
log(Monsoon) (contemporaneuous)
District FE
Region x Phase x Year FE
State-specific Phase trend
Mean of the DV
Observations
Number of Groups
-0.030
(0.027)
Yes
Yes
Yes
4.11
2496
338
Yes
Yes
Yes
4.11
2496
338
Yes
Yes
Yes
4.11
2496
338
Yes
Yes
Yes
4.11
2496
338
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in the parentheses are clustered at the district level.
6
Conclusion
This paper has set out to investigate the impact of the NREGA Workfare Program
on the dynamics of violence in Indian intra-state conflicts. I find that the income
dependency of violence has decreased significantly following the introduction of the
public employment scheme, suggesting that one of the key drivers of insurgency
violence can be moderated through the effective introduction and provision of social
insurance.
The paper shows that the observed NREGA effects are plausible when studying
NREGA participation data; furthermore, there appears to be a general equilibrium
effect of NREGA on agricultural wages as well - stabilising wages when these otherwise would be depressed due to adverse weather conditions.
References
Azam, M. (2011). The Impact of Indian Job Guarantee Scheme on Labor Market
Outcomes: Evidence from a Natural Experiment. SSRN Electronic Journal .
Berg, E., S. Bhattacharyya, R. Durgam, and M. Ramachandra (2012). Can Rural
Public Works Affect Agricultural Wages? Evidence from India. CSAE Working
Paper .
Besley, T., T. Fetzer, and H. Mueller (2012). The Welfare Cost of Lawlessness:
Evidence from Somali Piracy. mimeo.
Blattman, C. and E. Miguel (2009). Civil War. NBER Working Paper 14801 .
Burke, M., J. Dykema, D. Lobell, E. Miguel, and S. Satyanath (2010). Climate and
Civil War: Is the Relationship Robust?
Burnicki, A. C., D. G. Brown, and P. Goovaerts (2007, May). Simulating error propagation in land-cover change analysis: The implications of temporal dependence.
Computers, Environment and Urban Systems 31 (3), 282–302.
Cameron, A. C. and P. K. Trivedi (1999). Essentials of Count Data Regression.
mimeo, 1–17.
Ciccone, A. (2011). Economic shocks and civil conflict: A comment. American
Economic Journal: Applied Economics (February).
33
Dube, O. and J. Vargas (2013). Commodity price shocks and civil conflict: Evidence
from Colombia. Review of Economic Studies.
Fan, J. (1992). Design-adaptive nonparametric regression. Journal of the American
Statistical Association 87 (420), 998–1004.
Ferrara, E. L. and M. Harari (2012). Conflict , Climate and Cells: A disaggregated
analysis. mimeo.
Fetzer, T. R. (2013). Developing a Micro-Level Conflict Data Set for South Asia.
(2010), 1–17.
Greene, W. (2004, December). Fixed Effects and Bias Due to the Incidental Parameters Problem in the Tobit Model. Econometric Reviews 23 (2), 125–147.
Gruber, J. (1997). The Consumption Smoothing Benefits of Unemployment Insurance. American Economic Review 87 (1), 192–205.
Haberlandt, U. (2007, January). Geostatistical interpolation of hourly precipitation
from rain gauges and radar for a large-scale extreme rainfall event. Journal of
Hydrology 332 (1-2), 144–157.
Imbert, C. and J. Papp (2012). Equilibrium Distributional Impacts of Government
Employment Programs: Evidence from India s Employment Guarantee. mimeo.
Jia, R. (2013). Weather Shocks , Sweet Potatoes and Peasant Revolts in Historical
China. Economic Journal .
Johnson, D. (2009). Can Workfare Serve as a Substitute for Weather Insurance?
The Case of NREGA in Andhra Pradesh. Centre for Micro Finance Working
Paper (32).
Kapur, D., K. Gawande, and S. Satyanath (2012). Renewable Resource Shocks and
Conflict in India’s Maoist Belt. mimeo.
Khanna, G. and L. Zimmermann (2013). Guns and Butter ? Fighting Violence with
the Promise of Development. (January).
Khoudour-Castéras, D. (2008). Welfare state and labor mobility: the impact of
Bismarck’s social legislation on German emigration before World War I. Journal
of Economic History 7.
34
Kudamatsu, M., T. Persson, and D. Strömberg (2012). Weather and Infant Mortality
in Africa. mimeo.
Leung, Y., Y. Ge, J. Ma, and J. Wang (2005). Measurement Errors and their
Propagation in the Registration of Remote Sensing Images. In Developments in
Spatial Data Handling, pp. 285–297. Springer Berlin Heidelberg.
Li, X.-H., Q. Zhang, and C.-Y. Xu (2012, March). Suitability of the TRMM satellite
rainfalls in driving a distributed hydrological model for water balance computations in Xinjiang catchment, Poyang lake basin. Journal of Hydrology 426-427,
28–38.
Miguel, E., S. Satyanath, and E. Sergenti (2004). Economic shocks and civil conflict:
An instrumental variables approach. Journal of Political Economy 112 (4).
Niehaus, P. (2011). Corruption Dynamics : The Golden Goose Effect .
Nilakantan, R. and S. Singhal (2012). Naxalite Insurgency and the Economic Benefits
of a Unique Robust Security Response .
Persson, A. (2012). The effect of welfare-to-work programs on crime: Evidence from
a natural experiment.
Rahman, H. and D. Sengupta (2007). Preliminary Comparison of Daily Rainfall
from Satellites and Indian Gauge Data. mimeo.
Santos Silva, J. and S. Tenreyro (2006). The Log of Gravity. The Review of Economics and Statistics 88 (4), 641–658.
Sarsons, H. (2011). Rainfall and conflict. mimeo (October).
Silva, J. M. C. S. (2011). poisson : Some convergence issues. mimeo, 1–8.
Silva, J. S. and S. Tenreyro (2011). Further simulation evidence on the performance
of the Poisson pseudo-maximum likelihood estimator. Economics Letters 44 (0),
1–8.
Skaugen, T. and J. Andersen (2010, July).
Simulated precipitation fields with
variance-consistent interpolation. Hydrological Sciences Journal 55 (5), 676–686.
Smith, S. R., M. A. Bourassa, and M. Long (2011, July). Pirate attacks affect Indian
Ocean climate research. Eos, Transactions American Geophysical Union 92 (27),
225.
35
Vanden Eynde, O. (2011). Targets of Violence: Evidence from India’s Naxalite
Conflict. mimeo (November), 1–50.
Zimmermann, L. (2012). Labor market impacts of a large-scale public works program: evidence from the Indian Employment Guarantee Scheme. mimeo.
36
A
Appendix
A.1
Agricultural Wages in India
This appendix describes the process of how the agricultural wage data was cleaned
and put in shape for the analysis in the paper. The data is of variable and sometimes
questionable quality.24
The raw data gives monthly wages for male, female and children, broken into
skilled- and unskilled agricultural labour and different types of labour. The types of
skilled labour are blacksmith, carpenter and cobbler, while unskilled labour combines
ploughman, reaper/harvester, sower, weeder, other agricultural labour. In some
states, these separate unskilled labour categories are not reported25 , but rather, a
category “Field Labour Wages” is reported. This is conceived to be an average of
the different categories.
In some districts these wages are reported throughout the year, while in others
the wages are reported only in the parts of the year, when particular activities are
actually carried out (i.e. sowing wages in the early Kharif season of May, June and
July), while harvesting wages are reported in the fall of a given year.
After digitising and entering the raw data, I proceed to construct a quarterly
level agricultural field-labour wage as my main dependent variable.
For each district, there may be multiple wage-observations in case there are
multiple reporting centres. I generate a balanced panel requiring each quarter of the
year to have at least one non-missing observation of agricultural wages belonging
to the particular category of unskilled labour. I then construct the simple average
across these wage-observations.
There are advantages and disadvantages to this approach. In particular, by
construction, this implies that within a year, some field labour wage observations
are noisier then others. This can be taken into account by adequately weighting the
observations.
As an alternative, I can impose the requirement that there be at least one observation for each different unskilled labour category within a quarter. This condition
is very stringent, as it fails to recognise the types of agricultural activities that are
pursued during a year. This approach reduces the number of districts significantly,
but the results remain the same.
The southwest monsoon typically enters the mainland over Kerala in the first
week of June. It moves northward to cover the whole of India by mid-July. It
24
25
Issues are...
The states for which this is the case are Andhra Pradesh, Karnataka and Maharashtra.
37
starts withdrawing from mid-September. The southwest monsoon is critical to the
development of Indian agricultural production. The southwest monsoon provides 80
percent of India’s total precipitation and is critical to the development of its major
food and commercial crops such as rice, coarse grains, pulses, peanuts, soybeans and
cotton. Planting of the largely rainfed Kharif (monsoon season) crops, which include
rice, sorghum, corn, millet, peanut, soybean and cotton will begin after the monsoon
firmly establishes itself over the major producing states and planting will continue
through July and early August. Farmers in the northern rice surplus states of Punjab
and Haryana, where irrigation is available, often complete rice transplanting prior
to the monsoon arrival.. This season’s pre-monsoon, or early season rains in central,
south and east India should provide a favorable early season planting conditions
for rice, soybeans, sorghum and corn. The country’s economy is to a large extent
dependent on monsoon rains.
A.2
NREGA Data Sources
There are two main sources for data on NREGA take-up. These are the district-level
monthly-progress reports (MPR) and data coming from the Management Information System (MIS). The latter is a completely non-paper based system that has only
become mandatory to use in the financial year but was still not fully operational
until 2010-2011.
There are a lot of issues regarding the reliability of either datasets, as there is
quite some mismatch between the two datasets, especially in the earlier years when
the MIS was introduced.26 This may be due to partial compliance in the MIS after it
had been introduced, but could be also because the MPR system is more subject to
manipulation. It is difficult to asses the underlying divergence in the two databases.
Only the MPR data is available continually from 2006 onwards, hence this paper
uses this data.27 The format of the MPR reports has changed over time, making it
difficult to define a set of variables that are consistent for the whole time period. I
focus on two main variables: monthly-cumulative person days provided and monthly
cumulative number of (distinct) households provided employment both at district
level. The former is very informative about the overall level participation and the intensity of participation in each month, which can be reconstructed from the monthly
totals. The latter is indicative about the extensive margin of take-up, yet it does
not allow the construction of the number of distinct households that were active
26
See for example mismatch between MIS data and National Sample Survey returns data highlighted
by http://www.indiatogether.org/2013/jun/gov-nregs.htm,accessed on 12.06.2013.
27
Thanks to Clement Imbert for sharing NREGA take-up data for the earliest years
38
in any particular month due to the cumulative nature. Hence, extensive margin
empirical exercises can only be performed at annual level and not at monthly level.
Furthermore, intensive margin responses, such as the number of days per household
are difficult to characterise, as take-up towards the end of the financial year (say
February) is physically constrained to the 60 days left in the financial year.
I construct the NREGA take-up and participation data to match the Monsoon
calendar as in the main exercises. In particular, I define a Monsoon year as lasting
from October to the following October. This implies that I can study the effect of
the contemporaneous Monsoon as well as the effect of the preceding Monsoon on
NREGA take-up, thus matching the main empirical exercises.
A.3
Additional Robustness
Table 7: How does Rainfall Affect Violence? OLS Results
Intensity
(1)
Base
log(Annual Rain)
(2)
Monsoon
(3)
by Phase
-0.104***
(0.036)
log(Monsoon)
-0.110***
(0.031)
log(Outside Monsoon)
-0.005
(0.010)
Phase 1 x log(Monsoon)
-0.151**
(0.067)
Phase 2 x log(Monsoon)
-0.252***
(0.075)
Phase 3 x log(Monsoon)
-0.111*
(0.066)
District FE
Region x Phase x Time FE
Mean Attacks
Observations
Number of Groups
Yes
Yes
.272
27693
543
Yes
Yes
.272
27693
543
Yes
Yes
.265
27489
539
Notes: The dependent variable is the number of incidents per district
and quarter. Significance levels are indicated as * 0.10 ** 0.05 *** 0.01.
Standard errors in the parentheses are clustered at the district level.
39
Table 8: How does Rainfall Affect Violence? OLS Results
Intensity
(1)
Base
log(Annual Rain)
(2)
Monsoon
(3)
by Phase
-0.104***
(0.036)
log(Monsoon)
-0.110***
(0.031)
log(Outside Monsoon)
-0.005
(0.010)
Phase 1 x log(Monsoon)
-0.151**
(0.067)
Phase 2 x log(Monsoon)
-0.252***
(0.075)
Phase 3 x log(Monsoon)
-0.111*
(0.066)
District FE
Region x Phase x Time FE
Mean Attacks
Observations
Number of Groups
Yes
Yes
.272
27693
543
Yes
Yes
.272
27693
543
Yes
Yes
.265
27489
539
Notes: The dependent variable is the number of incidents per district
and quarter. Significance levels are indicated as * 0.10 ** 0.05 *** 0.01.
Standard errors in the parentheses are clustered at the district level.
40
Table 9: The Effect of NREGA on the Relationship between Rainfall
and Violence - OLS Results
Intensity
(1)
Base
log(Annual Rain)
NREGA x log(Annual Rain)
(2)
Monsoon
(3)
by Phase
-0.117***
(0.038)
0.064
(0.052)
log(Monsoon)
-0.127***
(0.032)
NREGA x log(Monsoon)
0.061
(0.041)
Phase 1 x log(Monsoon)
-0.213***
(0.076)
NREGA x Phase 1 x log(Monsoon)
0.250**
(0.115)
Phase 2 x log(Monsoon)
-0.211**
(0.101)
NREGA x Phase 2 x log(Monsoon)
0.157
(0.102)
Phase 3 x log(Monsoon)
-0.080**
(0.037)
NREGA x Phase 3 x log(Monsoon)
-0.061
(0.041)
District FE
Region x Phase x Time FE
Mean Attacks
Observations
Number of Groups
Yes
Yes
.272
27693
543
Yes
Yes
.272
27693
543
Yes
Yes
.272
27693
543
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard
errors in the parentheses are clustered at the district level. All rainfall measures
are for the preceding year or preceding growing season.
41
Table 10: How does Rainfall Affect NREGA Take-Up for the Whole of
India?
log(Monsoon)
(contemporaneuous)
(1)
Extensive
(2)
Extensive
(3)
Overall
(4)
Overall
-0.024
(0.017)
-0.023
(0.017)
-0.229***
(0.080)
-0.230***
(0.080)
log(Monsoon)
(lagged)
-0.094***
(0.020)
-0.362***
(0.084)
log(Outside Monsoon)
(contemporaneuous)
0.030***
(0.006)
0.030***
(0.006)
0.082***
(0.028)
0.083***
(0.027)
log(Outside Monsoon)
(lagged)
-0.006
(0.006)
-0.004
(0.006)
-0.036
(0.031)
-0.038
(0.032)
Phase 1 x
log(Monsoon) (lagged)
-0.153***
(0.031)
-0.270***
(0.103)
Phase 2 x
log(Monsoon) (lagged)
-0.070
(0.046)
-0.596***
(0.170)
Phase 3 x
log(Monsoon) (lagged)
-0.015
(0.026)
-0.344*
(0.178)
District FE
Region x Phase x Year FE
Mean of the DV
Observations
Number of Groups
Yes
Yes
.365
2310
521
Yes
Yes
.365
2310
521
Yes
Yes
14.7
2304
521
Yes
Yes
14.7
2304
521
Notes: Significance levels are indicated as * 0.10 ** 0.05 *** 0.01. Standard errors in
the parentheses are clustered at the district level. The panel is naturally unbalanced
due to the phase-wise roll-out.
42
-4
-2
0
2
4
Phase 1 Districts - Effect of Rainfall on Violence over Time
2000q3
2003q3
2006q3
Time
2009q3
2012q3
-4
-2
0
2
Phase 2 Districts - Effect of Rainfall on Violence over Time
2000q3
2003q3
2006q3
Time
2009q3
2012q3
-4
-2
0
2
Phase 3 Districts - Effect of Rainfall on Violence over Time
2000q3
2003q3
2006q3
Time
2009q3
2012q3
Figure 7: Phase Specific Effects over Time - 90% confidence intervals are dotted lines,
the blue line presents results from less demanding specification in case the effect was
significant.
43