Weather Shocks, Agriculture, and Crime: Evidence
from India
David S Blakeslee⇤ Ram Fishman†
January 31, 2017
Abstract
We use detailed crime, agriculture, and weather data from India during the years 1971-
2000 to conduct a systematic analysis of the relationship between weather shocks and
multiple categories of crime. We find that drought and heat exert a strong impact
on virtually all types of crimes, that the impact on property crimes is larger than for
violent crimes, and that this relationship has been relatively stable over three decades
of economic development. We then use seasonal and geographical disaggregations of
weather and agricultural cultivation to examine the prevailing hypothesis that agricul-
tural income shocks drive the weather-crime relationship in developing countries. The
patterns we find are consistent with this hypothesis in the case of rainfall shocks, but
suggest additional mechanisms may play an important role in driving the heat-crime
relationship, consistent with evidence from industrialized countries.
Keywords: Crime, Income Shocks, Weather Shocks, Climate, Agriculture
JEL Codes: O10, O13, Q54
⇤New York University - Abu Dhabi. Email: [email protected]
†Tel Aviv University and George Washington University. Email: [email protected]
1
1 Introduction
Growing interest in the potential impacts of climate change has spurred a substantial liter-
ature on the connection between weather conditions and conflict. This literature has now
produced a substantial body of evidence that weather shocks increase civil conflict (Hsiang,
Burke, and Miguel, 2013; Burke et al., 2009), which has dovetailed neatly with a related
literature on the economic determinants of civil conflict (Collier and Hoeffler, 1998, 2004;
Miguel, Satyanath, and Sergenti, 2004). A smaller number of papers, focused mostly on in-
dustrialized countries, have examined the relationship between weather and personal crime
(Ranson, 2014).
In this paper, we use weather and crime data spanning several hundred Indian districts
over three decades (1971-2000) to conduct a systematic analysis of the connections between
heat, drought, and a range of property and violent crimes that is unique in its spatial
and temporal scale; in the range of crimes investigated; and in the integrated analysis of
potentially correlated temperature and rainfall variability.
We find strong evidence that both elevated heat and drought lead to substantial increases
in crime rates of virtually all categories. During years in which precipitation is more than
one standard deviation below the local long-term mean, crime rates increase by about 5%,
while temperatures one standard deviation above the local mean are associated with a
3% increase in crime. We also find that, though average crime rates have declined over
time, the relationship between crime and weather has remained remarkably stable, despite
the considerable economic development and structural change occurring across these years.
This stability suggests that economic development, as it has hitherto unfolded, may have
limited ability to shield vulnerable sectors of society from the warming and increased rainfall
variability that is expected to occur over the coming decades as a result of climate change.
Having established the effect of weather shocks on crime, we make use of the spatial and
temporal extent of our data to make headway in understanding the mechanisms at work,
an issue which remains a major gap in the literature (Hsiang, Burke, and Miguel, 2013).
Previous studies of weather-crime connections in industrialized countries have tended to be
confined to temperature fluctuations, and to explain their correlation with crime through
a psychological mechanism relating heat to aggression (Anderson et al., 2000; Auliciems
and DiBartolo, 1995; Card and Dahl, 2011; Cohn and Rotton, 1997; Jacob, Lefgren, and
Moretti, 2007; Kenrick and MacFarlane, 1986; Larrick et al., 2011; Mares, 2013; Ranson,
2
2014; Rotton and Cohn, 2000). In contrast, the few studies of weather-crime connections
in developing countries have generally been confined to rainfall fluctuations, and invoked
an income mechanism to explain their results (Miguel, 2005; Mehlum, Miguel, and Torvik,
2006; Sekhri and Storeygard, 2014; Fetzer, 2014). The latter hypothesis is based on the sen-
sitivity of agricultural productivity, wages, and employment to rainfall fluctuations in many
developing countries, including in India (Guiteras, 2009; Jayachandran, 2006; Kaur, 2014;
Fishman, 2016), and the large share of agriculture in both rural income and employment.
It is also well grounded in the economic theory of crime, pioneered by Becker (1968), which
predicts that individuals experiencing negative income shocks will be more likely to resort
to criminal activity.1
As plausible as these hypotheses may be, previous studies have tended to assume rather
than test them.2 Moreover, their validity has increasingly been called into question by the
growing literature on the multi-faceted impacts of weather shocks. Temperature shocks in
particular are known to have deleterious effects on a range of economic outcomes (Dell, Jones,
and Olken, 2014), including not only agricultural output (Deschenes and Greenstone, 2007;
Lobell, Schlenker, and Costa-Roberts, 2011; Schlenker and Lobell, 2010; Guiteras, 2009;
Fishman, 2016; Colmer, 2016), but also labor markets and productivity in non-agricultural
sectors (Hsiang, 2010; Dell, Jones, and Olken, 2012; Sudarshan and Tewari, 2013; Zivin and
Neidell, 2014; Deryugina and Hsiang, 2014); while also affecting a number of psychological
(Anderson et al., 2000), cognitive (Zivin, Hsiang, and Neidell, 2015; Park, 2016), and health-
related (Burgess et al., 2013) outcomes. Because of the multiple channels through which
temperature shocks may affect crime and other forms of conflict, particularly in developing1Beginning with Becker (1968), an extensive literature has invoked individual utility optimization to
explain the economic factors driving the decision to engage in criminal activities. Particular emphasis is
given to the opportunity costs of engaging in crime, which consist of the foregone income and other penalties
in the event of capture, with the general prediction that reductions in legal income due to economic shocks
will reduce the opportunity cost of crime and thereby increase its incidence. A number of studies have
subsequently sought to empirically test the posited relationship between economic incentives and crime
(see Freeman (1999), for a review); and have generally found a relationship between economic distress and
the incidence of crime (Gould, Weinberg, and Mustard, 2002; Machin and Meghir, 2004; Grogger, 1997).
Virtually all of these studies are based on industrialized countries, however, with the result that little is
known about how income shocks affect crime in the developing world, precisely those places most susceptible
to highly disruptive economic shocks, and in which individuals are least able to insure themselves against
large drops in income. Moreover, since economic conditions can themselves be influenced by the prevalence
of crime (Bourguignon, 2000), and because unobservable characteristics such as institutions and culture may
influence both the dependent and independent variables, identifying the causal relationship between the two
is a perennial challenge.
2In fact, the well established relationship between rainfall and agricultural income has led some re-
searchers (e.g., Miguel, Satyanath, and Sergenti, 2004) to use rainfall incidence as an instrumental variable
for estimating the causal relationship between income and civil conflict, which effectively assumes that rain-
fall does not affect conflict by any channel other than agricultural income. Recent studies have sought
to test this assumption more closely by inspecting whether infrastructure (Sarsons, 2015) or social policy
(Fetzer, 2014) that can protect income from rainfall shocks weaken the rainfall-civil conflict relationship,
and find conflicting results.
3
countries, it essential to gather evidence that would shed light on the actual mechanisms at
work.
Our analysis focuses on the role and relative importance of agricultural income in driving
the associations we observe between heat, drought, and crime. Our approach consists of
a systematic comparison of various types of heterogeneity in the patterns relating weather
(temperature and precipitation) and crime, on the one hand, and those relating weather and
agricultural production, on the other. Disagreement between these patterns is counted as
evidence against the agricultural income hypothesis, while correspondences are interpreted
as strengthening it. Four types of heterogeneity are examined. The first makes use of the
seasonal structure of the Indian agricultural calendar, wherein agricultural production de-
pends most vitally on weather during the monsoon season. This approach contrasts with
that of previous studies, which have tended to rely on annual weather, thereby facilitating a
closer analysis of the relevant mechanisms at play, while simultaneously improving precision
by removing the noise introduced through the inclusion of weather events of little economic
importance.3 The second makes use of geographical variation in climatic conditions, includ-
ing average precipitation, which varies widely across different regions, and the existence of
a second monsoon in parts of southern and eastern India. The third utilizes geographical
differences in cultivation practices, including in the extent of dry-season cultivation and the
use of irrigation, both of which have strong implications for the predicted effects of weather
shocks on crime. The fourth compares property and non-property crime rates, and tests
the simple theoretical prediction that property crimes should be more strongly responsive
to income shocks than violent crimes
In the case of precipitation, these comparisons generally yield results consistent with the
agricultural income hypothesis. Drought during the principal rainy season, the Southwest
Monsoon, which is also the main cultivation period, has strong negative effects on agricul-
tural production and strong positive effects on crime rates. The effect on property crimes is
larger, in a statistically significant way, than on non-property crimes. Similar patterns are
found for drought occurring during a second rainy season, the Northeast Monsoon, in those
parts of India subject to it:4 when the Northeast Monsoon is weak, agricultural production
suffers and crime rates increase. Excessive rainfall during the primary rainy season also3In results not shown, we find that the estimated impacts of weather shocks on crime are reduced by the
use of the annual, as opposed to seasonal, measures of weather variation.
4This monsoon sweeps along the eastern seaboard up through northeast India just as the summer monsoon
is subsiding across much of the subcontinent.
4
leads to reductions in agriculture and increases in crime, though the effects are smaller than
for negative rainfall shocks. Moreover, this result obtains only in those areas having drier
climates and therefore specializing in “dryland crops” which are more susceptible than other
crops to damage in the face of extreme rainfall.5 Finally, irrigation dams, which are asso-
ciated with lower weather-sensitivity of agricultural output in downstream districts, appear
to lower the impact of negative rainfall shocks on crime as well, though insufficient power
prevents us from establishing this result conclusively.
The parallel evidence in the case of temperature shocks is also broadly consistent with
the existence of an agricultural mechanism. High temperature shocks that occur during
the primary agricultural season, which lead to substantial declines in agricultural output,
also lead to an increase in crime, and the estimated impact is larger for property than non-
property crime. In addition, the increase in crime with rainy season temperature shocks
is substantially larger than that found with positive temperature shocks occurring during
equally hot periods of the year that lie outside the primary agricultural season, suggesting
that an agricultural mechanism is at least partially responsible for the primary season re-
sults. Even the smaller impact on crime outside the primary season is consistent with an
agricultural mechanism: while the extent of cultivation during these periods is lower, yields
also suffer from the same temperature shocks that increase crime; and the effect of these
off-monsoon temperature shocks on crime is larger in those parts of the country where off-
monsoon cultivation is more extensive. These comparisons are strikingly consistent with an
agricultural income channel, though power limitations render verdicts of varying statistical
precision.
The agricultural income channel, however, is unable to explain all of the patterns we
observe for the effect of temperatures. Positive temperature shocks during the second mon-
soon, for example, lead to an increase in crime despite there being no corresponding de-
crease in agricultural income. In addition, negative temperature shocks occurring during
the monsoon season, which if anything increase agricultural income, simultaneously lead
to an increase in certain types of crime. The latter result, though not readily explicable
in terms of income-based channels, could potentially be explained through the effects of
favorable temperatures on social interactions, a phenomenon for which evidence has been5
Papaioannou (2016) provides historical evidence in favor of such an interpretation.
5
found in the U.S. (Ranson, 2014).
We interpret the sum of the evidence emerging from these various comparisons to be
consistent with the agricultural income hypothesis, i.e. that agricultural productivity shocks
drive at least part of the observed effects of weather on crime . Even though our analysis
doest not claim to, and is unable to fully explain each and every impact of weather shocks,
it constitutes an important step forward in the establishing some of the the mechanisms
mediating the weather-crime relationship. Our results should not be taken as precluding
the presence of channels of influence other than agricultural income, but they do support the
proposition that agricultural incomes are a central, if not exclusive, mechanism mediating
the weather-crime relationship.
The remainder of the paper is organized as follow. Section 2 describes our data and the
empirical approach. Section 3 presents results relating weather shocks to both agriculture
and crime. Section 4 investigates the agricultural income hypothesis further by comparing
agro-climatic heterogeneities in the impacts of weather shocks on crime and agriculture.
Section 5 discusses how the effects of weather on crime have changed across time, and
section 6 concludes.
2 Data and Empirical Strategy
2.1 Data Sources
2.1.1 Crime Data
Data on crime rates was obtained from India’s National Crime Records Bureau (INCRB),
housed under the Ministry of Home Affairs. INCRB produces annual documents on national
and sub-national crime trends, including detailed statistics on the annual incidence of various
types of crimes at the district levels, beginning in 1971.6 The types of crime for which data is
available and we will use here include burglary, robbery, banditry, theft, riots, murder, rape,
and kidnapping.7 Our main outcome variable is the logarithm of the number of incidents
occurring per 100,000 people in each district and year (population data are extrapolated
linearly from Indian census data). Figure A.1 displays the geographical distribution of the6Crime data is also available before 1971, but only at the state level.
7Robbery is distinguished from theft in that it involves violence in the commission of the crime. Banditry
is distinguished from robbery in that it is committed by five or more individuals.
6
average rates of the different crime categories in our data. Figure A.2 presents plots of India-
wide average crime rates over time, and table 1 presents some summary statistics. The three
columns report the incidence of the indicated crime across the three decades spanning 1971-
2000, and indicate a general decline in the incidence of most property crimes, particularly in
burglary, banditry, thefts, and robbery. Kidnapping was relatively stable across this period,
while riots have slightly declined. Murder and rape show a small increase.
A small number of observations (less than 2% of the sample for most crimes)8 report
zero crime incidence, making the outcome variables, the logarithm of crime rates, undefined.
We follow the approach of Pakes and Griliches (1980), who replace this outcome variable
with zero where the crime incidence is zero, and then include a binary indicator of this
sub-sample in the regressions.
2.1.2 Weather Data
Weather data is based on gridded daily precipitation and daily mean temperature data pro-
duced by the Indian Meteorological Department (Rajeevan et al., 2005; Srivastava, Rajeevan,
and Kshirsagar, 2009) and converted to district-wise figures by area-weighted averaging over
grid points falling within a given district (see Fishman, 2016 for more details). We make
use of this daily data in order to calculate summary measures of the annual weather real-
ization (precipitation and temperature) in each district, which we then use in the regression
analysis.
Our measure of precipitation consists of the total seasonal amount. Almost all rainfall
in India occurs during the Southwest Monsoon season (figure 1), which lasts from June to
September, though some parts of southern and eastern India receive a substantial portion
of their rainfall during a second monsoon season, the Northeast Monsoon, which occurs
between October and December (figure A.3). We follow the terminology of the Indian
meteorological department and denote June-September as the “monsoon” season, October-
December as the “post-monsoon” season, and March-May as the “pre-monsoon” season.9
Our interest in the linkage between weather, agriculture, and crime leads us to make
use of the growing season Degree Days as our measure of heat exposure. This measure is
considered to be more relevant for crop growth than average temperatures, and is commonly
used in empirical studies of weather-agriculture relationships, including in India (Schlenker,8The one exception to this is banditry, for which zero values are encountered 7% of the time.
9This classification scheme is described at https://en.wikipedia.org/wiki/Climate_of_India, which gives
useful details of the Indian seasonal calendar.
7
Hanemann, and Fisher, 2006; Guiteras, 2009; Fishman, 2016). Degree Days are defined by:
DDS =X
d
D(Tavg,d
)
D(T ) =
8>>>>>><
>>>>>>:
0 if T 8oC
T � 8 if 8oC < T 32oC
24 if T > 32oC
,
where Tavg,d
is the mean temperature in day d. The daily resolution of our temperature data
allows us to calculate degree days explicitly in each district, over the time period of interest,
which will usually consist of the monsoon season, but we will also make use of degree days
calculated during the pre and post-monsoon periods. We note that the pre-monsoon season
is the hottest period of the year in most of India (figure 1).
2.1.3 Agricultural Data
The main agricultural growing season in India, the Kharif, coincides with the monsoon
(June to September), when rain-fed cultivation is possible. For the poor in particular, who
lack access to irrigation, this season is the primary source of agricultural income. Rainy
season crops, of which rice is the predominant one, are typically planted in June to July
and harvested between September and December, depending on the specific crop and region
(figure 1). Production during this season is sensitive to both monsoon precipitation and
temperatures.
The secondary agricultural season, during which the Rabi crop is cultivated,10 begins in
November-January (sowing times vary by crop and region) and can last until as late as May.
Crops grown during this season, of which wheat is the predominant one, rely on monsoon
precipitation for soil moisture (or irrigation water), but they are also sensitive to in-season
(pre-monsoon) heat. In those parts of India that receive an additional rainy season in the
post-monsoon period (October-December), the agricultural calendar may differ somewhat,
since these secondary rains enable either a longer duration Kharif crop, or an additional
crop outside the monsoon period. As a result, production in these regions in both the main
and secondary seasons can be affected by low rainfall in the post-monsoon period.
District level agricultural data is obtained from two sources. The first consists of the10
In some regions there is even a third crop, which is grown at the end of the Rabi season during the hot
months prior to the monsoon
8
data used by Duflo and Pande (2007), which contains information on aggregate production,
yields (production per unit cultivated area) and wages. Production and yield are measured
in Rupees, and calculated as the total value of the production of the principal crops, using
constant 1961-5 prices. These data are defined on an annual basis, and are aggregated over
the main season of the year in question and the following secondary season, even though
the latter technically occurs in the following calendar year.
To complement this data with seasonally disaggregated agricultural output measures,
which are of importance to our analysis, we also make use of the India Harvest database
produced by the Center for the Monitoring of the Indian Economy (see Fishman, 2016 for
a complete description), which contains information on the crop-specific area, production,
and yield of major crops during both of the two agricultural seasons.11 Summary statistics
reported in table 1 show the substantial improvements in agricultural production which
occurred during the period of our study. We note that each of the two agricultural data sets
contain substantial numbers of missing observations (see Duflo and Pande, 2007).
These two sets of measures each have advantages and disadvantages as indicators of
agricultural performance. While outcomes for individual crops provide a richer and more
precise seasonal disaggregation, they may fail to capture shifts in cropping decisions across
crops, which may be correlated with weather. In addition, numerous crops are cultivated in
India, and it is not clear how to balance potentially divergent responses to weather shocks
across crops. Annually aggregated production provides a summary measure of output that
accounts for shifts in cropped areas and cross-seasonal impacts, but relies on the specific
weights used for aggregation (we follow Duflo and Pande, 2007, and Guiteras, 2009, in using
1960-65 prices).
Determining the appropriate agricultural indicator for an analysis of crime poses addi-
tional challenges. Ideally, one would observe the distribution of household-level agricultural
income within a district-year, and compare income shocks and crime incidence. Because such
data is unavailable at the required spatio-temporal resolution, we instead use district-level
data on agricultural output and wages, which have proven elsewhere to capture important
dimensions of rural welfare in India (Jayachandran 2006; Kaur 2014). Without information
on the timing of crime and the identity of its perpetrators, it is difficult to know which agri-
cultural outcome is most relevant. For example, if landless agricultural laborers are those
most likely to resort to crime, then agricultural wages or employment may be the most11
This data set has the drawback that it is unbalanced for all but the few most important crops.
9
relevant agricultural variable. If crime is instead typically being committed by land-owning
smallholders, then the yields of the crops typically cultivated by these farmers might be of
primary interest. Uncertainty regarding the identity of perpetrators also has implications
for the relevant timing of weather shocks as well: wage laborers may be most economically
vulnerable to early season weather fluctuations, when cropping decisions are made; whereas
smallholder farmers would be vulnerable to shocks occurring throughout the agricultural
cycle.
In what follows, we remain agnostic on the specific household-level mechanisms driving
the crime response, and report estimates for several agricultural outcomes, including aggre-
gate annual production, yield, and wage, and the production of the primary crop in each
season (main and secondary).12 The various agricultural indicators tend to display effects
that are well correlated in sign. In the appendix we report impacts on the production of
specific crops.
2.1.4 District Partitioning
An important issue in empirical studies on India for which the unit of observation is the
district is the substantial partitioning of districts that has occurred over time. This makes
it challenging to combine district-level data sets and to assign spatial data to observations
correctly. It also raises the question of the correct determination of the relevant unit for
capturing time-invariant unobservables. One approach is to fix district boundaries at a
certain early date, and then aggregate outcomes occurring after districts are partitioned
to the original boundaries. We prefer to follow a more conservative approach, in which a
distinct district is observed only during the continual period of time in which its boundaries
remained un-modified. Whenever a district is partitioned into two new districts, for example,
we consider the two new districts as separate from the original one, assigning separate fixed
effects to these separate districts in our panel regressions. In this way, we do not make any
assumptions about the resulting changes in institutions, district population, resources, and
local governance, which can be substantial. To implement this approach, we use detailed
records of the partitioning and formation of districts over the period of the study Kumar and
Somanathan (2009), and assign every district the appropriate agricultural and crime data.
Weather data is coded using the year 2000’s district boundaries, but we employ weighted12
The primary crop for a given district-season is defined as the crop that occupies the largest average area
in a district in a given season.
10
averaging to assign it to historical districts. Overall, our data spans 590 districts and a 30
year time period. However, given the way we define districts, every district is “observed”
only during its period of actual existence. The average number of observations per district
is 16, and the total sample size is about 9200.
2.2 Empirical Strategy
The basic structure of the models we estimate takes the form:
Log(Zist
) = ↵+ �W
ist
+ �
t
+ �
i
+ f
s
(t) + ✏
ist
. (1)
Here, Z is an outcome variable of interest, indexed by district (i), state (s), and year (t).
These outcomes will consist of the rates of various crime, and of agricultural outcomes
(production, yield, etc.). The main explanatory variable, W , is a vector of summary mea-
sures of annual weather (temperature and precipitation) constructed from daily data. The
model also includes district fixed effects, year fixed effects, and state specific time trends
(�i
, �t
, and f
s
(t), respectively),13 in order to absorb time-invariant district characteristics,
sample-wide annual shocks, and secular time trends that can lead to spurious correlations
between productivity, crime, and weather. As explained above, separate district fixed ef-
fects are assigned to every administrative district, and districts created from partitions or
combinations of existing districts are considered to be distinct, and assigned independent
fixed effects. Because our main outcome variable is defined in terms of per-capita crime oc-
currence, we weight each observation by that district’s population, as reported in the 1971
census.
In calculating the regression’s standard errors, we employ a two-way clustering procedure
in order to simultaneously account for within-district serial correlation in errors and within-
year spatial correlation in weather shocks (Conley and Molinari, 2007; Hsiang, 2010; Fetzer,
2014). In our benchmark specification, we allow for arbitrary serial correlation, and spatial
correlations extending up to 300 KMs.
2.2.1 Specifying Weather Shocks
There are two aspects of the manner in which we specify weather realizations that warrant
discussion. First, we choose to specify weather variables not in terms of their physical levels,13
Are main specification uses linear, state specific time trends, Specifications using quadratic time trends
were also estimated, and yield nearly identical results.
11
but in terms of the magnitude of their deviation from the local mean of this weather variable,
calculated in terms of the localized standard deviation over the 1970-2005 period. We refer
to these as the standardized variables.14 This seems a natural approach to the modeling of so
extreme a behavioral response as crime, to which individuals presumably resort only when
other, less extreme coping mechanisms that are typically available to economically distressed
households have been exhausted. From this perspective, it is reasonable to expect crime
to respond more highly to shocks that are extreme in terms of their rarity, rather than in
terms of their physical levels, as one might expect for the purely physical response of crop
yields to weather. In an appendix, we test the robustness of our main results to alternative
specifications.
Second, we choose to specify the functional form of these standardized weather deviations
through the use of two binary variables (for each weather variable), which we denote by
W
+and W
�, that indicate whether they differ from the local long-term mean by more
than one standard deviation in either direction.15 This specification is motivated by the
non-linear relationship between weather shocks, crime, and agriculture, suggested by figures
5-7, which portray local linear regressions of crime incidence and agricultural output on
our two main weather variables (standardized). We also estimate richer semi-parametric
regressions that include a larger number of binary indicators to capture more sub-ranges at
which standardized weather deviations may occur. Both of these aspects of our specification
are similar to those used in other papers that investigate the impacts of rainfall shocks on
district level outcomes in India, which make use of binary indicators of the extremity of a
rainfall shock in terms of the long-term localized rainfall distribution (e.g. Jayachandran,
2006; Kaur, 2014). For completeness, however, we also test additional weather specifications
used in other studies (e.g., Duflo and Pande, 2007; Jayachandran, 2006) in an appendix.
As explained above, we decompose annual weather into the three seasons of the monsoon,
the pre-monsoon, and the post-monsoon, and include separate shock indicators for each of
them.14
Note that while, as discussed above, we observe crime and agriculture in a given district only during the
years in which the district actually exists as an identical administrative units, we calculate these standardized
weather deviations by using the long-term (35 years) weather observed in the same location.
15More explicitly, the binary variable indicate if the standardized variable is higher than 1 or lower than
-1.
12
2.2.2 Temporal Structure
Figure 2 displays the temporal structure of our model specification. As depicted at the
bottom of the figure, our principal agricultural outcome, annual production in year t, is
defined as the sum of production during the primary agricultural season (generally June-
December of year t) and the following secondary agricultural season (January-May of year
t + 1). Production in the primary season is affected by rainfall and temperatures during
the monsoon; while production in the secondary season is affected both by weather in the
preceding monsoon16 as well as temperatures during the pre-monsoon period (as depicted
in figure 2), with little rainfall occurring during the secondary season itself. In order to
properly capture the effects of weather shocks on agriculture, our agricultural regressions
therefore use monsoon weather shocks occurring during the rainy season of year t, and
pre-monsoon weather shocks that occur during the calendar year t + 1 as the explanatory
variables. As we will discuss below, in some parts of India there is a secondary monsoon
season that occurs from October to December. In these areas, rainfall during this period
can affect crops grown during both the primary season and the following secondary season.
Our crime outcome variable consists of crime incidence over the course of a given calendar
year t. In our main specifications, we include weather shocks occurring within the same
calendar year in which crime is reported. These include: weather shocks occurring during
the monsoon season of year t (June-September), which impact the production of crops that
are planted around June and harvested up to December of year t; pre-monsoon weather
shocks occurring during March-May of calendar year t, which affect the production of crops
planted around January of year t and harvested up to May; and, in some specifications, post-
monsoon shocks occurring in October-December of year t, which impact the production of
crops grown during both the main season of year t and secondary season of year t+ 1.
There is a possibility that weather shocks occurring in the previous calendar year, t� 1,
may also affect crime rates in year t. For example, in areas experiencing a secondary
monsoon season (Northeast Monsoon) during October-December, weather shocks occurring
during the latter season of year t � 1 may affect crops that are grown and harvested in
the secondary season of year t; which, in turn, may lead to an increase in crime in year
t . The primary monsoon season (Southwest Monsoon) too may exert a lagged effect on
crime, as weather shocks during June-September of year t�1 also affect agricultural output16
The secondary season crops rely on moisture derived from the accumulation of rain water in soils,
aquifers, and surface water bodies.
13
during the dry season (January-May) of year t, though their influence is smaller than that of
secondary monsoon shocks. We therefore examine these effects in additional specifications
that control for “lagged” weather shocks.
2.2.3 Specification of Crime Regressions
We estimate two types of regressions in which crime rates are the outcome variable. In the
first, we pool together all crime categories in a single regression; in the second we estimate
the same specification separately for each crime, with fixed effects and time effects adjusted
accordingly. To be explicit, the specification employed is the pooled model is:
ln(ycist
) = ↵+ �1 · P+it
+ �2 · P�it
+ �3 · T+it
+ �4 · T�it
+ �
c,t
+ �
c,i
+ f
c,s
(t) + "
cist
(2)
where the natural log of the incidence of crime c (per 100k people) in district i, state s,
year t is regressed on binary indicators of positive and negative precipitation (P+, P
�) and
degree-days shocks ( T
+, T
�) in year t, district i. In estimating these pooled regressions,
we allow errors associated with different types of crime observations to be correlated, in
addition to employing temporal and spatial clustering.
As discussed above, simple theoretical considerations suggest income shocks should have
a larger impact on property crimes. We will therefore also separately estimate regressions
for property and non-property crime rates. Of the crimes included in the data, we classify
burglary, banditry, theft, robbery, and riots as property crimes, and murder, rape, and kid-
napping as non-property crimes. Of these, kidnapping and rioting would seem to occupy
an ambiguous place; however, closer scrutiny justifies this classification. Kidnapping, for
example, is disproportionately targeted against women, for reasons not entirely, or even
principally, economic.17 Riots are known to occur during times of economic duress, par-
ticularly in response to heightened food prices, and are often characterized by widespread
looting.18
17The data indicates that the 74 % of kidnappings are targeted against women, though this disaggregation
is only reported after 1987. Though others have classified this as an economic crime, such a classification
overlooks this important non-economic component.
18Note that these are distinct from the Hindu-Muslim riots analyzed by Bohlken and Sergenti (2010).
14
3 The Effects of Weather Shocks on Agriculture and
Crime
In this section, we examine whether weather shocks that disrupt agricultural production also
tend to increase crime rates. To do so, we regress both agricultural production and crime
rates on a range of weather shocks and compare the estimates. In most tables that follow,
we facilitate the comparison by reporting, side by side, regressions of the two outcomes on
the same set of weather shocks. Our analysis unfolds by gradually increasing the range of
weather shocks considered, thereby scrutinizing the thesis that agriculture meditates the
effects on crime in greater detail. For crime outcomes, we report results that pool all types
of crimes into a single regression; results that pool all property crimes together; results
that pool all non-property crimes together; and separate estimates for individual crime
categories. For agricultural outcomes, we focus on measures of agricultural production used
by Duflo and Pande (2007): the logarithm of the total value of production of the main crops,
calculated using 1960-65 prices, the logarithm of the yield (production per area), and the
agricultural labor wage rate. In addition, for the purpose of seasonal disaggregation, we also
report impacts on the (logarithm) of the yield of the primary crop in both the main and
secondary seasons, the primary crop in each district being defined as the crop that occupies
the largest average area in that district in the season in question.19 We also report, in
Appendix table A.1 results for the yields of the most important specific crops.
3.1 Weather Shocks in the Main Growing Season (Monsoon)
We begin by estimating a model that follows much of the weather-agriculture literature in
India in that it is focused on deviations in two variables: total precipitation and degree-days
in the main growing season (the monsoon season of June-September). We also control for the
same weather indicators calculated separately during the pre-monsoon season (March-May),
for reasons that will become clearer when we discuss their estimates in the next sub-section.
Regression results are reported in table 2. Column (1) reports estimates that pool all
types of crime into a single regression. Columns (2)-(6) report results of similar models
in which the dependent variables are the agricultural outcomes mentioned above. The
statistically significant estimates in columns (1) and (2) indicate that the same monsoon19
We are unable to calculate a complementary measures of seasonal aggregate production because pro-
duction data is missing for a substantial number of observations for all but the few most important crops.
15
weather shocks that reduce agricultural production also tend to increase crime. Negative
rainfall shocks are associated with a 15% decrease in production, a 9.5% decline in yields,
a 2.5% decline in the wage rate, and a 4.8% increase in crime incidence. Positive rainfall
shocks are associated with an (imprecise) 2.8 % decrease in production, 2.4% decrease in
yields, 2.6% decrease in wages, and a 1.5% increase in crime. Positive temperature shocks
are associated with a 8.4% decrease in production, a 4.8% decline in wages, and a 3.3%
increase in crime. In appendix table (A.1) we report the impact of these weather shocks on
the yields of specific crops, and find similar effects.
The impacts we observe on agricultural production are in line with basic agronomic
considerations and with previous statistical studies of weather-agriculture linkages in India
and elsewhere (Guiteras, 2009; Auffhammer, Ramanathan, and Vincent, 2012; Fishman,
2016). Drought (negative rainfall shocks) limits moisture availability for crops grown during
both the monsoon season and the subsequent season; excessive rainfall (positive rainfall
shocks), under certain conditions, can damage crop yields (more on this later); and excessive
heat (positive shocks to degree days) harms crop maturation and reduces yields.20
In table 3 we report results that disaggregate property and non-property crimes. Column
(1) reports estimates that only includes property crimes, column (2) reports estimates that
only includes non-property crimes, and column (3) reports the p-value for a test of the
equality of the coefficients for property and non-property crimes. The estimates in columns
(1) and (2) exhibit point estimates that tend to be larger for property crimes than for
non-property crimes: whereas negative rainfall shocks lead to a 5.8% increase in property
crimes, they lead to an almost 50% smaller (3.1%) increase in non-property crimes, with the
difference between the two being statistically significant at the 10% level (p-value=0.07).
Similarly, whereas positive temperature shocks lead to a 3.8% increase in property crimes,
they lead to a smaller (and statistically insignificant) 2.4% increase in non-property crimes,
though the difference is imprecisely estimated (p-value=0.32).
We next estimate the effects of the same weather shocks on each individual crime cate-
gory. In addition to the inherent intrigue of the effects for individual crimes, these estimates
are instructive for two reasons. First, they allow us to assess whether the differential effect
on property versus non-property crimes is sensitive to the classification adopted for kidnap-
ping and riots, which, as noted in the introduction, are of somewhat ambiguous economic20
Though the yield effect of high temperatures isn’t evidenced in this table, appendix table (A.1) shows
the expected effect when looking at individual crops.
16
content. Second, they allow us to explicitly observe effects on two crime categories, murder
and robbery, which are considered to be less susceptible to under-reporting (Fajnzylber, Le-
derman, and Loayza, 1998),21 and thereby provide a partial check of concerns about biased
reporting of other crime categories. The results are reported in columns (4)-(11) of table 3,
and plotted in figure 5. All five property crimes and two of the three non-property crimes
display a statistically significant increase with negative rainfall shocks. As was found in the
pooled regressions, the statistically significant effects on property crimes tend to be larger
than the effects on non-property crimes: riots increase by 7.0%, burglary by 6.7%, banditry
by 7.3%, thefts by 3.5%, and robbery by 4.2%. Among non-property crimes, murder in-
creases by 3.0%, kidnapping by 3.4%, and rape by a statistically insignificant 2.2%. Positive
rainfall shocks only increase banditry in a statistically significant way (by 4.7%). Positive
temperature shocks are associated with a 7.6% increase in banditry, a 5.8% increase in riots,
and a 3.0% increase in murder. We do not find evidence that the impacts on robbery and
murder stand out in comparison to other types of crime, suggesting that under-reporting is
unlikely to be systematically biasing our estimates.
We also estimate the relationship between weather shocks and crime using a richer
model that categorizes weather deviations into a larger number of “shock categories,” or
z-score bins: shocks that are greater than 0.5, 1, and 1.5 standard deviations above and
below the local mean. Table 4 reports the results of these semi-parametric regressions for
the pooled crime categories. Appendix figures A.4 and A.5 and table A.2 report similar
estimates for individual crime categories. Crime rises steadily with larger negative rainfall
deviations, while positive rainfall 1.5 standard deviations above the mean is associated with
a 3.8% increase in crime. The relationship between crime and temperature shocks is noisier:
while crime increases with positive temperature shocks in the 1.0-1.5 z-score bin, it drops
precipitously thereafter, to become indistinguishable from zero.
There is also an evident increase in crime with low temperature shocks, which is driven
by increases in burglary, banditry, robbery, and murder. This increase is not accompanied
by similar responses of agricultural production; and while we are unable to explain its
occurrence through an income mechanism, we note that previous studies have hypothesized
that favorable weather conditions may also lead to increases in crime by increasing social21
Though under-reporting is generally a concern in such contexts, there are good reasons to think that it
would, if anything be biasing our results downwards: specifically, it is not implausible that under-reporting
would increase at precisely those times when crime rates are high, which would have the effect of reducing
the absolute magnitude of the coefficient for adverse weather shocks.
17
interactions (Ranson, 2014).22 Whatever the merit of this thesis, we stress again that we
do not claim here to explain all aspects of the weather-agriculture relationship. Rather,
our purpose is to test the hypothesis that weather shocks that affect agriculture also affect
crime rates, and whether this occurs because they affect agriculture. This hypothesis does
not preclude the possibility that weather shocks have effects on crime which operate through
additional channels. The relative magnitude of the estimated impacts of weather shocks on
agriculture and crime are noteworthy. Figure 6 displays a scatter plot of the estimated
impacts of the four weather shocks that are found to have significant effects on crime (table
2). Though not supporting a definitive conclusion, the correlation of both sets of effects
is rather striking, and consistent with a single underlying mechanism linking the effects of
weather on income and crime. Thus interpreted, all four sets of point estimates point to an
elasticity of crime with respect to agricultural product which is in the order of 0.3-0.55.
3.2 Weather Shocks in the Secondary Growing Season (Pre-Monsoon)
The pre-monsoon period (March-May) is as hot, if not more so, than the June-September
period (figure 1), and experiences little rain. Consequently, the extent of cultivation during
this period is substantially lower than during the monsoon, and weather deviations have
a concomitantly smaller effect on annual agricultural incomes. The estimates reported in
table 2 show that pre-monsoon positive temperature shocks have a significant effect on crime
rates (1.9% increase). The smaller magnitude of this impact in comparison to monsoon
temperature shocks, despite average temperatures across the two periods being equally hot,
is therefore suggestive of the presence of an agricultural mechanism driving at least part
of the effect of monsoon temperatures on crime (though the difference between the two is
imprecisely estimated).
It is important to note, however, that even though cultivation during this time is not as
prevalent as during the main growing season, crops grown during the secondary season are
still likely to be in the field during the pre-monsoon period, and their yields can therefore
be sensitive to temperatures during this time. Indeed, while the estimated impact of pre-
monsoon positive temperature shocks on annual agricultural production is a negative but
imprecise 3.4%, crop-specific estimates, reported in appendix table A.1, show significant
impacts on the yields of of prominent dry-season crops. Therefore, the increase in crime22
In the hot climate prevailing across most of India, cooler than normal temperatures, especially during the
Monsoon, often qualify as favorable, and therefore are likely to increase the incidence of social interactions.
18
with pre-monsoon positive temperature shocks may also be consistent with an agricultural
mechanism, with cultivation in the pre-monsoon period simply constituting a far smaller
share of annual agricultural income. Below, we examine this possibility more carefully by
comparing the impacts on crime and agriculture in areas which have intensive dry season
cultivation to those that do not.
3.3 The Second Monsoon and Weather Shocks in the Post-Monsoon
We now turn to an examination of the impacts of post-monsoon weather shocks (October-
December) on crime and agriculture. In most of India, rainfall is almost entirely concen-
trated within the June-September monsoon season (the “Southwest Monsoon”). In those
areas, October-December temperatures can affect main season crops that are in the final
stages of their maturation, as well as secondary season crops in early stages of cultivation,
depending on the exact local agricultural calendar in each area. Some parts of southern
and eastern India also experience a second monsoon (the “Northeast Monsoon”), which
commences in October, just as the summer monsoon is subsiding, and continues through
November and December. Rainfall during the second monsoon facilitates the cultivation of
a longer main season crop, or of a second crop outside the monsoon (without the neces-
sity of using irrigation), which constitutes an important component of annual agricultural
incomes in these areas. Appendix figure A.3 plots the location of areas experiencing this
second monsoon.23 In these regions, rainfall during the post-monsoon period can have a
substantial effect on agricultural incomes, and potentially crime, and we therefore include
it as an explanatory variable in an analysis that is limited to these areas .
Table 5 reports the estimated impacts of post-monsoon weather shocks on crime, agricul-
tural product and yields. Estimates are conducted separately for the full sample (columns
1-3) and the second monsoon sample (columns 4-6); and, for the latter, second monsoon
rainfall shocks are also controlled for. The results establish that negative rainfall shocks
during the second monsoon reduce agricultural production by 6.4% and also increase crime
by 5.5%, lending further support to the agricultural income hypothesis for rainfall shocks.
On the other hand, post-monsoon positive temperature shocks display divergent effects on
agricultural product and crime. In the full sample, positive temperature shocks reduce pro-
duction but do not seem to affect crime. In the second monsoon sample, they increase crime23
To be precise, we define the second monsoon sample to consist of those districts receiving more than
150 millimeters of rainfall during the post-monsoon
19
rates (5.7%), even though they increase agricultural output. This divergence is hard to
reconcile with either a purely agricultural or a purely psychological channel.
3.4 The Effect of Lagged Weather Shocks
Our principal specification only includes contemporaneous weather shocks, i.e., those occur-
ring within the same calendar year in which crime is observed. Here, we examine the impacts
of weather shocks occurring in the previous calendar year. There are several reasons why
these “lagged” weather shocks might affect crime. First of all, criminal activity may display
serial correlation over time (Jacob, Lefgren, and Moretti, 2007), and could thus be affected
by lagged weather shocks that have increased crime in the previous year.24 Conversely, if
weather shocks merely shift the timing of criminal acts forward, lagged weather shocks may
have an effect on crime which is opposite in sign to the contemporaneous shocks (Jacob,
Lefgren, and Moretti, 2007). Both of these effects are not specific to any one weather shock,
and can therefore be expected to be similarly reflected in the lagged versions of all weather
shocks that are found to impact crime contemporaneously.
Second, weather can also have delayed direct effects on crime through the timing of
economic outcomes. Negative income shocks may be temporarily offset through drawing
down savings or accessing credit, with individuals resorting to crime only after these means
of smoothing consumption have been exhausted. In addition, weather shocks occurring in
calendar year t � 1 may affect the production of crops that are harvested in year t, and
thereby affect crime rates in that year as well;25 Areas experiencing a secondary monsoon
season (October-December) are a particularly plausible candidate for such a delayed effect,
though even weather shocks during the Monsoon period (June-September of year t�1) may
generate a lagged increase in crime, as they too affect crops grown during the dry season
(January-May) of year t, though to a smaller extent.
Table A.3 reports estimates of a specification similar to the baseline specification, ex-
cept that it also includes weather shocks occurring in the previous calendar year. Column
(1) reports estimates that include the full sample, and column (2) reports estimates that
use the “second monsoon” sample discussed above, and include variables for post-monsoon24
Persistence could occur, for example, if criminal acts reduce the opportunity cost of engaging in addi-
tional illicit activities, whether by generating crime-specific human capital, or reducing outside employment
options.
25This can happen if individuals or markets fail to fully predict and capture these future yield effects.
While planting decisions and agricultural labor markets, for example, are likely to react to early season
weather shocks, it seems plausible that some of the impacts of within season weather shocks cannot be fully
predicted and may only become manifest at the time of harvest.
20
weather shocks. We do not find statistically significant evidence for a lagged effect of Mon-
soon rainfall or temperature shocks on crime (column 1). It should be noted, however, that
the point estimate of the lagged Monsoon negative rainfall shock for crime, though impre-
cisely estimated, is indeed negative, and of a magnitude that is consistent with its relative
contribution to annual agricultural production (secondary season crops not only occupy a
smaller area, but as evident from table A.1, they also suffer smaller yield losses as a result of
monsoon rainfall shocks). There is also no indication of a lagged effect of positive Monsoon
temperature shocks on crime: the point estimate is very small and imprecise. However,
there is less reason to expect these lagged temperature shocks to affect secondary season
production in the following year than in the case of rainfall. It is therefore difficult to
interpret this estimate as pointing against an agricultural income channel.
When we examine the lagged effect of the secondary monsoon on crime in the appropriate
sub-sample (column 4), we find a statistically significant estimate of the lagged negative
rainfall shock (6.7%) that is remarkably close in magnitude to the contemporaneous effect.
The appearance of both contemporaneous and delayed effects of this shock on crime is
supportive of an agricultural income channel, whereby declines in secondary monsoon rainfall
affect income both before and after yields are realized, over a a physiological channel that
would be expected to only have contemporaneous impacts on crime. In contrast, there is
no apparent effect of lagged post-monsoon temperature shocks on crime in the following
year. This casts further doubt on the possibility that an agricultural income mechanism
is responsible for the contemporaneous effect of these shocks on crime, discussed in the
previous sub-section.
4 Agro-Climatic Heterogeneities
The analysis of the previous section compared the effects of seasonal rainfall and temperature
shocks on crime and agricultural production, and showed that shocks occurring during the
monsoon and pre-monsoon seasons had corresponding effects on the two outcomes, consis-
tent with an agricultural income mechanism as the driver of the weather-crime relationship.
In this section, we test this correspondence further by comparing agro-climatically derived
heterogeneity in the impacts of weather on these two outcomes. We first examine whether
crime in areas in which pre-monsoon cultivation is more prevalent is more responsive to
pre-monsoon temperature shocks. We then test whether crime is more responsive to posi-
21
tive rainfall shocks in areas that are drier and are therefore mostly cultivated with dryland
crops that are more vulnerable to excessive rain. Finally, we examine whether irrigation
attenuates the effect of negative rainfall shocks on crime.
4.1 Areas with Extensive Secondary Season Cultivation
In the previous section, we found that positive temperature shocks occurring during the
pre-monsoon periods increase crime, and raised the possibility that this effect is driven
by the corresponding impact of these temperature shocks on secondary season agricultural
production. This agricultural effect is manifest in the crop specific estimates reported in
table A.1. To test this possibility more carefully, we decompose our sample into districts
in which (average) secondary season cultivation does and does not constitute a significant
share of annual cultivation, and re-estimate the effect of weather shocks on agriculture and
crime in each.
The results are reported in table 6. Column (1) repeats the prior estimates of the
effect of summer temperatures on crime in the full sample. In column (2), the sample is
restricted to districts where the extent of secondary season cultivation relative to primary
season cultivation is below the median, and in column (3) to districts where it is above the
median. We find that positive pre-monsoon temperature shocks are only associated with a
statistically significant increase in crime in those districts in which dry season cultivation is
substantial, with little evidence of any such effect where this cultivation is low (though the
difference between the estimates in these two sub-samples is statistically insignificant).
4.2 Dry and Wet Areas
As was noted in our baseline results, there is some evidence of an increase in crime with
positive rainfall shocks, which we hypothesized to be due to income losses sustained from
excess rainfall or flooding (Hidalgo et al., 2010; Papaioannou, 2016). To explore this possi-
bility further, we exploit heterogeneities in the responsiveness of different crops to excessive
rainfall, and examine whether crime responds to these shocks in a stronger manner in areas
in which more susceptible crops tend to be grown. We first disaggregate our sample into
districts that are characterized by mean rainfall that is above and below the median, which
we will refer to as the ‘dry’ and ‘wet’ samples. Cropping patterns in India differ substan-
tially between areas with high and low mean rainfall: in areas with low rainfall, the so
22
called ‘dryland crops’ of Sorghum (Jowar), Millet (Bajra), and Maize are more commonly
cultivated; whereas in areas with higher rainfall, the cultivation of rice is more common.
In appendix table A.1, we examined the sensitivities of individual crops’ yields to weather
shocks, and found that the yields of dryland crops are more susceptible to the negative
effects of excessive rainfall shocks than other crops. This disparity is even more apparent
in semi-parametric estimates, reported in appendix table A.4, which show dryland crops
suffering steep declines in yields with excessive rainfall. This suggests that in dry areas,
agricultural production should be more sensitive to positive rainfall shocks; and that crime
may be expected to respond accordingly.
In table 7, we report some differences between the dry and wet samples (columns 1 and
2). In the first row we compare the (long-term average) fraction of cultivated area that is
sown with these crops across the two samples; and, as anticipated, find that this fraction is
larger by 34% in dry areas (56% in dry areas vs. 22% in wet areas). If the effect of positive
rainfall shocks on crime estimated in the previous section operates through an agricultural
channel, one would expect the effect of positive rainfall shocks on crime to also be larger
in the dry sample. This prediction is supported by the regression estimates: while positive
rainfall shocks increase crime by 3.5% in the dry sample, they have a small and insignificant
effect on crime in the wet sample (the difference between the two estimates is statistically
significant, p=0.016).
The semi-parametric estimation, reported in appendix table A.4, shows that both the
decline in agricultural production and the increase in crime in the dry sample occur when
rainfall exceeds 1.5 standard deviations above the mean, consistent with the flooding hy-
pothesis suggested above. In the dry districts, excessive rainfall is associated with a 10.5%
decrease in agricultural yields, and a 5.8% increase in crime. In wet districts, there is a
far smaller decline in yields and no evidence for an increase in crime. These results are
consistent with historical accounts of increases in crime with both drought and floods in
British colonial Asia (Papaioannou, 2016).
4.3 Irrigated and non-Irrigated Areas
Access to irrigation simultaneously increases agricultural income and reduces its susceptibil-
ity to rainfall shocks. If the effect of rainfall shocks on crime is mediated by an agricultural
channel, this may imply that crime in irrigated districts will simultaneously experience a
23
level reduction and a lower sensitivity to negative rainfall shocks. Testing this hypothesis,
though for civil conflict rather than individual crime, Sarsons (2015) finds that Hindu-
Muslim riots are in fact no less likely to occur with negative rainfall shocks in districts that
are located downstream from irrigation dams (and therefore having a greater extent of irri-
gation), thereby calling into question the validity of the agricultural income hypothesis as
driving civil conflict, and by implication the use of weather shocks as instrumental variables
in the income-conflict literature.
Here we follow a similar approach to Sarsons (2015), expanding the analysis to the
domain of individual crime incidence. Like Sarsons (2015), we make use of the data and
identification approach used by Duflo and Pande (2007), who employ geographical variation
to identify the causal effects of irrigation on agriculture, income, and poverty. Calculating
the predicted number of dams in a district (see Duflo and Pande, 2007, for details),26 we
estimate a regression that includes an interaction of our rainfall shocks with this variable.27
The results are reported in appendix table A.5.
Like Duflo and Pande (2007) and Sarsons (2015), we find that the adverse effects of
negative rainfall shocks are substantially mitigated by the presence of upstream dams: the
interaction term of negative rainfall shocks with upstream dams is 9.3% for agricultural
product (marginally insignificant) and 11.1% for agricultural yield (significant at the 5%
level). However, we estimate the interaction of negative rainfall shocks and the presence
upstream dams crime to also be negative, albeit imprecisely. While this result is a sug-
gestively consistent with the agricultural income mechanism emphasized throughout this
paper, limits of statistical power prevent us from testing this difference more conclusively.
It should be noted that the prediction that irrigation should break the rainfall-crime
relationship may depend on strong assumptions about the micro-dynamics of crime. For
example, irrigation coverage, as measured by the fraction of gross cultivated area that is
irrigated, is quite low almost everywhere in India, including in districts served by upstream26
The first-stage in Duflo and Pande (2007) is
Dist = ↵1 +4X
k=2
↵2k(RGrki ⇤Dst) + ↵3(Mi ⇤Dst) +4X
k=2
↵rk(RGrki ⇤ lt) + vi + µst + !ist,
where, Dist is the number of dams in district i of state s at time t, �i are district fixed effects, µst are state-
year fixed effects. RGrki are the river gradient variables, which determine the suitability of the district
for dam placement, and the other variables account for other relevant characteristics. Deriving imputed
dams from this regression,
dDist, we then interact the continuous measure of imputed dams, both within the
district, as well as in upstream districts, with our rainfall variables.
27Our specification differs from the one used by these authors in the way rainfall shocks are specified,
namely that we use the same binary indicators of rainfall shocks used throughout this paper rather than a
continuous linear function of rainfall.
24
dams. Indeed, on average, districts that are downstream from a dam irrigate just 23.5%
of the cultivated land, meaning that even ostensibly ‘irrigated’ districts have a substantial
amount of un-irrigated cultivation. With larger landowners being most likely to capture
the benefits of irrigation expansions, poorer farmers may continue to be disproportionately
dependent on rainfed cultivation, thereby muting the income-smoothing effect of irrigation
for the most vulnerable populations,28 and the potential for a lower impact on crime. While
the absence of detailed micro-data on criminal offenders prevents us from directly testing
these hypotheses, it is important to highlight that the prediction that crime incidence should
be less sensitive to rainfall shocks in more irrigated districts is not a priori self-evident.
5 Crime Responsiveness across Time
We finally turn to an analysis of whether there have been changes over time in the effects of
weather shocks on crime. Given the substantial changes occurring in the Indian economy, it
is natural to expect that there will have been changes in the sensitivity of crime to rainfall
and temperature deviations. For example, with more individuals living in urban areas and
employed in non-farm related activities, fewer households will be directly dependent on
weather for their incomes, while in the agricultural sector, technological and infrastructural
advancement may shield productivity from weather deviations . These changes have the
potential to reduce the sensitivity of crime to weather deviations, as has been found elsewhere
with respect to health outcomes (Barreca et al., 2016; Burgess et al., 2013).29 Countervailing
forces exist, however, that might work against these benefits; such as, for example, higher
urbanization rates, which generate new governance challenges and social antagonisms, or
the expansion of irrigation, which may increase rural inequality or attract landless laborers
from elsewhere.
Whether the remarkably consistent effects of climate on conflict found across a broad
literature might be mitigated by increased prosperity and reduced agricultural dependence
is a question of considerable import (Hsiang, Burke, and Miguel, 2013). Most previous
research has focused on economically static countries, or on short time spans, so that it has
not been possible to identify how the estimated effects might change with economic and28
In fact, such farmers might ultimately suffer larger income declines when rainfall is low, as the lower
effect on average yields in these districts might also reduce the accompanying increase in prices that partially
compensates for the reduction in output.
29Barreca et al. (2016) find that mortality rates in the US have become less responsive to high temperatures
due to the expansion of air conditioning.
25
human development. Climate change increases the urgency of understanding this issue, as
individuals will be increasingly exposed to abnormal weather deviations. Mean temperatures
in India expected to increase by an amount which corresponds to several standard deviations
above their long-run mean, and precipitation is expected to become more variable (Fishman,
2016). To shed light on the this matter, we examine how the effects of weather on crime
have varied across the three decades covered by our data.
Figure 7 shows the relationship between crime and weather shocks across the three
intervals: 1971-1980; 1981-1990; and 1991-2000. The effect of negative rainfall shocks is
relatively stable, remaining around 3-4% across these decades, though with a small increase
across time. In contrast, the effect of temperature shocks drop from 6.3% in the 1970s,
down to 0.5% in the 1980s, then 4.1% in the 1990s. Table A.6 reports the evolution of the
disaggregated impacts of property and non-property crimes, as well as those of specific crime
types. The decline of the temperature effect occurs primarily for property crimes; violent
crimes, in contrast, are as responsive to positive temperature shocks in the1990s as in the
1970s (3.8%), though have no effect in the 1980s.
6 Conclusion
We provide some of the first evidence on the relationship between weather and crime in
developing countries, using a data set that covers a population of nearly a billion people,
and spanning three decades of economic and social transformation. Variation in temperature
and rainfall are found to cause substantial increases in the incidence of both property and
violent crime. Weather shocks associated with increases in crime are also generally found
to reduce agricultural incomes, plausibly suggesting a role for agricultural income shocks in
mediating the observed relationship, consistent with a lengthy literature on the economic
determinants of crime and group conflict.
To test this hypothesis more rigorously, we explore the weather-crime-agriculture rela-
tionship across an array of seasonal and geographical disaggregations of our data. None of
these tests yield evidence at odds with the agricultural hypothesis; and, for precipitation
shocks, they reveal patterns in the responsiveness of agriculture and crime that are in re-
markable accord. In the case of temperature shocks, the evidence is more mixed, providing
some evidence for the operation of an agricultural income channel, but also for the pres-
ence of other mechanisms. Due to limitations in the temporal precision of the data, we are
26
unable to draw more definitive conclusions about the relative roles of agriculture and other
mechanisms in driving the effects of temperature on crime. Additional studies, perhaps uti-
lizing micro-data on crime and income, would facilitate further progress in resolving these
questions.
Finally, our results show that, despite the higher incomes, greater access to consumption
smoothing instruments, and reduced susceptibility of agriculture to climatic variability that
accompany economic growth, crime has not become substantially less responsive to extreme
weather than it was prior to these improvements. This may be taken as evidence that,
despite India’s remarkable gains in human and economic development, the poorest members
of society continue to remain highly vulnerable to aggregate economic shocks, reinforcing
existing concerns about the equity of India’s remarkable economic growth.
27
References
Anderson, Craig A, Kathryn B Anderson, Nancy Dorr, Kristina M DeNeve, and Mindy
Flanagan. 2000. “Temperature and aggression.” Advances in experimental social psychol-
ogy 32:63–133.
Auffhammer, Maximilian, Veerabhadran Ramanathan, and Jeffrey R Vincent. 2012. “Cli-
mate change, the monsoon, and rice yield in India.” Climatic Change 111 (2):411–424.
Auliciems, A and L DiBartolo. 1995. “Domestic violence in a subtropical environment: Police
calls and weather in Brisbane.” International Journal of Biometeorology 39 (1):34–39.
Barreca, Alan, Karen Clay, Olivier Deschenes, Michael Greenstone, and Joseph S Shapiro.
2016. “Adapting to Climate Change: The Remarkable Decline in the US Temperature-
Mortality Relationship over the Twentieth Century.” Journal of Political Economy
124 (1):105–159.
Becker, Gary S. 1968. “Crime and Punishment: An Economic Approach.” Journal of
Political Economy 76 (2):169–217.
Bohlken, Anjali Thomas and Ernest John Sergenti. 2010. “Economic growth and ethnic
violence: An empirical investigation of Hindu-Muslim riots in India.” Journal of Peace
research .
Bourguignon, Francois. 2000. “Crime, violence and inequitable development.” In Annual
World Bank Conference on Development Economics 1999. 199–220.
Burgess, Robin, Olivier Deschenes, Dave Donaldson, and Michael Greenstone. 2013. “The
unequal effects of weather and climate change: Evidence from mortality in india.” Un-
published working paper .
Burke, Marshall B, Edward Miguel, Shanker Satyanath, John A Dykema, and David B
Lobell. 2009. “Warming increases the risk of civil war in Africa.” Proceedings of the
national Academy of sciences 106 (49):20670–20674.
Card, David and Gordon B Dahl. 2011. “Family violence and football: The effect of
unexpected emotional cues on violent behavior.” The Quarterly Journal of Economics
126 (1):103.
28
Cohn, Ellen G and James Rotton. 1997. “Assault as a function of time and temperature:
A moderator-variable time-series analysis.” Journal of Personality and Social Psychology
72 (6):1322.
Collier, Paul and Anke Hoeffler. 1998. “On economic causes of civil war.” Oxford economic
papers 50 (4):563–573.
———. 2004. “Greed and grievance in civil war.” Oxford economic papers 56 (4):563–595.
Colmer, Jonathan. 2016. “Weather, labour reallocation, and industrial production: Evidence
from india.” Tech. rep., Mimeo.
Conley, Timothy G and Francesca Molinari. 2007. “Spatial correlation robust inference with
errors in location or distance.” Journal of Econometrics 140 (1):76–96.
Dell, Melissa, Benjamin F Jones, and Benjamin A Olken. 2012. “Temperature shocks and
economic growth: Evidence from the last half century.” American Economic Journal:
Macroeconomics 4 (3):66–95.
———. 2014. “What do we learn from the weather? The new climate–economy literature.”
Journal of Economic Literature 52 (3):740–798.
Deryugina, Tatyana and Solomon M Hsiang. 2014. “Does the environment still matter?
Daily temperature and income in the United States.” Tech. rep., National Bureau of
Economic Research.
Deschenes, Olivier and Michael Greenstone. 2007. “The economic impacts of climate change:
evidence from agricultural output and random fluctuations in weather.” The American
Economic Review 97 (1):354–385.
Duflo, Esther and Rohini Pande. 2007. “Dams.” The Quarterly Journal of Economics
122 (2):601–646.
Fajnzylber, Pablo, Daniel Lederman, and Norman Loayza. 1998. Determinants of crime
rates in Latin America and the world: an empirical assessment. World Bank Publications.
Fetzer, Thiemo. 2014. “Social insurance and conflict: evidence from India.” Job Market
Paper, November 14.
Fishman, Ram. 2016. “More uneven distributions overturn benefits of higher precipitation
for crop yields.” Environmental Research Letters 11 (2):024004.
29
Freeman, Richard B. 1999. “The economics of crime.” Handbook of labor economics 3:3529–
3571.
Gould, Eric D, Bruce A Weinberg, and David B Mustard. 2002. “Crime rates and local
labor market opportunities in the United States: 1979–1997.” Review of Economics and
statistics 84 (1):45–61.
Grogger, Jeff. 1997. “Market wages and youth crime.” Tech. rep., National Bureau of
Economic Research.
Guiteras, Raymond. 2009. “The impact of climate change on Indian agriculture.”
Manuscript, Department of Economics, University of Maryland, College Park, Maryland
.
Hidalgo, F Daniel, Suresh Naidu, Simeon Nichter, and Neal Richardson. 2010. “Economic
determinants of land invasions.” The Review of Economics and Statistics 92 (3):505–523.
Hsiang, Solomon M. 2010. “Temperatures and cyclones strongly associated with economic
production in the Caribbean and Central America.” Proceedings of the National Academy
of sciences 107 (35):15367–15372.
Hsiang, Solomon M, Marshall Burke, and Edward Miguel. 2013. “Quantifying the influence
of climate on human conflict.” Science 341 (6151):1235367.
Jacob, Brian, Lars Lefgren, and Enrico Moretti. 2007. “The dynamics of criminal behavior
evidence from weather shocks.” Journal of Human resources 42 (3):489–527.
Jayachandran, Seema. 2006. “Selling labor low: Wage responses to productivity shocks in
developing countries.” Journal of political Economy 114 (3):538–575.
Kaur, Supreet. 2014. “Nominal wage rigidity in village labor markets.” Tech. rep., National
Bureau of Economic Research.
Kenrick, Douglas T and Steven W MacFarlane. 1986. “Ambient temperature and horn
honking a field study of the heat/aggression relationship.” Environment and behavior
18 (2):179–191.
Kumar, Hemanshu and Rohini Somanathan. 2009. “Mapping Indian districts across census
years, 1971-2001.” Centre for Development Economics Working Paper (176).
30
Larrick, Richard P, Thomas A Timmerman, Andrew M Carton, and Jason Abrevaya. 2011.
“Temper, temperature, and temptation heat-related retaliation in baseball.” Psychological
Science .
Lobell, David B, Wolfram Schlenker, and Justin Costa-Roberts. 2011. “Climate trends and
global crop production since 1980.” Science 333 (6042):616–620.
Machin, Stephen and Costas Meghir. 2004. “Crime and economic incentives.” Journal of
Human Resources 39 (4):958–979.
Mares, Dennis. 2013. “Climate change and levels of violence in socially disadvantaged neigh-
borhood groups.” Journal of Urban Health 90 (4):768–783.
Mehlum, Halvor, Edward Miguel, and Ragnar Torvik. 2006. “Poverty and crime in 19th
century Germany.” Journal of Urban Economics 59 (3):370–388.
Miguel, Edward. 2005. “Poverty and witch killing.” The Review of Economic Studies
72 (4):1153–1172.
Miguel, Edward, Shanker Satyanath, and Ernest Sergenti. 2004. “Economic shocks and civil
conflict: An instrumental variables approach.” Journal of political Economy 112 (4):725–
753.
Pakes, Ariel and Zvi Griliches. 1980. “Patents and R&D at the firm level: A first report.”
Economics letters 5 (4):377–381.
Papaioannou, Kostadis J. 2016. “’Hunger Makes a Thief of Any Man’: Poverty and Crime
in British Colonial Asia.” Available at SSRN 2780577 .
Park, Jisung. 2016. “’Human Stress and Human Capital Production.” Unpublished
Manuscript, Harvard University Economic Department, Working Paper .
Rajeevan, M, Jyoti Bhate, JD Kale, and B Lal. 2005. “Development of a high resolution
daily gridded rainfall data for the Indian region.” Met. Monograph Climatology 22:2005.
Ranson, Matthew. 2014. “Crime, weather, and climate change.” Journal of environmental
economics and management 67 (3):274–302.
Rotton, James and Ellen G Cohn. 2000. “Violence is a curvilinear function of temperature
in Dallas: a replication.” Journal of personality and social psychology 78 (6):1074.
31
Sarsons, Heather. 2015. “Rainfall and conflict: A cautionary tale.” Journal of development
Economics 115:62–72.
Schlenker, Wolfram, W Michael Hanemann, and Anthony C Fisher. 2006. “The impact of
global warming on US agriculture: an econometric analysis of optimal growing conditions.”
Review of Economics and statistics 88 (1):113–125.
Schlenker, Wolfram and David B Lobell. 2010. “Robust negative impacts of climate change
on African agriculture.” Environmental Research Letters 5 (1):014010.
Sekhri, Sheetal and Adam Storeygard. 2014. “Dowry deaths: Response to weather variability
in India.” Journal of development economics 111:212–223.
Srivastava, AK, M Rajeevan, and SR Kshirsagar. 2009. “Development of a high resolution
daily gridded temperature data set (1969–2005) for the Indian region.” Atmospheric
Science Letters 10 (4):249–254.
Sudarshan, Anant and Meenu Tewari. 2013. “The economic impacts of temperature on
industrial productivity: Evidence from indian manufacturing.” Tech. rep., Working Paper.
Zivin, Joshua Graff and Matthew Neidell. 2014. “Temperature and the allocation of time:
Implications for climate change.” Journal of Labor Economics 32 (1):1–26.
Zivin, Joshua S Graff, Solomon M Hsiang, and Matthew J Neidell. 2015. “Temperature
and human capital in the short-and long-run.” Tech. rep., National Bureau of Economic
Research.
32
0"
50"
100"
150"
200"
250"
300"
0"
5"
10"
15"
20"
25"
Jan" Feb" Mar" Apr" May" Jun" Jul" Aug" Sep" Oct" Nov" Dec"Precipita
?on"(m
m)"
Tempe
rature"(D
egrees"Celsiu
s)"
Figure 1: Average climate in India. Monthly average temperature (black drop lines) andprecipitation (bars). Monsoon crops are usually sown June-September and harvested be-tween October and December. The blue and green shades represent the breakup of weatherindicators used in the empirical analysis
33
Crime&Observa,on&(Annual)&
}&Pre6Monsoon&
Temp.&&Shocks&
}&Monsoon&&Rain+Temp.&
Shocks&
Second&Monsoon&&Rain&
&Shocks.&&
}&}&
Monsoon&&Rain+Temp.&&
Shocks&&
}&Jun&&Jul&&Aug&&Sep&&Oct&&Nov&&Dec&&&Jan&&Feb&&Mar&&Apr&&May&&Jun&&Jul&&Aug&&Sep&&Oct&&Nov&&Dec&&&&&
Main&&Season&
Secondary&&Season&
Main&&Season&
Secondary&&Season&
Second&Monsoon&&Rainfall&&Shocks.&
&
Ag.&Produc,on&Observa,on&(Annual)&Ag.&Produc,on&Observa,on&(Annual)&
Figure 2: Temporal Structure of the model specification.
34
Figure 3: Local regressions of crime incidence on standardized (z-score) Monsoon temper-ature (degree-days, top panel) and rainfall (bottom panel) deviations. To produce theseplots, the pooled (logarithm of) crime rate and the weather deviation in question are bothregressed on all other explanatory variables in equation (2). The residual terms from thecrime regression is then locally regressed on the residual from the weather deviation re-gression using a locally weighted polynomial regression with Epanechnikov kernel functions(using the STATA command lpolyci). Note: 95% confidence intervals are based on errors(dashed lines) that are not clustered as in our main specifications.
35
Figu
re4:
Loca
lreg
ress
ions
ofag
ricu
ltur
alpr
oduc
tion
(top
pane
ls)a
ndw
ages
(bot
tom
pane
ls)o
nst
anda
rdiz
ed(z
-sco
re)M
onso
onte
mpe
ratu
re(d
egre
e-da
ys,r
ight
pane
ls)
and
rain
fall
(lef
tpa
nels
)de
viat
ions
.To
prod
uce
thes
epl
ots,
the
agri
cultur
alou
tcom
ean
dth
ew
eath
erde
viat
ion
inqu
estion
are
both
regr
esse
don
allo
ther
expl
anat
ory
vari
able
sin
equa
tion
(1).
The
resi
dual
term
from
the
agri
cult
ural
outc
ome
regr
essi
onis
then
regr
esse
don
the
resi
dual
from
the
wea
ther
devi
atio
nre
gres
sion
usin
ga
loca
llyw
eigh
ted
poly
nom
ialr
egre
ssio
nw
ith
Epa
nech
niko
vke
rnel
func
tion
s(u
sing
the
STA
TAco
mm
and
lpol
yci).
Not
e:95
%co
nfide
nce
inte
rval
sar
eba
sed
oner
rors
(das
hed
lines
)th
atar
eno
tcl
uste
red
asin
our
mai
nsp
ecifi
cation
s.
36
Estim
ated
Impa
ct o
n C
rime
Rat
es
-0.05
0
0.05
0.1
0.15
Burglar
y
Bandit
ryThe
fts
Robbe
ryRiot
s
Kidnap
ping
Rape
Murder
Estim
ated
Impa
ct o
n C
rime
Rat
es
-0.05
0
0.05
0.1
0.15
Burglar
y
Bandit
ryThe
fts
Robbe
ryRiot
s
Kidnap
ping
Rape
Murder
Negative Rainfall Shocks
Positive Temperature Shocks
Figure 5: Estimated coefficients from regressions of disaggregated crime rates on negativemonsoon rainfall shocks (top) and positive monsoon temperature shocks (bottom). Propertycrimes coefficients are indicated by green markers, and violent crime coefficients are indicatedby red markers Error bars depict 95% confidence intervals.
37
Figure 6: Estimated coefficients from regressions of pooled crime rates (horizontal axis) andagricultural product (vertical axis) for Monsoon and Pre-Monsoon weather shocks for whichsignificant impacts on crime were found. Each square represents one point estimate. Allestimates are obtained from table 2
38
Figure 7: Coefficients from regressions of pooled crime incidence on Monsoon negative rain-fall and positive temperatures variables indicating weather deviations more than 1 standarddeviation from the mean. To produce these plots equation 2 was estimated for each decadeindependently. The coefficients from the regressions were then plotted, along with the 95%confidence intervals for these coefficients.
39
Table 1: Summary Statistics1970-2000 1970s 1980s 1990s
Mean S.D. Mean Mean Mean
Crime Rate (per 100,000 pop.)
Burglary 20.5 (19.1) 33 20.2 13.2
Banditry 1.4 (2.0) 2.1 1.6 0.9
Theft 42.2 (46.2) 64.2 42.7 28.2
Robbery 3.2 (3.7) 4 3.4 2.5
Riots 11.6 (11.7) 11.9 12.8 10.5
Kidnapping 2 (2.2) 1.9 1.9 2.2
Rape 1.2 (1.3) 0.7 1.1 1.6
Murder 3.8 (2.8) 3.3 3.7 4.2
Weather
Monsoon Rainfall (mm) 855 (490) 837 846 873
Monsoon Temerpature (Degree Days per day) 20.86 (1.47) 20.68 20.88 20.93
Agriculture
Annual Production (Rupees, 1960-65 prices) 22,801 (23,091) 17,582 22,854 26,845
Yield (Annual Production per Hectare, Rupees) 76 (51) 55 72 98
Agricultural Wage Labour (Ruppes per hour) 5.62 (3.25) 4.29 5.41 7.29
40
Tabl
e2:
The
Effe
cts
ofW
eath
erSh
ocks
onA
gric
ultu
rean
dC
rim
eC
rim
eP
rodu
ctY
ield
Wag
eSe
ason
alY
ield
Ann
ual
Ann
ual
Mai
nSe
cond
ary
(1)
(2)
(3)
(4)
(5)
(6)
Monsoon
Neg
Rai
n0.
048*
**-0
.150
***
-0.0
95**
*-0
.025
**-0
.159
***
-0.0
47**
*(0
.010
)(0
.023
)(0
.015
)(0
.011
)(0
.020
)(0
.013
)Po
sR
ain
0.01
5*-0
.028
-0.0
24-0
.026
**-0
.017
0.01
3(0
.009
)(0
.021
)(0
.015
)(0
.013
)(0
.013
)(0
.010
)N
egTe
mp
0.01
40.
039*
*0.
029*
*0.
023
0.00
1-0
.007
(0.0
11)
(0.0
16)
(0.0
13)
(0.0
16)
(0.0
16)
(0.0
13)
Pos
Tem
p0.
033*
*-0
.080
**-0
.016
-0.0
48**
*-0
.002
0.00
1(0
.015
)(0
.032
)(0
.018
)(0
.015
)(0
.022
)(0
.014
)P
re-M
onsoon
Neg
Tem
p0.
010
0.03
9**
0.00
70.
004
0.02
70.
015
(0.0
13)
(0.0
20)
(0.0
13)
(0.0
12)
(0.0
19)
(0.0
13)
Pos
Tem
p0.
019*
-0.0
34-0
.029
-0.0
050.
004
-0.0
36**
*(0
.011
)(0
.036
)(0
.019
)(0
.014
)(0
.014
)(0
.014
)
R-S
quar
ed0.
131
0.09
20.
089
0.15
50.
117
0.09
9N
o.of
Obs
erva
tion
s73
295
6503
6502
3892
5920
7294
Col
umn
1re
port
ses
tim
ates
ofre
gres
sion
(2)
whi
chin
clud
esw
eath
ersh
ocks
inth
eM
onso
onan
dP
re-M
onso
onse
ason
s.C
olum
ns2-
4re
port
esti
mat
esof
regr
essi
ons
assp
ecifi
edin
equa
tion
(1)in
whi
chth
eou
tcom
eva
riab
les
are
the
loga
rith
mof
annu
alag
ricu
ltur
alpr
oduc
tion
(col
umn
2),y
ield
(col
umn
3)an
dw
ages
(col
umn
4).
Inco
lum
ns(5
)-(6
),si
mila
rre
gres
sion
sar
ees
tim
ated
inw
hich
the
outc
omes
are
the
yiel
dsof
the
prim
ary
crop
inth
em
ain
and
seco
ndar
yse
ason
s.St
anda
rder
rors
,clu
ster
edin
tim
ean
dsp
ace
(see
text
),ar
ere
port
edin
pare
nthe
ses.
Star
sin
dica
test
atis
tica
lsig
nific
ance
:*
p<
0.1,
**p
<0.
05,*
**p
<0.
01.
41
Tabl
e3:
The
Effe
cts
ofW
eath
erSh
ocks
onIn
divi
dual
Cri
me
Cat
egor
ies
Pro
pert
yN
on-P
rope
rty
Bur
glar
yB
andi
try
The
fts
Rob
bery
Rio
tsK
idna
ppin
gR
ape
Mur
der
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
Monsoon
Neg
Rai
n0.
058*
**0.
031*
**0.
07*
0.06
7***
0.07
3***
0.03
5**
0.04
2*0.
070*
**0.
034*
*0.
022
0.03
0***
(0.0
13)
(0.0
11)
(0.0
18)
(0.0
22)
(0.0
18)
(0.0
25)
(0.0
26)
(0.0
17)
(0.0
20)
(0.0
11)
Pos
Rai
n0.
016
0.01
20.
700.
008
0.04
7**
0.01
80.
004
0.00
30.
021
0.00
80.
003
(0.0
11)
(0.0
10)
(0.0
12)
(0.0
21)
(0.0
12)
(0.0
18)
(0.0
18)
(0.0
16)
(0.0
17)
(0.0
10)
Neg
Tem
p0.
023
-0.0
010.
180.
013
0.04
50.
011
0.05
3**
-0.0
09-0
.002
-0.0
050.
008
(0.0
15)
(0.0
12)
(0.0
16)
(0.0
31)
(0.0
16)
(0.0
25)
(0.0
24)
(0.0
19)
(0.0
21)
(0.0
13)
Pos
Tem
p0.
038*
*0.
024
0.32
-0.0
030.
076*
*0.
026
0.04
50.
058*
*0.
018
0.03
00.
030*
(0.0
18)
(0.0
16)
(0.0
19)
(0.0
33)
(0.0
22)
(0.0
31)
(0.0
28)
(0.0
22)
(0.0
24)
(0.0
15)
Pre-M
onsoon
Neg
Tem
p0.
011
0.00
80.
840.
033*
0.05
2*-0
.004
-0.0
19-0
.009
0.02
4-0
.021
0.02
1(0
.017
)(0
.013
)(0
.018
)(0
.027
)(0
.023
)(0
.032
)(0
.027
)(0
.020
)(0
.022
)(0
.014
)Po
sTe
mp
0.02
00.
017
0.80
0.00
10.
010
0.02
20.
051*
*0.
019
0.00
90.
036*
0.01
0(0
.013
)(0
.012
)(0
.019
)(0
.028
)(0
.014
)(0
.024
)(0
.023
)(0
.018
)(0
.021
)(0
.012
)
R-S
quar
ed0.
203
0.11
00.
420
0.50
70.
182
0.23
00.
234
0.17
90.
355
0.05
0N
o.of
Obs
erva
tion
s45
725
2757
091
7191
1791
1791
7391
4791
8691
9191
93
Col
umns
1-2
repo
rtes
tim
ates
ofre
gres
sion
(2)
whi
chin
clud
esw
eath
ersh
ocks
inth
eM
onso
onan
dP
re-M
onso
onse
ason
s,re
stri
cted
topr
oper
ty(c
olum
n1)
and
non-
prop
erty
(col
umn
2)cr
imes
.C
olum
n3
repo
rts
p-va
lues
for
diffe
renc
esbe
twee
nth
etw
ose
tsof
coeffi
cien
ts.
Col
umns
4-11
repo
rtes
tim
ates
ofre
gres
sion
sas
spec
ified
ineq
uati
on(1
)in
whi
chth
eou
tcom
eva
riab
les
are
the
loga
rith
ms
ofth
era
tes
ofin
divi
dual
crim
eca
tego
ries
.St
anda
rder
rors
,cl
uste
red
intim
ean
dsp
ace
(see
text
),ar
ere
port
edin
pare
nthe
ses.
Star
sin
dica
test
atis
tica
lsig
nific
ance
:*
p<
0.1,
**p
<0.
05,
***
p<
0.01
.
42
Table 4: The Effects of Weather Shocks on Agriculture and Crime (Multiple Bins)Crime Product Yield Wage
(1) (2) (3) (4)Rainfall
< -1.5 sd 0.064*** -0.278*** -0.172*** -0.038**(0.020) (0.046) (0.034) (0.015)
-1.5 to -1.0 0.051*** -0.109*** -0.077*** -0.014(0.012) (0.023) (0.015) (0.012)
-0.5 to -1.0 0.016 -0.045** -0.039*** -0.009(0.010) (0.020) (0.010) (0.009)
05. to 1.0 0.010 -0.006 -0.011 0.010(0.010) (0.014) (0.011) (0.012)
1.0 to 1.5 0.006 -0.014 -0.009 -0.022(0.012) (0.020) (0.014) (0.015)
> 1.5 sd 0.038*** -0.062* -0.066*** -0.031*(0.013) (0.034) (0.025) (0.019)
Temperature
< -1.5 sd -0.009 0.039 0.029 0.045**(0.021) (0.025) (0.021) (0.022)
-1.5 to -1.0 0.038*** 0.040** 0.035** 0.003(0.012) (0.018) (0.015) (0.022)
-0.5 to -1.0 0.025** 0.005 0.015 -0.005(0.012) (0.020) (0.014) (0.014)
05. to 1.0 -0.006 -0.018 -0.043** -0.008(0.011) (0.028) (0.020) (0.012)
1.0 to 1.5 0.041** -0.055 -0.020 -0.026(0.018) (0.037) (0.019) (0.016)
> 1.5 sd 0.019 -0.121** -0.043 -0.091***(0.023) (0.051) (0.029) (0.024)
R-squared 0.131 0.098 0.099 0.164No. of Observations 73295 6503 6502 3892
Column 1 reports estimates of regression (2) which includes multiple bins for weather shocksin the Monsoon and Pre-Monsoon seasons. Columns 2-4 report estimates of regressions asspecified in equation (1) in which the outcome variables are the logarithm of agriculturalproduction (column 2), yield (column 3) and wages (column 4). Standard errors, clustered intime and space (see text), are reported in parentheses. Stars indicate statistical significance:* p < 0.1, ** p < 0.05, *** p < 0.01.
43
Table 5: The Effects of Second Monsoon Weather Shocks on Agriculture and CrimeFull sample Second Monsoon Sample
Crime Product Yield Crime Product Yield(1) (2) (3) (4) (5) (6)
Monsoon
Neg Rain 0.045*** -0.150*** -0.098*** 0.055** -0.147*** -0.051**(0.010) (0.023) (0.015) (0.021) (0.045) (0.024)
Pos Rain 0.016* -0.026 -0.023 0.003 -0.006 0.005(0.009) (0.021) (0.015) (0.019) (0.040) (0.022)
Neg Temp 0.015 0.034** 0.025** 0.009 0.012 0.015(0.011) (0.016) (0.013) (0.028) (0.041) (0.023)
Pos Temp 0.035** -0.069** -0.016 0.002 -0.105* 0.008(0.016) (0.032) (0.018) (0.036) (0.057) (0.029)
Post-Monsoon
Neg Rain 0.055** -0.064* -0.052**(0.022) (0.036) (0.023)
Pos Rain 0.016 0.018 0.046**(0.021) (0.035) (0.022)
Neg Temp -0.008 0.071*** 0.064*** -0.026 0.049 0.030(0.014) (0.019) (0.015) (0.028) (0.035) (0.025)
Pos Temp 0.009 -0.086* -0.021 0.057** 0.150* 0.035(0.012) (0.052) (0.022) (0.028) (0.085) (0.030)
R-Squared 0.125 0.096 0.094 0.149 0.075 0.184No. of Observations 75071 6531 6530 15265 1242 1241
Columns 1-3 reports estimates of regressions for which the outcomes are crime (pooled),agricultural product and yield, and which control for Monsoon (June-September) rainfalland temperature shocks and for Post-Monsoon (October-December) temperature shocks.Columns 4-6 report similar estimates except that the sample is limited to those districtsthat receive a second monsoon (figure A.3) and the regressions also control for Post-Monsoonrainfall weather shocks. Standard errors, clustered in time and space (see text), are reportedin parentheses. Stars indicate statistical significance: * p < 0.1, ** p < 0.05, *** p < 0.01.
44
Table 6: The Heterogenous Effects of Weather Shocks on Crime by Dry Season CultivationAll Low Rabi High Rabi(1) (2) (3)
Monsoon
Neg Rain 0.048*** 0.034*** 0.059***(0.010) (0.013) (0.013)
Pos Rain 0.015* 0.007 0.025**(0.009) (0.013) (0.011)
Neg Temp 0.014 0.023 0.008(0.011) (0.015) (0.015)
Pos Temp 0.033** 0.054*** 0.021(0.015) (0.019) (0.018)
Pre-Monsoon
Neg Temp 0.010 0.011 0.045***(0.013) (0.016) (0.017)
Pos Temp 0.019* 0.010 0.033**(0.011) (0.012) (0.015)
R-Squared 0.131 0.134 0.138No. of Observations 73295 33374 39178
Columns 1-3 report estimates of regression 2 for the full sample (column 1), areas with high(column 2) and low (column 3) extent of cultivation in the dry season. All regression controlfor monsoon and pre-monsoon season weather shocks. Standard errors, clustered in timeand space (see text), are reported in parentheses. Stars indicate statistical significance: * p< 0.1, ** p < 0.05, *** p < 0.01.
45
Table 7: The Effects of Weather Shocks on Crime in Wet and Dry AreasCrime
Dry Wet(3) (4)
Share, Dryland Crops 0.56 0.22
Neg Rain 0.058*** 0.028**(0.016) (0.013)
Pos Rain 0.035** -0.007(0.014) (0.011)
Neg Temp -0.014 0.042***(0.019) (0.012)
Pos Temp 0.008 0.063***(0.024) (0.018)
R-Squared 0.145 0.110No. of Observations 38058 35395
Column 1-2 compare estimates of regression (2) which includes monsoon season weathershocks between the “wet” and “dry” sub-samples (see text for definitions). The first rowreports the average share occupied by “dryland crops” in cultivated area in the two sub-samples. Standard errors, clustered in time and space (see text), are reported in parentheses.Stars indicate statistical significance: * p < 0.1, ** p < 0.05, *** p < 0.01.
46
Additional Figures and Tables
47
Figu
reA
.1:
The
spat
iald
istr
ibut
ion
ofav
erag
e19
71-2
000
crim
era
tes
(cri
mes
com
mit
ted
per
100,
000
popu
lati
on).
48
Figure A.2: Plots of the (logarithm) of India-wide average crime incidence (per 100,000population) over time.
49
Figure A.3: A map of the districts included in the “Second Monsoon” analysis.
50
Figu
reA
.4:
Loca
lreg
ress
ions
ofin
divi
dual
crim
eca
tego
ryin
cide
nce
onst
anda
rdiz
ed(z
-sco
re)
mon
soon
tem
pera
ture
(deg
ree-
days
).To
prod
uce
thes
epl
ots,
the
pool
ed(log
arithm
of)
crim
era
tean
dth
ew
eath
erde
viat
ion
inqu
estion
are
both
regr
esse
don
allo
ther
expl
anat
ory
vari
able
sin
equa
tion
(2).
The
resi
dual
term
from
the
crim
ere
gres
sion
isth
enre
gres
sed
onth
ere
sidu
alfr
omth
ew
eath
erde
viat
ion
regr
essi
onus
ing
alo
cally
wei
ghte
dpo
lyno
mia
lreg
ress
ion
wit
hE
pane
chni
kov
kern
elfu
ncti
ons
(usi
ngth
eST
ATA
com
man
dlp
olyc
i).
Not
e:95
%co
nfide
nce
inte
rval
sar
eba
sed
oner
rors
(das
hed
lines
)th
atar
eno
tcl
uste
red
asin
our
mai
nsp
ecifi
cati
ons.
51
Figu
reA
.5:
Loca
lreg
ress
ions
ofin
divi
dual
crim
eca
tego
ryin
cide
nce
onst
anda
rdiz
ed(z
-sco
re)
mon
soon
rain
fall.
Topr
oduc
eth
ese
plot
s,th
epo
oled
(log
arit
hmof
)cr
ime
rate
and
the
wea
ther
devi
atio
nin
ques
tion
are
both
regr
esse
don
allo
ther
expl
anat
ory
vari
able
sin
equa
tion
(2).
The
resi
dual
term
from
the
crim
ere
gres
sion
isth
enre
gres
sed
onth
ere
sidu
alfr
omth
ew
eath
erde
viat
ion
regr
essi
onus
ing
alo
cally
wei
ghte
dpo
lyno
mia
lreg
ress
ion
wit
hE
pane
chni
kov
kern
elfu
ncti
ons
(usi
ngth
eST
ATA
com
man
dlp
olyc
i).
Not
e:95
%co
nfide
nce
inte
rval
sar
eba
sed
oner
rors
(das
hed
lines
)th
atar
eno
tcl
uste
red
asin
our
mai
nsp
ecifi
cation
s.
52
Tabl
eA
.1:
The
Effe
cts
ofW
eath
erSh
ocks
onIn
divi
dual
Cro
ps’Y
ield
Mon
soon
Cro
psD
rySe
ason
Cro
psR
ice
Jow
arB
ajra
Mai
zeA
rhar
Ric
eJo
war
Baj
raM
aize
Whe
at(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)(9
)(1
0)M
onsoon
Neg
Rain
-0.1
76**
*-0
.110
***
-0.2
02**
*-0
.105
***
-0.0
92**
*-0
.021
-0.1
09**
*-0
.146
**-0
.025
-0.0
59**
*(0
.023
)(0
.028
)(0
.044
)(0
.032
)(0
.021
)(0
.016
)(0
.033
)(0
.064
)(0
.027
)(0
.014
)Pos
Rain
0.01
0-0
.054
***
-0.0
41-0
.076
**0.
007
0.01
3-0
.008
-0.0
230.
011
0.02
3**
(0.0
12)
(0.0
19)
(0.0
36)
(0.0
32)
(0.0
17)
(0.0
18)
(0.0
22)
(0.0
35)
(0.0
17)
(0.0
10)
Neg
Tem
p0.
024
0.01
10.
163*
*0.
017
0.07
1***
0.00
1-0
.016
0.07
20.
017
0.02
2*(0
.018
)(0
.025
)(0
.066
)(0
.040
)(0
.027
)(0
.019
)(0
.042
)(0
.076
)(0
.026
)(0
.012
)Pos
Tem
p0.
001
-0.0
10-0
.075
-0.0
48-0
.024
0.02
70.
043
-0.0
67-0
.047
*0.
019
(0.0
25)
(0.0
35)
(0.0
61)
(0.0
36)
(0.0
25)
(0.0
26)
(0.0
41)
(0.0
66)
(0.0
27)
(0.0
14)
Pre-M
onsoon
Neg
Tem
p0.
028
0.07
3**
0.00
50.
011
-0.0
35-0
.006
0.04
9-0
.039
0.00
1-0
.004
(0.0
20)
(0.0
30)
(0.0
47)
(0.0
35)
(0.0
22)
(0.0
17)
(0.0
31)
(0.0
59)
(0.0
32)
(0.0
13)
Pos
Tem
p0.
013
0.00
60.
043
0.12
0***
0.02
7-0
.039
**-0
.037
-0.0
040.
013
-0.0
41**
*(0
.017
)(0
.025
)(0
.041
)(0
.033
)(0
.022
)(0
.016
)(0
.027
)(0
.058
)(0
.023
)(0
.016
)
R-Squared
0.15
70.
046
0.05
60.
062
0.09
60.
111
0.08
20.
117
0.34
10.
156
N57
8447
0512
9524
7667
3723
0019
6638
611
0869
85
Col
umns
1-10
repo
rtes
tim
ates
ofre
gres
sion
ofth
eyi
elds
ofin
divi
dual
crop
son
mon
soon
and
Pre
-mon
soon
wea
ther
shoc
ks.
Stan
dard
erro
rs,c
lust
ered
inti
me
and
spac
e(s
eete
xt),
are
repo
rted
inpa
rent
hese
s.St
ars
indi
cate
stat
isti
cals
igni
fican
ce:
*p
<0.
1,**
p<
0.05
,***
p<
0.01
.
53
Tabl
eA
.2:
The
Effe
cts
ofW
eath
erSh
ocks
onIn
divi
dual
Cri
me
Cat
egor
ies
(Mul
tipl
eB
ins)
Pro
pert
yN
on-P
rope
rty
Bur
glar
yB
andi
try
The
fts
Rob
bery
Rio
tsK
idna
ppin
gR
ape
Mur
der
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Rainfall
<-1
.5Sd
0.08
2***
0.03
5*0.
045
0.13
0***
0.05
3**
0.06
30.
119*
**0.
030
0.01
00.
045*
*(0
.025
)(0
.020
)(0
.030
)(0
.045
)(0
.026
)(0
.048
)(0
.042
)(0
.032
)(0
.031
)(0
.021
)-1
.5To
-1.0
0.06
2***
0.03
3**
0.07
7***
0.08
3***
0.03
7*0.
054*
*0.
059*
0.04
8**
0.02
10.
024*
*(0
.015
)(0
.013
)(0
.019
)(0
.027
)(0
.022
)(0
.026
)(0
.033
)(0
.019
)(0
.024
)(0
.012
)-0
.5To
-1.0
0.01
40.
017
-0.0
010.
032
0.01
30.
016
0.01
10.
024
0.01
30.
014
(0.0
12)
(0.0
10)
(0.0
16)
(0.0
23)
(0.0
13)
(0.0
19)
(0.0
19)
(0.0
15)
(0.0
17)
(0.0
11)
05.
To1.
00.
024*
*-0
.013
-0.0
120.
075*
**0.
018
0.04
6**
-0.0
030.
004
-0.0
45**
-0.0
03(0
.012
)(0
.011
)(0
.015
)(0
.023
)(0
.012
)(0
.020
)(0
.018
)(0
.017
)(0
.018
)(0
.011
)1.
0To
1.5
0.00
8-0
.000
-0.0
120.
029
-0.0
01-0
.006
0.01
80.
009
-0.0
11-0
.009
(0.0
14)
(0.0
13)
(0.0
17)
(0.0
28)
(0.0
15)
(0.0
23)
(0.0
19)
(0.0
21)
(0.0
21)
(0.0
13)
>1.
5Sd
0.04
4***
0.02
8*0.
021
0.11
9***
0.05
6***
0.04
5*-0
.013
0.04
5*0.
017
0.02
0(0
.016
)(0
.014
)(0
.017
)(0
.030
)(0
.018
)(0
.027
)(0
.028
)(0
.024
)(0
.024
)(0
.014
)Tem
perature
<-1
.5Sd
-0.0
07-0
.013
-0.0
19-0
.037
-0.0
040.
028
-0.0
13-0
.004
-0.0
320.
001
(0.0
28)
(0.0
23)
(0.0
27)
(0.0
48)
(0.0
29)
(0.0
43)
(0.0
40)
(0.0
29)
(0.0
38)
(0.0
23)
-1.5
To-1
.00.
050*
**0.
019
0.04
8***
0.07
1**
0.02
60.
097*
**0.
012
0.00
10.
022
0.03
5**
(0.0
16)
(0.0
15)
(0.0
17)
(0.0
36)
(0.0
16)
(0.0
28)
(0.0
28)
(0.0
24)
(0.0
25)
(0.0
15)
-0.5
To-1
.00.
027*
0.02
1*0.
034*
*0.
001
0.01
40.
059*
*0.
027
0.01
30.
011
0.03
3***
(0.0
15)
(0.0
12)
(0.0
16)
(0.0
30)
(0.0
16)
(0.0
24)
(0.0
22)
(0.0
18)
(0.0
22)
(0.0
13)
05.
To1.
0-0
.011
0.00
3-0
.021
-0.0
300.
013
-0.0
02-0
.017
-0.0
280.
012
0.02
2*(0
.013
)(0
.012
)(0
.017
)(0
.025
)(0
.014
)(0
.026
)(0
.020
)(0
.018
)(0
.019
)(0
.012
)1.
0To
1.5
0.05
2**
0.02
60.
005
0.08
6**
0.04
5*0.
071*
0.06
7**
0.00
70.
041
0.02
7(0
.021
)(0
.020
)(0
.025
)(0
.036
)(0
.027
)(0
.042
)(0
.032
)(0
.026
)(0
.030
)(0
.021
)>
1.5
Sd0.
015
0.02
6-0
.022
0.03
90.
015
0.02
40.
023
0.01
10.
018
0.06
3***
(0.0
26)
(0.0
23)
(0.0
32)
(0.0
46)
(0.0
31)
(0.0
46)
(0.0
38)
(0.0
32)
(0.0
38)
(0.0
20)
R-S
quar
ed0.
204
0.11
10.
421
0.50
90.
184
0.23
10.
235
0.18
00.
356
0.05
3N
4572
527
570
9171
9117
9117
9173
9147
9186
9191
9193
Col
umns
1-2
repo
rtes
tim
ates
ofre
gres
sion
(2)w
hich
incl
udes
wea
ther
shoc
ksin
the
mon
soon
and
Pre
-mon
soon
seas
ons,
rest
rict
edto
prop
erty
(col
umn
1)an
dno
n-pr
oper
ty(c
olum
n2)
crim
es.
Col
umns
3-10
repo
rtes
tim
ates
ofre
gres
sion
sin
whi
chth
eou
tcom
eva
riab
les
are
the
loga
rith
ms
ofth
era
tes
ofin
divi
dual
crim
eca
tego
ries
.In
allr
egre
ssio
ns,m
ulti
ple
bins
for
the
valu
eof
stan
dard
ized
wea
ther
shoc
ksin
the
mon
soon
and
Pre
-mon
soon
seas
ons
are
incl
uded
.St
anda
rder
rors
,clu
ster
edin
tim
ean
dsp
ace
(see
text
),ar
ere
port
edin
pare
nthe
ses.
Star
sin
dica
test
atis
tica
lsig
nific
ance
:*
p<
0.1,
**p
<0.
05,*
**p
<0.
01.
54
Table A.3: The Effects of Lagged Weather Shocks on CrimeEntireSample
SecondaryMonsoon
(1) (2)Same Year
Monsoon
Neg Rain 0.047*** 0.051**(0.010) (0.021)
Pos Temp 0.031** 0.005(0.015) (0.038)
Post-Monsoon
Neg Rain 0.067***(0.022)
Pos Temp 0.059**(0.029)
Previous Year
Monsoon
Neg Rain 0.012 0.020(0.010) (0.019)
Pos Temp -0.005 -0.001(0.012) (0.025)
Post-Monsoon
Neg Rain 0.067***(0.022)
Pos Temp 0.005(0.031)
R-Squared 0.131 0.153No. of Observations 73295 15041
Columns 1 and 2 reports estimates of regression (2) which include weather shocks in the yearof observation and the previous year. Regressions reported in column 1 include Monsoonseason shocks, and those reported in column 2 also include Post-Monsoon season shocks. Incolumns 2, the sample is limited to those districts that receive the second monsoon (figure??). Standard errors, clustered in time and space (see text), are reported in parentheses.Stars indicate statistical significance: * p < 0.1, ** p < 0.05, *** p < 0.01.
55
Tabl
eA
.4:
The
Effe
cts
ofW
eath
erSh
ocks
onC
rim
ean
dA
gric
ultu
rein
Wet
and
Dry
Are
as,M
ulti
ple
Rai
nfal
lBin
sC
rim
eA
gric
ultu
ralP
rodu
ctIn
divi
dual
Agr
icul
tura
lYie
lds
Dry
Wet
Dry
Wet
Dry
land
Cro
psW
etC
rops
Jow
arB
ajra
Mai
zeR
ice
Arh
ar(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)(9
)R
ainfall
<-1
.5Sd
0.06
2*0.
056*
*-0
.345
***
-0.1
94**
*-0
.232
***
-0.3
34**
*-0
.224
***
-0.2
31**
*-0
.136
***
(0.0
33)
(0.0
22)
(0.0
78)
(0.0
36)
(0.0
53)
(0.0
59)
(0.0
71)
(0.0
42)
(0.0
43)
-1.5
To-1
.00.
069*
**0.
022
-0.1
10**
*-0
.083
***
-0.1
06**
*-0
.209
***
-0.0
83**
-0.1
74**
*-0
.076
***
(0.0
18)
(0.0
15)
(0.0
30)
(0.0
32)
(0.0
28)
(0.0
52)
(0.0
32)
(0.0
26)
(0.0
24)
-0.5
To-1
.00.
012
0.01
3-0
.045
**-0
.036
-0.0
68**
*-0
.101
***
-0.0
00-0
.065
***
-0.0
33**
(0.0
14)
(0.0
13)
(0.0
22)
(0.0
34)
(0.0
19)
(0.0
36)
(0.0
23)
(0.0
11)
(0.0
17)
05.
To1.
00.
003
0.02
0*-0
.034
*0.
014
-0.0
350.
056
-0.0
350.
026*
*0.
033*
(0.0
15)
(0.0
12)
(0.0
19)
(0.0
17)
(0.0
21)
(0.0
36)
(0.0
28)
(0.0
11)
(0.0
17)
1.0
To1.
50.
021
-0.0
10-0
.024
-0.0
15-0
.050
**-0
.000
-0.0
67*
0.01
1-0
.016
(0.0
18)
(0.0
12)
(0.0
29)
(0.0
25)
(0.0
22)
(0.0
42)
(0.0
36)
(0.0
14)
(0.0
21)
>1.
5Sd
0.05
8***
0.01
0-0
.093
**-0
.001
-0.0
99**
*-0
.081
*-0
.098
*-0
.010
0.03
8(0
.019
)(0
.015
)(0
.045
)(0
.026
)(0
.030
)(0
.048
)(0
.050
)(0
.017
)(0
.024
)
R-S
quar
ed0.
146
0.11
10.
106
0.11
10.
052
0.07
70.
058
0.17
00.
100
N38
058
3539
535
8729
2847
0512
9524
7657
8467
37
Aco
mpa
riso
nof
the
effec
tsof
rain
fall
shoc
kson
crim
ean
dag
ricu
ltur
ein
the
“wet
”an
d“d
ry”
sub-
sam
ples
(see
text
for
defin
itio
ns)
usin
gm
ulti
ple
shoc
kca
tego
ries
(or
z-sc
ore
bins
).C
olum
ns1-
2re
port
esti
mat
esin
whi
chth
eou
tcom
eva
riab
leis
the
pool
edcr
ime
rate
.C
olum
ns3-
4re
port
esti
mat
esin
whi
chth
eou
tcom
eva
riab
leis
the
loga
rith
mof
annu
alag
ricu
ltur
alpr
oduc
t.C
olum
ns5-
9re
port
esti
mat
esin
whi
chth
eou
tcom
esar
eth
eyi
elds
ofth
em
ain
“dry
land
crop
s”(c
olum
ns5-
7)an
d“w
etcr
ops”
(col
umns
8-9)
.St
anda
rder
rors
,cl
uste
red
inti
me
and
spac
e(s
eete
xt),
are
repo
rted
inpa
rent
hese
s.St
ars
indi
cate
stat
istica
lsig
nific
ance
:*
p<
0.1,
**p
<0.
05,*
**p
<0.
01.
56
Table A.5: The Effects of Weather Shocks in Irrigated DistrictsCrime Product Yield Wage
(1) (2) (3) (4)Neg Rain 0.051*** -0.046*** -0.024** -0.016
(0.016) (0.017) (0.012) (0.011)Pos Rain 0.024** -0.033** -0.005 -0.013
(0.012) (0.015) (0.011) (0.011)
Upstream Dam ⇥ Neg Rain -0.037 0.093 0.111** 0.019(0.045) (0.065) (0.050) (0.019)
Upstream Dam ⇥ Pos Rain -0.018 0.028 0.002 -0.015(0.027) (0.047) (0.036) (0.019)
R-Squared 0.911 0.917 0.906 0.876No. of Observations 55322 6487 6486 3881
Column 1 reports estimates of regression (2) which includes monsoon rainfall shocks andtheir interaction with the presence of an irrigation dam upstream (the presence of damswithin the district is also controlled for). Columns 2-4 report similar regressions in which theoutcome variables are the logarithm of agricultural production (column 2), yield (column 3)and wages (column 4). Standard errors, clustered in time and space (see text), are reportedin parentheses. Stars indicate statistical significance: * p < 0.1, ** p < 0.05, *** p < 0.01.
57
Tabl
eA
.6:
The
Effe
cts
ofM
onso
onW
eath
erSh
ocks
onIn
divi
dual
Cri
me
Cat
egor
ies
Acr
oss
Dec
ades
Pro
pert
yN
on-P
rope
rty
Bur
glar
yB
andi
try
The
fts
Rob
bery
Rio
tsK
idna
ppin
gR
ape
Mur
der
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Negative
Rainfall
1970
S0.
036
0.01
60.
003
0.08
7**
-0.0
100.
029
0.07
0**
0.03
20.
006
0.00
7(0
.022
)(0
.022
)(0
.037
)(0
.040
)(0
.039
)(0
.035
)(0
.033
)(0
.027
)(0
.044
)(0
.021
)
1980
S0.
043*
*0.
014
0.09
0***
0.08
8**
0.03
2-0
.001
0.00
40.
015
-0.0
010.
029*
*(0
.017
)(0
.013
)(0
.023
)(0
.035
)(0
.022
)(0
.032
)(0
.049
)(0
.023
)(0
.022
)(0
.013
)
1990
S0.
039*
*0.
031*
*0.
044*
0.02
80.
027*
0.05
5**
0.04
90.
061*
*0.
006
0.02
7(0
.018
)(0
.016
)(0
.025
)(0
.031
)(0
.016
)(0
.027
)(0
.032
)(0
.024
)(0
.028
)(0
.019
)
Positive
Tem
perature
1970
S0.
079*
*0.
038
0.04
30.
123*
0.08
0**
0.01
00.
150*
**0.
020
0.07
30.
026
(0.0
31)
(0.0
26)
(0.0
32)
(0.0
64)
(0.0
39)
(0.0
47)
(0.0
51)
(0.0
35)
(0.0
46)
(0.0
25)
1980
S0.
012
-0.0
060.
019
-0.0
15-0
.044
0.11
6*-0
.018
-0.0
360.
036
-0.0
18(0
.028
)(0
.023
)(0
.035
)(0
.047
)(0
.043
)(0
.066
)(0
.047
)(0
.041
)(0
.035
)(0
.021
)
1990
S0.
046*
**0.
038*
*0.
019
0.06
3*0.
030
0.09
7***
0.03
00.
081*
**-0
.021
0.05
4***
(0.0
18)
(0.0
15)
(0.0
22)
(0.0
34)
(0.0
18)
(0.0
27)
(0.0
23)
(0.0
21)
(0.0
23)
(0.0
19)
Est
imat
edco
effici
ents
ofne
gati
vera
in(t
op6
row
s)an
dpo
siti
vete
mpe
ratu
re(b
otto
m6
row
s)sh
ocks
onpr
oper
ty(c
olum
n1)
,non
-pro
pert
y(c
olum
n2)
and
indi
vidu
alcr
ime
cate
gori
es(c
olum
ns3-
10)
inea
chof
the
thre
ede
cade
sin
the
sam
ple
(eac
hes
tim
ate
isfr
oma
sepa
rate
regr
essi
on).
Stan
dard
erro
rs,c
lust
ered
intim
ean
dsp
ace
(see
text
),ar
ere
port
edin
pare
nthe
ses.
Star
sin
dica
test
atis
tica
lsig
nific
ance
:*
p<
0.1,
**p
<0.
05,*
**p
<0.
01.
58
7 Alternative Rainfall Specifications
In table B.1, we report results of estimates that use alternative specifications of rainfall
deviations in the monsoon season. First, we replace binary indicators of rainfall totals
exceeding one localized standard deviation above and below the local long-term mean, with
similar indicators of rainfall totals falling in the 80th and 20th percentiles of the localized
long-term distribution, following Jayachandran (2006)and Kaur (2014). The results (panel
A) are similar, and indicate that both types of shocks increase crime in a statistically
significant manner, but that the impact of negative rainfall shocks is larger. Second, we
use a continuous measure of the fractional deviation of total rainfall from its long-term
local mean, following Duflo and Pande (2007). We find a negative impact, as expected
(panel B), but it is imprecisely estimated, possibly because of the non-monotonic effect of
rainfall. We therefore repeat the estimation, but separately estimate slopes in the positive
and negative shock domains; and find, once again, that both are positive and significant,
but that the negative effect is larger. Third, we conduct a similar estimation, but now using
the localized continuous z-score instead of the fractional deviation. Once again, we find
negative and positive shocks to have significant impacts, with a stronger effect for negative
shocks (panel C). Finally, we test the effect of an alternative measure of precipitation that
replaces the total rainfall amount with the number of rainy days (days having at least 0.1
mm of precipitation). Such measures have been recently shown to perform well as predictors
of agricultural productivity (Fishman, 2016). We divide the range of possible values of this
variable into 7 intervals and include binary indicators of each of those in the regression. We
find that small number of rainy days below 55 (out of a maximum possible 120) lead to
increases in crimes. The lack of impacts of positive rainfall shocks in this specification is
consistent with the nature of the variable. While excessive total rainfall can damage crops,
large numbers of rainy days imply a smooth distribution of rainfall and can therefore have
monotonic positive impacts on crop yields, as shown by Fishman (2016).
59
Table B.1: Alternative Rainfall Specifications(Logarithm) of Pooled Crime Rates
(1) (2) (3) (4) (5)A Bottom Quintile 0.035***
(0.009)Top Quintile 0.014*
(0.008)B Fractional Deviation -0.020
(0.013)Neg Fract Deviation 0.134***
(0.029)Pos Fract Deviation 0.055***
(0.021)C Neg Z-Score Deviation 0.034***
(0.008)Pos Z-Score Deviation 0.016**
(0.006)D 0 To 45 Wet Days 0.057**
(0.028)45 To 55 Wet Days 0.025*
(0.013)55 To 65 Wet Days 0.020
(0.012)65 To 75 Wet Days -0.013
(0.010)85 To 95 Wet Days 0.012
(0.010)95 To 105 Wet Days 0.000
(0.013)>105 Wet Days -0.043
(0.027)
R-Squared 0.132 0.130 0.131 0.131 0.131N 73131 73295 73295 73295 73295
Estimated effects of rainfall shocks on crime in alternative specifications. Panel A: Ex-planatory variables are binary indicators for the top and bottom quintiles of the localizeddistribution of rainfall over time (Jayachandran, 2006; Kaur, 2014). Panel B: Explanatoryvariable is the continuous fractional deviation (Duflo and Pande, 2007)of rainfall in relationto its long-term localized mean (column 2), or the same deviation, estimated separately fornegative and positive deviations (column 3). Panel C: The continuous z-score of rainfallin relation to its long-term localized mean and standard deviation, estimated separatelyfor negative and positive deviations. Panel D: binary indicators for the number of rainydays (precipitation above 0.1 mm) occurring during the monsoon (Fishman, 2016). Allregressions control for temperature shocks as in the main sepcification. Standard errors,clustered in time and space (see text), are reported in parentheses. Stars indicate statisticalsignificance: * p < 0.1, ** p < 0.05, *** p < 0.01.
60