Is there an economic dimension to suicide?

Evidence from climate and agriculture in India

Tamma Carleton & Ceren Baysan∗

Preliminary draft: Do not cite or circulate

December 2, 2015

Abstract

The multifaceted determinants of suicide have long been studied by psychologists and sociolo-gists, but only a small and more recent literature in economics has contributed to understanding thisphenomenon. We utilize a comprehensive set of suicide data from India, one of the countries mostpublicly struggling with rising suicide rates today, to look for evidence of economically motivatedsuicide. Using state-level panel data on suicide rates between the years 1967 and 2013 and exogenousvariation in climate, we (a) identify the causal effect of variations in temperature and precipitationon suicide rates across the country; and (b) find empirical support for an economic channel throughwhich these shocks affect suicide. We employ a fixed effects framework to show that temperaturehas a positive and significant effect on suicide, but only during India’s main agricultural growingseason, when high temperatures also lower crop yields. For growing season days above 10◦C, a 1◦Cincrease in daily temperature causes an additional 370 suicides. The magnitude and robustnessof this result, in tandem with additional mechanism tests, suggests the presence of an importanteconomic component to suicide. Contrary to public discussions of drought and suicide in India,growing season precipitation has a minimal impact on suicide rates. We find evidence that bothtemperature and precipitation effects operate, at least in part, through economic channels. Ourresults provide some justification for the development of suicide prevention policies that amelioratethe negative economic shocks suffered by agriculture-dependent households in India, such as debtrelief. We also contribute to a growing literature on the social and economic impacts of climatechange. Suicide prevalence, an indicator of severe hardship, may capture facets of welfare previouslyunmeasured in the climate change literature, and thus suggests a new avenue through which we canassess the climate-welfare relationship.

∗Carleton: Agricultural and Resource Economics, UC Berkeley. Email: tcarleton@berkeley.edu. Baysan: Agricul-tural and Resource Economics, UC Berkeley. Email: cbaysan@berkeley.edu.

1 Introduction

That suicide may often be consistent with interest and with our duty to ourselves, no onecan question, who allows that age, sickness, or misfortune, may render life a burden, andmake it worse even than annihilation. I believe that no man ever threw away life, while itwas worth keeping. - Hume

Humans have long debated the rationality and morality of suicide. Stoic philosophy regarded

suicide as noble when chosen through a deliberative process, placing rational decision-making at the

heart of the act (Seidler, 1983). Monotheistic religions, contrarily, generally view suicide as morally

reprehensible, and most current legal systems render it a crime. Since the seminal work of Durkheim

(1951), modern psychologists and sociologists have balanced the multiplicity of causes behind the

choice to take one’s own life, emphasizing to varying degrees moral, cultural and often economic

factors. However, only recently, and rarely, has the economic rationality of suicide been modeled and

empirically tested by economists (Becker and Posner, 2004; Hamermesh and Soss, 1974). In this paper,

we utilize a comprehensive set of suicide data from India, one of the countries most publicly struggling

with rising suicide rates today, and ask: how economic is suicide?

India’s suicide death rate has nearly doubled since 1980. One fifth of global suicides occur in India:

at least 135,000 lives were lost to suicide in 2010. Suicide amongst the rural population has gained

particular academic and media attention; recent data from India’s National Crime Records Bureau

(NCRB) shows that in some states the share of suicides committed by farmers currently exceeds 75%,

and has risen over the last decade. Blame is often placed on a range of agrarian, and clearly economic,

changes. Nearly all cited causes of recent suicide trends are agricultural in nature, and range from

trade liberalization of the economy in the 1990s to farmer debt induced by high input costs of Green

Revolution technologies like fertilizer and high-yield variety seeds (Mohanty, 2005).1 A common causal

mechanism in these arguments is that shocks to agricultural profitability, especially drought events and

crop failure, induce suicide as a coping response to the economic stress placed on farming households

(Deshpande, 2002; Gruere and Sengupta, 2011). While India has slowly been urbanizing over time,

economic welfare continues to be closely tied to agricultural outcomes. The fact that a majority of

Indian suicides are committed by ingestion of pesticides (WHO, 2014), which some have argued is a

symbolic act (Patel, 2007), further suggests that these deaths are occurring as responses to agricultural

1Recent media coverage of Indian suicides include a 2014 article in The Guardian titled “India’s farmer suicides: aredeaths linked to GM cotton?”, 2013 coverage by BBC News asking “Indian farmers and suicide: How big is the problem?”,a 2006 New York Times article titled “On India’s Farms, a Plague of Suicide”, and hundreds of pages of coverage in anarray of Indian media outlets.

1

hardship.

However, it is not empirically clear that economic strife is behind many Indian suicides. In 2010,

data from the NCRB show that only 4% of total suicides were attributed to bankruptcy or poverty,

while 24% were due to family problems and 19% to physical or mental illness. Agriculture’s role

in particular is debatable: while some states have exceptionally high farmer suicide rates, the mean

share of male suicides committed by farmers over the last decade is only 15%, and for females it is

only 4%. While these numbers are likely poorly measured and perhaps biased, they cast doubt on

the prominence of economic motivations. Moreover, existing evidence of an agricultural mechanism

is anecdotal (Mohanty, 2005; Patel, 2007), correlational (Gruere and Sengupta, 2011), or covers only

a small region of the country (Hebous and Klonner, 2014). In this paper, we use national state-level

panel data on suicide rates between the years 1967 and 2013 and exogenous variation in climate to first

identify a causal effect of temperature and precipitation on suicide rates, and then to seek empirical

support for an economic channel through which these shocks drive suicide.

We find that temperature has a positive and significant effect on suicide, but only during India’s

main agricultural growing season. The magnitude of this result, robust to a variety of specifications,

suggests the presence of an important economic component to suicide. To further explore the relevance

of agriculture to suicide prevalence, we use yield data to provide suggestive “fingerprinting” evidence

(Hsiang, Burke, and Miguel, 2013) of the shared pattern between suicides and yield losses, as they

relate to temperature. We find that yields and suicides have matching responses: hot days during the

growing season both damage yields (as has been shown by Guiteras (2009) and Burgess et al. (2014))

and increase suicide rates, while non-growing season heat has a minimal effect on both outcomes.

In a final test of the plausibility that economic motives drive our reduced form results, we use cross-

sectional patterns in geographic heterogeneity of effects. We show that states with more severe histories

of farmer suicide have stronger responses to temperature, and that states with higher crop losses due

to temperature also exhibit more suicides under heat stress. Contrary to common wisdom, growing

season precipitation has a very minimal impact on suicide rates, despite generating yield gains, while

drought appears insignificant. We explore possible reasons for this result, showing that we may be

underpowered to identify an agriculturally-driven impact of rainfall on suicides, given the small effect

rain has on yields relative to temperature, a finding consistent with existing literature (Burgess et al.,

2014; Guiteras, 2009).

We conclude that over our sample time period in India, there appears to be a strong economic

2

component to suicide, exhibited by significant impacts of exogenous climate shocks, particularly hot

temperatures, operating through agricultural channels. Our results contribute to two bodies of ex-

isting research, with distinct policy implications. First, we add to a the literature on the economic

determinants of suicide, supporting existing studies that also find evidence of economic motives. Our

findings provide some justification for the development of suicide prevention policies that ameliorate

the negative economic shocks suffered by agriculture-dependent households in India, such as debt re-

lief. Second, we contribute to a growing literature on the welfare impacts of climate change. Suicide

prevalence, an indicator of severe hardship, may capture facets of welfare previously unmeasured in the

climate change literature, and thus suggests a new avenue through which we can assess the climate-

welfare relationship. Our suicide findings can contribute to an accurate assessment of the impacts of

climate variability on human welfare, a critical step in the design of policy responses to climate change.

In this paper, we use the term climate in reference to observations of temperature and precipitation.

While climate is generally defined as the long-run distribution of climatic variables and not to short-run

observations, this distribution is inherently composed of short-run realizations. Societies experience

and respond to these short-lived outcomes, implying the frequency of these events is an important

economic facet of the climate (Burke, Hsiang, and Miguel, 2014). This is a view particularly relevant to

anthropogenic climate change: as the distribution of climatic variables shifts, the frequency with which

a population experiences a given climate observation (e.g. a hot day) also changes. Understanding the

impacts of these short-run events is therefore critical to understanding how societies are impacted by

climate change.

The paper proceeds as follows: we briefly review relevant literature in Section 2, outline a conceptual

framework in Section 3, and describe our data in Section 4. Section 5 details our empirical strategy,

Section 6 shows our results, and we conclude in Section 7.

2 Previous Literature

The causes of suicide are strongly debated across the social sciences, and the extent to which economic

welfare plays a role has remained unclear. In a review of the sociological literature, Stack (2000) de-

scribes inconsistent findings on the relationship between unemployment and suicide, emphasizing that

studies rarely control for many unemployment covariates that also affect suicide. Similarly, Rehkopf

and Buka (2006) review “widely divergent” findings on the impacts of economic variables on suicide

3

in the psychological literature, again focusing on methodological challenges. To our knowledge, only

a small number of economists have sought to use theoretical and empirical tools to rigorously address

this question.

Hamermesh and Soss (1974) were the first to directly model and empirically explore suicide in eco-

nomics, using time series and cross-sectional variation in the U.S. to argue that suicide rates increase

with aggregate measures of unemployment. Becker and Posner (2004) further developed an economic

theory of suicide, and a small body of recent empirical work has sought to better identify the role of

economics. However, despite the fact that 75% of suicides occur in low-income countries (WHO, 2014),

this literature is almost exclusively focused on OECD nations (e.g. Andres (2005); Inagaki (2010)).

Studies generate conflicting results (Andres, 2005; Botha, 2012), and tend to employ time series and

cross sectional data instead of panel (Lloyd and Yip, 2001; Snipes, Cunha, and Hemley, 2012). Panel

studies that do exist identify macroeconomic impacts, such as business cycles, rather than microeco-

nomic shocks (Botha, 2012; Pierard and Grootendorst, 2014; Snipes, Cunha, and Hemley, 2012). We

contribute to this literature by utilizing panel data in India, focusing explicitly on the importance of

and challenges involved in using exogenous variation to isolate the impact of microeconomic variation

on suicide.

India’s suicide trends have garnered widespread media and political attention, complemented by a

large body of academic work. While much has been written on the influence of economics and agricul-

ture on Indian suicide, nearly all papers use qualitative methods to detail particular circumstances of

rural suicides in isolated areas of India (Deshpande, 2002; Herring, 2008; Mohanty and Shroff, 2004;

Mohanty, 2005; Rao and Gopalappa, 2004; Sarma, 2004). While these studies are valuable, they have

low external validity, and without quantitative results are difficult to integrate into policy. In contrast,

Hebous and Klonner (2014) use panel data for two states over 6 years and an instrumental variables

framework to estimate the impact of poverty on suicide. While they argue that poverty increases the

male suicide rate, their results are significant for only one of two states, and the climate variables

used as instruments are unlikely to satisfy the exclusion restriction (Hsiang, Burke, and Miguel, 2013).

The only national-level studies we have identified are descriptive in nature (Nagaraj, 2008) or make

inferences based on cross-sectional correlations (Jonathan Kennedy, 2014). In this paper, we use data

covering all of India over nearly 50 years and employ a fixed effects empirical strategy to identify causal

effects of exogenous shocks to income and suicide. While we do not claim to directly identify economic

suicide motives, we use a variety of techniques to assess the plausibility that our reduced form results

4

operate through an agricultural, and hence economic, channel.

Our work also contributes to the literature on the welfare impacts of climate change. The startling

pervasiveness of suicide makes it a singularly important component of the well-known impact of climate

on mortality (Barreca et al., 2013; Burgess et al., 2014; Deschenes and Greenstone, 2011): each year,

suicide claims over 800,000 lives globally and is the second most common cause of death for people

aged 15 to 29. The majority of these deaths occur in low- and middle-income countries, many of

which, like India, have seen suicide numbers rise in recent years (WHO, 2014). Although a growing

literature on climate and violent conflict exists, reviewed in Burke, Hsiang, and Miguel (2014), Basyan

et al. (2014) is the only study in economics that directly estimates the reduced form impact of climate

variables on suicide rates.2 In contrast to our findings, the authors conclude that in Mexico, suicide

does not appear to have an economic component. Our results therefore provide a second estimate

in the economics literature that can help inform climate change policy, while highlighting geographic

differences in economic drivers of suicide.

3 Conceptual Framing

Suicide is a multifaceted phenomenon determined by the interaction of many individual and environ-

mental factors including mental health, social norms, beliefs, religion and, arguably, economic status.

In this section, we outline the possible ways in which economic shocks may affect suicide rates in the

Indian context, explicitly tracing the causal claims suggested in qualitative research and popular me-

dia. We then demonstrate how our empirical approach will allow us to use data on climate variability

and agricultural output to identify evidence of economic drivers of suicide, should they exist.

Economic outcomes throughout India remain closely tied to agriculture. In 2001, 57% of workers

were identified as employed in agriculture and allied activities, and this share fell to only 51% in the

latest national census of 2011. 70% of India’s population currently lives in rural areas (Government

of India, 2011). While modern technologies such as high-yielding variety (HYV) seeds and various

irrigation methods have been increasingly adopted through time, the profitability of agriculture in

India, as in most nations, remains subject to unpredictable variations (Foster and Rosenzweig, 1996;

Rosenzweig and Udry, 2013). Importantly, climate variability critically affects agricultural incomes,

and with over 30% of the population lying below the international poverty line of US$1.25 per day,

2The impact of climate on suicide has been identified in psychiatry (Deisenhammer, Kemmler, and Parson, 2003) andmeteorology (Jessen, Jensen, and Steffensen, 1998).

5

climate shocks can have dire consequences for households relying on agriculture. The probability that

any individual will contemplate, attempt or commit suicide depends on many, often unobservable,

factors. However, negative income shocks like those induced by agricultural income variability in India

may increase the likelihood for some individuals by increasing stress (Chemin, De Laat, and Haushofer,

2013), or by exacerbating existing risk factors such as family challenges or mental and physical health

(Rehkopf and Buka, 2006; Stack, 2000).

To seek evidence of economic motivations of suicide, we exploit random variation in a central

determinant of economic outcomes in India: climate variability. We identify the impact of economically

meaningful climate shocks - that is, climate outcomes that we demonstrate have substantial causal

impacts on agricultural yield values - on suicide rates, looking for evidence that the same types of

climate variations that reduce agricultural income are those that elevate suicide rates. We refrain from

an instrumental variables (IV) approach in which climate variables instrument for crop yields, as the

exclusion restriction is unlikely to hold due to the direct psychological impacts that climate factors

plausibly have on suicide rates (Anderson et al., 1999; Davidson, Putnam, and Larson, 2000; Jessen,

Jensen, and Steffensen, 1998; Lovheim, 2012; Seo, Patrick, and Kennealy, 2008). While many factors

unmeasured in our analysis undoubtedly affect suicides in India, the exogeneity of climate realizations

in any given location allow us to identify a causal impact of climate on suicide, and the distinction

between economically meaningful and economically unimportant climate outcomes allows us to uncover

an agricultural channel through which these climate variations drive suicides.

4 Data

We combine suicide and climate data at the state level for the years 1967-2013. To compare suicide

and yield responses to climate, we use an agricultural yield dataset at the district level for the period

1956-2000, combined with district-level climate data for the same years. Summary statistics for key

variables of interest are provided in Table 1.

4.1 Suicides

Suicide records in India are publicly available only at the state level. These data are in the “Accidental

Deaths and Suicides in India” report, published annually by the Indian National Crime Records Bureau

(NCRB) since 1967 for 27 of India’s 29 states and 5 of its 7 Union Territories. The dataset includes

6

total number of state suicides per year, with gender, occupation and cause of death information avail-

able after 2001. We calculate suicide rates as the number of total suicides per 100,000 people, with

population data linearly interpolated between Indian censuses. We do not exploit data on occupation

or gender, as this information is available only after 2000. Thus, we are measuring the overall suicide

rate, which will include farmers, non-farmers, and the unemployed.

Deaths in general are under-reported in India (Burgess et al., 2014), and the suicide data provided

by the NCRB are particularly problematic in this regard. The data are aggregated from district police

reports. Since attempted suicide is a criminal offense punishable under the Indian Penal Code, there

is likely to be significant under-reporting of suicide as a cause of death, as surviving family members

may experience social stigma from the criminality of the event. As evidence of this, the NCRB reports

135,000 suicides in India in 2010, while data from a randomly sampled survey of cause of death

estimates the 2010 value at 187,000 (Patel et al., 2012). This under-reporting is likely uncorrelated

with temperature and precipitation (otherwise, we would have to argue that hotter or drier conditions

induce better reporting, which seems implausible). Thus, our estimates of the response of suicide to

climate provide lower bounds on the true marginal effect, due to attenuation bias.

The evolution over time and space of state level suicide rates in India during our sample period is

shown in Figure 1; darker shades indicate higher suicide rates.3 There is clear spatial heterogeneity,

with southern India experiencing the highest suicide rates and largest increases over time. These

geographic differences inform our empirical strategy, which relies on within-state variation in order to

avoid conflation of climate impacts with unobservables, such as cultural norms, political structures,

and religious influences. Moreover, we account for regional differences in time trends, due to clearly

distinct patterns over time across India.

4.2 Climate

Both Deschenes and Greenstone (2011) and Burgess et al. (2014) emphasize that estimating the rela-

tionship between mortality and climate requires daily temperature data, as the relationship is nonlinear

and annual average temperatures obscure such nonlinearities. While existing studies on temperature

and suicide in the epidemiology literature do not explore nonlinearities, there are two reasons why they

are likely to occur.4 First, the growing literature on climate and interpersonal conflict reviewed by

3As a point of reference, suicide rates in the United States are currently approximately 12.5 per 100,000.4The only existing paper in the economics literature identifying climate’s impact on suicide is Basyan et al. (2014).

The authors do not allow for nonlinear temperature effects.

7

Burke, Hsiang, and Miguel (2014) often identifies nonlinearities in the effect of temperature on violent

crime. If we view suicide as a type of violence against oneself, it’s likely that a similar relationship

exists here. Second, Schlenker and Roberts (2009), among others, have identified a strongly nonlin-

ear response of staple crop yields to temperature. If suicide in India is indeed related to agricultural

productivity, then capturing this nonlinearity is critical.

Daily weather data are not publicly available for India with adequate geographic and temporal

coverage. Thus, we rely on the temperature data used by Guiteras (2009) and Burgess et al. (2014):

the National Center for Environmental Protection (NCEP) gridded daily dataset, a reanalysis product

that provides observations in a 1◦×1◦ grid. These data are produced by NCEP and the Climatic

Research Unit at the University of East Anglia and include daily mean temperature for each grid

over our entire sample period. Because there are multiple grid cells per state, we aggregate grids to

state-level daily temperature observations using an area-weighted average.

To convert daily temperature into annual observations without losing intra-annual variability in

daily weather, we use the agronomic concept of degree days. Degree days are calculated as follows,

where t∗ is a selected cutoff temperature value and t is a realized daily temperature value:

Dt∗(t) =

{0 if t ≤ t∗

t− t∗ if t > t∗(1)

Degree days allow temperature to affect an outcome variable only once its value surpasses the threshold

t∗. When degree days are summed over a time, regressing an outcome on cumulative degree daysT∑t=1

D(t) imposes a piecewise linear relationship, in which the outcome response has zero slope for all

temperatures less than t∗. While a body of literature identifies biologically-determined cutoffs t∗ for a

variety of major crops, there is no empirical support to draw on in selecting t∗ for suicides. Thus, we

use a range of plausible cutoffs based on the distribution of our temperature data to estimate a flexible

piecewise linear function that imposes minimal structure on the response function.

Because reanalysis models are less reliable for precipitation data, and because nonlinearities in

precipitation that can’t be captured with a polynomial appear to be less consistently important both

in the violent crime literature (Burke, Hsiang, and Miguel, 2014) and in the agriculture literature

(Schlenker and Roberts, 2009), we use the University of Delaware monthly cumulative precipitation

data to compliment daily temperature observations (Willmott and Matsuura, 2014). These data are

gridded at a 0.5◦×0.5◦ resolution, with observations of total monthly rainfall spatially interpolated

8

between weather stations. We again aggregate grids up to states using area-based weights.5

4.3 Agriculture

We use agricultural data from Duflo and Pande (2007), who provide district-level annual yield estimates

for major crops between 1956 and 2000, compiled from Indian Ministry of Agriculture reports and

other official sources. These data cover 271 districts in 13 major agricultural states, although they

omit Kerala and Assam, two large agricultural producers with high rates of suicide. We match these

district data to the same climate data described above, area-weighting over districts instead of states.

We use this merged climate-yield dataset to generate a response function of log annual yield values

of five major crops (in constant Indian Rupees) to temperature and rainfall, allowing us to identify

similarities and differences in how suicides and yields are impacted by climatic shocks. These data are

limited in their ability to identify heterogeneity of suicide impacts, as they cover only 13 states, do

not extend past 2000 and do not contain recent irrigation data. However, they allow for more precise

estimation of impacts, given their high spatial resolution.6

5 Empirical Strategy

5.1 Reduced form

To identify the impact of temperature and precipitation on annual suicide rates, we estimate two

versions of a panel fixed effects model in which the identifying assumption is the exogeneity of within-

state annual variation in degree days and cumulative precipitation. Heterogeneity in suicide rates and

trajectories across states, due to an interplay between unobservable cultural, political and economic

factors, implies that cross-sectional variation in climate is endogenous. Thus, we use state and year

fixed effects with regional time trends to control for time-invariant state-level unobservables, national-

level temporal shocks and region-specific time trends.

There is minimal precedent for the functional form of suicide’s relationship to climate. Our first

estimation approach therefore employs the most flexible model our data will allow, and one that is

consistent with the broader climate impacts literature. We use a piecewise linear response function

51% of the observations in all precipitation variables were Winsorized (Hastings Jr et al., 1947) on the right tail only,due to the presence of a small number of outliers. Results are robust to different levels of Winsorizing, and to leaving thedata complete.

6It is important to keep in mind that while they have significant overlap, our agriculture and suicide data cover slightlydifferent years and have distinct spatial coverage.

9

with respect to temperature, with kinks at 10◦C intervals, and a cubic polynomial function of cumu-

lative precipitation.7 To capture the distinct impact of economically meaningful climate variation, we

separately identify the temperature and precipitation response functions by agricultural seasons. Crop

seasonality in India is determined by the onset and withdrawal of the southwest monsoon, which signals

the beginning and close of the main growing season, called the kharif. Monsoon rainfall arrives on the

southern coast of India around June 1st and moves northward through the summer, withdrawing in a

reverse geographic pattern. For simplicity and to be consistent with prior work on Indian agriculture

(Auffhammer, Ramanathan, and Vincent, 2006; Duflo and Pande, 2007; Guiteras, 2009), we define the

growing season as June 1st through September 30th for all states. However, results are robust to using

state-specific growing season dates derived from long-run average monsoon patterns reported by the

Indian Meteorological Department.8

Let suicide rateit be the number of suicides per 100,000 people in state i in year t, s ∈ {1, 2} be

the season (growing and non-growing). Then Tidst is state i’s temperature in ◦C on day d in season s

in year t, and Pimt is cumulative precipitation during month m in season s in year t. We use k = 4

bins of temperature, each of width 10◦C, with the final bin containing all degree days above 30◦C. Our

main empirical model is thus:

suicide rateit = α+2∑

s=1

4∑k=1

βks∑d∈s

Tidst +2∑

s=1

λ1s∑m∈s

Pimt +2∑

s=1

λ2s∑m∈s

P 2imt (2)

+2∑

s=1

λ3s∑m∈s

P 3imt + δi + ηt + τrt+ εit

Equation 2 generates a piecewise linear response function for temperature, with four distinct temper-

ature slope parameters in each season. State fixed effects δi account for time-invariant unobservables

at the state level, while year fixed effects ηt account for India-wide time-varying unobservables. In

most specifications, we include region-by-year time trends τrt to control for differential regional trends

in suicide driven by time-varying unobservables. For these suicide regressions, we define our regions

using a classification system derived by Nagaraj (2008), which groups states based on the prevalence

of suicide as well as trends in total suicides and farmer suicides through time. For yield regressions, we

follow Guiteras (2009) in dividing states into four geographically-determined regions. Results are also

robust to alternative regional definitions, and in particular are upheld when we apply the same regions

7Results are very similar, yet less precise, when we use temperature bins of width 5◦C. These results are availableupon request.

8These results are available upon request.

10

to both the yield and suicide data. These regional trends absorb much of the variation in suicide

rates, so we show results for models with and without the τrt term. Our identifying assumption is

that, conditional on δi, ηt, and τrt, variations in daily temperature and monthly rainfall are as good

as random.

Separately for each season, Equation 2 allows us to identify each βk slope parameter. The marginal

effects are interpreted as follows: βk estimates the change in the annual number of suicides per 100,000

people induced by one day in bin k becoming 1◦C warmer, estimated separately for each season s. This

annual response to a daily forcing variable is similar to that estimated and described in Deryugina and

Hsiang (2014). The λ parameters estimate a polynomial response function of the annual suicide rate to

an additional millimeter of rainfall, again estimated seasonally. Due to likely correlation between errors

within states, we cluster standard errors at the state level. This strategy assumes spatial correlation

across states in any time period is zero, but flexibly accounts for within-state, across-time correlation.

A secondary specification that we employ to improve readability of tables, reduce the statistical

requirements placed on the data, and run heterogeneity tests, is a simple cumulative degree days model.

We estimate the following equation, where DDX,idt is degree days above a threshold of X◦C. We test

robustness of results for degree day thresholds of X = 10, 15, 20 and 25◦C.

suicide rateit = α+2∑

s=1

βs∑d∈s

DDX,idt +2∑

s=1

λ1s∑m∈s

Pimt +2∑

s=1

λ2s∑m∈s

P 2imt (3)

+2∑

s=1

λ3s∑m∈s

P 3imt + δi + ηt + τrt+ εit

This model is also piecewise linear in temperature, but assumes a zero marginal response of suicide

to temperature for daily values below X◦C, and a constant linear response with a slope of βs for

temperatures above X in season s. βs estimates the change in the annual suicide rate caused by a one

degree increase in daily temperature, conditional on temperature being above X◦C.

Note that both of these specifications do not include lagged climate variables, consistent with

literature on climate and mortality in India (Burgess et al., 2014). However, if climate shocks lower

yields during the growing season in year t, suicide in year t+ 1 could plausibly be affected. To account

for this possibility, we run robustness checks including up to two years of lags in temperature and

precipitation variables. Figure 2 in the Appendix shows that lagged coefficients are generally not

significant, while the impacts of contemporaneous climate remain. We also check for the presence of

cumulative climate shocks (e.g. two consecutive years of low rainfall) in Section 6, and find no evidence

11

of their influence. We therefore conduct our main analyses without lagged variables. However, we

acknowledge that lagged effects may be important, but simply unidentifiable in our data.

5.2 The role of agriculture

With ideal data, we would estimate separate response functions for farmers and non-farmers to isolate

the importance of an agricultural channel. Because our data do not provide the occupation of suicide

victims prior to 2001 (and because using only post-2001 data at the state level leaves us with an excep-

tionally small sample size in which no climate effects are statistically significant), we utilize a variety

of other methods to investigate the validity of the oft-cited agricultural mechanism. The primary ap-

proach we take is to compare the significance and magnitude of each coefficient βk in Equation 2 across

seasons. Temperatures and rainfall in June through September have been shown to be most critical for

agricultural productivity (Burgess et al., 2014; Guiteras, 2009), and thus should dominate the climate-

suicide relationship if the agricultural channel is important. In a similar exercise, Fetzer (2014) and

Blakeslee and Fishman (2013) demonstrate that monsoon-season precipitation impacts civil conflict

and interpersonal crime in India, respectively, more than precipitation outside the growing season.

Just as they use these findings as evidence of an agricultural channel through which climate affects

crime and conflict, we use our results to identify the presence of an agricultural channel for suicide.

An additional method for examining mechanisms is to “pattern match” response functions (Burke,

Hsiang, and Miguel, 2014). For example, Hidalgo et al. (2010) show that the nonlinear relationship

between agricultural income and rainfall in Brazil is nearly a perfect (inverse) pattern match to the

relationship between land-invasion risk and rainfall. Similarly, Hsiang and Meng (2014) match the

responses of conflict and income to the timing of the El Nino Southern Oscillation (ENSO), arguing

results provide support for an income channel. We follow this approach by estimating Equation 2

using the log value of yield as the dependent variable in place of suicide rates, employing the Duflo &

Pande agricultural data at the district level. This regression essentially replicates results in Guiteras

(2009), but employs more years of data and follows a slightly different degree days methodology.9 We

compare our response functions of suicide and yield to identify matching patterns.

Finally, we look for further support of economic motives by exploring spatial and temporal het-

erogeneity of impacts. For temperature shocks, we estimate a model that allows each of India’s 32

states and Union Territories to have a distinct suicide rate response function. We then look at correla-

9Guiteras (2009) uses both binned temperature using 1◦C bins and a degree days approach with cutoffs of 8◦C and32◦C.

12

tions between these state-level temperature responses and other variables related to agriculture, such

as the ratio of farmer suicides to total suicides (for the limited years in which these data exist) and

the vulnerability of state crop yields to temperature. For precipitation, we decompose the seasonal

response function into monthly effects, highlighting the importance of specific months of rainfall on

suicide rates. Together, season-specific regressions, pattern matching and heterogenous effects provide

multiple sources of evidence that the observed impact of climate variability on suicide rates likely

operates, at least in part, via economic channels.

6 Results

6.1 Temperature

Our main results for the impact of climate on suicide rates, estimated from Equation 2, are shown

in Table 2. We begin by considering temperature. Temperature response functions for suicide and

yield are plotted in Figure 3, with tabular results for yields shown in Table 3. Both functions have

been centered at their respective means, showing the predicted annual response to a daily increase in

temperature relative to mean suicides and the mean log value of yield.

Figure 3a demonstrates a clear distinction between the growing season and non-growing season

response of suicide rates to temperature, with temperature in the former significantly and positively

impacting suicide rates. The marginal effects in both the 10-20◦C and 20-30◦C bins are positive

and statistically significant: for the fixed effects model, a one degree increase in daily temperature,

conditional on that day being in the 10-20◦C bin, increases suicides by 0.03 per 100,000 people. This is

a 9.4% increase in the suicide rate per standard deviation (σ) increase in daily within-bin temperature,

which increases to 10.1% when regional time trends are included. Using 2010 population values, these

marginal effects imply that a 1◦C increase in a day’s temperature within the 10-20◦C bin causes an

additional 370 suicides. The corresponding results for the 20-30◦C bin are a 5.5%/σ increase and

170 more suicides per 1◦C. In contrast, warm temperatures in the non-growing season have a slightly

negative to zero impact on suicide rates, with only the hottest days having a significant positive

marginal effect on suicides (the effect for the hottest bin in the non-growing season is 3.5%/σ). These

findings are robust to our simpler degree day empirical model, as shown in Table 4. They are also

consistent with related literature on suicide, as well as other types of violent conflict. In the only other

empirical estimate of the reduced form impact of temperature on suicide rates, Basyan et al. (2014)

13

find a standardized effect of 7.4%/σ using average monthly temperatures and suicide rates in Mexico.

Importantly, while our temperature result depends critically on agricultural seasonality, theirs does

not; nor do other tests reveal evidence of economically-motivated suicide in Mexico. In a hierarchical

Bayesian meta-analysis of the literature on climate and interpersonal conflict, Burke, Hsiang, and

Miguel (2014) find an average effect of 2.4%/σ across 18 papers, each of which estimates the impact

of temperature on an outcome such as rape, assault, or murder. To the extent that suicide can be

seen as a type of violence related to interpersonal conflict, our estimate corresponds to a large impact,

relative to the mean effect across this literature.

The distinct response of suicide to temperature in the growing and non-growing seasons is consistent

with an economic story of suicide in which heat damages crops, placing pressure on farming households,

in turn increasing suicide rates via heightened levels of poverty. The response of yields to temperature,

shown in Figure 3b, provides further evidence of such a channel. In a mirror image of the suicide

response function, yields fall as growing season temperature rises. For daily temperatures in the 10-

20◦C bin, log annual yields fall by 1%/σ when regional time trends are included. The pattern match

holds in the off-season as well, where both annual yields and suicides respond minimally, if at all, to

temperature. Thus, precisely the same temperature shocks that cause agricultural losses are those that

significantly increase suicide rates.

One concern with using pattern matching is that temperature may impact suicide during the

growing season months only, but for reasons unrelated to agriculture. In particular, there is strong

evidence that suicide is directly impacted by heat through a psychological mechanism (Basyan et al.,

2014; Deisenhammer, Kemmler, and Parson, 2003). If there were few hot days outside the growing

season, our model may have insufficient statistical power to identify a true off-season impact, and

growing season temperature would simply be capturing the psychological effect. However, this is not

empirically a concern. Many hot days do occur outside the growing season (see Figure 4), and Table 2

shows that we do identify a significant off-season effect of days above 30◦C, suggesting that temperature

may impact suicide both through direct (psychological) and indirect (economic) channels.

6.1.1 Pattern matching in the cross-section

Reduced form results suggest the presence of an agricultural channel for the impacts of temperature

on suicide. To further support this argument, we use geographic heterogeneity in impacts to show

correlations between suicide responses and other key agricultural variables. We first disaggregate sui-

14

cide response functions by states: Figure 5 shows states colored according to their individual marginal

effect, estimated from a version of Equation 3 in which growing season degree days are interacted with

an indicator for each state. State coefficients are expressed as a fraction of the average treatment effect.

Due to our sample size, these are noisy estimates. However, there is a clear geographic pattern that is

implausibly random. Southern states (which are generally hotter, have higher suicide rates, and show

and steeper trends in suicide over time) have much stronger responses to climate.

Figure 6 provides two pieces of evidence that these differential state impacts are driven, at least

in part, by agriculture.10 First, Panel 6a shows the correlation between suicide’s response to growing

season temperature and the average farmer suicide ratio, defined as the number of farmer suicides

divided by the total number of suicides across all occupations, with the mean taken over the years

2001-2010 (the years for which data on occupations are available). Note that each point represents one

of India’s 32 states. The positive correlation evident in this figure shows that states where farmers make

up a larger share of suicides suffer higher marginal temperature effects, suggesting that agriculture is a

key component of suicide’s response to temperature. Second, Panel 6b shows the correlation between

each state’s yield response to growing season temperature and its suicide response to growing season

temperature. Unfortunately, we can only estimate 13 distinct state-level yield responses given the

coverage of our agricultural data. Nonetheless, the correlation between these 13 marginal effects of

yield and suicide in Figure 6b is positive, suggesting that states with agricultural yields that are

relatively more damaged by high temperatures are also states in which these temperatures increase

suicide rates by a relatively substantial magnitude. Three states that have been at the center of

India’s popular debates regarding agricultural influences on suicide - Maharashtra, Karnataka and

Andhra Pradesh - are shown to not only have high suicide responses to temperature, but also to have

high farmer suicide ratios and large negative impacts of temperature on yield.

We conclude that seasonality in response functions, pattern matching with yields, and cross-

sectional evidence on the heterogeneity of impacts all suggest that temperature has a positive impact

on suicide rates that is at least partially realized through agricultural losses.

10The state-level differences also may be due to local suicide prevalence and/or long-run climate. There is ampleevidence of the social contagiousness of suicide (Cutler, Glaeser, and Norberg, 2001; Gould, Jamieson, and Romer, 2003;Velting and Gould, 1997), and Figure 7b shows that average rates of suicide are correlated with temperature responses,consistent with this theory. Figure 7a also shows that hotter states have steeper responses, suggesting an increasedvulnerability for states with warmer average climates.

15

6.2 Precipitation

The precipitation response functions are shown in Figures 8a and 8b, both estimated as a cubic polyno-

mial as in Equation 2 and as a simple linear function of cumulative rainfall for ease of interpretation. As

expected, precipitation during the growing season has a positive impact on yields, while non-growing

season rainfall has a statistically significant but economically minimal effect. Consistent with other

literature on agricultural impacts of climate variability (Guiteras, 2009; Schlenker and Roberts, 2009),

these precipitation results are smaller in magnitude than those for temperature. In the linear model,

yields increase 0.9%/σ for an increase in growing season rainfall, and just 0.3%/σ for off-season rain-

fall. Were a simple agricultural channel to be identified for precipitation, we would expect a negative

impact of growing season rainfall on suicide rates, and a minimal or zero impact in the non-growing

season. Figure 8 shows a small negative growing season suicide response, despite the clear yield gains

of monsoon rains. While non-growing season rainfall has a positive impact on suicide, this effect is

statistically insignificant.

The null result for growing season rainfall is robust to many alternative specifications. First,

Figure 2 shows that lagged rainfall has no significant effect on suicide rates, suggesting that the lack

of a response is not due to a delay in the effect. Second, we create a drought indicator equal to one

when annual rainfall in a state is below the 20th percentile for that state over the sample. Columns (1)

through (3) in Table 5 show results for a regression including this drought variable, as well as a surplus

rainfall variable indicating a year of exceptionally high rainfall. None of the coefficients on drought

are significant. In results not shown, we create indicators for a state suffering two or three consecutive

years of drought, but no cumulative effects are significant. Finally, we test whether irrigation is causing

this null result. We use Ministry of Agriculture data to classify states as heavily irrigated if their share

of crop area under irrigation ever exceeds 50% during our sample period. Figure 9 shows precipitation

response functions for irrigated and rain-fed states separately, and demonstrates that accounting for

irrigation does not change the findings of our main model.

There are two key reasons we may fail to identify a growing season response in suicides. First,

characterizing monsoon rainfall at the state level is inherently difficult, as there are important within-

state differences in monsoon arrival and withdrawal (Burgess et al., 2014). For example, some districts

in states along the eastern coast of India receive a substantial portion of their rainfall in a late north-

eastern monsoon period between October and December, after the close of the conventional growing

season. Our need to define growing seasons at the state level due to the nature of the suicide data

16

may create important attenuation bias through measurement error. Figure 10 suggests measurement

error may be at play: this plot of monthly rainfall effects suggests a consistently negative, yet often

insignificant, impact of rainfall on suicide rates during the main growing season months. Second, many

suicide victims whose incomes are tied to agriculture may also have crop insurance, the payouts of

which are nearly always tied to rainfall in India (Gine, Townsend, and Vickery, 2007). Even though

low precipitation years clearly damage yields, agriculturally-dependent households may receive suffi-

cient insurance funds to mitigate negative income impacts, ameliorating a potential suicide response.

While we do not formally address the impact of insurance on suicide in this paper, we see this as a

fruitful area for subsequent work. In sum, our results appear to contradict the dominant focus on

drought in most media coverage of suicide in India.

7 Discussion

In this paper we address the long-debated question of the role of economics in suicide, using exogenous

variation in climate to identify shocks to income and suicide in a developing country context. While we

cannot precisely identify the effect of income on suicide rates, we estimate the causal effect of exogenous

variation in climate, as well as provide a range of empirical evidence suggesting that climate shocks alter

suicide rates, at least in part, via economic channels. The magnitudes of climate impacts on agricultural

yield values and suicides are both economically and socially significant. Annual suicides per 100,000

people increase by 10.1%/σ, and annual yields fall by 1%/σ, for an increase in daily temperature

within the 10-20◦C range. Growing season precipitation increases annual yields by 0.9%/σ, but has

no identifiable impact on suicide rates. We use seasonal “fingerprinting” and yield pattern matching

to show that temperature impacts suicides and yields similarly, suggesting the presence of economic

motives for suicide. For both temperature and precipitation, we employ a variety of heterogeneity

tests and temporal disaggregation to establish further correlational evidence that agricultural losses

motivate suicide in India.

However, our approach has important limitations. Primarily, we face the same challenge confronted

in the climate and violence literature: without the ability to shut down the economic channel through

which climate affects suicides, we cannot concretely identify the relevance of this mechanism. Although

an instrumental variables methodology would likely be inappropriate due to a failure of the exclusion

restriction, another option would be to include yields and climate jointly in a regression. However,

17

our agricultural and suicide data do not sufficiently overlap spatially or temporally to enable us to use

this strategy.11 Secondly, the low spatial resolution of our suicide data limits the statistical power of

our model and precludes us from exploring more detailed levels of heterogeneity, such as the impact

of annual variation in irrigation or the differential responses in rural and urban areas. Perhaps due to

this limited sample size, some of our results are not robust to the inclusion of regional time trends.

Finally, although our temperature results clearly demonstrate a pattern match between agricultural

yields and suicides, our rainfall findings are less clear. Due to extensive public discussion on growing

season drought and suicide in India, our results suggests that further exploration of rainfall’s impacts

on welfare is needed.

Despite these shortcomings, our results have important policy implications. Individual states in

India, as well as the national government, have debated, enacted and sometimes repealed legislation

seeking to prevent suicides by providing transfers or debt relief to subsistence farmers. Our results

suggest that these policies could be effective, particularly in years of agriculturally damaging climate

shocks. Moreover, our findings on the large magnitude of climate impacts on suicide is important for

climate change policy design. Given the predicted increase in both temperature and rainfall for India

over the coming century (Stocker et al., 2014), our results suggest a climate-driven rise in suicide rates

in coming years for a country already battling a growing number of lost lives each year.

References

Adam, Christopher, David Bevan, and Douglas Gollin. 2013. “Rural-Urban Linkages, TransactionCosts, and Poverty Alleviation: The Case of Tanzania.”

Anderson, Steven W, Antoine Bechara, Hanna Damasio, Daniel Tranel, and Antonio R Damasio. 1999.“Impairment of social and moral behavior related to early damage in human prefrontal cortex.”Nature neuroscience 2 (11):1032–1037.

Andres, Antonio Rodriguez. 2005. “Income inequality, unemployment, and suicide: a panel dataanalysis of 15 European countries.” Applied Economics 37 (4):439–451.

Auffhammer, Maximilian, V Ramanathan, and Jeffrey R Vincent. 2006. “Integrated model shows thatatmospheric brown clouds and greenhouse gases have reduced rice harvests in India.” Proceedingsof the National Academy of Sciences 103 (52):19668–19672.

Barreca, Alan, Karen Clay, Olivier Deschenes, Michael Greenstone, and Joseph S Shapiro. 2013.“Adapting to climate change: the remarkable decline in the US temperature-mortality relationshipover the 20th century.” National Bureau of Economic Research Working Paper .

11Using the states and years for which both yield and suicide data are a available leads to a sample size of 273. Resultsfrom regressions of suicide on climate variation controlling for yield are available upon request, but all climate variablesare statistically insignificant, likely due to the small sample.

18

Basyan, Ceren, F. Gonzalez, M. Burke, S. Hsiang, and E. Miguel. 2014. “Economic and non-economicfactors in violence: Evidence from organized crime, suicides, and climate in Mexico.”

Becker, Gary S and Richard A Posner. 2004. “Suicide: An economic approach.” University of Chicago.

Blakeslee, David and Ram Fishman. 2013. “Rainfall shocks and property crimes in agrarian societies:Evidence from India.” Available at SSRN 2208292 .

Botha, Ferdi. 2012. “The Economics Of Suicide In South Africa.” South African Journal of Economics80 (4):526–552.

Burgess, Robin, Olivier Deschenes, Dave Donaldson, and Michael Greenstone. 2014. “The UnequalEffects of Weather and Climate Change: Evidence from Mortality in India.” Working Paper .

Burke, Marshall, Solomon M Hsiang, and Edward Miguel. 2014. “Climate and Conflict.” Tech. rep.,National Bureau of Economic Research.

Chemin, Matthieu, Joost De Laat, and Johannes Haushofer. 2013. “Negative Rainfall Shocks IncreaseLevels of the Stress Hormone Cortisol Among Poor Farmers in Kenya.” Available at SSRN 2294171.

Cutler, David M, Edward L Glaeser, and Karen E Norberg. 2001. “Explaining the rise in youth suicide.”In Risky behavior among youths: An economic analysis. University of Chicago Press, 219–270.

Davidson, Richard J, Katherine M Putnam, and Christine L Larson. 2000. “Dysfunction in the neuralcircuitry of emotion regulation–a possible prelude to violence.” Science 289 (5479):591–594.

Deisenhammer, EA, G Kemmler, and P Parson. 2003. “Association of meteorological factors withsuicide.” Acta Psychiatrica Scandinavica 108 (6):455–459.

Deryugina, Tatyana and Solomon M Hsiang. 2014. “Does the Environment Still Matter? Daily Tem-perature and Income in the United States.” Tech. rep., National Bureau of Economic Research.

Deschenes, Olivier and Michael Greenstone. 2011. “Climate Change, Mortality, and Adaptation: Ev-idence from Annual Fluctuations in Weather in the US.” American Economic Journal: AppliedEconomics :152–185.

Deshpande, RS. 2002. “Suicide by farmers in Karnataka: agrarian distress and possible alleviatorysteps.” Economic and Political Weekly :2601–2610.

Duflo, Esther and Rohini Pande. 2007. “Dams*.” The Quarterly journal of economics 122 (2):601–646.

Durkheim, Emile. 1951. “Suicide: A study in sociology (JA Spaulding & G. Simpson, trans.).” Glencoe,IL: Free Press.(Original work published 1897) .

Fetzer, Thiemo. 2014. “Can workfare programs moderate violence? evidence from india.” .

Foster, Andrew D and Mark R Rosenzweig. 1996. “Technical change and human-capital returns andinvestments: evidence from the green revolution.” The American economic review :931–953.

Gine, Xavier, Robert Townsend, and James Vickery. 2007. “Statistical analysis of rainfall insurancepayouts in southern India.” American Journal of Agricultural Economics 89 (5):1248–1254.

Gollin, Douglas and Richard Rogerson. 2010. “Agriculture, roads, and economic development inUganda.” Tech. rep., National Bureau of Economic Research.

Gould, Madelyn, Patrick Jamieson, and Daniel Romer. 2003. “Media contagion and suicide among theyoung.” American Behavioral Scientist 46 (9):1269–1284.

Government of India. 2011. “Census of India.” URL http://censusindia.gov.in.

Gruere, Guillaume and Debdatta Sengupta. 2011. “Bt cotton and farmer suicides in India: an evidence-based assessment.” The journal of development studies 47 (2):316–337.

Guiteras, Raymond. 2009. “The impact of climate change on Indian agriculture.” Manuscript, De-partment of Economics, University of Maryland, College Park, Maryland .

19

Hamermesh, Daniel S and Neal M Soss. 1974. “An economic theory of suicide.” The journal of politicaleconomy :83–98.

Hastings Jr, Cecil, Frederick Mosteller, John W Tukey, and Charles P Winsor. 1947. “Low momentsfor small samples: a comparative study of order statistics.” The Annals of Mathematical Statistics:413–426.

Hebous, Sarah and Stefan Klonner. 2014. “Economic Distress and Farmer Suicides in India: AnEconometric Investigation.” .

Herring, Ronald J. 2008. “Whose numbers count? Probing discrepant evidence on transgenic cotton inthe Warangal district of India.” International Journal of Multiple Research Approaches 2 (2):145–159.

Hidalgo, F Daniel, Suresh Naidu, Simeon Nichter, and Neal Richardson. 2010. “Economic determinantsof land invasions.” The Review of Economics and Statistics 92 (3):505–523.

Hsiang, Solomon M, Marshall Burke, and Edward Miguel. 2013. “Quantifying the influence of climateon human conflict.” Science 341 (6151):1235367.

Hsiang, Solomon M and Kyle C Meng. 2014. “Reconciling disagreement over climate–conflict resultsin Africa.” Proceedings of the National Academy of Sciences 111 (6):2100–2103.

Inagaki, Kazuyuki. 2010. “Income inequality and the suicide rate in Japan: evidence from cointegrationand LA-VAR.” Journal of Applied Economics 13 (1):113–133.

Jessen, Gert, BØrge F Jensen, and Peter Steffensen. 1998. “Seasons and meteorological factors insuicidal behaviour.” Archives of suicide research 4 (3):263–280.

Jonathan Kennedy, Lawrence King. 2014. “The political economy of farmers’ suicides in India: in-debted cash-crop farmers with marginal landholdings explain state-level variation in suicide rates.”Globalization and Health 10 (16).

Lloyd, Chris J and Paul SF Yip. 2001. “A comparison of suicide patterns in Australia and HongKong.” Journal of the Royal Statistical Society: Series A (Statistics in Society) 164 (3):467–483.

Lovheim, Hugo. 2012. “A new three-dimensional model for emotions and monoamine neurotransmit-ters.” Medical hypotheses 78 (2):341–348.

Mirza, M Monirul Qader. 2002. “Global warming and changes in the probability of occurrence of floodsin Bangladesh and implications.” Global environmental change 12 (2):127–138.

———. 2011. “Climate change, flooding in South Asia and implications.” Regional EnvironmentalChange 11 (1):95–107.

Mohanty, BB and Sangeeta Shroff. 2004. “Farmers’ suicides in Maharashtra.” Economic and PoliticalWeekly :5599–5606.

Mohanty, Bibhuti B. 2005. “‘We are Like the Living Dead’: Farmer Suicides in Maharashtra, WesternIndia.” Journal of Peasant Studies 32 (2):243–276.

Nagaraj, K. 2008. Farmers’ suicides in India: Magnitudes, trends and spatial patterns. BharathiPuthakalayam.

NDTV. 2015. “Crops Hit by March Rains, Farmer Commits Sui-cide in Uttar Pradesh.” URL http://www.ndtv.com/india-news/crops-in-five-states-hit-by-march-rains-farmer-commits-suicide-in-uttar-pradesh-744078.

Patel, Raj. 2007. Stuffed & starved. Black Inc.

Patel, Vikram, Chinthanie Ramasundarahettige, Lakshmi Vijayakumar, JS Thakur, Vendhan Gajalak-shmi, Gopalkrishna Gururaj, Wilson Suraweera, and Prabhat Jha. 2012. “Suicide mortality in India:a nationally representative survey.” The Lancet 379 (9834):2343–2351.

Paul, Bimal Kanti and Harun Rasid. 1993. “Flood damage to rice crop in Bangladesh.” GeographicalReview :150–159.

20

Pierard, Emmanuelle and Paul Grootendorst. 2014. “Do downturns cause desperation? The effect ofeconomic conditions on suicide rates in Canada.” Applied Economics 46 (10):1081–1092.

Rao, VM and DV Gopalappa. 2004. “Agricultural Growth and Farmer Distress: Tentative Perspectivesfrom Karnataka.” Economic and Political Weekly :5591–5598.

Rehkopf, David H and Stephen L Buka. 2006. “The association between suicide and the socio-economiccharacteristics of geographical areas: a systematic review.” Psychological medicine 36 (02):145–157.

Rosenzweig, Mark and Christopher R Udry. 2013. “Forecasting profitability.” Working paper .

Sarma, EAS. 2004. “Is rural economy breaking down? Farmers’ suicides in Andhra Pradesh.” Economicand Political Weekly :3087–3089.

Schlenker, Wolfram and Michael J Roberts. 2009. “Nonlinear temperature effects indicate severedamages to US crop yields under climate change.” Proceedings of the National Academy of sciences106 (37):15594–15598.

Seidler, Michael J. 1983. “Kant and the Stoics on Suicide.” Journal of the History of Ideas :429–453.

Seo, Dongju, Christopher J Patrick, and Patrick J Kennealy. 2008. “Role of serotonin and dopaminesystem interactions in the neurobiology of impulsive aggression and its comorbidity with other clinicaldisorders.” Aggression and violent behavior 13 (5):383–395.

Snipes, Michael, Timothy M Cunha, and David D Hemley. 2012. “Unemployment Fluctuations andRegional Suicide: 1980-2006.” Journal of Applied Economics & Business Research 2 (2).

Stack, Steven. 2000. “Suicide: a 15-year review of the sociological literature part I: cultural andeconomic factors.” Suicide and Life-Threatening Behavior 30 (2):145–162.

Stocker, Thomas, Dahe Qin, Gian-Kasper Plattner, M Tignor, Simon K Allen, Judith Boschung,Alexander Nauels, Yu Xia, Vincent Bex, and Pauline M Midgley. 2014. Climate change 2013: Thephysical science basis. Cambridge University Press Cambridge, UK, and New York.

Velting, Drew M and Madelyn S Gould. 1997. “Suicide contagion.” .

WHO. 2014. “Preventing suicide: a global imperative.” World Health Organization. .

Willmott, Cort and Kenji Matsuura. 2014. “Global land precipitation and temperature.” URL http://climate.geog.udel.edu/~climate/.

21

8 Figures

(a) 1967 - 1977 (b) 1978 - 1987

(c) 1988 - 1997 (d) 1998 - 2013

Figure 1: Suicide rates over time

Notes: Suicide rates are measured annually per 100,000 people. Suicide count data are from India’s National CrimeRecords Bureau; population data are from the decadal Indian Census, linearly interpolated between census years. Thefull sample mean suicide rate is 9.47.

22

-.01

-.005

0

.005

.01

.015

Tem

pera

ture

0 1 2

Growing Season

-.01

-.005

0

.005

.01

.015

0 1 2

Non-Growing Season

-.01

-.005

0

.005

.01

.015

Prec

ipita

tion

0 1 2Lag

-.01

-.005

0

.005

.01

.015

0 1 2Lag

Point Estimate 95% Confidence Interval

Figure 2: Lagged effects of climate on suicide

Notes: This figure shows coefficients for the impact of climate variables on suicide rates, estimated from Equation 3 usinga degree day cutoff of 15◦C and including two lags on all climate variables. Only coefficients for the first order of theprecipitation polynomial only are shown.

23

(a) Suicide-temperature response

9

9.4

9.8An

nual

Sui

cide

Rat

e pe

r 100

,000

5 15 25 35Daily Temperature (Celcius)

Fixed EffectsFE + Region by Year Trend

Growing Season

9

9.4

9.8

5 15 25 35Daily Temperature (Celcius)

Non-Growing Season

95% confidence intervals for FE model. Mean annual suicide rate is 9.47 per 100,000.

3.92

3.93

3.94

Log

Yiel

d (R

upee

s pe

r ha)

5 15 25 35Daily Temperature (Celcius)

Fixed EffectsFE + Region by Year Trend

Growing Season

3.92

3.93

3.94

5 15 25 35Daily Temperature (Celcius)

Non-Growing Season

95% confidence intervals for FE model. Mean log annual yield is 3.93.

(b) Yield-temperature response

Figure 3: Pattern matching of piecewise linear temperature response functions

Notes: Figure plots response functions using temperature bins of width 10◦C from (a) a suicide regression with annualdata for 32 Indian states between 1967 and 2013, and (b) an agricultural regression with annual data for 271 Indiandistricts between 1956 and 2000. Growing season is June-September, non-growing season contains all other months.Regional time trends use regions defined by Nagaraj (2008) for suicide and Guiteras (2009) for agriculture. Standarderrors are clustered at the state level.

24

0

.2

.4

.6

.8

Den

sity

0 5 10 15Daily Degree Days

Growing Season Non-Growing Season

Figure 4: Distribution of cumulative degree days above 20◦C in the growing and non-growing seasons

Notes: This figure shows the distribution of daily degree days above 20◦C for the growing and non-growing seasons, usingdaily mean temperature for 32 of India’s states between 1967-2013. The growing season is June through September, whilethe non-growing season is all other months.

Figure 5: Geographic heterogeneity in the suicide-temperature response

Notes: Marginal effects are for the growing season only, and expressed as the percentage of the average coefficient acrossall states. They are estimated using a degree day model with a cutoff of 15◦C. Darker states correspond to largercoefficients; yellow indicates a negative effect.

25

-5

0

5

10R

elat

ive

Effe

ct o

f Tem

pera

ture

on

Suic

ides

-.1 0 .1 .2 .3 .4Average Farmer Suicide Ratio 2001-2010

Chhattisgarh

Kerala

Karnataka

Maharashtra

Andhra Pradesh

(a) Farmer suicide intensity

-2.5

0

2.5

5

Rel

ativ

e Ef

fect

of T

empe

ratu

re o

n Su

icid

e

-5 -2.5 0 2.5 5Relative Effect of Temperature on Yield

Andhra Pradesh

Karnataka

Tamil Nadu

Maharashtra

(b) Yield response

Figure 6: Correlations between agricultural variables and state-level suicide responses to temperature

Notes: Panel (a) shows the correlation between state-level suicide responses to growing season temperature and theaverage farmer suicide ratio, defined as farm suicides divided by total suicides. Panel (b) shows the correlation betweenstate-level suicide responses and yield responses to growing season temperature, where only 13 states are included due tolimited agricultural data. Temperature effects are expressed as each state’s proportion of the average treatment, wheretemperature is measured as growing season degree days above 15◦C. Note that Kerala and Chhattisgarh are not statesincluded in the agriculture data and therefore are not plotted in Panel (b).

-4

-2

0

2

4

6

Frac

tion

of th

e Te

mpe

ratu

re A

TE

0 1000 2000 3000 4000 5000Average Degree Days Above 15C

(a) Average temperature

-2

0

2

4

6

Frac

tion

of th

e Te

mpe

ratu

re A

TE

0 10 20 30 40Average Suicide Rate

(b) Social contagion

Figure 7: Potential drivers of vulnerability to temperature

Notes: These figures plot the marginal effect of temperature on suicide for each of 32 Indian states, relative to (a)long-run average temperature and (b) long-run average suicide rates. Annual suicide data cover 1967-2013 and marginaleffects are calculated using a degree days specification with a cutoff of 15◦C.

26

(a) Suicide-precipitation response

5

10

15

0 1000 2000 3000Cumulative Precipitation (mm)

Fixed EffectsFE + Region by Year Trend

Growing Season

5

10

15

0 1000 2000 3000Cumulative Precipitation (mm)

Non-Growing Season

95% confidence interval. Dashed line shows linear model. Mean annual suicide rate is 11.4 per 100,000.

3.6

4

4.4

Log

Yiel

d (R

upee

s pe

r ha)

0 1000 2000 3000Cumulative Precipitation (mm)

Fixed EffectsFE + Region by Year Trend

Growing Season

3.6

4

4.4

0 1000 2000 3000Cumulative Precipitation (mm)

Non-Growing Season

95% confidence interval. Dashed line shows linear model. Mean log annual yield is 3.93

(b) Yield-precipitation response

Figure 8: Pattern matching of nonlinear precipitation response functions

Notes: Figure plots response functions using third degree polynomials in precipitation from (a) a suicide regressionwith annual data for 32 Indian states between 1967 and 2013, and (b) an agricultural regression with annual data for271 Indian districts between 1956 and 2000. Growing season is June-September, non-growing season contains all othermonths. Regional time trends use regions defined by Nagaraj (2008) for suicide and Guiteras (2009) for agriculture.Standard errors are clustered at the state level.

27

5

10

15

20

Annu

al S

uici

de R

ate

per 1

00,0

00

0 1000 2000 3000

Fixed EffectsFE + Region by Year Trend

Growing Season, Irrigated

5

10

15

20

0 1000 2000 3000

Non-Growing Season, Irrigated

5

10

15

20

Annu

al S

uici

de R

ate

per 1

00,0

00

0 1000 2000 3000Cumulative Precipitation (mm)

Growing Season, Rain-fed

5

10

15

20

0 1000 2000 3000Cumulative Precipitation (mm)

Non-Growing Season, Rain-fed

95% confidence interval. Dashed line shows linear model. Mean annual suicide rate is 9.47 per 100,000.

Figure 9: Heterogeneity of suicide-precipitation response by irrigated area

Notes: This figure plots response functions from a regression using annual data for 32 Indian states between 1967 and2013, where all precipitation variables are interacted with an indicator for being an “irrigated” state. A state is consideredirrigated if its share of cultivated area under irrigation exceeded 50% for any year in my study sample (1967-2013), usingdata from the Indian Ministry of Agriculture.

-.015

-.01

-.005

0

.005

.01

Cha

nge

in S

uici

de R

ates

per

mm

of R

ainf

all

jan feb mar apr may jun jul aug sep oct nov dec

Figure 10: Suicide-precipitation response by month

Notes: This figure plots the coefficients from a regression of annual suicide rates on cumulative millimeters of rainfall inindividual months using data from 32 of India’s states between 1967-2013. Marginal effects were calculated controllingfor a piecewise linear function of temperature in four 10◦C bins, state and year fixed effects, and region-specific timetrends with errors clustered at the state level. The growing season months are shaded in pink, harvest months shaded ingreen.

28

9 Tables

Variable Mean (Std. Dev.) Min. Max. N

Suicide Data: 1967 - 2013 (32 States)

Suicide Rate (per 100,00 people) 11.4 (11.9) 0 73.23 1472Average Daily Temp (C) 20.94 (6.93) 0.73 29.39 1645Average Growing Season Daily DD > 20C 5.32 (2.93) 0 12.23 1645Average Non-Growing Season Daily DD > 20C 3.85 (2.83) 0 9.56 1645Growing Season Precip. (mm) 1176.38 (614.96) 108.28 3176.3 1598Non-Growing Season Precip. (mm) 437.35 (359.49) 5.76 1609.95 1598

Agricultural Data: 1956 - 2000 (13 States)

Log Annual Yield (Rupees per ha) 3.93 (0.72) -1.87 6.45 11379Daily Temperature (◦C) 22.42 (7) 0 28.47 11705Growing Season Degree Days > 20◦C 754.51 (365.04) 0 1799.64 11705Non-Growing Season Degree Days > 20◦C 1043.49 (458.49) 0 2069.25 11705Growing Season Precipitation (mm) 853.81 (491.8) 13.15 3036.64 11870Non-Growing Season Precipitation (mm) 201.83 (186.74) 0.37 934.18 11870

Table 1: Summary statistics

Note: Suicide data are from India’s National Crime Records Bureau and are reported annually at the state level. Yielddata are from Duflo and Pande (2007) and are reported annually at the district level, valued in constant Rupees. Growingseason is June-September, non-growing season contains all other months. Precipitation is measured cumulatively.

29

Dependent variable: Suicide rate per 100,000

(1) (2) (3) (4) (5)OLS Year FE State FE Year and State FE All FE + Trends

Growing Season Temp. 0-10◦ C 0.00162 0.00111 -0.0105*** -0.00848** -0.00986**(0.00792) (0.00881) (0.00294) (0.00349) (0.00383)

Growing Season Temp. 10-20◦ C 0.0284* 0.0370*** 0.0303 0.0360** 0.0399**(0.0155) (0.0130) (0.0190) (0.0170) (0.0163)

Growing Season Temp. 20-30◦ C 0.00837 0.00751 0.0257*** 0.0119 0.00358(0.00949) (0.00974) (0.00869) (0.00817) (0.00804)

Growing Season Temp. >30◦ C -0.00196 0.00140 -0.00968 0.000821 0.00851(0.0190) (0.0195) (0.00967) (0.00738) (0.00677)

Non-Growing Season Temp. 0-10◦ C -0.0214* -0.0253*** -0.0168 -0.0181** -0.0180*(0.0106) (0.00892) (0.0114) (0.00872) (0.00891)

Non-Growing Season Temp. 10-20◦ C 0.00120 0.000981 0.00609 -0.000756 0.000415(0.00608) (0.00615) (0.00533) (0.00429) (0.00414)

Non-Growing Season Temp. 20-30◦ C 0.0137*** 0.0141*** 0.000528 0.00111 0.000446(0.00482) (0.00487) (0.00278) (0.00376) (0.00355)

Non-Growing Season Temp. >30◦ C -0.0339 -0.0273 -0.000936 0.00476 0.00569(0.0237) (0.0265) (0.00461) (0.00568) (0.00542)

Growing Season Precip. (mm) 0.0228 0.0233 0.00279 0.00359 0.00320(0.0143) (0.0154) (0.00270) (0.00304) (0.00305)

Growing Season Precip.2 (mm) -1.29e-05 -1.34e-05 -1.49e-06 -1.36e-06 -1.21e-06(1.12e-05) (1.20e-05) (1.97e-06) (2.27e-06) (2.25e-06)

Growing Season Precip.3 (mm) 1.93e-09 2.06e-09 1.01e-10 8.89e-11 6.61e-11(2.31e-09) (2.44e-09) (4.08e-10) (4.84e-10) (4.76e-10)

Non-Growing Season Precip. (mm) -0.00680 -0.00268 -0.00538 0.00180 0.00131(0.0175) (0.0194) (0.00505) (0.00537) (0.00489)

Non-Growing Season Precip.2 (mm) 2.67e-05 2.32e-05 5.86e-06 -1.24e-06 -1.07e-06(2.61e-05) (2.84e-05) (7.03e-06) (7.17e-06) (6.59e-06)

Non-Growing Season Precip.3 (mm) -1.24e-08 -1.12e-08 -1.71e-09 6.27e-10 6.36e-10(1.10e-08) (1.18e-08) (2.73e-09) (2.79e-09) (2.60e-09)

Constant -3.593 -6.336 11.23 29.43 34.32(5.809) (7.132) (9.702) (20.56) (22.02)

Observations 1,434 1,434 1,434 1,434 1,434R-squared 0.560 0.583 0.876 0.896 0.899Year FE NO YES NO YES YESState FE NO NO YES YES YESRegion × Year Trend NO NO NO NO YES

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Table 2: Climate’s impact on suicide

Notes: Regression includes annual data for 32 Indian states between 1967 and 2013. Growing season is June-September,non-growing season contains all other months. Regional time trends use regions defined by Nagaraj (2008) and describedin the text. Standard errors are clustered at the state level.

30

Dependent variable: Log yield value (Rupees per ha)

(1) (2) (3) (4) (5)OLS Year FE State FE Year and State FE All FE + Trend

Growing Season Temp. 0-10◦ C -0.000953*** -0.000943*** -0.000943*** -0.000157 -0.000221(0.000334) (0.000330) (0.000330) (0.000409) (0.000460)

Growing Season Temp. 10-20◦ C 0.000660 0.000336 0.000336 -0.000806** -0.000665*(0.00102) (0.00112) (0.00112) (0.000367) (0.000388)

Growing Season Temp. 20-30◦ C 0.000310 0.000258 0.000258 -0.000310*** -0.000225**(0.000202) (0.000201) (0.000201) (0.000117) (0.000111)

Growing Season Temp. >30◦ C -0.00110 -0.00109 -0.00109 -3.46e-05 -1.77e-05(0.000731) (0.000778) (0.000778) (0.000265) (0.000259)

Non-Growing Season Temp. 0-10◦ C 0.000426 0.000611 0.000611 0.000507*** 0.000527***(0.000654) (0.000716) (0.000716) (0.000187) (0.000182)

Non-Growing Season Temp. 10-20◦ C -0.000412** -0.000426** -0.000426** 0.000101* 5.90e-05(0.000175) (0.000185) (0.000185) (5.80e-05) (5.20e-05)

Non-Growing Season Temp. 20-30◦ C -2.05e-05 -3.91e-05 -3.91e-05 -6.39e-05 -0.000163**(0.000174) (0.000177) (0.000177) (8.22e-05) (7.06e-05)

Non-Growing Season Temp. >30◦ C -0.000924** -0.000984** -0.000984** 0.000532*** 0.000657***(0.000358) (0.000429) (0.000429) (0.000128) (0.000124)

Growing Season Precip. (mm) 0.00132** 0.00136** 0.00136** 0.000673*** 0.000687***(0.000533) (0.000524) (0.000524) (0.000147) (0.000140)

Growing Season Precip.2 (mm) -1.03e-06** -9.96e-07** -9.96e-07** -4.45e-07*** -4.40e-07***(4.39e-07) (4.22e-07) (4.22e-07) (1.18e-07) (1.12e-07)

Growing Season Precip.3 (mm) 2.24e-10** 2.09e-10** 2.09e-10** 9.97e-11*** 8.67e-11***(9.58e-11) (9.11e-11) (9.11e-11) (0) (0)

Non-Growing Season Precip. (mm) 0.00438*** 0.00412*** 0.00412*** 0.000895*** 0.000867***(0.000458) (0.000509) (0.000509) (0.000141) (0.000135)

Non-Growing Season Precip.2 (mm) -7.73e-06*** -6.71e-06*** -6.71e-06*** -1.92e-06*** -1.81e-06***(1.31e-06) (1.35e-06) (1.35e-06) (3.73e-07) (3.54e-07)

Non-Growing Season Precip.3 (mm) 4.69e-09*** 3.84e-09*** 3.84e-09*** 1.24e-09*** 1.15e-09***(9.52e-10) (9.33e-10) (9.33e-10) (2.70e-10) (2.57e-10)

Constant 3.190*** 2.559*** 2.559*** 3.158*** -2.272(0.210) (0.233) (0.233) (0.0717) (3.450)

Observations 11,218 11,218 11,218 11,218 11,218R-squared 0.213 0.408 0.408 0.847 0.849Year FE NO YES NO YES YESDistrict FE NO NO YES YES YESRegion × Year Trend NO NO NO NO YES

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Table 3: Piecewise linear model for value of log yield (Rupees per hectare)

Notes: Regression includes annual data for 13 Indian states between 1956 and 2000. Growing season is June-September,non-growing season contains all other months. Regional time trends use regions defined by Guiteras (2009). Standarderrors are clustered at the state level.

31

Dependent variable: Suicides per 100,000

(1) (2) (3) (4) (5)OLS Year FE State FE Year and State FE All FE + Trend

Growing Seas. Degree Days -0.00694* -0.00717* 0.0154*** 0.00827* 0.00632(0.00378) (0.00400) (0.00320) (0.00422) (0.00408)

Non-Growing Seas. Degree Days 0.00905*** 0.00904*** 0.000870 -0.00177 -0.00181(0.00228) (0.00235) (0.00121) (0.00284) (0.00308)

Growing Season Precip. (mm) 0.00221 0.00357 0.00366 0.00414 0.00341(0.0104) (0.0110) (0.00234) (0.00308) (0.00308)

Growing Season Precip.2 (mm) -1.34e-06 -2.57e-06 -1.92e-06 -1.33e-06 -9.32e-07(9.49e-06) (9.94e-06) (1.77e-06) (2.28e-06) (2.25e-06)

Growing Season Precip.3 (mm) 0 3.10e-10 1.83e-10 7.19e-11 -0(2.10e-09) (2.18e-09) (3.72e-10) (4.74e-10) (4.66e-10)

Non-Growing Season Precip. (mm) 0.00301 0.00205 -0.00425 0.000628 0.000217(0.0150) (0.0167) (0.00509) (0.00547) (0.00503)

Non-Growing Season Precip.2 (mm) 1.86e-05 1.98e-05 6.29e-06 1.57e-06 1.39e-06(2.55e-05) (2.79e-05) (7.05e-06) (7.24e-06) (6.69e-06)

Non-Growing Season Precip.3 (mm) -9.36e-09 -9.31e-09 -2.28e-09 -6.70e-10 -4.80e-10(1.19e-08) (1.27e-08) (2.73e-09) (2.80e-09) (2.61e-09)

Constant -0.325 -0.192 7.465 22.89 27.39*(5.527) (6.431) (7.730) (15.64) (15.66)

Observations 1,434 1,434 1,434 1,434 1,434R-squared 0.471 0.499 0.867 0.888 0.892Year FE NO YES NO YES YESState FE NO NO YES YES YESRegion × Year Trend NO NO NO NO YES

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Table 4: 15◦C degree day model of climate’s impact on suicide

Notes: Regression includes annual data for 32 Indian states between 1967 and 2013. Growing season is June-September,non-growing season contains all other months. Degree day cutoff is 15◦C. Regional time trends use regions defined byNagaraj (2008) and described in the text. Standard errors are clustered at the state level.

32

Dependent variable: Suicides per 100,000

(1) (2) (3)Year FE Regional Trends Year FE + Region Trends

Growing Seas. Degree Days 0.00906** 0.00577*** 0.00735**(0.00375) (0.00201) (0.00307)

Non-Growing Seas. Degree Days 0.000650 -0.00133 0.000950(0.00217) (0.00111) (0.00213)

Drought: 20th percentile -0.469 -0.191 -0.317(0.310) (0.242) (0.317)

Surplus: 80th percentile 0.414 -0.458** 0.211(0.343) (0.217) (0.299)

Constant 59.02** -21.39*** 113.7***(21.82) (6.636) (10.38)

Observations 1,472 1,472 1,472R-squared 0.892 0.908 0.916State FE YES YES YESYear FE YES NO YESRegion × Year Trend NO YES YES

Standard errors clustered at the state level*** p<0.01, ** p<0.05, * p<0.1

Table 5: Climate’s impact on suicide: indicators for drought and surplus

Notes: Regression includes annual data for 32 Indian states between 1967 and 2013. Growing season is June-September,non-growing season contains all other months. Drought is defined as an indicator equal to one when annual rainfall is inthe 20th percentile or below, while surplus is equal to one when annual rainfall is above the 80th percentile. Standarderrors are clustered at the state level.

33