Climate Change, Rainfall Variability, and Adaptation through Irrigation: Evidence from Indian Agriculture ⇤ Ram Mukul Fishman

†

January 30, 2012

⇤ I thank Wolfram Schlenker, Upmanu Lall, Douglass Almond, Bernard Salanie, Je↵rey Sachs, Solomon Hsiang, Chandra Kiran Krishnamurti, Naresh Divedani, Shama Parveen, Tobias Siegfried, Jesse Anttila-Hughes, Gordon McCord, Kapil Narula, Avinash Kishore, Brian Dillon, Ruth Defries, Noel Michele Holbrook, Matthew Gilbert and the participants of the Sustainable Development Seminar in Columbia University. I thank the Columbia Water Center for Technical and Financial support. † Harvard Kennedy School. email: ram [email protected]

1

Abstract In many developing countries, especially in the semi-arid tropics, rain fed agriculture is still highly vulnerable to irregular rainfall patterns. Climate models predict precipitation to become even more variable with global warming, with potentially severe consequences for food production, but there is little quantitative estimates of these linkages. I combine detailed data on daily rainfall, crop yields and irrigation from over three decades in India to provide first estimates of the impacts of intraseasonal rainfall variability on crop yields, and of the the adaptive resilience achieved by the expansion of irrigation in the country during this time period. I then use these estimates to simulate the impact of climate change on rice production in India under various irrigation scenarios. I find that the intra-seasonal distribution of rainfall has an impact on rice yields with a magnitude and significance that rival that of total seasonal rainfall, and that taking it into account flips the sign of the projected precipitation-related climate change impact on crop production from positive (due to increased total rainfall) to negative (due to the simultaneous projected increase in the number of rainless days). I also provide evidence that farmers expand irrigation in response to adverse rainfall conditions and that irrigated yields are less vulnerable to the impacts of rainfall irregularities, and estimate that the expansion of irrigation to cover all the rice cultivating farmland in India could eliminate 80% of the projected precipitation-driven climate change impact. However, given that the current excessive use of groundwater in India may deplete the country’s aquifers, the projected impact may actually increase by 40%. I find no evidence to suggest irrigation attenuates the impact of increased heat exposure on rice yields. Since I find temperature increase to drive most of the projected decline in yield, this suggests the expansion of irrigation, often mentioned as a primary adaptation strategy, to have limited overall potential to achieve this goal.

2

1

Introduction

Perhaps more than any other sector of the economy, agriculture has always been and still remains vulnerable to changing weather conditions, because of the strong e↵ect temperatures and moisture availability have on the key processes involved in crop growth. Weather related volatility in agricultural production is still a primary concern for food security and economic development, especially in tropical regions (Rosenzweig and Hillel [2008], Bank [2007]). Furthermore, there are growing concerns that altered weather patterns associated with climate change may have significant negative consequences for food production in developing countries (Parry [2007]). While most studies of the impact of climatic variability and change on food production have focused on temperatures increases and shifts in mean precipitation amounts, crop yields are also highly sensitive to irregular precipitation events like dry spells and intense downpours, and much of the volatility in food production in developing countries, especially in the semi arid tropics, is thought to be driven by such precipitation irregularities. Moreover, there is a general expectation in the climate science literature that global warming will intensify the hydrological cycle, and result in more uneven intra-seasonal distribution of precipitation (Trenberth et al. [2005], Hennessy et al. [1997], Goswami et al. [2006]). The IPCC’s fourth assessment report

1

concludes that

“...in a warmer world, precipitation tends to be concentrated into more intense events, with longer periods of little precipitation in between. Therefore, intense and heavy downpours would be interspersed with longer relatively dry periods...” but that

2

1 ‘‘http://www.ipcc.ch/publications_and_data/ar4/wg1/en/faq-10-1.html’’ 2 ‘‘http://http://www.ipcc.ch/pdf/assessment-report/ar4/wg2/ar4-wg2-chapter5. pdf’’

3

“Although a few models since the Third Assessment Report have started to incorporate e↵ects of climate variability on plant production, most studies continue to include only e↵ects on changes in mean variables”. This paper provides what are, to the best of my knowledge, first econometric estimates of the impact of intra-seasonal rainfall variability on crop yields. My analysis uses detailed weather and agricultural data spanning three decades from across India, where food for a sixth of the world’s population is produced, and where agriculture has always been vulnerable to the vagaries of the monsoon3 (Krishna Kumar et al. [2004]), and finds a strong, significant impact on crop yields that rivals that of total rainfall. For example, I estimate that each additional rainless day reduces rice yields in the rainy season by 0.4% (keeping total rainfall fixed). In comparison, the impact of the associated decrease in total rainfall (about 10mm per day on average) would reduce yields by only about 0.15%, and the associated average increase in temperature degree days (about 1 degree-day lower in a rainy day, on average) would reduce yields by 0.09%. These estimates, together with climate model simulations of changes in the frequency of rainless days (for example Krishna Kumar et al. [2003]) flip the sign of the precipitation related climate change impact on rice production in India from a modest gain to a roughly 10% loss (in comparison to a simulation based solely on mean rainfall increases). In general, though, predicting the impacts of climate change on the basis of the short term e↵ects of temporary weather shocks neglects to take into account possible adaptation strategies that can make production less vulnerable. The paucity of evidence on the feasibility and costs of such adaptation strategies has thus far made the cost benefit analysis of climate change mitigation incomplete (Tol [2009]). 3 the period lasting roughly from June to September during which the great majority of India’s rainfall takes place, and on which its entire annual agricultural production depends for its water requirement.

4

In the second part of the paper, I partially address this gap by studying the usage of irrigation in India and the degree to which the expansion of irrigation can be an e↵ective means of adapting agriculture to the changing climate, in particular to rainfall irregularity. For millennia, irrigation has been used to supply crops with a more reliable source of water: irrigation water can substitute for deficient rainfall (especially in terms of its inter-temporal distribution, but potentially also in terms of quantity) and supply the additional evapo-transpirative water demand of crops that are exposed to increased temperatures. In India especially, the high variability and uneven temporal distribution of rainfall within and across years has made irrigation essential for agricultural productivity and food security, and its expansion has for decades been, and still is, one of the central pillars of India’s agricultural development policy (Shah [2008])(figure 1): India is now the world’s largest consumer of groundwater, with over twenty million wells, and the world’s third largest dam builder, with over 4,000 large dams. The expansion of irrigation coverage is therefore an often mentioned means of adaptation to climate variability and change (Howden et al. [2007], Parry [2007], Mendelsohn and Dinar [1999]). For example, crop simulation suggest that the expansion of irrigation can potentially reduce the agricultural losses resulting climate change, especially in developing countries (Rosenzweig and Parry [1994]). However, like other agricultural adaptation strategies, there is little evidence to suggest how e↵ective irrigation is when deployed on a large scale and used by farmers in actual field conditions. Such evidence is especially needed in developing countries, where a large fraction of the population is employed in agriculture, often as small holder subsistence farmers with little access to credit and insurance markets, inputs, extension services, and modern agricultural tech-

5

nologies. This makes the welfare consequences of yield losses more severe and adaptation more difficult, and means that evaluating the potential and large scale e↵ectiveness of agricultural adaptation technologies purely on the basis of their performance in simulations or controlled plot experiments insufficient (for example, in many developing countries, large gaps persist between yields achieved by farmers and those achieved in agricultural research stations). I find evidence that farmers attempt to respond to rainfall deficiencies through both cropping and irrigation decisions, undermining the basis of ‘dumb farmer’ assumption that no adaptation will take place (Patt et al. [2009], Tol et al. [1998]). The potential for the former is limited, since most cropping decisions have to be made before rainfall conditions are known, but irrigation can be applied at any time in the season, and I find that irrigated area is expanded in response to uneven intra-seasonal rainfall distributions. Since the amount of accumulated rainfall a↵ects both the demand for and the supply of irrigation water, irrigation usage can in principle respond to deficiencies in total rainfall in either direction. However, the evidence I find seems to agree with a simple model of irrigation usage that is based on the flat rate pricing of irrigation water that is a defining characteristic of India’s irrigation economy. I also find evidence that irrigated yields are higher (by 20% for rice and 34% for wheat) and significantly less sensitive to uneven intra-seasonal rainfall distributions than their rain-fed counterparts. The estimates suggest that the expansion of irrigation to cover the entire rice cultivated areas of India would reduce the impact of climate change related precipitation shifts by 80% (but this calculation should be interpreted with caution since it involves out of sample predictions). I also calculate, however, that under a ‘business as usual’ scenario, the depletion of India’s aquifers will worsen the projected impact by about 40%. Moreover, I find no evidence that irrigation reduces rice yields’

6

vulnerability to heat shocks, and considering that my estimates suggest that projected temperature increases will be responsible for most of the negative impact on rice yields, this casts further doubt on the e↵ectiveness of irrigation as an agricultural adaptation strategy to climate change.

1.1

Relation to Previous Literature

Estimates of the projection of climate change impacts on agriculture can roughly be divided into the crop model approach and the statistical approach, which can be further divided into Ricardian (cross-sectional) estimates and panel estimates. Crop models employ mathematical simulations of the basic processes of plant growth using detailed inputs of environmental conditions, including weather conditions, usually on the daily scale. By using simulated weather conditions in the future climate, obtained from climate models simulations, crop models can be used to project crop yields in the future climate. A major advantage of the crop model approach is that it can make a joint assessment of crop yields under modified weather conditions as well as elevated CO2 levels. Perhaps the major disadvantage is the reliance of crop models on numerous assumptions, both physiological but also on farming practices, the sensitivity to the input data, which is not always of high quality, and apart from calibration, the disassociation from actual observed yields in field conditions (Lobell and Burke [2009]). A review of crop model based studies of the impact of climate change on agriculture in India is presented in Mall et al. [2006]. In the cross sectional approach, first used to estimate the impacts of climate change on U.S. agriculture by Mendelsohn et al. [1994], and in India by Krishna Kumar et al. [2004], Dinar [1998], Kumar and Parikh [1998], Sanghi and Mendelsohn [2008] and Kumar and Parikh [2001], average land value or farm

7

profits are regressed on long-term climate variables like mean temperature and precipitation to project the impacts of long-term changes in those variables. A potential weakness of this approach is that it may fail to disentangle climate variables from unobserved confounding factors that impact farm profits. This is especially problematic in a developing country like India, where a host of endogenous economic factors are likely to influence crop yields. This difficulty is partially overcome in the panel approach, which uses random year-to-year fluctuations in weather to estimate the impacts on crop yields, and controls for unobserved, time independent spatial fixed e↵ects. The panel approach was first used to estimate agricultural climate change impacts in the U.S. (Deschenes and Greenstone [2007], Schlenker and Roberts [2009]) and has been recently applied in India (Guiteras [2008]). One weakness of the panel approach in comparison to the cross-sectional approach, is that it relies on short term changes, and therefore neglects to take into account adaptation methods that would have been potentially taken by farmers in response to long-term shifts in weather but may not be taken by them in response to short-term fluctuations. This paper follows the panel approach, but makes two principal contributions that help bridge some of the gaps between this approach and the other two approaches. First, most statistical studies of climate change’s impact have only considered seasonal or monthly precipitation totals (for example Krishna Kumar et al. [2004], Au↵hammer et al. [2006] and Guiteras [2008] in India, and Lobell et al. [2011] globally). These models neglect to capture dimensions of the intra-seasonal distribution of daily rainfall that can be important for crop growth - for example, figure 2 shows the time series of daily rainfall in two di↵erent years from the same location in India that have almost identical total rainfall, but a very di↵erent distribution - raising concerns that the statistical

8

models may miss important aspects of precipitation that are simulated in detail by crop models. Indeed, unlike some crop models (for example Nelson [2009]), while statistical models often find a statistically significant positive impact of total rainfall, it’s usually dominated by that of temperature increases. This paper incorporates data on the intra-seasonal distribution of daily rainfall and while the analysis reveals it has an important and large impact on climate change projections (in comparison to a projection that only considers total rainfall), the temperature e↵ect still turns out to be dominant because of the large projected rise in temperatures. Second, this paper explicitly incorporates data on irrigation coverage, a prominent adaptation method, in the analysis in order to test whether it explains the geographical heterogeneity in crops’ response to weather, and weather it is undertaken by farmers as a (short-term) adaptation strategy. Previous papers addressed the issue of irrigation within the cross sectional approach: Schlenker et al. [2005] shows that the response in irrigated and non-irrigated areas in the U.S. is di↵erent and runs the analysis of land values separately in the two sub samples; Kurukulasuriya and Mendelsohn [2007] also finds di↵erences in the response of yields in a cross sectional analysis in Africa (the authors also address the potential bias due to endogenous irrigation use, but find the results are not significantly altered). A notable dynamic result is provided by (Duflo and Pande [2007]), who show that downstream from a dam, the sensitivity of agricultural production to total rainfall is reduced after the dam’s construction. The remainder of the paper is organized as follows. Section 2 describes the data and some broad features of agriculture in India. Section 3 analyzes the overall impact of weather fluctuations on crop yields. Section 4 studies responses in cropping and irrigation decisions to weather fluctuations. Section 5 estimate the degree to which irrigation is e↵ective in reducing the impact of

9

weather fluctuations on yields. Section 6 uses these estimates to project the consequences of climate change for rice and wheat production in a few stylized climate and irrigation scenarios. Section 7 concludes.

2

Data

2.1

Agricultural Data

The Indian agricultural calendar is organized around the monsoon, during which most of India’s rainfall occurs. The two main growing seasons are the rainy season (Kharif, June to September) and the drier season that follows it (Rabi, roughly October to February) 4 . Crops in both seasons rely on the monsoon rainfall to satisfy their water requirements, either directly, in the rainy season, or by drawing on accumulated rainwater, be it as soil moisture or in artificial storage, in the dry season. My analysis is focused on rice and wheat, India’s two dominant crops. Today, rice cultivation covers more than half the cropped area in the rainy season and wheat cultivation covers more than half the cultivated area in the dry season. These two crops are relatively water intensive in comparison with the other crops in their respective seasons (especially rice which is mostly cultivated in flooded plots), and were the main focus of the green revolution, which consisted of the use of high yielding varieties and increased application of irrigation water and fertilizer. As a result, they also have the largest irrigation coverage in comparison to other crops in their respective seasons. Data on the yield, production and area (production divided by yield) of these crops is obtained from the Indian Harvest data set, produced by the Center for the Monitoring of the Indian Economy (CMIE). Observations are reported on 4 In some areas, a third summer crop is also grown, but I have no data on it, and it entirely irrigated

10

the levels of districts (Indian sub-state administrative units, of which there are 586 in the sample) in 18 states for the years 1970-2004. The data also provides figures on irrigated areas.

2.2

Weather Data

I use gridded daily precipitation and temperature figures produced by the Indian Meteorological Department through a re-analysis (with physical climate models) of observational data (Srivastava et al. [2009], Rajeevan et al. [2005]). The gridded data is converted to district level data, in order to match it with agricultural data, by averaging (area weighted in the case of precipitation) over grid points falling within a given district. To capture the impact of temperatures, I use growing degree days, a measure of heat exposure used to predict crop yield in phenology (and shown by Schlenker et al. [2006] to provide a better fit of observed yields), and defined by

DDS =

X

D(Tavg,d )

(1)

d

where Tavg,d is the average daily temperature in day d, the summation is over the days of the growing season and D(T ) reflects the ability of crops to absorb heat in the temperature range from 8 C to 32 C degrees, i.e.

D(T ) =

8 > > > 0, > > > <

T 8, > > > > > > :24,

if T  8 C if 8 C < T  32 C if T

(2)

32 C

Even though nonlinear e↵ects have been detected by using discrete temperature bins in the U.S. context (Schlenker and Roberts [2009]), Guiteras [2008], in his analysis of Indian agriculture and temperature, does not find strong di↵er-

11

ence in climate predictions when using the two methods. I therefore only include seasonal degree days in each of the two seasons in my regressions, depending on the crop. From the daily precipitation figures, I construct ten summary statistics. The first five include total monsoon rainfall and monthly rainfall for each of the monsoon months, namely June, July August and September. In addition, I use measures of the dispersion of rainfall within the monsoon season that are used in the climate science literature to capture increased rainfall variability in climate change simulations. These include the frequency of rainy days (with precipitation over 0.1 mm, May [2004]), the duration of the longest dry spell (Tebaldi et al. [2006]), and the shape parameter of the gamma distribution which is fitted to the distribution of daily rainfall within the rainy days of the season. Gamma distributions are commonly used to describe the distribution of observed and model-generated daily rainfall (Gregory and Mitchell [1995],Stephenson et al. [1999]), and the shape parameter measures the skewness of the distribution: higher values of the shape parameter indicate a more even distribution of rainfall between the season’s rainy days (dry days are not incorporated into the gamma fit). Figure 3 displays mean (1970-2004) annual rainfall and the number of rainy days over India. 5 . These weather parameters are all correlated to some degree. On average, an increase of one wet day tends to increase total rainfall by 10mm, reduce degree days by about 1 degree per day, and increase the shape coefficient by 0.003 (panel regression with fixed districts e↵ects). 5 Since I control for total seasonal rainfall, I do not incorporate rainfall intensity (total rainfall divided by the number of rainy days) or the scale parameter of the fitted gamma distribution (which is related to the rainfall intensity), both of which are also often used in the climate science literature

12

3

Weather Fluctuations and Crop Yields

3.1

Empirical Strategy

My empirical strategy follows the panel approach (Deschenes and Greenstone [2007],Schlenker and Roberts [2009],Guiteras [2008]) to estimate the proportional impact on annual changes in weather patterns on crop (rice and wheat) yields. The basic regression I run is

log Ysdt = v · Wdt + fs (t) +

t

+ pd + ✏sdt

(3)

Here, Ydst is the yield of a given crop in district d, state s and year t; Wdt is a vector of weather variables including total Monsoon rainfall, seasonal degree days (DDS), calculated in either the dry or rainy season, depending on the crop in question, and variables describing the intra-seasonal distribution of daily rainfall, described in section 2.2: the number of rainy days, the duration of the longest dry spell, and the shape parameter of the fitted gamma distribution. The estimated coefficients vector v describes the vulnerabilities of crop yields to these weather variables. The regressions also include controls for unobservable, time invariant district attributes pd that may a↵ect yields, such as soil quality or other geographical attributes; Quadratic time trends fs (t) that reflect technological progress and productivity gains, which I allow to di↵er from state to state because of the large variance in agricultural performance across India; and aggregate unobservable year e↵ects

t,

to separate annual country-wide factors,

such as climatic fluctuations (e.g. ENSO) from changes in the weather patterns that I include in the regressions. Because of potential spatial and serial correlation in both weather outcomes and yields, I allow the unobserved errors ✏sdt to be correlated across years and districts in the same state. I will mostly report regression results estimated with the use of standardized

13

weather variables 6 , which will facilitate an easier interpretation and comparison of the estimates across crops. I also run the regressions using absolute values of these variable, and the pattern of the results is unchanged.

3.2 3.2.1

Results Regression Estimates

Regression estimates are reported in table 1. Consistently with other studies, rainy season rice yield is negatively a↵ected by declines in total rainfall (6.3% decline per standard deviation), but with a declining marginal damage reflected by the negative coefficient on the square precipitation term (the average impact is estimated at 4%), a familiar result related to saturation of water requirements and the damage from extremely high rainfall. Yields are also negatively impacted by a increased heat (3.2% per standard deviation increase in rainy season degree days), with rising marginal damages reflected by the negative coefficient on the square degree days term. The inter-temporal distribution of precipitation within the monsoon also turns out to have significant impacts. The largest of which is the impact of the number of rainy days during the monsoon, which tends to decrease rice yields by 3.4% per standard deviation decrease in the number of rainy days. The duration of the longest dry spell has a negative, but small and statistically insignificant e↵ect once the number of rainy days is controlled for. The shape parameter of the fitted gamma distribution has a statistically significant impact on rice yields, indicating that a more even distribution of rain within the rainy days improves yields, but the impact is small in relation to that of the number 6 For

a variable Xdt that varies with district d and year t, the standardized form I use is Xdt

µd

(4)

d

where µd is the district specific mean (taken over time) and deviation.

14

d

is the district specific standard

of rainy days. To put these numbers in perspective, the average annual growth rate of rice yields over the entire period was about 2% per year. However, this growth has slowed down considerably to about 0.5% past 1990, and there are concerns over stagnating yields past 2000. Moreover, as we shall see later, the estimated impact of rainfall deficiencies hides a great deal of variation related to irrigation coverage - the impact will be about double for rain fed crops. Similarly to rice, dry season wheat yield is negatively impacted by total monsoon rainfall (3.9% per standard deviation), again with a declining marginal damage. Wheat yields display a sizable response to dry season heat (4.2% decline per standard deviation increase in dry season degree days), but the e↵ect is not statistically significant. We shall see later that when irrigation is controlled for, the impact becomes significant - wheat is a highly irrigated crop, and irrigation seems to be mitigating some of the impact of heat. Unlike rainy season rice yields, wheat yields show low sensitivity to the distribution of rainfall within the rainy season - in particular, there is no apparent impact of the number of rainy days, and as for rice, only weak impact of the shape parameter of the fitted gamma distribution. The di↵erence in the relative impact of the intra-seasonal rainfall distribution and its seasonal total during the monsoon on rainy and dry season crops is easy to explain. Rainy season crops experience the flow of irrigation directly, and are therefore sensitive to its distribution within the season. Dry season crops, in contrast, rely on the accumulation of rainy season precipitation in either natural (as soil moisture or groundwater) or in or artificial (reservoirs, tanks) storage. The efficiency with which rainfall is captured in storage is likely to be higher when the rainfall is more evenly distributed, but this e↵ect is probably secondary to the direct e↵ect of total rainfall amount.

15

3.2.2

Non Parametric Fits

Figure (4) presents non parametric plots (kernel regressions) for the impact of total rainfall, degree days and the frequency of rainy days on the yields of rainy season rice and dry season wheat. The plot shows clearly the lack of impact of rainy day frequency in the dry season, and the familiar concave response to total rainfall in both seasons. 3.2.3

Robustness Checks

Tables 2 and 3 display the results of sensitivity checks for rainy season rice and dry season wheat, respectively. Column 1 reports regression estimates with only the traditional measures of degree days and total rainfall and their squares. Column 2 repeats the basic regression. In column 3 I remove year fixed e↵ects. In column 4 I control for the previous year’s yield. Because this can introduce a bias in panel estimation when errors are serially correlated, I also use the Arellano-Bond estimator (Arellano and Bond [1991]) in column 5. In Column 6 I replace district fixed e↵ects with a smaller set of geographical fixed e↵ects (districts have split up over time, and in the basic regressions I include a separate fixed e↵ect for a district before and after it splits). In Column 7 I run the regression on a sub sample of that of column 6 in which each location has at least 25 years of observation, in order to address the concern that missing observations introduce a bias in the estimation. For wheat, the regression reported in column 8 includes the (log) rice yield of the preceding rainy season to ensure it is not the only channel through which rainy season rainfall impacts wheat yields. The estimates are mostly robust to these di↵erent specifications. In addition, I also re-estimate the basic regressions while omitting each of the states from the sample (results not shown). The coefficient estimates remain stable and significant.

16

4

Land and Irrigation Use Responses to Weather Fluctuations

If weather shocks a↵ect crop yields, it is natural to examine weather farmers are able to respond to these shocks in ways that can reduce their impacts. Two principal forms of response, or of short-term adaptation, are changes in cropping decisions, and as discussed in the introduction, the application of irrigation to cultivated land. While basing conclusions for long-term adaptation on shortterm responses can be incomplete (I will return to this issue below), observing these responses can provide valuable information on the degree to which farmers are aware of and are able to undertake them, and indirectly, the e↵ectiveness of these responses.

4.1 4.1.1

Land use decisions Conceptual Framework

I will begin by presenting a highly stylized, simple model of land use decisions, and build upon it further when discussing irrigation choices. Consider a vector of weather variables W that influences the supply of moisture to a crop (total rainfall, rainy day frequency or the negative of temperature related evapo-transpirative demand), so that the yield (per unit area), in the absence of irrigation, can be described as a rising, concave function of W:

Yf = pf y(W)

(5)

where p is a basic farm total factor productivity coefficient (e.g. soil, capital or technology), which I assume will vary across farmers: I will order farmers by a continuos parameter 0 < f < 1 that orders them according to their basic productivity factor pf . 17

I will assume that the cost of all inputs, other than irrigation, is CIN per unit area, so that the net profit per unit of (unirrigated) area, given a weather realization W, is

Pf (W) = pf y(W)

CIN

For simplicity, all other inputs have been suppressed

(6) 7

I will assume, for simplicity, that farmers are risk-neutral. Since farmers need to make cropping decisions before having full knowledge of stochastic weather conditions during the full growing season, they choose to crop a plot if

EWU Pf (WN , WU )

0

(8)

where I have separated the weather variables into those that are known to the farmer at the type of cropping WN and those that are unknown WU . For example, considering, as I have above, the three weather variables of total rainfall, rainfall distribution (e.g. rainy days), and growing degree days, then in the case of the rainy season, none of these are known at the time of cropping (except early season rainfall in June-July), whereas in the case of the dry season, both the total and the distribution of rainfall during the preceding rainy season are known, but of the two, only total rainfall is relevant for yield, as we have seen in the previous section. Expected profit is therefore a function of those weather variables known at the beginning of the season. If expected yield rises in one of these known weather variables, one would expect the threshold level of productivity pf at which a crop is cultivated to move down in response to a higher realization of 7 One

could start with an additively separable profit function Pf = pf y(W) + g(z)

cIN z

(7)

where z is a vector of other inputs, and maximize over input use to obtain the production function 6 with CIN = maxz g(z) cIN z.

18

this weather variable. In these cases only, if farmers take weather information into account in their cropping decisions, one would expect the cropped area to respond positively to positive fluctuations in these variables. The only case in which this is relevant is for the impact of total rainy season rainfall on the area cropped with wheat in the dry season. One would not expect to see such a response for the other variables. 4.1.2

Empirical Results

To test these predictions, I estimate a regression of a similar form to the yield regression (3), i.e.

log Asdt = v · Wdt + fs (t) +

t

+ ⇡d + ✏sdt

(9)

The results are displayed in table 4. Column 1 reports regression estimates for rainy season rice cropped areas. There is no statistical significance to any of the weather variables. Column 3 reports regression estimates for dry season wheat. As expected from the model, total rainfall accumulation in the rainy season has a statistically significant positive impact on wheat cropped area (about 5% per standard deviation increase in total rainfall). Column 5 repeats the same regression, controlling for rainy season rice yields in order to ensure the impact of total rainfall is not entirely driven by higher rainy season incomes. In columns 2 (for rice) and 4 (for wheat) I repeat the regressions but separate total rainy season precipitation to four monthly precipitation totals. This decomposition reveals a significant response of rice cropped areas to early season (July) rainfall, and that the response of wheat cropped area to total monsoon rainfall is largely driven by late (September) monsoon rainfall - these results are consistent with the timing of cropping decisions for rainy season rice and dry season wheat. The higher response to September rainfall can be a result

19

of both natural (soil moisture during the dry season is more highly dependent on September precipitation) and subjective factors (farmers may have a better recollection of late rainy season rainfall when they make their cropping decisions for the dry season).

4.2

Usage of Irrigation

The high variability and uneven temporal distribution of rainfall within and across years has made irrigation essential for agricultural productivity and food security in India, and its expansion has been and still is one of the central pillars of India’s rural development and agricultural growth policies for many decades (Shah [2008])(figure 1). About half of India’s cultivated area is now irrigated, and the government is continuing to make large investments in public infrastructure and subsidized electricity supply to continue expanding irrigation coverage. India is now the world’s largest consumer of groundwater, with over twenty million wells, relaying, by some estimates, on more than 20% of the country’s entire electricity consumption to pump water, and the world’s third largest dam builder, with over 4,000 large dams. While public investment determines the potentially irrigated area, actually irrigated area often falls short of this technical potential. For example, according to the 3rd Minor Irrigation Census 8 , only 60%-70% of the irrigation potential was utilized at 2000-1, with, for example, 7% of tube wells temporarily out of use that year, and with 65% of the remaining ones being under-utilized due to a variety of reasons like inadequate power supply or insufficient water yield. Short-term supply and demand factors seem to have a sizable impact on actual irrigation usage. Weather patterns impact both the supply and demand, with high rainfall amounts increasing the water availability in the existing infrastructure (e.g. water tables in wells, water flow in canals), and favorable rainfall and 8 ‘‘http://mowr.gov.in/micensus/mi3census/nt_level.htm’’

20

potentially temperature potentially reducing demand for irrigation. 4.2.1

Conceptual Framework

I now extend the simple model of the previous section to include irrigation choices and the impact that weather conditions can have on them. The two prominent forms of irrigation in India are surface irrigation and groundwater irrigation (which has steadily grown since the 1970s to displace surface irrigation as the largest share of irrigated area). A unique feature of both types of irrigation in India is that the dominant cost of irrigation is flat, i.e. independent of the amount of water used. In the case of surface (canal) irrigation, users are normally charged a certain flat price per unit of land, per season, and there is no volumetric charge. Groundwater is not directly priced or regulated. The cost of access is the cost of drilling a well, and the only running costs are the costs of energy for pumping, which can come from either electricity or diesel. Almost everywhere in India where groundwater use has boomed (the Northwest and much of the interior of peninsular India), it has done so by using highly subsidized, publicly provided electricity that is priced at a flat (if any) tari↵ (Shah [2008]), whereas diesel users (mostly located in eastern India) pay much higher and sticky marginal costs, (Shah et al. [2009]) and therefore irrigate substantially less area. Not all irrigation water is priced at a flat tari↵: water markets are prevalent in many parts of India (Saleth [1998]). Still, access to these markets is often distorted by various social and other factors (Anderson [2011]) and water prices are often not a balance of supply and demand (Dubash [2002]), partly because of the difficulty of transporting water over substantial distances and the resulting localized monopolies. Assuming a flat rate costing structure for irrigation, per unit land, is therefore a likely reasonable approximation and this is the modeling approach I take here.

21

Let non-irrigated profits be described, as above, by the profit function

Pf,N I = pf y(W )

CIN

(10)

where N I stands for non-irrigated. Assume that a fixed cost CIR , per season, per unit area, provides an amount of water ! per unit area which I assume to be a perfect substitute to moisture availability from precipitation W . I will also assume that the cost of other inputs is identical to the unirrigated case 9 . In other words, the irrigated yield is

Yf,IR = pf y(W + !)

(11)

and the net profit per unit of irrigated area is

Pf,IR = pf y(W + !)

CIN

CIR

(12)

Unlike cropping decisions, irrigation can be applied at any time of the growing season. Farmers will decide to irrigate a plot if

pf (y(W + !)

y(W ))

CIR > 0

(13)

I will assume that the irrigation entry cost CIR is independent of the current weather conditions. As discussed above, this is probably a reasonable first approximation in India, where most costing is flat and insensitive to actual water usage. For this reason, it is reasonable to assume irrigating farmers will utilize the entire amount of available water. However, the amount of water available per unit area to those farmers who choose to irrigate can depend on 9 this follows by using the additively separable production function mentioned in the footnote to equation (6)

22

weather conditions. For example, in the case of groundwater irrigation, the amount of water available to farmers with access to wells and a certain energy supply depends on on the water table, which in turn depends on the rainy season’s rainfall (this is shown by Fishman et al. [2011] using rainfall and water table data from two regions of India). Similarly, the amount of water that flows in irrigation canals depends on reservoir levels, which in turn determined by rainfall capture. While the cost of irrigation is independent of weather, the supply of irrigation water ! need not be. Specifically, higher rainfall totals impact the water table and water yields in irrigation wells, and also water levels in reservoirs and therefore the water flow in irrigation canals. In contrast, it would seem plausible that the availability of water is less dependent, to first order, on the intra-seasonal inter-temporal distribution of rainfall than it on total amount. In response to a positive change in weather W , the threshold productivity level justifying paying for irrigation will increase, so irrigated area will decrease, if d(y(W + !) dW

y(W ))

= y 0 (W + !)(1 + ! 0 (W ))

y 0 (W ) > 0

(14)

When W rises, the concavity of the yield function reduces the di↵erence between irrigated and non-irrigated yields, therefore driving the demand for irrigation downwards. However, a rise in ! increases the supply of irrigation, for the same fixed cost. In the case where W describes dimensions of the rainfall distribution that impact crop yields (the number of rainy days), then since one would expect, to first order, only a weak impact on supply, i.e. ! 0 (W ) ⇡ 0, the demand e↵ect should dominate, and irrigated area should decline in response to an increase in the number of rainy days, to the degree that farmers are able to respond.

23

In the case where W describes total rainfall, then since ! 0 (W ) > 0, the inequality can go either way and irrigated area can either decrease or increase in response to higher total rainfall. But one might expect the supply e↵ect to be stronger in the dry season than in the rainy season, when the supply is at least partially supplied by rainfall accumulation from previous years and potentially distant locations, and the supply from in-season rainfall is only gradually built up. 4.2.2

Empirical Results

Table 5 reports estimates of the regression

log Asdt = v · Wdt + fs (t) +

t

+ ⇡d + ✏sdt

(15)

Column 1 reports regression estimates for rainy season rice irrigated area. As expected, the ‘demand’ e↵ect seems to drive a negative dependency of irrigated area on rainy season rainfall distribution, as well as totals. In Column 2 I decompose total seasonal precipitation on a monthly basis, and find that the negative response is concentrated at the later stages of the season (AugustSeptember), while higher early season (June) rainfall has a positive, but only marginally significant e↵ect. This is consistent with the model presented above: early season rainfall contributes to the accumulation of irrigation water, which mostly takes place late in the season when the monsoon is less reliable and crops are sensitive to water availability. Columns 3-5 reports regression estimates for dry season wheat. They show a significant and large impact of total rainfall on irrigation for wheat, which is equally distributed throughout the rainy season, consistently with the ‘supply e↵ect’. In column 5 the regression controls for rainy season rice yields in order to ensure the impact of total rainfall is not entirely driven by higher rainy season

24

incomes.

5

The Impact of Irrigation on Crop Yields

Despite the importance of the expansion of irrigation to boost and protect agricultural production from rainfall irregularities, there is a shortage of evidence on its actual performance. A notable exception is Duflo and Pande [2007], who show that downstream from a dam, the sensitivity of agricultural production to total rainfall is reduced after the dam’s construction. Since irrigation is often accompanied by the use of di↵erent crop varieties, farming practices, and potentially other forms of risk preserving behavior, because it relies on water resources that are themselves dependent on rainfall, and because of pervasive inefficiencies in water management and distribution, its performance in the field and in controlled experiment may di↵er, and scholars actually debate its e↵ectiveness in stabilizing agricultural growth in India (Moench [1992], Hanumantha Rao et al. [1988])

10

.

In this section, I examine evidence for the impact of irrigation on crop yields and their sensitivity to weather conditions, including seasonal degree days, total rainfall and the frequency of rainy days. The ability of irrigation to mitigate heat shocks is not easy to predict but moisture availability can not be expected to completely o↵set the e↵ect of excessive heat. I will return to this issue in more detail in section 5.3. When it comes to precipitation, simple agronomic consideration would sug10 For

example, Moench (Moench [1992]) echoes the common notion that: “... In addition to serving as a regular source of supply, ground water plays a critical bu↵ering role during periods of drought when surface flows are limited and unreliable.” In contrast, Rao (Hanumantha Rao et al. [1988]), in one of the few existing works on the issue, uses state level crop statistics, and argues that “The instability in agricultural production has increased on account of rise in the sensitivity of output to variations in rainfall [due to the] high complementarity of new seed-fertilizer technology with wateralthough area under irrigation has increasedthis irrigation is itself dependent on rainfallthe uncertainty of irrigation has been increasing in the recent period”

25

gest irrigated yields to be less sensitive the intra-seasonal distribution of rainfall. Keeping total rainfall fixed, storing rainwater and using it for irrigation should enable farmers to smooth the supply of moisture to crops, at least to first order (i.e. ignoring very long dry spells or other impacts of the temporal arrangement of the rainless days within the season). The application of augmentative irrigation from additional sources (e.g. capturing rainfall runo↵ or even mining groundwater accumulated from other years and other geographical locations) should also be able to reduce the dependency of crop yields on local contemporaneous (same year) total precipitation, but this capacity is probably more limited. However, the amount of irrigation water can itself depend on precipitation, so the relative sensitivity of irrigated yields, as actually managed in farmers’ fields, can be more complicated. Recall that I model irrigated and non-irrigated yields as a rising concave function (justified by the results of section 3) of weather related moisture availability W (determined by total rainfall and number of rainy days for rainy season rice, and by total rainfall for dry season wheat) and irrigation water !:

Yf,N I = pf y(W )

(16)

Yf,IR = pf y(W + !)

(17)

First, because y(W ) is rising, irrigated yields are expected to be higher, on average, than their non-irrigated counterparts. Second, as discussed in section 4.2, since the amount of water applied per unit irrigated area, !, can depend on W itself, irrigated yields need not necessarily be less sensitive to weather. The sensitivity of yields to weather can be measured through the derivative of the yield function with respect to W , so irrigated 26

yields are less sensitive, in the sense that dYf,IR dYf,N I < dW dW

(18)

d(y(W + !) = y 0 (W + !)(1 + ! 0 (W )) < y 0 (W ) dW

(19)

if

a condition which is identical to the condition for irrigated area to respond negatively to positive changes in W , equation (14), and the same observations apply here: in the case where W describes dimensions of the rainfall distribution that impact crop yields (the number of rainy days), then since one would expect, to first order, only a weak impact on supply, i.e. ! 0 (W ) ⇡ 0, irrigated yields should be less sensitive to changes in the number of rainy days, keeping total rainfall fixed. In the case where W describes total rainfall, then since ! 0 (W ) > 0, the inequality can go either way and irrigated yields can be either more or less sensitive to total rainfall, with the sensitivity ‘gap’ declining the stronger the ‘supply e↵ect’ becomes. Again, one might expect the ‘supply’ e↵ect to be stronger in the dry season than in the rainy season.

5.1

Empirical Strategy

Unfortunately, the CMIE dataset does not report irrigated and unirrigated yields separately, nor is there any data available on the amount of water applied per unit area. The CMIE data does report the gross (annual total) area irrigated per crop in a given district in a given year, where irrigation is defined as the application of water to the field at least once during the growing season. I will therefore follow an indirect approach and examine weather higher irrigation coverage implies a lowered sensitivity of aggregate yield to weather variables. To measure the extent of irrigation coverage, I define the fraction of area

27

irrigated as

F IAsdt =

GIAsdt GAsdt

(20)

where GIA is gross irrigated area (annual totals of both seasons) and GA is gross cropped area (annual totals of both seasons).

11

I will test the hypothesis that higher irrigation coverage reduces the sensitivity to weather shocks in two di↵erent ways. The first approach is to estimate a model with interaction terms

log Ysdt = v0 · Wdt + fs (t) + in which the coefficient

t

+ ⇡d + F IAsdt +

· Wdt + ✏sdt

(21)

estimates the (proportional) increase in irrigated yields

in response to an increase in the fraction of irrigated areas (i.e. the proportional increase in irrigated yields in comparison to un-irrigated yields), and the three coefficients in the vector

estimate the change in the sensitivity of crop yields to

the three weather variables in response to an increase in the fraction of irrigated areas. A negative value of a particular component of

will suggest that yields in

more irrigated districts are less sensitive to the e↵ects of that particular weather variable. To separate the impact of irrigation coverage from that of unobservable state specific attributes, at the level of which most water and energy policies are determined, for example, I will estimate the regression above while also including terms ⌫ s · Wdt , where ⌫ s are un-observable state dummies. To separate the 11 For those crops, like what, that are exclusively cultivated in only one of the seasons, this measure provides the season specific, and not just annual, ration of area that is irrigated. For those crops that are grown in both seasons, like rice, the fraction of gross area irrigated provides a measure of irrigation coverage in specific season only under the assumption that the fraction of cropped area that is irrigated is equal in both seasons. Because this is a strong assumption, I will run the regressions below on a subsample of districts in which these crops are only grown in one of the seasons (the rainy season).

28

impact of irrigation from technological progress, possibly state dependent, that reduces crop sensitivity to weather shocks, I will also re-estimate the regression while including terms µs · Wdt t. The second approach will be to first estimate the basic yield model separately in each district, i.e. regress log Ysdt = vd · Wdt + ⌧d t + ⇠d t2 + ⇡d + ✏sdt

(22)

separately within the time series of each district in the data, and obtain district specific estimates vd (a vector that includes coefficients related to total rainfall, the number of rainy days, and seasonal degree days), and then regress them, between districts, against the mean irrigation ratio in that district F IAd (averaged across time): vd = v0 + F IAd + ⇠ d Again, a negative value of a particular component of

(23) will suggest that yields in

more irrigated districts are less sensitive to the e↵ects of that particular weather variable. I also try to separate the impact of growth in irrigation over time (within districts), by by dividing the sample into two time periods, before and after 1985, an approximate watershed period for irrigation coverage in many parts of India, and then to estimate the regression (22) separately, in each district, in each of the two time periods. This estimation produces district and period specific coefficients !d,i where i = 1, 2 and period 1 (2) refers to the years before (after) 1985. Similarly, one can also average the fraction of irrigated area within each district and each period to obtain F IAd,i . One can then run the regression:

vd,i = vd,0 + F IAd,i + ⇠ d,i

29

(24)

which control for unobservable, period independent district attributes vd,0 . The estimates of the components of

will then only reflect the impact of changes

in irrigation coverage within districts, over time, on the sensitivities of yields to weather variables, and allow us to rule out the possibility that any other time independent district property is driving the result.

5.2 5.2.1

Results Interacted Model

Estimates of the interacted model, equation (21) are reported in tables 6 (for rice) and 7 (for wheat). Column 1 repeats the estimation of the basic yield model, equation (3). In the case of rice, the estimation is restricted to those locations in which rice is not grown in the dry season (because only in those districts can gross irrigated area be attributed to one of the seasons). In column 2 the estimation is restricted to a sub-sample in which data on irrigated areas is available. Column 3 estimates the basic interacted model, equation (21). In column 4, state dummies are allowed to interact with weather variables to separate the impact of irrigation from that of unobservable state attributes. In column 5, state specific time trends are also allowed to interact with weather variables to further separate state specific trends in the ability to mitigate the impacts of weather shocks from the impact of irrigation. Since the impact of the rainfall variables is on average positive, a negative estimated coefficient for the interaction terms of F IA with both total rainfall and the number of rainy days indicates that irrigated yields are less sensitive to these weather variables than their non-irrigated counterparts. In the case of rainy season rice, both of the coefficients are statistically significantly negative. In the case of wheat, which is only sensitive to total rainfall, the interaction

30

term of total rainfall with F IA is also negative. Taken at face value, the estimates suggest that as irrigation coverage moves from none (F IA = 0) to full (F IA = 1), the response of rice yields to a one standard deviation drop in the number of rainy days falls from 8% reduction in yield to none; the response to a one standard deviation drop in total rainfall drops from 11% to 3%; and the response of wheat yields to a one standard deviation drop in total rainfall drops from 7% to 3%. However, only the mitigating e↵ect of irrigation on the number of rainy days can be separated from state dependent (with time trends) unobservable properties that might be driving the result: for both rice and wheat, the interaction coefficient estimate drops in size and significance in columns 2 and 3. In contrast to its mitigating e↵ect on the impact of rainfall shocks on rice yields, I find no evidence for a similar e↵ect of irrigation on rainy season heat shocks - the interaction coefficient is small and insignificant. In the case of wheat, however, irrigation does seem to be e↵ective as indicated by the positive sign on the coefficient of the interaction between degree days and irrigation coverage (even though the estimate loses statistical significance in columns 2 and 3, the magnitude of the coefficient does not change much. In Columns 6 and 7 I address the concern that correlations between irrigated area and weather fluctuations could be driving the results. In Column 6 I first regress F IA on the weather variables, and then plug the residual from the regression into (21) instead of the original F IA. In column 7 I replace F IA by its district mean (taken over time). Both measures should remove any correlations between the irrigation coverage and the same year’s weather anomalies. Columns 8-13 reports sensitivity checks that are similar to those performed for rice and wheat yields. In column 8 I remove year e↵ects. In column 9 I control for the previous year’s yield. Because this can introduce a bias in panel

31

estimation when errors are serially correlated, I also use the Arellano-Bond (Arellano and Bond [1991]) estimator in column 10. In Column 11 I replace district fixed e↵ects with a smaller set of geographical fixed e↵ects (districts have split up over time, and in the basic regressions I include a separate fixed e↵ect for a district before and after it splits). In Column 12 I run the regression on a sub sample of that of column 11 in which each location has at least 25 years of observation, in order to address the concern that missing observations introduce a bias in the estimation. For wheat, column 13 also controls for the rice yield in the preceding rainy season. The mitigating e↵ect of irrigation on the impact of rainfall totals and distribution on rice yields, and on the impact of rainfall totals and increased heat on wheat yield are quite robust to these alternative models. The coefficient estimates for the un-interacted variable F IA are robust across the models, positive, and highly significant. The estimates suggest that irrigated rice and wheat yields are, on average (in average weather conditions) about 30% higher than their non-irrigated counterparts. 5.2.2

Comparisons Between and Within Districts

Table 8 present estimates of the relationship between district specific sensitivity of yields to weather and the extent of irrigation coverage in that district. Columns 1-3 are devoted to rainy season rice yields, and columns 4-6 to dry season wheat. Each row describes results of a regression of the sensitivity of yields to a particular weather variable on the fraction of area irrigated in that district. Columns 1 and 4 present estimates of the basic cross-sectional model in equation (23). Columns 2 and 5 include state e↵ects in the cross sectional regression. Columns 3 and 6 present estimates of the model (24 which attempts to separate changes within districts from di↵erences between districts. Negative estimates in the case of the two rainfall variables and positive 32

estimates in the case of degree days indicate a mitigating e↵ect of irrigation for the impact of these weather variables. Only the impact of irrigation on the sensitivity to the number of rainy days is robust and significant across all three models. The mitigating impact on total rainfall is significant for both rainy season rice and dry season wheat yields in the basic cross section (these are plotted in figure 5), but diminishes in size and statistical significance when state e↵ects are included (wheat) or when estimated within time (rice). The positive estimate for dry season degree days is significant and positive in the simple cross section as well as within districts, indicating evidence for the mitigating impact of irrigation.

5.3

Discussion

Attributing the reduction in the sensitivity of rice and wheat yields to weather fluctuations to an expansion of irrigation can be erroneous if irrigation coverage (and the host of farming practices that accompany it) is correlated with other, unobserved factors that are actually driving this reduction. In the absence of an exogenous variation in irrigation cover, I am unable to completely overrule this possibility, but several lines of evidence reduce its likelihood. First, I provide evidence to show that the e↵ect of irrigation on crop vulnerability to rainfall variability and dry season heat exposure is independent of arbitrary unobserved factors that are well described by state dependent time trends. Any unobservable attribute that might be driving the reduced sensitivity to weather would therefore have to be correlated even after state dependent time trends are removed from the two variables. Second, I have also provided evidence that irrigation usage is expanded in response to uneven rainfall distributions (the number of rainy days) suggesting it is e↵ective in addressing such weather shocks.

33

Third, the di↵erent impacts I find irrigation has on the vulnerability of rainy season rice and dry season wheat to the di↵erent types of weather shocks are in broad qualitative agreement with phenological predictions, further restricting the range of ‘candidates’ for confounding correlates of irrigation that may be driving the results. As discussed above, irrigation should be especially e↵ective in overcoming gaps in rainfall, keeping total rainfall constant. Irrigation should also be expected to be e↵ective in addressing deficiencies in total rainfall, but this capacity is more limited and depends on the type and origin of the source of irrigation water. Things are more complicated when it comes to heat shocks, since moisture availability is not a perfect substitute for temperature. The impacts of high temperatures on crop growth occur through a number of channels and are not limited to increased evapo-transpirative water demands. For example, accumulated heat exposure a↵ects the length of a crop’s growing period and therefore its yield, irrespective of water availability. For this reason, it is unlikely that even unlimited water availability would mitigate all the e↵ects increased heat exposure can have on crop yields. Moreover, even when soil moisture is not limiting, a plant’s capacity to cool itself through evapo-transpiration depends on a variety of factors, including vapor pressure deficit, the atmospheric demand for water from a leaf. This pressure deficit is lower in humid conditions, when rainy season rice is cultivated, suggesting that even ample irrigation might be of low e↵ectiveness in protecting rice crops from high temperatures. Furthermore, and somewhat paradoxically, the flooded conditions in which rice is cultivated in India (paddy rice) also restricts rice crops’ ability to evapo-transpire. For these reasons, there is little basis to expect irrigation to be e↵ective in mitigating the impact of high heat exposure on rainy season rice crops. In dryer conditions, such as during the wheat season, the vapor pressure deficit is potentially

34

higher, making moisture availability a more limiting factor in plant’s ability to withstand high temperatures, and suggesting irrigation can help wheat crops manage some of the adverse impacts of high temperatures. Observations of yields in controlled conditions in research stations in Asia also show that irrigated rice yields are significantly a↵ected by temperatures (Peng et al. [2004], Welch et al. [2010]), whereas Guoju et al. [2005] shows that irrigating spring wheat in arid regions of China can o↵set some of the impacts of increased temperatures, in broad agreement with my findings. Crop model simulations also indicate potentially severe impacts on irrigated yield drops. For example, Nelson [2009] simulate a drop of about 15% in irrigated rice yields (but hardly any reduction in rainfed rice yields because of the compensating impact of increased rainfall) and 40%-50% drop in both irrigated and rainfed wheat yields in South Asia. Nevertheless, using the above estimates to project the impact of irrigation expansion to areas which are currently rain fed, as I will do in the following section, should be performed with caution. This is especially true if irrigation coverage is endogenously determined. For example, if current irrigation coverage is high at those places in which it is the most e↵ective in mitigating weather shocks, then the impact of expanding it to additional areas will be lower than that projected on the basis of my estimates. On the other hand, if investment in irrigation is higher at those regions where yields, and hence income, is more stable to begin with, then my estimates of the mitigating e↵ect of irrigation may be conservative.

6

Simple Climate Change Simulations

In this section, I apply the empirical results to estimate some stylized implications of climate change to rice yields in India. For illustration, I use the median 35

projections of the IPCC’s scenario A1B for the period 2080-2099, which include a precipitation increase over South Asia of about 11%, and an increase in mean daily temperature of about 2.7 C during the rainy season and about 3.5 C during the dry season. While changes in the frequency of rainless days are not reported by the IPCC, several studies cited therein report increases in this frequency simultaneously with the increase in total rainfall, including, May [2004] and Krishna Kumar et al. [2003], who report a decrease of up to 15 days in the number of rainy day across much of India in the 2050s. I therefore add an absolute uniform decline of 15 rainy days during the monsoon to the A1B projection. I assume total precipitation rises by an average 1mm per day, roughly 10% of the 1980-2000 mean. I restrict the estimation to the impacts on yields, assuming cropped areas remain identical to their 1980-2000 levels. I first re-estimate the statistical model for crop yields with the use of the absolute (vs. standardized) weather variables, and then combine the estimates with the climate scenario described above. To first order, the over all impact can be decomposed in to the three components related to changes in total rainfall, the number of rainy days, and growing degree days. Results for India averaged rainy season rice yields are summarized in figure 6. The top panel displays results from an estimation of weather impacts on yields. The impact of the increase in total rainfall is modest: when rainy day frequency is not taken into account in the regressions, it amount to a increase of about 2% in yield. When rainy day frequency is included, since the coefficient on total rainfall decreases, the yield increase declines to just a 1% increase in yields. However, the amount of the decrease in the number of rainy days is negative and much larger, at a 9% loss. Thus, taking rainy days frequency into account in the climate change projection flips the sign of the projected precipitation

36

related impact of climate change on rice yields in India from a modest gain to a 8% decline. The bottom panel of figure 6 shows how the projected impact of the decrease in rainy days changes in various irrigation scenarios. Irrigation scenarios give the share of rice cultivated area which is irrigated in each district, and I use estimates of an interacted model (like the model in equation (21)) to calculate the resulting climate change impact on the yield (only from the decrease in the number of rainy days) in each of these districts. Since to first order, the relative impact in di↵erent irrigation scenarios is independent of the actual climate scenario chosen, I display the size of the negative impact in each irrigation scenario in relation to the impact in a scenario in which irrigation coverage remains identical to current coverage. For example, the simple estimate, in which irrigation is not taken into account at all, gives an estimate which is 35% too high in relation to the impact calculated on the basis of current irrigation coverage. This is because districts which tend to be more irrigated also have a high relative level of rice production, and the impact in these districts will be lower. If irrigation were to be expanded to cover all of the rice cultivated area in India, the calculation shows that the size of the impact will be reduced by 80% in comparison to a maintainable of current irrigation coverage. Instead, if irrigation were to cease altogether, the impact would grow by 139%. None of these scenarios is particularly realistic. In particular, a full expansion of irrigated area is unlikely to be feasible in the long run with present irrigation practices. Already, India’s water resources are under a great deal of stress. In particular, the country’s groundwater aquifers, on which about half of the irrigated area depends, are being over-exploited in many parts of the country and their water tables are falling (Shah [2008],Fishman et al. [2011]). I therefore

37

also simulated the impacts in two additional scenarios: a ‘business as usual’ scenario, in which aquifers are exhausted in 2080 in those parts of India in which groundwater is currently over-exploited and irrigation from groundwater ceases (a depletion scenario); and a sustainable scenario, in which groundwater extraction shrinks to sustainable levels from now on and groundwater irrigated area shrinks proportionally. To simulate these scenarios, I used data on the share of area irrigated by groundwater at each district 12 and the extent of overexploitation of groundwater

13

(figure 7). I find that in a scenario of depletion,

the climate change impact is increased by 38%, but a transition to sustainable extraction would only result in yield losses that are 12% worst than would take place if current irrigation coverage were somehow maintained in the future. While these calculations are simplified, they shed additional light on the cost benefit analysis associated with India’s management of its water resources. But even if irrigation could be expanded substantially, the lack of evidence for its ability to attenuate the impact of increasing temperatures sheds further doubt on its e↵ectiveness as an adaptation strategy. As figure 6 shows, the simulated impact of increasing temperatures dominates the precipitation e↵ect. I estimate this impact at almost 40% yield losses, consistently with Guiteras [2008], even though, as he point out, the prediction needs to be interpreted with caution because the associated rise in temperature is much larger than the temperature fluctuations with which the crop response model was estimated. Indeed, it is because the projected temperature increase in larger than typical intra-annual fluctuations in temperatures that the impact is so large, and not because crop yields are more sensitive to typical temperature fluctuations. It is also worth mentioning that an important draw back of the statistical approach for simulating climate change impacts is in its inability to take into 12 ‘‘http://mowr.gov.in/micensus/mi3census/nt_level.htm’’ 13 url“http://cgwb.gov.in/”

38

account the potentially positive impacts of increased Carbon Dioxide concentrations on crop growth.

7

Conclusion

The analysis of this paper shows that intra-annual variability in rainfall can be important for crop yields, and taking it into account it can flip the sign of the projected climate change impact on crop production. This is a reminder that even though the panel approach can overcome biases resulting from unobservable omitted and time-independent variables, it may still be biased in important ways because of omitted weather attributes. Furthermore, despite incorporating measures of daily moisture supply that have been emphasized by crop model simulations but neglected in the statistical approach, and finding that they indeed have a significant and large impact on crop yields, the projected impact of temperature rise still remains, by virtue of its large magnitude (in relation to typical inter-annual fluctuations), the dominant driver of climate change’s impact, as previous statistical studies have also found. I also found evidence that farmers expand irrigation usage in response to more uneven distribution of rainfall, and that irrigated yields are less vulnerable to the impact of this variability. The expansion of irrigation in India has been a hallmark of its agricultural development, and the findings of this paper suggest that this expansion has been quite e↵ective in increasing yields and in protecting food production in the country from deficiencies in the amount or the timing of rainfall. Putting aside a complete cost-benefit analysis, the contribution of irrigation to economic development has likely been substantial. Moreover, the results suggest that the further expansion of irrigation can help adapt agriculture to this particular aspect of climate change. But yet again, since temperature shifts will dominate climate change’s im39

pact on crop yields, and since I found irrigation is found to be less e↵ective in protecting crops from temperature shifts than from rainfall deficiencies, its overall e↵ectiveness as an adaptation strategy seems limited. This does not mean, however, that irrigation will discontinue to be an e↵ective strategy for addressing inter-annual variability in precipitation patters. In fact, there are some predictions that this variability itself may increase, and a complete welfare evaluation of the benefits of irrigation expansion will have to take this into account. The success of India’s irrigation as a bu↵ering strategy has come at high costs. Irrigation, as it is practiced in India at present is unsustainable. Surface irrigation has proven highly inefficient and the construction of large scale dams of questionable social merit (Duflo and Pande [2007], on Dams [2000]). Groundwater irrigation, which has emerged as the principal and most successful irrigation source, consumes enormous amount of energy and is unsustainably depleting the country’s aquifers (Fishman et al. [Forthcoming]). The results of this paper should not be interpreted as suggesting that water resources have little value for India’s future agricultural performance. On the contrary. They confirm the valuable services these water resources can provide, services that may be compromised at a time in which they may turn out to be even more important.

References S. Anderson. Caste as an impediment to trade. American Economic Journal: Applied Economics, 3(1):239–263, 2011. M. Arellano and S. Bond. Some tests of specification for panel data: Monte

40

carlo evidence and an application to employment equations. The Review of Economic Studies, 58(2):277, 1991. M. Au↵hammer, V. Ramanathan, and J.R. Vincent. Integrated model shows that atmospheric brown clouds and greenhouse gases have reduced rice harvests in india. Proceedings of the National Academy of Sciences, 103(52): 19668, 2006. World Bank. Agriculture for Development. World Bank and Oxford University Press, 2007. O. Deschenes and M. Greenstone. The economic impacts of climate change: evidence from agricultural output and random fluctuations in weather. The American Economic Review, 97(1):354–385, 2007. A. Dinar. Measuring the impact of climate change on Indian agriculture, volume 23. World Bank Publications, 1998. N.K. Dubash.

Tubewell capitalism: groundwater development and agrarian

change in Gujarat. Oxford University Press, 2002. E. Duflo and R. Pande. Dams*. The Quarterly Journal of Economics, 122(2): 601–646, 2007. R. M. Fishman, T. Siegfried, P. Raj, V. Modi, and U. Lall. Over-extraction from shallow bedrock versus deep alluvial aquifers: Reliability versus sustainability considerations for india’s groundwater irrigation. Water Resources Research, Forthcoming. R.M. Fishman, T. Siegfried, P. Raj, V. Modi, and U. Lall. Over-extraction from shallow bedrock versus deep alluvial aquifers: Reliability versus sustainability considerations for india’s groundwater irrigation. Water Resources Research, 47(12):W00L05, 2011. 41

BN Goswami, V. Venugopal, D. Sengupta, MS Madhusoodanan, and P.K. Xavier. Increasing trend of extreme rain events over india in a warming environment. Science, 314(5804):1442, 2006. JM Gregory and JFB Mitchell. Simulation of daily variability of surface temperature and precipitation over europe in the current and 2⇥ co2 climates using the ukmo climate model. Quarterly Journal of the Royal Meteorological Society, 121(526):1451–1476, 1995. R. Guiteras. The impact of climate change on indian agriculture. University of Maryland Mimeo, 2008. X. Guoju, L. Weixiang, X. Qiang, S. Zhaojun, and W. Jing. E↵ects of temperature increase and elevated co2 concentration, with supplemental irrigation, on the yield of rain-fed spring wheat in a semiarid region of china. Agricultural water management, 74(3):243–255, 2005. CH Hanumantha Rao, S.K. Ray, and K. Subbarao. Unstable agriculture and droughts: Implications for policy, volume 47. Vikas Pub. House (New Delhi), 1988. KJ Hennessy, JM Gregory, and JFB Mitchell. Changes in daily precipitation under enhanced greenhouse conditions. Climate Dynamics, 13(9):667–680, 1997. S.M. Howden, J.F. Soussana, F.N. Tubiello, N. Chhetri, M. Dunlop, and H. Meinke. Adapting agriculture to climate change. Proceedings of the National Academy of Sciences, 104(50):19691, 2007. K. Krishna Kumar, NR Deshpande, PK Mishra, K. Kamala, and K.R. Kumar. Future scenarios of extreme rainfall and temperature over india. In Proceed-

42

ings of the Workshop on Scenarios and Future Emissions, Indian Institute of Management (IIM), Ahmedabad, July, volume 22, pages 56–68, 2003. K. Krishna Kumar, K. Rupa Kumar, RG Ashrit, NR Deshpande, and JW Hansen. Climate impacts on indian agriculture. International Journal of Climatology, 24(11):1375–1393, 2004. K. Kumar and J. Parikh. Climate change impacts on indian agriculture: the ricardian approach. 1998. KS Kumar and J. Parikh. Indian agriculture and climate sensitivity. Global environmental change, 11(2):147–154, 2001. P. Kurukulasuriya and R. Mendelsohn. Endogenous irrigation: The impact of climate change on farmers in africa1. World, 2007. D. Lobell and M. Burke. Climate Change and Food Security: Adapting Agriculture to a Warmer World, volume 37. Springer Verlag, 2009. D.B. Lobell, W. Schlenker, and J. Costa-Roberts. Climate trends and global crop production since 1980. Science, 2011. RK Mall, R. Singh, A. Gupta, G. Srinivasan, and LS Rathore. Impact of climate change on indian agriculture: A review. Climatic change, 78(2):445–478, 2006. W. May. Simulation of the variability and extremes of daily rainfall during the indian summer monsoon for present and future times in a global time-slice experiment. Climate dynamics, 22(2):183–204, 2004. R. Mendelsohn and A. Dinar. Climate change, agriculture, and developing countries: does adaptation matter? The World Bank Research Observer, 14 (2):277, 1999.

43

R. Mendelsohn, W.D. Nordhaus, and D. Shaw. The impact of global warming on agriculture: a ricardian analysis. The American Economic Review, pages 753–771, 1994. M.H. Moench. Chasing the watertable: Equity and sustainability in groundwater management. Economic and Political Weekly, pages 171–177, 1992. G.C. Nelson. Climate change: Impact on agriculture and costs of adaptation. Intl Food Policy Research Inst, 2009. World Commission on Dams. Dams and development: a new framework for decision-making: the report of the World Commission on Dams. Earthscan/James & James, 2000. M.L. Parry. Climate Change 2007: impacts, adaptation and vulnerability: contribution of Working Group II to the fourth assessment report of the Intergovernmental Panel on Climate Change. Cambridge Univ Pr, 2007. A.G. Patt, D. Schr¨ oter, A.C. de la Vega-Leinert, and R.J.T. Klein. Vulnerability research and assessment to support adaptation and mitigation: common themes from the diversity of approaches. Assessing vulnerability to global environmental change: making research useful for adaptation, decision making and policy, page 1, 2009. S. Peng, J. Huang, J.E. Sheehy, R.C. Laza, R.M. Visperas, X. Zhong, G.S. Centeno, G.S. Khush, and K.G. Cassman. Rice yields decline with higher night temperature from global warming. Proceedings of the National Academy of Sciences of the United States of America, 101(27):9971, 2004. M. Rajeevan, J. Bhate, JD Kale, and B. Lal. Development of a high resolution daily gridded rainfall data for the indian region. Met. Monograph Climatology, 22:2005, 2005. 44

C. Rosenzweig and D. Hillel. Climate variability and the global harvest: impacts of El Ni˜ no and other oscillations on agroecosystems. Oxford University Press, USA, 2008. C. Rosenzweig and M.L. Parry. Potential impact of climate change on world food supply. Nature, 367(6459):133–138, 1994. R. Saleth. Water markets in india: Economic and institutional aspects. Markets for Water, pages 187–205, 1998. A. Sanghi and R. Mendelsohn. The impacts of global warming on farmers in brazil and india. Global Environmental Change, 18(4):655–665, 2008. W. Schlenker and M.J. Roberts. Nonlinear temperature e↵ects indicate severe damages to us crop yields under climate change. Proceedings of the National Academy of Sciences, 106(37):15594, 2009. W. Schlenker, W.M. Hanemann, and A.C. Fisher. Will us agriculture really benefit from global warming? accounting for irrigation in the hedonic approach. The American Economic Review, 95(1):395–406, 2005. W. Schlenker, W.M. Hanemann, and A.C. Fisher. The impact of global warming on us agriculture: an econometric analysis of optimal growing conditions. Review of Economics and Statistics, 88(1):113–125, 2006. T. Shah. Taming the anarchy: groundwater governance in South Asia. Earthscan, 2008. T. Shah, M. Ul Hassan, M.Z. Khattak, P.S. Banerjee, OP Singh, and S.U. Rehman. Is irrigation water free? a reality check in the indo-gangetic basin. World Development, 37(2):422–434, 2009.

45

AK Srivastava, M. Rajeevan, and SR Kshirsagar. Development of a high resolution daily gridded temperature data set (1969–2005) for the indian region. Atmospheric Science Letters, 10(4):249–254, 2009. DB Stephenson, K.R. Kumar, FJ Doblas-Reyes, JF Royer, F. Chauvin, and S. Pezzulli. Extreme daily rainfall events and their impact on ensemble forecasts of the indian monsoon. Monthly weather review, 127(9):1954–1966, 1999. C. Tebaldi, K. Hayhoe, J.M. Arblaster, and G.A. Meehl. Going to the extremes. Climatic Change, 79(3):185–211, 2006. R.S.J. Tol. The economic e↵ects of climate change. The Journal of Economic Perspectives, 23(2):29–51, 2009. R.S.J. Tol, S. Fankhauser, and J.B. Smith. The scope for adaptation to climate change: what can we learn from the impact literature? Global Environmental Change, 8(2):109–123, 1998. K.E. Trenberth, J. Fasullo, and L. Smith. Trends and variability in columnintegrated atmospheric water vapor. Climate Dynamics, 24(7):741–758, 2005. J.R. Welch, J.R. Vincent, M. Au↵hammer, P.F. Moya, A. Dobermann, and D. Dawe. Rice yields in tropical/subtropical asia exhibit large but opposing sensitivities to minimum and maximum temperatures. Proceedings of the National Academy of Sciences, 107(33):14562, 2010.

46

Table 1: Dependent Variable: Log Yields, Standardized Weather Variables (1) (2) Rice Wheat Rainy Days 0.034⇤⇤ -0.001 (0.010) (0.005) Shape Parameter

0.009+ (0.005)

0.005⇤ (0.002)

Longest Dry Spell

-0.001 (0.005)

-0.001 (0.003)

Rainfall, Total

0.063⇤⇤ (0.017)

0.039⇤⇤ (0.011)

DDS

-0.032⇤ (0.014)

-0.042 (0.026)

-0.033⇤⇤⇤ (0.005)

-0.012⇤ (0.004)

-0.010⇤ (0.004) 8131 0.777

-0.006 (0.008) 10149 0.838

Rainfall, Total, sq. DDS, sq. N adj. R2

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

47

⇤⇤⇤

p < 0.0001

48 -0.033⇤⇤⇤ (0.005) -0.010⇤ (0.004) 8131 0.777 p < 0.0001

-0.035⇤⇤⇤ (0.005) -0.012⇤ (0.005) 8131 0.775

Rainfall, Total, Sq.

DDS, Sq.

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤

-0.032⇤ (0.014)

-0.042⇤ (0.016)

DDS

N adj. R2

0.063⇤⇤ (0.017)

0.071⇤⇤ (0.017)

-0.001 (0.005)

Longest Dry Spell

Rainfall, Total

0.009+ (0.005)

Shape Parameter

-0.003 (0.005) 8131 0.766

-0.039⇤⇤⇤ (0.006)

-0.034⇤ (0.013)

0.078⇤⇤ (0.020)

-0.007 (0.007)

0.017⇤ (0.006)

-0.009 (0.006) 6575 0.780

-0.029⇤⇤⇤ (0.005)

-0.038⇤ (0.018)

0.058⇤⇤ (0.015)

-0.003 (0.006)

0.014⇤⇤ (0.004)

-0.014⇤⇤ (0.004) 5470

-0.030⇤⇤⇤ (0.003)

-0.043⇤⇤⇤ (0.007)

0.058⇤⇤⇤ (0.006)

-0.006 (0.005)

0.010⇤ (0.005)

-0.009⇤ (0.004) 8082 0.776

-0.032⇤⇤⇤ (0.005)

-0.033⇤ (0.013)

0.060⇤⇤ (0.017)

-0.001 (0.004)

0.008+ (0.005)

-0.008 (0.006) 4922 0.745

-0.027⇤⇤ (0.005)

-0.035⇤ (0.015)

0.040⇤ (0.017)

-0.009 (0.006)

0.005 (0.004)

Table 2: Log rice Yield, Rainy Season, Standardized Weather Variables, Additional Regressions (1) (2) (3) (4) (5) (6) (7) Rainy Days 0.034⇤⇤ 0.045⇤⇤ 0.042⇤⇤ 0.034⇤⇤⇤ 0.034⇤⇤ 0.032⇤ (0.010) (0.012) (0.014) (0.007) (0.010) (0.014)

49 -0.005 (0.008) 10149 0.838

DDS, Sq.

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

N adj. R2

-0.012⇤ (0.004)

-0.012⇤ (0.004)

Rainfall, Total, Sq.

⇤⇤⇤

-0.011⇤ (0.004) -0.009 (0.008) 9420 0.854

-0.012⇤⇤ (0.003) 10149 0.833

-0.048+ (0.027)

-0.053⇤⇤ (0.016) -0.013⇤⇤ (0.003)

0.038⇤⇤ (0.011)

0.003 (0.002)

0.005+ (0.003)

0.045⇤⇤ (0.010)

0.001 (0.003)

0.010⇤⇤ (0.003)

p < 0.0001

-0.006 (0.008) 10149 0.838

-0.042 (0.026)

-0.041 (0.026)

DDS

0.039⇤⇤ (0.011)

0.038⇤⇤ (0.012)

-0.001 (0.003)

Longest Dry Spell

Rainfall, Total

0.005⇤ (0.002)

Shape Parameter

-0.027⇤⇤⇤ (0.006) 8796

-0.012⇤⇤⇤ (0.002)

-0.041⇤⇤ (0.011)

0.041⇤⇤⇤ (0.004)

0.002 (0.003)

0.004 (0.003)

-0.006 (0.009) 10048 0.836

-0.012⇤ (0.004)

-0.041 (0.025)

0.039⇤⇤ (0.011)

-0.001 (0.003)

0.005⇤ (0.002)

-0.008 (0.009) 8594 0.839

-0.011⇤ (0.005)

-0.048+ (0.026)

0.038⇤⇤ (0.011)

-0.001 (0.003)

0.004+ (0.002)

-0.018 (0.021) 6770 0.844

-0.008 (0.006)

-0.011 (0.030)

0.031⇤ (0.012)

-0.000 (0.004)

0.001 (0.003)

Table 3: Log wheat Yield, Dry Season, Standardized Weather Variables, Additional Regressions (1) (2) (3) (4) (5) (6) (7) (8) Rainy Days -0.001 -0.002 0.008 0.010⇤ -0.001 0.002 -0.010⇤ (0.005) (0.003) (0.006) (0.004) (0.005) (0.005) (0.004)

Table 4: Log Cropped Areas, Standardized Weather Variables (1) (2) (3) (4) (5) Rice Rice Wheat Wheat Wheat Rainy Days 0.011 0.008 0.010 0.005 -0.004 (0.010) (0.009) (0.013) (0.013) (0.021) Rainfall, Monsoon

0.015 (0.019)

DDS

-0.035 (0.021)

0.049⇤ (0.017) -0.031 (0.021)

-0.012 (0.037)

0.040⇤ (0.017) -0.016 (0.035)

Rainfall, Jun

0.010 (0.007)

0.006 (0.011)

Rainfall, Jul

0.028+ (0.016)

0.014+ (0.007)

Rainfall, Aug

-0.003 (0.014)

0.014 (0.011)

Rainfall, Sep

0.007 (0.009) 7307 0.948

0.051⇤⇤ (0.014) 10100 0.943

N adj. R2

7307 0.948

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

⇤⇤⇤

10100 0.943

p < 0.0001

50

0.006 (0.044)

6732 0.948

Table 5: Log Irrigated Areas, Standardized Weather Variables (1) (2) (3) (4) (5) Rice Rice Wheat Wheat Wheat Rainy Days -0.037⇤⇤ -0.036⇤ -0.017 -0.011 -0.023 (0.012) (0.014) (0.011) (0.011) (0.013) Rainfall, Monsoon

-0.027+ (0.015)

DDS

-0.038 (0.038)

0.057⇤⇤ (0.019) -0.022 (0.037)

-0.021 (0.049)

0.030⇤ (0.013) -0.021 (0.048)

Rainfall, Jun

0.031 (0.018)

0.000 (0.013)

Rainfall, Jul

0.000 (0.010)

0.023⇤ (0.011)

Rainfall, Aug

-0.007 (0.006)

0.025⇤⇤ (0.007)

Rainfall, Sep

-0.055⇤ (0.021) 4228 0.941

0.021+ (0.011) 5814 0.954

N adj. R2

4228 0.940

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

⇤⇤⇤

p < 0.0001

51

5814 0.954

-0.008 (0.041)

4220 0.964

52

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

5531 0.783 ⇤⇤⇤

p < 0.0001

2979 0.795

0.01 (0.02) 2979 0.803

FIA X DDS

N adj. R2

-0.02 (0.03)

-0.08⇤⇤ (0.03)

FIA X Total 0.01 (0.01) 2979 0.821

-0.13⇤ (0.06)

-0.13⇤ (0.06)

-0.08+ (0.04)

FIA X Rainy Days

0.01 (0.01) 2979 0.822

-0.02 (0.03)

0.27⇤⇤ (0.06)

0.27⇤⇤ (0.07)

0.28⇤⇤ (0.06)

0.02 (0.04)

0.81 (2.73)

FIA

0.11⇤⇤⇤ (0.02) -0.06⇤⇤ (0.01)

-0.06+ (0.03)

-0.04⇤ (0.02)

DDS

0.11⇤⇤ (0.03) -0.06+ (0.03)

0.07⇤ (0.03)

0.06⇤⇤ (0.02)

Rainfall, Total

Rainy Days

0.01 (0.02) 2979 0.803

-0.08⇤⇤ (0.03)

-0.08+ (0.04)

0.28⇤⇤ (0.06)

-0.05⇤ (0.02)

0.06⇤ (0.02)

-0.01 (0.02) 5148 0.791

-0.11⇤⇤ (0.03)

-0.06⇤ (0.03)

-0.03 (0.02)

0.11⇤⇤ (0.03)

0.02 (0.02) 2979 0.787

-0.07+ (0.03)

-0.08+ (0.04)

0.30⇤⇤ (0.06)

-0.05+ (0.02)

0.12⇤⇤ (0.04)

-0.00 (0.02) 2344 0.809

-0.09⇤⇤ (0.02)

-0.09+ (0.05)

0.20⇤ (0.07)

-0.06+ (0.03)

0.09⇤⇤ (0.03)

0.01 (0.02) 1667

-0.14⇤⇤⇤ (0.02)

-0.07⇤⇤ (0.03)

0.29⇤⇤ (0.09)

-0.05⇤⇤ (0.02)

0.12⇤⇤⇤ (0.02)

Table 6: Log rice Yield, Rainy Season, Interacted Model with Standardized Weather Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) 0.04⇤⇤ 0.05⇤ 0.08⇤ 0.13⇤⇤ 1.37 0.03⇤ 0.07⇤⇤ 0.09⇤ 0.09⇤⇤ 0.07⇤⇤⇤ (0.01) (0.02) (0.03) (0.04) (4.45) (0.01) (0.02) (0.03) (0.03) (0.02)

0.01 (0.02) 2899 0.808

-0.09⇤⇤ (0.03)

-0.07+ (0.04)

0.24⇤⇤ (0.05)

-0.06+ (0.03)

0.12⇤⇤ (0.03)

(11) 0.08⇤⇤ (0.03)

0.03 (0.03) 1694 0.804

-0.07⇤ (0.02)

-0.12 (0.07)

0.23+ (0.10)

-0.07 (0.05)

0.08+ (0.03)

(12) 0.10+ (0.05)

53 -0.01 (0.04) -0.02 (0.03) 0.08 (0.07) 5623 0.879

-0.01 (0.02) -0.04+ (0.02) 0.14⇤⇤ (0.04) 5623 0.876

FIA X Rainy Days

FIA X Total

FIA X DDS

Standard errors in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01,

10149 0.838

0.35⇤⇤⇤ (0.05)

0.34⇤⇤⇤ (0.05)

FIA

N adj. R2

-0.19⇤⇤ (0.05)

⇤⇤⇤

p < 0.0001

5623 0.870

-0.03 (0.02)

-0.14⇤⇤ (0.04)

-0.04 (0.03)

DDS

0.07⇤⇤ (0.02)

0.04⇤ (0.01)

0.07⇤⇤ (0.02)

0.04⇤⇤ (0.01)

Rainfall, Total

Rainy Days

0.13⇤ (0.04) 5623 0.880

0.00 (0.03)

-0.03 (0.04)

0.36⇤⇤⇤ (0.05)

-0.12⇤ (0.05)

3.23⇤ (1.50)

0.14⇤⇤ (0.04) 5623 0.876

-0.04+ (0.02)

-0.01 (0.02)

0.34⇤⇤⇤ (0.05)

-0.02 (0.02)

0.04⇤⇤ (0.01)

0.18⇤⇤ (0.03) 5623 0.872

-0.04⇤ (0.02)

-0.04⇤ (0.02) 0.04⇤⇤ (0.01) 9307 0.844

-0.01 (0.02)

0.34⇤⇤⇤ (0.05)

-0.19⇤⇤⇤ (0.03)

0.07⇤⇤ (0.01)

-0.02+ (0.01)

-0.07⇤⇤ (0.02)

0.07⇤⇤ (0.02)

0.14⇤⇤ (0.03) 5275 0.881

-0.04 (0.03)

-0.01 (0.03)

0.28⇤⇤ (0.05)

-0.14⇤⇤ (0.04)

0.07⇤⇤ (0.02)

0.10⇤⇤ (0.03) 4457

-0.04⇤⇤ (0.01)

-0.01 (0.01)

0.22⇤⇤⇤ (0.04)

-0.11⇤⇤ (0.03)

0.07⇤⇤⇤ (0.01)

Table 7: Log wheat Yield, Dry Season, Interacted Model with Standardized Weather Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) -0.00 -0.00 0.01 0.01 -4.58⇤ -0.01 0.02+ 0.00 0.02 0.02+ (0.00) (0.01) (0.01) (0.03) (1.88) (0.01) (0.01) (0.02) (0.02) (0.01)

0.13⇤⇤ (0.04) 5606 0.876

-0.04+ (0.02)

-0.01 (0.02)

0.34⇤⇤⇤ (0.05)

-0.13⇤⇤ (0.04)

0.07⇤⇤ (0.02)

(11) 0.01 (0.02)

0.15⇤⇤ (0.04) 5164 0.872

-0.04⇤ (0.02)

0.00 (0.02)

0.34⇤⇤⇤ (0.05)

-0.15⇤⇤ (0.04)

0.08⇤⇤ (0.02)

(12) 0.00 (0.01)

0.12⇤ (0.05) 4075 0.874

-0.01 (0.02)

-0.02 (0.03)

0.31⇤⇤⇤ (0.05)

-0.11⇤ (0.05)

0.05⇤⇤ (0.02)

(13) 0.01 (0.02)

54

0.02 (0.57) -0.06 (0.38)

-0.08⇤⇤ (0.00) -0.19⇤ (0.01)

-0.09⇤ (0.01) 0.02 (0.73)

State FE

Over time

N 265 265 265 p-values in parentheses + p < 0.1, ⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.0001

-0.06⇤⇤⇤ (0.00)

-0.05⇤⇤ (0.00)

-0.09⇤⇤⇤ (0.00)

Simple

Degree Days

Rainy Days

Total Rainfall

365

0.04 (0.33)

-0.01 (0.72)

-0.05⇤⇤⇤ (0.00)

Total Rainfall

365

-0.04 (0.26)

-0.04 (0.63)

-0.00 (0.66)

Rainy Days

365

0.06+ (0.09)

0.07 (0.38)

0.03⇤⇤ (0.00)

Degree Days

Table 8: District Comparisons - Impact of Fraction Area Irrigated Rice, Rainy Season Wheat, Dry Season

Figure 1: Time series of all india total food grain yield (dark green) and area (light green), relative to 1970 levels, the fraction of area irrigated (blue line), and the standardized deviation of annual rainfall (blue bars, right axis). The correlation between yields, and to a lesser degree area, with total rainfall may have declined somewhat in recent years, and while cultivated area has stopped increasing, irrigation is continuing to expand.

55

Ahmedabad, Gujarat Precipitation (mm)

250 1996: Total Rainfall: 642mm Rainy Days: 77

200 150 100 50 0

0

50

100

150

200

Precipitation (mm)

250 2000: Total Rainfall: 635mm Rainy Days: 47

200 150 100 50 0

0

50

100 Days (May 1st to Nov 1st)

150

200

Figure 2: Time series of daily rainfall during the monsoon season (May to October) at Ahmedabad district, Gujarat, in the years 1996 and 2000. While total rainfall in the two years is almost identical, the intra-seasonal distributions of rainfall are clearly very di↵erent, with rainfall in 2000 concentrated in fewer days of more heavy rainfall.

56

57

Figure 3: Mean annual rainfall (left) and the number of rainy days (right), 1970-2004, in 586 Indian districts (2000 district boundaries).

>1500

1500

1400

1300

1200

1100

1000

900

800

700

600

500

Mean Annual Precipitation (mm)

Figure 4: Local polynomial fits (kernel regressions), with 95% confidence intervals (errors not spatially correlated), of rainy season anomalies of rice (top) and wheat (bottom) yields (residuals from regressions on year e↵ects, state specific quadratic time trends, other weather variables and district fixed e↵ects) on rainy day frequency (left), total rainfall (center) and seasonal degree days (right) anomalies. Histograms display the distributions of the observed anomalies.

58

Figure 5: Quadratic fits of the district-wise estimated weather coefficients (total rainfall in blue, frequency of rainy days in green, and seasonal degree days in red) against the mean fraction of irrigated area in the district.

59

Simulated CC Impact on Rice Yield 1%

Percent Change

-3%

Precipitation

Wet Days

-13%

DDS

-9%

-23% -33% -38%

-43% -53%

Simulated Impact of Rainy Days Decrease on Rice Yield Relative to the current irrigation coverage

200%

139%

150% 100% 50% 0% -50% -100%

38%

35% 0% Simple Estimate

Current

Full

None

12%

Depleted Sustainable

-80% Irrigation Coverage

Figure 6: Simulation results for climate change related impact on rice yields in India using the IPCC’s median A1B scenario for 2080-2100. Top: simulated impact on rice yield due to each of the climatic variables considered in the simulation, in percentage losses to rice yields in relation to current yields. Error bars indicate one standard deviation. Bottom: the impact of the decrease in rainy days in various irrigation scenarios: the result for each scenario is displayed in relative percentage terms to the impact 60 calculated in an irrigation scenario that preserves current coverage of irrigation. The scenarios are defined in the text.

61

Figure 7: Share of irrigated area that is irrigated by groundwater (left) and stage of groundwater exploitation (right), defined as the ratio between estimated usage and estimated renewable supply through recharge. Stage of exploitation above 100% indicates unsustainable over-extraction or mining. Source: The 3rd Minor Irrigation Census and the Central Groundwater Board of India.

86% - 100%

66% - 85%

44% - 65%

22% - 43%

0% - 21%

Legend

Percentage of Area Irrigated by Groundwater