Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

H OW D O A FRICAN F ARM H OUSEHOLDS

R ESPOND TO C HANGES IN C URRENT AND P AST


W EATHER P ATTERNS ? A S TRUCTURAL P ANEL
D ATA A NALYSIS FROM M ALAWI
J UAN S ESMERO , J ACOB R ICKER -G ILBERT , AND A ARON C OOK

We use three waves of nationally representative household-level panel data from Malawi to estimate a
structural model characterizing the response of smallholder farm households to current and past
weather patterns, and the subsequent impacts on household net income. We also quantify heterogene-
ity among households along the wealth spectrum regarding their ability to adapt to evolving weather
patterns. This approach yields two key findings. First, adverse weather history prompts households to
devote more time to maize cultivation on their own farms, to the detriment of other, possibly more re-
munerative income sources. Households also reduce application of productivity-enhancing inputs, such
as fertilizer and improved maize varieties, in response to adverse weather history. Our results are ro-
bust to different clustering structures and falsification tests aimed at ruling out alternative explanations
to observed trends. Second, we find that, by maintaining a more diversified income structure, wealthier
households are better able to adapt to adverse weather history. Consequently, adverse changes in past
weather may be regressive in nature, creating a “climate-induced” poverty trap that locks poor small-
holder households into low-value maize cultivation from season to season. This finding suggests that de-
veloping more weather-resilient maize varieties and promoting smallholder livelihood diversification
strategies may help mitigate the effects of adverse weather on the most vulnerable households.

Key words: Adaptation, climate, income diversification, maize, poverty trap, structural model,
wealth, weather.

JEL codes: D13, J43, O13, Q12, Q54.

Most scientists agree that Earth’s climate is projected to increase by 4 C by the end of the
changing, and the 5th Assessment of the twenty-first century (Niang et al. 2014). In ad-
Intergovernmental Panel on Climate Change dition, southern Africa is expected to receive,
(IPCC) projects that sub-Saharan Africa on average, less precipitation and more rain-
(SSA) will be severely affected. Temperatures fall variability in the future. The changing
across SSA have increased by 0.5 Celsius (C) weather patterns will have a major impact on
over the past 50 to 100 years, and are rain-fed agriculture. In turn, this will affect the
welfare of hundreds of millions of people in
SSA, as up to 90% of the population in some
countries are engaged in farming activities.
Juan Sesmero and Jacob Ricker-Gilbert are associate professors Given the potential ramifications that
of agricultural economics at Purdue University. Aaron Cook is a
PhD Candidate in the Department of Agricultural Economics,
changes in weather trends will have for human
Sociology, and Education, The Pennsylvania State University. welfare in SSA, it is critical to accurately quan-
This study received support from the USDA National Institute tify its effects on agriculture so that policy mak-
of Food and Agriculture (NIFA), Grant no. 2014-51130-22492,
and the Purdue Research Foundation. The authors would like to ers can understand the benefits of climate
thank the editor Travis Lybbert and three anonymous referees policies or, conversely, the cost of inaction.
for their helpful comments. We would also like to thank Jerry There is a rapidly growing literature on how
Shively for valuable comments on an earlier draft. All remaining
errors are our own. The Department for International households may or may not adapt to changes in
Development (DFID)/Malawi, USAID/Malawi, and USAID’s weather patterns in developing countries.
Environment, Growth, and Technology Division funded the col-
lection of the data used in this article. Correspondence to be sent
However, scientists’ understanding of adapta-
to: jsesmero@purdue.edu. tion mechanisms such as switching crop

Amer. J. Agr. Econ. 100(1): 115–144; doi: 10.1093/ajae/aax068


Published online November 17, 2017
C The Authors 2017. Published by Oxford University Press on behalf of the Agricultural and Applied Economics
V
Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
116 January 2018 Amer. J. Agr. Econ.

varieties or diversifying income sources, as well studies focus on relatively drastic changes in
as the factors that facilitate or inhibit house- management practices (e.g., adopting irriga-
holds’ ability to adapt to changes in rainfall and tion, building water-harvesting schemes, and
temperature patterns remains limited. adopting soil conservation measures such as
With this in mind, we use three waves of terracing or planting trees). Yet these studies
nationally-representative panel data on small- do not consider the potential influence that
holder farm households in Malawi, collected in changes in weather patterns may have on
2002/2003, 2006/2007 and 2008/2009, to esti- more subtle, but equally important adapta-
mate a structural model of how smallholders tion mechanisms such as those mentioned in
adapt to changing weather patterns and the the previous paragraph. Nor do they distin-
resulting impact on net income.1 Malawi guish between weather occurrences during
makes an excellent case study to measure be- the current growing season and past (histori-
havioral responses to evolving weather pat- cal) weather patterns that may influence
terns for two main reasons. First, perceivable smallholder expectations.3 We quantify these
changes in weather patterns, such as shortening more subtle behavioral responses while dis-
of the growing season as well as higher temper- tinguishing between reactions to past weather
ature and rainfall variability, appear to be tak- patterns and weather that occurs over the
ing place in Malawi (Tadross et al. 2009). current growing season.
Second, the majority of Malawi’s population is Our second contribution is to model het-
engaged in subsistence agriculture, focused on erogeneity in households’ ability to respond
maize production. This makes households par- to changing weather patterns, along with the
ticularly vulnerable to adverse changes in cli- mechanisms underpinning such heterogene-
matic conditions, and also limits their ability to ity. Our framework recognizes and accommo-
make drastic changes in production decisions. dates the generally-accepted notion that
Therefore, it is essential to understand the role behavioral responses to changing weather
that more subtle mechanisms, such as adjusting patterns may vary across wealth levels. To
input quantities and/or reallocating time our knowledge, Dercon and Christiaensen
invested in alternative income sources, play in (2011), using smallholder data from Ethiopia,
smallholder decision making. and FAO (2016a), using smallholder data
We make two important contributions to from Tanzania, are the only other studies to
the literature on climate change and agricul- consider this issue, albeit in ways that differ
ture in SSA.2 First, many studies have esti- from ours. Dercon and Christiaensen esti-
mated the welfare implications of changes in mate how historical rainfall distributions cre-
weather patterns considering households’ ad- ate consumption risk for smallholders and
aptation to such changes either implicitly how that risk affects their use of inorganic
(Deressa 2007; Mendelsohn, Nordhaus, and fertilizer, conditional on their livestock
Shaw 1994; Sanghi and Mendelsohn 2008; wealth. FAO (2016a) estimates how
Wang et al. 2009) or explicitly (Seo and weather shocks (anomalies relative to an
Mendelsohn 2008; Seo 2010a, 2010b; Hassan average) affect welfare across the wealth
and Nhemachena 2008; Deressa et al. 2009; spectrum. Our analysis builds upon these
Deressa, Hassan, and Ringler 2011; Fosu- studies by examining how assets shape be-
Mensah, Vlek, and MacCarthy 2012; havioral responses across multiple, interde-
Di Falco and Veronesi 2013). However, these pendent production decisions such as
applying fertilizer, hiring labor, adopting
improved maize varieties, and working in
non-maize-related activities. We then
1
The first wave includes information on the 2002/2003 and quantify the resulting impact of weather
2003/2004 growing seasons. This wave was collected as part of
the Second Integrated Household Survey (IHS2). The second patterns on household net income.
and third waves span the 2006/2007 and 2008/2009 growing sea- We recognize that measuring behavioral
sons, respectively, and come from the Agricultural Inputs responses to changing weather patterns is far
Support Surveys I and II.
2
Our discussion of the literature focuses on previous studies from trivial. In recent reviews of the climate
examining, either implicitly or explicitly, behavioral responses to
changing weather patterns. Another important segment of the
existing literature measures the direct impact of weather on crop
3
yields (Adams et al. 1988; Adams 1989; Kaylen, Wade, and Burke and Emerick (2016) constitutes a notable exception.
Frank 1992; Easterling et al. 1992; Rosenzweig and Parry 1994; The authors distinguish longer-term weather trends from year-
Schlenker and Roberts 2006; Deschenes and Greenstone 2007; to-year fluctuations; they find that crop yields are as susceptible
Schlenker and Lobell 2010; Di Falco, Veronesi, and Yesuf 2011; to long-term patterns (or even more so in the case of rainfall), as
Di Falco and Veronesi 2013; Arslan et al. 2015). they are to year-to-year fluctuations.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 117

change literature, Auffhammer and labor and other scarce resources. Results also
Schlenker (2014) and Burke and Emerick suggest that changes in weather patterns per-
(2016), discuss the main identification chal- ceived as lasting (i.e., changes that house-
lenges emerging when researchers attempt to holds perceive to be recurrent over time)
measure households’ behavioral responses to seem to influence their expectations and are
changes in climatic patterns. First, a house- more likely to prompt behavioral changes
hold’s production decisions are interdepend- than are those that are considered to be
ent and weather patterns may be correlated short-lived. This not only indicates that there
with observable time-varying market condi- is ample heterogeneity in behavioral
tions (i.e., prices and profitability). Second, responses along the wealth spectrum, but also
there is a substantial risk of omitted-variable reveals that adverse changes in past weather
bias as many unobservable factors (both force households with fewer assets to spend
time-constant and time-varying) may be cor- more effort cultivating their staple crop—
related with weather patterns and production maize—likely in an effort to ensure their own
decisions. Third, analyses based on spatially- food security. Therefore, our results point to
aggregated data (e.g., production decisions at a “climate-induced” poverty trap that locks
the village or district levels) risk aggregation poor smallholder households into low-value
bias due to large spatial heterogeneity on cli- maize cultivation from season to season, at
mate change impacts (Fezzi and Bateman the expense of engaging in other possibly
2015). more remunerative activities. The challenge
We address the main identification issues of creating policies that can alleviate this pov-
highlighted by these studies in the following erty trap, along with the hardships associated
ways. First, our structural model of household with being stuck in this low-outcome equilib-
behavior includes data on observable market rium, highlight the need for longer-term strat-
conditions, and also accommodates profit- egies that focus on helping a significant
maximizing, interdependent production deci- number of smallholder farmers move out of
sions through cross-equation restrictions and agriculture and into non-farm sources of live-
correlated error terms. Second, we exploit lihood (Collier and Dercon 2014; Henderson,
panel data to estimate the models via house- Storeygard, and Deichmann 2017).
hold fixed effects (FE), which reduces omit-
ted variable bias by removing time-invariant
unobservable factors such as household’s
ability and risk aversion from the model. We Smallholder Agriculture in Malawi
also conduct falsification tests to examine
whether the effects of weather patterns on Malawi is a landlocked country in SSA with
behavior are confounded by other time- an annual per capita income of US $750 per
varying unobservables. Third, since we model year in 2013 purchasing power parity terms.
and quantify decisions at the household level, Roughly 85% of Malawians live in rural areas
and explicitly model heterogeneous behav- and earn a living through subsistence agricul-
ioral responses, we greatly reduce the risk of ture. Maize is by far the main crop grown by
measurement error that is common in more smallholders, and accounts for 60% of per
aggregate datasets. Finally, we conduct a capita caloric intake (Ecker 2009).
number of robustness tests to assess the Agricultural productivity is low, with yield
strength of our results. These include, (a) estimates of 1.5 tons per hectare, and an offi-
clustering the structure of the error terms, (b) cial poverty rate of 50.7% in 2010
changing the timeframe on which past (Government of Malawi 2012). Low produc-
weather is defined, and (c) adjusting the tivity and high poverty rates are compounded
threshold used to define “extreme” by high population growth of 2.7% per year
temperature. and one of the highest population densities in
Our results are largely consistent across all of SSA, with 128 people per square kilo-
multiple specifications and robustness checks. meter (Government of Malawi 2008).
In fact, our findings support our prior ex- Increasing population density, especially in
pectation that adjustments in input applica- rural areas, is leading to shrinking farm sizes,
tion and re-allocation of income sources declining fallow land, continuous cropping of
constitute an important behavioral response maize, and mass depletion of soil fertility.
to changes in weather patterns, likely because The threat of severe weather in the form of
they require relatively small investments in droughts, floods, and extreme heat due to

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
118 January 2018 Amer. J. Agr. Econ.

changes in the broader climatic regime program (for more details on Malawi’s FISP,
underlies all of the issues surrounding small- see Chirwa and Dorward 2013; and Lunduka,
holder agriculture in Malawi. The climate is Ricker-Gilbert, and Fisher 2013). For the
sub-tropical, with a uni-modal rainfall pattern purposes of our study it is important to con-
that allows for one rain-fed growing season trol for the effect of Malawi’s large-scale in-
per year, which takes place between put subsidy program, which clearly changes
November and April. At the same time, the availability and relative profitability of us-
meaningful investment in irrigation during ing inorganic fertilizer and improved seed va-
the dry season is not an option for most rieties. The empirical section of this article
smallholders in Malawi.4 While rainfall typi- discusses how we control for the impact that
cally ranges between 725 mm to 2,500 mm, FISP has on both improved maize and inor-
droughts are not uncommon, particularly ganic fertilizer use so that the program’s im-
during El Ni~ no years. In fact, major droughts pact is not mistakenly attributed to changes
occurred in Malawi in 1987, 1990, 1992, 2002, in past weather.
2005, and 2007 (RMSI 2009).5 Flooding is
also an issue, particularly in low-lying areas
of southern Malawi in the Lower Shire River Differentiating Current and Past Weather
floodplain and along Lake Malawi. RMSI Patterns
reports that significant floods occurred in
Malawi in 1991, 1997, 2001, 2002, 2003, and The history of weather occurrences may
2007. Finally, average temperature has been shape smallholder households’ expectations
rising, which increases the risk of harmful ex- of future weather patterns and, consequently,
treme heat events (Tadross et al. 2009). influence their production decisions.
In an effort to boost productivity and com- Growing-season weather may also influence
bat the many challenges facing smallholder their behavior. However, to the extent that
agriculture, the Malawian government scaled some production decisions are made before
up its input subsidy program for inorganic growing season weather takes place, behav-
fertilizer and improved maize seed in 2005/ ioral responses to past and current weather
2006. Known as the Farm Input Subsidy may differ. Therefore, we distinguish histori-
Program (FISP), this program distributes pa- cal weather patterns from their occurrence
per vouchers to smallholder households who over the current growing season, and include
meet certain criteria, as determined by local both measures in the empirical model. This
leaders. Recipient households can redeem allows us to assess the potentially different
the vouchers for 100 kilograms of fertilizer responses that smallholders make to expected
(one 50 kilogram bag of NPK, and one 50 ki- and unexpected changes in weather.
logram bag of urea), and for 2–4 kilograms of
improved maize seed. The subsidy was sub- Past Weather Patterns and Weather
stantial during the second and third waves of Expectation
our survey in 2006/2007 and 2008/2009. At
that time, more than half of the smallholder The behavioral premise behind the inclusion
farm population was targeted to receive pa- and measurement of past weather (and their
per vouchers that entitled them to acquire distinction from current growing season
inputs at a substantially reduced price. weather) is that households make many pro-
Fertilizer was subsidized at 72% of the com- duction decisions based on expectations of
mercial price in 2006/2007, and at 91% of the conditions in the upcoming growing season
commercial price in 2008/2009. Maize seed that are in turn informed by their perceptions
was subsidized at 100% in all years of the of general weather patterns. This is a com-
mon premise in the literature (e.g., Di Falco
and Veronesi 2013). Unfortunately, the
4 length of an agricultural producer’s memory
About 35% of the households earn positive income from dry
season (dimba) or tree/orchard cultivation. Much of it is very lim- of past weather is unobservable to the re-
ited as 50% of those earning this income source make less than searcher, as is the procedure by which the
US $20 from it, and most of the formal discussion of irrigation in-
vestment in Malawi (e.g. Greenbelt Initiative) has happened
producer maps past weather to expectations
since the most recent year of our data in 2008/09. of future weather (the expectation formation
5
As discussed by Pauw, Thurlow, and van Seventer (2010), process). We follow previous literature in
“RMSI is the registered name of a private consultancy and is no
longer considered an acronym for “Risk Management Solutions, SSA (e.g., Di Falco, Veronesi, and Yesuf
India,” as the firm was previously known.” 2011), and take the average of past weather

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 119

occurrences as a proxy of weather expecta- 30 C (the threshold above which increased
tions. As historical averages change over temperature hampers maize growth), and
time, so do weather expectations, and farm- subsequently adding up days over the grow-
ers update their beliefs based on the weather ing season. The SDDs are included to capture
that they observe and use them to inform adverse effects of excess heat and calculated
decisions in subsequent years. by subtracting 30 C (the threshold suggested
Given that our panel data comprises three by Schlenker and Lobell 2010) from the max-
waves spanning eight years, and uses rela- imum daily temperature, and subsequently
tively long historical periods to calculate past adding up days over the growing season. The
weather averages, we end up with overlap in source of weather data used in our analysis is
past-year weather when constructing these discussed in the data section.
averages. This reduces time-series variability,
but shortening the length of the period based Current Growing Season Weather
on which households form expectations may
also be misleading, as it may not capture their Current growing season weather refers to an-
perception of stable patterns. Most previous nual weather occurrence. As growing season
studies in SSA have defined past weather pat- weather unfolds, households may still be able
terns as average climatic conditions over the to adjust production decisions. For example,
previous 20 years (Di Falco, Veronesi, and they may decide not to apply inorganic fertil-
Yesuf 2011; Di Falco and Veronesi 2013; izer if conditions are severe at the beginning of
Arslan et al. 2015). However, Rao et al. the growing season, or they may decide to
(2011) suggest that ten years may be a more reallocate labor resources during the growing
appropriate boundary for maximum memory, season as weather unfolds. In contrast to
while Burke and Emerick (2016) use five longer-term patterns, households’ ability to re-
years of weather history. Our main specifica- spond to growing season weather may be more
tion uses 10 years of historical weather data, limited, so current-year weather patterns can
as it gives a mid-range of what is considered substantially affect household income.
relevant historical weather in the literature. Consistent with historical weather, we cap-
However, we test the robustness of our ture current growing season by GSP,
results by estimating the model under both CVGSP, GDD, and SDD. However, in con-
20-year and 5-year weather histories. trast to historical weather, these are mea-
We proxy historical weather in our analysis sured over the growing season corresponding
using four variables, (a) historical average of to each wave, and not averaged over the pre-
growing season precipitation (GSP), (b) his- ceding years. Another difference between
torical average of between-year coefficient of historical and current weather is in the mea-
variation of growing season precipitation surement of CVGSP. Historical CVGSP cap-
(CVGSP), (c) historical average of growing tures variability between seasons, while
degree days (GDD), and (d) historical aver- current CVGSP is measured as the coefficient
age of stress degree days (SDD).6 Growing of variation of monthly precipitation within
season precipitation is rainfall measured in the survey wave’s growing season.
millimeters from November to April of each
year, and GSP is the average of this annual
measure over the last T years. The annual Adaptation Strategies for Malawian
variability of precipitation is measured by the Smallholders
ratio of standard deviation of growing-season
rainfall over a period of T years, to mean Adaptation measures that appear most fre-
growing-season rainfall over a period of T quently in previous studies estimating small-
years. Growing-degree days are calculated by holder farm household responses to past
subtracting 10 C (the threshold for maize be- weather conditions fall broadly into three cate-
low which little plant growth occurs) from the gories: (a) diversifying income between on-
lowest of average daily temperature and farm and off-farm activities (Shewmake 2008;
Molua 2011; Fosu-Mensah, Vlek, and
MacCarthy 2012; Silvestri et al. 2012), between
crop and livestock activities (Thomas et al.
6
There was too little variation in the volatility of temperature 2007; Hassan and Nhemachena 2008; Seo
during our sample period to have any influence in production
and consumption decisions. Therefore, our regression does not 2010a, 2010b), and among multiple crops
include a variable for the coefficient of variation in temperature. (Dinar et al. 2008; Kurukulasuriya and

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
120 January 2018 Amer. J. Agr. Econ.

Mendelsohn 2008; Seo and Mendelsohn 2008; equilibrium model and find that shifting 10%
Mertz et al. 2009); (b) switching crop varieties of cultivated area from local maize varieties
(Molua 2002; Benhin 2006; Maddison 2007; to improved varieties fully offsets yield losses
Below et al. 2010); and (c) adjusting the quan- associated with a mild drought, and partially
tity of inputs applied in the production process. offsets those associated with more severe
While there are many possible adaptation droughts in Malawi.
mechanisms available to smallholder house- Furthermore, discussions with focus groups
holds (Asfaw et al. 2014 and Di Falco, of smallholders conducted across Malawi in
Veronesi, and Yesuf 2011), our goal is not to 2011 by one of the authors suggests that farm-
create an exhaustive list of such mechanisms ers’ perception of improved maize variety
and include them in the model. Instead, we fo- properties are consistent with the literature,
cus on those adaptation strategies that seem and that their adoption is an attractive alter-
most realistic for Malawian smallholders. native to alleviate vulnerability to unpredict-
able weather patterns. Therefore, it may be
reasonable to expect a shift from local varie-
Adopting Improved Maize Seed Varieties ties to improved varieties if Malawian house-
Maize is by far the main crop grown by holds perceive increased volatility in rainfall
Malawian smallholders, and they choose how patterns, provided they have proper weather
much land to allocate between local varieties information and access to improved seeds.
and improved varieties that are either open- However, households may be deterred from
pollinated varieties (OPV) or hybrids. Previous using improved varieties if their agronomic
literature in Malawi suggested that farmers pre- benefits (higher or more stable yields) under
fer local varieties for their consumption charac- expected weather patterns are hampered by
teristics such as taste, ease of pounding, and their relatively high cost. Therefore, adoption
storability, while improved varieties are pre- of improved varieties in response to expected
ferred for their production characteristics of or actual adverse weather is ultimately an
higher yields, early maturity, and drought resis- empirical question.
tance (Smale, Heisey, and Leathers 1995;
Lunduka, Fisher, and Snapp 2012). Diversification of Income Away from Maize
Local maize varieties are relatively low-
yielding but farmers can save the seeds and Diversifying income sources away from sub-
replant them in multiple seasons. Hybrid va- sistence maize production represents a sec-
rieties have the highest yields but lose signifi- ond relevant adaptation strategy for
cant productivity if the seeds are recycled and households in SSA, and diversifying into
used in more than one season. OPV varieties other crops may be feasible for some small-
fall between local and hybrid varieties in holders in Malawi. For example, tobacco has
terms of yield potential but can be used in historically been the main cash crop for
multiple seasons. Given differential perform- smallholders in Malawi, and it is more
ances between improved and local varieties drought resistant than maize. RMSI (2009)
under alternative climatic conditions, the estimates that in a severe drought, tobacco in
adoption of improved maize seed varieties Malawi will only lose 7% of its productivity.
may very well constitute an important climate Raising livestock is another option for small-
change adaptation strategy in Malawi.7 RMSI holders, although it is fairly limited in Malawi
(2009) estimates that during a severe drought due to severe land constraints. Working off-
in Malawi, local maize varieties lose 80% of farm, either by running or working for a
their productivity, while hybrid maize varie- micro-enterprise or working as an agricul-
ties and OPV varieties lose 25% and 16% of tural laborer on another farm is also an im-
their productivity, respectively. Pauw, portant income diversification source for
Thurlow, and van Seventer (2010) use esti- smallholder households. Income from work-
mates from RMSI in a computable general ing as an agricultural laborer on other farms
(called ganyu in Malawi) is estimated to
make up 10% of rural household income on
7
In this article improved maize seeds are defined as hybrid va-
average in Malawi, while non-farm income
rieties and open-pollinated varieties (OPV). Although small- makes up 26% of household income, on aver-
holder farm households in Malawi report that more than 95% of age (Jayne et al. 2010). Generally, poorer
the improved maize seed they acquire is hybrid, anecdotal evi-
dence from Malawi indicates that most households refer to any households are more likely to work as ganyu
improved seed as hybrid. laborers. They do this either to earn money

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 121

to purchase inputs at planting, to earn cash to context, net income (p) is affected by growing
buy food, or to earn maize directly as an in- season weather, denoted by W (which includes
kind payment to make up for production rainfall, rainfall variability, growing degree
shortfalls on their own farms (Alwang and days, and stress degree days), input and output
Siegel 1999). prices (p), a vector of other conditioning varia-
It also seems clear that smallholder house- bles (z), and input demands and output sup-
holds may change time allocated to different plies (y). Decisions on inputs and outputs are
activities in response to changing weather not only a function of growing season weather,
patterns. This may also translate into changes but also past weather denoted by PW as de-
in demand for hired labor, which can be an at fined before. Netputs are also functions of pri-
least partial substitute for own-farm labor. ces and other control variables. Net income can
Therefore, our structural model explicitly generally be expressed as
considers non-maize income sources and
hired labor. In combination, these categories ð1Þ p ¼ f fW; p; z; y½W; PW; p; zg:
allow us to make inference on households’
time allocation decisions and the effect of Note that past weather does not affect net in-
weather patterns on those decisions. come directly, but does so indirectly through
production decisions. Specifically, past
Inorganic Fertilizer Intensity weather patterns influence choices of inputs
and outputs, and ultimately net income. In
Altering the intensity of input utilization con-
contrast, current weather affects net income
stitutes a third means by which households
directly through its effects on yields, prices,
reduce production risk in the face of changing
and other sources of income. Including cur-
weather patterns in Malawi. Downing et al.
rent growing season and past weather allows
(1997) list “incremental adjustment in inputs”
us to distinguish the potential differences in
among the potential response strategies to
smallholders’ behavioral responses to actual
climate change in Africa. Specifically, Benhin
and expected weather.
(2006) find that farmers in South Africa in-
Adaptation decisions are likely to be made
crease the application of chemicals and ma-
jointly. For instance, the amount of inorganic
nure in order to reduce soil moisture loss and
fertilizer applied or time allocated to maize
maintain fertility. We explicitly model and
production by the household may be influ-
discuss changes in inorganic fertilizer applica-
enced by the type of variety planted as dis-
tion in response to growing season weather
cussed by Nkonya, Schroeder, and Norman
and past weather patterns. We also consider
(1997) and Ogunlade et al. (2010). In turn,
whether or not households have access to or-
households jointly determine the time that
ganic manure through livestock ownership.
they allocate to maize production and the
This is important to consider because small-
amount of labor that they hire in. Our struc-
holders in drier areas use manure to conserve
tural model accommodates such interdepen-
water in the soil, and livestock tends to be
dencies, and by doing so, it explicitly models
owned by relatively richer households, so
the channels through which weather affects
wealthier households may be more likely to
households’ net income.
use manure on their soils.8

Conceptual Framework Estimation Strategy

We start with a separable farm household We specify household net income as a trans-
model where production decisions are not af- log function due to its flexibility, and to the
fected by consumption preferences.9 In this

by a female, and number of adult equivalents in the household)


8
We thank an anonymous reviewer for making this point. are added as covariates in the estimated model. Under the null
9
While relaxing the separability assumption may improve es- hypothesis of separability, we find that the coefficient estimates
timation, it only does so in the absence of mis-specification in the on these variables are not statistically significant (p-values are
production/consumption model. On the other hand, the separa- 0.98 for the female-headed household dummy and 0.32 for the
bility assumption increases model parsimony and provides a clear household size variable, respectively). Failure to reject the null
starting point for our structural model. Following Benjamin lends some credibility to the separability assumption in our
(1992), we run a test for separability where socio-demographic model. With this in mind, the household is assumed to maximize
variables (such as a binary variable ¼1 if the household is headed net income with respect to a vector of inputs and outputs.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
122 January 2018 Amer. J. Agr. Econ.

ease with which elasticity of net income with growing season as the base year. Year dum-
respect to climate variables can be calculated. mies are included to capture year-specific fac-
We choose non-maize income as the numer- tors that may affect a household’s net
aire and normalize all other prices by it. A income, such as macroeconomic shocks.
full expression for the translog system with Finally, Wm and PWm represent weather in
error terms added (subscripts depicting the the current growing season and past weather
cross-sectional and time-series units have history, respectively.10
been removed for notational convenience) is Net income, netput shares, and variables in
as follows: vector zk are household- and time-specific;
P weather variables and prices are location-
ð2Þ ln p ¼ a0 þ i ai lnpi and time-specific. Household net income is
1P P measured in real Malawi Kwacha/year and is
þ i h cih lnpi lnph defined as total income from agricultural and
2 P
P non-agricultural sources minus expenditure
þ i k wik lnpi ln zk on inputs.11 Income from agricultural sources
P
þ m bm ln Wm include: (a) maize production; and (b) non-
P P maize income, including on-farm income
þ i m #im lnpi ln Wm
P P P (e.g., tobacco, livestock sales, and livestock
þ i m j uimj lnpi ln Wm ln z1;j product sales) along with agricultural and
P P non-agricultural off-farm work (e.g., working
þ i m dim lnpi ln PWm
P P P as a laborer on another farm, selling fire-
þ i m j wimj lnpi ln PWm ln z1;j wood, trading, fishing, or operating a non-
P farm related enterprise). Fertilizer quantity is
þ k hk ln zk þ ep
the total quantity of inorganic fertilizer ac-
1P
ð3Þ si ¼ ai þ cii lnpi þ h6¼i cih lnph
quired by the household from subsidized and
P 2 P commercial sources. Labor is the number of
þ k wik ln zk þ m #im ln Wm days of non-family labor that the household
P P hires in to work on their farm. We measure
þ m j uimj ln Wm ln z1;j
P quantity of improved maize seed varieties as
þ m dim ln PWm the total kilograms a household acquires and
P P plants in the growing season from subsidized
þ m j wimj ln PWm ln z1;j
and commercial sources. We do not explicitly
þ es i i ¼ 1; 2; 3; 4 model demand for local maize seed varieties,
as households often re-use them in multiple
where p denotes normalized net income, si seasons and there is no formal, developed
denotes the share of netput i on net income market for them.
(i ¼ 1 if maize, i ¼ 2 if fertilizer, i ¼ 3 if labor The overall impact of changes in past
demand, and i ¼ 4 if improved maize seed), weather patterns on households’ well-being is
pi denotes normalized prices for maize (i¼1), measured by the elasticity of net income with
fertilizer (i¼2), labor (i¼3), and improved respect to weather variables,
seeds (i¼4). Normalization of prices guaran-
tees that the net income function is homoge- @lnp
neous of degree zero in prices. We impose @lnPWm
symmetry in the net income function by forc- ð4Þ ¼P dim lnpi þP P w lnpi lnz1;j ;PWm
ing cih ¼ chi for all i and h (where both i i i j imj
and h denote netputs). The vector of covari- 2fGSPGS;CVGSPGS;GDDGS;SDDGSg
ates, denoted by zk , includes the value of
household assets, landholdings, quantity of
where GSPGS, CVGSPGS, GDDGS, and
subsidized fertilizer acquired by the house-
SDDGS are rainfall, rainfall variability,
hold, quantity of subsidized seed acquired by
the household, and age of the household
head, which is our proxy for farming experi-
ence. The vector zk also includes a dummy 10
We conduct an inverse hyperbolic sine transformation of
variable that equals one if the household SDD so that we can take logarithms of the variable.
11
owns livestock in the current year, which is Input costs in agricultural production activities include fer-
tilizer, seed, hired labor, and land rental costs. Households are
intended to proxy for manure use. We also asked to quantify input costs in livestock and other agricultural
include year dummies, treating the 2002/2003 activities as well as in non-agricultural activities.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 123

growing degree days, and stress degree days season weather does. This is of course
for the current growing season, respectively. intended to capture the effect of growing
The behavioral responses to changes in season weather on agricultural production.
past weather are characterized by the partial We treat income elasticities with respect to
derivative of input and output shares (si ) with current and past weather asymmetrically by
respect to historical weather variables. Such including stand-alone current growing sea-
partial derivatives portray how changing past son weather variables and excluding
weather patterns affect the relative intensity stand-alone past weather variables in equa-
with which households use inputs and rely on tion (2).
alternative income sources. These derivatives We characterize behavioral responses in
are equations (5) and (7), and we model hetero-
geneity in these responses based on the value
@si of household assets by including three-way
@lnPWm interaction terms between prices, weather
P
ð5Þ ¼dim þ w lnz1;j ; i¼1;2;3;4 PWm (past and current), and asset variables in the
j imj net income equation. Applying Hotelling’s
2fGSPGS;CVGSPGS;GDDGS;SDDGSg: lemma to the interaction terms generates in-
put and output shares on net income that are
We denote the share of net income from influenced by weather, and also depend upon
non-maize activities in total net income by s5 . assets. This permits us to assess the relative
Since non-maize income is the normalizing vulnerability of asset-poor households to
factor and given that input and output shares changes in weather. These mechanisms are
sum to 1 by definition, the partial effect of modeled through the partial derivatives in
past weather on the share of non-maize in- equations (5) and (7) that are functions of
z1;j , representing household asset holdings.
P4
come on net income is @ln@sPW5
m
¼  @si
@ln PWm . Therefore, parameters wimj and /imj capture
i¼1 heterogeneity in behavioral responses to cur-
The impact of changes in current (growing rent and past weather, respectively, along the
season) weather on net income can be wealth spectrum. In turn, dim and #im capture
depicted by the components of behavioral responses to
weather patterns that are common across
@ln p X X households.
ð6Þ ¼ bm þ #im lnpi
@ln Wm mX X i
þ uimj lnpi ln z1;j Wm Omitted Variable Bias and Identification
i j
The challenge of identifying the effects of
2 fGSP; C VGSP; GDD; SDDg: current and past weather comes from the pos-
sibility that many factors can be correlated
Further, behavioral responses to changes in with weather patterns, as well as with produc-
growing season weather are tion decisions and net income. We control for
a number of time-varying observables as de-
@si X scribed in the previous section. These observ-
¼ #im þ /imj ln z1;j ; ables include prices, household
@ln Wm
ð7Þ j
characteristics, and assets. We also control
i ¼ 1; 2; 3; 4
for institutional factors embedded in
Wm 2 fGSP; CVGSP; GDD; SDDg: Malawi’s FISP. This is important, as weather
patterns have been changing in Malawi at the
The partial derivative of the share of non- same time that FISP has been scaled up.
maize income with respect to growing season Furthermore, the presence of unobservable
P
4
factors influencing behavior and outcomes
weather is @ln@sW
5
¼  @ln@sWi m .
m
i¼1 and being correlated with weather patterns
Equations (4), (5), (6), and (7) clarify the may also bias our estimates. The system
mechanisms through which past and current formed by the net income equation and net-
weather affect net income and behavior. put shares is estimated via seemingly unre-
First, a comparison between equations (4) lated regression, and a household FE
and (6) reveals that past weather does not af- estimator is used to remove any correlation
fect net income directly, but current growing between the covariates and time-constant

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
124 January 2018 Amer. J. Agr. Econ.

unobservable factors.12 Unfortunately, FE es- wave of surveys was administered as part of


timation does not eliminate the risk of unob- the Second Integrated Household Survey
servable time-varying factors correlated with (IHS2) conducted by Malawi’s National
weather and the dependent variables. Statistical Office (NSO) following the 2002/
Therefore, we conduct a falsification test 2003 and 2003/2004 growing seasons. The
designed to rule out the possibility that the IHS2 covers 26 districts and 11,280 house-
estimated effect of weather patterns is being holds in Malawi, while waves of data come
confounded with other factors. In particular, from the 2007 Agricultural Inputs Support
we change the timing of weather histories as Survey (AISS1) conducted after the 2006/
described in more detail in the robustness 2007 growing season by the NSO. Due to
check section. If mismatched weather histo- budget constraints, the AISS1 survey was
ries do not explain dependent variables, then smaller than the IHS2, so of the 11,280 house-
it is unlikely that the effect of past weather is holds interviewed in IHS2, only 3,485 of them
being confounded with some underlying lived in enumeration areas that were re-
time-varying factor that affects outcomes. sampled in 2007. From these 3,485 house-
Finally, our identification strategy is en- holds, 2,968 were found for re-interview in
hanced by a number of robustness tests. First, 2007, at an attrition rate of 14.8%. The third
spatial correlation of the error terms can in- wave of data were collected as part of the
validate the standard errors of our estimates. 2009 Agricultural Inputs Support Survey II
In this case, we can only obtain reliable infer- (AISS2) conducted after the 2008/2009 grow-
ence under an appropriate choice of standard ing season by Wadonda Consultants. Of the
error clustering structure. Therefore, we esti- 2,968 households first sampled in 2003 and
mate the model under a range of clustering again in 2007, 1,642 of them lived in enumer-
structures. Second, we examine the robust- ation areas that were revisited in 2009. Of the
ness of our results by adjusting the time pe- 1,642 households in revisited areas, 1,375
riod on which past weather is defined, and were found for re-interview in 2009. This
third, we adjust the threshold based on which gives an attrition rate of 16.3% between 2007
SDD is defined. We include the data files and and 2009.
estimation codes used in this article in the on- The sample used in this analysis is based
line supplementary appendix. on the 2,968 households that were inter-
viewed in waves 1 and 2, along with 1,375
households that were interviewed in all three
Data waves. After removing unrealistic outliers
and the 185 households who experienced neg-
ative net income, we end up with an unbal-
In this section, we describe the structure of anced panel of 6,701 observations used in the
our household-level data, including attrition analysis.13
in our unbalanced panel, along with the na-
ture of historical weather information. We
also present descriptive statistics with a focus Controlling for Potential Attrition Bias
on within-household variability of key
variables. The attrition rate between each of the three
survey waves is 15% to 16%. If the house-
Household Survey Data holds in the survey attrite for non-random
reasons, this could bias the coefficient esti-
We use data from three rounds of household mates in our analysis. Fortunately, our esti-
surveys conducted in Malawi. The sampling mation strategy should be robust to most
frame is stratified random, gives broad geo- types of attrition bias, as the household FE
graphic coverage across Malawi, and is con- estimator removes any correlation between
sidered nationally representative. The first time-invariant unobservable factors that

12 13
A random effects model would be inappropriate in our con- We recognize that the IHS3 and IHS3-panel dataset created
text as it assumes that all covariates are uncorrelated with time- by the World Bank have been collected more recently than the
constant unobserved household-level factors, an assumption that data we use. However, those data have the downside of being
is unlikely to hold with observational data. Furthermore, in a lin- collected over a shorter time period (2009/10 – 2012/13), while
ear model such as ours, estimates of weather coefficients should ours span the 2002/03–2008/09 period. This provides larger vari-
be equivalent under fixed- and correlated-random effects ability in weather patterns over time, while maintaining a great
approaches. deal of cross-sectional variation.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 125

affect attrition and the covariates in our In contrast, SDD is relatively high in the first
model (Wooldridge 2010).14,15 and third waves.
So far we have discussed how measures of
central tendency of the cross-sectional distri-
Weather Data bution of households vary over waves.
Time-series data on rainfall and temperature However, an important issue given our fixed
come from the University of East Anglia’s effect estimation strategy is the degree of
Climate Research unit (CRU)-TS 3.1 within-observation variability of key past and
Climate Database (CRU 2011). These data current weather variables. Table 1 shows sub-
are interpolated using a spline function to stantial within-observation variability in
match the location where the households re- weather. As previously argued, within-
side. Specifically, we match the household- observation variability of past weather
level information in our dataset with monthly increases as the time over which averages are
rainfall and temperature totals specified at 44 calculated shortens. Values reported in table 1
locations across Malawi. Households are reveal that stress degree days and rainfall var-
assigned rainfall and temperature data iability are the most volatile measures. In
according to their spatial proximity to these particular, the within-observation coefficient
collection points. Therefore, diverse locations of variations of SDD are 6.9%, 7.3%, and
have different weather histories, providing 12% for the 20-, 10-, and 5-year periods, re-
cross-sectional variability in weather varia- spectively. For rainfall variability, these val-
bles. In addition, households experience tem- ues are 7.7%, 14.8%, and 53%, respectively.
poral variation in weather patterns over Rainfall displays limited within-observation
survey waves. The weather data are publi- variability in the 20- and 10-year specifica-
cally available, and further details can be tions (2% coefficient of variation), and more
found in Mitchell and Jones (2005).16 substantial variability in the 5-year specifica-
tion (8.2% coefficient of variation). Finally,
current weather displays substantial within-
Data Description and Summary observation variability over time.
Table 1 presents the means and medians of
the variables used in the analysis by survey
wave. The top of the table reports past Impact of Current and Past Weather
weather patterns calculated by averaging his-
torical weather for the 20, 10, and 5 years pre-
Results from the simultaneous equation esti-
ceding each wave. The table indicates a
mation of the translog system under alterna-
substantial amount of variation in cross-
tive definitions of past weather are reported
sectional mean and median of past CVGSP
in table A.1 in appendix A. After estimating
and SDD across survey waves, while at the
these parameters, we calculate the net in-
same time such variation is more limited for
come and partial derivatives of shares as de-
GSP. In terms of general trends, historic rain-
scribed in equations (4) to (7), and report the
fall variability increases slightly over survey
comparative statics with respect to the main
waves, as does SDD. Both of these trends are
specifications of past weather (10-year histor-
consistent with observations in Tadross et al.
ical average) in table 2, while results with re-
(2009), but no clear trend emerges in GSP.
spect to current growing season weather are
Cross-sectional mean and median of growing
reported in table 3. We present median val-
season weather variables also display varia-
ues, as they are less susceptible than mean
tion over waves. GSP and CVGSP are rela-
values to outliers in regressors, and their
tively low in the first and third waves.
bootstrapped standard errors.
It is likely that the independent variables
14 of interest (current and past weather) exhibit
Inverse probability weights are not valid with FE estimation
(Wooldridge 2010). spatial dependence. In such cases, either fail-
15
Recent studies by Mason and Ricker-Gilbert (2013) use the ure to control for clustered errors or a
same dataset as the present study, and both found little or no evi-
dence of attrition bias after controlling for time-constant unob-
poorly-chosen clustering structure can result
served factors. in misleading inference, such as artificially
16
The study details the sources of the raw temperature and narrow confidence intervals (Bertrand, Duflo,
rainfall data, the methods for dealing with missing data points,
and the way in which data were interpolated using a spline func- and Mullainathan 2004). The nature of our
tion, and homogenized across weather stations. data, along with administrative, market, and

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
126 January 2018 Amer. J. Agr. Econ.

Table 1. Descriptive Statistics of Variables Used in the Analysis


Survey wave Wave 1 Wave 2 Wave 3
Season (2002/2003 & 03/2004) (2006/2007) (2008/2009)
Weather variables mean median mean median mean median
20-year average GSP in mm 974 952 946 927 971 956
Within-observation variation over time 2.2%
(coefficient of variation)
20-year average SDD in degrees Celsius 78.77 38.61 86.02 44.01 77.09 22.70
Within-observation variation over time 6.9%
(coefficient of variation)
20-year average CVGSP 0.21 0.21 0.20 0.21 0.24 0.26
Within-observation variation over time 7.7%
(coefficient of variation)
10-year average GSP in mm 979 949 1002 961 1020 981
Within-observation variation over time 1.9%
(coefficient of variation)
10-year average SDD in degrees Celsius 86.66 47.02 90.65 48.09 78.24 24.67
Within-observation variation over time 7.3%
(coefficient of variation)
10-year average CVGSP 0.22 0.21 0.19 0.19 0.23 0.26
Within-observation variation over time 14.8%
(coefficient of variation)
5-year average GSP in mm 1,030 1,001 928 886 1,028 1,000
Within-observation variation over time 8.2%
(coefficient of variation)
5-year average SDD in degrees Celsius 81.19 43.38 96.84 50.40 82.09 27.60
Within-observation variation over time 12%
(coefficient of variation)
5-year average CVGSP 0.14 0.15 0.13 0.15 0.31 0.37
Within-observation variation over time 53%
(coefficient of variation)
GSP in mm 994 946 1603 1632 1102 1081
Within-observation variation over time 30.6%
(coefficient of variation)
SDD in degrees Celsius 99 63 61 27 124 84
Within-observation variation over time 21.7%
(coefficient of variation)
Coefficient of variation of within-season 0.56 0.56 0.80 0.83 0.60 0.63
monthly rainfall
Within-observation variation over time 55.5%
(coefficient of variation)
Share of area planted to improved maize 0.392 0.333 0.450 0.333 0.385 0.273
Maize output household level in kg 580 327 703 373 627 410
Real farm-gate maize price (Mk/kg) 21 20 14 13 28 30
Real total household net income (MK) 49,815 26,454 31,239 14,793 67,518 39,813
Days of agricultural labor hired in 14.06 0.00 5.49 0.00 5.13 0.00
Days of agricultural labor hired out 43.05 5.00 10.89 0.00 25.15 0.00
Real value of livestock and durable assets (MK) 33,326 10,771 47,300 10,434 56,297 13,600
Household landholding, (Ha) 1.32 0.86 1.06 0.81 1.08 0.81
¼1 if household owns livestock 0.31 0 0.31 0 0.34 0
Previous year’s real retail maize price (Mk/Kg) 23 22 22 22 40 40
Real commercial fertilizer price (Mk/Kg) 62 61 81 80 139 133
Real off-farm wage rate (Mk/Day) 179 184 207 208 282 278
Quantity of subsidized fertilizer acquired (Kg) 12.55 0.00 62.42 50.00 56.15 50.00
Quantity of subsidized improved maize seed 0.74 0.00 2.81 0.00 2.02 0.00
acquired (Kg)
Note: Real prices are in 2008/2009 Malawi Kwacha (MK); U.S. $1.00  140 Malawi Kwacha during 2008/2009.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 127

Table 2. Effect of Past Weather on Median Net Income and Netput Shares
Dependent Weather Elasticity Standard errors
variable variable
Unclustered Clustered
year region-by- district-by- district
year year
i Net income Rainfall (GSP) 1.31** (0.351) (0.508) (0.491) (0.533) (0.597)
ii Net income Rainfall 0.31*** (0.066) (0.150) (0.138) (0.094) (0.121)
variability
(CVGSP)
iii Net income Stress Degree 0.04 (0.043) (0.022) (0.113) (0.083) (0.148)
Days (SDD)
Change in net income shares with respect to Rainfall (GSP)
iv Maize output GSP 0.89*** (0.235) (0.326) (0.321) (0.304) (0.364)
v Non-maize output GSP 1.72** (0.551) (0.726) (0.716) (0.634) (0.773)
vi Inorganic fertilizer GSP 0.71 (0.344) (0.361) (0.397) (0.346) (0.505)
vii Hired labor GSP 0.01 (0.123) (0.121) (0.136) (0.160) (0.142)
viii Improved maize seed GSP 0.11 (0.109) (0.127) (0.144) (0.147) (0.210)
Change in net income shares with respect to Stress Degree Days (SDD)
ix Maize output SDD 0.03 (0.027) (0.020) (0.064) (0.050) (0.082)
x Non-maize output SDD 0.10 (0.058) (0.051) (0.125) (0.102) (0.155)
xi Inorganic fertilizer SDD 0.05 (0.032) (0.026) (0.049) (0.041) (0.052)
xii Hired labor SDD 0.03 (0.013) (0.005) (0.021) (0.017) (0.031)
xiii Improved maize seed SDD 0.01 (0.010) (0.005) (0.010) (0.015) (0.022)
Change in net income shares with respect to Rainfall Variability (CVGSP)
xiv Maize output CVGSP 0.20** (0.044) (0.073) (0.066) (0.057) (0.086)
xv Non-maize output CVGSP 0.48** (0.122) (0.144) (0.137) (0.139) (0.223)
xvi Inorganic fertilizer CVGSP 0.10 (0.079) (0.038) (0.091) (0.090) (0.139)
xvii Hired labor CVGSP 0.10** (0.038) (0.040) (0.044) (0.045) (0.055)
xviii Improved maize seed CVGSP 0.08 (0.032) (0.029) (0.041) (0.040) (0.058)
Note: Asterisk * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. Significance level
is based on standard errors obtained via bootstrapping at 200 repetitions and clustered at the district level since they provide the most conservative (highest
standard errors) estimates.

agro-environmental boundaries in Malawi, standard errors vis-a-vis unclustered ones,


suggest a number of plausible clustering suggest substantial correlation within cluster.
structures for our study. We believe errors Therefore, our results suggest the presence of
and regressors of interest may be correlated contemporaneous and non-contemporaneous
contemporaneously in areas within districts, correlation within districts. We rely on the
and even at higher levels such as districts most conservative estimates, which, for virtu-
themselves, or regions. In addition, errors ally all elasticities, are those obtained by clus-
may be serially correlated within relatively tering at the district level.
smaller geographic areas such as district. Changes in shares portray changes in the
Therefore, we estimate our main model un- relative intensity of netputs in net income.
der a range of plausible clustering structures:
cluster by year, by region and year, by district
and year, and by district.17,18 Larger clustered
time period. With our data, this would imply inverting a 6,701 by
6,701 matrix, which is infeasible with readily available computing
technology. To overcome this dimensionality issue, Belotti,
17
Serial correlation within regions is highly unlikely as many Hughes, and Mortari (2016) proposed different methods for im-
HHs within this cluster would be too far away geographically and putation of missing data to create a balanced spatial panel.
over time. However, such imputation methods come with great limitations
18
One possible approach is to adopt a spatial panel model. of their own. Another option would be to discard observations
Unfortunately, implementing a spatial panel model with our data that do not appear in all periods, thus creating a balanced panel.
is problematic. When data constitute an unbalanced panel, there However, in our case that would mean dropping 1,593 observa-
is a different spatial relationship between observations in each tions, which would imply a substantial loss of information.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
128 January 2018 Amer. J. Agr. Econ.

Table 3. Effect of Current Growing Season Weather on Median Net Income and Netput
Shares
Dependent Weather Elasticity Standard errors
variable variable
Unclustered Clustered by
year region-by- district-by- district
year year
i Net income Rainfall (GSP) 0.74* (0.206) (0.309) (0.227) (0.321) (0.467)
ii Net income Rainfall 0.17 (0.134) (0.123) (0.143) (0.163) (0.250)
variability
(CVGSP)
iii Net income Stress Degree 0.013 (0.014) (0.013) (0.019) (0.016) (0.019)
Days (SDD)
Change in net income shares with respect to Rainfall (GSP)
iv Maize output GSP 0.82*** (0.201) (0.440) 0.312) (0.235) (0.291)
v Non-maize output GSP 1.5*** (0.439) (0.850) 0.628) (0.472) (0.552)
vi Inorganic fertilizer GSP 0.51* (0.227) (0.273) 0.269) (0.206) (0.271)
vii Hired labor GSP 0.002 (0.061) (0.074) 0.068) (0.078) (0.076)
viii Improved maize seed GSP 0.14 (0.063) (0.069) 0.068) (0.077) (0.095)
Change in net income shares with respect to Stress Degree Days (SDD)
ix Maize output SDD 0.03 (0.017) (0.012) (0.013) (0.021) (0.028)
x Non-maize output SDD 0.08 (0.040) (0.032) (0.034) (0.045) (0.059)
xi Inorganic fertilizer SDD 0.07** (0.023) (0.020) (0.023) (0.023) (0.031)
xii Hired labor SDD 0.004 (0.006) (0.006) (0.006) (0.007) (0.008)
xiii Improved maize seed SDD 0.013 (0.007) (0.007) (0.008) (0.013) (0.020)
Change in net income shares with respect to Rainfall Variability (CVGSP)
xiv Maize output CVGSP 0.47* (0.174) (0.215) (0.195) (0.198) (0.273)
xv Non-maize output CVGSP 0.91* (0.027) (0.020) (0.064) (0.050) (0.082)
xvi Inorganic fertilizer CVGSP 0.49* (0.059) (0.068) (0.062) (0.068) (0.064)
xvii Hired labor CVGSP 0.07 (0.069) (0.063) (0.060) (0.086) (0.114)
xviii Improved maize seed CVGSP 0.05 (0.349) (0.373) (0.340) (0.394) (0.516)
Note: Asterisk * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. Significance level
is based on standard errors obtained via bootstrapping at 200 repetitions and clustered at the district level since they provide the most conservative (highest
standard errors) estimates.

For instance, all else being constant, net in- number of behavioral responses (rows iv-
come from maize and non-maize production xviii) that are, in turn, associated with lower
activities may both increase, but if the former net income (rows i-iii). Specifically, results
increases more relative to the latter, then the show that decreased past GSP and increased
relative intensity of maize in the household’s past CVGSP seem to induce households to
net income increases while the relative inten- rely relatively more on maize and less on
sity of non-maize activities decreases. Our other sources of income (rows iv and v, xiv
analysis focuses on changes in relative inten- and xv). The change in the relative impor-
sities because they capture the degree to tance of income sources in the face of adverse
which households’ livelihoods depend upon rainfall history suggests a re-allocation of la-
specific netputs. We should also note that if bor resources across activities. In particular,
weather has a statistically significant effect on as the expectation of weather becomes more
net income but not on the share of netput j, pessimistic, the median household seems to
this means that the expenditure on or reve- devote relatively more time to own-farm
nue from netput j changes proportionally maize cultivation. This notion is reinforced
with net income. by the fact that expenditure in hired-in labor
Results in table 2 show that adverse decreases proportionally with net income in
weather history (decreases in GSP, and response to an increased scarcity in past rain-
increases in CVGSP and SDD) prompts a fall (row vii), and more than proportionally

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 129

in response to increased variability in past season rainfall is associated with a 0.74% re-
rainfall (row xvii). Expenditures on fertilizer duction in net income (row i). Changes in
and improved varieties also seem to decrease shares reveal that households are limited in
proportionally with net income in the face of their ability to adjust their input use in re-
adverse rainfall history, as indicated by statis- sponse to unexpected changes in growing sea-
tically insignificant changes in shares in rows son rainfall (rows iv-viii). In particular,
vi, viii, xvi, and xviii. These results are consis- reduced rainfall decreases net income, and
tent with decreased expected marginal value increases the share of fertilizer, and improved
product of these inputs due to pessimistic maize varieties (though only marginally sta-
weather expectations. tistically significant) in net income. In other
Expenditure on improved maize varieties words, expenditure on these inputs does not
is reduced when households have experi- decrease proportionally with net income, sug-
enced scarce and volatile rainfall in the past. gesting a limited behavioral response. This is
This is evidenced in table 2 by statistically in- in contrast to a reduction in intensification in
significant changes in shares in rows viii and response to past rainfall, due to proportional
xviii, respectively, in combination with posi- (or more than proportional) changes in ex-
tive and negative net income elasticities in penditure. These results, along with a reduc-
rows i and ii, respectively. This result suggests tion in the share of maize in total net income,
that the agronomic benefits of improved vari- suggests that growing season rainfall has a di-
eties may not be large enough to cover their rect impact on net income through maize pro-
costs. However, we should note that we can- ductivity, rather than an indirect impact
not differentiate between hybrids and OPV; through other production decisions.
OPV seeds do not necessarily display the Conversely, the median household does seem
same drought resistance as hybrids, and to respond to increased rainfall variability by
households often adopt them for their high relying more on non-maize sources of income
yielding properties, which can be fully real- (row xv, table 3). These responses seem to
ized under more favorable weather condi- offset maize productivity losses (row xiv,
tions. Therefore, the positive link between table 3) so that within-season rainfall vari-
past rainfall and adoption of improved varie- ability does not seem to have a statistically
ties may be explained by increased adoption significant impact on net income in row ii.
of high-yielding as opposed to drought Households’ limited responsiveness to cur-
resistant varieties. rent growing season weather occurrences is
As a result of such behavioral responses, a to be expected, as decisions on input applica-
percentage point decrease in past rainfall tions are largely made before such weather
results in a 1.31% reduction in net income occurrences unfold. But the link between al-
(row i). A 1% increase in past rainfall vari- ternative income sources and rainfall vari-
ability results in a 0.31% reduction in net in- ability are somewhat different from the
come (row ii). One possible explanation for others. Rainfall variability over the current
these results is that adverse weather expecta- growing season is typically associated with
tions shrink households’ opportunities (i.e., relatively long dry spells or significant rainfall
their feasible netput set), forcing them to shocks that cause flooding. These events can
make production choices that attain a lower have substantial impacts on maize productiv-
net income. Alternatively, households may ity and income. As households observe these
still operate within a similar feasible netput events, they can adjust their decision of
set but nonetheless adopt production deci- whether to allocate more time to maize pro-
sions that provide protection to adverse duction on their own farms or not. For exam-
weather, while reducing mean net income. ple, this decision can be made during the
These decisions need not be rooted in irratio- growing season for weeding and harvesting
nal choices. Instead, they are perhaps activities. Adjusting behavior to ongoing
explained by households’ conviction that weather occurrences during the growing sea-
such reallocations reduce risk exposure son is an issue closely linked to the timing of
enough to outweigh losses in mean returns input application, and has been recognized
(Bellemare, Barrett, and Just 2013). since at least Antle (1983). These adjustments
Results in table 3 show that current grow- are in contrast to decisions that occur at the
ing season rainfall can have a substantial im- start of the season, such as deciding how
pact on a household’s net income, which is much area to allocate to improved maize
not surprising. A 1% reduction in current seeds.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
130 January 2018 Amer. J. Agr. Econ.

Results from tables 2 and 3 confirm our mismatched past weather patterns (but not
prior expectation that re-allocation of inputs “future past weather”).
and income sources are an important part of The results of the falsification test are
a smallholder household’s response to reported in table 4, and show that mis-
changes in current and past rainfall (we do matched past weather variables are not statis-
not see evidence of a significant response to tically significant in predicting the outcomes
temperature in the form of SDD). In particu- of interest. While the absence of false rela-
lar, results indicate that lower and more vari- tionships is not formal proof of a causal link
able rainfall in the past either forces or between past weather and outcomes, it does
induces households to focus attention on suggest that our analysis is not confounding
growing relatively low-value maize, using the effects of past weather with unobservable
more of their own labor, while engaging less time-varying factors, thereby solidifying our
in potentially more remunerative non-maize main results and insights.
opportunities. In contrast, adverse rainfall
conditions during the current growing season Changing Weather History Recall
push households to search for opportunities
outside maize. Results reported in table 2 are based on a
model in which past weather variables are
calculated by taking averages over the past
Robustness Checks ten years. However, the period over which
one should calculate these averages to most
In this section we test the robustness of our appropriately reflect smallholders’ expecta-
results to the possible presence of unobserv- tions remains a subject of debate. So far we
able time-varying factors, the arbitrarily- have focused our discussion on the 10-year
defined length of households’ memory of past period because it seems to strike the right
weather occurrences, and the arbitrarily-de- balance between patterns that are consistent
fined thresholds based on which SDD is enough over time to influence expectations,
computed. and a period that is short enough to display
adequate time series variability. To examine
Weather Falsification Tests the robustness of our results to alternative
specifications of historical weather, we re-
Statistical estimation of the within-household estimate elasticities under two additional
relationship between weather patterns and specifications of past weather: averages over
outcomes of interest may generate a spurious the past 5 years, and averages over the past
correlation, rather than detecting a causal as- 20 years. Results from these specifications are
sociation (Brainerd and Menon 2014). This is reported in table B.1 of Appendix B, along-
especially true when the dependent variable side those from the 10-year specification.
is behavioral in nature (Kneller and Reported values reveal that our main insights
McGowan 2012). In the context of our study, are largely robust to the length of the period
past weather may simply correlate with unob- over which historical averages are calculated.
servable changes in the underlying profitabil-
ity of specific production decisions over time. Adjusting the SDD Threshold
To examine this possibility, we follow
Brainerd and Menon (2014), and Kneller and Results in tables 2 and 3 show that our vari-
McGowan (2012) to use falsification tests to able for stress degree days (SDD) does not
change the timing of weather patterns. In par- have a statistically significant impact on
ticular, we move the independent variables smallholder behavior or net income. Our
(past weather) forward one wave. We do not main model assumes an SDD threshold of
have sufficient information on weather after 30 C. However, the threshold above which
wave 3, so we match outcomes in wave 3 with plants are particularly vulnerable to heat may
past weather from wave 1. In sum, outcomes vary by crop, variety, and region. Therefore,
in wave 1 are regressed on past weather of we re-estimate the model under alternative
wave 2, outcomes in wave 2 are regressed on thresholds for the SDD variable, and report
past weather of wave 3, and outcomes of results in table B.2 of appendix B. Our analy-
wave 3 are regressed on past weather of wave sis shows that our main insights are robust to
1. Therefore, waves 1 and 2 are run on “future reductions in the SDD threshold to 28 C,
past weather”, while wave 3 is run on and increases in the threshold to 32 C. Table

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 131

Table 4. Falsification Test: Effect of Mismatched Past Weather


Dependent Weather Elasticity with respect to Std. Err.
variable variable mismatched past weather
i Net income Rainfall (GSP) 0.61 (0.79)
ii Net income Rainfall variability (CVGSP) 0.09 (0.19)
iii Net income Stress Degree Days (SDD) 0.03 (0.11)
Change in net income shares with respect to Rainfall (GSP)
iv Maize output GSP 0.55 (0.49)
v Non-maize output GSP 1.17 (1.02)
vi Inorganic fertilizer GSP 0.72 (0.58)
vii Hired labor GSP 0.10 (0.18)
viii Improved maize seed GSP 0.01 (0.17)
Change in net income shares with respect to Stress Degree Days (SDD)
ix Maize output SDD 0.03 (0.07)
x Non-maize output SDD 0.12 (0.14)
xi Inorganic fertilizer SDD 0.06 (0.06)
xii Hired labor SDD 0.03 (0.03)
xiii Improved maize seed SDD 0.01 (0.02)
Change in net income shares with respect to Rainfall Variability (CVGSP)
xiv Maize output CVGSP 0.11 (0.12)
xv Non-maize output CVGSP 0.38 (0.28)
xvi Inorganic fertilizer CVGSP 0.18 (0.15)
xvii Hired labor CVGSP 0.06 (0.05)
xviii Improved maize seed CVGSP 0.02 (0.06)
Note: Asterisk * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. Significance level
is based on standard errors obtained via bootstrapping at 200 repetitions and clustered at the district level.

B.2 shows that across thresholds, higher SDD Wealth and Behavioral Responses to Past
in the current growing season leads to a reduc- Weather Patterns
tion in inorganic fertilizer application but has
no discernible impact on net income. All other Results in table 2 are important as they un-
potential impacts with respect to current and derscore the fact that, while households can
past SDD are statistically insignificant. and do respond to changes in weather pat-
These results appear to be inconsistent terns, adaptation strategies can only partially
with several previous studies indicating a neg- alleviate their adverse effects for smallhold-
ative and strong effect of heat stress on agri- ers in Malawi. However, focusing attention
cultural productivity (e.g., Schlenker and on the median household may mask substan-
Roberts 2006; Schlenker and Lobell 2010). tial heterogeneity in behavioral responses, as
However, as pointed out by Carter et al. well as the magnitude of losses associated
(2016), further investigation finds that a high with adverse climatic patterns. As a result, we
air temperature is indicative of high water examine the extent to which relatively weal-
evaporation. This implies that negative thy households, defined as those who have a
effects of temperature on maize productivity higher value of livestock and durable assets,
found by Schlenker and co-authors likely re- have greater adaptation capabilities. This
flect underlying soil moisture limitations in- means that they are more able to protect
stead of direct effects from heat stress. In their net income against the harmful effects
other words, the effects of high temperature of adverse changes in weather patterns.
tend to vanish when heat stress correlates Figure 1.a displays the elasticity of net in-
positively with rainfall. Exploration of our come with respect to past weather and
data reveals a positive correlation between figure 1.b displays it with respect to current
SDD and rainfall. This, in combination with weather, along the households’ wealth spec-
the arguments advanced by Carter et al. trum. These relationships are calculated by
(2016), offers a plausible explanation for the simulating equations (4) and (6) for the range
lack of a discernible effect of SDD on income of asset values observed in our sample. The
and most netputs. figures show that net income elasticities with

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
132 January 2018 Amer. J. Agr. Econ.

Rainfall Rainfall Variability Rainfall Stress Degree Days


Stress Degree Days Rainfall Variability
2.0
2.0

Elascity of net income with respect to


Elascity of net income with respect to

1.5 1.5

growing season weather


long term weather

1.0 1.0

0.5 0.5

0.0 0.0

10.8

12.0

13.3
5.2

5.8

6.4

7.1

7.9

8.8

9.7
10.8
12.0
13.3
5.2
5.8
6.4
7.1
7.9
8.8
9.7

-0.5 -0.5

-1.0 -1.0
Assets Assets
(logarithm of livestock and durable goods) (logarithm of livestock and durable goods)

Figure 1a. Past weather and net income Figure 1b. Current growing season weather and net
income

Figure 1. Net income effects of past and current weather along wealth spectrum

respect to rainfall and rainfall variability de- In either case, our findings speak to the possi-
crease in absolute value as the value of a ble existence of a “climate-induced” poverty
household’s assets increase. This indicates trap that locks poor smallholder households
that relatively poorer households are less into low value maize cultivation from season
able to buttress themselves from the negative to season.
ramifications of more sporadic rainfall, both In order to further understand the differen-
past and current. These results are consistent ces in income diversification strategies among
with those in Dercon and Christaensen households with different levels of wealth,
(2011) and FAO (2016a). table 5 displays how sources of household crop
Regarding the mechanisms underlying in- income and income shares vary across wealth
creased resilience to unreliable rainfall, groups. The results from this disaggregation
results show that richer households can better also support the idea of a possible “climate-
diversify away from maize and towards non- induced” poverty trap, as the poorest 20% of
maize activities (or at the very least, alleviate smallholders derive a higher share of crop in-
increased reliance on maize) in response to come from maize (66%) compared to the rich-
adverse weather, especially past weather (fig- est 20% of smallholders (61%). Although
ures 2.a and 2.b). Unfortunately, we are lim- maize has a relatively high share for most
ited in our ability to identify whether this Malawian smallholders, the richest households
increased reliance on maize production of seem to be relatively more diversified into the
poorer households is due to preferences (risk production of other, more highly-valued crops
aversion) or constraints. It is possible that such as tobacco, groundnuts, and/or cotton.
households adopt a safety-first approach by In addition, and perhaps more importantly,
which they allocate more resources to subsis- the poorest quintile of smallholder house-
tence maize cultivation to ensure their own holds derive a much higher share of their to-
food security, rather than engaging in other tal income from working as agricultural
activities that may provide better remunera- laborers on other people’s farms (20%) com-
tion, on average. It is also possible that pared to the richest quintiles of households
households search for non-maize income op- who only earn 5% of their income from
portunities in the face of changing climatic
patterns, but that these opportunities, such as
working as a laborer on another farm, are es- 19
This seems like a particularly relevant subject for further
pecially limited for asset-poor households.19 investigation.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 133

Rainfall Rainfall Variability Rainfall Stress Degree Days


Stress Degree Days Rainfall Variability
3.0 2.0
Derivave of non-maize income share

with respect to growing season weather


2.5
with respect to long term weather

Derivave of non-maize income share


1.5
2.0
1.0
1.5
1.0 0.5

0.5 0.0

10.8
5.2

5.8

6.4

7.1

7.9

8.8

9.7
0.0 -0.5

10.8
5.2

5.8

6.4

7.1

7.9

8.8

9.7
-0.5
-1.0
-1.0
-1.5 -1.5
Assets -2.0
(logarithm of livestock and durable goods) Assets
(logarithm of livestock and durable goods)

Figure 2a. Past weather and non-maize income Figure 2b. Growing season weather and non-maize
income

Figure 2. Non-maize income effect of past and current weather along wealth spectrum

Table 5. Mean Income and Income Share by Source, Across Asset Quintiles
asset Share of maize Share of ag. Share of non- Value of
quintile income on total labor income on farm income in household
crop income total income total income6 assets
Poorest 1 0.66*** 0.20*** 0.28*** 1,405***
2 0.65*** 0.18*** 0.28*** 5,114***
3 0.62 0.13*** 0.30** 11,808***
4 0.62 0.08*** 0.31 25,078***
Richest 5 0.61 0.05 0.33 183,171
Total 0.63 0.13 0.30 44,538
Note: Value of assets are in real 2009 Malawi Kwacha; U.S. $1.00 ¼140 Malawi Kwacha during period of study; 6 indicates that non-farm income does not in-
clude agricultural labor income; asterisks *, **, and *** indicate that the mean for the corresponding asset quintile is statistically different from the richest as-
set quintile at the 10%, 5%, and 1% significance levels, respectively, obtained using Bonferroni multiple-comparison test.

working as agricultural laborers. The relative directly affected by changing weather pat-
dependence on maize production and agricul- terns over time.
tural wage income leaves the poorest house-
holds in Malawi doubly vulnerable to
worsening weather patterns through the di- Study Limitations
rect negative production impacts on their
own farms, and through the likely reductions The results presented in this article have sev-
in agricultural labor opportunities on other eral caveats. First, data with a longer time-
farms, which themselves are negatively af- series would greatly enhance the reliability of
fected by worsening weather. Conversely, the the relationships we uncovered between past
richest quintile of households earns 33% of weather and net income by eliminating or re-
its income from non-farm sources, compared ducing the overlap in periods over which past
to 28% for the poorest quintile of house- weather is averaged. Second, the perfor-
holds. This category of income excludes agri- mance of maize is highly dependent on the
cultural labor income, and includes salaried timing of rainfall relative to its phase in the
employment, business, and investment in- growing cycle (Tadross et al. 2009). Since our
come, all of which are much less likely to be rainfall data only includes monthly totals, its

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
134 January 2018 Amer. J. Agr. Econ.

coarseness prevents us from drawing any con- this issue deserves attention in future re-
clusions about these agronomic issues. search as it is of particular importance to
Moreover, our weather information is spa- evaluate the relative merits of future policy
tially coarse. As pointed out by two anony- responses to climate change. If in fact
mous reviewers, using high resolution changes in weather patterns affect behavior
weather data would be preferred to interpola- only through changes in the probability distri-
tions of data from spatially-sparse weather bution of prices, price-stability policies may
stations. The increasing availability of high- constitute an effective tool to alleviate the
resolution spatial data creates a range of re- effects of adverse weather trends.
search opportunities in the future. Finally, Our results are also consistent with the
our results point to the need for more re- presence of market barriers to adaptation as
search to elucidate the mechanisms by which the elasticity of net income with respect to
assets steer behavioral responses and welfare weather history is sensitive to asset owner-
outcomes, which may inform identification of ship. Interestingly, households with lower as-
promising policy alternatives. set levels tend to rely more on maize-related
income as past weather patterns become
more adverse compared to their wealthier
counterparts. The positive link between
Conclusions households’ wealth and their resilience to ad-
verse weather patterns could be explained by
The present article estimates how small- a more dramatic reduction in non-maize op-
holder farm households in Malawi respond to portunities for poorer households. The rela-
changes in current and past weather patterns. tive dependence on maize income in the face
While much of the previous literature on be- of adverse weather could also be explained
havioral responses to changing weather pat- by a deliberate choice on the part of poorer
terns considers the adoption of entirely new households. Maize is by far the main staple
farming practices, little quantitative work has crop grown in Malawi, so the expectation of
been undertaken to determine the relevance adverse weather patterns may encourage
of simpler methods of adaptation such as poorer farmers to devote more resources into
adjusting the use of existing inputs or reallo- maize production, perhaps in an attempt to
cating resources across alternative income ensure their own food security. But regard-
sources. Employing a structural model of ad- less of the cause, these results suggest that
aptation that controls for economic and pol- wealthier households have access to a greater
icy factors, we examine the relevance of these diversity of earning opportunities as crop pro-
adaptation mechanisms and the importance duction becomes riskier.
of assets in facilitating these types of adjust- Our finding that poorer households tend to
ments and their distributional impacts in rely more on maize production in response to
Malawi. adverse changes in climate suggests that there
We find that the expectation of adverse may be a “climate induced” poverty trap that
weather patterns prompts households to in- locks poor smallholder households into low-
crease reliance on low-value maize cultiva- value maize cultivation from season to sea-
tion, while forgoing other potentially more son. Our results are consistent with Dercon
remunerative opportunities elsewhere. and Christiaensen (2011), who find Ethiopian
Nonetheless, we caution the reader against smallholder households with more livestock
attributing these reallocations to changes in are more likely to apply inorganic fertilizer in
expected weather alone. Changes in weather the face of rainfall risk, allowing them to
patterns can send ripple effects through mar- avoid what they call a “risk induced” poverty
kets and, ultimately, change expected price trap that captures more limited resource
patterns. For example, Bellemare, Barrett, farmers. In total, these results provide further
and Just (2013) find that changes in expected support for the concern that adverse changes
price patterns can prompt behavioral in climate are regressive in nature, dispropor-
responses, and that such responses can result tionally affecting the poorest and most vul-
in reduced profits as long as they sufficiently nerable households, and raising the need for
reduce risk exposure. While our framework is more focused intervention from
of limited usefulness in disentangling these policymakers.
market-mediating effects from the direct In addition, these insights underscore the
effects of weather expectations, we believe need for further research into better maize

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 135

varieties that produce more stable yields un- 2014. Climate Variability, Adaptation
der inconsistent moisture conditions. This Strategies and Food Security in Malawi.
finding is consistent with other recent studies ESA Working Paper: 14–08.
(e.g., Kurukulasuriya and Mendelsohn 2008; Auffhammer, M., and W. Schlenker. 2014.
FAO 2016b), and it also suggests that policies Empirical Studies on Agricultural
facilitating income diversification mechanisms Impacts and Adaptation. Energy
for limited resource households can help alle- Economics 46: 555–61.
viate the “climate-induced” poverty trap pre- Bellemare, M.F., C.B. Barrett, and D.R. Just.
viously discussed. Such policies may include 2013. The Welfare Impacts of
mechanisms for consumption smoothing, fi- Commodity Price Volatility: Evidence
nancial support for livestock activities, or from Rural Ethiopia. American Journal
training for off-farm employment opportuni- of Agricultural Economics 95 (4): 877–99.
ties. Ultimately, as climatic conditions worsen, Belotti, F., G. Hughes, and A.P. Mortari.
none of these strategies may be sufficient to 2016. Spatial panel data models using
assist the most limited-resource farmers in es- Stata. Stata Journal 17 (1): 139–180.
caping their poverty trap. In such a case, the Below, T., A. Artner, R. Siebert, and S.
most viable longer-term policy may be to help Sieber. 2010. Micro-level Practices To
these people migrate out of agriculture, and Adapt To Climate Change for African
move into non-farm activities that have their Small-Scale Farmers. Washington DC:
own challenges but provide more stable in- International Food Policy Research
come in the face of adverse weather. Institute Discussion Paper 953.
Benhin, J.K.A. 2006. Climate Change and
South African Agriculture: Impacts and
Adaptation Options, Vol. 21. Centre for
Supplementary material Environmental Economics and Policy in
Africa discussion paper.
Supplementary material are available at Benjamin, D. 1992. Household Composition,
American Journal of Agricultural Economics Labor Markets, and Labor Demand:
online. Testing for Separation in Agricultural
Household Models. Econometrica 60 (2):
287–322.
References Bertrand, M., E. Duflo, and S. Mullainathan.
2004. How Much Should We Trust
Adams, R.M., B.A. McCarl, D.J. Dudek, and Differences-in-Differences Estimates?
J.D. Glyer. 1988. Implications of Global Quarterly Journal of Economics 119 (1):
Climate Change for Western 249–75.
Agriculture. Western Journal of Brainerd, E., and N. Menon. 2014. Seasonal
Agricultural Economics, 13 (2): 348–56. Effects of Water Quality: The Hidden
Adams, R.M. 1989. Global Climate Change Costs of The Green Revolution to Infant
and Agriculture: An Economic and Child Health in India. Journal of
Perspective. American Journal of Development Economics 107: 49–64.
Agricultural Economics 71 (5): 1272–9. Burke, M.B., and K. Emerick. 2016.
Alwang, J., and P.B. Siegel. 1999. Labor Adaptation to Climate Change: Evidence
Shortages on Small Landholdings in from U.S. Agriculture. American
Malawi: Implications for Policy Reforms. Economic Journal: Economic Policy
World Development 27 (8): 1461–75. 8 (2): 106–40.
Antle, J.M. 1983. Sequential Decision Making Carter, E.K., J. Melkonian, S.J. Riha, and
in Production Models. American Journal S.B. Shaw. 2016. Separating Heat Stress
of Agricultural Economics 65 (2): 282–90. from Moisture Stress: Analyzing Yield
Arslan, A., N. McCarthy, L. Lipper, S. Response to High Temperature in
Asfaw, A. Cattaneo, and M. Kokwe. Irrigated Maize. Environmental Research
2015. Climate Smart Agriculture? Letters 11 (9): 094012.
Assessing the Adaptation Implications in Climate Research Unit, University of East
Zambia. Journal of Agricultural Anglia. 2011. CRU-TS 3.1 Climate
Economics 66 (3): 753–80. Database. CRU Time Series (TS) High
Asfaw, S., N. McCarthy, L. Lipper, A. Resolution Gridded Datasets, NCAS
Arslan, A. Cattaneo, and M. Kachulu. British Atmospheric Data Centre.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
136 January 2018 Amer. J. Agr. Econ.

Available at: http://badc.nerc.ac.uk/view/ Adaptation Strategies for Global


badc.nerc.ac.uk__ATOM__dataent_ Change 2 (1): 19–44.
1256223773328276. Easterling, W.E., N.J. Rosenberg, K.M.
Chirwa, E.W., and A. Dorward. 2013. Lemon, and M.S. McKenney. 1992.
Agricultural Input Subsidies. The Recent Simulation of Crop Responses to
Malawi Experience. Oxford UK: Oxford Climate Change: Effects with Present
University Press. Technology and Currently Available
Collier, P., and S. Dercon. 2014. African Adjustments (The “Smart Farmer”
Agriculture in 50 Years: Smallholders in Scenario). Agricultural and Forest
a Rapidly Changing World?. World Meteorology 59: 75–102.
Development 63: 92–101. Ecker, O. 2009. Economics of Micronutrient
Dercon, S., and L. Christiaensen. 2011. Malnutrition: The Demand for Nutrients
Consumption Risk, Technology in Sub-Saharan Africa. In
Adoption and Poverty Traps: Evidence Development Economics and Policy, vol.
from Ethiopia. Journal of Development 64, ed. F. Heidhues, J. von Braun, and M.
Economics 96 (2): 159–73. Zeller. Frankfurt, Germany: Peter Lang.
Deressa, T.T. 2007. Measuring the Economic Fezzi, C., and I. Bateman. 2015. The impact of
Impact of Climate Change on Ethiopian climate change on agriculture: Nonlinear
Agriculture: Ricardian Approach. effects and aggregation bias in Ricardian
Washington DC: World Bank Policy models of farmland values. Journal of the
Research Working Paper 4342. Association of Environmental and
Deressa, T.T., R.M. Hassan, C. Ringler, T. Resource Economists 2 (1): 57–92.
Alemu, and M. Yesuf. 2009. Food and Agriculture Organization of the
Determinants of Farmers’ Choice of United Nations. 2016a. Welfare Impacts
Adaptation Methods To Climate Change of Climate Shocks: Evidence from
in the Nile Basin of Ethiopia. Global Tanzania. ESA Working Paper No. 16-
Environmental Change 19: 248–55. 04. Rome: FAO. Available at: http://
Deressa, T.T., R.M. Hassan, and C. Ringler. www.fao.org/3/a-i6000e.pdf.
2011. Perception of and Adaptation To ———. 2016b. The state of food and agricul-
Climate Change by Farmers in the Nile ture. Vol. 37. Rome: Food and
basin of Ethiopia. The Journal of Agriculture Organization of the United
Agricultural Science 149: 23–31. Nations. Available at: http://www.fao.
Deschenes, O., and M. Greenstone. 2007. org/3/a-i6030e.pdf.
The Economic Impacts of Climate Fosu-Mensah, B.Y., P.L.G. Vlek, and D.S.
Change: Evidence from Agricultural MacCarthy. 2012. Farmer’s Perception
Output and Random Fluctuations in and Adaptation to Climate Change: A
Weather. American Economic Review Case Study of Sekyedumase District in
97: 354–85. Ghana. Environment Development and
Di Falco, S., M. Veronesi, and M. Yesuf. Sustainability 14: 495–5.
2011. Does Adaptation to Climate Government of Malawi. 2008. Population and
Change Provide Food Security? A Housing Census 2008. Zomba: National
Micro-Perspective from Ethiopia. Statistical Office.
American Journal of Agricultural ———. 2012. Integrated Household Survey
Economics 93 (2): 1–18. 2010-11. Household Socio-Economic
Di Falco, S., and M. Veronesi. 2013. How Characteristics Report. National
African Agriculture Can Adapt to Statistical Office, Zomba.
Climate Change? A Counterfactual Hassan, R., and C. Nhemachena. 2008.
Analysis from Ethiopia. Land Economics Determinants of African farmers’
89 (4): 743–66. Strategies for Adapting to Climate
Dinar, A., R. Hassan, R. Mendelsohn, and J. Change: Multinomial Choice Analysis.
Benhin. 2008. Climate Change and African Journal of Agricultural and
Agriculture in Africa: An Assessment of Resource Economics 2: 83–104.
Impacts and Adaptations. Earthscan. Henderson, J.V., A. Storeygard, and U.
Downing, T.E., L. Ringius, M. Hulme, and D. Deichmann. 2017. Has Climate Change
Waughray. 1997. Adapting to Climate Driven Urbanization in Africa? Journal
Change in Africa. Mitigation and of Development Economics 124: 60–82.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 137

Jayne, T.S., N. Mason, R. Myers, J. Ferris, D. Database of Monthly Climate


Mather, M. Beaver, N. Lenski, A. Observations and Associated High-
Chapoto, and D. Boughton. 2010. Resolution Grids. International Journal
Patterns and Trends in Food Staples of Climatology 25: 693–712.
Markets in Eastern and Southern Africa: Molua, E.L. 2002. Climate Variability,
Toward the Identification of Priority Vulnerability and Effectiveness of Farm-
Investments and Strategies for Level Adaptation Options: The
Developing Markets and Promoting Challenges and Implications for Food
Smallholder Productivity Growth. Security in Southwestern Cameroon.
Michigan State University International Environment and Development
Development Working Paper, East Economics 7 (3): 529–45.
Lansing, MI. United States. ———. 2011. Climate Change and African
Kaylen, M.S., J.W. Wade, and D.B. Frank. Agriculture: Review of Impact and
1992. Stochastic Trend, Weather and Adaptation Choices, Chapter 10. In
U.S. Corn Yield Variability. Applied Africa in Focus: Governance in the 21st
Economics 24 (5): 513–8. Century, eds. Kwandiwe Kondlo and
Kneller, R., and D. McGowan. 2012. Tax Chinenyengozi Ejiogu, 659–707.
Policy and Firm Entry and Exit Location: HSRC Press.
Dynamics: Evidence from OECD Niang, I., O.C. Ruppel, M.A. Abdrabo, A.
Countries. University of Nottingham Essel, C. Lennard, J. Padgham, and P.
Discussion Papers in Economics 12/01. Urquhart. Chapter 22. Africa. In Climate
Kurukulasuriya, P., and R. Mendelsohn. Change 2014: Impacts, Adaptation, and
2008. Crop Switching as a Strategy for Vulnerability. Part B: Regional Aspects.
Adapting to Climate Change. African Contribution of Working Group II to the
Journal of Agricultural and Resource Fifth Assessment Report of the
Economics 2 (1): 105–26. Intergovernmental Panel on Climate
Lunduka, R., M. Fisher, and S. Snapp. 2012. Change, ed. V.R., Barros, C.B. Field,
Could Farmer Interest in Diversity of D.J. Dokken, M.D. Mastrandrea, K.J.
Seed Attributes Explain Adoption Mach, T.E. Bilir, M. Chatterjee, K.L.
Plateaus for Modern Varieties in Ebi, Y.O. Estrada, R.C. Genova, B.
Malawi? Food Policy 37 (5): 504–10. Girma, E.S. Kissel, A.N. Levy, S.
Lunduka, R., J. Ricker-Gilbert, and M. MacCracken, P.R. Mastrandrea, and L.L.
Fisher. 2013. The Consistent and White, 1199–1265. Cambridge, United
Contrasting Results of Research Kingdom, and New York, NY
Examining the Economic Impacts of Cambridge University Press.
Malawi’s Fertilizer Subsidy Program. Nkonya, E., T. Schroeder, and D. Norman.
Agricultural Economics 44 (6): 563–79. 1997. Factors Affecting Adoption of
Maddison, D. 2007. The Perception of and Improved Maize Seed and Fertilizer in
Adapation to Climate Change in Africa. Northern Tanzania. Journal of
World Bank Policy Research Working Agricultural Economics 48 (1): 1–12.
Paper, 4308. Washington DC: The World Ogunlade, I., G. Olaoye, D. Tologbonse, and
Bank. O.E. Ayinde. 2010. On-Farm Evaluation
Mason, N.M., and J. Ricker-Gilbert. 2013. of Drought-tolerant Maize Varieties and
Disrupting Demand for Commercial Hybrids in the Southern Guinea
Seed: Input Subsidies in Malawi and Savanna Zones of Nigeria. Paper pre-
Zambia. World Development 45I: 75–91. sented at the South African Society for
Mendelsohn, R., W. Nordhaus, and D. Shaw. Agricultural Extension Conference, The
1994. The Impact of Global Warming on Impact of Climate Change on
Agriculture: A Ricardian Analysis. The Extension Delivery in South Africa,
American Economic Review 84 (4): 753–71. Langebaan, Western Cape, South Africa,
Mertz, O., C. Mbow, A. Reenberg, and A. May 4–7.
Diouf. 2009. Farmers’ Perceptions of Pauw, K., J. Thurlow, and D. van Seventer.
Climate Change and Agricultural 2010. Droughts and Floods in Malawi:
Adaptation Strategies in Rural Sahel. Assessing the Economywide Effects.
Environmental Management 43: 804–16. IFPRI Discussion Paper 00962.
Mitchell, T., and P. Jones. 2005. An Rao, K.P.C., W.G. Ndegwa, K. Kizito, and A.
Improved Method of Constructing a Oyoo. 2011. Climate Variability and

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
138 January 2018 Amer. J. Agr. Econ.

Change: Farmer Perceptions and Shewmake, S. 2008. Vulnerability and the


Understanding of Intra-Seasonal Impact of Climate Change in South
Variability in Rainfall and Associated Africa’s Limpopo River Basin. Washington
Risk in Semi-Arid Kenya. Experimental DC: IFPRI Discussion Paper 00804.
Agriculture 47 (2): 267–91. Silvestri, S., E. Bryan, C. Ringler, M.
RMSI. 2009. Malawi: Economic Herrero, and B. Okoba. 2012. Climate
Vulnerability and Disaster Risk Change Perception and Adaptation of
Assessment. Draft final report prepared Agro-Pastoral Communities in Kenya.
for the World Bank, August. Mimeo. Regional Environmental Change 12 (4):
Rosenzweig, C., and M.L. Parry. 1994. 791–802.
Potential Impact of Climate Change on Smale, M., P. Heisey, and H. Leathers. 1995.
World Food Supply. Nature 367: 133–8. Maize of the Ancestors and Modern
Sanghi, A., and R. Mendelsohn. 2008. The Varieties: The Micro-Economics of
Impacts of Global Warming on Farmers High-yielding Variety Adoption in
in Brazil and India. Global Malawi. Economic Development and
Environmental Change 18 (4): 655–65. Cultural Change 43 (2): 351–68.
Schlenker, W., and M. Roberts. 2006. Tadross, M., P. Suarez, A. Lotsch, S.
Nonlinear Effects of Weather on Corn Hachigonta, M. Mdoka, L. Unganai, F.
Yields. Applied Economic Perspectives Lucio, D. Kamdonyo, and M.
and Policy 28 (3): 391–8. Muchinda. 2009. Growing-season
Schlenker, W., and D. Lobell. 2010. Robust Rainfall and Scenarios of Future
Negative Impacts of Climate Change on Change in Southeast Africa:
African Agriculture. Environmental Implications for Cultivating Maize.
Research Letters 5 (1): 014010 Climate Research 40: 147–61.
Seo, N. 2010a. Is an Integrated Farm More Thomas, D.S.G., C. Ywyman, H. Osbahr, and
Resilient against Climate Change? A B. Hewitson. 2007. Adaptation to
Microeconomic Analysis of Portfolio Climate Change and Variability: Farmer
Diversification in African Agriculture. Responses to Intro-Seasonal
Food Policy 35: 32–40. Precipitation Trends in South Africa.
———. 2010b. A Microeconometric Analysis Climatic Change 83: 301–22.
of Adapting Portfolios To Climate Wang, J., R. Mendelsohn, A. Dinar, J.
Change: Adoption of Agricultural Systems Huang, S. Rozelle, and L. Zhang. 2009.
in Latin America. Applied Economic The Impact of Climate Change on
Perspectives and Policy 32 (3): 489–514. China’s Agriculture. Agricultural
Seo, N., and R. Mendelsohn. 2008. An Economics 40: 323–37.
Analysis of Crop Choice: Adapting To Wooldridge, J.M. 2010. Econometric Analysis
Climate Change in South American of Cross Section and Panel Data. 2nd
Farms. Ecological Economics 67: 109–16. Edition. Cambridge MA: MIT Press.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 139

Appendix A

Table A.1. in this appendix reports parameter estimates from the fixed-effect SUR model,
with standard errors clustered at the district level.

Table A.1 Estimated Parameters of the Structural Model


Variable Coefficient Clustered
(at district level)
standard errors
log(maize price) 22.645 12.663
log(fertilizer price) 25.717 16.949
log(labor price) 3.161 11.199
log(improved variety price) 13.217 9.446
log(maize price)*log(maize price) 0.052 0.051
log(maize price)*log(fertilizer price) 0.036 0.062
log(maize price)*log(labor price) 0.006 0.026
log(maize price)*log(improved variety price) 0.045 0.044
log(fertilizer price)*log(fertilizer price) 0.030 0.035
log(fertilizer price)*log(labor price) 0.009 0.033
log(fertilizer price)*log(improved variety price) 0.031 0.029
log(labor price)*log(labor price) 0.006 0.006
log(maize price)*log(improved variety price) 0.006 0.007
log(improved variety price)*log(improved variety price) 0.005 0.004
log(maize price)*log(current growing degree days) 0.446 0.474
log(maize price)*log(current stress degree days) 0.054 0.146
log(maize price)*log(current rainfall) 6.558 1.983
log(maize price)*log(current rainfall variability) 4.962 1.841
log(maize price)*log(assets) 2.291 0.686
log(maize price)*log(landholding) 1.411 1.409
log(maize price)*log(subsidized fertilizer quantity) 0.005 0.002
log(maize price)*log(age of HH head) 0.338 2.814
log(maize price)*cattle ownership 0.048 0.041
log(maize price)*log(subsidized improved seed) 0.000 0.000
log(maize price)*log(past rainfall) 5.351 2.878
log(maize price)*log(past growing degree days) 1.154 0.740
log(maize price)*log(past rainfall variability) 1.420 0.385
log(maize price)*log(past stress degree days) 0.100 0.208
log(fertilizer price)*log(current growing degree days) 2.588 0.843
log(fertilizer price)*log(current stress degree days) 0.042 0.239
log(fertilizer price)*log(current rainfall) 7.409 1.727
log(fertilizer price)*log(current rainfall variability) 1.847 2.447
log(fertilizer price)*log(assets) 2.100 1.149
log(fertilizer price)*log(landholding) 2.425 1.974
log(fertilizer price)*log(subsidized fertilizer quantity) 0.006 0.002
log(fertilizer price)*log(age of HH head) 1.004 3.840
log(fertilizer price)*cattle ownership 0.047 0.050
log(fertilizer price)*log(subsidized improved seed) 0.000 0.001
log(fertilizer price)*log(past rainfall) 5.649 3.617
log(fertilizer price)*log(past growing degree days) 3.302 1.287
log(fertilizer price)*log(past rainfall variability) 1.112 0.741
log(fertilizer price)*log(past stress degree days) 0.147 0.431
log(labor price)*log(current growing degree days) 0.756 0.202
log(labor price)*log(current stress degree days) 0.079 0.089
log(labor price)*log(current rainfall) 1.326 0.316
log(labor price)*log(current rainfall variability) 2.111 0.573
log(labor price)*log(assets) 0.098 0.256
Continued

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
140 January 2018 Amer. J. Agr. Econ.

Table A.1 continued


Variable Coefficient Clustered
(at district level)
standard errors
log(labor price)*log(landholding) 0.582 0.356
log(labor price)*log(subsidized fertilizer quantity) 0.001 0.001
log(labor price)*log(age of HH head) 0.337 2.254
log(labor price)*cattle ownership 0.060 0.014
log(labor price)*log(subsidized improved seed) 0.000 0.000
log(labor price)*log(past rainfall) 0.771 1.125
log(labor price)*log(past growing degree days) 0.381 0.447
log(labor price)*log(past rainfall variability) 0.175 0.381
log(labor price)*log(past stress degree days) 0.086 0.147
log(improved seed price)*log(current growing degree days) 0.392 0.256
log(improved seed price)*log(current stress degree days) 0.018 0.053
log(improved seed price)*log(current rainfall) 0.137 0.726
log(improved seed price)*log(current rainfall variability) 0.810 0.748
log(improved seed price)*log(assets) 0.827 0.397
log(improved seed price)*log(landholding) 0.300 0.262
log(improved seed price)*log(subsidized fertilizer quantity) 0.001 0.001
log(improved seed price)*log(age of HH head) 1.477 1.882
log(improved seed price)*cattle ownership 0.008 0.015
log(improved seed price)*log(subsidized improved seed) 0.000 0.000
log(improved seed price)*log(past rainfall) 1.063 1.427
log(improved seed price)*log(past growing degree days) 0.187 0.251
log(improved seed price)*log(past rainfall variability) 0.073 0.343
log(improved seed price)*log(past stress degree days) 0.017 0.072
log(maize price)*log(current growing degree days)*log(assets) 0.041 0.022
log(maize price)*log(current growing degree days)*log(landholding) 0.090 0.043
log(maize price)*log(current growing degree days)*log(age of HH head) 0.031 0.127
log(maize price)*log(current stress degree days)*log(assets) 0.012 0.008
log(maize price)*log(current stress degree days)*log(landholding) 0.026 0.021
log(maize price)*log(current stress degree days)*log(age of HH head) 0.035 0.036
log(maize price)*log(current rainfall)*log(assets) 0.154 0.088
log(maize price)*log(current rainfall)*log(landholding) 0.052 0.109
log(maize price)*log(current rainfall)*log(age of HH head) 1.555 0.634
log(maize price)*log(current rainfall variability)*log(assets) 0.076 0.063
log(maize price)*log(current rainfall variability)*log(landholding) 0.030 0.138
log(maize price)*log(current rainfall variability)*log(age of HH head) 1.238 0.547
log(maize price)*log(past rainfall)*log(assets) 0.090 0.079
log(maize price)*log(past rainfall)*log(landholding) 0.093 0.114
log(maize price)*log(past rainfall)*log(age of HH head) 1.878 0.888
log(maize price)*log(past growing degree days)*log(assets) 0.026 0.026
log(maize price)*log(past growing degree days)*log(landholding) 0.076 0.032
log(maize price)*log(past growing degree days)*log(age of HH head) 0.177 0.208
log(maize price)*log(past rainfall variability)*log(assets) 0.009 0.015
log(maize price)*log(past rainfall variability)*log(landholding) 0.081 0.037
log(maize price)*log(past rainfall variability)*log(age of HH head) 0.296 0.093
log(maize price)*log(past stress degree days)*log(assets) 0.001 0.005
log(maize price)*log(past stress degree days)*log(landholding) 0.026 0.009
log(maize price)*log(past stress degree days)*log(age of HH head) 0.016 0.051
log(fertilizer price)*log(current growing degree days)*log(assets) 0.112 0.044
log(fertilizer price)*log(current growing degree days)*log(landholding) 0.145 0.063
log(fertilizer price)*log(current growing degree days)*log(age of HH 0.387 0.185
head)
log(fertilizer price)*log(current stress degree days)*log(assets) 0.007 0.014
log(fertilizer price)*log(current stress degree days)*log(landholding) 0.057 0.019
log(fertilizer price)*log(current stress degree days)*log(age of HH head) 0.008 0.055
log(fertilizer price)*log(current rainfall)*log(assets) 0.193 0.201
Continued

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 141

Table A.1 continued


Variable Coefficient Clustered
(at district level)
standard errors
log(fertilizer price)*log(current rainfall)*log(landholding) 0.128 0.214
log(fertilizer price)*log(current rainfall)*log(age of HH head) 1.611 0.724
log(fertilizer price)*log(current rainfall variability)*log(assets) 0.071 0.124
log(fertilizer price)*log(current rainfall variability)*log(landholding) 0.415 0.213
log(fertilizer price)*log(current rainfall variability)*log(age of HH head) 0.453 0.712
log(fertilizer price)*log(past rainfall)*log(assets) 0.063 0.170
log(fertilizer price)*log(past rainfall)*log(landholding) 0.564 0.297
log(fertilizer price)*log(past rainfall)*log(age of HH head) 1.844 1.077
log(fertilizer price)*log(past growing degree days)*log(assets) 0.068 0.055
log(fertilizer price)*log(past growing degree days)*log(landholding) 0.013 0.080
log(fertilizer price)*log(past growing degree days)*log(age of HH head) 0.666 0.284
log(fertilizer price)*log(past rainfall variability)*log(assets) 0.038 0.038
log(fertilizer price)*log(past rainfall variability)*log(landholding) 0.043 0.091
log(fertilizer price)*log(past rainfall variability)*log(age of HH head) 0.363 0.174
log(fertilizer price)*log(past stress degree days)*log(assets) 0.000 0.016
log(fertilizer price)*log(past stress degree days)*log(landholding) 0.026 0.021
log(fertilizer price)*log(past stress degree days)*log(age of HH head) 0.026 0.091
log(labor price)*log(current growing degree days)*log(assets) 0.012 0.008
log(labor price)*log(current growing degree days)*log(landholding) 0.007 0.010
log(labor price)*log(current growing degree days)*log(age of HH head) 0.160 0.048
log(labor price)*log(current stress degree days)*log(assets) 0.001 0.004
log(labor price)*log(current stress degree days)*log(landholding) 0.007 0.007
log(labor price)*log(current stress degree days)*log(age of HH head) 0.020 0.016
log(labor price)*log(current rainfall)*log(assets) 0.033 0.024
log(labor price)*log(current rainfall)*log(landholding) 0.038 0.033
log(labor price)*log(current rainfall)*log(age of HH head) 0.269 0.094
log(labor price)*log(current rainfall variability)*log(assets) 0.066 0.027
log(labor price)*log(current rainfall variability)*log(landholding) 0.049 0.052
log(labor price)*log(current rainfall variability)*log(age of HH head) 0.376 0.137
log(labor price)*log(past rainfall)*log(assets) 0.031 0.046
log(labor price)*log(past rainfall)*log(landholding) 0.114 0.055
log(labor price)*log(past rainfall)*log(age of HH head) 0.136 0.226
log(labor price)*log(past growing degree days)*log(assets) 0.023 0.017
log(labor price)*log(past growing degree days)*log(landholding) 0.008 0.012
log(labor price)*log(past growing degree days)*log(age of HH head) 0.022 0.131
log(labor price)*log(past rainfall variability)*log(assets) 0.020 0.011
log(labor price)*log(past rainfall variability)*log(landholding) 0.026 0.020
log(labor price)*log(past rainfall variability)*log(age of HH head) 0.065 0.083
log(labor price)*log(past stress degree days)*log(assets) 0.001 0.004
log(labor price)*log(past stress degree days)*log(landholding) 0.012 0.009
log(labor price)*log(past stress degree days)*log(age of HH head) 0.028 0.034
log(improved seed price)*log(current growing degree days)*log(assets) 0.021 0.008
log(improved seed price)*log(current growing degree 0.008 0.013
days)*log(landholding)
log(improved seed price)*log(current growing degree days)*log(age of 0.052 0.065
HH head)
log(improved seed price)*log(current stress degree days)*log(assets) 0.006 0.003
log(improved seed price)*log(current stress degree 0.001 0.006
days)*log(landholding)
log(improved seed price)*log(current stress degree days)*log(age of HH 0.023 0.015
head)
log(improved seed price)*log(current rainfall)*log(assets) 0.081 0.041
log(improved seed price)*log(current rainfall)*log(landholding) 0.017 0.066
log(improved seed price)*log(current rainfall)*log(age of HH head) 0.274 0.157
log(improved seed price)*log(current rainfall variability)*log(assets) 0.051 0.046
Continued

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
142 January 2018 Amer. J. Agr. Econ.

Table A.1 continued


Variable Coefficient Clustered
(at district level)
standard errors
log(improved seed price)*log(current rainfall 0.084 0.089
variability)*log(landholding)
log(improved seed price)*log(current rainfall variability)*log(age of HH 0.353 0.154
head)
log(improved seed price)*log(past rainfall)*log(assets) 0.175 0.067
log(improved seed price)*log(past rainfall)*log(landholding) 0.006 0.059
log(improved seed price)*log(past rainfall)*log(age of HH head) 0.181 0.283
log(improved seed price)*log(past growing degree days)*log(assets) 0.029 0.019
log(improved seed price)*log(past growing degree 0.019 0.016
days)*log(landholding)
log(improved seed price)*log(past growing degree days)*log(age of HH 0.020 0.062
head)
log(improved seed price)*log(past rainfall variability)*log(assets) 0.031 0.016
log(improved seed price)*log(past rainfall variability)*log(landholding) 0.018 0.007
log(improved seed price)*log(past rainfall variability)*log(age of HH 0.078 0.083
head)
log(improved seed price)*log(past stress degree days)*log(assets) 0.004 0.004
log(improved seed price)*log(past stress degree days)*log(landholding) 0.007 0.006
log(improved seed price)*log(past stress degree days)*log(age of HH 0.017 0.020
head)
log(current growing degree days) 1.464 1.149
log(current stress degree days) 0.104 0.282
log(current rainfall) 5.907 3.211
log(current rainfall variability) 5.716 2.235
log(assets) 1.302 0.634
log(landholding) 0.847 1.944
log(subsidized fertilizer quantity) 0.005 0.004
log(age of HH head) 18.392 7.493
cattle ownership 0.077 0.088
log(subsidized improved seed) 0.000 0.000
log(current growing degree days)*log(assets) 0.029 0.040
log(current growing degree days)*log(landholding) 0.000 0.046
log(current growing degree days)*log(age of HH head) 0.529 0.265
log(current stress degree days)*log(assets) 0.010 0.011
log(current stress degree days)*log(landholding) 0.006 0.031
log(current stress degree days)*log(age of HH head) 0.045 0.057
log(current rainfall)*log(assets) 0.171 0.074
log(current rainfall)*log(landholding) 0.125 0.166
log(current rainfall)*log(age of HH head) 1.671 0.782
log(current rainfall variability)*log(assets) 0.106 0.093
log(current rainfall variability)*log(landholding) 0.159 0.196
log(current rainfall variability)*log(age of HH head) 1.332 0.550
yr2003_04 0.056 0.054
yr2006_07 1.038 0.071
yr2008_09 0.401 0.105

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Sesmero, Ricker-Gilbert, and Cook Adaptation to Weather Patterns by Farm Households 143

Appendix B

Table B.1. Effect of Changes in Past Weather on Median Net Income and Netput Shares
Dependent Weather Past Current Past Current Past Current
variable variables (5 years) growing (10 growing (20 years) growing
season years) season season
(based (based (based
on 5 on 10 on 20
years) years) years)
i Net incomeRainfall (GSP) 0.94** 1.30** 1.31** 0.74* 1.50** 0.92*
ii Net incomeRainfall 0.06 0.06 0.31*** 0.17 0.37 0.14
variability
(CVGSP)
iii Net income Stress Degree 0.01 0.00 0.04 0.013 0.10 0.00
Days (SDD)
Change in shares with respect to Rainfall (GSP)
iv Maize output GSP 0.63*** 0.54* 0.89*** 0.82*** 0.95** 0.70***
v Non-maize GSP 0.03 0.12 1.72** 1.5*** 0.09 0.19
output
vi Inorganic GSP 0.54 0.30 0.71 0.51* 0.73 0.41*
fertilizer
vii Hired labor GSP 0.00 0.01 0.014 0.002 0.04 0.01
viii Improved GSP 0.06 0.12* 0.11 0.14 0.10 0.11
maize seed
Change in shares with respect to Stress Degree days (SDD)
ix Maize output SDD 0.01 0.03 0.03 0.03 0.06 0.04
X Non-maize SDD 0.04 0.02 0.10 0.08 0.03 0.02
output
xi Inorganic SDD 0.03 0.06* 0.05 0.07** 0.05 0.08**
fertilizer
xii Hired labor SDD 0.02 0.00 0.03 0.004 0.04 0.00
xiii Improved SDD 0.01 0.01 0.007 0.013 0.01 0.02
maize seed
Change in shares with respect to Rainfall Variability (CVGSP)
xiv Maize output CVGSP 0.06 0.33 0.20** 0.47* 0.32* 0.37
xv Non-maize CVGSP 0.04 0.03 0.48** 0.91* 0.22 0.00
output
xvi Inorganic CVGSP 0.07 0.39* 0.10 0.49* 0.33 0.40*
fertilizer
xvii Hired labor CVGSP 0.00 0.07 0.10** 0.07 0.11 0.07
xviii Improved CVGSP 0.03 0.05 0.07 0.05 0.10 0.04
maize seed
Note: Asterisk * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. Significance is
based on standard errors obtained via bootstrapping at 200 repetitions and clustered at the district level. Standard errors are nor reported for expositional
ease.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
144 January 2018 Amer. J. Agr. Econ.

Table B.2. Effect of Changes in Past Weather on Median Net Income and Netput Shares
Dependent Weather SDD 28 SDD 30 (main model) SDD 32
variable variables
Past Current Past Current Past Current
growing growing growing
season season season
i Net income Rainfall (GSP) 1.51*** 0.86** 1.31** 0.74* 1.37** 0.81**
ii Net income Rainfall 0.32** 0.11 0.31*** 0.17 0.37** 0.09
variability
(CVGSP)
iii Net income Stress Degree 0.01 0.01 0.04 0.013 0.05 0.01
Days (SDD)
Change in shares with respect to Rainfall (GSP)
iv Maize output GSP 0.99*** 0.94*** 0.89*** 0.82*** 0.90*** 0.88***
v Non-maize GSP 1.92** 1.73*** 1.72** 1.5*** 1.72** 1.57***
output
vi Inorganic GSP 0.75 0.63** 0.71 0.51* 0.65 0.52**
fertilizer
vii Hired labor GSP 0.10 0.02 0.014 0.002 0.06 0.00
viii Improved GSP 0.09 0.13 0.11 0.14 0.12 0.15*
maize seed
Change in shares with respect to Stress Degree days (SDD)
ix Maize output SDD 0.01 0.02 0.03 0.03 0.03 0.02
X Non-maize SDD 0.09 0.05 0.10 0.08 0.03 0.04
output
xi Inorganic SDD 0.06 0.06** 0.05 0.07** 0.00 0.04*
fertilizer
xii Hired labor SDD 0.04 0.00 0.03 0.004 0.01 0.01
xiii Improved SDD 0.03 0.02 0.007 0.013 0.01 0.02
maize seed
Change in shares with respect to Rainfall Variability (CVGSP)
xiv Maize output CVGSP 0.21** 0.56** 0.20** 0.47* 0.25*** 0.53**
xv Non-maize CVGSP 0.52** 1.10** 0.48** 0.91* 0.57*** 1.00**
output
xvi Inorganic CVGSP 0.12 0.59** 0.10 0.49* 0.13 0.52**
fertilizer
xvii Hired labor CVGSP 0.11** 0.06 0.10** 0.07 0.12** 0.07
xviii Improved CVGSP 0.08 0.05 0.07 0.05 0.08 0.06
maize seed
Note: Asterisk * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. Significance is
based on standard errors obtained via bootstrapping at 200 repetitions and clustered at the district level. Standard errors are nor reported for expositional
ease.

Downloaded from https://academic.oup.com/ajae/article-abstract/100/1/115/4638892


by Adam Ellsworth, Adam Ellsworth
on 18 December 2017
Copyright of American Journal of Agricultural Economics is the property of Oxford
University Press / USA and its content may not be copied or emailed to multiple sites or
posted to a listserv without the copyright holder's express written permission. However, users
may print, download, or email articles for individual use.

You might also like