ON ADAPTATION TO CLIMATE CHANGE AND RISK EXPOSURE

IN THE NILE BASIN OF ETHIOPIA

Salvatore Di Falco

London School of Economics

Marcella Veronesi

Institute for Environmental Decisions, ETH Zurich

Abstract. This study investigates the impact of climate change adaptation on farm households’ downside
risk exposure (e.g., risk of crop failure) in the Nile Basin of Ethiopia. The analysis relies on a moment-
based specification of the stochastic production function. We estimate a simultaneous equations model
with endogenous switching to account for the heterogeneity in the decision to adapt or not, and for
unobservable characteristics of farmers and their farm. We find that (i) climate change adaptation reduces
downside risk exposure; farm households that implemented climate change adaptation strategies get
benefits in terms of a decrease in the risk of crop failure; (ii) farm households that did not adapt would
benefit the most in terms of reduction in downside risk exposure from adaptation; and (iii) there are
significant differences in downside risk exposure between farm households that did and those that did not
adapt to climate change. The analysis also shows that the quasi-option value, that is the value of waiting
to gather more information, plays a significant role in farm households’ decision on whether to adapt to
climate change. Farmers that are better informed may value less the option to wait to adapt, and so are
more likely to adapt than other farmers.

Keywords: adaptation, climate change, endogenous switching, Ethiopia, risk exposure, stochastic
production function, skewness.

JEL classification: D80, Q18, Q54

1
1. Introduction

One consequence of climate change in sub Saharan Africa is that farmers will be more exposed to risk.

More erratic and scarce rainfall and higher temperature imply that farmers will be facing a larger extent

of uncertainty. A prime example is Ethiopia. Rainfall variability and associated drought have been major

causes of food shortage and famine in Ethiopia. During the last forty years, Ethiopia has experienced

many severe droughts leading to production levels that fell short of basic subsistence levels for many

farm households (Relief Society of Tigray, REST and NORAGRIC at the Agricultural University of

Norway 1995, p. 137). Harvest failure due to extreme weather events is the most important cause of risk-

related hardship of Ethiopian rural households, with adverse effects on farm household consumption and

welfare (Dercon 2004, 2005). Future prospects of climate change are likely to exacerbate these issues.

The implementation of adaptation strategies is very important. Farmers may need to implement

adaptation measures to invest in soil conservation measures in the attempt of retaining soil moisture.

Alternatively they can plant trees to procure some shading on the soil or resort to water harvesting

technologies. On the other hand, if the conditions become far too challenging, then farmers may see less

of a scope for investment (i.e., prospects are too gloomy), and they might be forced out of agriculture and

migrate with very important implications in terms of livelihoods.

This paper investigates whether the set of strategies (e.g., change crops, soil and water

conservation) implemented in the field by farm households in response to long-term changes in

environmental conditions (e.g., temperature and rainfall) affect production risk exposure. In other words,

are farm households that implemented climate change adaptation strategies getting benefits in terms of a

reduction in risk exposure? Are there significant differences in risk exposure between farm households

that did and those that did not adapt to climate change? Looking at the risk implications of adaptation to

climate change is a novel contribution to the literature. There is a very large and growing body of

literature assessing the impact of climate change in agriculture. This literature, however, focuses on the

implications of climatic variables on land values, revenues or productivity (e.g., Mendelsohn et al. 1994;

2
Kurukulasuriya and Rosenthal 2003; Seo and Mendelsohn 2008; Deressa and Hassan 2010; Di Falco,

Veronesi and Yesuf, 2011). To our knowledge the empirical assessment of the role of adaptation on risk

exposure has not been investigated yet. We aim to fill this gap.

We define risk exposure in terms of downside risk (e.g., probability of crop failure). The analysis

relies on a moment-based specification of the stochastic production function (Antle 1983; Antle and

Goodger 1984). This method has been widely used in the context of risk management in agriculture (Just

and Pope 1979; Kim and Chavas 2003; Koundouri et al. 2006; and Di Falco and Chavas 2009). The focus

on crop failure seems natural in our setting. Avoiding crop failure is indeed the major preoccupation of

farmers in Ethiopia. Moreover, since the variance does not distinguish between unexpected good and bad

events, we consider the skewness in risk analysis, that is we approximate downside risk exposure by the

third moment of the crop yield distribution. If the skewness of yield increases then it means that

downside risk exposure decreases, that is the probability of crop failure decreases (Di Falco and Chavas

2009). This approach can thus capture the full extent of risk exposure, thus including the implications of

more extreme events In addition, we assume that uncertainty comes from random climate variables,

incomplete information, and from future profit flows, which depend on price and production uncertainty

due to sudden changes in global markets (e.g., in agricultural commodities).

We investigate the effects of adaptation on risk exposure in an endogenous switching regression

framework by using data from a survey undertaken in the Nile Basin of Ethiopia in 2005. The survey

collected information on both farm households that did and did not adapt plus on a very large set of

control variables. We take into account that the differences in risk exposure between those farm

households that did and those that did not adapt to climate change could be due to unobserved

heterogeneity. Indeed, not distinguishing between the casual effect of climate change adaptation and the

effect of unobserved heterogeneity could lead to misleading policy implications. We account for the

endogeneity of the adaptation decision by estimating a simultaneous equations model with endogenous

switching by full information maximum likelihood estimation. Finally, we build a counterfactual

3
analysis, and compare the expected downside risk exposure under the actual and counterfactual cases of

wheather the farm household did or did not adapt to climate change. Treatment and heterogeneity effects

are calculated to understand the differences in downside risk exposure between farm households that

adapted and those that did not adapt.

Key findings of our analysis are (i) that adaptation to climate change decreases downside risk

exposure, and thereby the risk of crop failure; (ii) that there are significant and non-negligible differences

in downside risk exposure between the farm households that adapted and those that did not adapt; (iii)

that farm households that did not adapt would benefit the most in terms of reduction in risk exposure

from adaptation; and (iv) that provision of information through radio, farmer-to-farmer extension, and

extension officers is a key driver of adaptation. This implies that the quasi-option value, that is the value

of waiting to gather more and better information, plays a significant role in farm households’ decisions on

whether to adapt to climate change.

The next section presents a description of the study sites and survey instruments. Sections 3 and 4

outline the model and the estimation procedure used. Section 5 presents the results, and Section 6

concludes the paper by offering some final remarks.

2. Survey Design and Data Description

This study relies on a survey conducted on 1,000 farm households located within the Nile Basin of

Ethiopia in 2005. The sampling frame considered traditional typology of agro-ecological zones in the

country (namely, Dega, Woina Dega, Kolla and Berha), percent of cultivated land, degree of irrigation

activity, average annual rainfall, rainfall variability, and vulnerability (number of food aid dependent

population). The sampling frame selected the woredas (an administrative division equivalent to a district)

in such a way that each class in the sample matched to the proportions for each class in the entire Nile

basin. The procedure resulted in the inclusion of twenty woredas. Random sampling was then used in

selecting fifty households from each woreda.

4
 One of the survey instruments was in particular designed to capture farmers’ perceptions and

understanding on climate change, and their approaches for adaptation. Questions were included to

investigate whether farmers have noticed changes in mean temperature and rainfall over the last two

decades, and reasons for observed changes. About 90 percent of the sample perceived long-term changes

in mean temperature or/and rainfall over the last 20 years. About 68, 4, and 28 percent perceived mean

temperature as increasing, decreasing and remaining the same over the last twenty years, respectively.

Similarly, 18, 62 and 20 percent perceived mean annual rainfall as increasing, declining and remaining

the same over the last twenty years, respectively. Overall, increased temperature and declining rainfall are

the predominant perceptions in our study sites.

Furthermore, some questions investigated whether farm households made some adjustments in

their farming practices in response to long-term changes in mean temperature and rainfall by adopting

some particular strategies. We define the undertaken strategies as “adaptation strategies,” and create the

variable adaptation equal to 1 if a farm household adopted any strategy in response to long-term changes

in mean temperature and rainfall, 0 otherwise. Changing crop varieties, adoption of soil and water

conservation strategies, and tree planting were major forms of adaptation strategies followed by the farm

households in our study sites. These adaptation strategies are mainly yield-related and account for more

than 95 percent of the adaptation strategies followed by the farm households who actually undertook an

adaptation strategy. The remaining adaptation strategies accounting for less than five percent were water

harvesting, irrigation, non-yield related strategies such as migration, and shift in farming practice from

crop production to livestock herding or other sectors. About 58 percent and 42 percent of the farm

households had taken no adaptation strategies in response to long-term shifts in temperature and rainfall,

respectively. More than 90 percent of the respondents who did not adapt indicated lack of information,

land, money, and shortages of labor, as major reasons for not undertaking any adaptation strategy. Lack

of information is cited as the predominant reason by 40-50 percent of the households.

5
 In addition, detailed production data were collected at different production stages (i.e., land

preparation, planting, weeding, harvesting, and post harvest processing). The area is almost totally

rainfed. Only 0.6 percent of the households are using irrigation water to grow their crops. Production

input and output data were collected for two cropping seasons, i.e., Meher (long rainy season), and Belg

(the short rainy season) at the plot-level. However, many plots have two crops grown on them annually

(one during each of the Meher and Belg seasons).

The farming system in the survey sites is very traditional with plough and yolk (animals’ draught

power). Labor is the major input in the production process during land preparation, planting, and post

harvest processing. Labor inputs were disaggregated as adult male’s labor, adult female’s labor, and

children’s labor. This approach of collecting data (both inputs and outputs) at different stages of

production and at different levels of disaggregation was chosen to reduce cognitive burden on the side of

the respondents, and increase the likelihood of retrieving better retrospective data. The three forms of

labor were aggregated as one labor input using adult equivalents. We employed the standard conversion

factor in the literature on developing countries where an adult female and children labor are converted

into adult male labor equivalent at 0.8 and 0.3 rates, respectively.

Finally, although a total of 48 annual crops were grown in the basin, the first five major annual

crops (teff, maize, wheat, barley, and beans) cover 65 percent of the plots. These are also the crops that

constitute the staple foods of the local diet. We limit the estimation to these primary crops. The final

sample includes twenty woredas, 941 farm households (i.e., on average about forty-seven farm

households per woreda), and 2,807 plots (i.e., on average about three plots per farm household). The

scale of the analysis is at the plot-level. The basic descriptive statistics are presented in table 1, and the

definition of the variables in table A1 of the appendix.

[TABLE 1 ABOUT HERE]

6
3. Adaptation to Climate Change and Risk Exposure

The climate change adaptation decision and its implications in terms of risk exposure can be framed in

the setting of a two stage framework.1 In the first stage, we use a selection model for climate change

adaptation where a representative risk adverse farm household i chooses to implement climate change

adaptation strategies if the expected utility from final benefits if she adapts U(π1) is greater than the

expected utility if she does not adapt U(π0), i.e.,

(1) E U ( 1 )   E U ( 0 )   0

where E is the expectation operator based on the subjective distribution of the uncertain variables facing

the decision maker, and U(.) is the von Neumann-Morgenstern utility function representing the farm

household’s preferences under risk.

In addition, we should consider the value that farm households assign to information. Farm

households may decide to delay the adoption of climate change adaptation strategies in order to collect

more information on climate change and on adaptation strategies, for example, through extension officers

and farmer-to-farmer extension. This implies that the farm household chooses to adapt iff

(2) E U ( 1 )   E U ( 0 )    ,

where I ≥ 0 represents the information value, which depends on the farm household’s characteristics, the

uncertainty concerning the adoption of new strategies, and the fixed costs of new investments (Koundouri

et al. 2006).

Let A* be the latent variable that captures the expected benefits from the adaptation choice with

respect to not adapting. We specify the latent variable as

1 if Ai*  0
(3) A  Z i α  i with Ai  
*
i ,
0 otherwise

1
A more comprehensive model of climate change adaptation is provided by Mendelsohn (2000).

7
that is farm household i will choose to adapt (Ai = 1), through the implementation of some strategies in

response to long-term changes in mean temperature and rainfall, if A* > 0, and 0 otherwise. The vector Z

represents variables that affect the expected benefits of adaptation. These factors can be classified into

different groups. First, we consider the characteristics of the operating farm (e.g., soil fertility and

erosion). For instance, farms characterized by more fertile soil might be less affected by climate change

and therefore relatively less likely to implement adaptation strategies. Since extension services are one

important means for farmers to gain information on this, access to extension (both government and

farmer-to-farmer) can be used as a measure of access to information. Particularly relevant in this setting

is that farmers received information on climate. Farm head and farm household’s characteristics (e.g.,

age, gender, education, marital status, if the farm head has an off-farm job, and farm household size), and

the presence of assets (e.g., machinery and animals) may in principle also affect the probability of

adaptation. Experience is approximated by age and education.

In the second stage, we model the effect of adaptation on risk exposure by relying on a moment-

based specification of the stochastic production function (Antle 1983; Antle and Goodger 1984). This is a

very flexible device that has been largely used in agricultural economics to model the implications of

weather risk and risk management (Just and Pope 1979; Kim and Chavas 2003; Koundouri et al. 2006;

and Di Falco and Chavas 2009). Consider a risk averse farm household that produces output y using

inputs x under risk through a production technology represented by a well-behaved (i.e., continuous and

twice differentiable) stochastic production function y = g(x, υ), where υ is a vector of random variables

representing risk, that is uncontrollable factors affecting output such as extreme weather events and

changes in temperature and rainfall.

Risk exposure is represented by the moments of the production function g(x, υ). The moments are

computed following Kim and Chavas (2003), and Di Falco and Chavas (2009). We consider the

following econometric specification for g(x, υ):

(4) g (x, υ)  f1 (x, γ 1 )  u

8
where f1 (x, γ 1 )  E  g (x, υ) is the mean of g (x, υ) , that is the first central moment, and

u  g (x, υ)  f1 (x, γ 1 ) is a random variable with mean zero whose distribution is exogenous to farmers’

actions. The higher moments of g(x, υ) are given by

(5)  
E  g (x, υ)  f1 (x, γ 1 )  |x  f k (x, γ k )
k

for k = 2, 3. This implies that f 2 (x, γ 2 ) is the second central moment, that is the variance, and f3 (x, γ 3 )

is the third central moment, that is the skewness. This approach provides a flexible representation of the

impacts of inputs, (e.g., seeds, fertilizers, manure, and labor), assets (e.g., machinery and animals), and

soil’s characteristics (e.g., soil fertility and erosion level) on the distribution of output under production

uncertainty. In this study, we go beyond standard mean-variance analysis by considering the effects of

skewness and downside risk exposure. An increase in skewness implies a reduction in downside risk

exposure, which implies, for example, a reduction in the probability of crop failure. Reducing downside

risk means decreasing the asymmetry (or skewness) of the risk distribution toward high outcome, holding

both means and variance constant (Menezes, Geiss, and Tessler 1980).

The simplest approach to examine the impact of adaptation to climate change on farm

households’ downside risk exposure would be to include in the skewness equation a dummy variable

equal to one if the farm household adapted to climate change, and then, to apply ordinary least squares.

This approach, however, might yield biased estimates because it assumes that adaptation to climate

change is exogenously determined while it is potentially endogenous. The decision on whether to adapt

or not to climate change is voluntary and may be based on individual self-selection. Farmers that adapted

may have systematically different characteristics from the farmers that did not adapt, and they may have

decided to adapt based on expected benefits. Unobservable characteristics of farmers and their farm may

affect both the adaptation decision and risk exposure, resulting in inconsistent estimates of the effect of

adaptation on production risk and risk of crop failure. For example, if only the most skilled or motivated

farmers chose to adapt and we fail to control for skills, then we will incur upward bias.

9
 We account for the endogeneity of the adaptation decision by estimating a simultaneous equations

model of climate change adaptation and risk exposure with endogenous switching by full information

maximum likelihood (FIML). For the model to be identified it is important to use as exclusion

restrictions, thus as selection instruments, not only those automatically generated by the nonlinearity of

the selection model of adaptation (1) but also other variables that directly affect the selection variable but

not the outcome variable. In our case study, we use as selection instruments the variables related to the

information sources (e.g., government extension, farmer-to-farmer extension, information from radio,

and, if received information in particular on climate), and the farm head and farm household

characteristics. We establish the admissibility of these instruments by performing a simple falsification

test: if a variable is a valid selection instrument, it will affect the adaptation decision but it will not affect

the risk exposure among farm households that did not adapt. Table A2 of the appendix shows that the

information sources and the farm head and farm household characteristics can be considered as valid

selection instruments: they are statistically significant determinants of the decision to adapt or not to

climate change (Model 1) but not of downside risk exposure among farm households that did not adapt

(Model 2).

To account for selection biases we adopt an endogenous switching regression model of downside

risk exposure where farmers face two regimes (1) to adapt, and (2) not to adapt defined as follows:

(6a) Regime 1: y1i  X1i β1  1i if Ai  1

(6b) Regime 2: y2i  X 2i β 2   2i if Ai  0

where yi is the third central moment f 3 (x, γ 3 ) of production function (4) in regimes 1 and 2, i.e., the

skewness, and Xi represents a vector of inputs (e.g., seeds, fertilizers, manure, and labor), and of the

soil’s characteristics, and assets included in Z.

The error terms in equations (3), (6a) and (6b) are assumed to have a trivariate normal

distribution, with zero mean and covariance matrix , i.e., (η, ε1, ε2)'  N(0, Σ)

10
   2  1  2 
 
with Σ   1  12 . ,
 2 .  2 
 2

where  2 is the variance of the error term in the selection equation (1), which can be assumed to be equal

to 1 since the coefficients are estimable only up to a scale factor (Maddala 1983, p. 223),  12 and  22 are

the variances of the error terms in the skewness functions (6a) and (6b), and  1 and  2 represent the

covariance of i and 1i and 2i.2 Since y1i and y2i are not observed simultaneously, the covariance

between 1i and 2i is not defined (reported as dots in the covariance matrix , Maddala 1983, p. 224). An

important implication of the error structure is that because the error term of the selection equation (1) i

is correlated with the error terms of the skewness functions (6a) and (6b) (1i and 2i), the expected values

of 1i and 2i conditional on the sample selection are nonzero:

 (Ziα )  (Zi α )
E 1i | Ai  1   1   1 1i , and E  2i | Ai  0   2   2 2i , where (.) is the
 (Z i α ) 1   (Z i α )

standard normal probability density function, (.) the standard normal cumulative density function, and

 (Ziα )  (Ziα )
1i  , and 2i   . If the estimated covariances ˆ1 and ˆ 2 are statistically
(Zi α) 1  (Ziα )

significant, then the decision to adapt and downside risk exposure are correlated, that is we find evidence

of endogenous switching and reject the null hypothesis of the absence of sample selectivity bias. This

model is defined as a “switching regression model with endogenous switching” (Maddala and Nelson

1975).

2
For notational simplicity, the covariance matrix  does not reflect the clustering implemented in the empirical analysis.

11
 An efficient method to estimate endogenous switching regression models is full information

maximum likelihood estimation (Lee and Trost 1978).3 The logarithmic likelihood function given the

previous assumptions regarding the distribution of the error terms is

N    
(7) ln Li   Ai  ln   1i   ln  1  ln  (1i ) 
i 1   1  
where  ji 
 Ziα   j ji /  j  , j  1, 2 , with  j
    1   2j
 (1  Ai )  ln   2i   ln  2  ln 1   ( 2i )   ,
  2  

denoting the correlation coefficient between the error term i of the selection equation (1) and the error

term ji of equations (6a) and (6b), respectively.

4. Conditional Expectations, Treatment and Heterogeneity Effects

The endogenous switching regression model can be used to compare the expected downside risk

exposure of farm households that adapted (a) with respect to farm households that did not adapt (b), and

to investigate the expected downside risk exposure in the counterfactual hypothetical cases (c) that the

adapted farm households did not adapt, and (d) that the non-adapted farm household adapted. The

conditional expectations for downside risk exposure in the four cases are presented in table 2 and defined

as follows:

(8a) E ( y1i | Ai  1)  X1i β1   1 1i

(8b) E ( y2i | Ai  0)  X 2i β 2   2 2i

(8c) E ( y2i | Ai  1)  X1i β 2   2 1i

(8d) E ( y1i | Ai  0)  X 2i β1   1 2i .

3
An alternative estimation method is the two-step procedure (see Maddala 1983, p. 224 for details). However, this method is

less efficient than FIML, it requires some adjustments to derive consistent standard errors (Maddala 1983, p. 225), and it

poorly performs in case of high multicollinearity between the covariates of the selection equation (3) and the covariates of the

skewness equations (6a) and (6b) (Hartman 1991; Nelson 1984; and Nawata 1994).

12
 [TABLE 2 ABOUT HERE]

Cases (a) and (b) along the diagonal of table 2 represent the actual expectations observed in the

sample. Cases (c) and (d) represent the counterfactual expected outcomes.

In addition, following Heckman et al. (2001), we calculate the effect of the treatment “to adapt”

on the treated (TT) as the difference between (a) and (c),

(9) TT  E ( y1i | Ai  1) - E ( y2i | Ai  1)  X1i (β1 - β 2 )  ( 1   2 )1i ,

which represents the effect of climate change adaptation on downside risk exposure of the farm

households that actually adapted to climate change. Similarly, we calculate the effect of the treatment on

the untreated (TU) for the farm households that actually did not adapt to climate change as the difference

between (d) and (b),

(10) TU  E ( y1i | Ai  0) - E ( y2i | Ai  0)  X 2i (β1 - β 2 )  ( 1   2 )2i .

We can use the expected outcomes described in (8a)-(8d) to calculate also the heterogeneity

effects. For example, farm households that did not adapt may have been exposed to lower downside risk

than farm households that adapted regardless of the fact that they decided not to adapt but because of

unobservable characteristics such as their abilities. We follow Carter and Milon (2005) and define as “the

effect of base heterogeneity” for the group of farm households that decided to adapt as the difference

between (a) and (d),

(11) BH1  E ( y1i | Ai  1) - E ( y1i | Ai  0) = (X1i - X 2i )β1i   1 (1i  2i ) .

Similarly for the group of farm households that decided not to adapt, “the effect of base

heterogeneity” is the difference between (c) and (b),

(12) BH 2  E ( y2i | Ai  1) - E ( y2i | Ai  0) = (X1i - X 2i )β 2i   2 (1i  2i ) .

13
 Finally, we investigate the “transitional heterogeneity” (TH), that is whether the effect of adapting

to climate change is larger or smaller for the farm households that actually adapted to climate change or

for the farm household that actually did not adapt in the counterfactual case that they did adapt, that is the

difference between equations (9) and (10), i.e., (TT) and (TU).

5. Results

Table 3 reports the estimates of the endogenous switching regression model estimated by full information

maximum likelihood with clustered standard errors at the woreda level without fixed effects.4 The first

column presents the estimation of downside risk exposure by ordinary least squares (OLS) with no

switching and with a dummy variable equal to 1 if the farm household adapted to climate change, 0

otherwise. The second, third and fourth columns present, respectively, the estimated coefficients of

selection equation (3) on adapting or not to climate change, and of downside risk exposure, which is

represented by skewness functions (6a) and (6b) (i.e., the third central moments of production function

(4) in regimes (1) and (2)), for farm households that did and did not adapt to climate change. Table A3 of

the appendix shows the estimation of the production function (4) in regimes (1) and (2) from which we

derived the third central moments.

[TABLE 3 ABOUT HERE]

The results of the estimation of equation (3) suggest that key drivers of farm households’ decision

to adopt some strategies in response to long-term changes in mean temperature and rainfall are

represented by the information sources farm households have access to, in particular the provision of

climate information both from formal and informal institutions, (table 3, column (2)). Farm households

that received information about future climate change, and had access to formal agricultural extension,

4
We use the “movestay” command of STATA to estimate the endogenous switching regression model by FIML (Lokshin and

Sajaia 2004).

14
farmer-to-farmer extension or the media are more likely to adapt. These positive effects may indicate that

farmers that are better informed may value less the option to wait, and so are more likely to adapt than

other farmers. This implies that waiting for opportunities to gather more and better information might

have a positive value (Koundouri et al. 2006).

The question now is whether farm households that implemented climate change adaptation

strategies got benefits in terms of a reduction in downside risk exposure, (e.g., a decrease in the

probability of crop failure). The simplest approach to answer this question consists in estimating an OLS

model of downside risk exposure that includes a dummy variable equal to 1 if the farm household

adapted, 0 otherwise (table 3, column (1)). An increase in skewness implies a reduction in downside risk

exposure. This approach would lead us to conclude that having adapted to climate change did not

significantly reduce farm households’ downside risk exposure (the coefficient of the dummy variable

adaptation is positive but insignificant). This approach, however, assumes that adaptation to climate

change is exogenously determined while it is a potentially endogenous variable. The estimation via OLS

would yield biased and inconsistent estimates. In addition, OLS estimates do not explicitly account for

potential structural differences between the skewness function of farm households that adapted to climate

change and the skewness function of farm households that did not adapt. The estimates presented in the

last two columns of table 3 account for the endogenous switching in the skewness function. Both the

estimated coefficients of the correlation terms  j are not significantly different from zero (table 3,

bottom row). Although we could not have known it a priori, this implies that the hypothesis of absence of

sample selectivity bias may not be rejected.

However, the differences in the coefficients of the skewness functions between the farm

households that adapted and those that did not adapt illustrate the presence of heterogeneity in the sample

(table 3, columns (3) and (4)). The skewness function of farm households that adapted to climate change

is significantly different from the skewness function of farm households that did not adapt (at the 1

percent statistical level, F-stat. = 612.71). Among farm households that adapted to climate change inputs

15
such as seeds and manure, and assets such as animals are significantly associated with an increase in the

skewness, and so in a decrease in downside risk exposure, while infertile soils are associated with an

increase in downside risk exposure. However, these factors do not significantly affect the downside risk

exposure of farm households that did not adapt.

[TABLE 4 ABOUT HERE]

As robustness check we present the estimates of the skewness functions including the fixed

effects by Mundlak approach (Table 4). The inclusion of the fixed effects is very important in this setting.

These effects would remove plot invariant characteristics (i.e., farmers skills). We test for the relevance

of the fixed effects, and we do not reject the null hypothesis that they are jointly equal to zero at the 1

percent statistical level (Table 4, bottom row). Tables 5 and 6 present the expected downside risk

exposure under actual (cells (a) and (b)) and counterfactual conditions (cells (c) and (d)). Cells (a) and (b)

represent the expected downside risk exposure observed in the sample if farm households adapted or not

to climate change. The last column presents the treatment effects of adaptation on downside risk

exposure. Our results show that adaptation to climate change significantly increases the skewness, that is

decreases downside risk exposure, and so the probability of crop failure. In addition, we find that the

transitional heterogeneity effect is negative, that is, farm households that did not adapt would have

benefited the most in terms of reduction in risk exposure from adaptation. Finally, the last row, which

adjusts for the potential heterogeneity in the sample, shows that farm households that actually did not

adapt are less exposed to downside risk than the farm households that adapted in both counterfactual and

actual conditions. This highlights that there are some important sources of heterogeneity that makes the

non-adapters less exposed to downside risk than the adapters irrespective to the issue of climate change.

[TABLE 5 ABOUT HERE]

16
 [TABLE 6 ABOUT HERE]

5. Conclusions

This paper investigated the implications of farm households’ decision to adapt to climate change on

downside risk exposure. We used a moment-based approach that captures the third moment of a

stochastic production function as measure of downside yield uncertainty. Then, we estimated a

simultaneous equations model with endogenous switching to account for unobservable factors that

influence downside risk exposure and the decision to adapt.

The first step of the analysis highlighted that access to information about climate change and

extension services are key determinants of adaptation. They significantly increase the likelihood that farm

households adapt to climate change. This finding is consistent with Koundouri et al. (2006) on irrigation

technology adoption under production uncertainty. Farm households that are better informed may value

less the option to wait, and so are more likely to adopt new technologies than other farmers. This implies

that waiting for gathering more and better information might have a positive value, and the provision of

information on climate change might reduce the quasi-option value associated with adaptation.

Koundouri et al. (2006) conclude that “policy makers may use information provision to induce faster

diffusion of adoption among farmers” (p. 659). They also emphasize that subsidies can be an alternative

instrument to incentivize adoption and diffusion of new technology. However, subsidy policies may

cause income transfers from other economic sectors with consequential welfare losses (Stoneman and

David, 1986).

In addition, we can draw four main conclusions from the results of this study on the effects of

climate change adaptation on downside risk exposure. First, climate change adaptation reduces downside

risk exposure. Farm households that implemented climate change adaptation strategies obtained benefits

in terms of a decrease in the risk of crop failure. Second, farm households that did not adapt would

benefit the most in terms of reduction in downside risk exposure from adaptation. Third, there are

significant differences in downside risk exposure between farm households that did and those that did not

17
adapt to climate change. These differences represent sources of variation between the two groups that the

estimation of an OLS model including a dummy variable for adapting or not to climate change cannot

take into account. Fourth, there are some important sources of heterogeneity that makes the non-adapters

less exposed to downside risk than the adapters irrespective to the issue of climate change.

These results are particularly important to design polices for effective adaptation strategies to

cope with the potential impacts of climate change. Public policies can play an important role in helping

farm households to adapt. The dissemination of climate change information and extension services are of

paramount importance in determining the implementation of adaptation strategies, which could result in

more food security for all farmers irrespective of their unobservable characteristics. The availability of

information on climate change may raise farmers’ awareness of the threats posed by the changing

climatic conditions. Extension services provide an important source of information and education, for

instance, on changing crops and specific soil conservation measures that can deliver food productivity

gains. Future research will investigate the role of different adaptation strategies, and whether the

beneficial effects of adaptation are sensitive to different rainfall areas.

[TABLE A1 ABOUT HERE]

[TABLE A2 ABOUT HERE]

[TABLE A3 ABOUT HERE]

18
References

Antle, J.M. .1983. “Testing the Stochastic Structure of Production: A Flexible Moment-Based

Approach,” Journal of Business and Economic Statistics 1:192-201.

Antle, J.M. and W.A. Goodger. 1984. “Measuring Stochastic Technology: The Case of Tulare Milk

Production,” American Journal of Agricultural Economics 66:342-350.

Carter, D.W., and J.W. Milon .2005. “Price Knowledge in Household Demand for Utility Services,” Land

Economics 81(2):265-283.

Dercon, S. 1996. “Risk, Crop Choice and Savings: Evidence from Tanzania.” Economic Development

and Cultural Change 44(3):385– 415.

Dercon, S. 2004. “Growth and Shocks: Evidence from Rural Ethiopia.” Journal of Development

Economics 74(2):309–329.

Deressa, T.T., and R.H. Hassan .2010. “Economic Impact of Climate Change on Crop Production in

Ethiopia: Evidence from Cross-Section Measures,” Journal of African Economies 18(4):529-554.

Di Falco, S., and J.-P. Chavas. 2009. “On Crop Biodiversity, Risk Exposure and Food Security in the

Highlands of Ethiopia. American Journal of Agricultural Economics 91(3):599-611.

Di Falco, S., M. Veronesi and M. Yesuf. 2011. Does Adaptation Provide Food Security? A Micro

Perspective from Ethiopia. American Journal of Agricultural Economics in press

Hartman, R.S. .1991. “A Monte Carlo Analysis of Alternative Estimators in Models Involving

Selectivity,” Journal of Business and Economic Statistics 9:41-49.

Heckman, J.J., J.L. Tobias, and E.J. Vytlacil .2001. “Four Parameters of Interest in the Evaluation of

Social Programs,” Southern Economic Journal 68(2):210-233.

Just, R.E. and R.D. Pope .1979. “Production Function Estimation and Related Risk Considerations,”

American Journal of Agricultural Economics 61:276-284.

Kim, K. and J.P. Chavas .2003. “Technological Change and Risk Management: An Application to the

Economics of Corn Production,” Agricultural Economics 29:125-142.

19
Koundouri, P., C. Nauges, and V. Tzouvelekas .2006. “Technology Adoption under Production

Uncertainty: Theory and Application to Irrigation Technology,” American Journal of Agricultural

Economics 88(3): 657-670.

Kurukulasuriya, P., and S. Rosenthal .2003. “Climate Change and Agriculture: A Review of Impacts and

Adaptations,” Climate Change Series 91. Environment Department Papers, The World Bank,

Washington, D.C..

Lee, L.F., and R.P. Trost .1978. “Estimation of Some Limited Dependent Variable Models with

Application to Housing Demand,” Journal of Econometrics 8:357-382.

Lokshin, M., and Z. Sajaia .2004. “Maximum Likelihood Estimation of Endogenous Switching

Regression Models,” Stata Journal 4(3):282-289.

Maddala, G.S. .1983. Limited Dependent and Qualitative Variables in Econometrics. Cambridge, U.K.:

Cambridge University Press.

Maddala, G.S., and F.D. Nelson .1975. “Switching Regression Models with Exogenous and Endogenous

Switching,” Proceeding of the American Statistical Association (Business and Economics

Section), pp. 423-426.

Mendelsohn, R. .2000. “Efficient Adaptation to Climate Change,” Climatic Change 45:583-600.

Mendelsohn, R., W. Nordhaus, and D. Shaw .1994. “The Impact of Global Warming on Agriculture: A

Ricardian Analysis.” American Economic Review 84(4):753–771.

Menezes, C., C. Geiss, and J. Tressler. 1980. “Increasing Downside Risk,” American Economic Review

70:921-932.

Mundlak, Y. (1978), “On the Pooling of Time Series and Cross Section Data,” Econometrica 46(1): 69-

85.

Nawata, K. .1994. “Estimation of Sample Selection Bias Models by the Maximum Likelihood Estimator

and Heckman’s Two-Step Estimator,” Economics Letters 45:33-40.

20
Nelson, F.D. .1984. “Efficiency of the Two-Step Estimator for Models with Endogenous Sample

Selection,” Journal of Econometrics 24:181-196.

Relief Society of Tigray (REST) and NORAGRIC at the Agricultural University of Norway.

1995. “Farming Systems, Resource Management and Household Coping Strategies

in Northern Ethiopia: Report of a social and agro-ecological baseline study in central

Tigray.” Aas, Norway.

Seo, S.N., and R. Mendelsohn .2008. “Measuring Impacts and Adaptations to Climate Change: A

Structural Ricardian Model of African Livestock Management,” Agricultural Economics

38(2):151-165.

Stige, L.C., J. Stave, K.Chan, L. Ciannelli, N. Pattorelli, M. Glantz, H. Herren, and N. Stenseth .2006.

“The Effect of Climate Variation on Agro-Pastoral Production in Africa,” PNAS, 103: 3049-3053.

Stoneman, P.L. and P.A. David. 1986. “Adoption Subsidies vs Information Provision as Instruments of

Technology Policy,” Economic Journal 96:142-150.

21
Table 1. Descriptive Statistics

Farm households Farm households

Variable name Total sample
that adapted that did not adapt
Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Dependent variables
adaptation 0.689 0.463 1.000 0.000 0.000 0.000
downside risk exposure 0.608 15.330 0.865 18.454 0.034 0.324
Explanatory variables

Soil characteristics
highly fertile 0.280 0.449 0.257 0.437 0.331 0.471
infertile 0.158 0.365 0.172 0.377 0.128 0.335
no erosion 0.484 0.500 0.472 0.499 0.510 0.500
severe erosion 0.104 0.305 0.114 0.318 0.081 0.274
Assets
machinery 0.019 0.136 0.024 0.153 0.007 0.084
animals 0.874 0.332 0.887 0.317 0.842 0.365
Inputs
labor 100.994 121.268 105.867 133.409 90.176 87.657
seeds 114.905 148.650 125.672 163.896 91.001 103.473
fertilizers 60.609 176.767 61.996 177.867 57.530 174.362
manure 198.148 831.347 254.560 951.670 72.758 438.123
Farm head and farm household
characteristics
literacy 0.489 0.500 0.524 0.500 0.412 0.492
male 0.926 0.263 0.932 0.252 0.912 0.284
married 0.927 0.261 0.930 0.256 0.920 0.272
age 45.717 12.550 46.239 11.926 44.556 13.770
household size 6.597 2.190 6.760 2.138 6.234 2.260
off-farm job 0.250 0.433 0.285 0.452 0.170 0.376
relatives 16.464 43.630 19.534 51.284 9.457 13.259
Information sources
government extension 0.609 0.488 0.761 0.426 0.269 0.444
farmer-to-farmer extension 0.516 0.500 0.660 0.474 0.196 0.397
radio information 0.307 0.461 0.382 0.486 0.139 0.346
climate information 0.422 0.494 0.563 0.496 0.110 0.313
Sample size 2,807 1,936 871
Note: The sample size refers to the total number of plots. The final total sample includes 20 woredas, 941 farm households and
2,807 plots.

22
Table 2. Conditional Expectations, Treatment and Heterogeneity Effects

Decision Stage

Sub-samples To Adapt Not to Adapt Treatment Effects

Farm households that adapted (a) E ( y1i | Ai  1) (c) E ( y2 i | Ai  1) TT

Farm households that did not adapt (d) E ( y1i | Ai  0) (b) E ( y2 i | Ai  0) TU

Heterogeneity effects BH1 BH2 TH

Note: (a) and (b) represent observed downside risk exposure, that is the third central moment f 3 ( x, γ 3 ) of

production function (4); (c) and (d) represent the counterfactual expected downside risk exposure.

Ai = 1 if farm households adapted to climate change; Ai = 0 if farm households did not adapt;

y1i: third central moment if farm households adapted;

y2i: third central moment if farm households did not adapt;

TT: the effect of the treatment (i.e., adaptation) on the treated (i.e., farm households that adapted);

TU: the effect of the treatment (i.e., adaptation) on the untreated (i.e., farm households that did not adapt);

BHi: the effect of base heterogeneity for farm households that adapted (i = 1), and did not adapt (i = 2);

TH = (TT - TU), i.e., transitional heterogeneity.

23
Table 3. Parameters Estimates of Climate Change Adaptation and Downside Risk Exposure

(1) (2) (3) (4)

a
Model OLS Endogenous Switching Regression

Regime 1 Regime 2
(Adaptation = 1) (Adaptation = 0)

Downside risk exposure Downside risk exposure

Dependent Variable Downside risk Adaptation 1/0
among farm households among farm households
exposure
that adapted that did not adapt
0.662
Adaptation 1/0
(0.389)
Soil characteristics
highly fertile -0.470 -0.209 -0.639 -0.018
(0.366) (0.108) (0.518) (0.011)
infertile -0.907* 0.090 -1.256** -0.007
(0.467) (0.163) (0.622) (0.020)
no erosion -0.181 0.065 -0.306 0.031
(0.567) (0.142) (0.812) (0.020)
severe erosion -0.248 0.189 -0.368 0.015
(0.751) (0.135) (1.057) (0.039)
Assets
machinery -0.832* 0.534 -1.186* -0.051
(0.426) (0.481) (0.677) (0.034)
animals 0.529 0.159 0.745* 0.025
(0.329) (0.189) (0.398) (0.021)
Inputs
labor -0.351 -0.468 0.013
(0.326) (0.406) (0.025)
squared labor 0.025 0.031 -0.004
(0.022) (0.025) (0.003)
seeds 0.689*** 0.848*** -0.037
(0.221) (0.267) (0.022)
squared seeds -0.032** -0.039** 0.008
(0.014) (0.016) (0.005)
fertilizers -0.087 -0.086 -0.001
(0.090) (0.122) (0.011)
squared fertilizers 0.003 0.002 -0.000
(0.002) (0.003) (0.0003)
manure 0.058* 0.051* 0.000
(0.028) (0.030) (0.003)
squared manure -0.001** -0.001*** 0.000
(0.0003) (0.000) (0.0001)
Farm head and farm
household characteristics
literacy 0.152
(0.152)
male -0.039
(0.329)
married -0.264
(0.295)
age 0.013**
(0.006)
household size 0.032
(0.035)
off-farm job 0.343***
(0.124)

24
relatives 0.009**
(0.004)
Information sources
government extension 0.592***
(0.113)
farmer-to-farmer extension 0.538***
(0.143)
radio information 0.500**
(0.203)
climate information 0.625***
(0.166)
constant -0.275 -1.269 0.612 -0.017
(0.358) (0.376) (1.129) (0.022)
i 18.170*** 0.312***
(6.730) (0.082)
j -0.048 -0.150
(0.035) (0.110)
Note: aEstimation by full information maximum likelihood at the plot-level. Sample size: 2,807 plots. Robust standard errors
clustered at the woreda level in parentheses. The dependent variable “downside risk exposure” refers to the third central moment
f3 ( x, γ 3 ) (i.e., the skewness) of production function (4);  i denotes the square-root of the variance of the error terms ji in the
outcome equations (6a) and (6b), respectively;  j denotes the correlation coefficient between the error term i of the selection
equation (3) and the error term ji of the outcome equations (6a) and (6b), respectively. The inputs coefficients have been
multiplied by 100. * Significant at the 10% level; ** Significant at the 5% level; *** Significant at the 1% level.

25
Table 4. Parameters Estimates of Climate Change Adaptation and Downside Risk Exposure

Including Fixed Effects by Mundlak (1978)’s approach

(1) (2) (3) (4)

a
Model OLS Endogenous Switching Regression

Regime 1 Regime 2
(Adaptation = 1) (Adaptation = 0)

Downside risk exposure Downside risk exposure

Dependent Variable Downside risk Adaptation 1/0
among farm households among farm households
exposure
that adapted that did not adapt
0.704
Adaptation 1/0
(0.415)
Soil characteristics
highly fertile -0.437 -0.209* -0.603 -0.024*
(0.369) (0.108) (0.528) (0.014)
infertile -0.865* 0.090 -1.172** -0.011
(0.430) (0.163) (0.568) (0.021)
no erosion -0.182 0.065 -0.299 0.030*
(0.565) (0.142) (0.801) (0.017)
severe erosion -0.287 0.189 -0.411 0.013
(0.752) (0.136) (1.033) (0.040)
Assets
machinery -0.774* 0.534 -1.121* -0.040
(0.375) (0.481) (0.596) (0.031)
animals 0.551 0.159 0.749* 0.029
(0.343) (0.189) (0.400) (0.026)
Inputs
labor -0.646 -0.746 0.027
(0.521) (0.567) (0.060)
squared labor 0.033 0.038 -0.004
(0.025) (0.027) (0.004)
seeds 0.796** 0.965*** -0.052*
(0.292) (0.336) (0.028)
squared seeds -0.034** -0.041** 0.008*
(0.015) (0.017) (0.005)
fertilizers -0.186** -0.253** -0.017
(0.084) (0.122) (0.027)
squared fertilizers 0.004* 0.004* 0.0002
(0.002) (0.003) (0.0005)
manure 0.085 0.084 -0.005
(0.050) (0.052) (0.005)
squared manure -0.001** -0.001** 0.00006
(0.0004) (0.0004) (0.00005)
Farm head and farm
household characteristics
literacy 0.152
(0.152)
male -0.039
(0.331)
married -0.264
(0.296)
age 0.013**
(0.006)
household size 0.032
(0.035)

26
off-farm job 0.343***
(0.124)
relatives 0.009**
(0.004)
Information sources
government extension 0.592***
(0.113)
farmer-to-farmer extension 0.538***
(0.143)
radio information 0.500**
(0.203)
climate information 0.625***
(0.167)
Fixed effects
average fertilizers 0.001 0.002 0.0002
(0.001) (0.001) (0.0004)
average seeds -0.002 -0.002 0.0002
(0.002) (0.002) (0.0002)
average manure -0.0005 -0.001 0.0001
(0.0005) (0.001) (0.0001)
average labor 0.004 0.004 -0.0003
(0.004) (0.004) (0.0004)
constant -0.349 -1.269*** 0.588 -0.012
(0.310) (0.376) (1.097) (0.021)
i 18.424*** 0.323***
(6.870) (0.089)
j -0.050 -0.122
(0.032) (0.100)
F-test on fixed effects 2.18 0.23 1.36
Note: aEstimation by full information maximum likelihood at the plot-level. Sample size: 2,807 plots. Robust standard errors
clustered at the woreda level in parentheses. The dependent variable “downside risk exposure” refers to the third central moment
f3 ( x, γ 3 ) (i.e., the skewness) of production function (4);  i denotes the square-root of the variance of the error terms ji in the
outcome equations (6a) and (6b), respectively;  j denotes the correlation coefficient between the error term i of the selection
equation (3) and the error term ji of the outcome equations (6a) and (6b), respectively. The inputs coefficients have been
multiplied by 100. * Significant at the 10% level; ** Significant at the 5% level; *** Significant at the 1% level.

27
Table 5. Average Expected Downside Risk Exposure; Treatment and Heterogeneity Effects

Decision Stage

Sub-samples To Adapt Not to Adapt Treatment Effects

(a) 0.867 (c) -0.013 TT = 0.880***

Farm households that adapted
(0.023) (0.003) (0.024)

(d) 1.624 (b) 0.037 TU = 1.588***

Farm households that did not adapt
(0.031) (0.001) (0.031)

BH1 = -0.757 BH2= -0.049*** TH = -0.708***

Heterogeneity effects
(0.041) (0.005) (0.041)

See note of table 2. (a) and (b) represent observed downside risk exposure f 3 ( x, γ 3 ) , that is the third central moment of

production function (4); (c) and (d) represent the counterfactual expected downside risk exposure. Standard errors in

parentheses. *** Significant at the 1% level.

Table 6. Average Expected Downside Risk Exposure; Treatment and Heterogeneity Effects

Including Fixed Effects by Mundlak (1978)’s approach

Decision Stage

Sub-samples To Adapt Not to Adapt Treatment Effects

(a) 0.865 (c) 0.005 TT = 0.860***

Farm households that adapted
(0.025) (0.004) (0.026)

(d) 1.746 (b) 0.037 TU = 1.709***

Farm households that did not adapt
(0.033) (0.001) (0.032)

BH1 = -0.881 BH2= -0.032*** TH = -1.124***

Heterogeneity effects
(0.044) (0.005) (0.022)

See note of table 2. (a) and (b) represent observed downside risk exposure f 3 ( x, γ 3 ) , that is the third central moment of

production function (4); (c) and (d) represent the counterfactual expected downside risk exposure. Standard errors in

parentheses. *** Significant at the 1% level.

28
Appendix

Table A1. Variables Definition

Variable name Definition

Dependent variables
adaptation dummy =1 if the farm household adapted to climate change, 0 otherwise
downside risk exposure skewness f 3 ( x, γ 3 ) : third central moment of production function (4) / 10,000,000,000

Explanatory variables
Soil characteristics
high fertility dummy =1 if the soil has a high level of fertility, 0 otherwise
infertile dummy =1 if the soil is infertile, 0 otherwise
no erosion dummy=1 if the soil has no erosion, 0 otherwise
severe erosion dummy=1 if the soil has severe erosion, 0 otherwise
Assets
machinery dummy =1 if machineries are used, 0 otherwise
animals dummy=1 if farm animal power is used, 0 otherwise
Inputs
labor labor use per hectare (adult days)
seeds seeds use per hectare (kg)
fertilizers fertilizer use per hectare (kg)
manure manure use per hectare (kg)
Farm head and farm household
characteristics
literacy dummy =1 if the household head is literate, 0 otherwise
male dummy =1 if the household head is male, 0 otherwise
married dummy =1 if the household head is married, 0 otherwise
age age of the household head
household size household size
off-farm job dummy =1 if the household head took an off-farm job, 0 otherwise
relatives number of relatives in the woreda
Information sources
government extension dummy =1 if the household head got information/advice from government extension workers, 0
otherwise
farmer-to-farmer extension dummy =1 if the household head got information/advice from farmer-to-farmer extension, 0
otherwise
radio information dummy =1 if the household head got information from the radio, 0 otherwise
climate information dummy =1 if extension officers provided information on expected rainfall and temperature, 0
otherwise
Fixed effects
average fertilizers average fertilizer use per hectare (kg)
average seeds average seed use per hectare (kg)
average manure average manure use per hectare (kg)
average labor average labor use per hectare (adult days)

29
Table A2. Parameter Estimates – Test on the Validity of the Selection Instruments

Model 1 Model 2
Adaptation 1/0 Third central moment of
production function among
farm households that did not adapt
Soil characteristics
highly fertile -0.208* -0.019
(0.108) (0.011)
infertile 0.091 0.001
(0.164) (0.022)
no erosion 0.066 0.032
(0.142) (0.022)
severe erosion 0.189 0.025
(0.136) (0.043)
Assets
machinery 0.535 -0.038
(0.481) (0.032)
animals 0.162 0.031
(0.189) (0.028)
Inputs
labor 0.0001
(0.0002)
squared labor /100 -0.00005
(0.00003)
seeds -0.0005*
(0.0002)
squared seeds /100 0.0001*
(0.00005)
fertilizers 0.00004
(0.0001)
squared fertilizers/100 -0.000001
(0.000003)
manure -0.00001
(0.00004)
squared manure /100 0.0000004
(0.000001)
Farm head and
farm household characteristics
literacy 0.153 -0.045
(0.150) (0.032)
male -0.040 0.096
(0.329) (0.074)
married -0.265 -0.134
(0.293) (0.100)
age 0.013** -0.001
(0.006) (0.001)
household size 0.032 0.006
(0.035) (0.007)
off-farm job 0.341*** -0.005
(0.125) (0.026)
relatives 0.009** 0.001
(0.004) (0.001)
Information sources
government extension 0.592*** -0.022
(0.116) (0.064)
farmer-to-farmer extension 0.539*** 0.030

30
 (0.141) (0.073)
radio information 0.501** -0.004
(0.203) (0.024)
climate information 0.625*** 0.105
(0.172) (0.091)
constant -1.272*** 0.065
(0.374) (0.056)
Wald test on farm head and
2 = 27.33*** F-stat. = 1.09
farm household characteristics
Wald test on information sources 2 = 93.16*** F-stat. = 1.49

Sample size 2,807 872

2 2
Notes: Model 1: Probit model (Pseudo R = 0.299); Model 2: ordinary least squares (R = 0.033). Estimation at the plot-
level. Standard errors clustered at the woreda level in parentheses. * Significant at the 10% level; ** Significant at the 5%
level; *** Significant at the 1% level.

31
Table A3. Parameters Estimates of Production Function (4) in the two Regimes

(Regime 1) (Regime 2)
Adaptation = 1 Adaptation = 0
Quantity produced per hectare by Quantity produced per hectare by
Dependent Variable
farm households that adapted farm households that did not adapt
Soil characteristics
highly fertile 208.522 113.086
(155.573) (74.408)
infertile -140.102** 25.388
(59.665) (60.644)
no erosion 72.225 -49.288
(101.778) (50.821)
severe erosion 68.605 32.007
(134.874) (90.459)
Assets
machinery -246.410 -190.328
(226.610) (134.288)
animals 230.069 145.616*
(164.245) (86.314)
Inputs
labor 3.220*** 3.819***
(1.088) (0.544)
squared labor /100 -0.119 -0.409***
(0.077) (0.083)
seeds 2.440*** -0.358
(0.903) (0.732)
squared seeds /100 0.043 0.386***
(0.041) (0.146)
fertilizers 0.720 0.884**
(0.480) (0.387)
squared fertilizers/100 -0.008 -0.025***
(0.013) (0.009)
manure 0.287*** 0.104
(0.109) (0.147)
squared manure /100 -0.003** 0.002
(0.001) (0.002)
Farm head and farm
household characteristics
literacy -70.585 -150.039***
(179.156) (54.392)
male 208.080 314.707***
(196.622) (74.994)
married -34.432 -295.703**
(249.831) (129.460)
age -2.425 -5.232***
(5.616) (1.394)
household size -9.798 -9.978
(33.486) (13.299)
off-farm job 195.416 7.494
(166.285) (87.423)
relatives 0.236 -0.976
(0.842) (2.861)
access to credit 1/0 -57.651 -265.288***
(238.195) (71.007)
having gold 1/0 -134.147 -96.294
(314.342) (89.900)
constant 218.021 527.073***

32
 (1,161.982) (178.647)
Note: Switching regression model with endogenous switching estimated by full information maximum
likelihood at the plot-level. Sample size: 2,807 plots. Robust standard errors clustered at the woreda level
in parentheses. * Significant at the 10% level; ** Significant at the 5% level; *** Significant at the 1%
level.

33