Content uploaded by Divina Gracia Rodriguez
Author content
All content in this area was uploaded by Divina Gracia Rodriguez on Oct 09, 2017
Content may be subject to copyright.
Impact of the alternate wetting and drying (AWD) water-saving irrigation
technique: Evidence from rice producers in the Philippines
Roderick M. Rejesus
a,
⇑
, Florencia G. Palis
b
, Divina Gracia P. Rodriguez
c
, Ruben M. Lampayan
d
,
Bas A.M. Bouman
d
a
Department of Agricultural and Resource Economics, North Carolina State University, Raleigh, NC 27695, United States
b
Social Science Division, International Rice Research Institute (IRRI), Los Baños, Laguna, Philippines
c
Department of Agricultural and Consumer Economics, University of Illinois at Urbana-Champaign, Urbana, IL 61801, United States
d
Crop and Environmental Sciences Division, International Rice Research Institute (IRRI), Los Baños, Laguna, Philippines
article info
Article history:
Received 17 March 2010
Received in revised form 24 November 2010
Accepted 29 November 2010
Available online 30 December 2010
Keywords:
Alternate wetting and drying
Water saving
Irrigation
Rice production
Technology
abstract
This article evaluates the impacts of a controlled irrigation technique in rice production called alternate
wetting and drying (AWD). Propensity score matching (PSM) and regression-based approaches applied to
farm-level survey data are used to achieve the objective of the study. The PSM and regression-based
approach accounts for the potential bias due to selection problems from observable variables. Results
of the impact analysis using both empirical approaches indicate that AWD, particularly the ‘‘Safe
AWD’’ variant, reduces the hours of irrigation use (by about 38%), without a statistically significant reduc-
tion in yields and profits. This reduction in irrigation time translates to corresponding savings in the
amount of irrigation water and pumping energy used. However, further analysis of the impact estimates
suggests that the potential magnitude of the selection bias based on unobservable variables may still be
able to eliminate the measured impact from the PSM and regression-based techniques that only control
for selection based on observable variables. Hence, the current impact results have to be interpreted with
caution and further data collection is needed to construct a panel data that would allow one to account
for selection problems due to unobservable variables and, consequently, better estimate the AWD impact.
Ó2010 Elsevier Ltd. All rights reserved.
Introduction
Irrigation water in Asia is becoming increasingly scarce. Rapid
population growth and multiple competing demands for water
(i.e., drinking, industrial uses) have contributed to irrigation water
scarcity in many Asian developing countries, including the Philip-
pines (Pingali et al., 1997; Van der Hoek et al., 2000; Tabbal
et al., 2002). Tuong and Bouman (2003) estimate that, by 2025,
about 2 million ha of Asia’s irrigated dry-season rice and 13 mil-
lion ha of its irrigated wet-season rice will experience physical
water scarcity. However, as population continues to rise in Asia,
more irrigation water may be needed to increase total food produc-
tion and meet growing food demand in the future (Rosegrant and
Ringler, 1998). Therefore, in order to help meet the food demand
of a rapidly growing population amidst this increasing water scar-
city, more efficient water management practices – water-saving
technologies – are needed so that rice production levels in Asia
(i.e., the main staple food in the continent) can still be maintained
or increased even with the use of less irrigation water.
One such water-saving technology that has been developed for
rice cultivation in Asia is the alternate wetting and drying (AWD)
irrigation approach (Belder et al., 2004; Bhuiyan, 1992; Bouman
and Tuong, 2001). AWD is an irrigation technique where water is
applied to the field a number of days after disappearance of ponded
water. This is in contrast to the traditional irrigation practice of
continuous flooding (i.e., never letting the ponded water disap-
pear). This means that the rice fields are not kept continuously sub-
merged but are allowed to dry intermittently during the rice
growing stage. The number of days where the field is allowed to
be ‘‘non-flooded’’ before irrigation is applied can vary from 1 day
to more than 10 days. The underlying premise behind this irriga-
tion technique is that the roots of the rice plant are still adequately
supplied with water for some period (due to the initial flooding)
even if there is currently no observable ponded water in the field.
In the Philippines, approximately 61% of the 3.4 million ha of
rice production area is under irrigation (IRRI, 1997). A key irrigated
rice production region is the Central Luzon area, which is about
70 km north of Manila (the capital city of the Philippines). In light
of the concerns about irrigation water scarcity in the area, the
International Rice Research Institute (IRRI), the National Irrigation
0306-9192/$ - see front matter Ó2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.foodpol.2010.11.026
⇑
Corresponding author. Address: Department of Agricultural and Resource
Economics, North Carolina State University, NCSU Box 8109, Raleigh, NC 27695-
8109, United States. Tel.: +1 919 513 4605; fax: +1 919 515 1824.
E-mail address: rod_rejesus@ncsu.edu (R.M. Rejesus).
Food Policy 36 (2011) 280–288
Contents lists available at ScienceDirect
Food Policy
journal homepage: www.elsevier.com/locate/foodpol
Administration (NIA) of the Philippines and the Philippine Rice Re-
search Institute (PhilRice) introduced the AWD irrigation tech-
nique to farmers in the Central Luzon area in 2001. As it has
been several years since the introduction of the AWD technique
in the area (at the time of our survey), and given the concerns
about food security in Asia and the Philippines, an evaluation of
the impact of this technology would be valuable.
The objective of this paper is to evaluate the impact of the AWD
irrigation technique on irrigation use (e.g., number of hours and
frequency), labor, yields, and profits. Cross-section farm survey
data from a major rice producing region in the Philippines is used
to achieve this objective. Since we only have cross-section data at
one point in time (i.e., no baseline survey data prior to AWD adop-
tion), we mainly rely on methods based on comparison of groups
(i.e., comparison of AWD adopters versus non-adopters, rather
than comparison before and after AWD adoption or a combined pa-
nel data approach). Given this limitation, we utilize propensity
score matching (PSM) and regression-based techniques to deal
with selection issues typically present in impact analysis that uti-
lizes comparison of groups.
There have been a number of studies that specifically examined
the decision to adopt various water-saving irrigation technologies
(see, among others, Casswell, 1991; Palis et al., 2004; Blanke et al.,
2005; Koundouri et al., 2006) and the potential impact of these
water-saving technologies (see, for example, Peterson and Ding,
2005). A recent paper by Abdulai et al. (2005), for example, studied
the adoption of a water-saving technology specifically for rice pro-
duction in China called the Ground Cover Rice Production System
(GCPS), although its impact was not evaluated.
In addition, there have been several studies that specifically
examined the impact of the AWD irrigation approach on water
use and/or rice yields primarily using field experiments. Zhi
(2001) explored the impact of AWD on water use and found that
irrigation water use was reduced by 7–25% with the AWD tech-
nique. Singh et al. (1996) reported that, in India, the AWD irriga-
tion approach can reduce water use by about 40–70% compared
to the traditional practice of continuous submergence, without a
significant yield loss. Studies by Cabangon et al. (2001) and Moya
et al. (2004) in China found similar results.
In contrast, a synthesis by Bouman and Tuong (2001) of 31 pub-
lished experiments on various forms of AWD revealed that 92% of
the AWD treatments resulted in yield reductions ranging from zero
to 70%, relative to the flooded control plots. The large range in re-
sults was caused mainly by the wide range in ‘severity’ of AWD
treatments, ranging from mild water savings with little stress on
the crop to large water savings with a severe impact on crop growth.
Other recent AWD studies that are based on field experiments also
reveal that some forms of AWD can reduce rice yields (see Tabbal
et al., 2002; Belder et al., 2004; Cabangon et al., 2004). There is a
specific form of AWD called ‘‘Safe AWD’’ that has been developed
to potentially reduce water inputs by about 30%, while maintaining
yields at the level of that of flooded rice (Bouman et al., 2007). In
Safe AWD, the ponded water on the field (also called ‘‘perched
water’’) is allowed to drop to 15–20 cm below the soil surface before
irrigation is applied. The depth of perched water is monitored using
a perforated or punctured water tube embedded in the soil. With
the threshold of 15–20 cm, roots are still able to extract water from
the perched water table and no stress to the plants develops. In Safe
AWD, each irrigation will flood the field to about 2–5 cm (in con-
trast to the 5–10 cm for traditional irrigation). During flowering,
the field is kept flooded so as to avoid spikelet sterility. This specific
AWD variant is the one typically used in the study area.
Notwithstanding all the academic papers that investigated
AWD in different countries, to the best of our knowledge none have
explicitly investigated the economic impact of the AWD water-sav-
ing irrigation technique in the context of rice production in a
developing country like the Philippines and none have utilized
the empirical approaches we use.
1
Furthermore, most of the AWD
yield impact studies cited above are based on controlled field exper-
iments that used simple ‘‘with AWD or without AWD’’ comparisons
without controlling for selection on observables. Hence, these stud-
ies do not utilize farm-level survey data to examine actual behavior
of farmers and the impact of the AWD irrigation approach. In this re-
gard, this paper contributes to the literature by utilizing PSM and
regression-based approaches that control for selection on observ-
ables in the impact analysis of the AWD irrigation method.
Background, survey design, and data description
AWD technology: overview and dissemination approach
As mentioned in the introduction, one of the main irrigated rice
production region in the Philippines is found in Central Luzon. The
Central Luzon region has an area of about 21,470 sq. km. spread
over seven provinces. Traditionally, irrigation in this area is done
by tapping into surface water through gravity, such as run-off
the river and reservoir-backed systems. Irrigation in this area is
now typically through shallow and deep tube wells. To develop
new water resources, the National Irrigation Administration
(NIA) developed 94 deep wells in Tarlac province between 1971
and 1983. The physical availability of irrigation water in the area
has been threatened over the last few decades due to the follow-
ing: (1) increased diversion of water from local reservoirs to the
country’s capital (Manila) for non-irrigated uses, (2) increased pol-
lution that makes water quality unsuitable for irrigation, and (3)
destruction and clogging of existing irrigation systems due to
earthquakes and the Mt. Pinatubo volcano eruption in 1991 (see
Tabbal et al., 2002). In response to the latter, NIA initiated a re-acti-
vation project for the deep wells in Tarlac in 1996, which eventu-
ally made the deep well system one of the main sources of
irrigation in the area.
To help address the irrigation water scarcity in Central Luzon,
various agencies (IRRI, NIA, and PhilRice) worked together to de-
velop practical water-saving techniques. In 2001, these local and
international agencies introduced the AWD technique to farmers
in Tarlac province who were members of a particular Irrigators’
Service Cooperative (ISC) with one of the deep well system. A par-
ticipatory approach (i.e., farmer–scientist partnerships) was used
to help validate and promote the AWD irrigation technique. In
2004, after several years of study in one ISC, AWD was then intro-
duced to several other deep well ISCs. The selection of farmers
(within ISCs), to initially introduce the AWD technology to, was
based on the farmer’s motivation and willingness to participate,
and on site criteria like accessibility, spread of farmers across the
site, position on the toposequence, and nearness to the irrigation
pump. A special effort was made to select farmers on different
toposequence positions (high, middle, low elevation) to capture
differences in groundwater status and soil type since these are ex-
pected to affect the actual number of days the crop can be without
standing water.
In most deep well rice irrigation systems, water is distributed to
the service area rotationally, where irrigation water is made avail-
able to groups of farmers once a week (after transplanting until
close to harvest), and the farmer usually uses this allocation to
maintain 6–8 cm of ponded water after irrigation. Irrigation sched-
ules for both AWD and traditional (non-AWD) systems follow the
rotational irrigation schedule of the ISCs (which is agreed upon
1
Although this study only focused on the Philippines, the conditions in the study
area are typical for irrigated rice farming in Asia (especially in other Asian developing
countries). Hence, our results may also provide implications that can be of use to
other irrigated rice producing regions in Asia.
R.M. Rejesus et al. / Food Policy 36 (2011) 280–288 281
by the member-farmers in the ISC). However, to differentiate the
water management of the AWD farms versus the non-AWD farms,
the amount of irrigation water made available to the AWD farms is
typically about 30–40% lower than the water made available to the
non-AWD farms at each ISC. This reduction of 30–40% is based on
an initial pilot study of the Safe AWD variant introduced in the first
ISC (Lampayan et al., 2004). Given the rotational schedule, the
AWD farmers monitor the depth of the water table in their fields
to check that the selected amount of water reduction did not lead
to perched water tables dropping too much below the 20 cm
depth that characterizes Safe AWD.
Note that both the AWD and non-AWD members of the ISC have
to buy the required fuel to operate the irrigation pump and irrigate
their fields at their scheduled times. If a farmer does not have the
money to buy the required fuel during his/her scheduled time (or if
he/she does not want to irrigate for that week), then the pump is
not turned on at his/her scheduled irrigation time and he/she will
not be able to irrigate his/her fields for that week. The farmer has
to decide whether or not he will irrigate in a particular week (i.e.,
he/she decides the number of times or number of weeks to irrigate
throughout the season) and the number of hours he/she wants to
irrigate, given the knowledge about the water made available at
his/her scheduled time every week.
Description of the sampling approach and data collection
The data set used in this study is from a 2005 farm-level survey
conducted by IRRI in Tarlac province of the Philippines. A two-
stage random sampling approach was employed in the selection
of farmer-respondents. The first stage was the random selection
of deep-well pumps in the province. There were 25 pumps ran-
domly selected out of 50 pumps in the province which were either
newly established or newly rehabilitated by the national irrigation
authority (NIA) of the Philippines. Each deep-well pump may serve
20–50 farmers. Second, a proportional random selection of farmer-
respondents was done for each of the deep-well pumps. Sample
sizes at each pump were proportioned based on the number of
farmers serviced within each of the pump. A total of 194 farmers
using deep-well irrigation were randomly selected and inter-
viewed. However, due to data constraints (i.e., missing informa-
tion, etc.), only data from 146 farmer-respondents are used in
the analysis. Thirty producers are AWD adopters and 116 are
non-adopters. AWD adopters are those farmers who currently
practice AWD; otherwise they are AWD non-adopters.
The data were collected for the 2005 dry season, from Decem-
ber 2004/January 2005 to March/April 2005. The data collected
has information about on-farm water management strategies such
as frequency of irrigation, number of hours of irrigation, and the
distance from the irrigation pump. Various input and output data
for rice production was also collected, such as input use (i.e. labor,
fertilizer, etc.), rice production output, input costs, and rice reve-
nues. In addition, information was collected about pertinent so-
cio-demographic characteristics of the farm like the household
size, tenure status, farm size, and exposure to irrigation training.
The farm survey was complemented by focus group discussions
(FGD) and key informant interviews (KII) of farmers, NIA officials,
and officers of the deep well ISCs.
2
Estimation strategies and empirical specification
Selection on observables and unobservables
The purpose of the estimation that follows is to measure the im-
pact of AWD adoption on several outcome variables of interest (see
‘Empirical specification’ below for a detailed description of these
variables). This is the so-called average treatment effect on the
treated (ATT) where the treatment in this case is adoption of the
AWD irrigation technique. To obtain an accurate assessment of
the AWD impact, one would ideally look at the difference between
the outcomes for farmers that use AWD and the outcomes from the
same farmers had they not adopted the AWD irrigation technique.
The empirical challenge we face is the typical one of filling missing
data on the counterfactual – what would have been the outcome if
the AWD adopters had not used the AWD irrigation method? In our
present context, and given the cross-sectional nature of our data,
one would need to identify a suitable comparison group of non-
AWD farmers whose outcomes on average provide an unbiased
estimate of the missing counterfactual. Given the non-random
adoption of AWD among farmers in Tarlac province, simple com-
parison of mean outcomes between AWD and non-AWD farmers
would yield biased estimates of impact.
There are two main sources of bias in simple comparison of
mean outcomes in the context of AWD adoption: (1) selection on
observables and (2) selection on unobservables.
3
First, AWD adopt-
ers are likely to differ from non-adopters in the distribution of their
observed characteristics leading to a bias from ‘‘selection on observ-
ables’’. This bias is likely to arise in our context since the selection
criteria for the initial dissemination of this technology in the region
(i.e., toposequence, nearness to irrigation pump) can also be ex-
pected to influence the outcome variables directly (i.e., direct effects
on irrigation levels, yields, and profits) even in the absence of the
program. As discussed further below, PSM techniques can control
selection on observables and provide an unbiased measure of AWD
impact on the AWD adopters (i.e., the ATT) under the assumption
of conditional mean independence. Conditional mean independence
is when pre-adoption outcomes are independent of AWD adoption
given the observable characteristics used in the matching procedure.
This is likely true in our case since AWD adoption rates in the area
are still quite low (for certain remote ISCs) and non-adopters in
these ISCs would adopt AWD if they have more information about
this technology.
The second source of potential bias in the simple comparison of
mean outcomes is from the difference in the distribution of unob-
servable characteristics between the AWD and non-AWD adopters,
where the unobservable characteristics affects both the decision to
adopt AWD and the outcomes of interest. For example, unobserved
motivation or unobserved managerial ability of farmers may affect
both the decision to adopt AWD, as well as the irrigation use or
yield outcomes. That is, the measured effect of AWD on the out-
comes may just be due to the difference in the unobservable
variables, rather than being due to AWD itself. In a cross-sectional
context, controlling for selection problems due to unobservables
would be straightforward if strong instrumental variables exists
that only affect AWD adoption but not the outcome of interest.
Unfortunately, we do not have any strong instruments in our data
2
In the interest of space, the results of the FGD and KII conducted in 2005 are not
reported here. But note that the results of our empirical analysis below are consistent
with the findings from the FGD and KII. The FGD and KII showed that farmers perceive
a significant reduction in water use without significant reduction in yields and profits
when AWD is used. However, the FGD and KII results show that the selected farmers/
key informants believe AWD significantly affects labor use, which we do not find
empirical evidence for. These FGD and KII results are consistent with the earlier FGD
and KII conducted in 2003 for the pilot area where AWD was initially introduced (see
Palis et al., 2004).
3
Another source of potential bias common in the program evaluation literature is
the bias due to diffusion. That is, the effect of treatment ‘‘diffuses’’ to the control
group such that treatment affects the outcome of the treated group directly and the
outcome of the control group is indirectly affected as well (i.e. the so-called Stable
Unit Value Assumption in the literature). In the context of this study, there is no
possibility of diffusion bias since the non-AWD farmers truly did not adopt the AWD
technology. Hence there can be no ‘‘contamination’’ in the outcomes of the controls
due to diffusion bias.
282 R.M. Rejesus et al. / Food Policy 36 (2011) 280–288
and are unable to explicitly control for selection on unobservables
in the estimation strategies discussed below. However, we follow
the procedures in Altonji et al. (2005) to assess the magnitude of
the potential bias due to selection on unobservables. This
additional analysis provides some guidance as to the relative size
of the selection bias from unobservables and whether this
selection bias can wipe out the measured effect from our
empirical approaches that only controls for selection on
observables.
Impact analysis by propensity score matching (PSM)
The PSM technique introduced by Rosenbaum and Rubin (1983)
is the primary approach used in this study to control for selection
bias based on observable characteristics. The basic idea behind the
PSM method is to find control observations (i.e., non-AWD farm-
ers) having observable characteristics as similar as possible to
the treatment group, to serve as valid surrogates for the missing
counterfactuals. Please see Wooldridge (2002, pp. 614–620) for a
more formal and detailed description of the PSM approach.
In this study, we follow common practice in the matching liter-
ature by using a parametric binary response model (a probit model
in our case) to estimate the propensity score for each observation
in the treatment (AWD adopters) and control (non-AWD adopters)
groups (see Sianesi, 2001; Becker and Ichino, 2002). A rich set of
observable covariates are used to estimate the propensity scores,
with special focus on the observable variables used as the criteria
for the initial dissemination of the AWD technology in the area
(e.g., distance from the pump and toposequence/elevation) and
the factors used in previous literature that studied irrigation tech-
nology adoption. The ‘‘balancing property’’ of the observables used
in the probit specification is then tested to ensure that observa-
tions with similar propensity scores have the same distribution
of observable characteristics independently of whether or not
AWD is adopted. In other words, for a given propensity score, sat-
isfying the balancing property indicates that AWD adoption is ran-
dom and, thus, the AWD adopters and non-adopters should be on
average observationally identical.
Using the estimated propensity score from the probit model, the
AWD adopters in the sample are then matched to non-adopters
with sufficiently similar propensity scores. Following usual prac-
tice in the PSM literature, we use two common matching ap-
proaches to determine the non-AWD adopters who match the
AWD adopters and calculate the ATT: (1) kernel matching and
(2) nearest-neighbor matching. Using different matching ap-
proaches helps us determine the robustness of our results. A ‘‘com-
mon support’’ constraint is imposed during the matching
procedure whereby AWD adopters are dropped when they have
propensity scores above the maximum or below the minimum pro-
pensity score for the comparison group (non-adopters). Imposing
the common support restriction often improves the quality of the
matches used to estimate the ATT (Becker and Ichino, 2002). Once
the matched estimates are chosen, the ATT can then be estimated
and used as the measure of the impact of AWD while controlling
for selection on observables.
Robustness check: regression-based impact analysis
Although we already undertake robustness checks on our im-
pact results by using alternative matching approaches in the PSM
procedure described above, we also utilize the regression-based
approach described in Godtland et al. (2004) that allows one to
consistently estimate the ATT while controlling for selection on ob-
servable characteristics (also see Wooldridge, 2002, pp. 611–613).
This regression-based method is a straightforward ordinary least
squares (OLS) estimation of the following model specification:
y
i
¼
c
þ
a
AWD
i
þbx
i
þdðx
i
xÞAWD
i
þ
e
i
;ð1Þ
where
xis a vector of the observable variable means of the treated
population (AWD adopters), and
c
,
a
,b,dare unknown parameters
to be estimated. In (1), the parameter estimate ^
ais the ATT measure
of the AWD impact. This regression-based approach allows us to
check the robustness of our impact estimates from the earlier
PSM approaches.
Assessing the selection bias from unobservables
The PSM approach and the regression-based approach de-
scribed above only allow us to account for selection bias due to ob-
servable variables. Given that we only have a cross-sectional data
set and we do not have any suitable instruments, we cannot explic-
itly control for selection on unobservables. Nonetheless, we use a
recently developed informal procedure by Altonji et al. (2005) to
assess the magnitude of the potential bias due to selection on
unobservables and whether this bias magnitude is large enough
to essentially eliminate the measured AWD impact that only con-
trols for selection on observables (i.e., the observed AWD impact
can possibly be due to selection bias from unobservables even if
we controlled for observables).
4
This procedure relies on a calculated ratio measure (called
s
)
that estimates the shift in the distribution of unobservables re-
quired to explain away the entire observed AWD effect based on
the observable variables. If
s
is substantially higher than 1.0, then
the shift in unobservables has to be substantially larger than shift
in the observables to invalidate the measured AWD impact. This
signifies that selection on unobservables may not be a big issue
in this case (i.e., it is likely that the observed impact is true and
not due to selection bias from unobservables) (for example, see re-
sults in Godtland et al., 2004 and Altonji et al., 2005). But if the
measure is close to 1.0 or less than 1.0, then this means that it
takes the same or a smaller shift in unobservables to eliminate
the AWD impact and the estimated AWD may impact may simply
be due to selection on unobservables (for example, see results in
Altonji et al., 2008).
Empirical specification
To implement the PSM approach described above, we first need
to identify the key explanatory variables in the probit model of
AWD adoption (AWD = 1 if adopter, zero otherwise). Following
previous studies of water-saving technology adoption in agricul-
ture (see Anderson et al., 1999 and Abdulai et al. (2005)) and the
information from the way AWD was disseminated in Tarlac prov-
ince, we include the following explanatory variables in the AWD
probit adoption model: age (Age), education (Educ), farming expe-
rience (Experience), land ownership (Own Land), elevation dum-
mies (High Elevation, Medium Elevation), whether or not the soil
is sandy (Sandy Soil), off-farm income (Off-farm income), household
size (Household Size), total rice area farmed (Rice Area), price of fuel
for irrigation (Irrig Fuel Price), attendance in a training about irriga-
tion (Training), size of the area served by the cooperative-owned
deep-well irrigation pump used (Service Area), and distance of
the farm from the pump (Distance). The age, education, farming
experience, household size, ownership dummy, and off-farm are
farmer/farm characteristics typically included in most general
technology adoption studies (as is done in the water-saving adop-
tion studies mentioned earlier as well). The service area, elevation,
sandy soil, and distance to pump variables are factors affecting the
4
The interested reader is referred to the original paper of Altonji et al. (2005) for a
formal treatment of this procedure and to the paper by Godtland et al. (2004) for a
more concise description of the rationale for this test.
R.M. Rejesus et al. / Food Policy 36 (2011) 280–288 283
effectiveness of the irrigation system used by a particular farmer
and we believe that his affects the decision of whether or not to
adopt an AWD system. The training dummy, on the other hand, re-
flects whether or not a producer has been exposed to training that
focus on irrigation management.
5
Once the parameters of the probit model have been estimated,
these are used to calculate a propensity score to implement the
PSM procedure. The outcome variables of interest investigated
are: the number of hours of irrigation for the season (Hours Irriga-
tion), the frequency of irrigation for the season (Irrigation Fre-
quency), total labor use (Labor), yield (Yield), and profits (Profit).
6
The hours and frequency of irrigation is of interest since we want
to know whether AWD does indeed reduce irrigation application
in practice. Labor is the main input of interest because the traditional
irrigation technique of continuous field submergence helps farmers
control weeds and helps reduce labor associated with weed control
(i.e., hand weeding). As shown in Moya et al. (2004), the AWD did
somewhat increase labor use in China although it did not seem to af-
fect the other inputs (which is why we focus on the labor input as an
outcome variable here). The yield and profit outcomes are also of
interest to determine whether the reduction in water application
through AWD adversely affects yields and profits.
The regression based empirical specification in (1) also uses the
four aforementioned outcome variables as the dependent variable,
with the explanatory variables in the probit adoption specification
used as the observable covariates included in vector x
i
and the
interaction term vector ðx
i
xÞAWD
i
. The explanatory variables in-
cluded in the regression-based modeling approach are all assumed
to be exogenous farm-level characteristics that are hypothesized to
affect the outcome variables of interest.
7
A detailed description of all the variables used in the empirical
analysis (i.e., the probit model and the regression-based impact ap-
proach) is shown in Table 1. Summary statistics for the pertinent
variables are presented in Table 2 for the whole sample (n= 146),
the AWD adopters (n= 30), and non-AWD adopters (n= 116).
Results and discussion
Probit model results: determinants of AWD adoption
The results of the AWD probit model are presented in Table 3.
Our results suggest that being exposed to irrigation training has
a statistically significant positive effect on the decision to adopt
AWD technology (at the 5% significance level). The strongly signif-
icant effect of this variable suggests that this may be the main fac-
tor in driving the adoption of AWD in the area. The statistically
significant positive effect of the irrigation training dummy is con-
sistent with our a priori expectations since these farmers have
more information about the benefits and costs of an AWD irriga-
tion technique. They are more aware of the potential irrigation cost
savings and water conservation benefits that can be achieved
through this technique. Hence, the trained producers are more
likely to adopt AWD even though there is a possibility of having
lower yields since less water is applied to the rice crop. In addition
to the statistically significant training effect, the sandy soils vari-
able and the off-farm income variable also tend to have a margin-
ally statistically significant effect (at the 10% significance level) on
the probability of adopting the AWD technique.
PSM impact results
Using the estimated parameters from the probit model above,
propensity scores are computed to be able to conduct the PSM
methods and find matched non-AWD adopters that serve as good
surrogates for the missing counterfactuals. The region of common
support for the probit model specification is between the interval
0.05 and 0.93. Given this common support, three AWD observa-
tions were outside of this of the common support and were not in-
cluded in the PSM impact analysis (i.e., only 27 observations are
considered in the PSM). As shown in Appendix Table A1, the bal-
ancing property is satisfied for the probit specification used in
the study since no observable variables have significantly different
means within the two different strata examined (the strata being
defined by the ordered propensity scores). That is, the observable
characteristics of the AWD and non-AWD adopters are not signifi-
cantly different after matching. This indicates that, based on the
observables and given a propensity score, the AWD and non-
AWD adopters are observationally identical.
Given that the balancing property is satisfied and the common
support imposed, we use the estimated propensity scores from
the probit model to conduct kernel matching and nearest-neighbor
matching that allows one to find matched AWD and non-AWD
Table 1
Variable definitions.
Variable name Definition
AWD =1 if adopt alternate wetting and drying (AWD); =0
otherwise
Age Age of household head
Educ Years of education for household head
Experience Years of farming experience
Own Land =1 if own land; =0 otherwise
High Elevation =1 if high elevation; =0 otherwise
Medium
Elevation
=1 if medium elevation; =0 otherwise (low elevation
omitted)
Sandy Soil =1 if soil is sandy; =0 otherwise
Off-farm Income Non-farm income (in ‘000 PhP)
Household Size No. of household members
Rice Area Total area planted to rice (in ha)
Irrig Fuel Price Price of fuel used for irrigation
Training =1 if attended water/irrigation training before; =0
otherwise
Service Area The area (in ha) served by the pump used by the farm
Distance Distance of the farm to the pump (in meters)
Hours Irrigation No. of hours irrigating per ha (for the season)
Irrigation Freq Frequency of irrigation per ha (for the season)
Labor Total labor used in production (in mandays/ha)
Yield Rice yield in kg/ha
Profits Profits (income less variable costs) in PhP/ha
5
In specifying the probit model for PSM analysis, the literature provides no specific
guidance as to what type of variables is to be included in the specification. Although,
including variables that fundamentally affect participation is often suggested to get
better propensity score estimates. In addition, ‘‘overparameterization’’ (i.e., including
as much variables as possible even if these variables do not intuitively affect
participation) is also commonly suggested to improve the propensity score estimates
and, ultimately, the matches to the treated group.
6
The irrigation, labor, and yield variables are the actual data collected from the
survey of the farmers. The profit variable is calculated by using the reported revenues
less the variable production costs (e.g., fertilizer, chemicals, labor, tractor rental, fuel
for irrigation). But note that the variable production costs did not include depreci-
ation costs and interest costs. The opportunity cost of the irrigation water used was
also not accounted for in the profit calculation (see Footnote 11). Profit in this case is
the net cash returns above variable cash costs. A detailed listing of all the variable cost
items included in the profit calculation is available from the authors upon request.
7
Note that the Training dummy variable may not be truly exogenous in the
empirical specification in (1). Controlling for the possible endogeneity of training
would require the use of instrumental variable (IV) techniques. But unfortunately, we
do not have strong instruments to properly account for this issue. This should be
recognized as a potential limitation of the study and a subject for future research.
However, we want to emphasize that the main objective of this study is to accurately
estimate the AWD impact and including the training variable in the regression based
analysis in (1) helps achieve this (i.e., this helps address the ‘‘selection on
observables’’ issue). Mor eover, as seen below, th e results from the PSM and
regression-based approach are fairly similar, which suggests that the bias in the
AWD parameter due to the potential endogeneity of the training variable may be
minimal.
284 R.M. Rejesus et al. / Food Policy 36 (2011) 280–288
adopters. For the kernel matching, an epanechnikov kernel is used
with a bandwidth at 0.06.
8
For the nearest-neighbor matching pro-
cedure, 10 matching nearest neighbors are utilized within the caliper
value of 0.05. After completing the matching procedures, comparison
of means for the observable variables are undertaken to validate that
the mean observed characteristics of the AWD adopters are not sig-
nificantly different from the non-AWD adopters (i.e., the matched
non-AWD are good surrogate counterfactuals). Results of the com-
parison of means for the matched sample indicate that there are
no observable characteristics that are significantly different (at the
5% level of significance) between the AWD and non-AWD adopters
(see Table 4).
9
In contrast, when the unmatched sample is used, the
Training and High Elevation variables are significantly different be-
tween AWD and non-AWD adopters.
10
Hence, the results in Table
4suggest that the matched non-AWD observations can serve as a
good surrogate for the missing counterfactual, which would allow
more accurate estimation of the AWD impact.
The impact of AWD on irrigation (hours and frequency), labor,
yields, and profits using the PSM methods are presented in Table
5. The results from the PSM method (i.e. matched sample) indicate
that there is a statistically significant difference between the AWD
and non-AWD adopters’ hours of irrigation (at the 5% level in the
kernel PSM and at the 10% level for the nearest neighbor PSM).
Our mean comparisons using the matched sample show that
AWD adopters use about 25 h less irrigation (about 38% less than
the mean for the full sample). The statistically significant negative
effect of AWD on hours of irrigation is expected given the nature of
the AWD technique where water is applied only when ponded
water disappears compared to the traditional irrigation approach
of continuous flooding. On the other hand, the frequency of irriga-
tion does not seem to statistically differ between the AWD and
non-AWD users based on the PSM techniques (although, for the
kernel matched sample, the difference in irrigation frequency is
marginally significant at the 10% level). Without using matching
and not controlling for selection on observables, the irrigation im-
pact results seem to be switched (i.e., strong statistical difference
in frequency of irrigation but not in the hours of irrigation). In gen-
eral, the PSM irrigation results suggest that when accounting for
selection on observables, adoption of the AWD method results in
a reduction in hours of irrigation, but it does not necessarily reduce
Table 2
Summary statistics: full sample, AWD adopters, and non-AWD adopters.
Variable Full sample (n= 146) AWD adopters (n= 30) Non-AWD adopters (n= 116)
Mean St. Dev. Mean St. Dev. Mean St. Dev.
AWD 0.20 0.40 – – – –
Age 48.80 13.75 51.37 12.72 48.14 13.98
Educ 3.94 1.90 3.97 2.01 3.93 1.88
Experience 23.88 15.18 26.50 13.29 23.21 15.61
Own Land 0.59 0.49 0.57 0.50 0.59 0.49
High Elevation 0.12 0.33 0.23 0.43 0.09 0.29
Medium Elevation 0.39 0.49 0.30 0.46 0.41 0.49
Sandy Soil 0.12 0.32 0.17 0.38 0.10 0.30
Off-farm Income 36.18 54.23 24.69 26.79 39.15 59.03
Household Size 3.70 2.13 3.36 1.83 3.79 2.20
Rice Area 1.28 0.89 1.46 1.04 1.24 0.85
Irrig Fuel Price 25.91 10.09 26.87 2.06 25.66 11.26
Training 0.56 0.50 0.87 0.34 0.48 0.50
Service Area 48.37 6.44 47.33 6.72 48.64 6.37
Distance 296.86 355.14 262.22 241.34 305.82 379.40
Hours Irrigation 60.63 77.45 40.53 22.30 65.83 85.46
Irrigation Freq 15.31 12.43 10.27 6.23 16.61 13.30
Labor 58.04 27.56 62.93 25.64 56.77 27.99
Yield 4898.39 1164.14 4670.28 1315.23 4957.88 1120.47
Profits 17800.25 13906.96 17732.29 12368.72 17817.83 14327.47
Table 3
Probit model results: determinants of AWD adoption (Dep. Var. = AWD dummy).
Variable Parameter estimate p-Value
Intercept 0.07 0.96
Age 0.01 0.69
Educ 0.03 0.61
Experience 0.003 0.98
Own Land 0.31 0.31
High Elevation 0.65 0.11
Medium Elevation 0.12 0.68
Sandy Soil 0.70 0.10
Off-farm Income 0.01 0.08
Household Size 0.10 0.17
Rice Area 0.08 0.60
Irrig Fuel Price 0.01 0.42
Training 1.37 <0.01
Service Area 0.02 0.36
Distance 0.0001 0.75
Log-likelihood 58.87
Pseudo-R
2
0.20
Akaike Information Criteria (AIC) 1.01
Bayesian Information Criteria (BIC) 535.101
8
Although not reported here (in the interest of space), we also used different
bandwidths based on rules of thumb reported in Pagan and Ullah (1999, p. 26). Using
these alternative bandwidths with the epanechnikov kernel does not qualitatively
change the results from the PSM analysis reported here. In addition, using a Gaussian
kernel instead of an epanechnikov kernel also does not qualitatively change the main
results from our analysis. Results from these runs provide additional evidence of the
robustness of our results. Detailed results of these runs are available from the authors
upon request.
9
Note that, in the interest of space, the comparison of means in Table 4 only
reports the tests using the kernel-matching procedure. The comparison of means
using the matched sample from the nearest-neighbor matching procedure also
provides similar results. These results are available from the authors upon request.
10
As one reviewer point out, it can be argued that AWD adoption in our sample may
be ‘‘almost random’’ because: (a) there are only two (out of the 14) observable
characteristics that are significantly different between the AWD and non-AWD
adopters, and (b) only a few variables were significant in the AWD probit adoption
model. This implies that a PSM approach may not be needed. However, we believe
that even if there are only two variables (High Elevation and Training ) that are
significantly different between AWD and non-AWD adopters, it is still possible that
these two could cause selection problems and can lead to misleading inferences.
Hence, we consider the PSM to still be an appropriate approach (i.e., not all observable
characteristics are insignificantly different.)
R.M. Rejesus et al. / Food Policy 36 (2011) 280–288 285
the frequency of irrigation by a substantial amount (i.e., maybe
lower duration of irrigation for each time water is applied).
For the labor, yield, and profit variables, however, the PSM pro-
cedure (and even the unmatched sample) shows that AWD do not
have a statistically significant impact on these variables. Hence, the
use of AWD do not significantly increase total labor use of the AWD
adopters, which indicates that using AWD do not necessarily exac-
erbate weed control problems that requires more labor for hand
weeding. Consistent with previous experimental studies by Singh
et al. (1996), Cabangon et al. (2001), and Moya et al. (2004), we find
that AWD adopters does not have statistically lower yields and
profits as compared to the non-AWD users.
11
Since AWD adoption
do not have a statistically significant effect on yields, it is likely that
AWD made farmers more technically efficient (i.e., they obtain sim-
ilar yields as non-AWD adopters even with less irrigation, ceteris
paribus). Taken all together, the PSM results in Table 5 imply that
AWD tend to reduce the hours of irrigation of AWD adopters without
increasing labor use (for controlling weeds) and without adversely
affecting yields and profits.
Robustness check: results from regression-based impact analysis
As a robustness check to the PSM results in Table 5, we imple-
ment the regression-based method for assessing the impact of
AWD as specified in Eq. (1). Results of these OLS regression models
are presented in Table 6. Consistent with the results in the PSM
analysis, the AWD adopters’ hours of irrigation tend be statistically
lower (at the 5% level) than the non-AWD adopters. The magnitude
of the estimated AWD effect on the farmers who adopted AWD in
the regression approach is similar to the PSM estimate (i.e., the
regression based estimate shows about a 20 h reduction, while
the PSM estimate shows an approximately 25 h reduction). On
the other hand, the ATT estimate for irrigation frequency in the
regression-based model is strongly significant as compared to the
marginal significance observed in the PSM. But note that the mag-
nitudes of the AWD effect on irrigation frequency are fairly similar
in both the regression-based and PSM approach (i.e., frequency
reduction of about 4 in the regression-based approach and about
3 for the PSM).
Table 5
Kernel and nearest neighbor propensity score matching results: the impact of AWD on irrigation, labor, yields, and profits.
Matching procedure/outcome variables Unmatched sample Matched sample
AWD mean Non-AWD mean Difference p-Value AWD mean Non-AWD mean Difference p-Value
A. Kernel matching
Hours Irrigation 40.53 65.82 25.29 0.11 40.64 65.43 24.79 0.04
Irrigation Freq 10.27 16.61 6.34 0.01 10.75 13.95 3.20 0.10
Labor 62.93 56.77 6.15 0.28 63.11 54.02 9.09 0.32
Yield 4670.28 4957.38 287.11 0.22 4703.02 5005.69 302.67 0.29
Profits 17732.28 17817.83 85.55 0.97 17917.59 18660.78 743.19 0.78
B. Nearest-neighbor matching
Hours Irrigation 40.53 65.82 25.29 0.11 40.64 65.92 25.28 0.07
Irrigation Freq 10.27 16.61 6.34 0.01 10.75 13.48 2.73 0.18
Labor 62.93 56.77 6.15 0.28 63.11 53.91 9.21 0.34
Yield 4670.28 4957.38 287.11 0.22 4703.02 4988.37 285.34 0.36
Profits 17732.28 17817.83 85.55 0.97 17917.59 18152.69 235.10 0.94
Table 4
Comparison of means of observable farm characteristics: unmatched and kernel matched samples.
Observable variables Unmatched sample Kernel matched sample
AWD
n=30
Non-AWD
n= 116
p-Value of difference AWD
a
n=27
Non-AWD
b
n= 116
p-Value of difference
Age 51.37 48.13 0.25 50.33 51.77 0.64
Educ 3.97 3.93 0.93 4.07 4.45 0.46
Experience 26.50 23.21 0.29 25.70 26.33 0.85
Own Land 0.57 0.59 0.78 0.59 0.65 0.63
High Elevation 0.23 0.09 0.04 0.14 0.11 0.66
Medium Elevation 0.30 0.41 0.26 0.33 0.45 0.30
Sandy Soil 0.17 0.10 0.34 0.11 0.12 0.87
Off-farm Income 24.69 39.16 0.19 27.06 25.14 0.78
Household Size 3.36 3.79 0.33 3.48 3.56 0.86
Rice Area 1.46 1.24 0.22 1.44 1.60 0.62
Irrig Fuel Price 26.87 25.66 0.55 26.68 33.78 0.23
Training 0.87 0.48 <0.01 0.85 0.84 0.96
Service Area 47.33 48.64 0.32 47.26 48.58 0.36
Distance 262.22 305.82 0.55 272.83 270.53 0.97
a
In the kernel matched sample, only 27 observations lie within the common support.
b
The means of the matched non-AWD sample is a weighted-average based on the weights produced in the kernel-matching procedure. That is, the weights given to each
non-AWD observation depends on how ‘‘close’’ its propensity score is to a particular AWD adopter.
11
As one reviewer indicated, the statistically insignificant profit difference between
AWD and non-AWD adopters may be due to the way profit was calculated in this
study. That is, the opportunity cost of water was not included in the profit calculation.
Note that the farmers do not actually pay for the water they use for irrigation (i.e., it is
a non-cash cost). As discussed in ‘AWD technology: overview and dissemination
approach’, farmers only actually pay for the fuel cost required to operate the ISC
pump at their scheduled irrigation time. These farmers also do not measure and
record the actual amounts of water they use for irrigation. Given that this information
is lacking, opportunity cost of water (defined as the amount of water used multiplied
by the price of water) cannot be accurately calculated and included in the profit
calculation. Given this data limitation, it is possible that when the opportunity cost of
water is accounted for, the profit impact of AWD could have been positive and
significant. Ultimately, this is an empirical question which could be an interesting
topic for future research (when more precise irrigation water data are available).
286 R.M. Rejesus et al. / Food Policy 36 (2011) 280–288
The result of no statistically significant difference in the labor,
yields, and profits for the AWD adopters versus non-adopters are
consistent in both the regression-based and PSM approach. The
magnitudes of the estimated (insignificant) impact of these vari-
ables are also fairly similar, which provides evidence of the robust-
ness of our results. Hence, the results in Table 6, also point to the
same robust conclusion as in the PSM model – that AWD reduces
the number of hours of irrigation without affecting labor, yields,
and profits (although our regression-based result also show a sig-
nificant reduction in irrigation frequency).
Assessment of the selection on unobservables: the Altonji, Elder and
Taber (2005) method
As described in the previous section (also see Footnote 4), we
assess the potential bias from selection on unobservables using
the procedure of Altonji et al. (2005) and implementing it on the
ATT estimate from the specification in (1). The estimated
s
using
the Altonji, Elder, and Taber (2005) procedure is reported in Table 7.
The
s
ratio of less than 1.0 for the irrigation variables, labor, and
yield indicate that it will only require a small shift in unobserva-
bles (when AWD is adopted) to invalidate the ATT estimates from
our procedures that only control for selection on observable vari-
ables. Only the ratio for the profit variable impact seems to indicate
a valid ATT estimate (where selection on unobservables will not be
a problem). Hence, the impact results for the irrigation variables
from the PSM and regression-based approaches have to be inter-
preted with caution. Selection bias due to unobservables may elim-
inate the significant ATT effect of the AWD technology on irrigation
(hours and frequency). But Table 7 shows that the impact result
where AWD do not adversely affect farmer profits may indeed be
valid. Given these results, we feel that further data collection is
needed to construct a panel data that would allow one to better
estimate the AWD impact by also accounting for selection prob-
lems that may be due to unobservable variables.
Conclusions and policy implications
Using cross-section farm-level survey data from the Philippines,
we empirically examine the impact of farmer adoption of the AWD
controlled irrigation technique. Propensity score matching (PSM)
approaches and a regression-based approach are employed to
determine the effect of the new irrigation technology on the adopt-
ers themselves (i.e., the average treatment effect on the treated
(ATT)), while controlling for potential selection bias due to obser-
vable variables. Our impact analysis suggests that AWD reduces
the number of hours of irrigation without a statistically significant
reduction in yields and profits. The estimated reduction in irriga-
tion hours is about 38% (relative to the mean irrigation hours in
the sample), which is consistent with the recommended water
reduction of 30–40% from the Safe AWD field experiments/pilot
study. The implication of this main result for our study area is that,
on average, rice producers who adopt AWD (particularly the Safe
AWD variant) can help conserve water (by about 38%) without ad-
versely affecting their yields and/or profitability.
However, even when controlling for selection problems based
on observables, further analysis of the impact estimates suggests
that the potential magnitude of the selection bias based on unob-
servable variables may still be able to invalidate the measured im-
pact from the PSM and regression-based techniques. Hence, the
current impact results have to be interpreted with caution and fur-
ther data collection is needed to construct a panel data that would
allow one to more definitively estimate the AWD impact by also
accounting for selection problems that may be due to unobservable
variables. Alternatively, having strong instruments would also en-
able one to control the selection on unobservables problem and
provide better impact estimates.
Notwithstanding the word of caution above, our results are still
suggestive of the fact that the AWD technique, specifically the Safe
AWD form, could be a promising water-saving technology that can
help address water scarcity in the area (and perhaps for most
major rice producing regions in Asia) because irrigation water
use can potentially be reduced without the adverse effect of lower-
ing rice producers’ yields or profits. The safe AWD irrigation tech-
nique would likely be most beneficial and applicable for rice
Table 6
Parameter estimates from the regression-based impact analysis approach.
Variable Dep. Var.: Hours Irrigation Dep. Var.: Irrigation Freq Dep. Var.: Labor Dep. Var.: Yield Dep. Var.: Profits
Coeff. p-Value Coeff. p-Value Coeff. p-Value Coeff. p-Value Coeff. p-Value
AWD 20.55 0.04 4.48 0.01 5.94 0.29 197.35 0.45 684.03 0.79
Age 0.74 0.48 0.02 0.91 0.29 0.32 12.51 0.43 164.81 0.36
Educ 4.33 0.27 0.40 0.50 2.15 0.16 17.61 0.81 609.55 0.47
Experience 0.19 0.82 0.09 0.47 0.16 0.57 15.24 0.25 80.93 0.57
Own Land 14.62 0.45 1.56 0.51 0.75 0.89 54.97 0.83 1242.64 0.69
High Elevation 12.64 0.49 2.11 0.45 6.69 0.42 208.42 0.64 6358.41 0.16
Medium Elevation 33.41 0.11 1.89 0.39 3.85 0.55 233.72 0.27 1487.32 0.62
Sandy Soil 21.69 0.53 1.07 0.74 7.96 0.36 439.44 0.30 1886.85 0.69
Off-farm Income 0.07 0.44 0.01 0.59 0.03 0.39 1.89 0.44 14.53 0.52
Household Size 0.58 0.92 0.26 0.63 0.21 0.87 27.79 0.63 169.75 0.79
Rice Area 1.18 0.91 9.05 <0.01 7.31 0.10 70.21 0.58 7011.82 <0.01
Irrig Fuel Price 1.85 0.12 0.02 0.77 0.13 0.33 0.93 0.92 239.14 0.02
Training 6.09 0.66 0.003 0.99 3.82 0.55 63.08 0.81 1579.32 0.62
Service Area 3.90 0.05 0.05 0.65 0.05 0.90 14.50 0.39 115.18 0.54
Distance 0.03 0.11 0.002 0.25 0.004 0.46 0.61 0.03 3.07 0.24
Intercept 90.99 0.45 29.84 <0.01 68.24 0.01 4506.01 <0.01 24277.33 0.05
R-squared 0.25 0.38 0.22 0.20 0.26
Notes: (1) The p-Values above are calculated using heteroskedasticity robust standard errors. (2) The parameter estimates of the interaction terms with the demeaned
explanatory variables (i.e. ðx
i
xÞAWD
i
in Eq. (1)) are not reported here in the interest of space, but are available from the authors upon request.
Table 7
Assessment of selection on unobservables: Altonji, Elder, and Taber (2005) procedure.
Outcome variable Ratio of estimated ATT parameter
to the estimated selection
on unobservables (
s
)
Hours Irrigation 0.21
Irrigation Freq 0.33
Labor 0.21
Yield 0.13
Profits 2.14
R.M. Rejesus et al. / Food Policy 36 (2011) 280–288 287
producing areas where pumping is used (as in the study site) be-
cause farmers do have private incentives to adopt the technology
(i.e., reduced marginal fuel costs). For rice farmers using gravity
flow irrigation systems, these private incentives may be less (i.e.,
irrigation water is free in this case) and public institutions may
be well-served to emphasize the ‘‘public good’’ aspect (or the ‘‘con-
servation benefits’’) of using the AWD technology – that is, water
resources will be more sustainably used and the onset of water
scarcity problems (for agriculture and urban users) could be de-
layed further into the future. The lack of a statistically significant
private profit effect also does not provide private incentives for
farmers to adopt AWD, which is why dissemination of information
about the water conservation externalities of AWD is important.
In light of our results above, insights from studies about the
flexible implementation of AWD (such as this one) need to be
effectively disseminated to rice producers through extension and
outreach programs. As shown in the probit model above, attending
a training program about irrigation has a strong positive impact on
the likelihood of AWD adoption. Extension programs that provide
general irrigation information and specific information regarding
AWD can potentially encourage further adoption of this irrigation
technique. As mentioned above the public benefit of adopting
water-conserving technologies such as AWD should be empha-
sized in these extension programs. Water conservation benefits
can potentially be realized if local government agencies (e.g., Phil-
Rice and NIA), as well as international agencies (e.g., IRRI), continue
to provide education and training about the latest research on
AWD and other water-conserving technologies to local extension
personnel, field technicians, ISCs, and farmers.
Appendix A
See Table A1.
References
Abdulai, A., Glauben, T., Herzfeld, T., Zhou, S., 2005. Water saving technology in
Chinese rice production – evidence from survey data. In: Selected Paper, 11th
Congress of the European Association of Agricultural Economists, Copenhagen,
Denmark, August 24–27, 2005.
Altonji, J.G., Elder, T.E., Taber, C.R., 2005. Selection on observed and unobserved
variables: assessing the effectiveness of Catholic schools. Journal of Political
Economy 113 (1), 151–184 (February).
Altonji, J.G., Elder, T.E., Taber, C.R., 2008. Using selection on observed variables to
assess bias from unobservables when evaluating Swan-Ganz catherization.
American Economic Review 98 (2), 345–350 (May).
Anderson, D.P., Wilson, P.N., Thompson, G.D., 1999. The adoption and diffusion of
level fields and basins. Journal of Agricultural and Resource Economics 24 (1),
186–203.
Becker, S.O., Ichino, A., 2002. Estimation of average treatment effects based on
propensity scores. The Stata Journal (2), 358–377 (fourth quarter).
Belder, P., Bouman, B.A.M., Spiertz, J.H.J., Cabangon, R., Guoan, L., Quilang, E.J.P., Li,
Y., Tuong, T.P., 2004. Effect of water and nitrogen management on water use
and yield of irrigated rice. Agricultural Water Management 65, 193–210.
Bhuiyan, S.I., 1992. Water management in relation to crop production: case study
on rice. Outlook on Agriculture 21, 293–299.
Blanke, A., Rozelle, S., Lohmar, B., Wang, J., Huang, J., 2005. Rural water saving
technology adoption in Northern China: an analysis of survey results. In:
Selected Paper, AAEA Annual Meeting, Providence, Rhode Island, July 24–27,
2005.
Bouman, B.A.M., Tuong, T.P., 2001. Field water management to save water and
increase its productivity in irrigated rice. Agricultural Water Management 49
(1), 11–30.
Bouman, B.A.M., Lampayan, R.M., Tuong, T.P., 2007. Water management in rice:
coping with water scarcity. International Rice Research Institute, Los Baños,
Philippines. 54 pp.
Cabangon, R.J., Castillo, E.G., Lu, G., Wang, G.H., Cui, Y.L., Tuong, T.P., Bouman, B.A.M.,
Li, Y.H., Chen, C.D., Wang, J.Z., 2001. Impact of alternative wetting and drying
irrigation on rice growth and resource-use efficiency. In: Barker, R., Loeve, R., Li,
Y.H., Tuong, T.P. (Eds.), Water Saving Irrigation for Rice: Proceedings of an
International Workshop. Wuhan, China. International Water Management
Institute, Colombo, Sri Lanka.
Cabangon, R.J., Tuong, T.P., Castillo, E.G., Bao, L.X., Lu, G., Wang, G.H., Cui, L., Bouman,
B.A.M., Li, Y., Chen, C., Wang, J., 2004. Effect of irrigation method and N-fertilizer
management on rice yield, water productivity and nutrient-use efficiencies in
typical lowland rice conditions in China. Paddy and Water Environment 2, 195–
206.
Casswell, M.F., 1991. Irrigation technology adoption decisions: empirical evidence.
In: Dinar, A., Zilberman, D. (Eds.), The Economics and Management of Water and
Drainage in Agriculture. Kluwer Academic Publishers, Boston, pp. 295–312.
Godtland, E.M., Sadoulet, E., De Janvry, A., Murgai, R., Ortiz, O., 2004. The impact of
farmer field schools on knowledge and productivity: a study of potato farmers
in the Peruvian Andes. Economic Development and Cultural Change 53
(October), 63–92.
International Rice Research Institute (IRRI), 1997. Rice almanac, second ed. IRRI, Los
Banos, Philippines, 181 pp.
Koundouri, P., Nauges, C., Tzouvelekas, V., 2006. Technology adoption under
production uncertainty: theory and application to irrigation technology.
American Journal of Agricultural Economics 88 (3), 657–670.
Lampayan, R., Bouman, B., de Dios, J., Lactaoen, A, Espiritu, A., Norte, T., Quilang, J.,
Tabbal, D., Llorca, L., Soriano, J., Corpuz, R., Malasa, R., Vicmudo, V., 2004.
Adoption of water saving technologies in rice production in the Philippines.
Food and Fertilizer Technology Center. Extension Bulletin 548, 15.
Moya, P., Hong, L., Dawe, D., Chongde, C., 2004. The impact of on-farm water saving
irrigation techniques on rice productivity and profitability in Zhanghe irrigation
system, Hubei, China. Paddy and Water Environment 2, 207–215.
Pagan, A., Ullah, A., 1999. Nonparametric Econometrics. Cambridge University Press,
New York, NY.
Palis, F.G., Cenas, P.A., Bouman, B.A.M., Hossain, M., Lampayan, R.M., Lactaoen, A.T.,
Norte, T.M., Vicmundo, V.R., Castillo, G.T., 2004. Farmer adoption of controlled
irrigation in rice: a case study in Canarem, Victoria, Tarlac. Philippine Journal of
Crop Science 29 (3), 3–12.
Peterson, J.M., Ding, Y., 2005. Economic adjustments to groundwater depletion in
the high plains: do water-saving irrigation systems save water. American
Journal of Agricultural Economics 87 (1), 147–159.
Pingali, P.L., Hossain, M., Gerpacio, R.V., 1997. Asian rice bowls: the returning crisis?
CAB International, Oxon, UK, 341 pp (in association with IRRI, Los Banos,
Philippines).
Rosenbaum, P., Rubin, D., 1983. The central role of the propensity score in
observational studies for causal effects. Biometrika 70 (April), 41–55.
Rosegrant, M.W., Ringler, C., 1998. Impact on food security and rural development
of transferring water out of agriculture. Water Policy 1 (6), 567–586.
Sianesi, B., 2001. Implementing Propensity Score Matching Estimators with Stata.
Paper Prepared for UK Stata Users group, VII Meeting, London, May 2001.
Singh, C.B., Aujla, T.S., Sandhu, B.S., Khera, K.L., 1996. Effects of transplanting data
and irrigation regime on growth, yield and water use in rice (Oryza sativa)in
northern India. Indian Journal of Agricultural Sciences 66, 137–141.
Tabbal, D.F., Bouman, B.A.M., Bhuiyan, S.I., Sibayan, E.B., Satar, M.A., 2002. On-farm
strategies for reducing water input in irrigated rice: case studies in the
Philippines. Agricultural Water Management 56, 93–112.
Tuong T.P., Bouman, B.A.M., 2003. Rice production in water-scarce environments.
In: Proceedings of the Water Productivity Workshop, International Water
Management Institute, Colombo, Sri Lanka, November 12–14, 2001.
Van der Hoek, W., Sakthivadivel, R., Renshaw, M., Silver, J.B., Birley, M.H., Koradsen,
F., 2000. Alternate wet dry irrigation in rice cultivation: saving water and
controlling malaria and Japanese encephalitis? IWMI Research Report 47,
Colombo Sri Lanka, International Water Management Institute (IWMI), 39 pp.
Wooldridge, J.M., 2002. Econometric Analysis of Cross-section and Panel Data. MIT
Press, Cambridge, MA.
Zhi, M., 2001. Water-efficient irrigation and environmentally sustainable irrigated
rice production in China. Unpublished Manuscript, Wuhan, China, Wuhan
University.
Table A1
Balancing test results: p-values of the mean comparisons of observable characteristics
(AWD versus Non-AWD adopters in two strata).
Observable variables Stratum 1 Stratum 2
Age 0.70 0.71
Educ 0.46 0.06
Experience 0.70 0.72
Own Land 0.58 0.16
High Elevation 0.60 0.52
Medium Elevation 0.94 0.20
Sandy Soil 0.88 0.91
Off-farm Income 0.67 0.39
Household Size 0.99 0.75
Rice Area 0.45 0.28
Irrig Fuel Price 0.07 0.22
Training 0.29 –
a
Service Area 0.39 0.95
Distance 0.59 0.15
No. of Obs. 93 11
a
All the AWD and non-AWD adopters in this stratum had water irrigation
training (Training = 1). Thus, no p-value is reported.
288 R.M. Rejesus et al. / Food Policy 36 (2011) 280–288