00074918.2012.728638
Bulletin of Indonesian Economic Studies
ISSN: 0007-4918 (Print) 1472-7234 (Online) Journal homepage: http://www.tandfonline.com/loi/cbie20
Too little too late: welfare impacts of rainfall
shocks in rural Indonesia
Emmanuel Skoufias , Roy S. Katayama & B. Essama-Nssah
To cite this article: Emmanuel Skoufias , Roy S. Katayama & B. Essama-Nssah (2012) Too little
too late: welfare impacts of rainfall shocks in rural Indonesia, Bulletin of Indonesian Economic
Studies, 48:3, 351-368, DOI: 10.1080/00074918.2012.728638
To link to this article: http://dx.doi.org/10.1080/00074918.2012.728638
Published online: 20 Nov 2012.
Submit your article to this journal
Article views: 218
View related articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=cbie20
Download by: [Universitas Maritim Raja Ali Haji]
Date: 18 January 2016, At: 00:28
Bulletin of Indonesian Economic Studies, Vol. 48, No. 3, 2012: 351–68
TOO LITTLE TOO LATE:
WELFARE IMPACTS OF RAINFALL
SHOCKS IN RURAL INDONESIA
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah*
World Bank, Washington DC
We use regression analysis to assess the potential welfare impacts in rural Indonesia of two types of shock: a delay in monsoon onset; and a signiicant shortfall
in rain during the 90-day post-onset period. Focusing on households with family
farm businesses, we ind that a delay in monsoon onset does not have a signiicant
effect on the welfare of rice farmers. However, rice farm households located in areas
exposed to low rainfall following the monsoon are negatively affected. Such households appear to be able to protect their food expenditure in the face of weather
shocks, but at the expense of their non-food expenditure. We also use propensity
score matching to identify community programs that might moderate the impact
of this type of shock. We ind that access to credit and public works projects has the
strongest moderating effect. This is an important consideration for the design and
implementation of adaptation strategies.
Keywords: welfare, rainfall shocks, social protection, agriculture, rice
INTRODUCTION
In Indonesia, annual rainfall patterns are critical to agricultural output and rural
livelihoods. In the cultivation of rice, the country’s most important crop, farmers
typically grow seedlings in a small plot and then transplant them to looded paddy
ields when rainfall is suficient. Thus, low cumulative rainfall at the beginning
of the wet season can delay transplanting and subsequently the harvest (Heytens
1991). Such climate-induced delays in crop harvests can mean an extended hungry season for poor farmers with limited savings or food stocks. They can also
undermine the prospects for a decent second harvest later in the year.
Empirical studies have shown that in Indonesia the amount of rainfall in
September–December, the period encompassing the early portion of the wet season, has a strong positive correlation with rice production through the area planted
and the area harvested in January–April. Between 1971 and 1998, the rainfall in
the September–December period explained over 80% of the variation in both the
* We wish to thank Will Martin, Syud Amer Ahmed, William Wallace and two anonymous
referees for useful comments on an earlier version of this paper. The indings, interpretations and conclusions are entirely those of the authors. They do not necessarily represent
the views of the World Bank, its executive directors or the countries they represent.
ISSN 0007-4918 print/ISSN 1472-7234 online/12/030351-18
http://dx.doi.org/10.1080/00074918.2012.728638
© 2012 Indonesia Project ANU
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
352
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
planted and the harvested rice area in January–April (Naylor et al. 2001, 2002).
These same studies made further linkages to sea surface temperature anomalies
and to El Niño/La Niña climate patterns, and proposed forecasting models to
inform food policy planning. Falcon et al. (2004) extended the climate–production
models down to the provincial level. He suggested that with improved forecasting models and timely dissemination, farmers could be notiied of recommended
cropping patterns to adapt to changing conditions, and agencies could be better positioned to mobilise relief efforts to assist poor and near-poor households
affected by the shocks.
Although there has been extensive research on rainfall–production linkages
at the aggregate level, very little is known about the welfare losses that households experience from rainfall shocks, whether induced by El Niño events or not.
Households with low levels of income are believed to be the most vulnerable to
the effects of negative shocks, including the rainfall shocks discussed above. This
is due to their geographical locations, limited assets, restricted access to resources
and services, low human capital and high dependence on natural resources for
income and consumption. While there is broad recognition of the threat to the
poor posed by climate-induced shocks, little attention has been given to quantifying the household-level effects of rainfall shocks. Our analysis considers the
household welfare implications of both a late monsoon onset and a low level of
rainfall. As we note later, a certain amount of rainfall is needed during the 90-day
post-onset period for rice to grow properly.
With projections pointing to a greater probability of rainfall shocks in the
future,1 policy makers will need to know what policies can either mitigate the
effects of the shocks or help households cope with them. As various social safety
nets and other social assistance programs are already in place, an assessment of
their role would be a good place to start. For instance, programs that give households greater access to credit may help them cope with delayed or poor harvests.
If community block grants are used to invest in more advanced irrigation infrastructure, they may help mitigate the effects of rainfall shocks. Or, if those grants
are used to support public works projects, they may generate nonfarm employment opportunities in the community. Evidence from within Indonesia conirming or refuting such claims could help policy makers identify instruments to help
protect vulnerable households. Using available data, we explore the moderating
effects of various programs.
The purpose of this paper is to analyse the potential effects on welfare of rainfall shocks in rural Indonesia, and to draw relevant policy lessons. The paper
is organised as follows. The next section presents the methodology, focusing on
the estimation of the effects of rainfall variability on household expenditure per
capita, our measure of welfare. The guiding view here is that the distribution of
welfare losses associated with such events depends on the degree of household
and community-level vulnerability and the moderating effect of existing assets
1 Adapting projections by the Intergovernmental Panel on Climate Change to local conditions, Naylor et al. (2007) predicted that the probability of a 30-day delay in monsoon
onset would increase from 9–18% at the time the study was undertaken to 30–40% by 2050.
Combined with increased temperatures, such a delay could reduce rice and soybean yields
by as much as 10%.
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
353
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
and social protection programs. Understanding these factors plays an important
role in designing policies to minimise exposure to rainfall shocks and reduce
their impact. The third section describes the available data and the fourth presents the analytical results. The paper ends with a discussion of the study’s policy
implications.
METHODOLOGY
This section describes the methodology and analytical framework used to estimate the effects of rainfall variability on household welfare in rural Indonesia and
the potential moderating effects of community-based programs. We need to make
our analytical framework consistent with the logic of vulnerability, the bedrock
concept for the study of the welfare impacts of weather shocks. The distribution
of economic welfare in any given society hinges crucially on individual endowments and behaviour and the socio-political arrangements that govern social
interaction. These factors (endowments, behaviour and social interaction) also
determine the distribution of vulnerability.2
Adger (1999) emphasises the connection between individual and collective
vulnerability, on the grounds that it is impossible to consider individual achievement in isolation from the natural and social environment. The vulnerability of an
individual or a household to livelihood stress depends crucially on both exposure
to and the ability to cope with and recover from the shock. Exposure is a function
of such things as climatic and topographical factors and the extent to which livelihoods are dependent on the weather. The ability to cope is largely determined by
access to resources, the diversity of income sources and social status within the
community.3 Increased exposure combined with a reduced capacity to cope with,
recover from or adapt to any exogenous stress on livelihood leads to increased
vulnerability.
In our analysis, we focus on exploiting cross-sectional variation in the data. We
link our welfare indicator, real per capita expenditure or some component thereof
(food versus nonfood expenditure), to a rainfall shock deined on the basis of
available rainfall data, focusing mainly on rural households. As noted in footnote
1, yields of crops such as rice and soybeans can be very much affected by changes
in precipitation patterns.
Because of the importance of rice farming in the rural economy of Indonesia,
we deine rainfall shocks with reference to this activity. In their study of the delay
in monsoon onset, Naylor et al. (2007) deine ‘onset’ as the number of days after
1 August for cumulative rainfall to reach 20 centimetres – the amount needed to
moisten the ground suficiently for planting – and ‘delay’ as the number of days
above the mean onset date over the 25year period from 1979 to 2004. Since farmers
2 Vulnerability is usually taken to mean the likelihood that, at a given point in time, individual welfare will fall short of some socially acceptable benchmark (Hoddinott and
Quisumbing 2008).
3 Hoddinott and Quisumbing (2008) make essentially the same point, noting that, at the
household level, vulnerability is determined by the nature of the shock, the availability of
additional sources of income, the functioning of labour, credit and insurance markets, and
the extent of public assistance.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
354
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
typically begin planting after monsoon onset, late onset may affect the prospects
for a second harvest later in the season and possibly change crop combinations,
with potentially signiicant consequences for production and market prices.
While delayed onset is an important determinant of harvest, we also need to
consider the amount of rainfall after monsoon onset. After farmers plant their rice
ields, 60–120 centimetres of rainfall is needed during the 3–4month growout
period (Naylor et al. 2002). Thus, the second dimension of our shock involves the
deviation of the amount of postonset rainfall from the 25year mean for each of
32 Indonesian weather stations. We deine the amount of postonset rainfall as the
total amount of rainfall during the 90-day period following the monsoon onset
date.
The timing of these events is considered in relation to the 2000 Indonesian
Family Life Survey (IFLS). The IFLS data on household consumption were collected from late June to the end of October 2000; we therefore focus on rainfall
patterns from mid to late 1999 that would affect planting and harvests during
the period immediately preceding the implementation of the IFLS. Considering
that the degree of rainfall variability can differ across areas and that households
may adjust their farming practices accordingly, we use standard deviations from
the inter-temporal mean to help account for such spatial differences. In the case
of delay in monsoon onset, we deine a negative shock as being more than one
standard deviation above the 25year mean. In the case of the amount of post
onset rainfall, we deine a negative shock as being more than two standard deviations below the 25year mean.
Given the connection between individual and collective vulnerability and
adaptive capacity, our empirical analysis uses regression analysis to link an indicator of household welfare – that is, real per capita total expenditure or its food
and non-food components – to some rainfall shock, while controlling for household characteristics and the province of residence. We estimate a regression equation of the form:
(
Yij = ß0 + ß1 Xi + ß2 S j + ß3 S j * Fi
)
(1)
where Yij represents per capita household expenditure of household i in community j; Xi represents various control variables; Sj represents the covariate rainfall
shocks; and Fi is a binary variable representing rice farm households. Standard
errors account for clustering at the community level and stratiication by province
and urban/rural sector, in line with the complex survey design of the IFLS.
After analysing the effects of rainfall shocks on welfare, we consider the potential moderating effect of various community-level social assistance programs.
Ideally we would like to measure, for the same household, per capita expenditure
with and without the program of interest at a particular point in time. As this is
not possible, we must seek alternatives.
If program placement were done randomly, simply comparing average per capita household expenditure between communities with and without the program
could be a good option for evaluating whether a certain program had helped
households exposed to shocks. But the placement of government programs is not
likely to be random (Pitt, Rosenzweig and Gibbons 1993). Many of the social safety
net programs that emerged following the 1997 inancial crisis were intended to
protect the poor and were thus targeted to poorer communities and households,
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
355
albeit with high leakage rates (Sumarto, Suryahadi and Widyanti 2002). Given
this potential for selection bias in program placement, the distribution of community and household characteristics is likely to differ between communities that
have and do not have a program. One consequence of the endogeneity in program
placement is that if this issue is not addressed in the analysis, it is likely to result in
biased estimates of program effects, especially when cross-sectional data are used.
Recognising that government assistance programs are often targeted to poor
areas, we use propensity score matching to investigate the role that various social
programs (for example, safety nets or credit) can play in moderating the effect of
a shock on household welfare, most likely by helping affected households cope
with the shock. The method comprises two main steps. The irst step is to construct a propensity score model that can be used to predict the likelihood of a
household or community receiving treatment, in this case, one of the social assistance programs. The predicted values are commonly referred to as the propensity scores. Assuming that, conditional on observable community characteristics,
program placement is as good as random, we can consider two households with
the same propensity score as being observationally equivalent. The second step is
to use the scores to match each household in the group that receives the program
to an equivalent household in the group without the program, such that the two
groups are comparable. Hereafter, we will refer to the group of households that
reside in communities with a speciic program as the ‘treatment group’ and the
constructed comparison group of households without the program as the ‘control
group’. With the treatment and control groups deined, the average difference in
the outcome variable can then be estimated.
We estimate propensity scores on covariates using probit regressions and
retrieve their predicted values to allow for the matching of ‘treated’ observations
with comparable ‘untreated’ observations in communities without the program.
For each program, a separate stepwise estimation of the probit speciication is
performed such that variables with a p-value of less than 0.2 are added to the
regression. The dependent variable is a binary variable indicating whether or not
a household resides in a community that is participating in the speciic program
of interest. The list of possible right-hand-side variables for the stepwise estimation includes household and community variables, and binary variables for the
different provinces. The household variables always included in the model are
household size; age of head; educational level of head; household use of electricity; ownership of farmland; ownership of a household nonfarm business; and
ownership of a household farm business. Candidate household variables are
marital status of head and gender of head. The candidate community variables
are availability of public transport; availability of piped water; predominance of
asphalt roads; share of households with electricity; distance to provincial capital; distance to district capital; shares of household heads with elementary, junior
high school, high school and universitylevel education; and share of households
with a ‘poverty certiicate’.4 All rural households included in the IFLS form the
sample for the probit regressions.
4 A poverty certiicate (Surat Keterangan Tidak Mampu) allows holders to get free treatment at hospitals and clinics. It is issued by local authorities to certify a household’s poverty status.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
356
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
After estimating the propensity scores, we trim the observations to obtain a
common range of scores for both ‘treated’ and ‘untreated’ observations. Then,
we match each ‘treated’ household to its three ‘nearest neighbours’ based on propensity scores from the group of ‘untreated’ households. The matching process
is done with replacement, such that all ‘untreated’ households are available as
candidates for each of the ‘treated’ households.To adjust for inexact matches of
the propensity scores, regression adjustments are performed as in Abadie et al.
(2004). We then compare the average outcome for households in the treatment
group (that is, in communities with a speciic program) to the average outcome
for similar households in the control group (that is, living in communities without
the program under consideration).
To describe this somewhat more formally, let Yi (1) denote the per capita
expenditure outcome of household i in the presence of some ‘treatment’ attribute
in the local community, such as a safety net program or a type of infrastructure,
and let Yi (0) denote the per capita expenditure outcome of household i in the
absence of the attribute in the local community. As it is not possible to observe
both Yi (1) and Yi (0) simultaneously for the same household, we must construct
a counterfactual group of households in communities that do not have the ‘treatment’ attribute of interest but do have a similar probability of having the attribute
based on observable community characteristics. Through a matching process, we
deine biascorrected matching estimators, Yˆi (0) , in place of Yi (0),5 and estimate
the sample average treatment effect for the treated (sub-population) (SATT):
1
SATT =
(2)
∑ Yi (1) − Yˆi (0)
n1 i|Wi =1
{
}
where Wi = 1 indicates that a household is in a community with the treatment
attribute, and n1 is the sample size of the treatment group.
DATA
We are able to study the effects of extreme weather events on rural households
by merging household and community-level data from the 2000 IFLS with daily
rainfall data covering the 25year period from 1979 to 2004. The combined data set
contains information on rainfall, household expenditure, household-level socioeconomic characteristics and community-level attributes. The IFLS household
and community surveys were ielded from late June to the end of October 2000.
The community surveys include data on whether various social programs were
available in the community in 1999–2000. It should be noted that the data do not
indicate which households actually participated in the programs.
The household-level data provide information on the value of aggregate consumption and its food and non-food components. The food component consists
of 37 food items and includes purchases, own production and gifts/transfers consumed within the last week. The non-food component consists of frequently purchased goods and services (utilities, personal toiletries, household items, domestic
services, recreation and entertainment, transport, sweepstakes); less frequent purchases and durables (clothing, furniture, health care, ceremonies, tax); housing;
5 See Abadie and Imbens (2002) and Abadie et al. (2004) for details.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
357
and educational expenditures for children living in the household. Transfers out
of the household are excluded. All values are calculated on a monthly basis and
are expressed in real terms. To obtain real values, we used both temporal and spatial delators, with prices in December 2000 in Jakarta as the base.6
Using daily rainfall data from 1979 to 2004 obtained from the Indonesian
Meteorological Agency,7 we calculated the 25year mean and standard deviations for monsoon onset and the amount of post-onset rainfall for 32 weather
stations. The rainfall data from these stations were then matched to communities
in the IFLS. Weather data were merged with household survey data at the community level based on proximity. Only weather stations with complete data for
the 25year period were used.
After the available precipitation data were merged and observations with missing data dropped, the sample size for the analysis was reduced to 6,188 households from a total of 10,292 in the 2000 IFLS. The matched data covered a total of
267 communities and 32 weather stations. In rural areas, 106 communities (a total
of 3,290 households) in nine provinces were matched to 27 stations. In rural Java,
66 communities (a total of 2,159 households) in four provinces were matched to
18 stations. The number of communities per station ranged from one to ten in
rural areas. Data from additional weather stations would beneit the analysis by
improving the level of disaggregation of weather data, but such data could not
be obtained.
Figure 1 shows the variation in monsoon onset and post-onset rainfall in 1999–
2000, by province. Only provinces in Java experienced a delay in monsoon onset
greater than one standard deviation from the 25year mean, and only the Javanese provinces experienced rainfall more than two standard deviations below the
25year mean during the 90day postonset period.
The summary statistics for household expenditure, household characteristics
and exposure to rainfall shocks in rural Java are shown in table 1. The majority
of household heads are married, male and do not have more than an elementary
education. The vast majority of households use electricity. Half of the households
own farmland, and 44% are engaged in non-farm businesses. Another 60% of
households are engaged in farm businesses, 38% with rice as the most valuable
crop and 22% with another crop as the most valuable. Turning to the rainfallshock variables, we see that 34% of households in the sample are exposed to
delayed monsoon onset and 45% to postonset low rainfall. The correlation coeficient between these two variables for our sample is not large, at 0.38.
EMPIRICAL RESULTS
We present our indings irst on the impact of rainfall shocks on per capita household consumption levels, and second on the role that various social programs may
have played in helping households cope with the negative effects of the shocks.
6 The spatial delator used is the ratio of the local poverty line (provincial, urban/rural
area) to the Jakarta poverty line, in December 2000 prices. Thus, the delator converts the
local December 2000 values into Jakarta December 2000 values.
7 Imputed values for missing data were provided by Centre Européen de Recherche et
d’Enseignement des Géosciences de l’Environnement.
358
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
FIGURE 1 Variation in Monsoon Onset and
Post-onset Rainfall by Province, 1999–2000 a
Monsoon onset
J
J
J
J
J
J
J
J
J
0
J
J
J
-1
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
C. Java
1
W. Java
Standard deviation from 25-year mean
J
2
-2
Bali
J
J
J
S. Kalimantan
E. Java
Yogyakarta
Jakarta
Lampung
S. Sumatra
W. Sumatra
N. Sumatra
-3
90-day post-onset rainfall
8
J
6
J
0
J
J
J
J
J
J
J
J
J
J
J
J
J
J
-2
J
-4
J
J
J
J
J
-6
J
J
J
J
J
J
J
J
J
J
J
E. Java
2
J
J
J
Yogyakarta
4
C. Java
Standard deviation from 25-year mean
S. Kalimantan
Bali
W. Java
Jakarta
Lampung
S. Sumatra
W. Sumatra
-8
N. Sumatra
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
3
a A negative shock is deined as more than one standard deviation above the 25year mean for mon-
soon onset, and more than two standard deviations below the 25year mean for postonset rainfall.
Source: Authors’ calculations based on the Indonesian Family Life Survey and rainfall data from the
Indonesian Meteorological Agency.
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
359
TABLE 1 Summary Statistics for Households in Rural Java, 2000
Variable
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Total expenditure (Rp per capita per month)
Food
Non-food
Household size (no. of members)
Age of head (years)
Married head
Female head
Highest level of education of head
Elementary
Junior high school
High school
University
Household uses electricity
Household owns farmland
Household owns non-farm business
Household owns farm business
Rice is most valuable crop
Other crop is most valuable
Exposure to shock in 1999–2000
Delay in monsoon onset (>1 standard deviation)
Delay in monsoon onset (>2 standard deviations)
Postonset low rainfall (
ISSN: 0007-4918 (Print) 1472-7234 (Online) Journal homepage: http://www.tandfonline.com/loi/cbie20
Too little too late: welfare impacts of rainfall
shocks in rural Indonesia
Emmanuel Skoufias , Roy S. Katayama & B. Essama-Nssah
To cite this article: Emmanuel Skoufias , Roy S. Katayama & B. Essama-Nssah (2012) Too little
too late: welfare impacts of rainfall shocks in rural Indonesia, Bulletin of Indonesian Economic
Studies, 48:3, 351-368, DOI: 10.1080/00074918.2012.728638
To link to this article: http://dx.doi.org/10.1080/00074918.2012.728638
Published online: 20 Nov 2012.
Submit your article to this journal
Article views: 218
View related articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=cbie20
Download by: [Universitas Maritim Raja Ali Haji]
Date: 18 January 2016, At: 00:28
Bulletin of Indonesian Economic Studies, Vol. 48, No. 3, 2012: 351–68
TOO LITTLE TOO LATE:
WELFARE IMPACTS OF RAINFALL
SHOCKS IN RURAL INDONESIA
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah*
World Bank, Washington DC
We use regression analysis to assess the potential welfare impacts in rural Indonesia of two types of shock: a delay in monsoon onset; and a signiicant shortfall
in rain during the 90-day post-onset period. Focusing on households with family
farm businesses, we ind that a delay in monsoon onset does not have a signiicant
effect on the welfare of rice farmers. However, rice farm households located in areas
exposed to low rainfall following the monsoon are negatively affected. Such households appear to be able to protect their food expenditure in the face of weather
shocks, but at the expense of their non-food expenditure. We also use propensity
score matching to identify community programs that might moderate the impact
of this type of shock. We ind that access to credit and public works projects has the
strongest moderating effect. This is an important consideration for the design and
implementation of adaptation strategies.
Keywords: welfare, rainfall shocks, social protection, agriculture, rice
INTRODUCTION
In Indonesia, annual rainfall patterns are critical to agricultural output and rural
livelihoods. In the cultivation of rice, the country’s most important crop, farmers
typically grow seedlings in a small plot and then transplant them to looded paddy
ields when rainfall is suficient. Thus, low cumulative rainfall at the beginning
of the wet season can delay transplanting and subsequently the harvest (Heytens
1991). Such climate-induced delays in crop harvests can mean an extended hungry season for poor farmers with limited savings or food stocks. They can also
undermine the prospects for a decent second harvest later in the year.
Empirical studies have shown that in Indonesia the amount of rainfall in
September–December, the period encompassing the early portion of the wet season, has a strong positive correlation with rice production through the area planted
and the area harvested in January–April. Between 1971 and 1998, the rainfall in
the September–December period explained over 80% of the variation in both the
* We wish to thank Will Martin, Syud Amer Ahmed, William Wallace and two anonymous
referees for useful comments on an earlier version of this paper. The indings, interpretations and conclusions are entirely those of the authors. They do not necessarily represent
the views of the World Bank, its executive directors or the countries they represent.
ISSN 0007-4918 print/ISSN 1472-7234 online/12/030351-18
http://dx.doi.org/10.1080/00074918.2012.728638
© 2012 Indonesia Project ANU
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
352
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
planted and the harvested rice area in January–April (Naylor et al. 2001, 2002).
These same studies made further linkages to sea surface temperature anomalies
and to El Niño/La Niña climate patterns, and proposed forecasting models to
inform food policy planning. Falcon et al. (2004) extended the climate–production
models down to the provincial level. He suggested that with improved forecasting models and timely dissemination, farmers could be notiied of recommended
cropping patterns to adapt to changing conditions, and agencies could be better positioned to mobilise relief efforts to assist poor and near-poor households
affected by the shocks.
Although there has been extensive research on rainfall–production linkages
at the aggregate level, very little is known about the welfare losses that households experience from rainfall shocks, whether induced by El Niño events or not.
Households with low levels of income are believed to be the most vulnerable to
the effects of negative shocks, including the rainfall shocks discussed above. This
is due to their geographical locations, limited assets, restricted access to resources
and services, low human capital and high dependence on natural resources for
income and consumption. While there is broad recognition of the threat to the
poor posed by climate-induced shocks, little attention has been given to quantifying the household-level effects of rainfall shocks. Our analysis considers the
household welfare implications of both a late monsoon onset and a low level of
rainfall. As we note later, a certain amount of rainfall is needed during the 90-day
post-onset period for rice to grow properly.
With projections pointing to a greater probability of rainfall shocks in the
future,1 policy makers will need to know what policies can either mitigate the
effects of the shocks or help households cope with them. As various social safety
nets and other social assistance programs are already in place, an assessment of
their role would be a good place to start. For instance, programs that give households greater access to credit may help them cope with delayed or poor harvests.
If community block grants are used to invest in more advanced irrigation infrastructure, they may help mitigate the effects of rainfall shocks. Or, if those grants
are used to support public works projects, they may generate nonfarm employment opportunities in the community. Evidence from within Indonesia conirming or refuting such claims could help policy makers identify instruments to help
protect vulnerable households. Using available data, we explore the moderating
effects of various programs.
The purpose of this paper is to analyse the potential effects on welfare of rainfall shocks in rural Indonesia, and to draw relevant policy lessons. The paper
is organised as follows. The next section presents the methodology, focusing on
the estimation of the effects of rainfall variability on household expenditure per
capita, our measure of welfare. The guiding view here is that the distribution of
welfare losses associated with such events depends on the degree of household
and community-level vulnerability and the moderating effect of existing assets
1 Adapting projections by the Intergovernmental Panel on Climate Change to local conditions, Naylor et al. (2007) predicted that the probability of a 30-day delay in monsoon
onset would increase from 9–18% at the time the study was undertaken to 30–40% by 2050.
Combined with increased temperatures, such a delay could reduce rice and soybean yields
by as much as 10%.
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
353
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
and social protection programs. Understanding these factors plays an important
role in designing policies to minimise exposure to rainfall shocks and reduce
their impact. The third section describes the available data and the fourth presents the analytical results. The paper ends with a discussion of the study’s policy
implications.
METHODOLOGY
This section describes the methodology and analytical framework used to estimate the effects of rainfall variability on household welfare in rural Indonesia and
the potential moderating effects of community-based programs. We need to make
our analytical framework consistent with the logic of vulnerability, the bedrock
concept for the study of the welfare impacts of weather shocks. The distribution
of economic welfare in any given society hinges crucially on individual endowments and behaviour and the socio-political arrangements that govern social
interaction. These factors (endowments, behaviour and social interaction) also
determine the distribution of vulnerability.2
Adger (1999) emphasises the connection between individual and collective
vulnerability, on the grounds that it is impossible to consider individual achievement in isolation from the natural and social environment. The vulnerability of an
individual or a household to livelihood stress depends crucially on both exposure
to and the ability to cope with and recover from the shock. Exposure is a function
of such things as climatic and topographical factors and the extent to which livelihoods are dependent on the weather. The ability to cope is largely determined by
access to resources, the diversity of income sources and social status within the
community.3 Increased exposure combined with a reduced capacity to cope with,
recover from or adapt to any exogenous stress on livelihood leads to increased
vulnerability.
In our analysis, we focus on exploiting cross-sectional variation in the data. We
link our welfare indicator, real per capita expenditure or some component thereof
(food versus nonfood expenditure), to a rainfall shock deined on the basis of
available rainfall data, focusing mainly on rural households. As noted in footnote
1, yields of crops such as rice and soybeans can be very much affected by changes
in precipitation patterns.
Because of the importance of rice farming in the rural economy of Indonesia,
we deine rainfall shocks with reference to this activity. In their study of the delay
in monsoon onset, Naylor et al. (2007) deine ‘onset’ as the number of days after
1 August for cumulative rainfall to reach 20 centimetres – the amount needed to
moisten the ground suficiently for planting – and ‘delay’ as the number of days
above the mean onset date over the 25year period from 1979 to 2004. Since farmers
2 Vulnerability is usually taken to mean the likelihood that, at a given point in time, individual welfare will fall short of some socially acceptable benchmark (Hoddinott and
Quisumbing 2008).
3 Hoddinott and Quisumbing (2008) make essentially the same point, noting that, at the
household level, vulnerability is determined by the nature of the shock, the availability of
additional sources of income, the functioning of labour, credit and insurance markets, and
the extent of public assistance.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
354
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
typically begin planting after monsoon onset, late onset may affect the prospects
for a second harvest later in the season and possibly change crop combinations,
with potentially signiicant consequences for production and market prices.
While delayed onset is an important determinant of harvest, we also need to
consider the amount of rainfall after monsoon onset. After farmers plant their rice
ields, 60–120 centimetres of rainfall is needed during the 3–4month growout
period (Naylor et al. 2002). Thus, the second dimension of our shock involves the
deviation of the amount of postonset rainfall from the 25year mean for each of
32 Indonesian weather stations. We deine the amount of postonset rainfall as the
total amount of rainfall during the 90-day period following the monsoon onset
date.
The timing of these events is considered in relation to the 2000 Indonesian
Family Life Survey (IFLS). The IFLS data on household consumption were collected from late June to the end of October 2000; we therefore focus on rainfall
patterns from mid to late 1999 that would affect planting and harvests during
the period immediately preceding the implementation of the IFLS. Considering
that the degree of rainfall variability can differ across areas and that households
may adjust their farming practices accordingly, we use standard deviations from
the inter-temporal mean to help account for such spatial differences. In the case
of delay in monsoon onset, we deine a negative shock as being more than one
standard deviation above the 25year mean. In the case of the amount of post
onset rainfall, we deine a negative shock as being more than two standard deviations below the 25year mean.
Given the connection between individual and collective vulnerability and
adaptive capacity, our empirical analysis uses regression analysis to link an indicator of household welfare – that is, real per capita total expenditure or its food
and non-food components – to some rainfall shock, while controlling for household characteristics and the province of residence. We estimate a regression equation of the form:
(
Yij = ß0 + ß1 Xi + ß2 S j + ß3 S j * Fi
)
(1)
where Yij represents per capita household expenditure of household i in community j; Xi represents various control variables; Sj represents the covariate rainfall
shocks; and Fi is a binary variable representing rice farm households. Standard
errors account for clustering at the community level and stratiication by province
and urban/rural sector, in line with the complex survey design of the IFLS.
After analysing the effects of rainfall shocks on welfare, we consider the potential moderating effect of various community-level social assistance programs.
Ideally we would like to measure, for the same household, per capita expenditure
with and without the program of interest at a particular point in time. As this is
not possible, we must seek alternatives.
If program placement were done randomly, simply comparing average per capita household expenditure between communities with and without the program
could be a good option for evaluating whether a certain program had helped
households exposed to shocks. But the placement of government programs is not
likely to be random (Pitt, Rosenzweig and Gibbons 1993). Many of the social safety
net programs that emerged following the 1997 inancial crisis were intended to
protect the poor and were thus targeted to poorer communities and households,
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
355
albeit with high leakage rates (Sumarto, Suryahadi and Widyanti 2002). Given
this potential for selection bias in program placement, the distribution of community and household characteristics is likely to differ between communities that
have and do not have a program. One consequence of the endogeneity in program
placement is that if this issue is not addressed in the analysis, it is likely to result in
biased estimates of program effects, especially when cross-sectional data are used.
Recognising that government assistance programs are often targeted to poor
areas, we use propensity score matching to investigate the role that various social
programs (for example, safety nets or credit) can play in moderating the effect of
a shock on household welfare, most likely by helping affected households cope
with the shock. The method comprises two main steps. The irst step is to construct a propensity score model that can be used to predict the likelihood of a
household or community receiving treatment, in this case, one of the social assistance programs. The predicted values are commonly referred to as the propensity scores. Assuming that, conditional on observable community characteristics,
program placement is as good as random, we can consider two households with
the same propensity score as being observationally equivalent. The second step is
to use the scores to match each household in the group that receives the program
to an equivalent household in the group without the program, such that the two
groups are comparable. Hereafter, we will refer to the group of households that
reside in communities with a speciic program as the ‘treatment group’ and the
constructed comparison group of households without the program as the ‘control
group’. With the treatment and control groups deined, the average difference in
the outcome variable can then be estimated.
We estimate propensity scores on covariates using probit regressions and
retrieve their predicted values to allow for the matching of ‘treated’ observations
with comparable ‘untreated’ observations in communities without the program.
For each program, a separate stepwise estimation of the probit speciication is
performed such that variables with a p-value of less than 0.2 are added to the
regression. The dependent variable is a binary variable indicating whether or not
a household resides in a community that is participating in the speciic program
of interest. The list of possible right-hand-side variables for the stepwise estimation includes household and community variables, and binary variables for the
different provinces. The household variables always included in the model are
household size; age of head; educational level of head; household use of electricity; ownership of farmland; ownership of a household nonfarm business; and
ownership of a household farm business. Candidate household variables are
marital status of head and gender of head. The candidate community variables
are availability of public transport; availability of piped water; predominance of
asphalt roads; share of households with electricity; distance to provincial capital; distance to district capital; shares of household heads with elementary, junior
high school, high school and universitylevel education; and share of households
with a ‘poverty certiicate’.4 All rural households included in the IFLS form the
sample for the probit regressions.
4 A poverty certiicate (Surat Keterangan Tidak Mampu) allows holders to get free treatment at hospitals and clinics. It is issued by local authorities to certify a household’s poverty status.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
356
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
After estimating the propensity scores, we trim the observations to obtain a
common range of scores for both ‘treated’ and ‘untreated’ observations. Then,
we match each ‘treated’ household to its three ‘nearest neighbours’ based on propensity scores from the group of ‘untreated’ households. The matching process
is done with replacement, such that all ‘untreated’ households are available as
candidates for each of the ‘treated’ households.To adjust for inexact matches of
the propensity scores, regression adjustments are performed as in Abadie et al.
(2004). We then compare the average outcome for households in the treatment
group (that is, in communities with a speciic program) to the average outcome
for similar households in the control group (that is, living in communities without
the program under consideration).
To describe this somewhat more formally, let Yi (1) denote the per capita
expenditure outcome of household i in the presence of some ‘treatment’ attribute
in the local community, such as a safety net program or a type of infrastructure,
and let Yi (0) denote the per capita expenditure outcome of household i in the
absence of the attribute in the local community. As it is not possible to observe
both Yi (1) and Yi (0) simultaneously for the same household, we must construct
a counterfactual group of households in communities that do not have the ‘treatment’ attribute of interest but do have a similar probability of having the attribute
based on observable community characteristics. Through a matching process, we
deine biascorrected matching estimators, Yˆi (0) , in place of Yi (0),5 and estimate
the sample average treatment effect for the treated (sub-population) (SATT):
1
SATT =
(2)
∑ Yi (1) − Yˆi (0)
n1 i|Wi =1
{
}
where Wi = 1 indicates that a household is in a community with the treatment
attribute, and n1 is the sample size of the treatment group.
DATA
We are able to study the effects of extreme weather events on rural households
by merging household and community-level data from the 2000 IFLS with daily
rainfall data covering the 25year period from 1979 to 2004. The combined data set
contains information on rainfall, household expenditure, household-level socioeconomic characteristics and community-level attributes. The IFLS household
and community surveys were ielded from late June to the end of October 2000.
The community surveys include data on whether various social programs were
available in the community in 1999–2000. It should be noted that the data do not
indicate which households actually participated in the programs.
The household-level data provide information on the value of aggregate consumption and its food and non-food components. The food component consists
of 37 food items and includes purchases, own production and gifts/transfers consumed within the last week. The non-food component consists of frequently purchased goods and services (utilities, personal toiletries, household items, domestic
services, recreation and entertainment, transport, sweepstakes); less frequent purchases and durables (clothing, furniture, health care, ceremonies, tax); housing;
5 See Abadie and Imbens (2002) and Abadie et al. (2004) for details.
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
357
and educational expenditures for children living in the household. Transfers out
of the household are excluded. All values are calculated on a monthly basis and
are expressed in real terms. To obtain real values, we used both temporal and spatial delators, with prices in December 2000 in Jakarta as the base.6
Using daily rainfall data from 1979 to 2004 obtained from the Indonesian
Meteorological Agency,7 we calculated the 25year mean and standard deviations for monsoon onset and the amount of post-onset rainfall for 32 weather
stations. The rainfall data from these stations were then matched to communities
in the IFLS. Weather data were merged with household survey data at the community level based on proximity. Only weather stations with complete data for
the 25year period were used.
After the available precipitation data were merged and observations with missing data dropped, the sample size for the analysis was reduced to 6,188 households from a total of 10,292 in the 2000 IFLS. The matched data covered a total of
267 communities and 32 weather stations. In rural areas, 106 communities (a total
of 3,290 households) in nine provinces were matched to 27 stations. In rural Java,
66 communities (a total of 2,159 households) in four provinces were matched to
18 stations. The number of communities per station ranged from one to ten in
rural areas. Data from additional weather stations would beneit the analysis by
improving the level of disaggregation of weather data, but such data could not
be obtained.
Figure 1 shows the variation in monsoon onset and post-onset rainfall in 1999–
2000, by province. Only provinces in Java experienced a delay in monsoon onset
greater than one standard deviation from the 25year mean, and only the Javanese provinces experienced rainfall more than two standard deviations below the
25year mean during the 90day postonset period.
The summary statistics for household expenditure, household characteristics
and exposure to rainfall shocks in rural Java are shown in table 1. The majority
of household heads are married, male and do not have more than an elementary
education. The vast majority of households use electricity. Half of the households
own farmland, and 44% are engaged in non-farm businesses. Another 60% of
households are engaged in farm businesses, 38% with rice as the most valuable
crop and 22% with another crop as the most valuable. Turning to the rainfallshock variables, we see that 34% of households in the sample are exposed to
delayed monsoon onset and 45% to postonset low rainfall. The correlation coeficient between these two variables for our sample is not large, at 0.38.
EMPIRICAL RESULTS
We present our indings irst on the impact of rainfall shocks on per capita household consumption levels, and second on the role that various social programs may
have played in helping households cope with the negative effects of the shocks.
6 The spatial delator used is the ratio of the local poverty line (provincial, urban/rural
area) to the Jakarta poverty line, in December 2000 prices. Thus, the delator converts the
local December 2000 values into Jakarta December 2000 values.
7 Imputed values for missing data were provided by Centre Européen de Recherche et
d’Enseignement des Géosciences de l’Environnement.
358
Emmanuel Skouias, Roy S. Katayama and B. EssamaNssah
FIGURE 1 Variation in Monsoon Onset and
Post-onset Rainfall by Province, 1999–2000 a
Monsoon onset
J
J
J
J
J
J
J
J
J
0
J
J
J
-1
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
J
C. Java
1
W. Java
Standard deviation from 25-year mean
J
2
-2
Bali
J
J
J
S. Kalimantan
E. Java
Yogyakarta
Jakarta
Lampung
S. Sumatra
W. Sumatra
N. Sumatra
-3
90-day post-onset rainfall
8
J
6
J
0
J
J
J
J
J
J
J
J
J
J
J
J
J
J
-2
J
-4
J
J
J
J
J
-6
J
J
J
J
J
J
J
J
J
J
J
E. Java
2
J
J
J
Yogyakarta
4
C. Java
Standard deviation from 25-year mean
S. Kalimantan
Bali
W. Java
Jakarta
Lampung
S. Sumatra
W. Sumatra
-8
N. Sumatra
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
3
a A negative shock is deined as more than one standard deviation above the 25year mean for mon-
soon onset, and more than two standard deviations below the 25year mean for postonset rainfall.
Source: Authors’ calculations based on the Indonesian Family Life Survey and rainfall data from the
Indonesian Meteorological Agency.
Too little too late: welfare impacts of rainfall shocks in rural Indonesia
359
TABLE 1 Summary Statistics for Households in Rural Java, 2000
Variable
Downloaded by [Universitas Maritim Raja Ali Haji] at 00:28 18 January 2016
Total expenditure (Rp per capita per month)
Food
Non-food
Household size (no. of members)
Age of head (years)
Married head
Female head
Highest level of education of head
Elementary
Junior high school
High school
University
Household uses electricity
Household owns farmland
Household owns non-farm business
Household owns farm business
Rice is most valuable crop
Other crop is most valuable
Exposure to shock in 1999–2000
Delay in monsoon onset (>1 standard deviation)
Delay in monsoon onset (>2 standard deviations)
Postonset low rainfall (