ArticlePDF Available

What prevents child diarrhoea? The impacts of water supply, toilets, and hand-washing in rural India

Authors:

Abstract and Figures

The authors apply three matching methods to estimate the impacts of water supply, toilet, and hand-washing interventions on child diarrhoea in rural India. Although propensity-score matching generally retains sample size, it can be associated with imbalance in the variables used to estimate propensity scores between treated and control groups. In contrast, exact matching is balanced over observables between treated and control units, but can result in considerable loss of observations. The authors also apply a novel ‘coarsened exact matching’ method that can potentially address the problem of sample attrition when matching. Their main finding using each of these three methods is that hand-washing, after defecating or before eating, significantly reduces prevalence and duration of a measure of overall diarrhea as well as acute watery diarrhoea among children under age five but not acute dysentery. In contrast, there may also be an effect of piped water on acute dysentery but not acute watery diarrhoea. Effects of improved water supply or improved toilets on different diarrhoeal outcomes are not observed consistently across matching methods.
Content may be subject to copyright.
Journal of Development Effectiveness
Vol. 3, No. 3, September 2011, 340–370
What prevents child diarrhoea? The impacts of water supply, toilets,
and hand-washing in rural India
Victoria Yue-May Fan
a
and Ajay Mahal
b
a
Harvard School of Public Health, Global Health and Population, 665 Huntington Avenue, Boston,
MA 02115, USA;
b
Monash University, School of Public Health and Preventive Medicine, Alfred
Centre, 99 Commercial Road, Melbourne, Australia
The authors apply three matching methods to estimate the impacts of water supply,
toilet, and hand-washing interventions on child diarrhoea in rural India. Although
propensity-score matching generally retains sample size, it can be associated with
imbalance in the variables used to estimate propensity scores between treated and con-
trol groups. In contrast, exact matching is balanced over observables between treated
and control units, but can result in considerable loss of observations. The authors also
apply a novel ‘coarsened exact matching’ method that can potentially address the prob-
lem of sample attrition when matching. Their main finding using each of these three
methods is that hand-washing, after defecating or before eating, significantly reduces
prevalence and duration of a measure of overall diarrhea as well as acute watery
diarrhoea among children under age five but not acute dysentery. In contrast, there
may also be an effect of piped water on acute dysentery but not acute watery diarrhoea.
Effects of improved water supply or improved toilets on different diarrhoeal outcomes
are not observed consistently across matching methods.
Keywords: diarrhoea; hand-washing; water; sanitation and hygiene; infrastructure;
environmental health; India
JEL classification:
H54; I12
1. Introduction
Diarrhoea is the second leading cause of child mortality in the world after pneumo-
nia, and diarrhoea alone kills more children than AIDS, malaria, and measles combined
(UNICEF/World Health Organisation [WHO] 2009). The most recent global estimates
for morbidity and mortality related to diarrhoea are for 2004, when nearly 2.5 billion
cases of diarrhoea occur red among children aged younger than five, of which 1.5 mil-
lion died (UNICEF/WHO 2009), despite the disease being both preventable and treatable.
Systematic reviews and meta-analyses of interventions focused on improved water qual-
ity and supply and improved sanitation facilities suggest that these can prevent diarrhoea
(Clasen et al. 2007, 2010, Waddington et al. 2009). Clasen et al. (2007) examined
42 studies, mostly randomised trials, in Latin America, Asia, and Africa. They concluded
that interventions to improve microbial quality of drinking water can reduce the risk of
diarrhoea among children under age five by about 30 to 40 per cent, and that water-quality
interventions in the household were more effective than interventions at the water source
*Corresponding author. Email: vfan@alum.mit.edu
ISSN 1943-9342 print/ISSN 1943-9407 online
© 2011 Taylor & Francis
http://dx.doi.org/10.1080/19439342.2011.596941
http://www.tandfonline.com
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 341
in preventing diarrhoea. Another review by Waddington et al. (2009) examined 65 impact
evaluations in 35 countries, concluding that interventions aimed at improving water quality
are significantly more effective in r educing the r isk of diarrhoeal disease compared with
interventions that improve water supply. A recent Cochrane systematic review by Clasen
et al. (2010) also examined the impacts of improved disposal of human excreta on occur-
rence of diarrhoea. Based on an analysis of 13 studies, they detected protective effects
ranging from 20 to 80 per cent compared with baseline risk.
These findings suggest that although improving water quality and sanitation are effec-
tive, improving the availability of water supply alone without considering the quality of
water accessed by the household is unlikely to be sufficient to prevent diarrhoea. Consider,
for example, piped water, which is categorised as an ‘improved water’ source (WHO
2009). The availability of a piped water facility may not prevent diarrhoea if the piped
water is susceptible to frequent shortages at the source, forcing households to seek out
alternate, potentially ‘unimproved’, drinking water sources such as unprotected springs
and surface water. (Similarly, if a toilet no longer functions, then a person may resort to
‘unimproved’ sanitation, which includes defecating in a bucket or in the open.) Moreover,
even if piped water supply is uninterrupted, the water quality ingested may be poor if
household containers for storing or for drinking water become contaminated over time;
this argument is strengthened by the reviews by Clasen et al. (2007) and Waddington
et al. (2009), who found that point-of-use treatments are more effective than those at the
source.
The transmission of diarrhoea is not entirely attributable to environmental transmis-
sion through water supply and sanitation facilities. An especially important intervention
for preventing diarrhoea that can be categorised as hygiene, rather than as water and san-
itation, is individual hand-washing behaviour. The transmission of acute watery diarrhoea
relies primarily on the ‘faecal–oral’ route via microbial agents such as viruses and bac-
teria that can survive on hands and hard surfaces (American Public Health Association
2008). This biological feature reinforces the fact that diarrhoea is not only transmitted
through contaminated drinking water. This feature is particularly relevant in cultures s uch
as those in India where large segments of the population eat directly with their hands. It is
possible for microbial concentration to be higher on a dry surface compared with con-
taminated water (Cash 2011). Indeed, to contract diarrhoea, a person would need to ingest
large amounts of water to be exposed to a pathogen, or ingest water with extremely high
microbial concentration (Cairncross et al. 2010).
A Cochrane systematic review by Ejemot et al. (2008) examined 18 randomised tri-
als of hand-washing on diarrhoea and found that interventions promoting hand-washing
resulted in a 32 per cent reduction in diarrhoeal episodes in children living in low-income
or middle-income countries, a conclusion supported by other reviews. Waddington et al.
(2009) assessed that interventions that promote hand-washing behaviour and improve san-
itation are both effective, although they were concerned about changing compliance rates
over time. A review by Zwane and Kremer (2007) also lends support to the protective
effects of point-of-use water treatments, some evidence of the effectiveness of piped water
and sanitation infrastructure, and little evidence of the impact of communal rural water
infrastructure. Finally, a systematic review by Cairncross et al. (2010) examined the effects
mainly on diarrhoeal morbidity of three interventions: hand-washing with soap, drinking-
water quality, and excreta disposal. They estimated diarrhoea risk reductions of 48 per cent
for hand-washing with soap, 17 per cent for water treatment, and 36 per cent for sanitation,
although they judge the available evidence to be of poor quality in general.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
342 V.Y. Fan and A. Mahal
Despite the large numbers of analyses included in these reviews, they suffer from
important weaknesses that affect both their validity and reliability. Most of the exper-
imental studies reported in these reviews of water quality, toilets, or hand-washing on
diarrhoea are characterised by small sample sizes, a focus on short-term outcomes, and
limited or no compliance information. As Cairncross et al. (2010, p. i194) write: ‘There
is a strong temptation to conduct evaluations of the health impact of water supply, sani-
tation, and hygiene interventions, but the challenges also are many’. Given the practical
difficulties involved in undertaking carefully designed randomised experiments relating
to water, sanitation, and hygiene interventions, an alternative is to examine the impact of
interventions using quasi-experimental approaches, such as matching methods, on cross-
sectional household survey data (Imai et al. 2008). Such observational data exist in many
countries with a sample size an order (or orders) of magnitude greater than any individual
study reported in the Cochrane reviews. Surprisingly, Waddington et al. (2009) found four
studies that used propensity-score matching to evaluate the impacts of water, sanitation,
and hygiene (Pradhan and Rawlings 2002, Jalan and Ravallion 2003, Khanna 2008, Bose
2009). Pradhan and Rawlings (2002), Jalan and Ravallion (2003) and Khanna (2008) evalu-
ated the impacts of water supply on child diarrhoea; the first two found a significant impact
of water supply on child diarrhoea, whereas Khanna (2008) did not. Pradhan and Rawlings
(2002) and Bose (2009) also evaluated the impacts of sanitation on child diarrhoea; the for-
mer did not find a significant effect, whereas the latter found that it significantly reduced
child diarrhoea.
In this paper, we go beyond these earlier quasi-experimental studies by using three dif-
ferent matching methods propensity-score matching (Rosenbaum and Rubin 1983), exact
matching and a novel ‘coarsened exact matching’ method (Iacus et al. 2011a) to esti-
mate the impacts of water supply, toilet, and hand-washing interventions, on the prevalence
of child diarrhoea in rural India. Choosing between exact matching and propensity-score
matching reflects a bias-efficiency trade-off ( Ho et al. 2007, Imai et al. 2008). Exact match-
ing methods do not require balance checking, but often lose many observations when
matching on many variables or dimensions. In contrast, propensity-score matching tends
to retain sample size but can lead to biased results without balance checking. Coarsened
exact matching seeks a middle ground between the two approaches (Iacus et al. 2011b).
Our reason for choosing India for investigation is that it is the country with the highest
number of both child cases and deaths from diarrhoea, with an estimated 237,482 deaths
from diarrhoea in 2008 (Black et al. 2010). Moreover, concerns about access to water and
sanitation facilities persist (see Figure 1). In 1990 it was estimated that just 7 per cent
of India’s rural population had access to an improved sanitation facility and 66 per cent
had access to an improved water source. Fifteen years later, these shares have increased
to 18 per cent for sanitation and 81 per cent for improved water, but with a popula-
tion of 1.2 billion people several hundred million people in rural India are still unable
to adequately access water and/or sanitation (UNICEF/WHO 2009).
India also has good data on this subject from multiple sur veys. For this paper, we rely
on a large observational rural household survey undertaken by the National Council for
Applied Economic Research (NCAER) in India during 1994. The survey data solicited
information on source of water supply, type of toilet, and hand-washing behaviours, whose
effects we are interested in, as well as a variety of village and household socio-economic
characteristics. In doing so, we are also able to contrast our approach to matching with
an earlier study that used the same dataset but that used only propensity-score matching
and studied only the effects of piped water (Jalan and Ravallion 2003). This study con-
cluded that piped water significantly reduced prevalence of under-five child diarrhoea by
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 343
0
10
20
30
40
50
60
70
80
90
1990 1992 1994 1996 1998 2000 2002 2004 2006
Improved drinking water source, rural
Improved sanitation facility, rural
Diarrhea prevalence within last two weeks
Figure 1. Trends in availability of improved water s upply, sanitation facility and diarrhoea
prevalence in India, 1990–2006.
Source: WHO (2011) and Macro International (2011).
21 per cent, which contrasts with findings from systematic reviews that tend to emphasise
the role of water quality, hand-washing and sanitation. Apart from assessing the impacts of
water, sanitation, and hygiene interventions in addition to piped water, our study builds on
the J alan and Ravallion (2003) framework in several ways. Two in particular are important.
In our study, we conduct balance diagnostics according to best practice for propensity-
score matching, at both individual and household levels by calculating a newly developed
measure of imbalance, L1 (see Iacus et al. 2011a). We pursue multiple matching methods
and thus assess and compare the robustness of results from propensity-score matching with
those from exact matching and coarsened exact matching.
Our main conclusion is that hand-washing, after defecating or before eating, signifi-
cantly reduces prevalence and duration of acute watery diarrhoea among children under
age five but does not affect acute dysentery. There may be an effect of piped water on acute
dysentery but not on acute watery diarrhoea, which is consistent with the study by Jalan and
Ravallion (2003). No significant effect of improved water supply (including piped water) or
improved toilets on different diarrhoeal outcomes is observed consistently across matching
methods.
2. Matching methods and observational data
Our analysis relies on the 1994 Human Development Index survey, undertaken by India’s
NCAER. This survey collected information on health status, hand-washing behaviours,
education, and other variables for 194,398 individuals including 25,117 children under age
ve living in rural areas. The survey collected data on socio-economic variables, source
of water supply, and type of toilet for 33,230 households including 16,245 households
with children under age five, and village-level variables such as the presence of village
development (for example, schools, health facilities, road quality) for 1762 villages in
16 major states of India. The survey was conducted during January through June 1994
(Shariff 1999). The survey collected a rich set of data at the village, household and individ-
ual levels, which were essential for this study. By contrast, the recently collected National
Family and Health Survey (2005/06) data did not include detailed village-level information
or information on hand-washing behaviours (although it collected information on handling
a child’s stools). Periodically collected National Sample Survey data on health status and
Downloaded by [Victoria Fan] at 04:38 26 September 2011
344 V.Y. Fan and A. Mahal
health-seeking behaviour also do not include this information. The NCAER data were col-
lected using a multistage sampling design, which involved the construction of region-level
strata by agricultural income and female literacy rates, followed by selection of districts
and villages, from which the households were sampled.
In observational data of the kind collected in the NCAER survey, assignment of treat-
ment is not randomised. Thus, the treated and untreated are likely to differ in characteristics
other than the treatment; that is, there may be imbalance in characteristics (both obser ved
and unobserved) between the treated and untreated groups. Matching is an attempt to con-
struct a sufficiently plausible counterfactual using information on observables and thereby
address the problem of selection bias or confoundedness. We use three matching methods
propensity-score matching (PSM), exact matching, and coarsened exact matching (CEM)
to estimate the effects of improved water supply, improved toilets, and hand-washing on
prevalence and duration of child diarrhoea in rural India. These methods are applied to
assess six ‘binary’ treatments: piped water; tube-well; improved water supply; improved
toilet; hand-washing done after defecation or handling stools, which we call ‘primary
hand-washing’; and hand-washing done before eating, feeding, or preparing food, which
we call ‘secondary hand-washing’. We used the WHO definition of ‘improved water sup-
ply’ and ‘improved t oilet’ (WHO 2009) and the categorisation of primary and secondary
hand-washing as described by Curtis et al. (2000).
2.1. Matching methods
PSM involves a cyclical two-stage process: estimating propensity scores for treated units
and control units (sometimes referred to as the ‘first stage’) with pre-treatment covariates
in a logit (or probit) model, followed by matching treated and control units to each other
with similar propensity scores (the ‘second stage’) using a matching algorithm such as
‘nearest neighbour’. In contrast, exact matching and CEM match directly on pretreatment
covariates and not on an intermediate score. In our study, we assume the placement of
treatments of interest to be a function of village-level and household-level (rather than
child-level) covariates that are plausible confounders of treatment, whereas the outcome
of interest, a diarrhoea episode and its duration, is at the individual child level. One set of
propensity scores were estimated at the household level and another set at the individual
child level.
1
We use only the 16,245 households with children under age five as they are
the population at risk of under-five child diarrhoea.
In the ‘second stage’, matching between the treatment and controls was undertaken at
both the household level and the individual level, using the nearest- n neighbour matching
algorithm without replacement. The nearest-n neighbour algorithm takes the average out-
come measure of the closest n number of matched controls as the counterf actual for each
treated unit. Conducting nearest-five neighbour matching, for example, depends on having
at least five controls to one treated case.
A major challenge in implementing PSM is that it requires that the treated and control
units matched on the (nearest) propensity score are balanced on pretreatment covariates.
Current best practice in PSM requires the researcher to generate propensity scores, match
observations on the scores, and then check balance, iteratively continuing until balance in
the matched data is maximised, by pruning the data and thereby excluding observations for
matching (Ho et al. 2007, Imai et al. 2008). In general, researchers can prune observations
using the propensity score in at least two ways: by applying ‘common support’ and by
using a calliper to match only neighbours within a certain ‘distance’. The term ‘common
support’ has been generally used to refer to the exclusion of control observations higher or
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 345
lower than the range of scores for treated observations. In the case of a calliper, matches
are only made within a certain distance to the propensity score. In PSM conducted in this
paper, both common support and a calliper are applied for nearest-neighbour PSM.
However, pruning through either common support or a calliper does not guarantee a
balanced matched sample under PSM, and thus the researcher must still check for balance
after matching. One consequence of this practical challenge is that researchers commonly
neglect to undertake or report balance diagnostics in applications of PSM (Ho et al. 2007,
Imai et al. 2008). (This iterative, cyclical nature of PSM is masked when referring to it
as a two-stage process.) Even when balance checks have been undertaken in the literature,
they have often taken the form of hypothesis testing to compare means of pretreatment
variables of the treated group and control group in the matched sample. However, Ho et al.
(2007) and Imai et al (2008) argue that one should ideally compare the joint distribution
of all covariates for the matched treatment and control units, and maximise balance as
much as possible. In the past it has been relatively difficult to accurately measure balance
under high dimensionality, that is, with many covariates to match on. Ho et al. (2007)
suggested using lower-dimensional balance diagnostics, including calculating the differ-
ences in mean in the matched sample and relative improvement in balance over the full
unmatched sample. This crude indicator of differences in means for a given variable or
distance measure (for example, the propensity score) between the matched and full sam-
ple can still mask imbalances between matched pairs. Thus a recent paper by Iacus et al
(2011a) proposed a new summary measure of multivariate imbalance, L1, and another
paper by King et al. (2011) recommends comparing the effectiveness of different matching
methods by choosing the matched sample with the smallest level of imbalance. We conduct
these best practices in balance diagnostics by conducting multiple iterations in PSM spec-
ifications of pretreatment variables, varying the size of the calliper, and applying common
support and in particular calculating the summary L1 measure (with further description of
these diagnostics in the next section).
Two alternatives that do not require iterative balance checking are exact matching and
CEM (Iacus et al. 2011b). Both of these methods match exactly on covariates and not
on an estimated propensity score. Unlike PSM, exact matching methods have the advan-
tage of not requiring any balance checking since units are matched exactly on all included
covariates and not on a one-dimensional score. However, exact matching is handicapped
by the smaller number of available units to match as the number of variables to match on
increases. CEM is a variation on exact matching that seeks to address this concern; the pro-
cedure matches observations exactly, but only after ‘coarsening’; that is, manually creating
categories for continuous variables or creating coarser categories from finer categories, that
are intuitive or natural (Iacus et al. 2011b). For example, one could consider the following
‘natural’ coarsening on education levels. Rather than using continuous years of education
of a household member, one could use a level of education such as primary school, middle
school, high school, and college or higher.
After applying each matching method, we take a sample weighted difference in out-
come means to calculate the ‘average treatment effect on the treated’ in each case. Ho et al.
(2007) argue that matching is a means to non-parametrically pre-process a dataset before
parametric analysis, thereby making estimates less dependent on (parametric) modelling
choices and specifications. Impact estimates should not vary much even when changing
parametric modelling assumptions, and in many cases matching also reduces variance
of the estimated causal effects. Thus, as a specification check after conducting PSM, we
conducted parametric analysis by including potential confounding variables to account for
Downloaded by [Victoria Fan] at 04:38 26 September 2011
346 V.Y. Fan and A. Mahal
any potentially remaining imbalances and compare those results with that from taking a
simple difference of means.
2
The key assumptions underlying any matching method are that treatment is uniform
in implementation for all units and that there are no spillovers. It is not obvious that the
treatment variables, especially those of water supply and sanitation facility, are uniform
across households or villages. For instance, the quality and reliability of supply of water
may vary across r egions. As long as the variation i n the inter ventions is not large, or if
the variation is due to a characteristic that is irrelevant to the health outcome (for instance,
depth of clean groundwater), this may not be a handicap. The risk of spillovers is expected
to be small given the relatively small number of households sampled in each village (on
average about 7 per cent of all households in each sampled village) and the small number
of villages (about 1760 villages) randomly sampled from a total of nearly 565,000 Indian
villages.
2.2. Measurement of treatment variables
Our water supply treatment variable is constructed from two questions posed to survey
respondents: one on the main source of drinking water in the non-summer season, another
on the main source of drinking water in the summer season. Respondents selected one of
the following: ponds; dug well; running stream/canals; protected wells; tanker truck; piped
outside the house; piped in the house; hand pump; others. We define piped water as any
piped water in any season.
3
Using the WHO definition of improved water (WHO 2009), we
construct ‘improved water supply’ as a binary variable that takes the value one for house-
holds using protected wells, tanker truck, piped water (inside or outside), and hand pump;
and ‘not improved’ (or zero) if the main source was a pond, dug well, running stream, or
others.
4
Our measure of sanitation relied on the type of toilet in the household (manual dis-
posal, septic tank, drainage, water-sealed pits, or others). Using the definition of improved
sanitation in WHO (2009), we define an ‘improved toilet’ as one with a septic tank,
drainage, or water-sealed pits. Finally, measures of individual hand-washing behaviours are
based on seven distinct questions on the sequence of hand-washing in relation to key events
affecting disease transmission: after using the latrine or defecating; after cleaning a child’s
stools; after disposing child’s stools; before cooking; before feeding food; before serving
food; and before eating food. We use measures of hand-washing as a primary barrier or
as a secondary barrier to diarrhoea depending on the sequencing of hand-washing (Curtis
et al. 2000). ‘Primary hand-washing’ is defined as hand-washing done after contact with
contaminated material from defecation or stools, and ‘secondary hand-washing’ is hand-
washing done before activities related to preparing, serving, or consuming food. Although
the hand-washing treatments are reported at the individual level, it is reasonable to assume
that hand-washing behaviours of children are a function of household-level pre-treatment
variables if we assume that hand-washing by (or of) children is a behaviour lear ned within
the household (see, for example, Luby et al. 2004, which promoted hand-washing with
soap for children as young as 30 months old). An alternative interpretation of this treat-
ment is that it may also reflect the hand-washing behaviours of the survey respondent (such
as the child’s mother).
Table 1 reports the distribution of households by treatment status. Nearly one-quarter
(23.4 per cent) of the sampled households with children under age five (that is, the
households of interest) had access to piped water in any season. Whereas 72.9 per cent
of households of interest had access to improved water, only 11.0 per cent of households
had access to an improved toilet. Self-reported primary hand-washing and secondary
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 347
Table 1. Distribution of households or individuals by treatment status.
Household subset
Treatment All households
With children
under age five
Without children
under age five
Number of households 33,216 16,245 16,971
Number of individuals 194,398 25,117 169,281
Household water supply
Piped water 25.8 23.4 27.8
Improved water supply 73.4 72.9 73.9
Tube-well (asset) 9.1 10.0 8.4
Household sanitation facility
Improved toilet 12.3 11.0 13.3
Individual hand-washing
Primary hand-washing 69.3 21.5 75.9
Secondary hand-washing 69.8 27.5 75.6
Household income quintile
Lowest 20.0 24.8 16.0
Second 20.0 23.4 17.1
Third 20.0 19.9 20.1
Four th 20.0 17.8 21.8
Highest 20.0 14.1 24.9
Total 100.0 100.0 100.0
Highest education of female in
household
Illiterate 47.0 48.0 46.1
Primary 29.4 30.3 28.6
Matriculation at most 19.8 18.2 21.2
Higher secondary or more 3.8 3.5 4.0
Total 100.0 100.0 100.0
Note: Of the 33,230 households in the household dataset, 14 households could not be matched to a village, hence
reducing the total number of all households to 33,216. Household weights were used here and throughout the
paper. Piped water and improved water supply both refer to that in any season, not all seasons. The hand-washing
variables (primary or secondary) here are individual, rather than household, prevalence.
hand-washing were lower among those under age five (21.5 per cent and 27.5 per cent,
respectively) compared with primary and secondary hand-washing among those aged ve
and higher (75.9 per cent and 75.6 per cent, respectively).
5
2.3. Measurement of diarrhoea
The survey included several questions to elicit short-duration diarrhoeal morbidity experi-
enced by household members in the past month and its duration. Diarrhoea was categorised
in six forms: acute watery diarrhoea, acute dysentery, persistent diarrhoea, food poisoning,
parenteral diarrhoea, and chronic diarrhoea. We construct three diarrhoea variables: ‘all
diarrhoea’ (which includes any one of the six forms r eported), acute watery diarrhoea, and
acute dysentery. Acute watery diarrhoea was defined as three or more loose motions with-
out blood in stools, lasting for 14 days or less. Acute dysentery was defined as three or
more loose motions with blood mixed with stools. We use only the first reported episode
of the illness in the reference period. That is, a child was defined as having the illness if
the child had a first episode of diarrhoea regardless of whether there was a second episode.
Given the reference period of one month (for reporting illnesses), the number of second
(or more) episodes was small and would not affect our conclusions.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
348 V.Y. Fan and A. Mahal
Table 2. Crude prevalence or duration of diarrhoea among children by treatment status.
Prevalence by type of diarrhoea
Treatment type
Treatment
status
All
diarrhoea
Acute watery
diarrhoea
Acute
dysentery
Duration of
diarrhoea
(days)
Overall sample 0.1086 0.0821 0.0095 0.5873
Piped water Untreated 0.1124 0.0855 0.0110 0.6102
Treated 0.0965 0.0714 0.0049 0.5136
Improved water Untreated 0.1127 0.0844 0.0105 0.6044
supply Treated 0.1071 0.0813 0.0092 0.5809
Tube-well Untreated 0.1087 0.0823 0.0090 0.5845
Treated 0.1083 0.0812 0.0137 0.6101
Improved toilet Untreated 0.1117 0.0848 0.0098 0.5957
Treated 0.0829 0.0605 0.0070 0.5181
Primary Untreated 0.1189 0.0928 0.0096 0.6418
hand-washing Treated 0.0710 0.0431 0.0091 0.3876
Secondary Untreated 0.1233 0.0972 0.0093 0.6699
hand-washing Treated 0.0700 0.0424 0.0102 0.3699
Note: Piped water and improved water supply refer to those treatments in any season.
Table 2 reports the crude prevalence of diarrhoeal disease, conditional on each of the
water supply, toilet, or hand-washing treatments. Of these three categories of diarrhoea,
acute watery diarrhoea was the most prevalent (8.2 per cent). The prevalence of acute
dysentery was 1 per cent. Prevalence of all diarrhoea was 10.9 per cent. Among children
under five in households with piped water, prevalence of all diarrhoea was 9.7 per cent com-
pared with 11.2 per cent among those without piped water. By contrast, the difference in the
crude prevalence and duration between those with improved toilet and those without was
more accentuated. Among children with an improved toilet, prevalence of all diarrhoea was
8.3 per cent compared with 11.2 per cent among those without. Children with an improved
toilet also had a shorter duration than those without (0.519 days compared with 0.642 days).
Even larger crude differences in outcomes are observed for hand-washing. Among those
who practiced secondary hand-washing, prevalence of all diarrhoea was 7.0 per cent com-
pared with 12.3 per cent among those who did not. Duration of diarrhoea among those with
secondary hand-washing was also much shorter compared with those without (0.370 days
compared with 0.670 days).
These estimates are almost i dentical to those reported in the official NCAER Human
Development Report based on the same dataset that we use (Shariff 1999) and are
also comparable with data from a separate report based on the National Family Health
Survey, 1992/93 (Macro International 2011). Shariff (1999) reports a prevalence of under-
ve child diar rhoea of 10.8 per cent, whereas our estimate of overall prevalence is
10.86 per cent. Similarly, Macro International (2011) estimated a diarrhoea prevalence of
11.5 per cent among children under age five in the past two weeks over 1992/93. Jalan and
Ravallion (2003) in Table 3 reported the prevalence of diarrhoea to be 1.08 per cent among
children under five living in households with piped water. It is possible that Jalan and
Ravallion (2003) made a simple error by shifting the decimal place to the left throughout
their results, or alternatively are referring to acute dysentery which indeed has a prevalence
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 349
Table 3. Summary of balance for treatment of piped water.
Method Model Model 1 Model 2 Model 3 Model 4
Exact matching Untreated 16, 816 10, 566 1170 665
Treated 6871 5420 938 528
CEM Untreated 17, 040 11, 146 2042 1173
Treated 6871 5462 1477 815
L1 0.964 0.951 0.715 0.428
%BI 99.6 99.6 99.4 99.7
PSM with narrowing callipers
PSM, calliper 0.25 Untreated 6302 6332 6337 6281
Treated 6302 6332 6337 6281
L1 0.968 0.969 0.963 0.966
%BI 88.6 87.1 86.4 86.8
PSM, calliper 0.1 Untreated 5977 5953 5995 5965
Treated 5977 5953 5995 5965
L1 0.966 0.966 0.965 0.967
%BI 97.2 95.9 95.6 95.5
PSM, calliper 0.001 Untreated 5587 5006 4934 4955
Treated 5587 5006 4934 4955
L1 0.967 0.957 0.958 0.96
%BI 100.0 100.0 100.0 100.0
PSM, calliper 0.0001 Untreated 5560 4257 2222 2156
Treated 5560 4257 2222 2156
L1 0.964 0.95 0.856 0.848
%BI 100.0 100.0 100.0 100.0
PSM, calliper 0.00001 Untreated 5560 4115 919 650
Treated 5560 4115 919 650
L1 0.963 0.947 0.589 0.368
%BI 100.0 100.0 100.0 100.0
Note: L1 here refers to multivariate L1, with zero indicating perfect balance and one indicating perfect imbalance.
In contrast, %BI refers to ‘percentage balance improvement’ in distance measure and approaches 100% to indicate
better balance. Model 1 includes state, household asset index (quintile), village infrastructure index (quintile).
Model 2 includes household income (quintile) and highest education of female in addition to Model 1 variables.
Model 3 includes scheduled caste, scheduled tribe, Muslim household, household size; household head is self-
employed, household head is married in addition to Model 2 variables. Model 4 includes electricity, improved
stove, and chimney in addition to Model 3 variables.
of 1 per cent in the under-five population, close to their estimate. (See the next section on
balance diagnostics for description of our replication of the results by Jalan and Ravallion
2003.)
2.4. Covariates for matching
There were several potential pretreatment variables for the first-stage PSM model available
in the dataset.
6
We briefly describe two important kinds of sets of variables used for match-
ing household wealth and village-level infrastructure as they are likely to be impor tant
predictors of the treatment variables. An index of household assets was constructed using
principal component analysis for the full sample of households (Filmer and Pritchett 2001).
Separately, we constructed an index of village infrastructure by using information on the
different facilities available in each village. In previous work, village wealth indices have
mostly been constructed by averaging the wealth of households living in a village (Pritchett
2010), presumably because of the lack of detailed data on village infrastr ucture (which are
available in the NCAER dataset). Village-level facility data include whether the village
has an Aaganwadi (daycare) centre, a primary school and other schools, library, bank,
Downloaded by [Victoria Fan] at 04:38 26 September 2011
350 V.Y. Fan and A. Mahal
market, and so forth. The construction of the household asset index and the village infras-
tructure index as summary measures of several variables also made exact and coarsened
exact matching feasible. The highest education of the female in the household variable was
coarsened to four categories (Table 1). Variables that were plausibly pre-treatment in the
causal pathway were included, whereas variables that were potentially post-treatment were
excluded in order to avoid post-treatment bias.
7
(See Tables 3 and 4 in the next section for
list of covariates used in matching specifications.)
3. Balance diagnostics
In this section we consider the trade-off between balance and sample size in pretreat-
ment variables in the matched samples from the three matching methods. In the context
of the PSM model, this means following an iterative process of propensity-score estima-
tion, matching and balance checking, as suggested by Gary King and his colleagues (Ho
et al. 2007, Iacus et al. 2011a, 2011b, King et al. 2011). Because balance on character-
istics can potentially worsen in the sample matched on the propensity score, the process
should ideally be repeated until the matching procedure suggests maximum balance, at the
cost of pruning observations. By contrast, exact matching and CEM are always balanced,
but sample attrition can be large with many matching variables, particularly continuous
variables.
The main measure of imbalance presented in this paper is the multivariate imbalance
L1 measure, which offers a relatively intuitive interpretation ‘[For any given set of bins],
if the two empirical distributions are completely separated, then L1 = 1; if the distribu-
tions exactly coincide, then L1 = 0’ (see Iacus et al. 2011a, p. 352). This measure can
be considered a common support restriction along a multi-dimensional histogram of mul-
tiple variables, not just a single variable or dimension such as the aggregate propensity
score. We also constructed a measure, ‘percent balance improvement’, which is defined
as 100(|A| |B|)/|A|, where (A) is the distance between the means of the treated and
untreated groups in the unmatched sample and (B) is the distance between the means in
the matched sample (Ho et al. 2007). In addition we considered percentage improvement
in the propensity score for PSM and the L1 measure for CEM.
The three matching methods for each treatment were executed using MatchIt (Ho et al.
2011) and CEM (Iacus et al. 2011a) in R 2.10. For PSM we applied common support and
started with a generous calliper of 0.25 standard deviations in propensity score. Nearest-
one neighbour PSM was conducted for every treatment; nearest-five neighbour PSM was
not possible for most treatments given a ratio of control to treated units of less than ve
(see Table 1).
In Table 3, we use piped water as an example to present a summary of balance in
matched sample size across the three matching methods (exact matching, CEM, and PSM),
which are shown in the first three sets of rows in the table. PSM retained a larger sample size
than CEM, which in turn retained a larger sample size than exact matching. As one moves
down a given column across rows in Table 3, we apply an increasingly narrow calliper for
PSM as a sensitivity analysis. When using PSM, one should prune the data manually; for
example, through ‘common support’ or applying a calliper or other procedure. In contrast,
exact matching methods automatically prune the data by virtue of matching exactly and
discarding those not matched. The pruning from general common support in PSM is gen-
erally limited. Table 3 shows that, by narrowing the calliper used for PSM, fewer units are
matched while balance improves. (The robustness of impact estimates to calliper tightening
and data pruning are shown in Section 6.)
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 351
In addition, Table 3 also presents summary measures of balance when increasing the
number of matching variables (moving right in a given row from Models 1 to 4). Exact
matching and CEM led to reductions in matched sample size when adding more vari-
ables to match on. In contrast, PSM retained sample size but did not balance as well on
the matching variables. The final covariate specification presented in this paper was the
result of successively adding four sets of variables that were plausibly pretreatment in the
causal pathway. The four sets of variables used were: village development index ( quin-
tile), household asset index (quintile), and state; household income per capita and highest
education of a female member in the household; demographic characteristics of scheduled
caste, scheduled tribe, Muslim household, household size, whether the household head is
married or not, and whether the household head is self-employed or not; and household
‘infrastructure’ that is, electricity, clean stove, and available chimney especially as elec-
tricity is highly correlated with piped water, as shown in Appendix 2). Although impact
estimates presented in the next section use Model 4, estimates from Models 1–3 were also
calculated. As a sensitivity check, we also included in some specifications (results not pre-
sented) piped water, tube-well and improved toilet as matching variables for hand-washing
given that water supply may affect whether one chooses to wash one’s hands.
For the six treatments, we present balance diagnostics in Table 4 for the fourth and final
covariate specifications using PSM and CEM. The table suggests that samples matched on
hand-washing interventions had better balance than samples matched on other water and
sanitation interventions. Moreover, for every intervention, CEM produced better balance
than PSM.
In addition to our main matching specifications and analyses, we replicated the Jalan
and Ravallion (2003) results using their pretreatment variables as well as their PSM
specification of matching on the odds ratio of the propensity score within a calliper
of 0.001. The first-stage model coefficients in our replication were highly comparable
with their study, which used all 33,000 households including those without children (see
Appendices 2 and 3). For our main analyses and for the second-stage of this replication
(when matching units by the generated propensity score), we restricted the sample to only
households with children who thus were at risk of the disease. After matching, imbal-
ance remained on several variables between treated and control groups (see Appendix 4).
As shown in Appendix 5, we estimated an effect of –0.019 of piped water on all diarrhoea,
–0.019 on acute watery diarrhoea, and –0.006 on acute dysentery, all highly significant at
p < 0.05. In comparison, Jalan and Ravallion (2003) found a piped water effect of –0.023
(if we assume a uniform shift in the decimal place, and thus their diarrhoea prevalence
estimate of 1.08 per cent would correspond to the correct estimate of 10.8 per cent). Thus
although we are able to replicate their results quite closely in t erms of the outcome preva-
lence and the significance level, as we show next, the effects on every form of diarrhoea
are not robust.
4. Main results and impact estimates
Impact estimates at the individual child level for water supply treatments on diarrhoeal
outcomes are reported in Table 5 and estimates for toilet and hand-washing treatments in
Table 6. Given limitations of space, the results presented use the fourth and final model;
results with the three matching specifications were consistent (see Tables 3 and 4 for
variables in the four covariate specifications). Impact estimates for exact matching and
CEM are the simple difference in the outcomes between treated and controls. After PSM,
no apparent differences in estimates were observed between simple differences (shown
in table) or after parametric adjustment (data not shown), suggesting that estimates are
‘doubly robust’ from matching or from parametric adjustment (Ho et al. 2007).
Downloaded by [Victoria Fan] at 04:38 26 September 2011
352 V.Y. Fan and A. Mahal
Table 4. Summary of balance on matching variables.
PSM CEM
Measure
Piped
water
Improved
water Tube-well
Improved
toilet
Primary
hand-
washing
Secondary
hand-
washing
Piped
water
Improved
water Tube-well
Improved
toilet
Primary
hand-
washing
Secondary
hand-
washing
Control 6281 6329 2562 2221 4779 6154 1173 1314 557 362 4258 4740
Treated 6281 6329 2562 2221 4779 6154 815 1458 429 238 2940 3587
Multivariate L1 0.968 0.944 0.971 0.987 0.804 0.756 0.428 0.419 0.485 0.412 0.130 0.128
Univariate L1
State 0.039 0.082 0.030 0.022 0.016 0.016 0.020 0.006 0.016 0.017 0.006 0.002
H. asset index 0.006 0.023 0.020 0.038 0.009 0.002 0.000 0.000 0.000 0.000 0.000 0.000
Village index 0.013 0.061 0.007 0.009 0.003 0.003 0.000 0.000 0.000 0.000 0.000 0.000
Income quintile 0.011 0.038 0.005 0.012 0.002 0.006 0.000 0.000 0.000 0.000 0.000 0.000
Female education 0.021 0.053 0.002 0.022 0.004 0.001 0.000 0.000 0.000 0.000 0.000 0.000
Scheduled tribe 0.002 0.040 0.003 0.002 0.004 0.009 0.000 0.000 0.000 0.000 0.000 0.000
Scheduled caste 0.001 0.006 0.009 0.014 0.005 0.001 0.000 0.000 0.000 0.000 0.000 0.000
Muslim 0.006 0.020 0.016 0.025 0.004 0.002 0.000 0.000 0.000 0.000 0.000 0.000
Household size 0.042 0.045 0.053 0.054 0.059 0.042 0.080 0.047 0.114 0.059 0.018 0.013
Household head:
self-employed
0.001 0.012 0.007 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000
Household head:
married
0.001 0.000 0.007 0.001 0.005 0.003 0.000 0.000 0.000 0.000 0.000 0.000
Electricity 0.027 0.067 0.008 0.032 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000
Improved stove 0.016 0.018 0.004 0.018 0.001 0.009 0.000 0.000 0.000 0.000 0.000 0.000
Chimney 0.000 0.050 0.000 0.014 0.004 0.004 0.000 0.000 0.000 0.000 0.000 0.000
Note: L1 here refers to multivariate L1, with zero indicating perfect balance and one indicating perfect imbalance.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 353
The results in Table 5 indicate that there may be an effect of piped water on acute
dysentery but not other diarrhoeal outcomes. Using PSM and CEM, there was a significant
effect at p < 0.05 of piped water on acute dysentery and using exact matching at 0.05
p < 0.1. As noted earlier, in replicating the analysis by Jalan and Ravallion (2003), we
found that piped water results in a negative effect on acute dysentery at the 5 per cent level
of significance. In considering improved water supply and tube-well as a treatment, no
significant effect on a given outcome is observed consistently across matching methods.
When using only CEM (and not other methods), improved water supply reduces preva-
lence of acute dysentery by 0.009 (p = 0.039); that is, 0.9 percentage points compared with
1.8 per cent in controls in this matched sample.
We next consider treatments of improved toilets and hand-washing in Table 6. An effect
of improved toilet on acute dysentery was detected only at 0.05 p < 0.1 using CEM but
not other methods. When using exact matching, improved toilets appear to significantly
reduce prevalence of all diarrhoea and its duration, but the sample suffered from high
attrition.
In sharp contrast, hand-washing appears to have a large, negative, and highly signifi-
cant effect (p < 0.0001) as estimated from every matching method on ‘all diarrhoea’, acute
watery diarrhoea, and diarrhoeal duration. This contrasts quite starkly with the detected
effects of water supply and sanitation for which the significance is both small and not
robust across methods. Moreover, sample size reductions were lower across all methods
when matching for hand-washing, compared with that for water and sanitation treatments.
The effect sizes detected are consistent with the experimental literature. Using PSM, pri-
mary hand-washing reduced acute watery diarrhoea by 0.058 and was highly significant
(p < 0.0001) or a 58 per cent reduction. Using PSM, secondary hand-washing reduced all
diarrhoea by 0.059 (p < 0.0001). Although the effect sizes under exact matching and CEM
are somewhat larger than those when using PSM, exact matching and CEM had higher
sample attrition and retained just a fraction of the treated units, whereas PSM retained
most treated units.
Although hand-washing appeared to reduce prevalence of acute watery diarrhoea and
diarrhoeal duration, no effect of hand-washing was detected on acute dysentery using
any matching method. Diarrhoeal duration was reduced by hand-washing by between
one-third and more than one-half of a day, depending on the matching method used.
Primary hand-washing reduces duration of diarrhoea by more than one-half a day (0.516,
p < 0.0001) when using exact matching, whereas it reduces duration by 0.317 days when
using PSM.
These results are consistent with the crude correlations observed in Table 2, which
suggested a negative correlation between improved toilet on all diarrhoea and diarrhoeal
duration as well as a negative correlation between hand-washing and different diarrhoeal
outcomes. We also conducted analyses of household-level treatments at the household
level; results are consistent with those at the individual child level (results not presented).
5. Sensitivity analyses and subgroup analyses
In this section we describe and present some of the sensitivity analyses conducted. There
is always the chance that the main analyses failed to detect a significant effect when a true
effect exists or detected a false positive. As mentioned earlier, a simple sensitivity check
is to compare estimates between using simple differences and using parametric adjust-
ment after matching; we found our estimates ‘doubly robust’ (Ho et al. 2007). Another
sensitivity analysis relates to the analyses in Table 5, which analysed piped water relative
Downloaded by [Victoria Fan] at 04:38 26 September 2011
354 V.Y. Fan and A. Mahal
Table 5. Impact estimates of water supply on diarrhoeal outcomes in children.
Exact CEM PSM
Treatment Outcome Const. Impact Const. Impact Const. Impact
Piped water All diarrhoea 0.131 0.011 0.142 0.019 0.099 0.000
(0.013) (0.019) (0.010) (0.016) (0.004) (0.005)
Acute watery 0.105 0.013 0.107 0.013 0.074 0.001
diarrhoea (0.012) (0.018) (0.009) (0.014) (0.003) (0.005)
Acute dysentery 0.014 0.011 0.019 0.015 0.009 0.004
(0.004) (0.006)† (0.003) (0.005)
(0.001) (0.001)
Diarrhoeal 0.710 0.167 0.714 0.139 0.552 0.027
duration (0.081) (0.123) (0.058) (0.093) (0.028) (0.040)
Control n
(of 18,006)
665 1173 6281
Treated n (of 6905) 528 815 6281
L1 n.a. 0.428 0.968
% BI n.a. 99.67 86.6
Improved water All diarrhoea 0.148 0.008 0.157 0.021 0.114 0.004
supply (0.012) (0.017) (0.010) (0.013) (0.004) (0.006)
Acute watery 0.117 0.004 0.118 0.011 0.085 0.000
diarrhoea (0.011) (0.015) (0.009) (0.012) (0.003) (0.005)
Acute dysentery 0.018 0.004 0.018 0.009 0.011 0.003
(0.004) (0.006) (0.003) (0.004)
(0.001) (0.002)
Diarrhoeal 0.747 0.089 0.847 0.092 0.607 0.008
duration (0.102) (0.140) (0.078) (0.106) (0.029) (0.041)
Control n
(of 18,384)
871 1314 6329
Treated n (of 6527) 931 1458 6329
L1 n.a. 0.419 0.944
% BI n.a. 99.94 68.77
Tube-well All diarrhoea 0.148 0.010 0.138 0.007 0.109 0.004
(0.020) (0.031) (0.014) (0.023) (0.006) (0.009)
Acute watery 0.116 0.002 0.099 0.016 0.084 0.001
diarrhoea (0.018) (0.027) (0.013) (0.020) (0.006) (0.008)
Acute dysentery 0.005 0.013 0.004 0.008 0.011 0.004
(0.006) (0.009) (0.003) (0.005) (0.002) (0.003)
Diarrhoeal 0.667 0.235 0.724 0.141 0.652 0.017
duration (0.119) (0.183) (0.100) (0.156) (0.053) (0.074)
Control n
(of 21,913)
313 557 2562
Treated n (of 2998) 247 429 2562
L1 n.a. 0.485 0.971
% BI n.a. 99.90 94.47
Note: Standard errors are listed below estimates. Sampling weights applied. L1 refers to the multivariate
imbalance measure; % BI refers to percentage balance improvement.
Significance at p < 0.05, †0.05 p < 0.1. CEM refers to coarsened exact matching. PSM refers to nearest-one
neighbour propensity score matching. The number of matched treated and control units are listed below estimates
for each matching method.
to any water s upply. As a sensitivity analysis, we analysed the effect of piped water rela-
tive to an unimproved water supply (Table 7). The results suggest that significance declines
somewhat, although an effect of piped water on acute dysentery remains significant at p <
0.05 using CEM.
There is also a concern that the results from PSM are sensitive to data pruning with
callipers. Whereas the main analyses above used a calliper of 0.25 standard deviations, in
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 355
Table 6. Impact estimates of improved toilet or hand-washing on diarrhoeal outcomes in children.
Exact CEM PSM
Treatment Outcome Const. Impact Const. Impact Const. Impact
Improved toilet All diarrhoea 0.165 0.085 0.131 0.017 0.089 0.006
(0.023) (0.035)
(0.017) (0.028) (0.006) (0.009)
Acute watery 0.121 0.057 0.088 0.029 0.068 0.002
diarrhoea (0.020) (0.031)† (0.014) (0.022) (0.005) (0.008)
Acute dysentery 0.014 0.014 0.021 0.017 0.004 0.005
(0.006) (0.009) (0.006) (0.010)† (0.002) (0.002)
Diarrhoeal 0.924 0.622 0.716 0.218 0.421 0.175
duration (0.137) (0.208)
(0.099) (0.159) (0.050) (0.070)
Control n
(of 22,090)
213 362 2221
Treated n (of 2821) 155 238 2221
L1 n.a. 0.412 0.987
% BI n.a. 99.65 96.53
Primary All diarrhoea 0.151 0.093 0.151 0.087 0.128 0.058
hand-washing (0.005) (0.008)
(0.005) (0.008)
(0.004) (0.006)
Acute watery 0.120 0.087 0.120 0.083 0.100 0.058
diarrhoea (0.005) (0.007)
(0.004) (0.007)
(0.004) (0.005)
Acute dysentery 0.013 0.003 0.014 0.004 0.012 0.002
(0.002) (0.003) (0.002) (0.003) (0.001) (0.002)
Diarrhoeal 0.836 0.516 0.814 0.468 0.702 0.317
duration (0.040) (0.061)
(0.036) (0.057)
(0.032) (0.045)
Control n
(of 20,131)
3639 4258 4779
Treated n (of 4780) 2730 2940 4779
L1 n.a. 0.130 0.804
% BI n.a. 99.78 93.25
Secondary All diarrhoea 0.155 0.100 0.154 0.093 0.125 0.056
hand-washing (0.005) (0.007)
(0.005) (0.007)
(0.004) (0.005)
Acute watery 0.127 0.097 0.127 0.092 0.101 0.059
diarrhoea (0.004) (0.006)
(0.004) (0.006)
(0.003) (0.005)
Acute dysentery 0.012 0.002 0.012 0.002 0.010 0.000
(0.002) (0.002) (0.002) (0.002) (0.001) (0.002)
Diarrhoeal 0.823 0.526 0.803 0.471 0.646 0.278
duration (0.037) (0.055)
(0.033) (0.051)
(0.025) (0.034)
Control n
(of 18,756)
4079 4740 6154
Treated n (of 6155) 3320 3587 6154
L1 n.a. 0.128 0.756
% BI n.a. 99.23 95.15
Note: Standard errors are listed below estimates. Sampling weights applied. L1 refers to the multivariate
imbalance measure; % BI refers to percentage balance improvement.
Significance at p < 0.05, †0.05 p < 0.1. CEM refers to coarsened exact matching. PSM refers to nearest-one
neighbour propensity score matching. The number of matched treated and control units are listed below estimates
for each matching method.
a sensitivity analysis we used progressively narrower callipers of 0.1, 0.001, 0.0001, and
0.00001 (with balance for these callipers reported in Table 3). Results with progressively
narrower callipers are shown in Table 8. The results indicate that the effects of piped water
on acute dysentery, and not other outcomes, are robust.
Another concern pertains to the coarsening of variables for CEM. In this study we
manually coarsened several variables for CEM, which includes the construction of sample
Downloaded by [Victoria Fan] at 04:38 26 September 2011
356 V.Y. Fan and A. Mahal
Table 7. Sensitivity analysis I: impact estimates of piped water on diarrhoeal outcomes in children
relative to control households with an unimproved water supply.
Exact CEM PSM
Outcome Const. Impact Const. Impact Const. Impact
All diarrhoea 0.168 0.010 0.189 0.049 0.114 0.011
(0.022) (0.032) (0.017) (0.025) (0.005) (0.007)
Acute watery 0.123 0.002 0.137 0.034 0.089 0.004
diarrhoea (0.020) (0.029) (0.015) (0.022) (0.004) (0.006)
Acute dysentery 0.012 0.008 0.027 0.022 0.008 0.003
(0.005) (0.006)† (0.005) (0.007)
(0.002) (0.002)†
Diarrhoeal duration 0.829 0.096 0.915 0.260 0.639 0.063
(0.134) (0.196) (0.102) (0.153)† (0.038) (0.054)
Control n (of 18,006) 282 482 4148
Treated n (of 6905) 247 386 4148
Multivariate
imbalance L1
n.a. 0.424 0.974
% balance improv. n.a. 99.8 81.3
Note: Standard errors are listed below estimates.
Significance at p < 0.05, †0.05 p < 0.1. Sampling weights applied.
Table 8. Sensitivity analysis II: impact estimates of piped water on diarrhoeal outcomes in children
on pruned samples.
PSM calliper 0.1 PSM calliper 0.001
PSM calliper
0.0001
PSM calliper
0.00001
Outcome Const. Impact Const. Impact Const. Impact Const. Impact
All diarrhoea 0.097 0.004 0.100 0.008 0.114 0.000 0.128 0.014
(0.004) (0.005) (0.004) (0.006) (0.007) (0.010) (0.013) (0.018)
Acute watery 0.069 0.007 0.075 0.005 0.084 0.005 0.094 0.003
diarrhoea (0.003) (0.005) (0.004) (0.005) (0.006) (0.009) (0.011) (0.016)
Acute dysentery 0.009 0.004 0.010 0.005 0.010 0.004 0.019 0.015
(0.001) (0.002)
(0.001) (0.002)
(0.002) (0.003) (0.004) (0.006)
Diarrhoeal
duration
0.526 0.008 0.513 0.053 0.574 0.024 0.732 0.184
(0.027) (0.038) (0.030) (0.042) (0.044) (0.062) (0.085) (0.119)
Control n
(of 18,006)
5965 4955 2156 650
Treated n
(of 6905)
5965 4955 2156 650
Multivariate
imbalance L1
0.967 0.960 0.848 0.368
% balance
improv.
95.5 99.9 100.0 100.0
Note: Standard errors are listed below estimates.
Significance at p < 0.05, †0.05 p < 0.1. Sampling weights applied.
quintiles of the village development index and asset index (see Section 2). As a separate
robustness check, we included the continuous index in place of quintiles and let CEM
automatically coarsen these variabl es. This automated coarsening of continuous variables
had a much smaller sample size than using the quintiles, but the results with this CEM
variant (data not presented) were consistent with CEM with quintiles. A related concern
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 357
to coarsening variables is whether impact estimates are sensitive to the specification of
covariates for matching. As one coarsens to the extreme, a variable is effectively excluded
and dropped (King et al. 2011). As noted in the results section, the results for all three
matching methods are consistent (data not presented) with results from the fourth and final
model. Given the large significance of the hand-washing results, as a robustness check
we added additional covariates of piped water, tube-well, and improved toilet, and found
robust effects of hand-washing (both primary and secondary).
Finally, we conducted subgroup analyses to estimate the impacts of primary hand-
washing and secondary hand-washing on subgroups of interest, namely: households in
the top three quintiles of assets compared with the bottom two quintiles; households
in the top three quintiles of village development compared with the bottom two quin-
tiles; and households where the highest education of a female member was secondary
schooling matriculation or higher compared with those with primary schooling or lower
education (Table 9). These subgroup analyses do not suggest major differences between
subgroups. High-asset-holding households may benefit more from hand-washing than low-
asset-holding households. In contrast, however, in low-developed villages, households
benefit more from hand-washing than high-developed villages. Similarly, hand-washing
may matter more in households with a low level of its highest educated female compared
with a high level.
6. Discussion and concluding remarks
Our quasi-experimental study assessed the impacts of water supply, sanitation and hand-
washing interventions for rural India on diarrhoeal disease using observational data. The
matching methods used in this paper are consistent with best practices in matching meth-
ods, particularly in the use of balance diagnostics, and apply a novel Coarsened Exact
Matching method. Our results are broadly consistent with the epidemiologic and clini-
cal literature that suggests a robust and large significant effect of hand-washing. We also
find smaller effects of piped water and improved toilets, although less robust to alternative
matching methods. Obviously, our results do not test the effect of water quality, which is
known to reduce diarrhoea. Moreover unlike epidemiologic studies that often lack infor-
mation on detailed household-level and village-level indicators, our study controlled for
confounding of household education, wealth, asset ownership, and village development.
Our results on the effect of piped water on diarrhoea are consistent with a previous study
by Jalan and Ravallion (2003) that used the same dataset, which found a significantly large
effect of piped water in reducing diarrhoea but only if the outcome measure used is acute
dysentery.
The paper contributes to the literature on the effects of various interventions on both
acute watery diarrhoea and acute dysentery. Rarely do studies measure both outcomes of
watery diarrhoea and dysentery; the handful of studies that measure both outcomes do not
definitively suggest larger effects on either outcome (see Fewtrell et al. 2005, Cairncross
et al. 2010). Similarly, there is no conclusive evidence suggesting that improved water
supply reduces one of these outcomes more than the other. Biologically, it may be that
pathogens spreading watery diarrhoea are less ‘sticky’ than those pathogens spreading
dysentery. If this is true then incidence of watery diarrhoea may be more amenable than
dysentery to hand-washing.
Furthermore, it is believed that dysentery (such as shigellosis) is more commonly
associated in areas with insufficient water supply and inadequate sanitation, including
wells, stagnant water sources, or sources of water that are contaminated by faecal matter
Downloaded by [Victoria Fan] at 04:38 26 September 2011
358 V.Y. Fan and A. Mahal
Table 9. Subgroup analyses: impact estimates of hand-washing on diarrhoeal outcomes in children, CEM.
Asset Index Village Development Index Female Education
Low High Low High Low High
Outcome Const. Impact Const. Impact Const. Impact Const. Impact Const. Impact Const. Impact
Treatment: primary hand-washing
All diarrhoea 0.162 0.095 0.179 0.120 0.211 0.123 0.134 0.097 0.181 0.115 0.136 0.088
(0.008) (0.012)
(0.007) (0.010)
(0.008) (0.012)
(0.006) (0.010)
(0.006) (0.009)
(0.011) (0.017)
Acute watery 0.138 0.096 0.146 0.108 0.187 0.131 0.100 0.076 0.154 0.109 0.097 0.075
diarrhoea (0.007) (0.011)
(0.006) (0.009)
(0.007) (0.011)
(0.006) (0.009)
(0.005) (0.008)
(0.009) (0.014)
Acute dysentery 0.007 0.004 0.009 0.001 0.009 0.007 0.006 0.002 0.007 0.003 0.010 0.001
(0.002) (0.003) (0.002) (0.003) (0.002) (0.004) (0.002) (0.003) (0.002) (0.002) (0.003) (0.005)
Diarrhoeal duration 0.850 0.468 0.938 0.640 1.037 0.614 0.771 0.528 0.948 0.594 0.707 0.471
(0.056) (0.087)
(0.044) (0.069)
(0.046) (0.073)
(0.051) (0.080)
(0.039) (0.061)
(0.070) (0.110)
Control n
(of 20,131)
1855 2433 2295 1993 3547 741
Treated n (of 4780) 1182 1747 1448 1481 2347 555
Treatment: secondary hand-washing
All diarrhoea 0.179 0.109 0.181 0.126 0.235 0.147 0.132 0.094 0.195 0.128 0.124 0.083
(0.007) (0.011)
(0.006) (0.010)
(0.008) (0.012)
(0.006) (0.009)
(0.006) (0.008)
(0.009) (0.014)
Acute watery 0.158 0.116 0.151 0.114 0.208 0.153 0.105 0.081 0.170 0.126 0.092 0.072
diarrhoea (0.007) (0.010)
(0.006) (0.009)
(0.007) (0.011)
(0.005) (0.008)
(0.005) (0.008)
(0.008) (0.012)
Acute dysentery 0.006 0.005 0.007 0.002 0.007 0.009 0.006 0.002 0.005 0.006 0.012 0.006
(0.002) (0.003) (0.002) (0.003) (0.002) (0.003) (0.001) (0.002) (0.001) (0.002) (0.003) (0.005)
Diarrhoeal duration 0.874 0.475 0.927 0.637 1.122 0.702 0.712 0.452 0.978 0.612 0.625 0.409
(0.049) (0.074)
(0.042) (0.063)
(0.046) (0.069)
(0.044) (0.067)
(0.037) (0.056)
(0.059) (0.090)
Control n
(of 18,756)
2497 3188 2752 2933 4559 1126
Treated n (of 6155) 1058 1566 1307 1319 2118 508
Note: Standard errors are listed below estimates. Sampling weights applied.
Significance at p < 0.05. CEM refers to coarsened exact matching. PSM refers to nearest-one neighbour propensity score matching. The number of matched treated and control
units are listed below estimates for each treatment.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 359
(Lichnevski 1996; DuPont 2010). Thus if piped water yields larger volumes of water, i s l ess
stagnant and has less faecal contamination than other forms of water supply (for example,
tube-wells), then it might be expected that piped water reduces dysentery more than tube-
wells reduce dysentery. It is believed that dysentery is transmitted in a small concentration
of bacteria relative to watery diarrhoea. Therefore watery diarrhoea, which is known to
require larger concentrations of pathogen than dysentery, would be less amenable to large
increases in water supply than dysentery.
The results from this study of rural India should be interpreted cautiously for policy
purposes. Where sample size was reduced by a large measure from exact or coarsened
exact matching, as in the case of water and sanitation, the results are likely to be less
externally valid, although internally valid for the smaller matched sample. In contrast, the
results of hand-washing are robust across matching methods, at levels of very high signif-
icance, and do not suffer from significantly reduced samples from matching. Yet with any
matching method there is the concern that the estimated effect may be biased if there are
unobserved variables predicting treatment. In the case of hand-washing, the effects may be
over-estimated if the outcomes are driven by unobserved knowledge and behaviour other
than hand-washing.
Our results may help to shed light on the long-term effects of policies relating to inter-
ventions on water supply and sanitation facilities, many of which are known to deteriorate
in quality over time or suffer from low ‘compliance’. Several reports have described the
increasingly complex challenges of water supply in everyday Indian life. A recent World
Bank report by Briscoe and Malik (2008) argued that India’s water sector has been facing
a major financing gap, which has led to crumbling or deteriorating infrastructure and the
decline in quantity and quality of public irrigation and water supply services. In response to
the decline, people have devised ingenious ways or ‘coping strategies’ around weak water
supply systems; for example, individuals drilling tube-wells or relying on others’ tube-
wells. It is estimated that some 20 million individual tube-wells are now installed. Similarly,
the urban middle class, in coping with irregular, unpredictable and often polluted and piped
public water services, have devised household storage and purification systems to improve
water quality, or private well access to tap groundwater. The challenges in infrastructure
and service delivery are likely to be compounded by corruption (for example, Davis 2004).
These common challenges of water supply were also described in a 2007 repor t pro-
duced by the Planning Commission of the Government of India, which suggests that ground
water is ‘open access common property natural resource and anyone can bore a well and
pump out water without limit’ (Planning Commission, Government of India 2007). The
variability in public provision of water has led to a National Rural Drinking Water Program,
for which a ‘good framework should consider different drinking water sources accessible in
different situations and different points of time’, and that drinking water should not rely on
a single source but rely on multiple sources to ensure water security including ground-
water, surface-water, and rainwater harvesting (Rajiv Gandhi National Drinking Water
Mission 2009). It is notable that neither of the two reports cited above made a distinction
between ‘piped water’ and other sources of water supply, such as boreholes or tube-wells.
These reports have focused primarily on socio-economic implications of India’s ‘private,
self-provision with groundwater’, and place a minor emphasis, if any, on the health implica-
tions of water supply. It is not obvious how increasing diversity of household water sources
would affect health outcomes.
The state of sanitation in India is dire. There has been an increasingly prominent focus
of policy attention on the situation of sanitation in the country (see Mara et al. 2010,
Downloaded by [Victoria Fan] at 04:38 26 September 2011
360 V.Y. Fan and A. Mahal
World Bank 2011). Mara et al. (2010) cited a few reports that found that toilets in India
are often not used for their intended purpose of defecation, but instead used for other pur-
poses such as firewood stores or storage sheds. This lack of ‘compliance’ may help to
explain why our study did not detect a consistently significant effect of improved toilets
on diarrhoea. Our study does not dispute the fact that improved sanitation, when sustained
and maintained over time, can prevent major health risks as suggested in the limited epi-
demiologic literature. The costs of unimproved sanitation to Indians are large. The World
Bank (2011) recently reported that ‘inadequate sanitation’ caused an annual economic loss
of US$53.8 billion in 2006, as measured by premature mortality and other health-related
impacts, as well as productive time lost, and impacts related to drinking water that are
primarily shouldered by children younger than age five.
Hand-washing has been one of many potential interventions suggested by
UNICEF/WHO (2009), including a number of other (what they call) ‘primary preven-
tive measures’ such as improved drinking water supply, community-wide sanitation, and
vaccination for rotavirus and measles. Despite this emphasis in that UNICEF/WHO
report, hand-washing prevalence is not a routinely reported indicator in its WHO Statistics
Database, unlike the indicators of percentage of individuals with improved water or with
improved sanitation (or even other self-reported indicators of immunisation coverage or
use of various family planning measures). Hand-washing is also not mentioned in a WHO
(2009) report.
8
Some aspects of hand-washing have been measured in the Demographic
and Health Surveys (which canvass detailed self-reported family planning behaviours) for
multiple surveys over 1999–2005 in sub-Saharan Africa (see Macro International 2011).
Thus, a key policy recommendation from the results of this paper would be that preva-
lence of hand-washing behaviour, even self-reported, is an important indicator for routine
health surveillance, and is arguably as important as measures of water supply, sanitation,
and other self-repor ted health-seeking behaviours.
The policy drivers of hand-washing behaviour in a population are not well understood.
Of course hand-washing also requires access to water, though not necessarily an ‘improved’
water supply. There are studies that suggest that educational interventions can change
a population’s behaviour at least in the short r un and can reduce diarrhoea incidence.
A study by Stanton and Clemens (1987) tailored an intervention to a population’s existing
behaviours to provide an eight-week community-based health education programme with
local experienced trainers. The intervention had three main messages: proper hand-washing
before food preparation, defecating away from the house at a proper site, and suitable
disposal of waste and faeces. They relied on public demonstrations, small-group discus-
sions and community-wide meetings as well as posters, games and stories to illustrate that
messages may be effective. As noted by Luby (2010), behaviour-change campaigns ‘often
require substantial resources (especially trained personnel, community organization, and
funding)’. Temporal features of these interventions (that is, how frequently the intervention
is offered and over what length of time) may determine whether the intervention produces
long-lasting behaviour changes in the community. Changing norms through community
leaders, social networks, and local community organisation is also likely to be an important
feature of long-lasting interventions.
As we noted earlier, progress in this sector of water, sanitation, and hygiene has been
slow in India. We have no evidence on whether hand-washing behaviours have improved
in the past two decades. We hope that the results from this study can spur needed attention
to improving water, sanitation, and hygiene, but particularly hygiene and hand-washing as
well as sanitation in order to improve the health of India’s children.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 361
Notes
1. Estimating propensity scores at either household or individual level in the first-stage model fol-
lowed by matching to the individual level will lead to equivalent matching results, if the same
pretreatment covariates village-level and household-level covariates are used. If the propensity
score is estimated at individual level, then individual children from the same village and house-
hold will be assigned the same propensity score if they come from the same household. If the
score is estimated at household level, then a single household will be assigned a single propensity
score and therefore children within the same household would receive the same score. Clustering
standard errors in the first-stage logit model will not affect the distribution of propensity scores,
only their standard errors.
2. Because exact matching is balanced on obser vables, estimates using parametric analyses are the
same as estimates taking a simple difference. See Ho et al. (2007).
3. We chose the definition of piped water that appeared to be closest to that used in Jalan and
Ravallion (2003). See Appendix 1.
4. A tube-well (or bore-well) relies on a mechanical or motorised pump, whereas a hand pump
is similar to a tube-well but relies on a hand pump. Piped water refers to water distributed by
pipeline, but does not refer to a kind of extraction method.
5. These estimates are consistent with a recent study of 288 households by Biran et al. (2009),
who found a prevalence of observed hand-washing with both hands on any occasion of
28–31 per cent. A separate study by Biran et al. (2008) found that self-reported hand-
washing with soap underestimated observed hand-washing with soap, whereas self-reported
hand-washing generally overestimated observed hand-washing with water.
6. Household-level variables include demographic characteristics, whether the household belongs
to a scheduled tribe, a scheduled caste, household religion, household size, occupation of
household head, radio and television listening behaviours of household members, proportion
of household members who are elderly, proportion of adults who are female, proportion of chil-
dren who are male, whether household head is male, whether household head is single, whether
household head is married, education level of household head, highest education level of female
in household, total household income (in quintiles), and asset index (in quintiles) constructed
from asset ownership (house, other property, bicycle, sewing machine, thresher, winnower, bul-
lock cart, radio, television, f an, livestock, nature of house, condition of house, number of rooms
in house, electricity, clean stove, chimney for cooking, ventilated kitchen, separate kitchen, use
of landholding for cultivation, landholding size, gross cropped area, and gross irrigated area).
Village-level infrastructure variables including proportion of gross cropped area that is irrigated,
whether the village has a daycare centre, a primary school, a middle school, a high school, type
of road approaching the village, whether the village has a bus stop, a railway station, a post
office, a telephone facility, a community television centre, a library, a bank, and a market as well
as village demographic variables of village population, the student–teacher ratio, male–female
student ratio, and male–female minority student ratio.
7. Three pretreatment variables used by Jalan and Ravallion (2003) the student–teacher ratio,
the male–female student ratio, and the male–female minority student ratio are not necessarily
pretreatment and were thus excluded. Moreover, they had large numbers of missing observations.
8. Hand-washing is not mentioned in a WHO report on global health risks (WHO 2009). The
report defines the category of ‘unsafe water, sanitation, and hygiene’ as ‘improved water’ and
‘improved sanitation’ with specific types of water sources (for example, piped water, borehole,
and so forth) and sanitation facilities (for example, septic tank, pit latrine, and so forth) without
mention of any ‘hygiene’-related behaviours such as hand-washing.
References
American Public Health Association, 2008. Control of communicable diseases manual: an offi-
cial report of the American Public Health Association, ed. D.L. Heymann. Washington, DC:
American Public Health Association.
Biran, A., et al., 2008. Comparing the performance of indicators of hand-washing practices in rural
Indian households. Tropical medicine & international health, 13 (2), 278–285.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
362 V.Y. Fan and A. Mahal
Biran, A., et al., 2009. The effect of a s oap promotion and hygiene education campaign on hand-
washing behaviour in rural India: a cluster randomised trial. Tropical medicine & international
health, 14 (10), 1303–1314.
Black, et al., 2010. Global, regional, and national causes of child mortality in 2008: a systematic
analysis. The lancet, 375, 1969–1987.
Bose, R., 2009. The impact of water supply and sanitation interventions on child health: evidence
from DHS surveys. Mimeo. New Delhi: International Initiative for Impact Evaluation.
Briscoe, J. and Malik, R.P.S., 2008. India’s water economy: bracing for a turbulent future. New Delhi:
World Bank. Available from: http://go.worldbank.org/R09M773280 [Accessed 3 January 2011]
Cairncross, S., et al., 2010. Water, sanitation, and hygiene for the prevention of diarrhoea.
International journal of epidemiology, 39, i193–i205.
Cash, R., 2011. Email. Personal communication, 7 June.
Clasen, T., et al., 2007. Interventions to improve water quality for preventing diarrhoea: systematic
review and meta-analysis. British medical journal, 334 (7597), 782.
Clasen, T.F., et al., 2010. Interventions to improve disposal of human excreta for preventing
diarrhoea. Cochrane database of systematic reviews, 6, CD007180.
Curtis, V., Cairncross, S. and Yonli, R., 2000. Domestic hygiene and diarrhoea pinpointing the
problem. Tropical medicine & international health, 5 (1), 22–32.
Davis, J., 2004. Corruption in public service delivery: experience from South Asia’s water and
sanitation sector. World development, 31 (1), 53–71.
Department of Drinking Water Supply, undated. Rajiv Gandhi National Drinking Water Mission,
Ministry of Rural Development, Government of India. Available from: http://www.ddws.nic.in/
DuPont, H.L., 2010. Shigella species (bacillary dysentery). In: G.L. Mandell, J.E. Bennett and
R. Dolin, eds. Mandell, Douglas, and Bennett’s principles and practice of infectious disease.
7th ed. Philadelphia: Churchill Livingston (Elsevier), 2905–2910.
Ejemot, R.I., et al., 2008. Hand washing for preventing diarrhoea. Cochrane database of systematic
reviews, 1, CD004265.
Fewtrell, L., et al., 2005. Water, sanitation, and hygiene interventions to reduce diarrhoea in less
developed countries: a systematic review and meta-analysis. Lancet infectious diseases, 5 (1),
42–52.
Filmer, D. and Pritchett, L.H., 2001. Estimating wealth effects without expenditure data-or tears: an
application to educational enrolments in states of India. Demography, 38 (1), 115–132.
Ho, D.E., et al., 2007. Matching as nonparametric preprocessing for reducing model dependence in
parametric causal inference. Political analysis, 15, 199–236.
Ho, D.E., et al., 2011. MatchIt: nonparametric pre-processing for parametric causal inference.
Journal of statistical software, 42.
Iacus, S.M., King, G. and Porro, G., 2011a. Multivariate matching methods that are monotonic
imbalance bounding. Journal of the American Statistical Association, 106 (493), 346–361.
Iacus, S.M., King, G. and Porro, G., 2011b. Causal inference without balance checking: coarsened
exact matching. Political analysis, forthcoming.
Imai, K., King, G. and Stuart, E.A., 2008. Misunderstandings between experimentalists and observa-
tionalists about inference. Journal of the Royal Statistical Society, Series A (statistics in society),
171, 481–502.
Jalan, J. and Ravallion, M., 2003. Does piped water reduce diarrhoea for children in rural India?
Journal of econometrics, 112, 153–173.
Khanna, G., 2008. The impact on child health from access to water and sanitation and other socioe-
conomic factors. Geneva: The Graduate Institute of International and Development Studies, HEI
Working Paper No. 02/2008.
King, G., et al., 2011. Comparative effectiveness of matching methods for causal inference. Working
paper. Cambridge, MA: Harvard Unviersity.
Lichnevski, M., 1996. Shigella dysentery and s higella infections. Eastern Mediterranean health
journal, 2 (1), 102–106. Available from: http://www.emro.who.int/publications/emhj/0201/
14.htm [Accessed 1 March 2011].
Luby, S.P., et al., 2004. Effect of intensive handwashing promotion on childhood diarrhoea in high-
risk communities in Pakistan: a randomised controlled trial. Journal of the American Medical
Association, 291 (21), 2547–2554.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 363
Luby, S.P., et al., 2010. A community-randomised controlled trial promoting waterless hand sanitizer
and handwashing with soap, Dhaka, Bangladesh. Tropical medicine & international health,15
(12), 1508–1516.
Macro International Inc, 2011. MEASURE DHS STATcompiler. Available from: http://www.
measuredhs.com [Accessed 7 January 2011].
Mara, D., et al., 2010. Sanitation and health. PLoS medicine, 7 (11), e1000363.
Planning Commission, Government of India, 2007. Report of the expert group on ground water
management and ownership. New Delhi: Government of India, iii. Available from: http://
planningcommission.nic.in/reports/genrep/rep_grndwat.pdf
Pradhan, M. and Rawlings, L.B., 2002. The impact and targeting of social infrastructure investments:
lessons from the Nicaraguan Social Fund. World Bank economic review, 16 (2), 275–295.
Prichett, L., 2010. Email. Personal communication, 31 March.
Rajiv Gandhi National Drinking Water Mission, 2009. National rural drinking water programme
movement towards ensuring people’s drinking water security in rural India: framework for
implementation, 2009–2012. New Delhi: Department of Drinking Water Supply, Rajiv Gandhi
National Drinking Water Mission, Ministry of Rural Development, Government of India.
Available from: http://ddws.gov.in/popups/RuralDrinkingWater_2ndApril.pdf
Rosenbaum, P.R. and Rubin, D.B., 1983. The central role of the propensity score in observational
studies for causal effects. Biometrika, 70, 41–55.
Shariff, A., 1999. India: human development report. a profile of Indian states in the 1990s.New
Delhi: National Council of Applied Economic Research and Oxford University Press, 300.
Stanton, B.F. and Clemens, J.D., 1987. An educational intervention for altering water-sanitation
behaviours to reduce childhood diarrhoea in urban Bangladeshi: a randomized trial to assess
the impact of the intervention on hygienic behaviours and rates of diarrhoea. American journal
of epidemiology, 125 (2), 292–301.
UNICEF/WHO, 2009. Diarrhoea: why children are still dying and what can be done. Geneva:
WHO and New York: UNICEF. Available from: http://www.childinfo.org/files/diarrhoea_hires.
pdf [Accessed 25 March 2010].
Waddington, H., et al., 2009. Water, sanitation, and hygiene interventions to combat childhood
diarrhoea in developing countries. New Delhi: The International I nitiative for Impact Evaluation
(3ie), 3ie Synthetic Reviews. Available from: http://www.3ieimpact.org/admin/pdfs2/17.pdf
[Accessed 8 January 2011].
WHO, 2009. Global health risks: mortality and burden of disease attributable to selected
major risks. Geneva: WHO, 23 and 38. Available from: http://www.who.int/healthinfo/
global_burden_disease/GlobalHealthRisks_report_full.pdf [Accessed 3 January 2011].
WHO, 2011. WHO Statistical Information System (WHOSIS). Available from: http://www.who.int/
gho/en/ [Accessed 1 May 2011].
World Bank, 2011. The economic impacts of inadequate sanitation in India: inadequate sanitation
costs India Rs. 2.4 trillion (US$53.8 billion). Water and Sanitation Program (World Bank), Asian
Development Bank, AusAID, and UKAID. Available from: http://www.wsp.org/wsp/sites/wsp.
org/files/publications/wsp-esi-india.pdf [Accessed 3 January 2001].
Zwane, A.P. and Kremer, M., 2007. What works in fighting diarrhoeal diseases in developing
countries? A critical review. World Bank Research Observer, 22 (1), 1–24.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
364 V.Y. Fan and A. Mahal
Appendix 1. Household access to piped water
Household access to piped water by income quintile and by highest education of female member.
Households with piped water stratified by highest
education of female members
Piped water treatment
variable (dataset/s
used) Income quintiles
Number of
households
Households
with piped
water Illiterate Primary Matriculation
Secondary
or higher
Percentage of
people with
piped water
Piped water in any
season (household
dataset only)
Lowest 6584 1789 789 686 278 36 27.17
Second 6509 1653 693 623 301 36 25.40
Third 6542 1765 680 605 414 66 26.98
Four th 6694 1984 685 646 542 111 29.64
Highest 6901 2320 688 630 766 236 33.62
Full sample 33,230 9511 3535 3190 2301 485 28.62
Piped water in any
season (merged
household and
village dataset)
Lowest 6579 1786 787 685 278 36 27.15
Second 6506 1653 693 623 301 36 25.41
Third 6539 1765 680 605 414 66 26.99
Four th 6692 1983 685 646 541 111 29.63
Highest 6900 2319 688 629 766 236 33.61
Full sample 33,216 9506 3533 3188 2300 485 28.62
Piped water in all
seasons (merged
household and
village dataset)
Lowest 6579 1508 672 582 224 30 22.92
Second 6506 1376 585 514 246 31 21.15
Third 6539 1489 579 508 346 56 22.77
Four th 6692 1665 593 535 445 92 24.88
Highest 6900 2025 598 553 667 207 29.35
Full sample 33,216 8063 3027 2692 1928 416 24.27
Jalan and Ravallion
(2003)
Lowest 6581 1707 768 655 251 33 27.18
Second 6508 1567 674 590 274 29 25.40
Third 6543 1658 667 560 371 60 26.96
Four th 6694 1814 660 602 462 90 29.62
Highest 6904 2081 665 593 638 185 33.63
Full sample 33,230 8827 3434 3000 1996 397 28.62
Notes: Piped water in any season refers to piped water use in either summer or non-summer seasons; piped water in all seasons refers to piped water use in both summer and
non-summer seasons. After merging the household dataset to the village dataset, the total number of available household observations decreases from 33,230 to 33,216. (Jalan
and Ravallion also used only 33,216 observations in the first-stage model [Table 2].) Note that the number of households with access to ‘piped water’ in Jalan and Ravallion
(2003) was between these two numbers. We think that Jalan and Ravallion were more likely to have used piped water in any season, rather than in all seasons, as the ‘% of people
with piped water’ match exactly at 28.62 per cent, even though the total number of households with piped water differs (9511 or 9506 compared with Jalan and Ravallion’s total
of 8827).
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 365
Appendix 2. Replication of Table 2 in the paper by Jalan and Ravallion (2003)
(1) (2) (3) J&R
Variable Coefficient z-stat Coefficient z-stat Coefficient z-stat Coefficient t-stat
Constant 1.150 4.24 1.588 5.50 2.740 7.66 1.495 5.40
Household variables
tribe 0.259 4.73 0.231 4.13 0.219 3.57 0.213 4.20
caste 0.028 0.69 0.017 0.42 0.010 0.23 0.010 0.29
hindu 0.237 1.50 0.254 1.57 0.306 1.88 0.242 1.71
muslim 0.206 1.22 0.207 1.20 0.272 1.55 0.216 1.43
christian 0.552 2.98 0.349 1.84 0.573 2.88 0.404 2.43
sikh 0.874 4.12 0.927 4.29 0.842 3.81
0.866 4.53
hhsize 0.010 1.62 0.010 1.63 0.013 1.83 0.003 0.57
cultland 0.154 2.94 0.055 1.03 0.037 0.63 0.171 1.91
houseown 0.189 2.55 0.172 2.29 0.153 1.88 0.190 2.85
propown 0.041 0.93 0.001 0.02 0.050 1.04 0.002 0.04
bicyown 0.246 7.04 0.254 7.15 0.238 6.20 0.265 8.24
sewmown 0.025 0.50 0.010 0.19 0.015 0.26 0.012 0.25
thresown 0.105 1.01 0.090 0.85 0.175 1.47 0.058 0.58
winnown 0.217 1.71 0.174 1.34 0.150 1.06 0.218 1.82
cartown 0.259 5.05 0.262 5.01 0.282 5.06 0.259 5.43
radioown 0.030
0.69 0.035 0.80 0.040 0.85 0.010 0.25
tvown 0.128 2.33 0.106 1.88 0.057 0.94 0.081 1.34
fanown 0.022 0.48 0.018 0.38 0.033 0.65 0.013 0.32
livstown 0.131 3.55 0.099 2.64 0.088 2.15 0.078 2.34
hsnat_kuccha 0.123 3.12 0.109 2.71 0.173 4.00 0.100 2.78
hsnat_pucca 0.106 2.19 0.103 2.09 0.135 2.53 0.120 2.71
hscond_good 0.036 0.52 0.030 0.42 0.010 0.13 0.002 0.04
hscond_liv 0.096 1.64 0.097 1.64 0.049 0.76 0.108 1.37
rooms_1 0.150 1.76 0.121 1.40 0.027 0.28 0.068 1.37
rooms_2 0.025 0.32 0.042 0.53 0.146 1.64 0.075 0.95
rooms_35 0.081 1.11 0.075 1.01 0.142 1.69 0.020 1.11
kitchen 0.006 0.15 0.008 0.19 0.003 0.06 0.081 0.53
vent 0.091 2.27 0.084 2.06 0.099 2.24 0.406 2.21
elec 0.471 11.80 0.433 10.69 0.435 9.97 0.024 11.22
occ1 0.082 1.31
0.050 0.79 0.008 0.11 0.024 0.48
occ2 0.062 0.87 0.086 1.20 0.082 1.06 0.146 0.43
occ3 0.135 1.73 0.095 1.20 0.007 0.08 0.069 2.25
occ4 0.031 0.41 0.043 0.56 0.002 0.03 0.201 0.96
radiom 0.081 1.31 0.089 1.42 0.071 1.05 0.124 3.48
radiof 0.049 0.79 0.045 0.72 0.068 1.01 0.094 2.18
tvm 0.173 2.80 0.122 1.95 0.186 2.76 0.039 1.29
tvf 0.037 0.57 0.021 0.31 0.069 0.98 0.090 0.49
newsm 0.011 0.24 0.006 0.12 0.047 0.93 0.041 1.81
newsf 0.125 2.39 0.092 1.74 0.130 2.27 0.114 0.63
propold 0.082 0.50 0.151 0.92 0.077 0.43 0.046 1.07
propfemad 0.044 0.34
0.023 0.18 0.075 0.53 0.084 0.33
propmalec 0.019 0.44 0.022 0.51 0.038 0.80 0.055 0.78
malehead 0.223 2.30 0.246 2.50 0.332 3.11 0.180 2.32
single 0.185 0.91 0.182 0.89 0.165 0.77 0.167 1.27
married 0.040 0.56 0.018 0.24 0.002 0.02 0.026 0.42
hhead_ed1 0.212 2.18 0.219 2.22 0.243 2.26 0.130 1.45
hhead_ed2 0.100 1.04 0.123 1.26 0.154 1.46 0.037 0.42
hhead_ed3 0.059 0.63 0.076
0.79 0.091 0.87 0.034 0.39
hhead_ed4 0.111 0.89 0.135 1.07 0.111 0.80 0.055 0.48
(Continued)
Downloaded by [Victoria Fan] at 04:38 26 September 2011
366 V.Y. Fan and A. Mahal
Appendix 2. (Continued)
(1) (2) (3) J&R
Variable Coefficient z-stat Coefficient z-stat Coefficient z-stat Coefficient t-stat
croparea 0.000 0.17 0.000 0.45 0.000 1.07 0.000 0.67
irrarea 0.000 0.61 0.000 0.64 0.000 0.46 0.001 1.34
landless 0.150 3.01 0.187 3.70 0.267 4.93 0.328 4.00
landmarg 0.187 3.05 0.219 3.52 0.325 4.86 0.311 3.99
landsmall 0.153 2.33 0.167 2.49 0.245 3.42 0.221 2.92
Village-level variables
vsize 0.117 4.43 0.082 4.27
pirrig_75 0.102 2.52 0.202 4.49 0.048 1.19
pirrig_50_75 0.108 2.08 0.145 2.56 0.194 4.18
angan 0.084 2.35 0.126 3.22
0.072 2.23
prim 0.070 1.27 0.223 2.30 0.081 1.43
mid 0.035 0.94 0.114 2.77 0.090 2.58
high 0.210 4.81 0.144 3.07 0.265 7.41
fmvill 0.005 3.16 0.106 3.01
fmmin 0.077 2.11
puccaroad 0.183 3.07 0.259 3.86 0.194 3.64
kuccharoad 0.044 0.80 0.033 0.53 0.002 0.03
busin 0.266 6.32 0.287 6.37 0.114 2.95
railin 0.115 1.34 0.141 1.54 0.009 0.18
postin 0.032 0.76 0.043 0.91 0.022 0.55
phonein 0.319 7.74 0.246 5.55 0.331 9.66
tvin 0.003 0.08 0.068 1.37 0.099 2.66
libin 0.023 0.47 0.048 0.93 0.042 1.12
bankin 0.252 5.59 0.165 3.43 0.191 4.66
mktin 0.361 6.25 0.468 7.29 0.317 6.09
stratio 0.002 5.30
n 26,876 26,867 22,396 33,216
Notes: The regressions above attempt to replicate the first-stage logit model by Jalan and Ravallion (J&R) (2003).
The regressions presented are done for all households including those without children. Regression (1) uses only
household-level variables, (2) uses both household-level and village-level variables after dropping four village-
level variables, and (3) adds in two village-level variables. The number of observations in Regression (1) is 26,876,
although the overall dataset has 33,216 households, because when one includes the variable ‘proportion of males
among children’ into the regression, households without children of any age are automatically dropped and not all
households have children of any ages. Regression (2) does not include four of the village-level variables (village
size, female to male students in the village, female to male students for minority groups, and student–teacher
ratio in the village) that Jalan and Ravallion (2003) used. There were only 26,653 observations with village
size, 22,555 observations with female–male student ratio in village, and 14,190 observations with female–male
minority (SC, ST) ratio in village, and 13,903 observations with student–teacher ratio in village. Regression
(3) does not include two of the village-level variables that Jalan and Ravallion (2003) used; these variables were
not included because of the large number of missing observations and if included, these variables would reduce
the total number of observations used in the regression. Except for the village size, these three variables were
marked as missing if the denominator was zero (that is, if the village did not have any male students, did not
have minority male students, or did not have any teachers). Village size and female-to-male ratio in village was
included in Regression (3) as it lost 4471 observations (as opposed to the other two variables, which would have
resulted in a loss of 12,964 observations). Jalan and Ravallion (2003) do not include the two stratifying variables,
household income and highest educational attainment by female in household, in the logit model, although they
estimate the subgroup effects separately.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 367
Appendix 3. Histograms of estimated propensity scores
Histograms of estimated propensity scores treated by piped water in any season (A) or
without piped water in any season (B) at household level
(A) Treated by piped water in any season
0 .02 .04 .06 .08
Fraction
0.2.4.6.81
psmatch2: Propensity Score
(B) Not treated by piped water in any season
0 .05
.1 .15
Fraction
0.2.4.6.81
psmatch2: Propensity Score
Note: These propensity scores correspond to the regression in column (3) in Appendix 2.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
368 V.Y. Fan and A. Mahal
Appendix 4. Summary of balance diagnostics using the specification by Jalan and
Ravallion (2003)
Full unmatched data Matched data % BI
Variable
Means
Treated
Means
Control
SD
Control
Mean
Diff
Means
Treated
Means
Control
SD
Control
Mean
Diff
Mean
Diff
distance 0.38 0.23 0.15 0.14 0.34 0.34 0.16 0.00 98.70
pirrig_75 0.23 0.23 0.42 0.01 0.25 0.25 0.43 0.00 47.62
pirrig_50_75 0.11 0.11 0.32 0.00 0.12 0.10 0.29 0.02 2496.71
angan 0.63 0.46 0.50 0.17 0.61 0.64 0.48 0.04 79.41
prim 0.90 0.87 0.33 0.02 0.90 0.90 0.30 0.00 97.71
mid 0.49 0.36 0.48 0.13 0.47 0.46 0.50 0.01 92.71
high 0.34 0.19 0.39 0.16 0.31 0.28 0.45 0.03 82.77
puccaroad 0.48 0.33 0.47 0.15 0.46 0.41 0.49 0.04 71.72
kuccharoad 0.42 0.57 0.50 0.15 0.46 0.48 0.50 0.02 87.22
busin 0.59 0.36 0.48 0.24 0.56 0.56 0.50 0.01 96.92
railin 0.04 0.03 0.17 0.02 0.04 0.04 0.19 0.00 96.43
postin 0.56 0.38 0.49 0.18 0.53 0.50 0.50 0.03 84.06
phonein 0.50 0.26 0.44 0.23 0.45 0.44 0.50 0.01 94.75
tvin 0.31 0.20 0.40 0.11 0.30 0.29 0.45 0.01 92.90
libin 0.20 0.12 0.33 0.08 0.19 0.16 0.37 0.02 69.65
bankin 0.24 0.17 0.38 0.07 0.23 0.20 0.40 0.03 60.26
mktin 0.11 0.08 0.27 0.03 0.10 0.07 0.26 0.03 0.27
tribe 0.09 0.15 0.36 0.06 0.10 0.10 0.31 0.01 89.80
caste 0.23 0.24 0.43 0.01 0.24 0.24 0.43 0.00 73.79
hindu 0.85 0.81 0.39 0.04 0.83 0.85 0.36 0.02 53.73
muslim 0.09 0.12 0.33 0.03 0.10 0.09 0.29 0.01 69.42
christian 0.03 0.02 0.15 0.01 0.03 0.02 0.14 0.01 59.00
sikh 0.02 0.03 0.18 0.01 0.02 0.03 0.16 0.00 69.34
hhsize 7.06 6.93 3.01 0.13 7.00 6.98 3.06 0.02 84.81
cultland 0.61 0.69 0.46 0.07 0.61 0.64 0.48 0.02 71.03
houseown 0.94 0.97 0.18 0.03 0.95 0.95 0.21 0.00 97.86
propown 0.17 0.12 0.33 0.05 0.15 0.15 0.36 0.00 98.75
bicyown 0.49 0.58 0.49 0.09 0.52 0.51 0.50 0.01 87.46
sewmown 0.20 0.13 0.33 0.08 0.17 0.18 0.39 0.01 83.66
thresown 0.03 0.04 0.18 0.01 0.03 0.03 0.16 0.00 31.33
winnown 0.02 0.01 0.10 0.01 0.02 0.02 0.13 0.00 92.49
cartown 0.12 0.15 0.35 0.03 0.13 0.15 0.35
0.01 48.22
radioown 0.44 0.37 0.48 0.07 0.42 0.41 0.49 0.02 78.70
tvown 0.22 0.12 0.32 0.10 0.19 0.17 0.37 0.03 75.38
fanown 0.35 0.21 0.41 0.14 0.32 0.31 0.46 0.01 92.20
livstown 0.64 0.69 0.46 0.05 0.64 0.68 0.47 0.04 22.64
hsnat_kuccha 0.43 0.57 0.49 0.14 0.45 0.48 0.50 0.02 82.87
hsnat_pucca 0.23 0.15 0.36 0.08 0.23 0.20 0.40 0.02 69.89
hscond_good 0.27 0.21 0.41 0.06 0.26 0.24 0.43 0.02 61.20
hscond_liv 0.65 0.68 0.47 0.03 0.65 0.66 0.47 0.01 62.59
rooms_1 0.26 0.26 0.44 0.01 0.26 0.30 0.46 0.04 566.26
rooms_2 0.37 0.37 0.48 0.00 0.37 0.37 0.48 0.01 102.06
rooms_35 0.31 0.32 0.47 0.00 0.32 0.28 0.45 0.04 1560.46
kitchen 0.54 0.40 0.49 0.14 0.50 0.49 0.50 0.01 94.32
vent 0.46 0.32 0.47 0.14 0.41 0.42 0.49 0.00 96.80
elec 0.67 0.41 0.49 0.26 0.61 0.62 0.49 0.01 96.90
occ1 0.55 0.60 0.49 0.05 0.55 0.58 0.49 0.03 32.46
occ2 0.16 0.17 0.38 0.01 0.17 0.17 0.37 0.01 45.31
occ3 0.10 0.08 0.27 0.02 0.10 0.09 0.28 0.01 36.32
(Continued)
Downloaded by [Victoria Fan] at 04:38 26 September 2011
Journal of Development Effectiveness 369
Appendix 4. (Continued)
Full unmatched data Matched data % BI
Variab le
Means
Treated
Means
Control
SD
Control
Mean
Diff
Means
Treated
Means
Control
SD
Control
Mean
Diff
Mean
Diff
occ4 0.10 0.08 0.28 0.01 0.10 0.08 0.28 0.01 10.41
radiom 0.62 0.52 0.50 0.10 0.61 0.58 0.49 0.03 70.52
radiof 0.55 0.45 0.50 0.10 0.54 0.51 0.50 0.03 71.32
tvm 0.46 0.32 0.47 0.14 0.44 0.40 0.49 0.04 74.11
tvf 0.40 0.26 0.44 0.14 0.38 0.35 0.48 0.03 77.74
newsm 0.39 0.28 0.45 0.11 0.36 0.34 0.47 0.02 79.51
newsf 0.20 0.14 0.35 0.06 0.20 0.17 0.37 0.03 46.24
propold 0.06 0.05 0.09 0.01 0.06 0.06 0.10 0.00 76.89
malehead 0.95 0.97 0.18 0.01 0.96 0.96 0.20 0.00 91.52
single 0.00 0.01 0.07 0.00 0.00 0.00 0.07 0.00 47.20
married 0.92 0.93 0.26 0.01 0.92 0.92 0.28 0.00 100.00
hhead_ed1 0.43 0.52 0.50 0.09 0.46 0.48 0.50 0.02 77.13
hhead_ed2 0.29 0.24 0.43 0.05 0.28 0.29 0.45 0.01 86.97
hhead_ed3 0.21 0.18 0.39 0.03 0.20 0.18 0.38 0.02 19.41
hhead_ed4 0.03 0.03 0.16 0.00 0.03 0.03 0.16 0.00 85.03
croparea 43.80 43.24 80.20 0.56 44.70 43.76 82.58 0.94 69.78
irrarea 20.31 21.83 59.40 1.52 21.99 21.16 54.78 0.84 44.76
landless 0.47 0.47 0.50 0.00 0.45 0.44 0.50 0.01 218.81
landmarg 0.16 0.20 0.40 0.04 0.16 0.17 0.38 0.01 74.29
landsmall 0.10 0.10 0.31 0.01 0.10 0.11 0.31 0.01 3.96
statecode 7.78 9.18 4.55 1.40 8.12 8.12 4.23 0.00 100.00
Notes: %BI refers to percentage balance improvement. Common support was applied and a calliper of
0.001 propensity score (0.0425 standard deviation [SD]) that Jalan and Ravallion (2003) used. In total 5538 treated
units (of 6905) and 12,800 control units (of 18,006) were matched. Although overall distance in propensity score
improved, balance decreased in the following variables: village size, proportion of irrigation land, presence of
bank in village, presence of market in village, caste, Christian, household size, ownership of cart or livestock,
number of rooms, occupation, land ownership, and education of household head.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
370 V.Y. Fan and A. Mahal
Appendix 5. Replication of Jalan and Ravallion (2003) with PSM
J&R specification F&M specification
Individual Household Individual Household
Outcome Coeff. SE Coeff. SE Coeff. SE Coeff. SE
(1) Nearest-five matching on OR (PS)
All diarrhoea 0.015 (0.005)
0.019 (0.007)
0.016 (0.005)
0.020 (0.007)
Acute watery
diarrhoea
0.015 (0.005)
0.019 (0.007)
0.017 (0.004)
0.021 (0.006)
Acute dysentery 0.004 (0.001)
0.006 (0.002)
0.005 (0.001)
0.007 (0.002)
Diarrhoeal duration 0.046 (0.037) .. .. 0.056 (0.036) .. ..
N
T
5666 3609 6655 4222
N
C
17,462 11,399 17,854 11,652
(2) Nearest-five matching on PS
All diarrhoea 0.010 (0.005)
0.012 (0.007) 0.013 (0.005)
0.016 (0.007)
Acute watery
diarrhoea
0.010 (0.004)
0.012 (0.006)
0.014 (0.004)
0.017 (0.006)
Acute dysentery 0.003 (0.001)
0.004 (0.002)
0.004 (0.001)
0.006 (0.002)
Diarrhoeal duration 0.025 (0.059) .. .. 0.042 (0.035) .. ..
N
T
6743 4272 6873 4350
N
C
17,857 11,648 17,960 11,721
Notes: For household-level analyses, the outcome is the probability of having at least one case of child diarrhoea
in the household. J&R, Jalan and Ravallion (2003); F&M, Fan and Mahal, present article.
Significance at
p < 0.05. Analyses in this table used Stata SE 11, pscore program. Nearest five matching was not possible given
insufficient numbers of control units. A calliper of 0.001 was used here.
Downloaded by [Victoria Fan] at 04:38 26 September 2011
... A longitudinal study in the slums of Ethiopia shows sanitation facilities and hygienic condition of households were associated with acute diarrhoea [41]. Studies have also indicated improved water, sanitation and hygiene conditions of the [42,43]. Unimproved sources of drinking water, quality of drinking water, absences of sanitation facilities and garbage collection was associated with stomach problem in urban India [16,44,45]. ...
Article
Full-text available
Background India suffers from a high burden of diarrhoea and other water-borne diseases due to unsafe water, inadequate sanitation and poor hygiene practices among human population. With age the immune system becomes complex and antibody alone does not determine susceptibility to diseases which increases the chances of waterborne disease among elderly population. Therefore the study examines the prevalence and predictors of water-borne diseases among elderly in India. Method Data for this study was collected from the Longitudinal Ageing Study in India (LASI), 2017–18. Descriptive statistics along with bivariate analysis was used in the present study to reveal the initial results. Proportion test was applied to check the significance level of prevalence of water borne diseases between urban and rural place of residence. Additionally, binary logistic regression analysis was used to estimate the association between the outcome variable (water borne diseases) and the explanatory variables. Results The study finds the prevalence of water borne disease among the elderly is more in the rural (22.5%) areas compared to the urban counterparts (12.2%) due to the use of unimproved water sources. The percentage of population aged 60 years and above with waterborne disease is more in the central Indian states like Chhattisgarh and Madhya Pradesh followed by the North Indian states. Sex of the participate, educational status, work status, BMI, place of residence, type of toilet facility and water source are important determinants of water borne disease among elderly in India. Conclusion Elderly people living in the rural areas are more prone to waterborne diseases. The study also finds state wise variation in prevalence of waterborne diseases. The elderly people might not be aware of the hygiene practices which further adhere to the disease risk. Therefore, there is a need to create awareness on basic hygiene among this population for preventing such bacterial diseases.
... Jalan and Ravallion (2003) find that in rural India, children under age 5 living on premises with piped water have a lower prevalence and duration of diarrhea compared to those without piped water. Fan and Mahal (2011), however, argue the opposite view, that piped water in rural India has no effect on acute watery diarrhea. An intriguing phenomenon is both of the Indian studies use propensity score matching (PSM) methods. ...
Article
Full-text available
The research on the protective effect of piped water on young children has been documented in developing countries. However, little is known about the effect of access to piped water (APW) on adolescent health. Based on China Education Panel Survey (CEPS) baseline data for 9,204 adolescents in rural China, we examine the causal effect of APW on adolescent health by employing a control function with ordered probit (CF‐oprobit) model. We find that the availability of piped water in rural households can significantly improve the adolescent health status and also that the effect of piped water on adolescent health is heterogeneous in different subgroups. The protective effect is more pronounced among minorities (rather than Han nationalities), the only‐child families, and left‐behind children during the preschool years. Consequently, piped water programs have irreplaceable significance in improving adolescent health in rural China. Furthermore, policy‐makers should pay more attention to households of ethnic minorities, left‐behind children, and other vulnerable groups in their implementation.
... In this paper, we investigate the impact of access to various drinking water sources and sanitation facilities on incidence of diarrhea among children aged below 5 years in Ethiopia. Similar studies have been conducted in other countries including India (Jalan & Ravallion, 2003;Fan & Mahal, 2011;Kumar & Vollmer, 2013), Bangladesh (Begum, Ahmed, & Sen, 2011), Senegal (Novak, 2014), Philippines (Capuno, Tan, & Fabella, 2015), and Egypt (Roushdy & Sieverding, 2016). While most of these studies focus on the impact of water sources, Begum et al. (2011) and Capuno et al. (2015), like this paper, investigate the impact of both water sources and sanitation facilities on incidence of diarrhea. ...
Article
In this paper, we investigate the impact of access to drinking water sources and sanitation facilities on the incidence of diarrheal diseases among children below 5 years of age in Ethiopia using the propensity score matching technique with a polychotomous treatment variable. We find that among the water sources traditionally considered as improved, only water piped into dwelling, yard or plot leads to a large percentage point reduction in diarrhea incidence. The other water sources, generally believed as clean, are not effective in reducing diarrhea even compared with some of the unimproved water sources. We also find that some unimproved water sources and sanitation facilities are less inferior than they are believed to be. These results suggest that the traditional way of categorizing different types of improved and unimproved water sources and sanitation facilities into a dichotomous variable, “improved” or “unimproved”, could be misleading as it masks the heterogeneous effects of the water sources and the sanitation facilities.
Article
Full-text available
Background Lack of access to and use of water, sanitation and hygiene (WASH) cause 1.6 million deaths every year, of which 1.2 million are due to gastrointestinal illnesses like diarrhoea and acute respiratory infections like pneumonia. Poor WASH access and use also diminish nutrition and educational attainment, and cause danger and stress for vulnerable populations, especially for women and girls. The hardest hit regions are sub-Saharan Africa and South Asia. Sustainable Development Goal (SDG) 6 calls for the end of open defecation, and universal access to safely managed water and sanitation facilities, and basic hand hygiene, by 2030. WASH access and use also underpin progress in other areas such as SDG1 poverty targets, SDG3 health and SDG4 education targets. Meeting the SDG equity agenda to “leave none behind” will require WASH providers prioritise the hardest to reach including those living remotely and people who are disadvantaged. Objectives Decision makers need access to high-quality evidence on what works in WASH promotion in different contexts, and for different groups of people, to reach the most disadvantaged populations and thereby achieve universal targets. The WASH evidence map is envisioned as a tool for commissioners and researchers to identify existing studies to fill synthesis gaps, as well as helping to prioritise new studies where there are gaps in knowledge. It also supports policymakers and practitioners to navigate the evidence base, including presenting critically appraised findings from existing systematic reviews. Methods This evidence map presents impact evaluations and systematic reviews from the WASH sector, organised according to the types of intervention mechanisms, WASH technologies promoted, and outcomes measured. It is based on a framework of intervention mechanisms (e.g., behaviour change triggering or microloans) and outcomes along the causal pathway, specifically behavioural outcomes (e.g., handwashing and food hygiene practices), ill-health outcomes (e.g., diarrhoeal morbidity and mortality), nutrition and socioeconomic outcomes (e.g., school absenteeism and household income). The map also provides filters to examine the evidence for a particular WASH technology (e.g., latrines), place of use (e.g., home, school or health facility), location (e.g., global region, country, rural and urban) and group (e.g., people living with disability). Systematic searches for published and unpublished literature and trial registries were conducted of studies in low- and middle-income countries (LMICs). Searches were conducted in March 2018, and searches for completed trials were done in May 2020. Coding of information for the map was done by two authors working independently. Impact evaluations were critically appraised according to methods of conduct and reporting. Systematic reviews were critically appraised using a new approach to assess theory-based, mixed-methods evidence synthesis. Results There has been an enormous growth in impact evaluations and systematic reviews of WASH interventions since the International Year of Sanitation, 2008. There are now at least 367 completed or ongoing rigorous impact evaluations in LMICs, nearly three-quarters of which have been conducted since 2008, plus 43 systematic reviews. Studies have been done in 83 LMICs, with a high concentration in Bangladesh, India, and Kenya. WASH sector programming has increasingly shifted in focus from what technology to supply (e.g., a handwashing station or child's potty), to the best way in which to do so to promote demand. Research also covers a broader set of intervention mechanisms. For example, there has been increased interest in behaviour change communication using psychosocial “triggering”, such as social marketing and community-led total sanitation. These studies report primarily on behavioural outcomes. With the advent of large-scale funding, in particular by the Bill & Melinda Gates Foundation, there has been a substantial increase in the number of studies on sanitation technologies, particularly latrines. Sustaining behaviour is fundamental for sustaining health and other quality of life improvements. However, few studies have been done of intervention mechanisms for, or measuring outcomes on sustained adoption of latrines to stop open defaecation. There has also been some increase in the number of studies looking at outcomes and interventions that disproportionately affect women and girls, who quite literally carry most of the burden of poor water and sanitation access. However, most studies do not report sex disaggregated outcomes, let alone integrate gender analysis into their framework. Other vulnerable populations are even less addressed; no studies eligible for inclusion in the map were done of interventions targeting, or reporting on outcomes for, people living with disabilities. We were only able to find a single controlled evaluation of WASH interventions in a health care facility, in spite of the importance of WASH in health facilities in global policy debates. The quality of impact evaluations has improved, such as the use of controlled designs as standard, attention to addressing reporting biases, and adequate cluster sample size. However, there remain important concerns about quality of reporting. The quality and usefulness of systematic reviews for policy is also improving, which draw clearer distinctions between intervention mechanisms and synthesise the evidence on outcomes along the causal pathway. Adopting mixed-methods approaches also provides information for programmes on barriers and enablers affecting implementation. Conclusion Ensuring everyone has access to appropriate water, sanitation, and hygiene facilities is one of the most fundamental of challenges for poverty elimination. Researchers and funders need to consider carefully where there is the need for new primary evidence, and new syntheses of that evidence. This study suggests the following priority areas: • Impact evaluations incorporating understudied outcomes, such as sustainability and slippage, of WASH provision in understudied places of use, such as health care facilities, and of interventions targeting, or presenting disaggregated data for, vulnerable populations, particularly over the life-course and for people living with a disability; • Improved reporting in impact evaluations, including presentation of participant flow diagrams; and • Synthesis studies and updates in areas with sufficient existing and planned impact evaluations, such as for diarrhoea mortality, ARIs, WASH in schools and decentralisation. These studies will preferably be conducted as mixed-methods systematic reviews that are able to answer questions about programme targeting, implementation, effectiveness and cost-effectiveness, and compare alternative intervention mechanisms to achieve and sustain outcomes in particular contexts, preferably using network meta-analysis.
Thesis
The separation of humans from fecal waste through sanitation is a crucial element of public health that has prevented countless deaths throughout history. However, health improvements from sanitation are not shared equally across populations. Almost 500,000 children under five die from diarrhea each year, mostly in low-income countries that depend on low-cost sanitation technologies that may not effectively prevent disease. Those diseases have been virtually eliminated in high-income countries through widespread coverage with sewerage and wastewater treatment, but many populations within wealthy countries, including rural communities, racial/ethnic minorities, and other marginalized groups, do not share equitable access to sanitation and experience poor health as a result. Furthermore, sewerage requires copious amounts of water and is not sustainable in an increasingly water-stressed world. One existing solution is the reuse of wastewater for irrigation, but without adequate treatment the practice poses health risks to exposed communities. Achieving global access to sanitation that protects health requires understanding the true health benefits of different sanitation solutions, improved safety and sustainability of waste management practices, and efforts to reach vulnerable populations. In this dissertation, I present three research aims on these topics with the goal of improving our understanding of sanitation and health across national income levels. In Aim 1, we conducted a literature review and meta-analysis of studies on sanitation and diarrhea. Three of four recent major trials on low-cost sanitation interventions found no effect on diarrhea, while historical average estimates have found strong effects. We evaluated literature reviews on sanitation and diarrhea to understand this discordance and found that consensus estimates included numerous flawed studies and inappropriately averaged across widely heterogeneous interventions and contexts. Our meta-analysis highlighted that average effects are largely driven by sewerage and interventions that improved more than sanitation alone. We found that there is no true overall effect of sanitation because variability between interventions and contexts is too complex to average and that the null effects of recent low-cost interventions are not surprising. In Aim 2, we conducted a spatial analysis on households in Central Mexico to understand routes of exposure between wastewater reuse and diarrhea. To test if these exposures have a spatial dependency, we estimated the association between diarrheal disease in children living where wastewater is reused and household proximity to wastewater canals. We constructed a multilevel logistic regression model accounting for spatial autocorrelation and found that children living closer to wastewater canals had substantially higher odds of diarrhea compared to children living farther away. This finding suggests that spatially dependent exposure routes, such as spread by domestic animals or through aerosolization, affect communities that reuse wastewater. In Aim 3, we characterized water and sanitation access among a marginalized population within a high-income country: the Bedouin of the Negev region in Israel. The Bedouin in Israel are formerly nomadic and have faced relocation, demolition, and forced sedentarization since the founding of Israel. Land disputes have resulted in some Bedouin living in historical villages that are not recognized as legal by the government. We conducted a household survey among planned, recognized, and unrecognized Bedouin communities. We found that Bedouin people, especially in unrecognized villages, face limited access to safely managed water and sanitation and have high rates of diarrhea in children. Our study emphasizes shortfalls in global sanitation access and the importance of reaching marginalized communities.
Article
This paper proposes a micro-based approach to investigate the impact of improved water provision on individual health outcomes in rural Uganda. We merge household and individual panel datasets with sub-county level administrative data on water supply projects. Our approach allows us to estimate fixed-effect panel data models which use temporal and spatial variation at the sub-county level as identifying variation. We find evidence that the installation of improved water sources leads to higher reported improved water usage, and shorter water collection times. However, increasing the sub-county rate of improved water sources per capita does not seem to be sufficient to lead to a statistically significant effect in the likelihood of individuals suffering from symptoms of illness associated with inadequate water supply. We argue that our micro-based approach provides a cost-effective means of evaluating development projects. The approach is scalable, i.e. it can be applied to other settings and countries.
Article
Background: Diarrhoea accounts for 1.8 million deaths in children in low- and middle-income countries (LMICs). One of the identified strategies to prevent diarrhoea is hand washing. Objectives: To assess the effects of hand-washing promotion interventions on diarrhoeal episodes in children and adults. Search methods: We searched CENTRAL, MEDLINE, Embase, nine other databases, the World Health Organization (WHO) International Clinical Trial Registry Platform (ICTRP), and metaRegister of Controlled Trials (mRCT) on 8 January 2020, together with reference checking, citation searching and contact with study authors to identify additional studies. Selection criteria: Individually-randomized controlled trials (RCTs) and cluster-RCTs that compared the effects of hand-washing interventions on diarrhoea episodes in children and adults with no intervention. Data collection and analysis: Three review authors independently assessed trial eligibility, extracted data, and assessed risks of bias. We stratified the analyses for child day-care centres or schools, community, and hospital-based settings. Where appropriate, we pooled incidence rate ratios (IRRs) using the generic inverse variance method and a random-effects model with a 95% confidence interval (CI). We used the GRADE approach to assess the certainty of the evidence. Main results: We included 29 RCTs: 13 trials from child day-care centres or schools in mainly high-income countries (54,471 participants), 15 community-based trials in LMICs (29,347 participants), and one hospit