1. SPRING CLEANING NOTES
Trickle Down: Diffusion of Chlorine for Drinking Water Treatment in Kenya
-- This work is joint with Michael Kremer of Harvard, Ted Miguel & Clair Null of U.C. Berkeley, and Alix Zwane of Google.Org.
Michael Kremer, Harvard University and NBER
Edward Miguel, U.C. Berkeley and NBER
Clair Null, U.C. Berkeley
Alix Zwane, google.org

2. The Economics of Rural Water
SPRING CLEANING NOTES
The Economics of Rural Water
Source water improvements vs point-of-use (POU)
Source water improvements serve many households simultaneously, thus require cooperation; POU is private decision by HH
Possibility of recontamination during storage & transport
Child death is one of the great health problems in the world, and water-borne disease is one of the main causes.
[The others are malaria, respiratory infections]
(2) In rural areas of Africa, improved water service almost always means some kind of communal site that is some distance from the home. Piped water to dispersed rural households is usually expensive, infeasible.

3. The Rural Water Project (RWP)
SPRING CLEANING NOTES
The Rural Water Project (RWP)
Randomized evaluation of alternative water interventions in rural western Kenya
Source water quality improvement
Point-of-use water treatment (chlorination)
Increased water quantity
Alternative institutions for community maintenance of water sources
This paper: we study distribution of 6-month supply of free sodium hypochlorite (WaterGuard) to a subset of households in 184 rural Kenyan communities
Spring protection does not create a new water point so the distance that a household must walk to get water is unchanged.
-- In contrast, well construction changes both the quality and quantity of water, so an evaluation of well provision cannot be used to help resolve this debate on whether it is more useful to have higher quality drinking water versus lots of water (for washing and bathing) that could be of lower quality. Distinction: water-borne disease versus water-washed disease

4. SPRING CLEANING NOTES
Project Background
Child mortality in Kenya is high at 120 per 1000 live births (2005), and even higher in rural areas
Diarrheal disease is a leading cause
Lack of knowledge about diarrhea & POU’s doesn’t seem to be a major problem:
72% of study households volunteer that “dirty water” is a cause of diarrhea
87% of study households have previously heard of WaterGuard
But take-up is low: only 3% of study households have chlorine in water prior to intervention
The majority of people in Kenya live in rural areas, without household piped water connections, and they get their drinking water from sources where there is a high risk of contamination because of environmental exposure – meaning basically that fecal matter and associated pathogens can be washed into the water.

5. SPRING CLEANING NOTES
Research Questions
1) What are the impacts of free chlorine distribution on:
-- Home water quality?
-- Child health?
-- Household behaviors?
2) What is the relationship between clean water & diarrhea?
3) How does information about chlorine spread through a community?
-- Is there a “tipping point” for network effects?
-- What sorts of relationships are relevant?
-- What types of people are influential?
4) How does the distribution of free chlorine affect social networks & conversation patterns in the community?
Close ties between this project and Spring Cleaning for cost-benefit analysis of source water quality improvements versus point of use.
First goals similar to spring cleaning – household & child level effects.
Novel aspect of this project is the ability to directly observe how information diffuses throughout the community, from those who got WG to those who didn’t. Will hopefully also be able to distinguish informational network effects versus spreading of the free chlorine by sharing (so far, not quite there).

6. Intervention Baseline survey (Aug 2004 – Feb 2005)
SPRING CLEANING NOTES
Intervention
Baseline survey (Aug 2004 – Feb 2005)
47 of 184 springs protected
Roughly 1300 HH’s in each survey round (7-8 at each spring of 184 springs)
695 HH’s given mL bottles of WaterGuard (approx. 6 month supply); 673 HH’s in comparison group
Two “intensity” levels of WaterGuard intervention:
at 92 springs, 6 of 8 HH in treatment group
at 92 springs, 2 of 8 HH in treatment group
Follow-up survey #1 (Apr – Aug 2005)
Pre-intervention social network data collected
93 of 184 springs protected
Follow-up survey #2 (Aug – Nov 2006)
WaterGuard intervention conducted
Timing:
WG intervention cross-cut with spring protection (springs protected between survey rounds, WG intervention conducted as part of survey round)
Networks data from 2nd survey round
One bottle of WG lasts approx. 1 month, depending on HH size, using only for drinking/kids, etc.; knew there would be pressure to share so gave sufficient to last until return visit, between 2-7 months after intervention
Bottle costs less than ½ day’s wage (about as much as a bottle of soda)
Follow-up survey #3 (Jan – Mar 2007)
Post-intervention social network data collected

7. SPRING CLEANING NOTES
Data
Water Quality
Tested for levels of fecal indicator bacteria E. coli at spring and in home (all 4 survey rounds)
Tested for residual chlorine in home water (last 2 survey rounds)
Household Survey
Water collection (source choice, number of trips, walking distance) and water-related behaviors
Hygiene knowledge, sanitation
Child health (diarrhea), anthropometrics
Household demographic, socioeconomic variables
Social networks data
all pair-wise combinations of study households within spring community
frequency of conversations about children’s health problems, drinking water, & chlorine

8. Take-Up Panel A: Dependent variable,
SPRING CLEANING NOTES
Take-Up
Panel A: Dependent variable,
Water tested positive for chlorine
Treatment
mean (s.d.)
Comparison
T – C (s.e.)
Before WaterGuard distribution
0.03
0.02
0.01
(0.18)
(0.15)
(0.01)
After WaterGuard distribution
0.59
0.07
0.52
(0.49)
(0.25)
(0.02)***
After – Before difference (s.e.)
0.55
0.04
0.51
(0.01)***
% Change in use/contamination
55% 4%
51%
Low baseline usage (randomization successful)
Suggestive evidence of spillover effects from T to C HH’s
Huge increase in use among T group
Still 40% of HH’s that didn’t adopt – why? Doesn’t seem to be for lack of free chlorine; not correlated with # of bottles left from HH accounting or with time between survey rounds. From baseline data, among those who had previously used WG (around 42% of sample), very favorable impressions of product. 94% able to volunteer at least one valid health benefit of WG. Taste often hypothesized as potential impediment to take-up but only 10% said tasted bad with the rest saying tasted good, and “sweetening” often volunteered as reason to use WG.
Usage data corroborate chlorine tests; according to HH accounting, seems that ½ of HH’s were chlorinating consistently & appropriately based on elapsed time & number of bottles used
Not seeing any synergies with hygiene or HH characteristics (mother’s education, # kids)

9. Household Water Quality
70% reduction in contamination (intention to treat effect)
Improvements even for households at springs with low pre-intervention contamination
But not all treatment households had evidence of chlorine in their water
How much did water quality improve among households who actually used the chlorine? (effect of the treatment on the treated)

10. Estimating the ToT
Choice to use free chlorine could be related to other decisions that affect water quality
Need to separate effect of chlorine from effects of other decisions
Can use instrumental variable technique – estimate causal effect of chlorine on water quality by using some source of exogenous variation in chlorine use (not related to other decisions)
Find a variable that is
correlated with chlorine use
but has no effect on water quality other than through its relationship with chlorine use

11. Assignment to Treatment as an Instrument
Probability that a household uses chlorine is affected by assignment to treatment group
But assignment to treatment doesn’t affect water quality other than through its effect on probability that a household uses chlorine (thanks to randomization)
Focus on variation in chlorine use induced by intervention in order to estimate the effect of chlorine on water quality (specifically for those who actually used the chlorine because of the intervention)
Since roughly half of treatment households used chlorine, we would expect water quality improvements for these households to be twice as large as the intention to treat effect
Still don’t know how chlorine would have affected water quality for treated households who didn’t use it

12. SPRING CLEANING NOTES
Child Effects
Diarrhea prevalence of 20% among kids 3 or younger in control households
Pre-intervention difference in diarrhea between treatment & control children of 4 percentage points (22% versus 18%, respectively; significant at 95%)
Treatment associated with ~8 percentage point reduction in diarrhea on average (significant at 95%)
No differential treatment effects for boys versus girls or on the basis of other household characteristics (latrines, hygiene knowledge, mother’s education, etc.)

13. Social Networks 75% of relationships same-tribe Types of relationships
SPRING CLEANING NOTES
Social Networks
75% of relationships same-tribe
Types of relationships
65% of relationships are familial
Non-familial relationships all categorized as neighbors
Frequency of contact: “close” if talk 2-3 times per week or more
60% of relationships are close
14% of pairs are with a household the respondent does not know
1.8 close contacts to treatment households on average
20% of households had no close contacts to treatment
Relatively ethnically homogenous (given recent political developments, important statistic)
Let HH’s volunteer descriptions of their relationships with one another
Common familial relationship types reflect survey protocol (mother of youngest child) and cultural tradition of moving to husband’s village: 20% are mother/daughter-in-law and 25% are wife of brother-in-law

14. 33 had at least one close contact in treatment group
SPRING CLEANING NOTES
Assuming linear effects, each close contact increases probability of take-up by 2%pts (from base of only 3%, so this is big)
Network effects aren’t significantly different for T HH’s (maybe positive reinforcement important for them, too)
Neither 2nd degree (close contacts of my close contacts) nor acquaintences (not shown) seem to matter
Potentially necessary to have multiple contacts
For T HH’s, network effects likely positive reinforcement (they already have the stuff, just need to use it)
For C HH’s, what is the mechanism for network effects?
(4 didn’t have network data)
35 is 81% - in Hm2 only 20% of HH’s report having purchased chlorine, and only 15% of comparison HH’s who didn’t test + for chlorine reported purchasing it
Among 43 comparison households with chlorine in their water at follow-up:
33 had at least one close contact in treatment group
35 reported purchasing chlorine in past six months
14 reported receiving WaterGuard as a gift

15. Take-Up Related to Networks
For each close contact in treatment group, household is 2 percentage points more likely to have chlorine in water
Regardless of the household’s own treatment status
Small effect relative to increase in take-up due to treatment, but huge for control households (50% increase)
Among 43 comparison households with chlorine in water at follow-up:
33 had at least one close contact in treatment group
35 reported purchasing chlorine in past six months
14 reported receiving WaterGuard as a gift
Suggestive of non-linearities (imprecisely estimated)
Community leaders particularly influential (households without latrines particularly non-compelling)

16. Changes in Conversation Patterns
SPRING CLEANING NOTES
Changes in Conversation Patterns
Treatment households are
Roughly 30% more likely to report talking about drinking water
Almost three times as likely to report talking about WaterGuard
If a household’s conversation partner was in treatment group, respondent was
Around 20% more likely to report talking about drinking water
Slightly more than twice as likely to report talking about WaterGuard
Treatment didn’t seem to affect close relationships, though slight evidence that if either of HH’s were treated, they were more likely to list one another as being at least acquaintances at follow-up. Statistically sig at 90% but not economically relevant.
Clearly very effective at prompting conversations about WG specifically and drinking water more generally.

17. Summary Intervention was successful (at least in the short run) at:
increasing water chlorination
reducing water contamination
preventing diarrhea
prompting conversations about WaterGuard & drinking water more generally
Social networks in the community do seem to influence take-up of the product
Possibly non-linear effects (low power to estimate)
Community leaders are key

18. Questions for Future Work
SPRING CLEANING NOTES
Questions for Future Work
Why is take-up so low / high?
Who isn’t using it?
Can we say anything about why they don’t use it? (externalities?)
What is the binding constraint to reducing diarrhea?
Chlorine doesn’t kill everything
Hygiene practices
What will happen in the long(er)-run? Adoption of free chlorine versus adoption of purchased chlorine
Coupon study

19. External Validity
Take-up rates would likely vary according to local perceptions
Water quality effects might be more stable
Scientific, rather than behavioral
Child health depends on many factors, including sanitation
Network effects likely context specific
Finding that community leaders are influential might be generalizable

20. Scaling Up Intervention conducted in order to:
Facilitate cost-benefit comparisons between alternate technologies
Track how information spreads through a community
Not designed with scale in mind
Related project examining potentially scale-able means of encouraging chlorine adoption
Infrastructure
Monitoring

21. Conclusion
Understanding leakage of intervention is explicit goal of study
Still don’t know exact channels for social network effects
Clear example of the differences between the:
intention to treat effect
averaging over all treated households, including both those who did and did not use the chlorine
effect of the treatment on the treated
using assignment to treatment as an instrument for chlorine use
Not always as easy to distinguish those who “take” the treatment from those who don’t
In this case, test for presence of chlorine in the water