We conducted the study in 171 communities eligible to receive the arsenic mitigation program. The mean community size is 153 households. The eligibility criteria were defined with respect to the baseline level of arsenic contamination: (i) communities where more than 25% of sources of drinking water are arsenic contaminated; or (ii) communities where fewer that 25% of sources are arsenic contaminated, but the arsenic-contaminated sources are highly geographically concentrated.
We randomly select 129 communities to receive the program by public lotteries. Treated communities are further randomly assigned to one of three requirements for co-funding the installation costs: no contributions; cash contributions; labor contributions.
We implemented the intervention using a Community Driven Development (CDD) approach. Communities are informed of the terms of the program, and select locations for the subsidized safe sources of drinking water and agree on how to divide the community contributions, where applicable. These decisions are taken by consensus at a community meeting with representation of women and the poor, and in the presence of project staff who act as moderators. Communities must then independently raise or coordinate the community contributions for the wells to be installed. Communities also take responsibility for maintenance, and if necessary repair, of the wells after they are installed.
To causally estimate changes in average household water quality and in behavior with respect to obtaining water for drinking and cooking, we primarily estimate reduced form “intent-to-treat" effects that exploit the random assignment of the program to treatment units.
In order to estimate the causal effects of water source quality and distance to safe drinking water on household water quality, we take two approaches. Our first analysis of mechanisms is a difference-in-difference approach where we evaluate how changes in household bacterial contamination vary with changes in source contamination, transport distance and storage practice. However, although assignment to the safe drinking water program is random, selection of locations for water source installation is endogenous, as it is determined by consensus at a community meeting. As a result, it is possible that changes in distance to water source, or water source contamination, are correlated with changes in other determinants of household drinking water contamination. These confounding factors might in principle bias the above analysis. To address this concern, we carry out a second, instrumental variables analysis.
The instrumental variables approach uses baseline data to predict where in the village a community will decide to install the new water source. We predict the location of the newly constructed water sources using two methods. First, we predict installation sites by visually inspecting the map of existing water sources in the community and selecting location(s) based on population density and existing source quality. Second, we program an algorithm to predict locations in treatment and control villages based on similar criteria. We then use the predicted source location, combined with household-level baseline data, to construct the following instruments: (i) predicted change in source fecal contamination; (ii) predicted change in distance to source of drinking water between baseline and follow-up. We use these predicted changes as instruments for observed changes in source fecal contamination and distance to drinking water.