Improving state effectiveness in environmental risk mitigation: Experimental evidence from Bangladesh

In Bangladesh, the DPHE (Department of Public Health Engineering) is the governmental body in charge of allocating and installing arsenic-safe public wells. Each fiscal year, villagers submit applications to the local DPHE office asking for the installation of DPHE wells in their neighborhood. The DPHE budget for public wells is shared across unions (administrative unit of 13 villages on average). Since applications often exceed the available slots, the Union Parishad chairperson (henceforth "union leader") has to make decisions on allocating wells. The decision making process directly involves the DPHE upazila-level DPHE engineer, who advises the union leaders and oversees the logistics of installing and maintaining the DPHE wells.

Despite their important role in providing arsenic-safe water, DPHE wells have been subject to suboptimal allocation–due to the lack of local knowledge on arsenic contamination status and political pressure. Further, while a large-scale, DPHE-led field testing campaign involving testing of 6.3 million wells allows a detailed understanding of within-union arsenic contamination levels, an evaluation of upazila-level engineers’ understanding of village-level contamination revealed large gaps between the DPHE engineers’ understanding of village-level arsenic contamination and the actual arsenic contamination in these villages.

This strongly suggests that two factors pose important barriers to optimal DPHE well installation that would maximize its public health benefits: first, the information gap between expected arsenic severity and the actual arsenic severity among the responsible agents (union leaders and engineers), and second, whether availability of correct information to responsible agents will lead to actual application in DPHE well allocation.

The intervention aims to study above barriers using a field experiment where we cross-randomize an “information” and an “incentives” treatment. The information intervention provides Union Parishad leaders with a trained-access dashboard displaying sub-union arsenic data, digitizing well applications, and simulating the health benefits of alternative allocations. In the incentive intervention, leaders rank four preferred wells and are randomized to either a bonus-well tournament rewarding arsenic-focused allocations or a truth-telling condition in which one preferred well is installed regardless of its potential impact.

Overall, the two interventions–incentives and information–will be cross-randomized, hence forming four treatment arms. This design will allow us to separately identify both informational and preference-based barriers to arsenic-reducing well allocations. A total of 250 unions will be randomized into the four treatment arms, plus one more arm for “backup” unions, in case of attrition from union leaders. A total of 200 unions will thus be surveyed.

2026-07-21

2026-09-30

Arsenic exposure reduction of well allocation, from elicited allocations

To compute the Arsenic Exposure Reduction, from elicited allocations, we first elicit the preferred allocation from the union leaders. We then will use information on the list of applications for wells in each union to calculate the minimum and maximum number of arsenic-drinking families that could be affected by an allocation. The maximum and minimum impacts are B_{max} and B_{min}, and then an allocation that reaches B_k households is scored as $(B_k - B_{min})/(B_{max} - B_{min})$.

Note that the calculation of arsenic exposure at the union level presumes that the number of wells allocated to each union will remain fixed. I.e. that the union parishad chairmen UNOs and DPHE engineers do not move wells across unions: their allocation decisions are all taken within unions. In all past years, each union has been allocated the same number of wells, and this is a strong equity norm that we do not expect will be violated during our study. In the very rare cases we do observe deviations from such rules, we will assess the robustness of our results to dropping such unions.

-Arsenic understanding (after using the application; the pre-app measurement in the website arms will be used for heterogeneity)
-Self-described motivations for the final allocation
-Average level of arsenic in the mauzas chosen by the union leader
-Average levels of household wealth, education, and other observable characteristics in the mauzas chosen by the union leader
-Average level of population density in the mauzas chosen by the union leader
-Proxies for political gains from, and political pressure for, well allocations
-(Well-level outcomes) Choice of the union leader's own mauza and the expected arsenic contamination levels of a well in the chosen mauzas
-Transformations of the primary outcome: unnormalized benefit, order-weighted score, and top-choice score

-Arsenic understanding. Union leaders' understanding of local arsenic contamination will be measured by asking them to rank three mauzas from their own union (a preloaded high/middle/low contamination triple) by well-water arsenic contamination, and computing Kendall's tau against the true ranking. The mauza pool is restricted to mauzas tested in the ARRP campaign, so the truth is measured rather than imputed. With three items, tau takes four values (−1, −1/3, 1/3, 1); we therefore also report two coarser binaries: an indicator for the ranking being exactly correct, and an indicator for correctly identifying the most contaminated of the three. Understanding is measured once (after elicitation) for union leaders who do not receive the website, and twice for those who do (once before and once after the website is introduced, on the same triple). The post measurement, which exists in all four arms, is the secondary outcome; the pre measurement, which exists only in the website arms, is used as a heterogeneity dimension rather than as a covariate.
-Self-described motivations. Immediately after the preferred allocation is elicited, union leaders are asked an open-ended question on what considerations entered their choices. Responses are coded into a fixed taxonomy: applicant need, water quality, village-level equity, opinions of elected representatives, opinions of unelected elders, own constituency, and other (additional categories may be added depending on responses). Each category yields an indicator; the water-quality indicator is of primary interest, since it tests whether the treatments change the stated objective and not only the chosen allocation. We report the indicators separately and as a standardized index.
-Average level of arsenic in the chosen mauzas. The (population-weighted) average of well-water arsenic contamination rates across the mauzas selected by the union leader for the DPHE well allocation, using ARRP field-kit readings recalibrated against DPHE laboratory analysis and, for untested mauzas, the random-forest downscaling model. This outcome drops the families-served weighting and within-union normalization of the primary outcome, isolating the contamination margin from the size margin. Individual well-level arsenic contamination will also be obtained from the mauza-level rate.
-Average levels of household wealth, education, and other observable characteristics. Population-weighted averages across the chosen mauzas of proxies for household wealth (e.g., household structure type), educational attainment, sectoral employment (agriculture/industry/services), religious composition, and other observable characteristics, from the 2022 Population and Housing Census of the Bangladesh Bureau of Statistics (BBS). These are the principal evidence on the distributional incidence of the treatments: because arsenic contamination and poverty are not strongly correlated, improved arsenic targeting may redirect wells away from poorer mauzas.
-Average level of population density. Population-weighted mean density of the chosen mauzas from the 2022 BBS census. Read alongside the average-arsenic outcome, this decomposes movement in the primary outcome into a contamination component and a population component.
-Proxies for political gains and political pressure. Polling-station-level vote margins in favor of the incumbent BNP (Bangladesh Nationalist Party) and voter turnout in the most recent national election (the first after the July 2024 student uprising), matched to mauzas via polling-station location and averaged across the four chosen mauzas. This tests whether arsenic-based targeting displaces politically based targeting. If polling-station returns cannot be obtained, we will use upazila-level election results as a driver of heterogeneous treatment effects rather than as an outcome, and will state which is being reported.
-(Well-level outcomes) Choice of the union leader's own mauza and the expected arsenic contamination levels of a well in the chosen mauzas. An indicator for whether any of the four named wells is sited in the mauza in which the union leader resides, elicited pre-treatment in the characteristics battery. We also report the analogous indicator for the union leader's childhood mauza, the count of named wells falling in either, and the indicator restricted to the first-named well. Because the leader's own mauza may itself be high-arsenic, the indicator is additionally reported separately for leaders whose own mauza is above versus below the union median contamination, distinguishing targeting from self-dealing that happens to look like targeting. We will also use the average contamination levels of wells in the chosen mauzas to obtain a well-level estimate of expected arsenic levels for each chosen well.
-Transformations of the primary outcome. (a) Unnormalized benefit: the raw number of arsenic-exposed families reached by the four named wells, in levels and in logs. This is the quantity of interest to a public-health planner and the natural scale for national-level extrapolation; it is not the primary outcome only because its across-union variance is dominated by the choice set. (b) Order-weighted score: the primary score recomputed with the first- through fourth-named wells weighted 4, 3, 2, 1, with the attainable maximum and minimum renormalized using the same weights. Because order is payoff-irrelevant under the lottery device, any treatment effect on this score beyond the effect on the unweighted score is informative about how the leader ranks, not only which set he names. (c) Top-choice score: the normalized score of the first-named well alone, which isolates the single choice least contaminated by list-completion behavior.

We randomly assign a primary sample of 200 unions to four treatment groups. This randomization is stratified by the arsenic exposure reduction of past public well allocations and by upazila (details below). We also randomly assign an additional 50 unions to a “backup” pool to serve as replacements for either treatment or control groups if needed.
The 250 total unions belong to 23 upazilas. All data required for stratification come from the ARRP (Arsenic Risk Reduction Program), which tested 6.3 million drinking-water wells across the country for arsenic. The data includes the type of the well (public vs. private), along with the arsenic testing results from field test kits. A random subset of the well water samples was analyzed in DPHE laboratories, the results of which were used to recalibrate the test kit results.
For each union, we compute the arsenic exposure reduction of past well allocations using the same formula we describe in the explanation for its use as a primary outcome.
For upazilas with a sufficient number of unions (i.e., at least 10 unions), we split the upazila’s unions into two strata based on their past well allocation score. For upazilas with fewer than 10 unions, we use the upazila itself as the stratum. This leads to a total of 38 strata, which we use to conduct the stratified randomization into the four primary groups. The 50 backup unions will be kept in a randomly generated order; upon not being able to work with a given primary union, we will replace it with a backup union using this pre-determined sequence.

Not available

Randomization done in office by a computer using R, using set.seed for reproducibility

Union

No

200 unions

200 unions / 800 chosen wells

Control - 50 unions / 200 chosen wells
T1 (incentive, no information) - 50 unions / 200 chosen wells
T2 (no incentive, information) - 50 unions / 200 chosen wells
T3 (incentive, information) - 50 unions / 200 chosen wells

Effect of incentive: 12.63 (unit: arsenic expsoure reduction score, see above for calculation method) Effect of information: 12.61 (unit: arsenic expsoure reduction score, see above for calculation method) Baseline standard deviation of the outcome: 23.61 (unit: arsenic expsoure reduction score, see above for calculation method) The values are obtained using baseline distribution of within-union allocation of DPHE wells in the tested unions.

Pre-analysis Plan for "Improving state effectiveness in environmental risk mitigation: Experimental evidence from Bangladesh"