The Economic Impacts of Rural Water Supply Infrastructures in Developing Countries: Empirical Evidence from Senegal

Abstract

The paper addresses the often-neglected economic impacts associated with the supply of hydraulic infrastructure in rural and under-serviced communities in developing countries. We rely on a rich panel dataset including 1319 Senegalese rural households collected in 2016 and 2020, during the deployment of the first phase of the Emergency Program for Community Development (PUDC). By combining propensity score matching (PSM), inverse probability weighting, difference-in-differences, and quantile regression, we find that access to piped water improves employment in the agricultural sector but has no significant impact on household expenditures. After controlling for attrition, through PSM, we find that the employment effect operates through access to a greater quantity of water and a reduction in the time women devote to water fetching chores. Moreover, when bundled with complementary infrastructure interventions such as the construction of rural roads, we find that access to water services generates an even higher impact. The quantile analysis shows that non-poor households seem to benefit more from the provided water supply infrastructure compared to poor households. Finally, when comparing the welfare effect of government-led PUDC water supply with that of community-led initiatives, our findings advocate for the widespread implementation of the former for reasons of cost-effectiveness.

1 Introduction

Among the 25% of the global population who lack access to water services, most people live the rural areas of Sub-Saharan Africa (SSA) (WHO and UNICEF 2021). Several factors may explain the low rural coverage with water infrastructure in SSA: the overall low performance of the water utilities is one of them (World Bank 2017). Low water prices hamper cost recovery and narrow the revenue basis that would be needed to attract the necessary investments. Even in instances where water is priced at market level, low consumption and the high prevalence of non-revenue water observed in many countries make the cost-of-service delivery quite onerous. As a result, several utilities rely on public funds not only to secure the costs needed for coverage expansion but also to guarantee reasonable operations and maintenance.

The enabling environment in which utilities operate is another contributing factor. The expansion of water infrastructures to remote and sparsely populated communities, which characterizes rural SSA, is hindered by low economies of scale (Abubakar 2019), making the achievement of universal access to sufficient water and sanitation services (Sustainable Development Goal, SDG6) seemingly unattainable, especially in the medium term (UNICEF and WHO 2023).

Governance challenges in the water sector are another contributing factor. Transparent processes would be necessary to secure long-term investments, particularly in the water sector that is characterized by considerable initial sunk costs, which are required to expand the network infrastructure. A recent OECD survey on water governance in African cities offers mixed results concerning perceptions about transparency and integrity in the water sector (OECD 2021).¹ While more than half of the respondent African cities confirm that clear procurement processes are applied when investment decisions are made, only a limited number of the surveyed cities (less than 25%) declare that random auditing, anti-corruption plans, and integrity charters are implemented. Failure to guarantee integrity not only undermines trust but also jeopardizes the quality of service delivery. Similar patterns can also be observed from the Afrobarometer surveys, in which 20% of the respondents from 34 countries declared having obtained water utility services only once some form of a bribe was paid (Howard and Han 2020).

This predicament has severe consequences, as limited access to quality water exposes households to waterborne diseases, often being fatal for vulnerable individuals (Kumar et al. 2022; Adams 2018; Ezeh et al. 2014; Ntouda et al. 2013). In households lacking indoor tap water, members have to dedicate a larger share of their time to fetching water from remote sources (Choudhuri and Desai 2021; Kremer et al. 2011). For example, in rural Senegal, where this study is located, women and children walk distances ranging from 500 m to 1 km on average to access potable drinking water, spending an average of 30 min to one hour on a round-trip to and from a water source (ANSD 2021). Concomitantly, water fetching has a high opportunity cost, since the activity reduces the time available for income-generating activities, employment, and educational pursuits (Adams 2018; Choudhuri and Desai 2021; Ortiz-Correa et al. 2016). Furthermore, it exposes women to sexual harassment, exacerbating the gender inequality gaps encountered in the global South (Dickin et al. 2021; Routray et al. 2017).

Senegal has made quite some progress in providing access to clean water to its populations over the past two decades (ANSD 2021). Nearly 97% of Dakar's population and 82% of the remaining urban dwellers have access to potable water, while access in rural areas is available for 65.7%, resulting in a national access rate of 78.7% (ANSD 2021). Despite noticeable progress – (for more details see Appendix A)—ample disparities exist across regions with more remote areas tending to be most underserved (for details see Fig. 4 in Appendix B). This situation motivated the Senegalese Government to implement the Emergency Program for Community Development (PUDC) including a targeted water intervention in 2016 (see Appendix A for details about PUDC). This study sets out to identify whether the public investments mobilized for the supply of rural populations with quality water services have yielded any economic impact. In addition, we explore whether there are differences in impacts between public and community-led interventions. The study offers a rigorous evaluation of the impact of access to quality water services on key economic outcomes (expenditures and employment, per sector) in Senegal. The economic dynamics resulting from the implementation of the centrally administered hydraulic infrastructure under PUDC are compared with the ones associated with previous alternative water supply initiatives. We term these alternative initiatives: community-led initiatives or non-PUDC water supply programs.² Before PUDC rollout, baseline data were collected in 2016, followed by a 2020 survey, resulting in a panel of 1319 households. Given the non-random implementation of PUDC and the pre-existence of community-led initiatives at baseline, quasi-experimental analysis techniques, such as propensity score matching (PSM) and inverse probability-weighting (IPW) are employed to evaluate the impact of PUDC water infrastructure on employment and household expenditure. Furthermore, based on Smith and Todd (2005), we combine PSM with a difference-in-difference estimator (DiD). Finally, quantile regressions are used to identify heterogeneous effects across the distribution of household expenditure. Both DiD quantile regression and Firpo's (2007) approach are applied to identify the distributional effects of PUDC and non-PUDC water infrastructure. The need to rely on PSM further arises from the high attrition rate we face between baseline and follow up (~ 42%) due to cost constraints. Thus, in interpreting our results, we need to consider that those areas that are more remote and have potentially benefitted the most have not been resurveyed making our impact estimates likely a conservative lower bound.

Our findings reveal that the PUDC water intervention stimulates agricultural employment, while non-PUDC water supply initiatives show no significant impact on both employment and expenditure. The exploration of the underlying mechanisms reveals that households connected to PUDC piped water consume larger amounts of water and women benefitting from the program allocate their saved time to farming and gardening. The PUDC water intervention is a large-scale development project with a huge budget to ensure not only the full operation but also the functional quality of the installed water infrastructure. PUDC hydraulic infrastructure was installed at the request of rural populations, guaranteeing the efficient and productive use of the installed water supply infrastructure through sustained maintenance. The demand-driven nature of PUDC water infrastructure and their maintenance, facilitating the population’s buy-in, may explain why the PUDC water intervention provides better economic opportunities to the users. Participatory and demand-driven water infrastructure has been deemed effective as communities feel responsible for them, ensuring their sustainability (Barde 2017). The quantile regressions further show a more pronounced effect of PUDC water infrastructure when combined with road infrastructure, with non-poor households benefiting more from the provided water.

We make two contributions to the existing literature. First, the study further enhances our understanding of the economic impacts associated with the adoption of piped water in a developing country context. This is important since it provides information on the value-for-money associated with public investments in water supply infrastructure (Trémolet et al. 2015; Prat et al. 2015; Mujica et al. 2015). For Senegal, the current literature has examined the social benefits associated with the expansion of water supply infrastructure (Daffe et al. 2022; Novak 2014; Ntouda et al. 2013), the effect of access to tap water on water consumption (Briand et al. 2010), the various productive uses of water (Hall et al. 2014; Houweling et al. 2012), and the determinants of piped water adoption (Briand et al. 2009). To the best of our knowledge, no study has examined the economic effects associated with access to water services in Senegal so far. Yet, there is growing interest in the role of water in supporting economic transformations and rural–urban disparity in terms of access to utility services. We close this gap by assessing the extent to which access to water influences economic activities within rural settings. We disaggregate the effects along the expenditure distribution to capture potential heterogeneity effects. Understanding such disaggregated effects is important for designing policies that aim at reducing water poverty and inequality in access to quality water. Finally, we determine the effects of water-related services on economic dynamics when combined with other interventions such as access to electricity and rural roads.

Second, the study compares the impacts of government-led (PUDC) versus community-led (non-PUDC) water supply interventions in a developing country context, where a combination of stakeholders operates in the water sector. Whether government-led interventions crowd community-led interventions in or out is an important empirical question that can guide the governance structures adopted in the water sector. Previous studies have explored the factors contributing to the effectiveness of community-led water deployment infrastructure (Ankon et al. 2022; Daniel et al. 2023; Marks and Davis 2012; Roekmi et al. 2018). Others have offered a direct comparison between community and government-led water interventions (Barde 2017; Newman et al. 2002). Barde (2017) and Newman et al. (2002) report that community-led water services enhance both water access and water quality in comparison to centrally supplied large-scale water infrastructure from public entities. With the increasing influx of donor-driven aid to fund small-scale water infrastructure in rural areas of developing countries, there is a revival of the notion that centrally supplied large-scale water infrastructure is ineffective in rural areas due to poor maintenance. The richness of our dataset provides a unique opportunity to contribute new insights to this debate.

The remainder of this article is structured as follows: Sect. 2 offers an exhaustive review of the related literature, Sect. 3 introduces the theoretical framework. Section 4 presents the data and the variables. The methodology is highlighted in Sect. 5, while Sect. 6 discusses the results. Finally, Sect. 7 concludes and offers policy recommendations for the deployment of future water supply initiatives.

2 Related Literature

Given the productivity gains that households connected to safe drinking water derive, greater economic outcomes (incomes, and employment) are expected to follow. Recent studies support this expectation (Winter et al. 2021; Zhou and Turvey 2018). Using cross-sectional data based on 434 rural Zambian households, Winter et al. (2021) found that households connected to a piped system were likely to expand the size of their cultivated areas. The expansion of the cultivated area supports higher crop yield, which leads to higher on-farm employment and enhanced income from the selling of agricultural products. Zhou and Turvey (2018) showed that off-farm employment increased after households gained access to piped water in China. Renwick (2007) concluded that the supply of water services to rural households in Zimbabwe had led to a 27% increase in their income.

Yet, existing findings are not clear-cut. Access to piped water was found to have no significant impact on income, expenditure, and labor supply in several studies (Devoto et al. 2012; Koolwal and van de Walle, 2013; Lokshin and Yemtsov 2003; Rauniyar et al. 2011; Viet and Vu 2013). Using community-level panel data obtained from 106 communities observed between 1998 and 2002, Lokshin and Yemtsov (2003) found that water interventions in rural Georgia had no significant employment impact. Similar results were found for Pakistan by Rauniyar et al. (2011) based on cross-sectional data from 2602 households and using propensity score matching. Viet and Vu (2013) found no significant effects of piped water on employment and income in Vietnam, based on panel data and the combination of propensity score matching (PSM) and difference-in-difference. Using a randomized design (based on a sample of 844 poor urban households), Devoto et al. (2012) identifies no significant improvement in household income following interventions in Morocco. A study carried out by Guzmán et al. (2016), based on panel data from 1200 Ghanaian households combining propensity score matching and analysis of covariance, showed that access to safe drinking water has no significant impact on households' income and assets.

Other studies have looked at social benefits from access to quality water such as health and education. The central message is that water containing pathogenic microorganisms (E. coli) and chemical substances poses a detrimental effect on users, necessitating targeted interventions aimed at improving water quality. Thus, most studies demonstrate that access to piped water (or improved water quality) has positive effects on health and education. Since the focus of this study is on economic effects, we placed the detailed discussion of the social benefits in Appendix C.

Only a few papers employed randomized controlled trials (RCTs) (Prasad et al. 2023), as water intervention, tap water notably, are challenging to randomize (Devoto et al. 2012). The few existing RCTs available mainly focused on water treatment and hygiene (Dupas et al. 2016; Ercumen et al. 2015; Heitzinger et al. 2016; Kremer et al. 2011; Lindquist et al. 2014; Shaheed et al. 2018; Luby et al. 2018; Chard et al. 2019; Quattrochi et al. 2021; Stevenson et al. 2016). The remaining past studies have made use of a variety of quasi-experimental methodological approaches due to the absence of experimental data. These include the logit model and ordinary least squares (Renwick 2007; Fink et al. 2011), fixed effect models (Komarulzaman et al. 2019; Koolwal and van de Walle 2013; Winter et al. 2021; Zhang and Xu 2016), propensity score matching (PSM) (Hasan and Gerber 2016; Jalan and Ravallion 2003; Rauniyar et al. 2011; Guzmán et al. 2016), double difference (DiD) (Chen et al. 2022), the combination of PSM and DiD (Klasen et al. 2012; Lokshin and Yemtsov 2003; Viet and Vu 2013), and the instrumental variable (IV) approach (Klasen et al. 2012; Mangyo 2008; Ortiz-Correa et al. 2016; Zhang and Xu 2016; Zhou and Turvey 2018). Each quasi-experimental approach has its own limitations. With a simple logit model and ordinary least squares, one cannot claim causality. PSM is unable to eliminate selection bias driven by unobserved factors. DiD is unable to control for time-varying heterogeneities. Although the IV approach remains a suitable method to correct for potential endogeneity bias, using poor instruments leads to biased estimates. We conclude from the existing literature that in the absence of exogenous and reliable instruments, the combination of PSM and DiD appears to be the most suitable empirical strategy.

Many of the above cited studies used cross-sectional data due to lack of panel data (Fink et al. 2011; Hasan and Gerber 2016; Jalan and Ravallion 2003; Klasen et al. 2012; Koolwal and van de Walle 2013; Rauniyar et al. 2011; Ortiz-Correa et al. 2016; Trinies et al. 2016; Winter et al. 2021; Zhang and Xu 2016). Yet, with cross-sectional data, time-varying characteristics are impossible to control for as well as dynamics over time and it becomes hard to establish causal relationships. These limitations are circumvented by studies that employ panel dataset consisting of baseline and follow-up data to explore the welfare effects of access to improved water sources (Chen et al. 2022; Devoto et al. 2012; Komarulzaman et al. 2019; Lokshin and Yemtsov 2003; Mangyo 2008; Viet and Vu 2013; Zhang 2012; Zhou and Turvey 2018).

Across the existing literature external validity has hardly been discussed. The country-specific nature of the estimates associated with water quality standards and the processes that govern the distribution of water infrastructure allow for limited generalizations (Wapenaar and Kollamparambil 2019). For instance, the effectiveness of water interventions was shown to be affected by local factors, such as hygiene, maternal education, and health facilities (Jalan and Ravallion 2003; Mangyo 2008; Wapenaar and Kollamparambil 2019; Esrey et al. 1991; Gamper-Rabindran et al. 2010). Similarly, findings for rural areas may not necessarily be applicable for urban zones (Klasen et al. 2012). Statistical reasons may also contribute to limited external validity, especially when RCTs are piloted. Moreover, small samples that are not representative affect generalization (Devoto et al., 2012; Guzmán et al. 2016; Heitzinger et al. 2016; Hasan and Gerber 2016; Stevenson et al. 2016; Quattrochi et al. 2021; Winter et al. 2021; Lokshin and Yemtsov 2003).

A noticeable observation is that no previous studies have examined the economic impact of tap water in Senegal. The existing studies conducted in Senegal looked at (i) the social benefits associated with access to potable water (Daffe et al. 2022; Novak 2014; Ntouda et al. 2013), (ii) the productive use of water (Hall et al. 2014; Houweling et al. 2012), and iii) the effect of access to tap water on water consumption (Briand et al. 2010). Additionally, most of the existing studies report mean impacts rather than distributional effects. Thus, there is a need for an in-depth exploration of the economic implications of piped water adoption in Senegal. Moreover, we unpack heterogeneous effects in relation to income, gender, and educational status.

Furthermore, an important observation from the literature is that studies focusing on the welfare impact of water interventions often overlook the role of suppliers or implementers. The studies rather emphasize the observed improvements in water quality and availability in the users' vicinity. We are not aware of any comparative assessment of the welfare effects of community versus government-led water interventions. Community-led water infrastructure, supplied by NGOs, operates on a limited budget and may charge higher prices if they try to retrieve the investment costs. In contrast, government-led water infrastructure is provided by public water utilities (or their decentralized agencies), which charge lower prices due to economies of scale inherent in large-scale public investment. In turn, the longstanding tradition of a top-down approach to water infrastructure provision, coupled with the failures of centrally planned water interventions, has prompted the exploration of alternatives, such as community-led water infrastructure (Barde 2017). The latter was expected to yield greater effects due to its participatory nature. Local communities are involved in the design, construction, and management of water projects, which fosters sustainability and community empowerment. Thus, it is key to understand which of such approaches is more effective. This is what the study offers by comparing the state-led Emergency Program for Community Development (PUDC) with community-led interventions supplied by private actors and non-governmental organizations (NGOs).

3 Theoretical Framework

It is expected that access to on the premises tap water increases household welfare that mainly works through three channels: (i) time-saving, (ii) greater use of water, and (iii) productivity gains. Households with indoor tap water have their time constraints alleviated as the drudgery of water collection is removed. The extra time gives these households the possibility to take up off-premises work and initiate income-generating activities. The presence of indoor tap water also allows to access a greater volume of water, thereby facilitating its productive uses. Importantly, the health and education benefits stemming from tap water further enhance productivity gains. All these effects –participation in the labor market, development of income-generating activities, productivity gains, and more use of water—translate into more income, expenditure, and increased overall welfare. Figure 5 in Appendix B shows the likely pathways between access to water and economic outcomes.

We further formalize the gains in a model of time allocation, which we draw from the work of Koolwal and van de Walle (2013). The model explores the relationships between access to water, time spent on water collection, and economic dynamics; the model further allows to derive the conditions under which the adoption of indoor tap water results in welfare improvements.

We assume a typical household allocates its available time, normalized to unity, to two types of domestic labor (${t}_{1}; {t}_{2})$, market wage work (${t}_{3}$), and leisure (${{t}_{4}=l}_{4})$. Domestic labor (${t}_{1}; {t}_{2})$ is utilized to produce two types of domestic goods (${g}_{1}; {g}_{2})$; type 1 is related to water collection and is subject to an exogenous productivity gain, represented by a strictly positive parameter $\delta$. Type 2 covers some additional domestic tasks such as firewood collection, tending to children's health and education, cooking, and washing. These chores are primarily performed by rural women in Senegal.

Market wage work (${t}_{3}$) is remunerated at a market wage rate of $w$. The household purchases market-produced goods (${g}_{3}$) with its total income, which comprises wage income ($w{g}_{3}$) generated by labor work (${g}_{3}$) at the market wage rate of $w$, as well as other incomes ($y$). Assuming that the utility function $\left(U(.)\right)$ is concave and increasing in the four arguments, ${g}_{1}$, ${g}_{2}$, ${g}_{3}$ and ${l}_{4}$, with diminishing marginal productivity (${U}{\prime}\left(.\right)>$ 0 and ${U}^{{\prime}{\prime}}(.)<0$), the typical household solves the following maximization problem to derive the Marshallian demand functions for the four goods:

$$Max U\left( {g_{1} , g_{2} , g_{3} , l_{4} } \right) \,{\text{subject }}\,{\text{to}} \,t_{1} + t_{2} + t_{3} + l_{4} = 1$$

(1)

As the focus here is to unpack how changes in the productivity of type 1 domestic labor affect time allocation and consumption patterns, the typical household solves the following equivalent program: $\text{Max U}\left({\text{g}}_{1}, {\text{g}}_{2}, {\text{g}}_{3}, {\text{l}}_{4}\right)$ subject to $(\text{w}/\updelta ){\text{g}}_{1}+{\text{wg}}_{2}+{\text{g}}_{3}+{\text{wt}}_{4}=\text{w}+\text{y}.$ Following Koolwal and van de Walle (2013), we define ${g}_{1}=\delta {t}_{1}$, ${g}_{2}={t}_{2}$ and ${g}_{3}=w{t}_{3}+y$. The market wage rate ($w$) captures the relative price of both type 2 domestic labor and leisure, $w+y$ represents the full income while $w/\delta$ is the relative price of the type 1 good. The Slutsky decomposition can be derived as follows:

$$\frac{\partial {g}_{i}}{\partial \left(w/\delta \right)}={\left(\frac{\partial {g}_{i}}{\partial \left(w/\delta \right)}\right)}_{u=\overline{u} }-{g}_{1}\frac{\partial {g}_{i}}{\partial \left(w+y\right)}\, for\,i=\text{1,2},3.$$

(2)

Thus, the change in the demand for the type 1 domestic good results from the substitution and the income effect. Assuming that the increase in the productivity of the type 1 domestic good affects its demand only through its relative price, and that type 1 domestic and market goods are normal goods, $\left(\partial {g}_{i}/\partial \left(w+y\right)>0\, for\,i=\text{1,2},3\right)$, the consumption of type 1 domestic goods (water collected from outside the premises) increases with its productivity.

Furthermore, if the type 1 good is a complement to both the type 2 domestic and the market good, $\left(\partial {g}_{i}/\partial \left(w/\delta \right)<0\,for\, i=\text{1,2},3\right)$, the increase in the productivity of type 1 domestic goods causes domestic and market goods to increase. Under these two conditions, access to indoor piped water increases the productivity of domestic goods, leading households to participate in the labor market and consume more market goods.

Overall, irrespective of the effect on time allocation, including market work, the increased productivity of domestic products unambiguously improves household well-being as long as the type 1 domestic good is valued. However, the expected welfare effects would be compromised if: (1) the time allocated to type 1 domestic goods does not increase with the productivity of type 1 domestic goods; (2) domestic goods are substitutes for market goods (the substitution effect might be superior to the income effect); (3) changes in productivity ($\delta$) affect both income ($w+y$) and the relative price ($w/\delta$). Thus, the welfare effect of access to piped water remains theoretically ambiguous and asks for an empirical examination. We attempt to offer such an examination for the case of Senegal.

4 Data and Variables

4.1 Data

We use a panel dataset consisting of 1,319 households. The baseline data were gathered in 2016, while the follow-up survey was conducted in 2020. The database provides information at both the household level (household composition, agricultural and non-agricultural activities and income, fixed assets, expenditure, food security, health, housing, and access to drinking water) and the village level (encompassing environment, economy, social life, taxation, and infrastructure). We use the data to evaluate the economic benefits associated with the deployment of water infrastructure within PUDC.

The data collection was undertaken by the Research Centre for Economic and Social Development (CRDES) which employed a two-phase sampling method (CRDES 2016). In the initial phase, the treatment group was established from the list of PUDC-eligible villages provided by the government to CRDES. Potential control villages encompassed all those that are not present in the list of PUDC-eligible villages. CRDES visited both the eligible treatment and potential control villages, collecting data on economic activities, infrastructure, and population size to develop a comprehensive overview of the villages and to ensure similarity between the potential control villages and the eligible treatment villages. The assessment of similarity was based on a simple comparison test. Essentially, the subgroup of potential control villages that was not statistically different from the eligible treatment villages was classified into the control group (CRDES 2016).

In the second phase, ten households were selected per village within the treatment and control group. The selection of these ten households was based on a random walk method (CRDES 2016). Given the dispersed nature of residences in rural Senegal, enumerators were instructed to ensure that selected households fall within the village limits. In cases where the boundaries were ambiguous, guidance was sought from the village chiefs. Data collection was done digitally on the enumerators' phones and tablets with CSPro. The use of electronic questionnaires ensured the quality of the collected data. In each surveyed village, one questionnaire was directed to the chief to assess village characteristics, the ten household questionnaires were administered to the heads of the households. Additional assistance from another two or three household members was sought to enhance the accuracy of the responses.

At baseline, 2,580 households and 270 villages were surveyed. The sample size was determined based on the minimum detectable effect (MDE). The MDE varied between 0.20 and 0.15 for a sample size ranging from 200 to 350 villages (CRDES 2016). At a sample size of 250 villages no large gains in MDE could be attained anymore. Yet, to mitigate attrition effects, the sample size was increased by 8%. As a result, another 20 villages were added resulting in a sample of 270 villages, with 135 control and 135 treatment villages, distributed across 10 regions (i.e., Diourbel, Fatick, Kaffrine, Kaolack, Kedougou, Louga, Matam, Sain-Louis, Tambacounda, Thiès). Figure 6 in Appendix B visualizes the geographical spread of treatment and control villages represented by green and red dots, respectively. Regions surveyed in both rounds are highlighted in bold. At follow-up, only seven regions (Diourbel, Fatick, Kaffrine, Kaolack, Louga, Saint-Louis, and Thiès) were covered due to financial constraints (CRDES 2016). In total, 1,500 households across 163 villages, with 74 target villages and 89 control villages, were revisited during the follow-up survey (CRDES 2021). A graphically representation of the study set up is shown in Fig. 7 (Appendix B).

Undoubtedly, the reduction in sample size from 2,580 to 1,500 households during the follow-up survey may introduce attrition bias since the three dropped regions are more remote compared to the resurveyed regions and likely to be systematically different. We assess attrition formally in Appendix D. We identify significant differences between attrited and non-attrited households that are largely driven by village characteristics. To account for attrition, we resorted to matching techniques to obtain a comparable sample, resulting in the exclusion of 134 observations of which 38 participated in both survey rounds and 96 only in the baseline survey. The careful examination of attrition allowed us to reduce any possible bias to its minimum as shown by Table 9 in Appendix D. Moreover, the analysis supported our field-related knowledge that differences largely stem from remoteness and are introduced by the decision not the resurvey three regions and not by non-responses. We are therefore confident that the threats posed by sample attrition and the resulting estimates are rather lower bounds of the effects (for more details see Appendix D).

Furthermore, we observe that some villages already had access to improved water before the intervention. These are the non-PUDC areas, and the water infrastructure being used is termed community-based water infrastructure as discussed earlier in the paper. This study setup requires a careful cleaning of the database. We dropped the 38 households identified by the attrition analysis as off the common support, resulting in a sample of 1406 households distributed as follows: 949 had access to piped water in 2016 and 2020, 237 had access to piped water only in 2020, 133 did not have access to piped water in 2016 or 2020, 65 had access to tap water in 2016 and no longer had access to it in 2020, and for 22 households information about access to water was missing.

Given the data structure, two treatment groups are defined. The first treatment group (${T}_{1}$) comprises 237 households that had access to tap water only in 2020, thus no access in 2016 (during the baseline). These households sourced their drinking water from the government-led water program under PUDC. The second treatment group (${T}_{2}$) includes 949 households that had access to piped water in both survey rounds from community-led water services. The control group consisted of 133 households that did not have access to piped water in both surveys. The other 87 (65 + 22 households) observations have been dropped, resulting in a final sample size of 1,319 households observed in 2016 and 2020. Qualitative surveys and discussions with Senegalese policymakers and officials in charge of implementing the PUDC program support our stratification into the two treatment groups. Table 10 in Appendix E summarizes the data structure.

The treatment categorization into ${T}_{1}$ and ${T}_{2}$ was determined based on the information provided by the households regarding the type of water source they use. Senegalese households in rural areas get their water from various sources such as tap water on the premises, tap water from neighboring households, communal and public tap water, protected wells, and non-protected wells (Appendix B, Fig. 8). Households with access to tap water are included in the treatment groups, while households consuming water from protected and non-protected wells constitute the control group. The terms "protected" and "unprotected" refer to underground water sources that do not originate from the extension of water supply pipelines. Households relying on other sources, such as water sellers and streams, were not considered in the study due to data limitations.

Being the primary water source for households in Senegal (ANSD 2021), tap water is considered safe for drinking purposes (Jalan and Ravallion 2003); it is potable water. Tap water is drinkable since taps are typically located inside or near the home (within 5 to 10 m), thus minimizing bacterial contamination that may occur during the transportation of water (in open and unprotected buckets) over long distances. Additionally, having a water connection at home ensures access to drinking water in sufficient quantities, enabling rural households to fulfill their hygiene needs (Ahankari et al. 2021). In turn, underground water tends to be contaminated rendering it unsuitable for consumption (Rango et al. 2012). To illustrate, the level of unhealthy substances (salt, iron, and fluoride) in Senegalese underground water is estimated to exceed the acceptable limits set by the World Health Organization (USAID and SWP 2021, p. 5). Hence, we only consider tap water as potable water that is safe for consumption.

4.2 Main Outcomes

The primary outcomes under consideration in this article are the logarithm of household expenditure per capita, expressed in the local currency (Franc CFA), and the proportion of employed household members. Household expenditures represent consumption expenditures over the four months preceding the surveys. Initially recorded on a quarterly basis, we transformed these expenditures into an equivalent annual figure by multiplying them by a factor of three. Subsequently, we calculated expenditures per capita by dividing the total expenditures by the household size and then took the logarithm of the result. The components of household expenditures consist of both food and non-food expenditures. Non-food expenditures include expenses for education, health, clothing, transportation, housing, construction, repairs, celebrations, and miscellaneous items (such as tobacco, alcohol, and gifts). We examine total household expenditure, as well as food and non-food expenditure, expressed in per capita terms.

A scholarly debate exists about the choice between expenditure and income as proxy for measuring household well-being. Deaton (1997) argues that expenditure is a superior indicator of household welfare compared to income, especially in the context of household surveys in developing countries. This preference arises because individuals may not fully disclose all their income during surveys (Meyer and Sullivan 2011). Therefore, income-related information tends to be considered incomplete (Adams 2018; Tirumala and Tiwari 2022). For this reason, we opt for expenditure as a proxy for household well-being.

During the surveys, household members were asked whether they were employed or not. If yes, it was asked in which sector they worked. Responses classify respondents as self-employed, public servants, or private sector employees. Occupational sectors were further classified into the agricultural and non-agricultural sector, the latter encompassing mining, manufacturing, education, health, and services. Based on this information we identify the employment status within the household by calculating the ratio of employed household members to the total household size, representing the share of household members with jobs in the week leading up to the surveys. Importantly, in this context, employment does not exclusively refer to formal job contracts or formal employment, as commonly understood in developed countries. Rather, it captures any form of occupation, whether formal or informal, as most rural workers engage in informal sector activities.

The primary economic outcomes considered are the logarithm of household expenditure per capita (both total and categorized into food and non-food expenditures) and the proportion of employment (comprising total, agricultural, and non-agricultural employment). Descriptive statistics for the main outcomes are presented in Tables 1 and 2 (Panel A) for PUDC-access (${T}_{1})$ and community-led access (${T}_{2})$, respectively, comparing the treatment groups, ${T}_{1}$ and ${T}_{2}$, to the control group. In Column 4 of Table 1, no statistically significant differences in mean outcomes are observed in 2016 for the eventual PUDC beneficiaries compared to the control group. However, in Column 8 of the same table, statistically significant differences emerge in the means of food expenditure per capita, the share of agricultural employment, and the share of total employment in 2020 after PUDC water access has been provided. This suggests that households that received treatment ${T}_{1}$ were initially comparable to the control group but diverged later, showing in the 2020 significant differences.

Table 1 Outcomes and observable characteristics by access to piped water status (T1).

Table 2 Outcomes and observable characteristics by access to piped water status (T2).

Moving to Table 2, Column 4 indicates significant differences in mean expenditure per capita (total, agricultural, and non-agricultural) already in 2016 as these households had already access to improved water in 2016. The significant difference persists in 2020, except for per capita non-food expenditure, which becomes insignificant.

Overall, these descriptives show that households who have access to potable drinking water (either treatment ${T}_{1}$ or ${T}_{2}$) had, on average, higher per capita total and food expenditure in 2020 compared to the control group. The share of agricultural and total employment was greater for households that received treatment ${T}_{1}$ in 2020. Yet, there is no employment difference in 2020 for households under treatment ${T}_{2}$ compared to the control group.

4.3 Observable Characteristics

In addition to the outcomes of interest, the study incorporates observed characteristics at both household and village levels. Household-specific characteristics include household size, the proportion of women, the percentage of children under five years, the age of the household head (HH), the gender of the HH, the literacy status of the HH, and housing conditions, as indicated by the wall material, roof material, and type of toilet. Village-specific characteristics cover the population size, distance to the market, distance to the nearest health center, and distance to the main roads. The rationale behind the selection of these variables is provided in Appendix F—all are motivated by the existing literature.

Panel B of Tables 1 and 2 accounts for observable heterogeneities between treatment groups (${T}_{1}$ and ${T}_{2}$) and the control group. Observed village characteristics show significant differences at baseline and follow-up in most cases. At the household level, households in the treatment group ${T}_{1}$ are statistically similar to households in the control group in both survey rounds, as shown in Columns 4 and 8 of Table 1 (Panel B). There are only two exceptions; treated households were more likely to have a cemented wall in 2016 and a larger size in 2020 compared to households in the control group. Table 2 presents the observed heterogeneities for treatment ${T}_{2}$ showing that at the household level most characteristics are significantly different from the control group at baseline and follow-up.

These observable differences highlight the need for a careful identification strategy that accounts for these pre-treatment heterogeneities between households that use tap water and households that do not. We turn to the empirical identification in the next section.

5 Methodology

5.1 Addressing Selection Bias

We denote ${Y}_{1i}$ the outcome—either employment or expenditure—of household $i$ with tap water (${T}_{i}=1$) and ${Y}_{0i}$ the outcome of household $i$ without it (${T}_{i}=0$). If both states ${Y}_{1i}$ and ${Y}_{0i}$ were known, the effect of access to water would be $\Delta Y=$ ${Y}_{1i}-{Y}_{0i}$. In reality, both states cannot be observed. For the same household $i$ only one state is observable: either ${Y}_{1i}$ for households that use tap water or ${Y}_{0i}$ for households that are not connected to water infrastructure. Therefore considering $\Delta Y$ as the impact of access to tap water on household $i$ is a biased estimate as ${Y}_{0i}$ is a contrefeit conterfactual for ${Y}_{1i}$. The average effect would be $AE=E\left({Y}_{1i}|{T}_{i}=1\right)-E\left({Y}_{0i}|{T}_{i}=0\right)$. By adding and subtracting $E\left({Y}_{0i}|{T}_{i}=1\right)$ to $AE$ – another way of estimating the counterfactual – the average effect becomes $AE=E\left({Y}_{1i}|{T}_{i}=1\right)-E\left({Y}_{0i}|{T}_{i}=0\right)+ E\left({Y}_{0i}|{T}_{i}=1\right)-E\left({Y}_{0i}|{T}_{i}=1\right)$. With such transformation, the average effect (AE) has two components: the average treatment effect (ATE) which is $E\left({Y}_{1i}|{T}_{i}=1\right)-E\left({Y}_{0i}|{T}_{i}=1\right)$ and the selection bias which is $E\left({Y}_{0i}|{T}_{i}=1\right)-E\left({Y}_{0i}|{T}_{i}=0\right)$. The bias stems from the fact that the outcomes of the untreated households are different in both states, implying that $E\left({Y}_{0i}|{T}_{i}=1\right) \ne E\left({Y}_{0i}|{T}_{i}=0\right)$. Thus, the potential sources of selection bias need to be explored to provide an appropriate identification strategy.

The process of program placement and the household decision to connect to tap water determine the extent of selection bias. Utility infrastructure tends to be supplied endogenously in developing countries. Government agencies typically deploy rural infrastructure after careful consideration of intrinsic characteristics in the treatment areas, such as proximity to a city, political ties, and institutional structures (Grogan 2016; Khandker et al. 2014; Ortiz-Correa et al. 2016). Previous evidence has emphasized the non-random placement of hydraulic equipment (Choudhuri and Desai 2021; Gamper-Rabindran et al. 2010; Jalan and Ravallion 2003). The argument often put forward is that policymakers build hydraulic infrastructure in areas already equipped with sanitary facilities (Jalan and Ravallion 2003) and in poor and vulnerable areas, especially when aiming to close inequality gaps (Gamper-Rabindran et al. 2010). Government officials are likely to extend water infrastructure to densely populated areas to guarantee high economies of scale (Barde 2017). Furthermore, remoteness is also a key determinant of water infrastructure placement or rather non-placement. The high costs associated with the extension of water infrastructure to remote areas motivate the preference for areas closer to urban zones. All these characteristics that affect water infrastructure are observable by their very nature and can be controlled. Yet, unobserved characteristics at the village or community level may also influence the construction of water plants. These are often referred to as omitted variables that may simultaneously affect community development outcomes and treatment status.

The decision of households to connect to water infrastructure is endogenous as well. After the provision of hydraulic infrastructure, households decide whether or not to use it based on several variables, either observed or unobserved. Households' financial capacity (income) and educational level are key predictors of connection to water pipelines (Rahut et al. 2015; Adams et al. 2016). As both variables (education and income) may jointly influence the adoption of potable drinking water and households' well-being, they emerge as potential confounders in assessing the true effect of adequate drinking water on welfare outcomes. The size of the household and its composition are also crucial factors for households' access to adequate sources of drinking water (Adams et al. 2016; Arouna and Dabbert 2010; Tshililo et al. 2022). Additionally, the households' ability to assess the advantages of having an improved water source on the premises and their preferences regarding health and time savings, are unobserved confounders (Rahut et al. 2015; Adams et al. 2016; Ortiz-Correa et al. 2016).

To ensure exogenous variability in treatment assignment, both randomized assignment and randomized promotion are often employed (Winter et al. 2021; Devoto et al. 2012). Randomizing infrastructure placement and household access to water would remove selection bias. In the absence of randomization, instrumental variables help correct selection bias. Ortiz-Correa et al. (2016) have instrumented access to potable water with a geophysical variable (measuring water availability) and the number of rivers in the village. Zhou and Turvey (2018) instrumented access to water with exogenous variabilities of irrigated land and hydro-geological traits. In cases where adequate instruments, satisfying the exclusion restriction and relevance assumptions, are not readily available, researchers have employed different empirical strategies to mitigate selection. These empirical strategies include propensity score matching (PSM) (Choudhuri and Desai 2021; Jalan and Ravallion 2003; Manalew and Tennekoon 2019; Novak 2014), the combination of PSM and the difference-in-difference approach (Klasen et al. 2012; Lokshin and Yemtsov 2003; Viet and Vu 2013), and panel data quantile regression (Gamper-Rabindran et al. 2010). This paper addresses potential threats from selection bias by using PSM and DiD. Wee further complement it with quantile regressions.

5.2 Empirical Strategies for Access to Non-PUDC Water Infrastructure (T₂)

The empirical strategies employed in this article depend on the structure of our data and the type of treatment considered. For treatment ${T}_{2}$, we mainly resort to propensity score matching (PSM) to estimate the mean impact of access to tap water. We further use it to account for the high rate of attrition bias (refer to Appendix D). PSM is applied both to baseline and follow-up data. The PSM approach addresses the selection bias by determining a statistical counterfactual based on three steps. First, propensity scores are determined using the Logit model:${\text{P}}_{i}\,(\text{T}=1)= {\beta }_{0}+{\beta }_{1}{\prime}{X}_{i,\text{j}}+{\beta }_{2}{\prime}{V}_{i,j}+{U}_{i,\text{j}}$. The model estimates the probability that a household has access to clean and potable water. This probability is a function of the observable variables at the level of the household ($X$) and the village ($V$), and the non-observable variables measured by the idiosyncratic error term (${U}_{i,\text{j}}$). Second, the predicted probabilities are used to match households from the treatment and control group. In the third and final step, considering that matched household pairs are identical from the point of view of observable variables, the mean impact of the program (${\lambda }_{PSM}$) is calculated by taking the difference between the average outcome of the treatment group and that of the matched control group:

$$\lambda_{PSM} = \frac{1}{N}\left[ {\mathop \sum \limits_{i*\epsilon *T} Y_{i}^{T} - \mathop \sum \limits_{j*\epsilon *C} {\Gamma }\left( {i,j} \right)Y_{J}^{C} } \right],$$

(3)

where $\Gamma (i,j)$ represents the weight used by the Kernel matching algorithm, ${Y}_{i}^{T}$ the outcome of household $i$ of the treatment group, ${Y}_{j}^{C}$ the outcome of household $j$ in the matched control group and $N$ the size of the treatment group.

To test the robustness of the PSM results, the IPW estimator is used. This estimator allocates a different weight to the outcome in the treatment and control group so that both groups have similar observable characteristics. Outcome (${\text{Y}}_{\text{i}}$) is weighted by $1/P(X)$ in the treatment group and by $1/(1-P(X))$ in the control group. The average treatment effect (${ATE}_{IPW}$) is then given by

$${ATE}_{IPW}=\frac{1}{\text{N}}\sum_{\text{i}=1}\left(\frac{{T}_{\text{i}}\times {\text{Y}}_{\text{i}}}{{\text{P}}_{\text{i}}(\text{X}) }-\frac{{(1-\text{T}}_{\text{i}})\times {\text{Y}}_{\text{i}}}{1 - {\text{P}}_{\text{i}}(\text{X})}\right),$$

(4)

where $N$ is the size of the matched control group.

Finally, we use quantile regression to assess the distributional effects of access to water services. Let ${Q}_{{Y}^{T}}^{\tau }$ and ${Q}_{{Y}^{1-T}}^{\tau }$ be the total expenditure per capita for quantile $\tau$ for treated (T) and non-treated households (1-T), respectively. For τ = 0.50, the average effect of this quantile measures the outcome difference of the 50th quantile between households that have access to clean water services and households in the control group. The impact of access to water in quantile $\tau$ is estimated based on the Firpo's (2007) quantile estimator:

$${\Delta }^{\tau }={Q}_{{Y}^{T}}^{\tau }-{Q}_{{Y}^{1-T}}^{\tau }$$

(5)

5.3 Empirical Strategies for Access to PUDC Water Infrastructure (T₁)

For the treatment group ${T}_{1}$, the PSM–DiD approach is used exploiting temporal variation in treatment as the households in treatment group ${T}_{1}$ gained access to clean water services only after the program was implemented (Heckman et al. 1997). The PSM estimator is invalid if treatment is affected by unobserved variables (innate capacities, aptitudes, and preferences). The PSM–DiD approach minimizes bias resulting from time-invariant unobserved traits and simultaneity bias that could emanate from the bi-directional causality between treatment and outcome (Ding et al. 2018). The following Equation is estimated on the matched sample to captures the average treatment effect:

$${Y}_{ijt}={\lambda }_{0}+{\lambda }_{1}t+{\lambda }_{2}{T}_{1ijt}+{\beta }_{1}{\prime}{X}_{i\text{j}}{+\varepsilon }_{ijt},$$

(6)

where ${Y}_{ijt}$ represents the outcome variable, ${X}_{i,\text{j}}$ the vector for observed characteristics, $t$ the binary time variable equal to 1 for the follow-up and 0 for the baseline survey, ${T}_{1ijt}$ the treatment variable (${T}_{1}$) for a household $i$ in village $j$ at time $t$ and ${\varepsilon }_{ijt}$ the idiosyncratic error term. ${\beta }_{1}{\prime}$ and ${\lambda }_{i}$ ($i=1,\dots ,5$) are parameters to be estimated. ${\lambda }_{0}$ measures the average outcome of the control group, ${\lambda }_{1}$ measures the average temporal difference in outcomes between the baseline and follow-up, ${\lambda }_{2}$ is the average effect of access to tap water that we are mainly interested in.

As highlighted above, a factual reality of rural areas of developing countries is that women and children spend a great share of their time in firewood collection and water fetching when they lack both indoor sources of drinking water and electricity. Both domestic activities are onerous for rural households as they take away time from income-generating activities and affect intra-household allocation of resources (Choudhuri and Desai 2021). As electricity and piped water are not substitutes, households still face a time burden if they only access one resource, either piped water or electricity. In that regard, access to electricity is expected to strengthen the expected gains from access to indoor piped water. In the theory of infrastructure-led development, one of Agénor's (2010) perspectives is that limited gains from infrastructure found in developing countries is driven not only by poor quality of infrastructure but also by their incompleteness. An instance of incomplete infrastructure is having good electricity and water infrastructure, but a poor road infrastructure. Given that the components of the infrastructure network complement each other, the joint availability of basic infrastructure is expected to generate bigger gains (Agénor 2010). Therefore, we interact access to road infrastructure with piped water. Access to road infrastructure also facilitates access to the market and eases economic exchange, contributing to enhancing the expected benefits from piped water infrastructure. To account for the interactions of access to a bundle of infrastructure services, we introduce two additional variables in Eq. 6 that capture access to electricity (${E}_{ijt}$) and access to road infrastructure (${D}_{ijt}$). ${E}_{ijt}$ is equal to 1 if the household has access to electricity and 0 if it does not, ${D}_{ijt}$ is the distance between the household and the closest main road. We interact these infrastructure variables with the treatment variable to estimate the average effect of access to tap water in electrified households and the average effect of access to tap water conditional upon the distance between the household and the closest main road.

To estimate the welfare impact of access to indoor tap water across the distribution of total expenditure, we employ a difference-in-differences (DiD) quantile regression ${(QTE}^{DD})$.

6 Main Results

6.1 Determinants of Access to Potable Water

We start the discussion of the results by introducing the determinants of access to potable water. Table 3 presents the outcomes of the Logit models for treatments ${T}_{1}$ (Column 1) and ${T}_{2}$ (Column 2). In line with the earlier discussed attrition, we observe that village characteristics are the key determinants of access to potable water in both PUDC and non-PUDC villages. The coefficient associated with village size is as expected—positive and significant. Larger villages (in terms of population size) have a higher likelihood that households are connected to piped water, confirming the strategic placement of hydraulic infrastructure in sizable communities to ensure favorable returns on investments (Barde 2017).

Table 3 Logit model results.

In contrast, the distance to the market exhibits a negative correlation with access to adequate drinking water, while the distance to the main road shows a positive correlation with access to potable water. The unexpected positive sign for the distance to the main road may be explained by the fact that in the case of the PUDC the water intervention targeted remote areas for placing the water infrastructure to expedite universal access to potable water and contribute to the achievement of SDG6.

In contrast to village characteristics, the relationship between household characteristics and access to potable water varies depending on the treatment. For households under treatment ${T}_{1}$, the age of the household head (HH) is negatively and significantly correlated with access to potable water. This negative coefficient can be explained by our definition of access to potable water, which predominantly involves public taps located outside households’ residences. Households with older heads may not fetch water from a distant place and are thus less likely to access public taps. This finding aligns with Boone et al. (2011), who reported a negative and significant correlation between access to public taps and the age of the HH. For households under treatment ${T}_{2}$, the age of the HH is not found to be significant.

The housing conditions, including roof and wall materials and toilet type, play a crucial role in determining access to potable water. For households under treatment ${T}_{2}$, those using flush toilets are more likely to have access to potable water. This aligns with prior research findings by Adams (2018) and Irianti et al. (2016). Irianti et al. (2016) identified for Indonesian data that households using improved toilets are more likely to have access to improved drinking water compared to those without toilet facilities. Similarly, Adams (2018) reported that households with improved toilet facilities allocate a higher percentage of income to water expenditure. Table 3 further illustrates that access to water is positively correlated with roof material and negatively correlated with wall material, regardless of the water supplier (PUDC or non-PUDC). The unexpected negative correlation with wall material may be attributed to the fact that having a cemented house is not necessarily an indicator of wealth compared to possessing a flush toilet or a roof made of tiles.

For both treatments, household size, composition (share of children and women), gender of the household head (HH), and literacy status of the HH are not found to be significant determinants of access to quality water. These findings indicate that these observable characteristics are well balanced across the treatment and control groups. Similar results have been reported by previous studies. For instance, Abubakar (2019) and Rahut et al. (2015) found no significant correlation between household size and access to piped water or public taps. Adams (2018) and Dhin Etia et al. (2022) even reported a negative correlation. While the need for a greater quantity of water may encourage larger households to adopt drinking water, the financial constraints of larger households may discourage them, as increased household size correlates with higher water expenditure. This balancing act between opposing forces may further explain the insignificant correlation between household size and access to drinking water. Similarly, and in line with Adams (2018) and Tshililo et al. (2022), the education level of the HH is not significantly correlated with access to drinking water. Like our findings, Arouna and Dabbert (2010) and Gebremichael et al. (2021) highlighted that the gender of the HH is not significantly correlated with access to drinking water in rural Ethiopia. The lack of correlation observed in our findings may be explained by the fact that we consider access to tap water in general without distinguishing between private and public taps. For instance, Briand et al. (2009) documented that female-headed households are more likely to obtain water from indoor pipes than from public taps in Dakar (Senegal). Similar results were reported by Briand and Loyal (2017) for Bamako (Mali).

In short, housing conditions, age of the HH, population size, and village remoteness emerge as significant determinants of a households' access to potable water implying that access to tap water is not exogenous, as highlighted in the literature. The observable characteristics identified as correlates of access to tap water may also be directly related to factors that affect the economic outcomes of interest, confounding the true impact of tap water. Therefore, careful analysis is required to reduce bias.

6.2 Distributions of Propensity Scores by Treatment Status

Figures 1 and 2 illustrate the distribution of propensity scores based on treatment status (${T}_{1}$ and ${T}_{2}$) before the matching process. In Fig. 1, the propensity scores of the control group are compared to those of the first treatment group (${T}_{1}$), representing households connected to PUDC water infrastructure. Figure 2 presents the distribution of propensity scores of the control group in contrast to the second treatment group (${T}_{2}$), which comprises households exposed to non-PUDC water infrastructure.

Fig. 1

Source: Authors, 2023

Distribution of propensity scores before matching by potable water adoption status (${T}_{1}$).

Fig. 2

Source: Authors, 2023

Distribution of propensity scores before matching by potable water adoption status (${T}_{2}$).

Upon initial observation, these figures reveal that households in the treatment groups do not share the same likelihood of connecting to tap water as the control group. The treatment groups display a higher probability of being connected to tap water compared to the control group. Put differently, the propensity score of the treatment groups is shifted to the right. While there is a common segment in both Figs. 1 and 2 where the propensity score distributions of the control and treatment groups overlap, it is important to note that units outside this shared segment were excluded from the subsequent analysis as households with propensity scores close to one in the treatment group and those with scores close to zero in the control group are not comparable.

Following matching and the exclusion of units outside the common support (as illustrated in Figs. 9 and 10, Appendix B), the post-matching distributions of the treatment groups now overlap with that of the control group, supporting the effectiveness of the matching process. The success of the matching procedure is contingent upon validating the un-confoundedness assumption, a validation that was conducted through a balance test.

The balance test, conducted via the standardized t-test, calculates the standardized bias for each matching variable. Standardized bias, expressed as a percentage of the square root of the average variance of the treated and non-treated households, gauges the extent to which a matching variable is balanced between the treatment and control groups. Minimal bias in a matching variable implies a balanced distribution between the two groups. Results of the standardized t-test are detailed in Tables 11 and 12 in Appendix E for treatments ${T}_{1}$ and ${T}_{2}$, respectively. These tables reveal a substantial reduction in bias after matching. The only exception is the distance to the main road that remains unbalanced in treatment ${T}_{1}$, while both the distance to the main road (in kilometers) and the proportion of women show imbalance for treatment ${T}_{2}$. However, this slight imbalance is deemed acceptable, as the standardized biases are below 5% (Caliendo and Kopeinig 2008). A visualization of the standardized bias across covariates is depicted in Fig. 11 and 12 of Appendix B (for ${T}_{1}$ and ${T}_{2}$, respectively), further supporting the success of the matching.

Additional diagnostic tests corroborate the bias reduction (Appendix E, Table 13). For treatment ${T}_{1}$, bias decreased from 85.1 (in the unmatched sample) to 23.3 (in the matched sample). Likewise, for treatment ${T}_{2}$, bias dropped from 79.4 to 14.2. The significant reduction in the Pseudo-R², approaching zero, serves as an indicator of the effectiveness of the matching process as well. Thus, we are confident that we can credibly employ the matched sample to assess the impact of the two treatments.

6.3 Impacts of Access to Tap Water on Expenditure and Employment in Rural Households

6.3.1 PUDC Water Supply Infrastructure

We start with the impact estimates using PSM and the inverse-probability weighted (IPW) estimators. Results are presented in Table 14 in Appendix E. Table 14 indicates a positive and statistically significant average treatment effect on the treated (ATT) for both per capita non-food expenditure and the share of agricultural employment. This finding is consistent across PSM and IPW estimates, suggesting that households connected to PUDC-based water infrastructure have experienced an increase in expenditure and agricultural employment. Yet, since PSM and IPW may not be robust to unobservable characteristics, we prefer the PSM–DiD estimates, which are presented in Table 4.

Table 4 Estimated impacts of access to PUDC water (${T}_{1}$) on employment and household expenditure.

Upon initial inspection, the results are similar to the IPW estimates, indicating that access to PUDC water supply infrastructure led beneficiary households to enhance both expenditure and agricultural employment. However, the ATT related to expenditure is no longer statistically significant, while the ATT related to agricultural employment remains significant. Moreover, the magnitude of the ATT linked to agricultural employment has increased in size. These differences reveal potential hidden bias primarily driven by unobserved heterogeneities. Taken together, the findings support the notion that access to PUDC water infrastructure (${T}_{1}$) is associated with increased agricultural employment within households connected to piped water, compared to observationally similar units using water from wells.

The finding that access to potable water increases the proportion of employed household members in the agricultural sector aligns with earlier research by Winter et al. (2021). Access to water services empowers rural households to employ water productively for irrigation, stimulating smallholder farmers to engage in practices that enhance income, such as breeding cattle (Hall et al. 2014; Houweling et al. 2012). Additionally, the availability of on the premises potable water may relax time constraints for women.

To explore the channels of transmission, we examine women's and men’s daily time and monthly water consumption. As this information is available only in the follow-up dataset, we conduct an exploratory analysis with mean comparison tests. The results are detailed in Tables 17, 18, and 21 in Appendix E. Results from Table 17, showing women's time use, indicate significant shifts in their time use. Due to access to water, women’s time spent on water fetching decreases, while the time devoted to leisure, gardening, sleeping, and farming activities significantly increases. Such reallocation of time is not observed among men (Table 18). This is consistent with previous studies indicating that women save time when they no longer have to walk long distances to collect water (Choudhuri and Desai 2021; Rahut et al. 2015; Winter et al. 2021; Guzmán et al. 2016; Hasan and Gerber 2016). This extra time creates opportunities for participation in the labor market and the development of income-generating activities (Houweling et al. 2012; Koolwal and van de Walle, 2013).

An outcome of the exploratory analysis is that treated households consumed a greater quantity of water compared to the control group (Appendix E, Table 21). These results concur with findings from Winter et al. (2021) and Briand et al. (2010) who reported significant increases in water consumption. Winter et al. (2021) showed for Zambia that households that accessed tap water increased their water consumption by 32% using it for domestic and productive purposes. Predicting water demand functions for Senegal, Briand et al. (2010) found that getting a tap connection induced an expected increase in water use of 26 L per capita per day.

Overall, the increased quantity of water and time use for purposes other than fetching water represent the primary channels through which access to PUDC tap water affects agricultural employment. It is noteworthy that findings from the current study are contrary to those of Koolwal and van de Walle (2013), Devoto et al. (2012), and Viet and Vu (2013), who found that access to piped water had no significant economic effects.

Next, we present results from incorporating interaction terms. The corresponding estimates are in Table 5. The positive coefficient associated with the treatment reinforces the finding that access to tap water is linked to increased agricultural employment. Additionally, the table reveals a positive but statistically insignificant coefficient for the interaction between access to water and access to electricity, regardless of the considered outcome. Among households with access to tap water, those with electricity do not exhibit a more pronounced improvement in well-being compared to households without electricity. This may be attributed to electricity primarily serving basic functions such as lighting, phone charging, and powering radio and television in rural Senegal. Cooking with electricity is rare, and electrified households still spend time collecting firewood. However, the noteworthy observation is the negative and significant coefficient of the interaction between access to tap water and distance to the main road. The farther households are from main roads, the less they benefit economically from access to tap water. Consequently, road infrastructure plays a pivotal role in enhancing the well-being of households. Road infrastructure facilitates rural households' access to markets and stimulates economic transactions, thus contributing to the improvement of social welfare in rural areas.

Table 5 Impacts of access to PUDC water (${T}_{1}$) on employment and household expenditure—interactions.

6.3.2 Non-PUDC Water Supply Infrastructure

Next, we examine the economic effects of non-PUDC tap water (${T}_{2}$). Findings are presented in Tables 15 and 16 in Appendix E. Table 15 provides estimates from PSM and IPW for 2016, and Table 16 displays estimates for 2020, using the same estimation strategies. Table 15 consistently indicates a positive and significant average treatment effect on the treated (ATT) related to per capita non-food expenditure for both the PSM and IPW estimator in 2016. While the ATTs related to per capita total expenditure and food expenditure are significant with PSM, they become insignificant when estimated with IPW. From Table 16 we see that in 2020 only the ATT related to food expenditure is significant for both PSM and IPW. Overall, no clear pattern emerges regarding the effect of treatment (${T}_{2}$) on household expenditure over the years. Although there was an increase in household non-food expenditure in 2016 and food expenditure in 2020, such an effect is not observed in total expenditure.

Regarding the employment effect of treatment (${T}_{2})$, Tables 15 and 16 show no significant ATT across all measures of employment for both years, 2016 and 2020. Thus, access to non-PUDC tap water has not led to additional jobs within households. Taken together, we conclude that access to non-PUDC water supply infrastructure has no robust and significant impact on total employment and expenditure.

The ineffectiveness of non-PUDC water supply infrastructure revealed by our findings contrasts with previous works by Bisung and Elliott (2019), who documented that community or non-government-led water interventions are well-being increasing. Similar work reported that non-governmental water interventions lead to increased time and water expenditure savings, which are both key ingredients to well-being (Devoto et al. 2012; Galiani et al. 2009; Winter et al. 2021). We also expected non-PUDC water infrastructure to enhance household well-being, given that community-led water interventions are often designed as an alternative to government-led water interventions. However, to ensure the provision of water infrastructure to rural households, community-led water interventions utilize financial resources from various stakeholders (the community itself, non-governmental organizations, and external private donors) to build the relevant infrastructure. The lack of financial resources for long-term maintenance of community-led infrastructure, coupled with insufficient community ownership and poor financial resource management, is likely to restrict the efficacy of non-state water interventions (Humphrey 2019; Jimenez-Redal et al. 2018; Klug et al. 2017). These factors may well explain the ineffectiveness of non-PUDC water facilities in Senegal. The PUDC water intervention, in contrast, is a large-scale project endowed with a substantial budget, guaranteeing also for the maintenance of the installed infrastructure. Importantly PUDC works with a demand-driven model, installing infrastructure at the request of rural populations, thereby ensuring the efficient and productive use of the installed water supply infrastructure and their maintenance.

The limited effectiveness of non-PUDC water infrastructure is further corroborated by findings from exploratory analyses on tariffs. We compared the average tariff (per cubic meter) across PUDC and non-PUDC water infrastructure and identify a noteworthy difference. The average self-reported tariff paid by households using non-PUDC water is higher than the average self-reported tariff for PUDC water infrastructure, as illustrated in Fig. 13 (Appendix B). The elevated tariffs contribute to a reduced quantity of water consumed. We found no significant difference in the average quantity of water consumed on a monthly basis between households connected to non-PUDC water and households in the control group (Appendix E, Table 21). Consequently, it is not surprising that we find that households with non-PUDC water infrastructure still dedicate a considerable amount of time to water collection (Table 19 and 20).

6.3.3 Distributional Effects of Access to Water Services

Finally, we turn to distributional effects across the income distribution. The results are presented in Table 22 (Appendix E). Panel A shows the estimates of quantile regressions for treatment ${T}_{1}$, while quantile estimates for treatment ${T}_{2}$ are presented in Panels B and C for the years 2016 and 2020, respectively. In Panel A, we demonstrate that access to PUDC-water has no positive and significant effect across the entire distribution of total expenditure. Yet, a noteworthy impact has been identified in the 7th, 8th, and 9th quantiles for food expenditure, and in the 3rd and 4th quantiles for non-food expenditure. Concerning treatment ${T}_{2}$, the average treatment effect on the treated for access to non-PUDC water is significant from the 4th quantile to the 7th quantile for food expenditure in 2020 (Panel C of Table 22). Similar impacts were not recorded in 2016 (Panel B of Table 22).

To gain a comprehensive understanding of the distributional effects of access to water, we categorize households above and below the 50th percentile of household expenditure. The former represents non-poor households, while the latter stands for poor households. Based on this classification, significant increases in household expenditures are observed among non-poor households as indicated by both Panels A and B of Table 22. This group derived greater economic benefits (in terms of increased expenditure among various sub-groups) from piped water adoption compared to poor households, irrespective of the type of water infrastructure—whether PUDC or non-PUDC.

This contrasts with the expectation that poor households would derive the greatest economic benefits from piped water adoption. The findings from our distributional analysis are contrary to Sekhri (2014), who posited that access to water reduces rural poverty. However, our results align with Beyene et al.’s (2018) perspective, suggesting that if access to water promotes poverty reduction, it does so marginally, as access to water infrastructure alone is insufficient to guarantee a substantial drop in the poverty rate. Additional public investment in road infrastructure is necessary, as emphasized by also by our findings. Jalan and Ravallion (2003) underscored the need for investments in oral rehydration, medical treatment, hygiene, and nutrition, which are often lacking among poor households. Higher maternal educational attainment (Mangyo 2008) and improved sanitary conditions (Gamper-Rabindran et al. 2010) are also relevant preconditions for the full manifestation of the well-being effects of access to tap water. Additionally, considering that access to water services is not free, poor households' limited financial capacity may be insufficient to cover water connection costs, jeopardizing the expected economic benefits (Carrard et al. 2019). Higher water bills for low-income households can lead to less disposable income for health-related expenditures, introducing an intra-household consumption trade-off that affects social welfare (Cory and Taylor 2015). In this regard, a pro-poor pricing system within the water sector may be a viable option, contributing to the improvement of welfare for the poor (Ruijs 2009).

It is worth recalling one of the major caveats of the current study which is the high attrition rate faced in the second round of data collection. High attrition rate engenders a significant drop in the sample size, which can reduce the statistical power of the study and make it harder to detect significant effects or differences. If not properly addressed, the non-randomness of such attrition could lead to significant bias, as the remaining participants may differ systematically from those who dropped out, potentially skewing the results and affecting the generalizability of the findings. Fortunately, the differences between the attrited households and the remaining ones have been considerably minimized before further analysis (for more details see Appendix D). As a result, we were able to detect significant economic impacts of tap water adoption. Though we retain confidence in the estimates, caution must be exercised when interpreting our results. Since the program being evaluated was directed toward remote areas, from which households attrited, our impact estimates are likely to be around a conservative lower bound.

7 Conclusion and Policy Recommendations

This study scrutinizes the economic repercussions that access to water services might exert, specifically focusing on household expenditure and employment in rural Senegal. Employing a comprehensive panel dataset comprising 1,319 rural households observed in 2016 and 2020, as part of the Senegalese Emergency Program for Community Development (PUDC), the study distinguishes two water access treatments: access to PUDC tap water versus non-PUDC piped water. The former pertains to government-led water infrastructure, while the latter refers to community-led water services allowing for a unique comparative assessment of both types of infrastructure provision.

We employ a combination of empirical strategies, including PSM, IPW, and PSM-DiD, to ascertain the average impacts of water service accessibility. Additionally, we apply quantile regression in DiD and that developed by Firpo (2007) to capture the impacts across the distribution of household expenditure. The findings reveal that access to non-PUDC water yields negligible economic benefits for households, while PUDC water services exhibit positive and significant effects. Access to PUDC tap water is correlated with increased agricultural employment, facilitated by a higher water consumption rate and reduced time spent on water fetching. The analysis of distributional effects indicates that non-poor households derive more substantial economic benefits from quality water access compared to their impoverished counterparts.

Our findings call for more, large-scale, well-funded government-led initiatives, akin to PUDC, to effectively propagate water-supply infrastructure across rural Senegal. Such initiatives are likely to provide the needed push to the nation's pursuit of universal access to water services in rural communities. In addition, the bundling of diverse interventions has been demonstrated to amplify the positive economic impacts. Therefore, policymakers should concurrently provide rural households with both road and water infrastructure, particularly given the secluded nature of many rural communities in developing countries. Constructing roads is paramount not only to facilitate access to water services but also to enhance economic benefits through improved market accessibility.

Importantly, we identify disparate economic benefits, with poor households benefitting less compared to their non-poor counterparts. This finding suggests that the overarching rural development policy of the Senegalese government to mitigate inequality and poverty through basic infrastructure provision needs to actively implement pro-poor policies. For instance, water infrastructure provision could be coupled with support mechanisms, such as subsidies or a pro-poor pricing structure, enabling impoverished households to bear the costs of water service. Similar initiatives, such as pro-poor pricing systems, have been successfully implemented in South Africa (Ruijs 2009).

We cannot conclude without pointing towards the limitations of the current study. First, our findings can only be generalized to similar context, i.e. a rural African setting. While the external validity is limited, the roll out of the second phase of the program starting in 2021 offers the opportunity for future research to further explore economic effects and corroborate the external validity of the current findings. Second, in the absence of a randomized experiment we had to resort to quasi-experimental methods. While we applied the full toolbox with all relevant tests, we cannot fully rule out unobserved confounders. Future studies may want to refine the identification strategy. Third, we are faced with attrition as three regions could not be re-survey during the follow up. Even though we account for possible impacts and attrition is not the result of non-response but of external, i.e. financial considerations, it has reduced the sample size, which further limits our analysis. Fourth, due to the small sample we cannot explore gender-based impacts or other types of sub-sample analyses. Finally, we have to rely on self-reported data about household consumption expenditure, monthly water expenditure, and daily water consumption and do not have any information about the quality of the water infrastructure. Future studies should aim to address these limitations by incorporating more detailed measurements, including technical information.

Appendices

Appendices

1.1 Appendix A: The Water Sector in Senegal and the First Phase of the PUDC

State supremacy over the supply of water services in Senegal has, for a long time, been a source of inefficiency. Although the country’s hydraulic potential (estimated at 4.750 m³/person/year) remains higher than the reference value for a country to be categorized as physically water scarce (estimated at 1.000 m³/person/year), the total water needs of the populations are not entirely covered by existing supply options (MHA 2016). To improve access to clean water services for rural populations, Senegal initiated reforms during the 1990s. The sanitation sector was separated from the water one, and the latter was subdivided into rural and urban water supply entities. The National Water Company of Senegal (SONES) and the Sénégalaise des Eaux (SDE) (AFD 2006) oversee urban water supply solely. They collect information on present and future urban water demands and design response mechanisms that guarantee the supply of water services to urban populations. For rural water supply, the Office of Rural Drilling (OFOR) was created in 2014 (MEA 2021) to deal with the management of water resources in rural and decentralized zones. This clear separation between urban and rural water management architecture was made to cater to rural landscapes often characterized by a low density of the population, higher levels of poverty, more presence of farming communities, and weaker willingness to pay water bills. These institutional reforms have also led to a change in the water tariff structure and magnitude. Monthly flat rates were replaced with volumetric rates based on actual water consumption. Additionally, initiatives that promoted public–private partnerships (PPPs) were also initiated through contractual arrangements between the State and the private sector. The particularity of these partnerships, to increase efficiency, consists in entrusting the management of public service of water supply to private actors (SDE in urban areas, SEOH, AQUATECH, SOGES, and FLEXEAU in rural areas). The funding of hydraulic equipment remains the responsibility of the delegating authority (the State).

It is within the framework of this institutional system that the hydraulic section of the PUDC was initiated. The PUDC, a vast program for the development and structural transformation of the Senegalese economy, is part of the Plan Senegal Émergent (PSE) that seeks to position Senegal as an emergent country by 2035. Initiated in 2015, the first phase of the program (PUDC I) was operational from 2016 to 2020. Placed at the village level, the PUDC I hydraulic interventions include the renovation and construction of new boreholes and water towers. Altogether, 238 boreholes and 159 water towers were built in the treatment villages. Drilling sites come with a five-thousand-litter storage tank that can meet the needs of more than five hundred people. This hydraulic infrastructure was built to service one or several villages at a time. Given the fact that the PUDC aimed primarily at reducing inequalities and poverty within rural areas, the water supply infrastructure was built in enclaved villages that expressed a strong need for access to water services to reduce poverty. However, once the equipment is installed or placed at the village level, households can decide whether they want to use the water that has been made available or not, provided they are charged according to consumption levels (Fig. 3). The type of water use depends on the needs expressed and the quantity required (domestic use, commercial use, market gardening, or drinking for animals).

Fig. 3

Source: MEA (2021, p. 139)

Average access rates (per cubic meter) per type of water source.

1.2 Appendix B: Additional Figures

See Figs. 4, 5, 6, 7, 8, 9, 10, 11, 12 and 13.

Fig. 4

Source: ANSD (2021)

Rate of access to potable water by region in Senegal.

Fig. 5

Source: Adapted from Winter et al. (2021)

Theory of change of access to adequate source of water.

Fig. 6

Source: CRDES baseline survey, 2023

Treatment and control villages of PUDC I.

Fig. 7

Source: Authors, 2023

Study set up.

Fig. 8

Source: Authors’ computations, 2023

Distribution of households according to the source of water consumed. Note The abbreviations PW and W stand for piped water and wells respectively. Other sources include river water, stream water, and water from hawkers.

Fig. 9

Source: Authors, 2023

Distribution of propensity scores after matching by tap water adoption status (${{\varvec{T}}}_{1}$).

Fig. 10

Source: Authors, 2023

Distribution of propensity scores after matching by tap water adoption status (${T}_{2}$).

Fig. 11

Source: Authors’ computations, 2023

Distribution of the standardized bias across covariates after matching (T₁).

Fig. 12

Source: Authors’ computations, 2023

Distribution of the standardized bias across covariates after matching (T₂).

Fig. 13

Source: Authors’ computations based on the follow-up survey data, 2023

Water tariffs in PUDC and non-PUDC areas. Note The tariffs were calculated based on the monthly quantity of water consumed and the total costs incurred. Households reported both their monthly costs for water consumption and their daily water consumption in liters. We determined the tariffs by calculating the monthly water consumption in cubic meters. This involved multiplying the daily water consumption by 0.03 (30 days/1000 L) and then dividing the monthly costs of water consumption by the monthly amount of water consumed (expressed in FCA per cubic meter). The mean comparison test between PUDC and non-PUDC areas indicates a significant difference, pointing towards higher costs in non-PUDC areas.

1.3 Appendix C: Social Impacts of Access to Water Services

The two major social outcomes associated with access to water are health and education. In this regard, previous studies assess the health-damaging effects arising from the use of water unfit for consumption (Akter 2019; J. M. Brown et al. 2008; Corbo et al. 2008; He et al. 2008; Luby et al. 2015; Yamakawa et al. 1992; Brown et al. 2023; Devoto et al. 2012; Fink et al. 2011; Hasan and Gerber 2016; Jalan and Ravallion 2003; Klasen et al. 2012; Komarulzaman et al. 2019; Koolwal and van de Walle, 2013; Mangyo 2008; Rauniyar et al. 2011; Trinies et al. 2016; Viet and Vu 2013; Wapenaar and Kollamparambil 2019; Rahman et al. 2021; Winter et al. 2021; Zhang 2012). The central message is that water containing pathogenic micro-organisms (E.coli) and chemical substances poses a detrimental effect on users, necessitating targeted interventions aimed at improving water quality. Most studies demonstrate that access to piped water (or other forms of quality water) reduces child stunting and the prevalence of diarrhea among both children and adults. For example, using a simple OLS model and panel data from 4500 rural households (in 152 villages) spanning from 1989 to 2006, Zhang (2012) identifies an 11 percent reduction in the incidence of illness among adults that is associated with access to improved water services. The weight-for-height of adults and children increased by 0.835 kg/m and 0.446 kg/m, respectively, following program implementation that targeted the deployment of water supply infrastructure. Employing propensity score matching, Jalan and Ravallion (2003) observed for the case of India that prevalence and duration of diarrhea spells were lower for households using piped water compared to observationally similar households that were not connected to piped water. Similar results were found by Trinies et al. (2016) in Mali, Winter et al. (2021) in Zambia, and Novak (2014) in Senegal. Another channel through which water scarcity affects human health is the way in which limited availability of water compels households to store this important resource under conditions that do not meet the required sanitary standards. Given the urgent need to secure water for the next day(s), unhealthy storage practices, combined with water containers used for transport and distribution that lack hygiene, result in the contamination and/or reduction of the water quality (Rufener et al. 2010).

Consequently, access to water does not necessarily have positive health impacts. Some studies report no impact on health (Hasan and Gerber 2016; Rauniyar et al. 2011; Viet and Vu 2013), while others condition the health effect of piped water on additional pre-existing requirements (Jalan and Ravallion 2003; Mangyo 2008; Wapenaar and Kollamparambil 2019). Access to water in sufficient quantity is deemed a necessary condition, but itdoes not guarantee improved health, either. For example, Jalan and Ravallion (2003) and Mangyo (2008) have shown that the effect of access to water on children's health is contingent on the income level of the household and the mother’s educational level. Similarly, Wapenaar and Kollamparambil (2019) show that only when the primary caregiver (often the mother) has a minimum level of education exceeding 7 years that children's health improves following access to piped water. At this level of education, mothers are more likely to understand the vital need of investing in oral rehydration, medical treatment, hygiene, and nutrition to deliver enhanced welfare (Jalan and Ravallion 2003). The presence of good-quality health services and latrines within communities also constitutes critical hygienic conditions that are necessary for water services to develop their full potential (Esrey et al. 1991; Gamper-Rabindran et al. 2010).

A second strand of the literature looks at the impact of access to water on education (Akter 2019; Asadullah and Chaudhury 2011; Bouchard et al. 2011; Gopu et al. 2022; Rocha-Amador et al. 2007). Children exposed to polluted water are expected to experience a decline in their cognitive ability, resulting in poor academic performance (Akter 2019; Bouchard et al. 2011; Rocha-Amador et al. 2007). Analyzing a child-level dataset from Bangladesh, Akter (2019) found that excessive water salinity had an adverse impact on the grade advancement of 7–12-year-old children. Drawing on cross-sectional data, Bouchard et al. (2011) report that the use of water contaminated by manganese reduced the intelligence quotient (IQ) of 6–13-year-olds. Similarly, Rocha-Amador et al. (2007) found that exposure to fluoride through drinking water impaired the IQ of 6-to-10-year-old school children. The identified underlying mechanisms are: early childhood exposure to unsafe water (Akter 2019) and the ostracization of children who become unhealthy due to the intake of unsafe water (Asadullah and Chaudhury 2011). Based on cross-sectional data, Ortiz-Correa et al. (2016) concluded that children living within households connected to a piped water system gained 0.7 years of schooling in Brazil. Employing a panel dataset of 259 Indonesian districts from 1994 to 2014 and both static and dynamic fixed-effect models, Komarulzaman et al. (2019) found that districts with a greater share of households having access to on-premises water facilities displayed higher school enrolment and lower school absenteeism. Choudhuri and Desai (2021) and Koolwal and van de Walle (2013) documented similar findings for India and across developing countries. Drawing on longitudinal data from the China Health and Nutrition Survey (CHNS) spanning from 1989 to 2011, Zhang and Xu (2016) found that early exposure to tap water increased the completed grade of education among youth by 1.1 years. Utilizing a DiD approach, Chen et al. (2022) similarly concluded that an extra year of exposure to piped water in early life increased the cognitive ability of 10-to-15-year-old children by 0.132. Other studies further support positive impacts of access to water on human capital (Chen et al. 2022; Choudhuri and Desai 2021; Komarulzaman et al. 2019; Koolwal and van de Walle, 2013; Ortiz-Correa et al. 2016; Zhang and Xu 2016).

1.4 Appendix D: Attrition Analysis

Attrition is a non-negligible concern for the study at hand. We face a reduction in sample size from 2580 to 1500 households during the follow-up surveys. This is mainly driven by the fact that three regions are not re-surveyed in the second round (Kédougou, Matam, Tambacounda). These three regions differed systematically from those in the seven regions covered in both survey rounds (Diourbel, Fatick, Kaffrine, Kaolack, Louga, Saint-Louis, and Thiès). Importantly, the three regions were not re-surveyed due to financial constraints. Thus, while it is reducing external validity, the main source of attrition is introduced by cost considerations and not non-compliance of initial survey respondents. Yet, the observed attrition can impact the estimates and, consequently, the analysis conducted if households systematically differ.

To assess attrition, we start by comparing the characteristics of households lost in the follow-up survey with those that participated in both rounds of data collection (Table 

Table 6 Mean comparison test between households that participated in both round and those lost in follow-up at the baseline.

6). We find significant differences between households which participated in both rounds of data collection and those who participated only in the baseline survey as shown in Column 3 of Table 6. As already discussed, the large attrition rate is not random but a result of financial constraints. In a logit regression we further identify the determinants of attrition (Table 

Table 7 Determinants of attrition in the second round of survey.

7). Unsurprisingly and in line with the systematic lack of follow-up data for three regions, village characteristics are the main determinants of attrition. The village population, distance to the market and main road are the key drivers of attrition. The further villages are located from the main road, the lower their probability of being selected for the second round of the survey. Thus, the data confirm that the follow-up survey has targeted villages that are closer to cities (and the capital cities) to minimize survey-related expenditures.

Therefore, we resort to matching techniques to balance both groups of households. After the matching, 134 households—of which 38 participated in both rounds—that are off the common support were dropped. The results of the balance tests after matching are presented in Table 

Table 8 Balance test on observable variables – Attrition (Yes = 1).

8 and indicate that most observed characteristics have been balanced, except for the household size, the share of women, the sex of the household head, the logarithm of village population and the distance to the nearest health center. Nonetheless, the balance tests further demonstrate a significant and sizable reduction of the attrition bias (Table 

Table 9 Diagnostic tests after matching households that attrited with those that did not.

9).

Based on the results from the attrition examination and our knowledge from the field about the three regions that were not re-surveyed, we are confident that the sample attrition does not inflate our estimates but rather attenuate them and thus, they represent a conservative lower bound of the impact.

1.5 Appendix E: Additional Tables

See Tables 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21 and 22.

Table 10 The data structure.

Table 11 Balance test on observable variables – Treatment ${T}_{1}$.

Table 12 Balance test on observable variables – Treatment ${T}_{2}$.

Table 13 Diagnostic tests after matching.

Table 14 Impact estimates of access to drinking water (${T}_{1}$) on employment and expenditure – PSM and IPW estimates.

Table 15 Impact estimates of access to drinking water (${T}_{2}$) on employment and expenditure (round 2016) – PSM and IPW estimates.

Table 16 Impact estimates of access to drinking water (${T}_{2}$) on employment and expenditure (round 2020)—PSM and IPW estimates.

Table 17 Women’s time use (T₁).

Table 18 Men’s time use by treatment (T₁).

Table 19 Women’s time use by treatment (T₂).

Table 20 Men’s time use by treatment (T₂).

Table 21 Water consumption per treatment (T₁ and T₂).

Table 22 Results from quantile regressions.

1.6 Appendix F: Selection of Independent Variables for the Treatment Equations

This section addresses the selection and justification of independent variables used in the selection equation. In other words, it addresses the selection of potential determinants of access to a quality water source. These variables consist of a selection of both village and household characteristics. Building on the discussion of the non-random placement of hydraulic infrastructure in sub-Sect. 6.1, we control for village characteristics such as population size and remoteness measures, including the distance of a village to the main road, the distance to the market, and the distance to the nearest health center. We anticipate a positive correlation between village population density and access to potable drinking water. The rationale is that water infrastructure is constructed in areas with greater population density due to the larger economies of scale they offer, ensuring the recovery of initial investment costs (Barde 2017). Moreover, in larger populations, the queuing time for water collection may be longer. Also, households in densely populated areas are likely to adopt piped water on the premises (Arouna and Dabbert 2010). In turn, remote areas often receive less infrastructure due to the prohibitive costs of the network expansion, leading to less adequate water facilities (Barde 2017). Consequently, we expect distance to the main road, the market, and the nearest health center to be negatively correlated with access to improved sources of drinking water.

Once villages are connected to water infrastructure, households decide whether to use it or not based on their characteristics. We control for the following observable household characteristics: the size of the household, the composition of the household (the gender of the household head, the share of women, and the share of children under five years of age), the age of the household head (HH), the education level of the HH, the occupation of the HH, and the household’s housing conditions (type of toilet, wall material, and roof material). The selection of household characteristics is informed by the existing literature on the determinants of access to safe and appropriate drinking water in rural zones of developing countries. We detail our choices in the following paragraphs.

The size of the household is a crucial factor influencing households' decisions to connect to water through a piped system. Larger households are more likely to install piped water on the premises due to their increased need for a greater quantity of water and the challenges associated with collecting water from distant sources to meet their demands (Adams et al. 2016; Arouna and Dabbert 2010; Briand et al. 2009; Rahut et al. 2015; Tshililo et al. 2022). Therefore, we anticipate a positive association between household size and the use of adequate drinking water (conditional on income, see below).

Household composition plays a significant role in water connectivity (Arouna and Dabbert 2010). As women and children are the primary users of potable water for their hygiene, the absence of adequate drinking water exposes them to water-related diseases (Bisung and Dickin 2019). The responsibility of fetching water also falls mainly on women, and the lack of an indoor water source diverts their time away from active labor market participation (Koolwal and van de Walle, 2013). Consequently, we anticipate that households with a higher share of women and children are more likely to connect to an improved water system (conditional on income, see below). We use the gender of the household head, the share of women, and the share of children to capture household composition.

Age is also a predictor of a household's connection to a piped water system (Abubakar 2019; Rahut et al. 2015). As household members age, they may lack the physical strength needed to collect water from distant sources, making them more likely to install water taps on the premises. Households with older household heads also tend to be better off. Therefore, we expect a positive correlation between the age of the household head and access to improved drinking water sources.

The education level of household members is another crucial parameter influencing a household's decision to install piped water on the premises, for at least two reasons. First, as the opportunity cost of water collection increases with alternative uses of time, educated household members would have a higher opportunity cost compared to the uneducated ones, preferring piped water on the premises (Rahut et al. 2015). Second, as the awareness of health benefits stemming from potable water increases with education, literate household members are more likely to make an informed decision to have piped water on the premises (Arouna and Dabbert 2010). We measure the household's education level by the literacy of the household head, i.e. whether s/he is literate or not. Our hypothesis is that literate household heads are more likely to install piped water on the premises and rely less on water obtained from wells.

Empirical findings suggest that the occupation of the household head is significantly related to the quantity of water used in the household. For instance, Abubakar (2019) reported that households with heads employed in the agricultural sector use less water compared to those with off-farm activities in Nigeria. We expect a similar finding in this study.

Finally, one of the crucial determinants of the use of piped water in developing countries is a household's income level. Wealthier households are more likely to bear the costs of connection to the water system, while poorer households are less likely to adopt tap water due to income constraints (Rahut et al. 2015; Carrard et al. 2019). We proxy the income level of households by their housing conditions. The argument is that wealthier households have better housing conditions compared to poorer households. To the extent that housing conditions are determined by the income level of households, they hold the potential to predict the household's water system connection status. Households with good housing conditions usually have flush toilets, requiring a greater quantity of water. We measure housing conditions by the type of toilet used (flush or pit), wall, and roof material.