Factors Influencing Water Point Functionality in Liberia: A Regression and Bayesian Network Analysis

Abstract

Maintaining functional rural community water supply is a persistent challenge across Sub-Saharan Africa, particularly in Liberia. This study examined the determinants of hand pump functionality in Liberia using a comprehensive dataset from the Liberian Government. We analyzed 11,065 Afridev hand pumps using regression and Bayesian network models. Water points managed by local and institutional entities had substantially higher odds of being functional than those with no management (adjusted OR 3.73 and 2.89), while WASH committees showed a smaller increase (OR 2.43). Pump part damage significantly reduced functionality (undamaged vs. damaged, OR: 10.46. Faster repair was an important determinant, with odds of functionality up to 6.37 times higher. The availability of a trained mechanic with a modest toolkit modestly improved odds (OR 1.25), and proximity to spare parts suppliers played a role (second quartile vs. farthest quartile, OR 1.57). We quantified the impact of service delivery: posterior odds that a water point is functional under the most ideal conditions were four times those under the least ideal conditions. These findings underscore the importance of effective management and prompt repairs to maintain the functionality of water infrastructure. These insights can guide improvements in Liberia and other regions facing similar challenges.

1. Introduction

Access to safe, reliable drinking water is a human right [1]. It is important for human health and well-being, and recognized in the Sustainable Development Goals (SDGs). Goal 6 of the United Nations Sustainable Development Goals (SDGs) aims to ensure the availability and sustainable management of water and sanitation for all [2]. Estimates suggest that 56% of the world population, approximately 4.4 billion people, lack safely managed drinking water [2,3]. Sub-Saharan African (SSA) countries face considerable challenges in providing populations with access to water services. Additionally, many countries experience substantial disparities in access to water services, with rural communities, low-income households, ethnic minorities, and people with disabilities often facing significantly lower levels of access compared to their urban, wealthier, and more socially advantaged counterparts [4,5]. Liberia is one such country with low access to water facilities. Approximately 25% of Liberians lack access to basic drinking water. Diarrhea remains the second leading cause of death in the country [6,7]. This situation forces communities to rely on distant, unreliable, or contaminated water sources, posing severe health risks and undermining the benefits of having a safe water supply [8].

Boreholes with hand pumps are a primary source of clean water in Liberia and throughout the subcontinent; estimates suggest that approximately 39% of the Liberian population relies on a hand pump for their primary source of drinking water [9]. However, these hand pumps are frequently plagued by mechanical failures and performance issues, compromising the water supply due to poor functionality. In SSA, one in four hand pumps is non-operational at any time [10]. Challenges in ensuring reliable functionality may also be exacerbated due to the increasing impact of the climate on precipitation patterns and aquifers [11]. Therefore, there is a pressing need to evaluate and address the critical issue of water sustainability and accessibility in Liberia and throughout Sub-Saharan Africa, including the performance of hand pumps, users of which may not have a safe alternative drinking water source.

Addressing water service delivery is a complex systems problem where the interplay of hardware (e.g., infrastructure) and software (e.g., governance and management) factors influences water service outcomes. Previous studies suggest that community-based management, higher monthly fees, monitoring and evaluation systems, and distance from the capital city are related to hand pump performance [10,12,13,14,15,16]. Large-scale analyses have been conducted in countries such as Malawi, Ghana, Sierra Leone, and Uganda, using both regression and Bayesian approaches to identify key predictors of functionality [12,17,18]. Our study extends this body of work by applying both multivariable logistic regression and Bayesian network analysis to a large, nationally representative dataset of over 11,000 Afridev hand pumps in Liberia. Unlike earlier studies, we integrate both modeling techniques to assess management and repair dynamics at scale, and we simulate how combinations of service delivery factors influence functionality outcomes. This approach helps clarify the mechanisms by which governance and operational capacity affect the sustainability of rural water services.

In this study, our goal is to assess the key factors influencing the performance of hand pumps in Liberia using a large dataset published by the Liberian Government’s WASH project. The findings could help pinpoint areas for improving water point conditions and support the long-term sustainability of Liberia’s water supply system.

2. Materials and Methods

2.1. Data Source

Data used in this study were obtained from Liberia’s WASH initiative website (see the Data Availability Statement). The WASH initiative was a joint effort by the Liberian Ministry of Public Works and its partners, including the World Bank, UNICEF, USAID, and Liberian water points map and surveying program Akvo, local districts, and clans [19]. The total number of water points mapped was 20,205, distributed across 15 counties. The survey included questions that assessed water point state and type, geographical information, and community resource availability [19].

Data were collected by the Water Point Mapping Technical Committee (WPMTC), which comprises two sub-committees: a field data collection sub-team and a data processing sub-team. In total, 16,286 enumerators participated in the data collection [19]. The enumerators were trained in data collection and processing to ensure data integrity and consistent methodology. Data were collected and aggregated using “Akvo Flow”—a proprietary cell phone app developed by the Akvo Foundation [19]. The survey data, published by the Liberia Ministry of Public Works, were downloaded from https://wash-liberia.org in an XLSX file format and subsequently imported into statistical software SAS Enterprise Guide (Version 8.5) for cleaning and analysis.

2.2. Data Processing and Cleaning

During data cleaning, best practices for data cleaning and processing were used [20]. Missing data induced by survey questions involving “skip-logic” were processed after being combined into a single variable (Table 1). For example, the survey question “Were trained mechanics provided with toolkits?” was skipped if the question “Is there a trained mechanic available at this point?” was answered “No”. The toolkit likely refers to a basic set of tools needed for routine maintenance of Afridev hand pumps, which are designed to be easily repaired with minimal equipment. Responses to the survey question about the availability of a trained mechanic and, if applicable, that mechanic’s access to a toolkit were combined into a single variable with multiple levels. Data were, therefore, only considered missing if the response to the higher-level question was missing.

Table 1. List of survey questions and their derived variables selected for modeling in this study, with different types, classifications, and response.

Type of Variable	Survey Question	Survey Response	Variable Name	Variable Levels or Units
Outcome	Water point functionality	Yes—Functional (and in use) Yes—Functional (but not in use) No—Broken down	Functionality	Yes  No
Predictor	Who is maintaining the water point (routine repairs)?	WASH management committee Other community group Private owner School NGO Other	Maintainer	No Management Local Institutional WASH committee
Predictor	Is there a trained mechanic available at this point?	Yes No	Mechanic and toolkit availability	No Mechanic Mechanic with Toolkit without Toolkit
	Were trained mechanics provided with toolkits?	Yes No
Predictor	Is the water paid for at this point?	Yes, by volume Yes, flat rate Only after breakdown No, it’s free	Water point payment status	Per jerry can Per liter Only after a breakdown No, it’s free
Predictor	Last time the water point broke down, how long did it take to repair?	Never Broken More than a year Over a month Less than a week	Time it took to repair	Never Broken More than a year Over a month Less than a week
Predictor	Spare parts supplier distance (minutes)	Continuous (min)	Spare parts supplier distance	Quartile 1 (0–10 min) Quartile 2 (11–42.57 min) Quartile 3 (42.57–60 min) Quartile 4 (60–148 min)
Predictor	Is water available throughout the year?	Always water Seasonal Never water/dry	Water availability	Always water  Seasonal
Mediator	Is the water point damaged?	Yes No	damage status	Yes  No

Notes: For the Maintainer variable: Local includes NGO, community, local group, and private owner. Institutional includes school, health facilities, and other management. For the Mechanic and toolkit availability variable: About 900 entries (7%) responses were imputed from Unknown. For the water point payment status variable: 30% (2539) missing entries were imputed.For the continuous variable Spare parts supplier distance variable: Entries greater than 2 Median Absolute Deviations (MAD) were removed For the water availability variable: Entries with Never water/dry (<1%) were excluded.

2.2.1. Inclusion and Exclusion Criteria

Data points from unimproved water sources (unprotected dug wells, unprotected springs, and unequipped boreholes) where hand pumps are infrequently installed were excluded from the analysis. Additionally, water points with water as “never available” were also excluded (<1%). Only those with hand pumps were considered among the improved water points. Specifically, the data were restricted to “Afridev” branded hand pumps, as they are the most common pump type in Liberia and account for 90% of all surveyed hand pumps (Table 2).

Table 2. Pump brands of Liberian water points observed among hand pump-equipped improved water points (n = 13,239). Frequencies and percentages are calculated among hand pumps; Afridev (11,796; 89.1%) was the dominant brand and was used for analysis.

Pump Brand	Frequency	Percent (%)
Afridev	11,796	89.10
Other	467	3.53
Consallen	322	2.43
India Mark	273	2.06
Kardia	183	1.38
Rope pump	143	1.08
Vergnet	55	0.42
Total	13,239	100.00

2.2.2. Missing Variables

Several variables in this study contained missing values, which we addressed using hot deck single imputation with the Approximated Bayesian Bootstrap (ABB) selection method. We assumed the missing data followed a Missing Completely at Random (MCAR) mechanism. This assumption was based on exploratory data analysis, including visual inspection of missingness patterns and comparisons of group means between missing and non-missing observations using statistical software. The ABB method was chosen because it handles complex survey designs, including non-probabilistic censuses, such as the water point mapping exercise used to collect the data in this study. It also provided more accurate imputation by incorporating uncertainty and maintaining the overall data distribution, thereby leading to more reliable and valid statistical inferences [21]. The default parameters for the donor in this hot deck imputation are simple random sampling with replacement (SRSWR), and there is one donor per recipient of missing values.

2.3. Exploratory Data Analysis

Geographic Distribution of Water Points

The geospatial distribution of water points included in the dataset is visualized in Figure 1, which shows clustering near population centers, particularly along the coast and in central counties. Table 3 summarizes the total number of mapped water points by county. The highest counts were found in Montserrado (6204), Nimba (2166), and Margibi (1601) counties. In contrast, River Gee (236), River Cess (364), and Grand Kru (414) had the fewest. This distribution likely reflects underlying differences in population density, infrastructure development, and possibly the completeness of reporting. These spatial patterns were important to account for in modeling water point functionality.

Figure 1. The distribution of water points reported in the dataset by county throughout Liberia in a study on water point functionality.

Table 3. Water points by county in Liberia (n = 20,205). Percentages are calculated as county count divided by 20,205.

County	Water Points	Percent (%)
Montserrado	6204	21.9
Nimba	2166	7.6
Margibi	1601	5.7
Lofa	1495	5.3
Bomi	1454	5.1
Bong	1402	4.9
Grand Bassa	1156	4.1
Grand Cape Mount	911	3.2
Grand Gedeh	909	3.2
Maryland	638	2.3
Gbarpolu	634	2.2
Sinoe	621	2.2
Grand Kru	414	1.5
River Cess	364	1.3
River Gee	236	0.8
Total	20,205	100.0

2.4. Statistical Analysis

2.4.1. Logistic Regression Model

A multivariable and univariable logistic regression model, constructed using SAS Enterprise Guide (Version 8.5), was employed to investigate the variables affecting the functionality of the water points reported in this study. The multivariable logistic regression model is referred to as the adjusted model, while the univariable logistic regression model is referred to as the unadjusted model. In subsequent analyses, the results from the multivariable regression model are used because it allows for the control of potential confounding variables and provides more accurate estimates of the associations between predictors and outcomes [22]. A literature review guided the selection of variables included in this regression model [12,23,24].

Additional model selection techniques used included the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). Model diagnostics and fit assessment was performed by determining the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve and calibration plot, which compared predicted probabilities with observed proportions across deciles to assess how well the model’s predicted risks matched actual outcomes.

2.4.2. Bayesian Network Model

A Bayesian network (BN) model (see Appendix A) was also used to predict outcome variables using discrete predictor variables. In a BN, each node represents a variable. The directed edges between nodes represent hypothesized causal relationships. The strength and nature of these relationships are quantified using conditional probability distributions (CPDs). BN models have demonstrated superior performance in modeling complex environmental systems, incorporating hypothesized causal links in addition to the correlation data reported by logistic models.

To build the BN model, we used Netica (Version 6.06), a proprietary software. Continuous variables were converted into quartiles. The data were randomly divided into 80% training and 20% testing datasets. Each observation in a BN is a case, and case files (datasets) were used to generate conditional probability tables for network learning. The compiled BN is referred to as the base case. Full (100%) adoption of specified states was simulated to conduct scenario analyses (e.g., “What would be the predicted functionality if a WASH committee managed 100% of hand pumps?”). As noted, the edges in the BN model were hypothesized based on a literature review and expert review. The most ideal and least ideal scenario simulations were determined based on the results of the multivariate logistic regression. The only variables included in the most and least ideal scenarios were those that were modifiable (e.g., availability of toolkits). The ideal scenario simulates modifiable conditions that correspond to the highest odds of water points being functional, as predicted by the logistic regression model. In this scenario, a water point is repaired within a week by mechanics equipped with toolkits and managed by a local authority. Conversely, in the least ideal scenario, repairs take more than a year, no mechanic is present in the community, and there is no managing party responsible for maintenance.

To measure the impact of key variables, we calculated the Posterior Odds Ratio (POR) to quantify the strength of association between the two scenarios. The POR compares the likelihood of an outcome by dividing the odds of that outcome in a favorable scenario by its odds in an unfavorable scenario. First, the posterior probability ($\mathbb{P}$) from each scenario was converted to odds using the formula odds = ${\displaystyle \frac{\mathbb{P}}{(1-\mathbb{P})}}$. The magnitude of the resulting odds ratio was then interpreted using established guidelines to classify the association’s strength.

$$\mathrm{POR}={\displaystyle \frac{\mathbb{P}(\mathrm{Functionality}\mid \mathrm{Most}\mathrm{Ideal})·\left(1-\mathbb{P}(\mathrm{Functionality}\mid \mathrm{Least}\mathrm{Ideal})\right)}{\mathbb{P}(\mathrm{Functionality}\mid \mathrm{Least}\mathrm{Ideal})·\left(1-\mathbb{P}(\mathrm{Functionality}\mid \mathrm{Most}\mathrm{Ideal})\right)}}$$

We interpret the Posterior Odds Ratio in the same way as a traditional odds ratio from a frequentist model whereas values greater than 1 suggest increased odds of functionality under the “most ideal” scenario compared to the “least ideal” scenario, while values less than 1 suggest the opposite. However, because this POR is derived from a Bayesian network, it reflects conditional probabilities under the model’s structure and assumptions, rather than marginal associations estimated via regression.

The validation dataset was used to assess the model’s performance. Parameters used to evaluate such performance included the ROC curve, error rate, logarithmic loss, quadratic loss, and spherical payoff. Sensitivity analysis was conducted to evaluate how changes in other variables impact the functionality variable, using the Netica software.

3. Results

3.1. Descriptive Statistics Multivariable Logistic Regression Model

A total of $11,065$ water points were included in the analysis. A multivariable logistic regression model using maximum likelihood estimators revealed statistically significant associations between the outcome and predictor variables (model-wide p < 0.0001). The model with the lowest AIC was selected, and model fitting was evaluated using residual analysis. Sensitivity analysis using the Receiver Operating Characteristic (ROC) curve showed an Area Under the Curve (AUC) of 0.87, and the calibration plot showed good agreement between predicted probabilities and observed outcomes across prediction deciles (Figure 2), indicating that the model is well-calibrated with good performance [25].

Figure 2. Calibration plot showing observed vs. predicted probabilities of water point functionality. Points with 95% CIs grouped by decile; the dashed line indicates perfect calibration.

3.1.1. Maintenance and Availability of Mechanic

Most functional water points (35.5%, Table 4) were maintained by institutional organizations, while non-functional water points were more frequently managed by local authorities (32.4%, Table 4). Water points managed by WASH committees constituted 29.7% (Table 4) of functional and 33.4% (Table 4) of non-functional water points. The odds of functionality for locally managed water points are approximately four times those for water points with no management (95% CI: 3.95–5.67, p < 0.0001, Table 5). WASH committees were the least effective in management (OR = 2.65, CI: 2.3–3.1, p < 0.0001, Table 5).

Table 4. Descriptive statistics of predictor variables cross-tabulated with the outcome variable; each variable is cross-tabulated with functionality status.

Variable	Functional (N = 9089)	Non-Functional (N = 1976)	p-Value **
Maintainer			<0.0001
Institutional	3225 (35.5%)	640 (32.4%)
WASH committee	2699 (29.7%)	660 (33.4%)
Local	2694 (29.6%)	370 (18.7%)
No Management	471 (5.2%)	306 (15.5%)
Mechanic and toolkit availability			<0.0001
No mechanic	4771 (52.5%)	1110 (56.2%)
Mechanic with toolkits	3576 (39.3%)	647 (32.7%)
Mechanic without toolkits	742 (8.2%)	219 (11.1%)
Water point payment status			0.2253
No payment – it’s free	7898 (86.9%)	1723 (87.2%)
After a system breakdown	671 (7.4%)	135 (6.8%)
Yes, Flat rate	342 (3.8%)	88 (4.5%)
Yes, by volume	178 (2.0%)	30 (1.5%)
Time it took to repair			<0.0001
Never Broken	3757 (41.3%)	210 (10.6%)
Over a month	2213 (24.3%)	665 (33.7%)
More than a year	767 (8.4%)	807 (40.8%)
Less than a week	1220 (13.4%)	117 (5.9%)
Over a week	1132 (12.5%)	177 (9.0%)
Spare parts supplier distance			<0.0001
Quartile 1 (0–10 min)	2427 (26.7%)	486 (24.6%)
Quartile 2 (11–42.57 min)	2143 (23.6%)	362 (18.3%)
Quartile 3 (42.57–60 min)	2091 (23.0%)	456 (23.0%)
Quartile 4 (60–148 min)	2428 (26.7%)	654 (33.4%)
Water availability			<0.0001
Always Water	6270 (69.0%)	1091 (55.2%)
Seasonal	2819 (31.0%)	885 (44.8%)
Damage Status			<0.0001
No	6771 (74.5%)	323 (16.3%)
Yes	2317 (25.5%)	1653 (83.7%)

** p-values are based on Mantel–Haenszel Chi-square p-values comparing functional and non-functional water points across each survey item. Bold values indicates statistical significance α = 0.05.

Table 5. Odds ratios with Wald 95% confidence intervals for water point functionality from both univariate and multivariate logistic regression models.

	Unadjusted			Adjusted
	OR	95% CI	p-Value **	OR	95% CI	p-Value **
Water point maintenance
Local vs No Management	4.730	3.950–5.665	<0.0001	3.733	2.993–4.657	<0.0001
Institutional vs No Management	3.274	2.770–3.869	<0.0001	2.893	2.349–3.562	<0.0001
WASH Committee vs No Management	2.657	2.248–3.140	<0.0001	2.426	1.946–3.026	<0.0001
Time it took to repair
Over a month vs. More than a year	3.501	3.070–3.993	<0.0001	2.893	2.491–3.360	<0.0001
Over a week vs. More than a year	6.729	5.583–8.110	<0.0001	4.896	3.982–6.020	<0.0001
Less than a week vs. More than a year	10.971	8.858–13.59	<0.0001	6.372	5.046–8.046	<0.0001
Never broken vs. More than a year	18.823	15.87–22.32	<0.0001	7.095	5.869–8.577	<0.0001
Mechanic & toolkit availability
Mechanic without toolkit vs no mechanic	0.788	0.669–0.929	0.004	0.896	0.726–1.106	0.307
Mechanic with toolkit vs no mechanic	1.286	1.156–1.430	<0.0001	1.246	1.084–1.431	0.002
Damage Status
No vs. Yes	14.955	13.16–17.00	<0.0001	10.458	9.107–12.01	<0.0001
Payment Status
By Volume vs. Free	1.294	0.876–1.912	0.195	0.886	0.552–1.421	0.614
Flat rate vs. Free	0.848	0.667–1.078	0.177	0.847	0.634–1.133	0.263
After a system breakdown vs. Free	1.084	0.895–1.314	0.408	1.336	1.066–1.673	0.011
Water Availability
Seasonal vs. Year-Around	1.804	1.634–1.992	<0.0001	1.059	0.939–1.195	0.351
Supplier Distance (minutes)
Q1 vs. Q4	1.358	1.192–1.546	<0.0001	1.220	1.041–1.431	0.014
Q2 vs. Q4	1.609	1.398–1.852	<0.0001	1.566	1.322–1.855	<0.0001
Q3 vs. Q4	1.215	1.064–1.386	0.004	1.151	0.980–1.352	0.085

** Bold p-values indicate statistical significance at $\alpha =$ 0.05. “Unadjusted” model refers to univariate model logistic model while “Adjusted” model refers to multivariate logistic model.

Additionally, 52.5% (Table 4) of functional water points had no mechanic available, whereas this was the case for 56.2% (Table 4) of non-functional water points. The presence of mechanics without toolkits was observed at 8.2% (Table 4) of functional water points and 11.1% (Table 4) of non-functional points. The presence of mechanics with toolkits was more common in functional water points (39.3%, Table 4) than in non-functional ones (32.7%) (Table 4). Mechanics’ availability with a toolkit was associated with 1.25 times the odds of water point functionality compared to no mechanic (95% CI: 1.10–1.42, p = 0.002, Table 5).

3.1.2. Damages and Repair

Non-functional water points remain disproportionately damaged, with about 83% reported to have damage compared to only about 25% of functional water points (Table 4). Regression models concur that damages to the water point were associated with a significantly lower likelihood of functionality, with the odds of a damaged water point being functional being 10.5 times lower compared to undamaged ones (95% CI: 9.1–12.0, p < 0.0001, Table 5).

About 41.3% of functional water points had never broken down (Table 4), compared to 10.6% of water points that are non-functional (Table 4), and the time to repair exceeded a week in 12.5% of functional water points and 9.0% of non-functional points (Table 4), with significant differences in the repair time distribution across categories (p = <0.0001, Table 4). Water points that took over a year to repair were $7.10$ times less likely to be functional than those that never broke down (95% CI: 5.87–8.57, p < 0.0001,Table 5). Similarly, water points repaired in less than a week were 6.73 times less likely to be functional than those that never broke down (95% CI: 5.05–8.05, p < 0.0001, Table 5).

3.1.3. Payment for Water

Payment for water was not required at 71.38% of functional water points and 15.6% of all non-functional water points. The payment by volume was uncommon, observed at 1.6% of all functional water points and 0.27% of all non-functional water points. Similarly, flat-rate payments were observed at 3.09% of all functional water points and 0.8% of all water points, which are not functional. Payment was required only after a system breakdown, affecting 6.0% of all functional water points and 1.2% of all non-functional water points with no significant difference between groups (p = 0.22, Table 5). However, none of the payment methods (by volume, flat rate, or after a breakdown) showed significant odds ratios for predicting water point functionality, with p-values of 0.64, 0.26, and 0.011, respectively (Table 5).

3.1.4. Water Availability

The majority of most functional and non-functional water points have year-round water availability rather than seasonal access, with 69% of functional and 59% of non-functional points reporting year-round availability (Table 4). However, water availability did not have a statistically significant association with functionality in the adjusted model (OR = 1.05, 95% CI: 0.94–1.20, p = 0.35, Table 5).

3.1.5. Distance to Spare Part Provider

Quartile analysis revealed that the distance to spare parts suppliers was associated with water point functionality. Water points in the first quartile (0–10 min) had marginally higher odds of functionality compared to those in the fourth quartile (60 min+) (OR = 1.22, 95% CI: 1.04–1.43, p = 0.014, Table 5). Water points in the second quartile showed significantly higher odds of functionality (OR = 1.55, 95% CI: 1.32–1.86, p < 0.0001, Table 5), whereas those in the third quartile did not have significantly increased odds of functionality compared to the fourth quartile (p = 0.08, Table 5).

3.2. Bayesian Network Model and Posterior Odds Ratio

Bayesian networks (BNs) were used to model the posterior probability of outcome variable likelihood based on most-ideal and least-ideal scenarios (Figure 3). The BN model was evaluated for performance using several metrics, including the Area Under the Receiver Operating ROC, AUC, and logarithmic loss (see Appendix B). The model had an AUC of 0.81 with an error rate of 16.4% and a logloss of 0.37.

Figure 3. Base BN structure used in the analysis. Each node represents a variable; arrows indicate hypothesized causal relationships. Peach-colored nodes represent modifiable or mediator variables, blue indicates the outcome (functionality), and rounded white boxes are non-modifiable predictors. Each variable’s box shows its categories along with prior or posterior probabilities. Variables on the left are exogenous inputs; those on the right reflect model-derived distributions.

Under the “most ideal” scenario (a local authority maintains a water point, and repairs are completed within a week by a mechanic with a toolkit), the posterior distribution of functionality shows a high likelihood of 77.7% Yes and 22.3% No (Figure 4).

Figure 4. Simulated least ideal scenario in the Bayesian network, with modifiable predictor variables shown in gray. Repair time is set to “more than a year”, mechanic and toolkit availability to “No Mechanic”, and maintainer to “No Management”. Resulting posterior probabilities reflect the updated functionality likelihood under these optimal conditions. Each variable’s box shows its categories along with prior or posterior probabilities. Variables on the left are exogenous inputs; those on the right reflect model-derived distributions.

Under the “least ideal” scenario, where the water point is not managed, repairs take more than a year, and no mechanic is available in the community, the posterior probability of functionality shifts to “Yes” and 54.5% “No” (Figure 5).

Figure 5. Simulated most ideal scenario in the Bayesian network, with modifiable predictor variables shown in gray. Repair time is set to “Less than a week,” mechanic and toolkit availability to “Mechanic with toolkit,” and maintainer to “Local.” Resulting posterior probabilities reflect the updated functionality likelihood under these optimal conditions. Each variable’s box shows its categories along with prior or posterior probabilities. Variables on the left are exogenous inputs; those on the right reflect model-derived distributions.

We calculated the Posterior Odds Ratio (POR) to quantify the strength of the association between the management scenarios and water point functionality. This metric compares the odds of a water point being functional in the “most ideal” scenario to its odds in the “least ideal” scenario.

The analysis yielded a POR = ${\displaystyle \frac{0.777/(1-0.777)}{0.455/(1-0.455)}}=4.19$. This suggests that the odds of a water point being functional are over four times higher under the “most ideal” conditions compared to the “least ideal” conditions. Thus, the model provides strong evidence that maintaining water points under the optimal conditions defined in our “most ideal” scenario significantly improves their functionality.

Sensitivity Analysis

Sensitivity analysis was conducted using Netica’s “GetMutualInfo_bn” function to assess how different predictor nodes influence the outcome node, “Functionality”. Table 6 summarizes the mutual information values, which shows the strength of dependence between each node and the outcome functionality. Higher mutual information suggests a stronger influence on the target variable.

Table 6. Sensitivity analysis of the outcome node “Functionality” in the Bayesian network. Mutual information quantifies how much each node informs functionality. Percent = 100 × Mutual Info/Mutual Info (Functionality). Variance of Beliefs shows how much the posterior probability of functionality changed when each node’s states were varied.

Node	Mutual Info	Percent	Variance of Beliefs
Functionality	0.854	100.000	0.201
Damage Status	0.072	8.390	0.020
Repair time	0.043	5.000	0.012
Payment Status	0.011	1.320	0.003
Maintainer	0.011	1.260	0.003
Mechanic & Toolkit	0.005	0.549	0.001
County	0.005	0.561	0.001
Water Availability	0.002	0.263	0.001
Supplier rank	0.001	0.151	0.000

By identifying which factors contribute most to the uncertainty reduction in functionality, decision-makers can prioritize interventions on the most influential determinants. As shown in Table 6, Damage Status and Repair Time have the highest mutual information after functionality itself. The Variance of Beliefs provides an additional measure of sensitivity by indicateing how much the posterior probability of functionality changes when each node’s states vary [26].

4. Discussion

We explored factors influencing water point functionality in Liberia by analyzing survey data using multivariable and univariable logistic regression and BNs. We analyzed data for 11,065 water points, one of the few large-scale functionality data analyses from Liberia or SSA. We found that most water points in Liberia use Afridev-branded pumps, and therefore, we restricted our analysis to these pump types. We found that functionality varied by location, water point payment availability, damage status, community management, and the availability of mechanics and toolkits.

4.1. Mechanic and Toolkit Availability

Like results from other studies conducted in Sub-Saharan Africa, the availability of trained mechanics with toolkits emerged as a significant determinant of functionality [27,28]. In our study, the presence of a mechanic with a toolkit increased the odds of a water point being functional, highlighting the importance of investing in training and equipping local mechanics to ensure the sustainability of water points. Additionally, the functionality of hand pumps with a mechanic but no toolkit was not significantly different than the functionality of hand pumps with no mechanic at all. This finding supports prior work from other countries in the SSA region, such as Malawi, Ghana, Ethiopia, and Uganda, which emphasized the need for both technical capacity and access to tools to avoid prolonged downtime and disuse [12]. This finding is also consistent with those of Truslove et al. [18] in Malawi, where trained area mechanics with adequate resources were associated with significantly fewer infrastructure failures and improved hand pump serviceability.

4.2. Timely Repairs

We found that water points that took over a month to repair had significantly lower odds of being functional than water points repaired within a month (Table 5). Longer downtime could be due to several factors not included in this analysis, such as seasonality, but may lead to reduced faith in management systems, affecting revenue generation and committees’ ability to conduct subsequent repairs. Other studies have suggested that delays in minor repairs can escalate into major breakdowns requiring external intervention, which prolongs downtime and increases the risk of water point abandonment [29]. This cycle can weaken community trust in local management, reduce user contributions for maintenance, and ultimately contribute to system failure and demands for entirely new water sources. These results point to the need for decentralized repair capacity, where local actors are empowered and resourced to act quickly.

4.3. Water Point Maintainer

The management structure of a water point was significantly associated with its functionality (Table 5). Although the presence of any management doubled the odds of a water point being functional, WASH committees had the lowest odds of success among all managed points (Table 5). This result is noteworthy because these committees are often assumed to improve outcomes. This aligns with critiques in the literature that point to the limited technical and financial capacity of community-based structures (e.g., Truslove et al. [18], Chowns [30], and Carter and Ross [31]).

This finding is important to the ongoing debate in the rural water supply sector regarding the merits of community-based management versus more professionalized service models. For the last few decades, community-based (and often volunteer-driven) WASH committees have been the dominant model in Sub-Saharan Africa [32]. However, a growing body of evidence highlights their limitations in delivering sustained services, particularly when technical oversight and financial planning are required [30,33,34]. Critics have also questioned how genuinely participatory these models are [35,36,37]. Our findings add to this evolving discussion by showing that water points managed by institutional actors were more likely to remain functional than those managed by community volunteers. This supports recent calls to strengthen professionalized service provision, especially when the complexity of the system and the need for ongoing oversight exceed what volunteer-led committees can reasonably provide [38].

4.4. Water Fee Collection

More than 90% of the water points included in this survey were free to use (i.e., that no money was ever collected, or that money was only collected after a breakdown) (Table 4). This observation contrasts with findings in other research, such as Nigeria and Tanzania, where fee collection was strongly associated with improved functionality [12]. In our analysis, however, payment structure did not have a significant association with functionality after we controlled for other variables (Table 5). This suggests that fee collection features alone is not a reliable indicator of performance, and that other aspects of service delivery may be more important.

Most water points in Liberia are managed by institutional actors who often operate with external support or fixed budgets instead of relying on user contributions (Table 5). This arrangement may explain why fee collection was not a significant factor in our model. The situation also reflects a wider shift in the rural water supply sector, which increasingly recognizes that financial sustainability cannot depend solely on informal, volunteer-led systems. Instead, building savings from external or institutional sources, rather than relying solely on fee collection, may offer a more reliable path to maintaining functionality [39].

4.5. BN Model Simulation

The BN model, which enables the examination of complex systems by graphically representing relationships between variables, revealed important simulation results of the interplay of various factors influencing water point functionality. The model demonstrated the substantial impact of optimizing conditions on functionality by simulating both “most ideal” and “least ideal” scenarios. Under the “most ideal” scenario, characterized by features like local community management, rapid repairs, and the presence of a mechanic with a toolkit, the probability of functionality reached 77%. While under the “least ideal” scenario, marked by no management, extended repair times, and the absence of mechanics, functionality plummeted to a mere 45%. This suggests that modifiable structures, such as management structure, prompt repairs, and the presence of a mechanic and toolkit in the community, can drastically impact functionality.

Findings from the BN further supplement results from the multivariable logistic regression model by providing a more comprehensive understanding of the causal pathway between predictor variables used in this study. Further sensitivity analysis via the BN reflected that the most influential variables to water point functionality are repair time, damage status, and payment structure, which further emphasize the importance of prompt management of the water point.

4.6. Study Limitations

4.6.1. Causality and Older Data

This study’s reliance on cross-sectional data means that while associations can be identified, causal relationships cannot be established. Both Marks et al. [40] and Morita et al. [15] on hand pump sustainability in Ghana and Mozambique highlight this specific challenge of reverse causality, where the presumed effect could actually be the cause. This design makes it impossible to determine if effective management practices lead to improved hand pump functionality, or if an already functional system encourages better management and more positive community perceptions of the planning process [15,40]. Longitudinal studies that track changes in functionality and associated factors over time would be necessary to draw stronger causal inferences.

Although the data are several years old, the underlying service delivery structures and institutional capacity constraints have remained largely consistent. Liberia’s Pro-Poor Agenda (2018–2023) set ambitious WASH goals, including 85% access to basic water supply, but recent sector reviews continue to identify major funding gaps, weak institutional coordination, and lack of a dedicated Water Ministry as ongoing challenges [41]. As such, the patterns observed in the 2017 dataset likely reflect persistent systemic issues and offer relevant insights for current planning and policy.

4.6.2. Missing Data and Limited Scope

The study excluded data points from unimproved water sources and focused solely on “Afridev” hand pumps, which accounted for approximately 90% of the reported hand pumps (Table 2). As a result of those restrictions, our study results are limited to improved water sources using Afridev hand pumps. We employed robust imputation methods under the assumption of a Missing Completely at Random (MCAR) pattern. We confirmed this pattern using visual inspection of missingness patterns and analyzing missing data matrices. However, further work is needed to fully assess the validity of this assumption and understand the potential influence of missingness on our findings. Additionally, we did not have access to the original survey instrument, so certain variable definitions, such as “toolkit”, could not be confirmed. We interpret this term as referring to the basic set of tools typically required for routine maintenance of Afridev hand pumps, which are designed for ease of repair. By contrast, the other dominant pump type in Sub-Saharan Africa, India Mark pumps, were designed to break down infrequently, which may differ in maintenance and supply chain requirements [42]. Therefore, the findings reported here may not be generalizable across pump types. We were also limited to the variables available in the dataset. The definition of “functionality” has been critiqued and may not fully encompass all aspects of water point “operationality” [43].

4.6.3. Data Quality

Despite our best efforts to investigate the data-gathering process, the data collection manual and protocol could not be found. In addition, inconsistencies in data collection and reporting, subjective interpretations of survey questions, or inaccurate/missing responses from community members could introduce bias and affect the reliability of the findings.

4.7. Study Outcomes

Our analysis of over 11,000 water points reveals clear, actionable priorities for improving water service reliability in Liberia. The findings indicate that policy interventions could focus on two critical areas: strengthening the technical support system for maintenance and building the financial capacity of local management structures.

One of the most significant factors impacting water point functionality is the availability of a local mechanic properly equipped with a toolkit. This shows that both human capacity and access to essential tools are prerequisites for routine maintenance. To address this, we recommend moving beyond general training programs to implement more targeted initiatives. A key step would be to establish a formal certification program for water point mechanics, ensuring a standardized level of skill. This could be paired with a government- or NGO-supported program to provide certified mechanics with access to subsidized or financed toolkits and a reliable supply chain for spare parts.

Our findings also show that such technical capacity may only be effective within a functional management structure. It is not enough to simply have a mechanism for revenue generation; our work highlights that functionality improves only when those funds are actively directed toward technical management. Therefore, the core of “effective management” is the ability to translate collected fees into technical support by compensating a trained and equipped mechanic. While WASH committees are a common management model, they currently have lower odds of successfully maintaining a functional water point compared to institutional actors and local actors. This does not suggest a gap in operational capacity. Therefore, we recommend targeted support to strengthen the financial management skills of WASH committees. Concrete steps would include providing training in basic bookkeeping, establishing standardized protocols for tariff collection, and creating clear operational guidelines for the timely disbursement of funds for repairs.

To continue optimizing functionality for local managers, we recommend that the professionalized mechanics and supply chains should be directly accessible to these high-performing local management entities. This ensures that the most effective managers can readily access the technical support needed to prevent prolonged downtime. This approach leverages existing local capacity and focuses resources where they will have the greatest impact.

By implementing these specific, evidence-based policies, Liberia can build a more resilient and sustainable framework for its rural water infrastructure, ensuring communities have reliable access to this essential resource.