Centralized or Onsite Testing? Examining the Costs of Water Quality Monitoring in Rural Africa

Abstract

Rural water systems in Africa have room to improve water quality monitoring. However, the most cost-effective approach for microbial water testing remains uncertain. This study compared the cost per E. coli test (membrane filtration) of four approaches representing different levels of centralization: (i) one centralized laboratory serving all water systems, (ii) a mobile laboratory serving all systems, (iii) multiple semi-centralized laboratories serving clusters of systems, and (iv) decentralized analysis at each system. We employed Monte Carlo analyses to model the costs of these approaches in three real-world contexts in Ghana and Uganda and in hypothetical simulations capturing various conditions across rural Africa. Centralized testing was the lowest cost in two real-world settings and the widest variety of simulations, especially those with water systems close to a central laboratory (<36 km). Semi-centralized testing was the lowest cost in one real-world setting and in simulations with clustered water systems and intermediate sampling frequencies (1-2 monthly samples per system). The mobile lab was the lowest cost in the fewest simulations, requiring few systems and infrequent sampling. Decentralized testing was cost-effective for remote systems and frequent sampling, but only if sampling did not require a dedicated vehicle. Alternative low-cost testing methods could make decentralized testing more competitive.

Keywords: cost modeling; membrane filtration; rural water quality; safely managed water; water quality monitoring; water quality testing.

Figure 1

Four approaches to water quality monitoring considered in this study. Each diagram provides an illustration of what sample collection might look like on a given day, as shown by the arrows and changing locations of the mobile laboratory.

Figure 2

Breakdown of microbial water quality monitoring costs associated with four microbial testing approaches across three settings in Ghana and Uganda, based on the results of 10,000 Monte Carlo simulations under uncertainty in each setting. The total height of each bar represents the median cost per water test, and the error bars represent the 25th and 75th percentile values from the Monte Carlo simulations (the values are also given in the text annotations above all bars). Each graph also indicates the portion of simulations in which each approach had the lowest cost per test.

Figure 3

Distribution of microbial water quality monitoring costs associated with four microbial testing approaches, from 10,000 Monte Carlo simulations across the generalized parameter space. For each approach, results from each simulation have been sorted from the lowest to highest total cost per test. Simulations showing no cost on the right side of a plot denote extreme cases where that microbial testing approach proved to be infeasible (e.g., water systems are too distant for the sampling car to travel to and from a single system in one 8 h workday). We constrained the y-axis due to the extremely high costs associated with simulations at the upper end of each approach’s distribution.

Figure 4

Lowest cost approaches from 10,000 Monte Carlo simulations, displayed within the multidimensional space produced by the five most important parameters influencing relative cost per test, as determined by 10 random forest models (Table S4). In the four plots, the background color reflects the lowest cost approach in the largest number of simulations within a grid cell. (Each parameter range was divided into 25 segments to create grid cells for easier visualization of trends.) Differently colored points show simulations where the lowest cost approach is different from the trend defined by the background color. Below the plots, we provide an illustrative decision tree involving the five most predictive parameters. This illustrative tree is one of the ten generated from separate sets of 10,000 Monte Carlo simulations (Figure S3), and we felt this tree provided a representative illustration of the structure and thresholds across all ten trees. It provides practical insight into the conditions controlling the lowest cost approach, though it does not accurately capture every situation (accuracy of only 66%; Table S4).