Article pubs.acs.org/est
Evaluating Cellular Instrumentation on Rural Handpumps to Improve Service DeliveryA Longitudinal Study in Rural Rwanda Corey Nagel,† Jack Beach,‡ Chantal Iribagiza,§ and Evan A. Thomas*,∥ †
Division of Epidemiology, Oregon Health & Science University, Portland Oregon 97239-3098, United States Independent Contractor, Living Water InternationalRwanda, Kigali, Rwanda § Living Water InternationalRwanda, Kigali, Rwanda ∥ Department of Mechanical Engineering, Portland State University, Portland, Oregon 97201, United States ‡
ABSTRACT: In rural sub-Saharan Africa, where handpumps are common, 10−67% are nonfunctional at any one time, and many never get repaired. Increased reliability requires improved monitoring and responsiveness of maintenance providers. In 2014, 181 cellular enabled water pump use sensors were installed in three provinces of Rwanda. In three arms, the nominal maintenance model was compared against a “best practice” circuit rider model, and an “ambulance” service model. In only the ambulance model was the sensor data available to the implementer, and used to dispatch technicians. The study ran for seven months in 2014−2015. In the study period, the nominal maintenance group had a median time to successful repair of approximately 152 days, with a mean per-pump functionality of about 68%. In the circuit rider group, the median time to successful repair was nearly 57 days, with a per-pump functionality mean of nearly 73%. In the ambulance service group, the successful repair interval was nearly 21 days with a functionality mean of nearly 91%. An indicative cost analysis suggests that the cost per functional pump per year is approximately similar between the three models. However, the benefits of reliable water service may justify greater focus on servicing models over installation models.
■
INTRODUCTION To provide access to safe drinking water in support of the Millennium Development Goals, international donors and governments have installed improved water point sources throughout developing countries resulting in an apparent increase in access to improved water supplies in rural areas from 58% in 1990 to 91% in 2015.1 However, reliable, sustained water service delivery remains a challenge.2 In rural sub-Saharan Africa, where hand pumps are a common technology, 10%−67% of improved water sources are nonfunctional at any one time, and many never get repaired.2,3 Furthermore, only about half of protected dug wells (inclusive of handpumps) adequately meet the implicit intent of MDG 7, targeting safe drinking water4 due to poor groundwater quality and temporary outages derived from seasonal fluctuations in water table levels.5 Finally, imprecise methods used by nonprofits, private companies, and water drillers to evaluate handpumps for a limited period of time after implementation add to the reduction of true water access improvement figures in developing world countries.6 Problems maintaining the functionality of water pumps result from social, logistical, and technical issues like the breakdown of community management structures, insufficient human resources to provide service, and lack of spare parts.7 Community-led activities commonly include creating user committees to determine access arrangements, water fees, and © 2015 American Chemical Society
maintenance provision. A community caretaker may provide service and repairs for small problems. For significant problems and pump failures, communities contact a pump mechanic, often one affiliated with the local water district. In the contract model, the community or water district hires an organization to manage O&M. The organization may make predetermined, periodic service visits, often every month to a year, to provide service and fix identified problems with water pumps (i.e., circuit rider service model) and/or communities may call for service as needed (i.e., ad-hoc, or nominal, service model). Importantly, this ad-hoc service model competes for budget and time with other activities technicians and managers are providing, including water pump installations. Downtime is substantial due to problems identifying and reporting pump failures and lack of funds to pay for service and repairs.6 Remotely reporting instrumentation may help address challenges in reliability of water supplies through enabling greater accountability and responsiveness of cognizant service providers. This approach is enabled by the impressive growth in cellular network population coverage, 60% across sub-Saharan Africa in the past 15 years.8 Received: Revised: Accepted: Published: 14292
August 26, 2015 October 29, 2015 November 23, 2015 November 23, 2015 DOI: 10.1021/acs.est.5b04077 Environ. Sci. Technol. 2015, 49, 14292−14300
Article
Environmental Science & Technology
network connectivity were recorded. Nine of 324 sites did not have confirmed cellular data connectivity. Site selection and cohort assignment was iterative, and semirandom. The research team agreed to restrict the Ambulance and Circuit models to the Ruhango (Southern Province) and Karongi (Western Province) districts such that technician routes could be geographically limited. Given the number of pumps available in these districts, this necessarily biased the Nominal model to pumps in the Central province. From the available LWIR managed pumps with cellular connectivity at baseline in these two districts, pumps were initially randomly assigned between Ambulance, Circuit, and Nominal cohorts. Baseline functionality was not an inclusion criteriasensors were installed on both functional and nonfunctional pumps, with the exception of pumps that were nonfunctional in a way that would expose the sensor to vandalism. Five sites were excluded because of poor cellular connectivity realized during sensor installation. Reserve sites were available from the randomization process to account for site exclusions. Instrumentation and Data Management. The sensor technology used in this study has been described, in earlier forms, elsewhere,12 and was designed and produced on a prototype basis for this study. The enclosure housed a AA sized battery compartment (allowing the use of disposable alkaline or rechargeable nickel−metal−hydride batteries), a control board, a cellular radio chip and SIM card holder, an accelerometer, and a differential water pressure transducer. The water pressure transducer had one port open to the atmosphere and the other submerged within the water pump overflow basin, thereby designed to record water level as pressure, regardless of ambient temperature or altitude. The sensor enclosure exterior included a high strength magnet for attachment, and a smartphone scannable barcode for tracking. The sensor includes a fully integrated cellular connectivity system, and reports data over cellular networks directly to an online platform. In both Afridev and India Mark 2 handpumps, the hand lever is attached to a riser main drawing water up through a PVC or galvanized iron pipe. The water overflows the pipe into a basin and then flows out from the tap. In both pump types, the sensor was installed within the overflow basin, attached by to the external pump wall. An image of the installation location is shown here (http://www.pdx.edu/sweetlab/cellpump-rwanda). Sensor reported pump usage data were matched to sites and maintenance cohorts based on sensor installation records. In most cases, sensor location, maintenance, and replacement were tracked through barcodes attached to both sensors and pumps, and scanned on smartphones, and included smartphone derived GPS location. A combination of MySQL, PHP, and R scripts interpreted the sensor reported data. Accelerometer derived event data were used in this analysis. Daily, an R script polled each sensor data table and identified gaps in accelerometer triggered data exceeding 60 s, an arbitrarily selected indicator of separate usage “events”, and tallied all events in the 24 h preceding the latest data timestamp. The script assigned an alarm status of “green” to pumps recording more than 100 events in 24 h, “yellow” to pumps between 10 and 100 events, and “red” to pumps recording fewer than 10 usage events. If the latest sensor timestamp was more than 7 days old, a status of “sensor fault” was recorded, suggesting a nonfunctional sensor. These arbitrary thresholds were periodically scrutinized and adjusted to these values within the first weeks of the study.
In 2013, Oxford University trialed mobile enabled technologies on 66 “smart handpumps”. The study demonstrated that handpumps with cellular network enabled sensors which received a repair in the trial saw a decrease in pump downtime from an average of 27 days to 2.6 days. Participating communities also increased their willingness to pay for pump services by over 3-fold.9 Study Design. In 2014, 181 cellular enabled water pump use sensors were installed in three provinces of Rwanda. In three arms, the status-quo nominal model of operation and maintenance (Nominal) was compared against a “best practice” circuit rider model wherein technicians were assigned to pump groups and visited pumps on a scheduled, periodic basis (Circuit), against an “ambulance service” model (Ambulance). In only the ambulance model was the sensor derived data made available to the implementer, and subsequently used to dispatch technicians when pump failures were identified. The objective of this study was to examine if (a) in situ cellular reporting sensors could enable increased pump functionality through greater responsiveness, and (b) of the three models, which was most cost-effective in delivering a water supply. The study ran for approximately seven months between November 1, 2015 and May 31, 2015. Typically, Rwanda has rainy seasons between March and May, and again in October and November. Therefore, roughly half the study was conducted in the rainy seasons and half in dry. A preliminary review of the data set after six months suggested that significant differences were observed between the service models, which allowed budgetary constraints to justify halting the study at seven months, while the operational trial continued throughout 2015. Study Context. The hand pump users in his study reside primarily in clustered or grouped rural settlements in the Central, Western, and Southern provinces of Rwanda. Twentyfour percent of the improved water sources in these provinces are hand pumps.10 While planning, regulation, hygiene promotion, monitoring, and oversight are the responsibility of the central government in Rwanda, the 30 districts are responsible for water and sanitation infrastructure including operation and servicing of improved water sources. While public−private partnerships manage 25% of water and sanitations services for piped rural systems in Rwanda,11 private organizations less commonly manage hand pumps. Instead, communities are responsible for operation and servicing for these smaller systems. However, the Karongi and Ruhango have contracted (at no direct cost) a nongovernmental organization, Living Water International Rwanda (LWIR), to provide the installation and maintenance services for the majority of the hand pumps. These Districts do not have direct financial resources or staffing presently available to manage these services directly and at present communities are not providing fees for service. The study included 181 rural communities with hand pumps. LWIR estimated that each hand pump serves on average 250 people per day Therefore, LWI estimated the direct user base was 45 250 people. While over 84% of households have access to a pit latrine that is not shared, few have access to electricity, with upward of 80% relying on kerosene, firewood, and candles for lighting and nearly all using firewood or other biomass for cooking fuel.10 Baseline Determination and Site Selection. A baseline evaluation was conducted by LWIR at 324 hand pump sites across 18 of the 30 districts in Rwanda between November 2013 and February 2014. Pump functionality and cellular 14293
DOI: 10.1021/acs.est.5b04077 Environ. Sci. Technol. 2015, 49, 14292−14300
Article
Environmental Science & Technology
Statistical Analysis. Data Sources. All sensor recorded data were parsed into pump specific data sets based on sensor maintenance records matching sensor IDs with pump sites between installation and, if applicable, removal dates. On this basis, for each day, sensor data were recorded while installed in a pump, the total events detected were tallied. Gaps in sensor data because of sensor nonfunctionality (average duration = 14.3 days) were interpolated based on the following criteria: (a) If a sensor detected pump functionality or nonfunctionality both before and after a data gap, then the pump was assumed to have persisted in that status through the period; (b) if a sensor gap suggested a change from functional to nonfunctional or vice versa, then the change was assumed to have occurred in the middle of the gap period. Functionality reports from both the LWIR and sensor maintenance team site visits were used to confirm that interpolation of data gaps was accurate and all interpolation of sensor records was conducted by an investigator blinded to the service model assignments. Repair events were entered on a smartphone form at the point of service by the LWIR O&M team and data were uploaded to a SQL server. The repair record included location, date, whether the pump was functioning on arrival, repair required, whether the repair was performed, and photos of the site. Preimplementation and Baseline. The pump failure rate (per pump month) was calculated using historical LWIR O&M data supplemented with network validation data. Univariate regression analyses with robust (Huber-White) standard errors to adjust for geographic clustering were used to identify pump and site characteristics associated with pump function and pump downtime. Among the sites included in the study, we used descriptive and bivariate analyses to assess for differences in pump characteristics between the service models. Chi-squared tests were used for categorical variables and Kruskal−Wallis test was used for continuous variables. Pairwise comparisons following Kruskal−Wallis test were performed using Dunn’s test with Hochberg’s correction.13 Functional Time. We used fractional logit regression to examine the relationship between pump functional time and service model. Because the number of observed days was not equivalent across all the study sites (due to gaps in the sensor record), the proportion of time that each site was operational was calculated by dividing the number of functional days by the total number of days under observation. As the resulting outcome variable was a proportion ranging from 0 to 1, we employed a quasi-likelihood approach to modeling fractional response variables described elsewhere.14 We constructed both univariable and multivariable models of the association between the proportion of functional days and service model. Multivariable models were adjusted for pump type, pump age, and well depth. To assess for potential bias resulting from (1) the use of maintenance team data to supplement the sensor record or (2) differences in baseline functionality between the study groups, we performed a sensitivity analysis using only sensor-recorded data and another after defining each pumps baseline as the first day in the study period the pump was operational. In both cases, the results showed the same directional patterns and significance as the primary analyses. Time to Repair. Accelerated failure time (AFT) models with a log−logistic distribution were used to examine the association between service model and time to repair among nonfunctional wells during the study period. Time to repair was calculated as
A separate script, running every 6 min throughout the study period, polled the LWIR operation and maintenance record database, collected by smartphone on iFormBuilder, parsed the field where technicians were instructed to scan a pump barcode if available, and identified any pumps visited by the LWIR O&M teams. If an O&M record was recorded within the preceding 14 days, then the location status was changed to “under repair”. Another script, running on a 6 min interval, tabulated the present recorded status of all pumps. These data were then presented visually on an online dashboard in both table and map form. The academic partners in this study had insight into all sensor data, while LWIR’s online login provided access only to the sensors associated with the Ambulance service model. A sensor maintenance team, supervised by the academic partners, were privy to all sensor statuses, in order to maintain a reasonable level of sensor functionality through site visits for battery and/or sensor replacements. Pump Servicing Models. As fee for service had not previously been instituted within these communities, and in many cases service level in the baseline was less than intended, this trial assumed a free maintenance service that, if value was sufficiently demonstrated, may later be institutionalized through local governments and fees secured from local communities. Ambulance. The Ambulance model was the only method that used sensor-derived pump functionality data operationally. Across 24 sites in Karongi and 23 in Ruhango, an LWIR O&M manager based in Kigali used the dashboard described above to identify water pumps that presumed to have failed, and to dispatch the dedicated Ambulance O&M team. The Ambulance team was composed of two technicians and a pickup truck. The team dispatched to the repair sites working in a single district before moving to the next district. If all assigned repairs were completed before the end of a week, then the LWIR manager provided new sites. Circuit. The Circuit rider model was modeled on best-known practices for rural water supply maintenance by organizations similar in structure and mission to LWIR. This model involved one LWIR technician based in the Karongi District covering 26 pumps and another covering the 24 pumps in the Ruhango District. These staff utilized Yamaha 100 cm3 motorcycles to travel on a routine service visit “circuit” stopping at every pump, regardless of the functional status. In this model, when a pump failure or preventative maintenance activity was required, the technicians attempted to perform the repair. When a repair required tools or materials not carried by the Circuit technicians, the LWIR manager dispatched a separate repair team, drawn from the Nominal maintenance model, described next. This team was usually dispatched in the week following an identified need. Nominal. The nominal maintenance model approximates the O&M servicing LWIR typically deployed prior to this study. LWIRs primary activity in Rwanda focused on pump installation and O&M services were provided only on a request basis, when communities or officials contacted LWIR. These requests would be weighed with other demands on technician and manager time and budget, including water pump installation obligations. As maintenance requests increased in a region, LWIR would schedule a servicing team. During the study period, this model included two part-time staff and a pickup truck. In this way, this model reflects, as best approximated, the typical servicing provided by LWIR historically. 14294
DOI: 10.1021/acs.est.5b04077 Environ. Sci. Technol. 2015, 49, 14292−14300
Article
Environmental Science & Technology Table 1. Sample Characteristics ambulance number of sites western province (Karongi District) southern province (Ruhango District) central province (Kigali) total observation days observation days per pump: Median (IQR) non-functional pumps at baseline: N(%) non-functional pumps at endline: N(%) pump type: N (%) AfriDev India Mark 2 U2 India Mark 2 Normal pump age (months): median (IQR) cylinder depth (feet): median (IQR) casing diameter (inches): median (IQR)
circuit
nominal
42 24 23 0 8509 208 (5) 13(30.95) 4(9.52)
44 26 24 0 8920 207 (9) 18(40.91) 5(11.36)
82 16 38 30 15360 197 (16) 22(26.83) 32(39.02)
20 (47.62) 18 (42.86) 4 (9.52) 43.02(41.93) 33(18) 125(50)
14 (31.82) 23 (52.27) 7 (15.91) 56.23(40.35) 30(19.5) 125(25)
48 (58.54) 15 (18.29) 19 (23.17) 56.77(65.03) 33(22.5) 125(25)
p-value
