Introduction

Cholera is caused by the etiological agent Vibrio cholerae O1 and O139 serogroups and is transmitted through the faecal-oral route by consuming food or water containing the bacteria [1]. The World Health Organization (WHO)’s Global Task Force on Cholera Control views cholera not just as a global threat to public health, but also an indicator of poverty and a mark of inequity [2]. Since 1970, cholera has become endemic in many countries of sub-Saharan Africa [3] with annual Case Fatality Ratios (CFRs) of 1.6%—the highest in the world [4]. While several Oral Cholera Vaccines (OCVs) now exist, they only protect against cholera infection for up to two years, and they address the symptoms and not the cause. Large outbreaks continue to plague several countries, especially in sub-Saharan Africa. Without addressing the pathways to contamination in these low-income settings with inadequate infrastructure, outbreaks will continue to be a risk.

Pathways for transmission can include contaminated water sources, both for drinking and domestic use [5–8], poor sanitation [9–11]; contaminated food, poor personal hygiene [11] and poor drainage [12,13]. Furthermore, spatio-temporal analyses have demonstrated how cases can cluster among close contacts and therefore be more prevalent in dense living environments where an uncontaminated water supply is not readily available for drinking, washing or cooking [14]. These same pathways can also lead to other water-borne diseases such as typhoid (from the etiological agent Salmonella enterica serovar Typhimurium [15], making it vital to understand which pathways are most important in a given setting in order to target resources effectively to reduce the spread of these diseases. Additionally, since cholera is a water-borne disease, the focus in terms of risk factors is usually on infrastructure and water sources, yet another important factor is mobility. If infected individuals travel from an area with a cholera outbreak to an area without any cases and the water/sewer networks are susceptible to contamination, the traveller could contaminate the water in this new location and lead to the spread of the disease through host shedding into water supplies that are subsequently ingested [16,17]. Studying the role of mobility could be important for helping to contain spread from host shedding [18–20].

In particular, outbreaks in emerging nation contexts are usually attributed to where households have low access to ‘improved’ water and sanitation infrastructure such as networked water and sewer systems (‘improved’ water and sanitation is as defined by the WHO Joint Monitoring program) [3,21]. However, poorly maintained, or operated network infrastructure can also fail to bring about the protection from cholera that it may seem to promise. There are examples from developed and middle-income countries where poor operation of sewerage infrastructure can lead to health and environmental hazards [22–24].

An understanding of the spatial distribution of cholera risk, and the factors driving risk, can support decision making around longer-term interventions to reduce the likelihood of outbreaks. Especially in cities with mixed service provision—where households may rely on piped water and point sources for drinking water, and on-site sanitation or sewer networks for sanitation—there are a variety of potential improvements that could be made. These range from investing in repair or construction of network infrastructure (water or sewer pipes), to ensuring better access to protected point sources (boreholes, protected wells) or improving the quality of on-site sanitation (such as pit latrines or septic tanks). While in general, improving access to piped water is considered an improvement in infrastructure with public health benefits, piped water networks can become health hazards if not properly maintained. For example, in Zambia, Ashraf et al. found that service interruptions were associated with an increased incidence of diarrheal disease, upper respiratory infections and typhoid fever [25]. Equally, sewer networks are considered to be the top of the sanitation service ladder according to the Joint Monitoring Program of the WHO [26] with risks from onsite facilities more well-documented in the literature [27]. Nevertheless, when bursts or leakages are not properly addressed, liquid sewage can potentially be even more hazardous than waste from onsite sanitation facilities [28].

This paper brings together a rich set of geolocated indicators on water and sanitation, population characteristics and mobility, that jointly are rarely available in a lower-income setting. These indicators are used to conduct a geo-spatial analysis to explore the role of water and sanitation infrastructure on cholera risk. The study focuses on Harare, Zimbabwe, using the cholera outbreak of September–December 2018 as a case study. We use spatial modelling, detailed data on infrastructure and mobility, and georeferenced case data to identify factors associated with cholera risk. Based on this correlational analysis, we discuss different scenarios of infrastructural improvements and how they may be associated with reduction in relative risk when targeting by ward across the city.

Methods

Spatial modelling

One common method to model point event data, such as locations of disease cases, is to aggregate over an arbitrarily sized grid [40].This results in a count per grid cell, which can then be modelled using a Poisson regression. However, given that the size of the grid used to aggregate points is essentially arbitrary, results are conditional on the choice of grid size. While in theory it would be possible to decrease the size of the grid used (i.e., increase the spatial resolution of the grid), this can quickly become computationally expensive, particularly if the data in question are to be modelled over space and time, making it necessary to aggregate spatially for each period.

A more suitable approach is to apply Poisson point process models which allow models to be fit to the point event data directly. Poisson point process models are suitable for modelling the distribution of points in space and time and have been used in a number of settings including modelling disease case data, species distributions [42] and locations of storm peaks [43].

An important aspect of the models of cholera cases is the inherent spatio-temporal variation in disease risk which should be accounted for as part of any analysis. Log Gaussian Cox processes offer one way to include a spatio-temporal component to the model which can be computationally expensive to fit, although more efficient methods based on integrated nested Laplace approximation and stochastic partial differential equation models have been proposed [44].

Similar to the approach taken by Youngman and Economou [45], here we fit Poisson point process models using a GAM framework [44]. GAMs provide a flexible and efficient way to fit regression models and, particularly those involving spatial and spatio-temporal data alongside covariate effects. The approach tries to estimate the relationship between covariates, including the spatio-temporal term, and the intensity surface that gave rise to the cases. A standard way to estimate this relationship is using numerical integration at observation (i.e., case) point locations combined with a set of quadrature points [46], which are essentially evenly distributed points across the spatio-temporal domain of interest. A critical aspect of this approach is assigning quadrature weights to each point which then act as an offset term in the regression model. For most Poisson point process modelling (e.g., species distribution modelling using presence only data) this would be the area of the domain represented by the quadrature point. Where quadrature points are distributed over a regular lattice, this weight would be equivalent to the spatial resolution of the cells formed by the lattice. When modelling disease cases, however, the quadrature weights (offset) should reflect the population in the area represented by the quadrature point. Here we set the quadrature weight equal to the population in the corresponding lattice cell, excluding any cells in which cases occurred. Weights for cases were set as the population in the cell in which the case arose. The number of quadrature points per seven-day period was set to be 4,000, which inspection of model fit statistics suggested sufficient.

To account for residual spatio-temporal effects, the GAM used to model intensity included a tensor product between a bivariate Gaussian process smooth on latitude and longitude and a cubic regression spline on time. While covariate effects could be modelled using non-linear splines using this approach, here we opted to fit linear effects to simplify interpretation of associations.

S1 Table lists details of the covariates examined in the modelling process.

The covariates explored were selected based on plausible relationships with cholera risk. We include water infrastructure due to the well-known association between contaminated drinking water and water used for domestic purposes with cholera risk [5–7]. Though piped water is considered to be safer by World Health Organization standards, it depends on a continuous service through sealed pipes and with water treatment included. This is not always the case with poor quality piped infrastructure [25]. Sanitation infrastructure and sanitation facilities are also included due to their known association with cholera risk. For onsite facilities contamination can occur through unlined or poorly constructed pit latrines which seep content into the groundwater supply. Thus, groundwater vulnerability is also included in the model because it feeds into the many borewells in the city [47]. Meanwhile, sewer networks where breakages or overflows are common are also known to contaminate the water bodies they discharge into [48]. There are also illegal connections that can cause negative pressures. However, those working closely with the utility report that intermittent supply is the main cause of pressure surges, compromising system integrity. Vandalism outstanding bursts and aging infrastructure also serve to exacerbate maintenance challenges. From sewer pipes this may result in groundwater contamination when pipes are buried, or surface water if untreated sewage is discharged directly into streams or rivers [9–11]. We include the market and public facility locations as well as the population density data due to the pathways to contamination that have been identified in contaminated food and poor personal hygiene [11]. In addition, cholera transmission has been associated with case clusters among close contacts respectively [14]. We finally include an indicator that accounts for mobility by calculating the expected imported cases that come in from other areas (Fig 1). Emigration could lead to the generation of additional cholera cases in an area, given research showing that population mobility can lead to spread of cholera to new areas [17–20]. While distance to small and large sized markets was also considered, these were excluded from the modelling process given high correlation (greater than 0.8) with other covariates. All covariates were produced at, or resampled to, approximately 100m resolution (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Estimated incidence of imported cholera cases for the first 6 weeks of the outbreak in Harare.

Map made with R package Leaflet v2.1.1 and map data from OpenStreetMap.

https://doi.org/10.1371/journal.pntd.0011353.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Covariates included in the point process modelling.

Map made with R package Leaflet v2.1.1 and map data from OpenStreetMap.

https://doi.org/10.1371/journal.pntd.0011353.g002

To integrate human movement into the model, we use a methodology drawn from Milusheva et al. to calculate the risk of imported malaria [39]. For every individual/SIM we calculated the probability that an individual traveling to a destination became infected in the last ten days. This calculation was based on the 1) locations the individual spent time in during those ten days, 2) duration of time spent in each location and 3) local level of risk at each location (estimated by calculating observed weekly incidence per tower catchment). We then sum probabilities across all individuals entering a given destination ward in each week to calculate the total expected number of imported cases of cholera that week.

This resulted in an estimate of the incidence of ‘imported’ cases per ward for each week. This was rasterized to the same resolution and extent as the other covariates and included in the model as a dynamic covariate lagged by 1 week (i.e., incidence of imported cases in the week prior). Given there is an incubation period and potentially additional wait time before a new case is recorded in a facility, using this lag was meant to capture new cases that may have been generated because of infected individuals that have entered the area and “imported” cholera into the given destination.

These static and dynamic covariates were included in the model as linear effects. A backwards stepwise approach was taken with covariates retained in the model if their inclusion led to a decrease in AIC of greater than two. In addition, to examine the hypothesis that faults in the sewer network may have led to contamination of the piped water network, we included an interaction between the proportion of households with piped water and distance to the sewer network.

Model fit was evaluated by plotting the observed number of cases per time period against the number predicted by the model. In addition, we constructed a hex grid of approximately 1km width and compared the number of observed cases across time periods per grid cell to the number predicted.

Results

Fig 3 shows the distribution of cases temporally and spatially. Cases were concentrated towards the first few weeks in the south-west area of the city in the Glen View and Budiriro neighbourhoods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of cholera cases across Harare temporally by week (left) and spatially across all weeks (right).

Basemap attribution: CARTO (basemaps linked here), OpenMapTiles, OpenStreetMap contributors (license linked here).

https://doi.org/10.1371/journal.pntd.0011353.g003

All covariates except for water burst density were retained in the final model, including the interaction between proportion with piped water and distance to the sewer network. Table 1 shows the coefficients for the final model. The final model had an AIC value of 56034. The same model without lag incidence of imported cases had an AIC of 56219 and a spatio-temporal only model (i.e., with no covariates) had an AIC of 57943. This indicates that the addition of covariates and lag incidence improved the model fit. Improvement of model fitting was made using Hastie et al. as a reference [49].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Results from the point process modelling of cholera risk in Harare.

Coefficients represent the difference in the log of the expected rates for a unit increase in the covariate.

https://doi.org/10.1371/journal.pntd.0011353.t001

Model validation showed that the fitted values showed good alignment with the observed numbers of cases, both per week and per 1km hex grid cell (Figs 4 and S1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4.

Observed versus predicted numbers of cases per week (left) and per ~1km hex grid cell (right).

https://doi.org/10.1371/journal.pntd.0011353.g004

The models suggested that sanitation risk, poverty, sewer burst density and incidence of imported cases were all risk factors for cholera (Table 1). In contrast, increasing geological vulnerability, distance to medium sized market, proportion with unimproved or other water source, private borehole density and distance to water network were all negatively associated with risk. To aid interpretation, Table 2 shows the relative risk per covariate using a scaled approach. Rather than showing the relative risk for every unit increase in the covariate, which is dependent on the scale of the covariate, here we calculate the relative risk between covariate strata, where strata are defined by dividing each covariate into 10 equal width strata across the range of covariate values found in populated grid cells. For example, if the values in populated grid cells of a given covariate range from zero-two hundred, the relative risk is shown for every increase of twenty in the covariate. This shows that differences in geological vulnerability, poverty and lag incidence of imported cases were associated with relatively small changes in risk with relative risks per incremental stratum of 0.92, 1.12 and 1.07 respectively. In contrast, other covariates, particularly proportion with unimproved water, sewer burst density and distance to water network were associated with larger differences in risk.

The models also suggested that there was an interaction between the proportion of households with piped water and distance to the sewer network. Fig 5 describes this interaction, showing the relationship between the proportion of households with piped water and cholera risk at different distances from the sewer network when all other covariates are set to their mean. This shows that adjacent to the sewer network (at zero km), risk is slightly higher, the higher the proportion of households with piped water. However, this relationship switches direction in areas greater than two km from the sewer network, where the higher the proportion of households with piped water, the lower the risk of cholera. (A log visualization of this relationship is provided in S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. The effect of proportion with piped water at different distances to the sewer network.

The y axis represents the risk (cases per person) when holding all other covariates at their mean value. Higher values indicate higher cholera risk.

https://doi.org/10.1371/journal.pntd.0011353.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Relative risk of moving into the next stratum of a covariate (excluding interaction between distance to sewer and proportion with piped water) where strata are formed by dividing the covariate values (in populated grid cells) into ten equal width strata.

For example, if the values in populated grid cells of a given covariate ranges from zero–two hundred, the relative risk is shown for every increase of twenty in the covariate.

https://doi.org/10.1371/journal.pntd.0011353.t002

Fig 6 shows the predicted incidence at 100m resolution across Harare for the first six weeks in which 88.6% of cases were recorded. Incidence is markedly higher towards the southwest part of the city around Glen View and Budiriro neighbourhoods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Predicted incidence of cholera across Harare for the first 6 weeks of the outbreak (September 1^st–October 13^th, 2018).

Note the log-scaled colour palette. Basemap attribution: CARTO (basemaps linked here), OpenMapTiles, OpenStreetMap contributors (license linked here).

https://doi.org/10.1371/journal.pntd.0011353.g006

Discussion

There are a few key findings that come out of this research and have important policy implications. As demonstrated by the coefficients in Table 2, the models found an association between sanitation risk, poverty, sewer burst density and incidence of imported cases with increased cholera risk. Meanwhile, increasing geological vulnerability, distance to a medium sized market, proportion with unimproved or other water source, private borehole density and distance to water network were associated with reduced cholera risk. While the correlation of poor sanitation with cholera risk [12] is well known in the literature, including in the context of a previous cholera outbreak in Zimbabwe [50], the rich geolocated data we have compiled allows us to identify and associate specific types of sanitation infrastructure. Our model results also align with previous literature demonstrating an association with poverty [51]. Associated factors with poverty and market size are often identified in the literature such as population density, overcrowding and low educational status [52,53]https://www.zotero.org/google-docs/?1kGWLd. The association with reduced risk of proximity to the water network also complements similar findings in Lusaka, Zambia [12].

A counterintuitive finding from this work is that the proportion of those with unimproved water source was associated with reduced cholera risk. The model suggests that in Harare, there was an association between the combination of the proportion of households with piped water and distance to the sewer network, with increased cholera risk. In other words, it suggests that the closer to the sewer network, the greater the association with increased cholera risk, for households with piped water. However, this relationship switches direction in areas greater than two km from the sewer network, where having piped water is associated with lower risk of cholera. It is unclear as to why those with a higher proportion of unimproved water are associated with lower cholera risk and may be explained by contamination of the piped water network which temporarily makes unimproved water sources relatively safer. In addition, there is undoubtedly error in the predicted proportion with unimproved water as these data are based on a small number of obfuscated DHS data points which could lead to a spurious relationship. The DHS has a policy of randomly displacing the GPS coordinates of households to protect the privacy of respondents by between zero and two km. Similarly, ordinarily we would expect that the relationship with geological vulnerability to be associated with increased cholera risk, especially if sewage or OSS effluent is proximate to unprotected drinking water sources. Therefore, in this case, the reasons for a lack of a negative association require further exploration.

The most accurate way of testing whether the piped water closer to the sewer network is more contaminated would be through water quality testing data at different points along the network. Although water quality testing data was provided for this analysis, there were insufficient data points to be able to make such a comparison. In the absence of such data, this analysis can still provide some useful insights into spatial targeting of investments. As shown in Fig 1, sewer bursts are spatially very concentrated in a few areas. Therefore, it could be possible to target sewer maintenance investments in these few areas, and particularly in areas which are closer to the piped water network. Assuming these areas represent locations of the sewer network more vulnerable to future bursts, doing so would thereby offer health benefits in terms of reducing cholera risk, while minimizing costs since it would require investments in only a few areas. Such targeting is illustrated by example scenarios provided in Annex 2 on how the effectiveness of different targeting schemes could be calculated. Though costing data was not available for us to extend the analysis in that direction, by virtue of targeting a few wards, rather than the whole city, the study offers policymakers some indication on the best use of resources.

We identify that spatially, the risk of sewer bursts on this occasion was concentrated in five wards in the southwest of the city (S2 Fig). It is likely that sewer bursts do not occur at random but are spatially correlated due to underlying weaknesses in the network in those areas. Therefore, targeting the five wards where the majority of sewer bursts were reported could be an efficient way to reduce cholera risk relative to repairing sewer bursts across the city. This is supported by our data, that a high density of sewer bursts was positively associated with cholera cases. Using the data from the bursts reported and analysed, there appeared to be relatively little increased benefit (in terms of reducing cholera risk) in investing in repairing sewer bursts throughout the city as opposed to just focusing on the five wards mentioned. The efficiency and cost effectiveness of reducing cholera risk is therefore much greater when concentrating on areas that fall within the top 25th percentile of sewer bursts. It is also more efficient to provide piped water to within 500 meters of the existing network, rather than to extend access to the whole city when considering their relative impact on reducing the number of cholera cases.

The association between sewer bursts and cholera risk is important given that according to the Joint Monitoring Program of the WHO, who set the standards for sanitation globally, sewered sanitation sits at the top of the improved sanitation ladder. However, it serves to highlight the importance of proper maintenance of existing sanitation networks rather than necessarily emphasizing new construction. This is particularly important in a context where the institutional capacity and financial resources to maintain large scale sanitation systems may be lacking.

An important limitation of the study is that the analysis presented is only able to draw correlations and associations, rather than drawing causal inferences given we are not able to use some form of exogenous variation to identify the causal impact of individual factors. Instead, we look at the relative predictive power of a small set of indicators, which can help to provide insights into factors that may be important for policy makers to consider in further analysis. This clearly has important implications when interpreting the scenario analysis. Secondly, when considering data quality and reliability, we must consider that a number of layers used in the analysis, such as the proportion of households with access to piped water, were generated using DHS data, which have been spatially obfuscated. As a result, this may have introduced error into the layer, which could have reduced the accuracy of predicted associations with cases. Sewer and water burst density was based on spatio-temporally referenced complaints data, which is obviously a proxy measure with a number of reported and recorded issues. Data from direct inspections of sewer and water networks would have been a more reliable measure but such data was not available to the researchers. Finally, the case data is based on the line list, which though geo-referenced to households, does rely largely on suspected rather than confirmed case data. Although this data collection method does not deviate from the WHO GTFCC guidelines, having a larger proportion of verified cholera cases may have changed the distribution of the data used in the analysis. A recommendation for further outbreaks may be to increase the proportion of confirmed cases with respect to suspected.

The focus of this analysis has been to better understand spatial variation in risk of cholera, using Harare, Zimbabwe as a case study that could help to provide insights for low-income urban settings, and to study what factors might be associated with cholera risk. A natural extension of the work would be to focus on some of the less intuitive associations and identify settings where it may be possible to study them in a causal framework. This could then lead to more direct policy implications of factors that need to be tackled, which could be combined with cost estimates to compare different targeted interventions. Furthermore, if additional data were available on factors driving sewer bursts, the analysis could be enhanced to have predictive capacity to reduce the risk of future outbreaks. Above all, this analysis shows the importance of emphasizing maintenance above and beyond new construction, especially when thinking of extending an already poorly maintained system.

Supporting information

Acknowledgments

The authors also gratefully acknowledge the cooperation of officials from City of Harare (CoH) Water and GIS Departments, in particular Toine Ramaker from VEI, Dr. Manangazira from the Ministry of Health, Zimbabwe; in country counterparts at the World Health Organization (WHO); and the Country Management Unit (CMU) of the World Bank in Zimbabwe, in particular then Country Manager Mukami Kariuki, Senior Operations Officer Tonderai Fadzai Naome Mukonoweshuro and Health Specialist Consultant Chenjerai Sisimayi for facilitating contact with counterparts and local dissemination of findings. From the Water GP, we also received invaluable support from Anna Cestari, Senior Water Resource Management Specialist, Ireen Mangoro, Water Consultant, Chris Heymans, Senior Water Supply and Sanitation Specialist. Without their support this work would not have been possible.

The findings, interpretations and conclusions expressed in this paper do not necessarily reflect the views of the World Bank, the Executive Directors of the World Bank, or the governments whom they represent. The World Bank does not guarantee the accuracy of the data included in this work.

Official delimitation of areas and borders do not necessarily reflect the official position of the World Bank Group. Country borders or names do not necessarily reflect the World Bank Group’s official position. These maps are for illustrative purposes and do not imply the expression of any opinion on the part of the World Bank, concerning the legal status of any country or territory or concerning the delimitation of frontiers or boundaries.