Measuring the spatial heterogeneity on the reduction of vaginal fistula burden in Ethiopia between 2005 and 2016

Vaginal fistula is a shattering maternal complication characterized by an anomalous opening between the bladder and/or rectum and vagina resulting in continuous leakage of urine or stool. Although prevalent in Ethiopia, its magnitude and distribution is not well studied. We used statistical mapping models using 2005 and 2016 Ethiopia Demographic Health Surveys data combined with a suite of potential risk factors to estimate the burden of vaginal fistula among women of childbearing age. The estimated number of women of childbearing age with lifetime and untreated vaginal fistula in 2016 were 72,533 (95% CI 38,235–124,103) and 31,961 (95% CI 11,596–70,309) respectively. These figures show reduction from the 2005 estimates: 98,098 (95% CI 49,819–170,737) lifetime and 59,114 (95% CI 26,580–118,158) untreated cases of vaginal fistula. The number of districts having more than 200 untreated cases declined drastically from 54 in 2005 to 6 in 2016. Our results show a significant subnational variation in the burden of vaginal fistula. Overall, between 2005 and 2016 there was substantial reduction in the prevalence of vaginal fistula in Ethiopia. Our results help guide local level tracking, planning, spatial targeting of resources and implementation of interventions against vaginal fistula.


Results
General description. Ethiopian Table 1 shows the number of women with vaginal fistula symptoms and the number who have not been treated for this condition by region. There was no significant difference in the median of prevalence of untreated fistula between rural and urban communities across regions neither in 2005 (Wilcoxon's test, p-value = 0.7545) nor in 2016 (Wilcoxon's test, p-value = 0.1843). Neither was there significant difference in the median of prevalence of lifetime fistula between rural and urban communities in 2005 (Wilcoxon's test, p-value = 0.439), although prevalence was found to be higher in rural than urban areas in 2016 (Wilcoxon's test, p-value < 0.05). When we compared the prevalence of lifetime fistula in rural and urban areas between years, we observed no difference in the prevalence in rural areas (Wilcoxon's test, p-value = 0.3381) whereas the prevalence in urban areas declined sharply from 2005 to 2016 (Wilcoxon's test, p-value < 0.05). However, the prevalence of untreated fistula declined significantly both in rural and urban areas (Wilcoxon's test, p-value < 0.05) in 2016. Figs. 1 and 2 shows a heterogeneous spatial distribution and number of cases of lifetime and untreated VF in Ethiopia. The highest prevalence of VF for 2005 is estimated in some parts of Amhara, Oromia, Southern Nations, Nationalities, and Peoples (SNNP) and Tigray regions. The predicted map showed a significant reduction of lifetime and untreated fistula in 2016, although this reduction is spatially heterogeneous.
We estimated the number of WCA who had ever had VF symptoms to be 98,098 (95% CI: 49,819-170,737) in 2005 (Table 4) (Table 5). Oromia, Amhara and SNNP bear the highest burden of WCA who had ever had VF symptoms and women who presently had symptoms of VF. The lowest burden of current symptoms of VF was predicted in Harari, Gambella and Dire Dawa (Tables 4 and 5).
Based on our models, the estimated prevalence of untreated vaginal fistula only exceeded 20 per 1000 childbearing age women in two districts of both Tigray and Amhara regions in 2005, and no single district was    www.nature.com/scientificreports www.nature.com/scientificreports/ estimated to exceed this prevalence in 2016 (Fig. 3). When exploring the burden of vaginal fistula by district, we found that the number of districts that were estimated to have more than 200 cases of untreated vaginal fistula dropped from 54 in 2005 to 6 in 2016. The majority of the districts with untreated cases were located in Amhara, Tigray, SNNP and eastern Oromia Regions (Fig. 4).
In 2016 lifetime VF prevalence of ranged from 2.3 per 1000 in Endegagn district (SNNP Region) to 6.7 per 1000 in Sigmo district (Oromia Region; Supplementary

Model validation.
We provide in a supplementary information (Figs. S9 and S10) the output of our validation procedure for the binomial mixed models as described under methods section. We simulated 500 data set from the fitted model and calculated the mean and standard deviation. We then compared the distribution of these two statistics with the ones observed in the data. For both lifetime and untreated OF models, there is agreement between the simulation from our models and the metrics observed in the empirical data, indicating a good predictive performance of the fitted models.

Discussion
This is the first geostatistical analysis to estimate the burden of VF at district level in a developing country setting. We used large scale survey data in combination with socio-economic, demographic and health related continuous data to obtain these estimates. The point and life time prevalence of VF exhibited considerable heterogeneity across districts. The prevalence of symptoms of vaginal fistula although decreasing, still is considerably high. Districts with high prevalence are clustered in Amhara, Oromia and SNNP Regions. According to our estimates, in 2016, there were 31,961 (95% CI: 11,596-70,309) untreated cases of VF in Ethiopia, with majority of the cases in Oromia, Amhara and SNNP and with reduction from 2005 of 72,533 (95% CI: 38,235-124,103) untreated cases.
The point prevalence of VF, 1.4 per 1000 WCA, is higher than the prevalence reported by Ballard 9 and colleagues (0.2 untreated fistula prevalence per 1000 WCA), and consistent with the 1.5 per 1000 WCA reported by Muleta and colleagues 10 and global estimates for Sub-Saharan (1-1.6 per 1000 WCA) 7,8 . The point prevalence of VF estimates in 2005 was slightly lower than a previous estimation which used the same data sets 7 . The difference between the two estimations is attributed to the definition of treated cases. We assumed all women who reported, "having received treatment", regardless of the outcome of the treatment, as treated cases. The previous analysis considered those women who did not report the outcome of the treatment as untreated cases.
We estimated that there were 31,961 (95% CI: 11,596-70,309) cases of untreated women with vagina fistula symptoms and 72,533 (95% CI: 38,235-124,103) lifetime vaginal fistula in 2016. There is a significant reduction from 2005 figures. The reduction in the prevalence of fistula between 2005 and 2016 could be explained by the introduction of health extension workers (HEW) at the community level, which facilitated identification, and referral of women with symptoms of vaginal fistula. It could also be attributed to the government's ambitious National Action Plan launched in 2014 to eliminate fistula in six years' time 11 .
The results also clearly show that vagina fistula is a ubiquitous and spatially heterogeneous health problem in Ethiopia. There is clear clustering of vagina fistula symptoms in some districts. Districts with high burden of untreated vagina fistula symptoms include Fareta, Ebenat and Dessie Zuria at the Amhara region. The number of untreated VF cases varies between districts from 0 to 291. Arguably, districts in the central Amhara, Oromia, and northern SNNP are the areas, which bear the highest burden of VF. These districts required more attention to reduce the burden of VF in the country. Districts surrounding Hawassa have to some extent access to emergency obstetric care. Thus, it is important to investigate the reason why these districts have high prevalence despite having access to emergency obstetric care. To achieve the elimination of obstetric fistula, the highest burden districts should primarily be targeted to clear the existing backlog of VF symptoms. In addition, investigating district-specific determinants with focus to high burden districts is warranted.
The strength of our analysis include use of large surveys with well powered sample size, covering large geographical areas and population groups and the use of standardize questionnaire for measuring VF. We generated estimates at district level. District level data allows benchmarking of subnational administrative units, which will help identify best and worst performing districts and design tailored interventions for the different scenarios. However, we must acknowledge some limitations of our analysis. First, VF symptoms is a highly stigmatizing condition so that substantial number of women suffering from this condition might have been expelled from their homes due to ostracism by their families and seek refuge in long term care facilities dedicated to patients with fistula. Considering that DHS is a household based survey, the aforementioned scenarios may lead to many www.nature.com/scientificreports www.nature.com/scientificreports/ cases not being recorded and therefore the point prevalence be underestimated 7,12 . Second, the DHS surveys only target WCA for self-reported vaginal fistula, although previous studies have documented low incidence of vaginal fistula among women older than age 50 years and those younger than 15 years 7 . This can also lead to underestimating the prevalence of VF. Finally, measurement bias could be another limitation 13,14 nonetheless, we have accounted for sensitivity and specificity of the DHS questionnaire based on previous studies 7, 13,15 . However, we do not believe that these limitations diminish the overall interpretations, program and policy implications and actions based on our findings.
Our findings also highlight the importance of tracking symptoms of vaginal fistula at granular level for monitoring and programme evaluation. These can be achieved through collection such data through HEW at regular interval or through routine community based surveillance at district level. Our findings have several importance for policy and programming. First, the findings here will be of help in benchmarking district level burden of VF, support monitoring current and future interventions against this condition. Second, identifying districts with high burden of VF will help prioritize target areas for intervention and optimize resource allocation. By addressing district with high burden districts, the country can speed up the elimination of VF. In addition, program planners could identify district level interventions based of the burden of VF.
Third, identification of district with low and high burden VF would open an important research avenue for social and behavioural scientist. Comparative studies on identifying contextual and sociocultural factors that contribute to low and high prevalence of VF will have important implications in addressing the root causes of VF as it has been in other studies [16][17][18] . Finally, we have demonstrated how spatial analysis can be used to estimate geographical disparity of the burden of vaginal fistula across Ethiopia. Similarly, this framework has the potential to be effectively integrated in the national health information system to track burden of other health conditions in the country.
In conclusion, our estimates have proven a sharp reduction of the VF between 2005 and 2016 in Ethiopia 7 . The number of districts with >200 cases of untreated vaginal fistula reduced from 54 in 2005 to 6 in 2016. The difference could be due to the improved health system and maternal health services, the increased availability and accessibility of surgical services and specific efforts to end fistula in the country. Despite the progress made, there are still significant number of untreated cases in the country. Therefore, decentralized efforts targeting high burden districts is required to stop fistula being a public health problem. Elimination of obstetric fistula by 2020 as stipulated in the Health Sector Transformation Plan of Ethiopia 11 is off track and effort should be intensified on prevention interventions and treatment of the existing ones. Strong surveillance system to identify old and new cases of vaginal fistula and linking them to the needed surgical services should be strengthened by focusing on districts with high burden. There is a favourable momentum in Ethiopia to eliminate fistula, identification and referral of women with vaginal fistula is tremendously facilitated by the existence of HEW 19,20 . The huge boost in the number of health centres, trained midwives help manage deliveries and avoid obstructed labour that causes obstetric fistula; expansion of hospitals and trained surgical officers over the years have expanded access to fistula surgery 20,21 . Social integration of those who suffered fistula that addresses the physical, mental, social, legal and psychological needs of women with this condition is also of paramount importance. Administratively, Ethiopia is divided into nine geographical regions and two administrative cities 6,22 . The sample for the EDHS was designed to provide estimates of key indicators for the country, for urban and rural areas separately, and for each of the nine regions and the two administrative cities 6,22 . Each region was stratified into urban and rural areas, yielding 21 sampling strata. Samples of enumeration areas (EAs) were selected independently in each stratum in two stages 6,22 . Implicit stratification and proportional allocation were achieved at each of the lower administrative levels by sorting the sampling frame within each sampling stratum before sample selection, according to administrative units in different levels, and by using a probability proportional to size selection at the first stage of sampling 6,22 .

Methods
In  Supplementary Information; Figs. S3 and S4) 6 . They were selected with probability proportional to EA size (based on the 2007 PHC) and with independent selection in each sampling stratum 6 . All women age 15-49 who were either permanent residents of the selected households or visitors who stayed in the household the night before the survey were eligible to be interviewed 6 .

Measurement of outcomes.
We aimed to estimate the district-level lifetime and untreated vaginal fistula among WCA and to explore changes overtime. The 2005 and 2016 EDHS included a specific thorough questionnaire of 11 questions intended to diagnose vaginal fistula and to ascertain the severity of the condition 6,22 . Women were asked if they had experienced vaginal fistula as defined by 'constant leakage of urine or stool from your vagina during the day and night 6,22 ?' Those who reported suffering from vaginal fistula were asked if they had ever been treated for this condition 6,22 . We assessed two main estimates of prevalence following the approach used by Maheu-Giroux and colleagues 7 : (i) lifetime prevalence of fistula, as the proportion of respondents who reported having ever had symptoms of vaginal fistula, and (ii), point prevalence (untreated fistula) of fistula symptoms by excluding those women who sought treatment for vaginal fistula 7 .

Scientific RepoRtS |
(2020) 10:972 | https://doi.org/10.1038/s41598-020-58036-0 www.nature.com/scientificreports www.nature.com/scientificreports/ Data analysis. Sources of covariates. As it is detailed below, models for lifetime and untreated vaginal fistula were constructed using a set of covariates that may potentially be associated with the risk for vaginal fistula 7,15,23 . We considered covariates that were available as continuous gridded maps for the whole country. From the Spatial Data DHS Program 24 , we obtained modelled surfaces of the following covariates indicated to be associated with a higher risk for vaginal fistula: (i) percentage of literate women 23,25 , (ii) proportion of live births in the five (or three) years preceding the survey delivered at a health facility 4,23,26 , (iii) proportion of women using any modern method of contraception and iv) proportion of women who had a live birth in the five (or three) years preceding the survey who had four or more antenatal care visits 15,23 (Figures 5 S & 6 S, Supplementary Information). These modelled surfaces are available for Ethiopia at a spatial resolution of 5 km × 5 km and have been produced using standardized geostatistical methods, publicly available EDHS data collected in 2016, and a standardized set of covariates 27 .
To account for disparities between rural and urban, we used gridded maps of urban accessibility obtained from the European Commission Joint Research Centre Global Environment Monitoring Unit (JRC) 28 and from the Malaria Atlas project 29,30 for 2000 and 2015, respectively (Fig. S6, Supplementary Information). Accessibility is time (minutes) taken to travel from a grid cell in the map to a city using land-based travel and employing the minimum cost. Cost-distance or impedance was determined by a friction surface that, although generated using different sources of data for the 2000 and 2015 accessibility models, in both cases contained information on the transport network and environmental and political factors that affect travel times between locations. Further details on the generation of this friction surfaces are provided elsewhere 28,30 . We later used the friction surface from 2015 accessibility model to construct a gridded map of distances (kilometres) to the nearest health facility (Fig. S5, Supplementary Information). For this, we used a list of health facilities with their geographic coordinates downloaded from the Humanitarian Data Exchange database 31 . This dataset, provided by the Ethiopian Ministry of Health, was created in 2012 and is regularly updated. The cost-distance tool available at the Spatial Analyst tool from ArcGIS 10.5 (ESRI Inc., Redlands CA, USA) was used to estimate the nearest distance from every cell grid in a 5 km × 5 km resolution gridded map from Ethiopia to the nearest health facility.
Population density estimates were obtained from the WorldPop project 32 for 2005 and 2015. These estimates were used to classify areas as urban (population densities ≥1,000 persons/km 2 , peri-urban (>250 persons/km 2 within a 15 km distance from urban extents edge)or rural areas (250 persons/km 2 and/or >15 km from the urban extents edge) 2,33 (Fig. S7, Supplementary Information).
Finally, night-light emissivity for 2005 and 2013 (most recent available) captured by the Operational Linescan System instrument on board a satellite of the Defence Meteorological Satellite Programme was used as a proxy measure of poverty across Ethiopia 34,35 (Fig. S7, Supplementary Information). This instrument measures visible and infrared radiation emitted at night-time, resulting in remote imagery of lights on the ground. This information has been correlated with gross domestic product in developed countries [35][36][37] and, although far from precise, would provide an indirect measure of poverty in developing countries 38,39 .
Input grids were resampled to a common spatial resolution of 1 square-km using a bilinear interpolation approach and clipped to match the geographic extent of a map of Ethiopia, and eventually aligned to it 35 . Raster manipulation and processing was undertaken using raster package in in the R software environment 35,39,40 and final map layouts created with ArcGIS 10.5 software (ESRI Inc., Redlands CA, USA). We followed a model selection procedure to identify an optimal suite of covariates to include in the fixed effects part of the binomial models. In order to reduce any potential collinearity and confounding effects, we first grouped the variables and use a formal model selection criterion to select one variable within each of the groups 41 (Tables S1 and S2, Supplementary Information). Variables showing some level of correlation or have similar nature (i.e. accessibility and urbanization-related variables) were grouped together. Within each group, we investigated the relationship between the risks for untreated and lifetime vaginal fistula and each potential explanatory variable by fitting univariate models relating the logit of the outcomes to each of the variables 41 (Tables S3 and S4, Supplementary Information). We compared the univariate models in terms of the Akaike Information Criterion (AIC) and selected the model with the lowest value of AIC within each group of variables (Eq. 1). AIC is defined as where θl( ) is the maximum log-likelihood function and k is the number of parameters.
Modelling analysis. We estimated district-level prevalence of lifetime and untreated vaginal fistula for 2005 and 2016 across Ethiopia using a binomial mixed model, accounting for fixed effects (covariates) and random effects, for potential unexplained variability between regions 42 . We chose this method over other analytic approaches based on the absence of spatial structure on the prevalence of lifetime and untreated fistula prevalence, explored by fitting an empirical variogram 42 (Fig. S8, Supplementary Information) where the d j (x i ) are geo-referenced covariates (described previously) and the Z k are independent and identically distributed zero-mean Gaussian variables with variance σ 2 . We fitted the model using the lme4 package in the R software environment 40,42 . Prior to construct the models, observed lifetime and untreated prevalence of vaginal fistula at location level were adjusted by assuming a sensitivity and specificity of fistula questionnaires common to all surveys of 97% and 99% respectively, as suggested by some authors 7,15 . Adjustment was implemented as suggested by Diggle (2011) 43 . Models misspecification and goodness-of-fit were tested by simulating dataset (n = 500) from the fitted models, and eventually comparing summary statistics computed on the simulated data with the empirical ones. As result of good model performance, we expect that our model is able to replicate the characteristics of the observed prevalence.
Fitted models were applied on the selected covariates to produce continuous predictions of lifetime and untreated vaginal fistula prevalence, predicted mean prevalence and corresponding 95% credible intervals, at 1 km 2 spatial resolutions. Gridded maps of both population density and age structure were obtained from the WorldPop project 42,44,45 . We used this gridded population surface for computing the estimates of affected population by pixel by multiplying prevalence of lifetime and untreated vaginal fistula in each square-km area with the corresponding population of women of childbearing age  year-old women) at the same spatial resolution 42