Oral rehydration therapies in Senegal, Mali, and Sierra Leone: a spatial analysis of changes over time and implications for policy

Background Oral rehydration solution (ORS) is a simple intervention that can prevent childhood deaths from severe diarrhea and dehydration. In a previous study, we mapped the use of ORS treatment subnationally and found that ORS coverage increased over time, while the use of home-made alternatives or recommended home fluids (RHF) decreased, in many countries. These patterns were particularly striking within Senegal, Mali, and Sierra Leone. It was unclear, however, whether ORS replaced RHF in these locations or if children were left untreated, and if these patterns were associated with health policy changes. Methods We used a Bayesian geostatistical model and data from household surveys to map the percentage of children with diarrhea that received (1) any ORS, (2) only RHF, or (3) no oral rehydration treatment between 2000 and 2018. This approach allowed examination of whether RHF was replaced with ORS before and after interventions, policies, and external events that may have impacted healthcare access. Results We found that RHF was replaced with ORS in most Sierra Leone districts, except those most impacted by the Ebola outbreak. In addition, RHF was replaced in northern but not in southern Mali, and RHF was not replaced anywhere in Senegal. In Senegal, there was no statistical evidence that a national policy promoting ORS use was associated with increases in coverage. In Sierra Leone, ORS coverage increased following a national policy change that abolished health costs for children. Conclusions Children in parts of Mali and Senegal have been left behind during ORS scale-up. Improved messaging on effective diarrhea treatment and/or increased ORS access such as through reducing treatment costs may be needed to prevent child deaths in these areas. Supplementary Information The online version contains supplementary material available at 10.1186/s12916-020-01857-7.

180 181 diarrhea who received any ORS, only RHF, or no ORT, respectively, as binomial count data, . The counts, , 183 probabilities, , predictions from the three child models , and residual terms * are all indexed at a space-time 184 coordinate. The term represents both the annual proportion and the annual probability that an individual child will 185 receive any ORS, only RHF, or no ORT, respectively, given the child resides at that particular location. The logit of temporal error term, ; and an independent error term, . Coefficients, , on the child models represent their 188 respective predictive weighting in the mean logit link and are constrained to sum to one. The country random effect 189 was not used in individual country models, with the exception of India where we set this term to be state-level 190 random effects. , is modelled as a three-dimensional Gaussian process in space-time centered at zero and with a 191 covariance matrix constructed from a Kroenecker product of spatial and temporal covariance kernels. The spatial 192 covariance, Σ space , is modelled using a Matérn covariance function (10), and temporal covariance, Σ time , as an 193 autoregressive order 1 (AR1) function represented in the model with 18 annual knots.

195
This approach leveraged the data's residual correlation structure to more accurately predict coverage estimates for 196 locations with no data, while also propagating the dependence in the data through to uncertainty estimates (11). The 197 posterior distributions were fit using computationally efficient and accurate approximations in R- INLA (12,13) 198 (integrated nested Laplace approximation) with the stochastic partial differential equations (SPDE) (14) 199 approximation to the Gaussian process residuals.

202
The following priors were used: We generated 95% uncertainty intervals around the mean of our estimates by taking the 2.5% and 97.5% quantiles 237 of each of the draws, at the grid-cell or administrative level. The proportion of children that received any ORS and 238 only RHF estimates were adjusted by draw using the following formulas to ensure that they together summed to 1; 239 the proportion of children that received no ORT in each location-year as follows:

256
To explore the in-sample validity of our models, we plotted our predictions vs. the observed data by modelling 257 regions and by year at the country-level, first administrative-level, and second administrative-level aggregations. We 258 also calculated mean error (ME, or bias), root-mean-squared-error (RMSE, which summarizes total variance), and 259 95% coverage of our predictive intervals (the proportion of observed in-sample data that fall within our predicted total of 166 facilities (33).   PRITECH introduced ORS in Senegal in 1985. ORS was donated by USAID and UNICEF, and was  ORS and RHF were introduced in community "health huts" starting in the 1990s. However, ORS was not 369 widely available due to weak distribution systems that caused frequent shortages (18,36,37). Health, Public Hygiene and Prevention. This covered the whole country, but was also accompanied 374 by increases in medicine prices in order to cover supply costs (38).

387
 Zinc was introduced in community "health huts" in 2009. There were more than 1600 of these in Senegal

414
The geospatial modelling process consists of four sections.

472
Estimates correspond to results in Figure 2a,b and Additional file 1: Table S10a.

476
Estimates correspond to results in Figure 3a,b and Additional file 1: Table S10b.

478 479 480
Estimates correspond to results in Figure 4a,b and Additional file 1: Identify and describe any categories of input data that have potentially important biases (e.g., based on characteristics listed in item 5).
Additional file 1: Section 2.0 For data inputs that contribute to the analysis but were not synthesised as part of the study: 7 Describe and give sources for any other data inputs. Manuscript: Methods Additional file 1: Sections 2.0, 3.0

8
Provide all data inputs in a file format from which data can be efficiently extracted (e.g., a spreadsheet rather than a PDF), including all relevant meta-data listed in item 5. For any data inputs that cannot be shared because of ethical or legal reasons, such as thirdparty ownership, provide a contact name or the name of the institution that retains the right to the data.
Global Health Data Exchange (link available upon publication)

9
Provide a conceptual overview of the data analysis method. A diagram may be helpful. Manuscript: Methods Additional file 1: Section 3.0 Additional file 1: Figure S1  10 Provide a detailed description of all steps of the analysis, including mathematical formulae. This description should cover, as relevant, data cleaning, data pre-processing, data adjustments and weighting of data sources, and mathematical or statistical model(s

565
Covariate any ORS only RHF no ORT

Distance to rivers or lakes
the fixed effects: the intercept (int) and the the covariates (gam, gbm, and enet) corresponding to the predicted ensemble rasters. Fitted values for the 569 spatio-temporal field hyperparameters and the precision parameters (inverse variance) for random effects are shown in the bottom four rows. 43 c) Senegal percentages in the format: mean (2.5%-97.5% uncertainty interval). Groups Northern, Central, and Southern correspond to regions shown in Figure 4.

a) Sierra Leone
greater than 95% posterior probability of increase or decrease in each time period. Groups and time periods correspond to tho se presented in Figure 2.