The study locations, data sources, spatial covariates, and statistical analyses for this study are consistent with those employed in our previous study [6], with a few modifications. First, we restricted the analysis to just three countries: Senegal, Sierra Leone, and Mali. We selected these countries because each (1) had experienced changes in ORS and/or RHF during the study period [6], (2) had implemented policies or interventions aimed at improving ORS coverage during the study period (see the “Analysis of policy changes” section below), and (3) had at least 6 years of survey data available. Second, we created mutually exclusive and collectively exhaustive ORT indicators such that each child was placed into exactly one of three categories (any ORS, only RHF, and no ORT). Third, we ran models separately for each country and restricted models to run from the first to the last year of study data. These methods are described below, with further details in Additional file 1 [9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26].
Survey data
We collected data from national, population-based household surveys—including Demographic Health Surveys and UNICEF Multiple Indicator Surveys—in which primary sampling units (PSU) could be geo-located below the country level. We included data from child modules where mothers were asked whether their children under 5 years of age had diarrhea in the past 2 weeks and, if so, whether they received ORS and/or RHF. In these surveys, diarrhea is defined as three or more loose or watery stools in a 24-h period, which corresponds to acute diarrhea (henceforth referred to as “diarrhea”). In total, we included six surveys in Sierra Leone (years 2000–2017), six surveys in Mali (2001–2018), and nine surveys in Senegal (2000–2017) (Additional file 1: Table S2).
Treatment categories
We determined the percentage of children with diarrhea in each PSU that received (1) any ORS treatment (“any ORS,” which included treatment with only ORS, or with ORS and RHF), (2) only RHF treatment (“only RHF”), or (3) no oral rehydration treatment (“no ORT,” not treated with ORS or RHF) by taking the population-weighted mean of all children sampled within that PSU.
We employed methods from our previous study to adjust for differences in RHF definitions and survey questions from 2000 to 2018 [6]. In brief, RHF questions were classified as including options of (1) recommended or acceptable home fluids, (2) sugar and salt solutions, (3) other home fluids, and/or (4) other liquid foods. We fitted a logistic regression model to surveys across all LMICs, regressing coverage on definition, country-level fixed effects, and a natural cubic spline on survey year. The global adjustment factor for each non-standard definition (including options other than “recommended or acceptable home fluids”) was the coefficient of the fixed effect for that definition. Coverage reported by non-standard surveys was multiplied by the adjustment factor in logit space (Additional file 1: Tables S3 and S4).
Statistical analyses
Analyses were carried out using R version 3.5.0. Coverage of any ORS, only RHF, and no ORT were modeled independently using the Bayesian model-based geostatistical framework from our previous ORT study [6]. Similar methodological details are available from additional previous mapping work [27,28,29]. Briefly, this framework used a hierarchical logistic regression model to predict coverage on a continuous surface, assuming similar coverage in locations closer together in space and time and with similar covariate patterns. Potential non-linear relationships between covariates (Additional file 1: Section 2.4) and coverage were incorporated through stacked generalization [30]. Posterior distributions of all model parameters and hyperparameters were estimated using R-INLA version 19.05.30.9000.30 [31]. Coverage estimates were obtained by taking 1000 draws from the posterior distribution and were adjusted by draw to ensure that the three categories summed to 1 in each location-year:
$$ \mathrm{Any}\ {\mathrm{ORS}}_{\mathrm{adjusted}}=\frac{\mathrm{Any}\ \mathrm{ORS}}{\mathrm{Any}\ \mathrm{ORS}+\mathrm{only}\ \mathrm{RHF}}\times \left(1-\mathrm{no}\ \mathrm{ORT}\right) $$
$$ \mathrm{Only}\ {\mathrm{RHF}}_{\mathrm{adjusted}}=\frac{\mathrm{Only}\ \mathrm{RHF}}{\mathrm{Only}\ \mathrm{RHF}+\mathrm{any}\ \mathrm{ORS}}\times \left(1-\mathrm{no}\ \mathrm{ORT}\right) $$
We calculated population-weighted aggregations of the 1000 draws by country and second administrative-level unit. Mean estimates are reported with 95% uncertainty intervals, which represent the 2.5th and 97.5th percentiles of the 1000 draws. All estimates are reported in Additional file 1: Tables S9–S11 and Additional file 2, with corresponding maps of mean estimates in Additional file 1: Figure S5–S6.
Analysis of policy changes
We conducted a non-systematic literature search to identify changes in health policies, interventions, or events that may have impacted ORS scale-up within each country during the study period. We initially searched Google and PubMed with broad terms such as “Senegal” and “oral rehydration,” or “Senegal” and “diarrhea treatment.” We reviewed research articles, policy reports, news articles, and their references. Based on the initial findings, we modified and expanded our initial search until we exhausted the information we could find.
We synthesized and discussed the findings with in-country experts, whom we identified through contacts in the Global Burden of Disease Collaborator Network. We created a report for each country that included (1) study methodology, (2) summary of key dates and data sources, and (3) preliminary results. In each interview, we (1) discussed the content of the report, (2) answered any questions they had, and (3) asked questions about whether there were key events and/or additional context surrounding changes in diarrhea treatment and access to healthcare at national or subnational levels that should be taken into account.
For the final analysis, we focused on two major events in each country, which are described below with additional details on all policies, programs, and events in the study period in Additional file 1: Section 5.0. The time periods for analysis were selected based on (1) the date(s) that the policy, intervention, or event occurred; (2) years for which we had data available to inform the estimates; and (3) deliberation among co-authors. For ease of interpretation, we set the time periods to start and end at the mid-point of each year (i.e., July 2).
In Sierra Leone, key events were a national policy implemented in 2010 to make healthcare free for pregnant women, new mothers, and children [32], and the Ebola outbreak from 2014 to 2016 [33] (Additional file 1: Section 5.1). We examined ORT changes during three periods: (1) July 2, 2000–July 1, 2009, before policy changes; (2) July 2, 2009–July 1, 2013, comprising policy changes; and (3) July 2, 2013–July 1, 2017, comprising the outbreak.
In Mali, key events were interventions in southern Mali between 2003 and 2004 that introduced ORS and zinc therapy in Bougouni [34, 35] and established mutual health organizations in Bla and Sikasso [36], and the war in North Mali in 2012 [37] (Additional file 1: Section 5.2). We examined ORT changes from (1) July 2, 2001–July 1, 2004, during intervention implementation; (2) July 2, 2004–July 1, 2011, following the interventions; and (3) July 2, 2011–July 1, 2018, comprising the war.
In Senegal, key events were a national policy implemented in 2008 to promote combined ORS and zinc treatment for diarrhea [38], and a national ORS and zinc intervention launched in 2012 [39, 40] (Additional file 1: Section 5.3). We examined ORT changes from (1) July 2, 2000–July 1, 2006, before policy changes; (2) July 2, 2006–July 1, 2012, comprising policy changes; and (3) July 2, 2012–July 1, 2017, during the intervention.
Analysis of changes over time
We analyzed changes over time in three distinct ways within each country. First, we described the absolute change in coverage of ORS and RHF within each time period (Additional file 1: Table S10). We compared results by location and year using mean and 95% uncertainty intervals. Second, we examined the annual rate of change in coverage (Additional file 1: Table S11). We calculated these rates at the draw-level to assess whether or not there was a high probability (posterior probability > 95%) that coverage increased or decreased within each second administrative-level unit and time period.
Finally, we examined whether or not the use of RHF was replaced with the use of ORS. RHF was considered not replaced when more than 95% of draws (posterior probability > 95%) showed that decreases in “only RHF” were greater than increases in “any ORS” (i.e., the percent of children that received no ORT increased). RHF was considered replaced when more than 95% of draws showed that the percent of children that received no ORT decreased.
Numbers of untreated children
Numbers of untreated children with diarrhea were draw-level estimates of treatment coverage at the second administrative-level unit multiplied by mean childhood diarrhea prevalence [7] and by the number of children under the age of 5 [41]. Numbers of children that did not receive any ORS in 2017/2018, but would have received some form of ORT in 2000/2001, were calculated by multiplying the number untreated with ORS in 2017/2018 by ORT coverage (or, 1 – “no ORT”) in 2000/2001.