Study design and population
We used individual-level data collected in the 2010 Rwanda Demographic and Health Survey (RDHS) and rainfall data from the Rwanda Meteorological Agency in the Rwanda Ministry of Infrastructure, the Famine Early Warning Systems (FEWS) Network and the US National Oceanic and Atmospheric Administration (NOAA) [16–19].
The RDHS is a cross-sectional survey conducted every 5 years by the Rwandan government with support from Independent Consulting Firm (ICF) International. In the 2010 RDHS, women aged 15 to 49 years answered detailed questions about themselves, their households and all of their children. Of the 13,790 women eligible, 13,671 (99.1 %) responded. The most common reasons for non-response were the woman not being home at the time of interview, the woman being incapacitated or her refusal to participate in the study. The 2010 RDHS used a two-stage sampling process. In the first stage, 492 villages were randomly selected with probability proportional to village size, stratified by district. In the second stage, 26 households from each village were randomly selected. A geographic (GPS) coordinate was collected in each village and randomly geo-displaced by up to 2 km for urban neighborhoods and 5 km for rural villages with one in every 100 rural coordinates displaced up to 10 km.
We obtained daily rainfall data collected at 14 stations across Rwanda between January 2009 and December 2011 from the Rwanda Meteorological Agency; 10-day estimates of the 30-year (1971–2000) long-term average rainfall for grid cells of approximately 30 km2 from the FEWS Network and monthly runoff estimates from NOAA which uses a global hydrological model for 30 km2 grid cells (described in ). We used the GPS data collected in the RDHS to determine the weather station and the 30 km2 grid closest in location and date to each household and assigned each child these rain-associated variables. Because household respondents in the same village were often interviewed over a period ranging from 2 days to 1 week, rainfall variables for children within a village did not vary.
In the 2010 RDHS, mothers were asked for each child under age 5 whether that child experienced one or more episodes of diarrhea during the 2 weeks preceding the survey.
Primary predictors: rainfall variables
We considered four aspects of rainfall that could impact the occurrence of diarrhea: total monthly rainfall, monthly rainfall intensity, runoff water and anomalous rainfall (Table 1). The first two variables, calculated from Rwanda Meteorological Agency data, were the total monthly rainfall, defined as the sum of the daily rainfall in last 30 days prior to the survey and monthly rainfall intensity, defined as the average daily rain for the month compared to the average daily rain for the year. While these two variables are correlated, the first variable is linked to the hypothesis that having more rain matters whereas the second is linked to the hypothesis that having more (or less) rain than usual matters. The third variable, runoff water, was the sum of the runoff from the NOAA database in the month prior to the date of data collection and finally, anomalous rainfall was defined as the rain for last 6 months compared to the long-term rainfall for the same 6 month window for the previous 30 years and calculated using FEWS data.
To facilitate the interpretation of the results by decision and policy makers, rainfall variables were categorized into three levels. For the total monthly rainfall variable, we calculated a mean and standard deviation for the entire country. We classified total monthly rainfall as normal if it fell within one standard deviation of the mean, low if it was less than the mean by more than one standard deviation, and high if it was greater than the mean by more than one standard deviation. We used an identical process for categorizing the three remaining rainfall-related variables.
Potential confounders and effect modifiers: sanitation variables
We considered the following household demographic factors ascertained in the DHS as potential confounders of the rainfall-diarrhea relationship: child’s age, sex, use deworming medication in the last 6 months; mother’s age, employment status, and education level; and household’s urban/rural location, number of under-five children and wealth quintile. We also considered the following sanitation factors, as confounders or effect modifiers: easy access to water, defined as having water on the premises or obtainable within 30 min; quality of drinking water, considered “improved” if water was drawn from a protected spring, public tap, or standpipe and “unimproved” otherwise; quality of toilet, considered “improved” if toilets are pit latrines with a slab or flush toilets connected to a piped sewer system or septic tank and “unimproved” if they are pit latrines without a slab or open pits; stool disposal, considered to be “adequate” if the child used a toilet or latrine or if the fecal matter was placed/rinsed/poured into a toilet or latrine and inadequate otherwise; and shared toilet, considered shared if used by more than just the members of the household.
We assessed the association between socio-demographic variables and diarrhea using Chi-squared tests. We used univariate logistic regression analysis to assess the association between household sanitation and rainfall variables and the outcome, diarrhea. All socio-demographic variables that were significant at the α = 0.1 level were included in the final multivariable logistic regression model. We constructed a multivariable logistic regression model to assess the adjusted impact of the four rainfall variables and diarrhea controlling for variables identified in the previous step. We selected the final model using manual backward stepwise regression. We first considered the rainfall-sanitation interaction terms and then the rainfall variables, removing variables one at a time until all remaining rainfall variables and interaction terms were significant at the α = 0.05 significance level. The analysis was completed in Stata v12 (StataCorpLP; 4905 Lakeway Drive, College Station, TX, USA), using survey commands to account for clustering of data, stratification and unequal probability of sampling.