Predicting Subnational Ebola Virus Disease Epidemic Dynamics from Sociodemographic Indicators

Abstract

Background: The recent Ebola virus disease (EVD) outbreak in West Africa has spread wider than any previous human EVD epidemic. While individual-level risk factors that contribute to the spread of EVD have been studied, the population-level attributes of subnational regions associated with outbreak severity have not yet been considered.

Methods: To investigate the area-level predictors of EVD dynamics, we integrated time series data on cumulative reported cases of EVD from the World Health Organization and covariate data from the Demographic and Health Surveys. We first estimated the early growth rates of epidemics in each second-level administrative district (ADM2) in Guinea, Sierra Leone and Liberia using exponential, logistic and polynomial growth models. We then evaluated how these growth rates, as well as epidemic size within ADM2s, were ecologically associated with several demographic and socio-economic characteristics of the ADM2, using bivariate correlations and multivariable regression models.

Results: The polynomial growth model appeared to best fit the ADM2 epidemic curves, displaying the lowest residual standard error. Each outcome was associated with various regional characteristics in bivariate models, however in stepwise multivariable models only mean education levels were consistently associated with a worse local epidemic.

Discussion: By combining two common methods-estimation of epidemic parameters using mathematical models, and estimation of associations using ecological regression models-we identified some factors predicting rapid and severe EVD epidemics in West African subnational regions. While care should be taken interpreting such results as anything more than correlational, we suggest that our approach of using data sources that were publicly available in advance of the epidemic or in real-time provides an analytic framework that may assist countries in understanding the dynamics of future outbreaks as they occur.

Fig 1. Weekly cumulative time series data at the ADM2 subnational level.

Plots for time since first case in each ADM2. A. Guinea; B. Sierra Leone; C. Liberia.

Fig 2. Polynomial model fit to the first 6 weeks of cumulative time series data at the subnational level for each country.

Time zero is the date of the first reported EVD case for the whole outbreak in Guinea. A. Guinea; B. Sierra Leone; C. Liberia.

Fig 3. Boxplots of dependent and independent variables.

Each plot shows mean, interquartile range and any outlier values; all are measured at the ADM2 subnational level.

Fig 4. Pairwise Spearman correlation plots between outcomes and covariates.

The direction of correlation is indicated by color type (blue: positive; red: negative). The strength of correlation is indicated by color intensity; uncertainty around the estimates is indicated by width of the ovals (wider: more uncertainty). A. Only correlations statistically significant at α = 0.05. B. All correlations. Outcomes: regional growth rate estimated using data from first 15 weeks of epidemic in each region (exponential fit, “ExpGowth”; logistic fit, “LogGrowth”; polynomial fit, “PolyGrowth”); total number of infections reported in each region (“EpidSize”), proportion of whole population infected (“EpidProp”). Socio-economic covariates: wealth index, “Wealth”; average years of education, “Education”; percent of population Christian, “%Christian”; percent of population Muslim, “%Muslim”; percent of population living in urban area, “%Urban”; number of weeks from start of EVD in West Africa to the first recorded case in each region, “StartWeek”; population size, “PopSize”; population density, “PopDensity”; average age, “Age”; percent of population female, “%Female”).

Fig 5. Geographic distribution of covariates across ADM2 units in Guinea, Liberia and Sierra Leone.

A: Mean education in years. B: Proportion of 18-49 year old population female. White regions are those which reported no EVD cases in the study period.