Spatiotemporal variability in case fatality ratios for the 2013-2016 Ebola epidemic in West Africa

Abstract

Background: For the 2013-2016 Ebola epidemic in West Africa, the largest Ebola virus disease (EVD) epidemic to date, we aim to analyse the patient mix in detail to characterise key sources of spatiotemporal heterogeneity in the case fatality ratios (CFR).

Methods: We applied a non-parametric Boosted Regression Trees (BRT) imputation approach for patients with missing survival outcomes and adjusted for model imperfection. Semivariogram analysis and kriging were used to investigate spatiotemporal heterogeneities.

Results: CFR estimates varied significantly between districts and over time over the course of the epidemic. BRT modelling accounted for most of the spatiotemporal variation and interactions in CFR, but moderate spatial autocorrelation remained for distances up to approximately 90 km. Combining district-level CFR estimates and kriged district-level residuals provided the best linear unbiased predicted map of CFR accounting for the both explained and unexplained spatial variation. Temporal autocorrelation was not observed in the district-level residuals from the BRT estimates.

Conclusions: This study provides new insight into the epidemiology of the 2013-2016 West African Ebola epidemic with a view of informing future public health contingency planning, resource allocation and impact assessment. The analytical framework developed in this analysis, coupled with key domain knowledge, could be deployed in real time to support the response to ongoing and future outbreaks.

Keywords: Case fatality ratio; Ebola; Spatiotemporal analysis; Variogram; West Africa.

Figure 1

Choropleth spatial distribution of case fatality ratio (CFR) for 2013-2016 Ebola epidemic in West Africa. a) Median predicted CFR adjusted for imputation (and standard deviation) for each district. b) Median observed district-level CFR. Dark grey shading denotes districts for which data were unavailable. c) Median district-level residuals (i.e. observed CFR minus predicted CFR adjusted for imputation). In each case the median is on the left and the standard deviation (sd) is on the right.

Figure 2

Semivariograms fitted with a Gaussian model to the residuals of district-level CFR adjusted for imputation based on the Boosted Regression Tree model (BRT). The red line is the fitted model for the region as a whole and the blue lines are fitted models for individual countries (Sierra Leone, Guinea and Liberia). Note that the x-axes vary.

Figure 3

Isopleth case fatality ratio (CFR) map for West Africa of a) median and b) standard deviation (sd) of kriged residuals (i.e. observed CFR minus predicted CFR adjusted for imputation). c) Isopleth map for kriged CFR (i.e. predicted CFR adjusted for imputation plus the kriged residuals).

Figure 4

Temporal distribution of case fatality ratio (CFR) for 2013-2016 Ebola epidemic in West Africa. a) Median predicted CFR adjusted for imputation (and 95% CI) for each quarter. b) Median observed quarterly CFR. c) Quarterly residuals (i.e. observed CFR minus predicted CFR adjusted for imputation). Note that there were too few cases in late 2013 and early 2016 to be meaningfully aggregated into quarter.

Figure 5

Semivariograms fitted with a Gaussian model to the residuals of quarter-specific CFR adjusted for imputations based on the Boosted Regression Tree model (BRT). The red line is the fitted model for the region as a whole and the blue lines are fitted models for individual countries (Sierra Leone, Guinea and Liberia). Note that the x-axes vary.