Optimizing the location of vaccination sites to stop a zoonotic epidemic

Abstract

Mass vaccinations are crucial public health interventions for curbing infectious diseases. Canine rabies control relies on mass dog vaccination campaigns (MDVCs) that are held annually across the globe. Dog owners must bring their pets to fixed vaccination sites, but sometimes target coverage is not achieved due to low participation. Travel distance to vaccination sites is an important barrier to participation. We aimed to increase MDVC participation in silico by optimally placing fixed-point vaccination locations. We quantified participation probability based on walking distance to the nearest vaccination site using regression models fit to participation data collected over 4 years. We used computational recursive interchange techniques to optimally place fixed-point vaccination sites and compared predicted participation with these optimally placed vaccination sites to actual locations used in previous campaigns. Algorithms that minimized average walking distance or maximized expected participation provided the best solutions. Optimal vaccination placement is expected to increase participation by 7% and improve spatial evenness of coverage, resulting in fewer under-vaccinated pockets. However, unevenness in workload across sites remained. Our data-driven algorithm optimally places limited resources to increase overall vaccination participation and equity. Field evaluations are essential to assess effectiveness and evaluate potentially longer waiting queues resulting from increased participation.

Figure 1

Alto Selva Alegre district with possible vaccination locations. The potential sites of fixed-location vaccination tents are depicted by white triangles. The optimization algorithms select the optimal locations among these possibilities. The map was created using R package ggmap version 4.0.0 (https://cran.r-project.org/web/packages/ggmap/).

Figure 2

Regression models of the effect of distance on vaccination coverage. Regression curves are shown for the fixed-effects (left) and mixed-effects (right) models used to estimate the relationship between walking distance to the closest vaccination site and MDVC participation probability. Historical vaccination coverage data (colored dots) are visualized using 30 m binned distances, where dots are colored by year, scaled by the number of houses per bin, and plotted as the mean distance for all houses within a bin vs. the proportion of houses that participated in the MDVC for that bin. *Poisson and negative binomial regression curves are shown with a single line because coefficients were nearly identical for both fixed- and mixed-effects models.

Figure 3

Vaccination point selection method comparison. Vaccination tent locations (white triangles) and subsequent catchment areas (different colored regions) are mapped based on different tent selection algorithms: A convenience, B p-center: minimized maximal walking distance, C p-median: minimized mean walking distance, and D p-probability: maximized participation probability. The middle-row histograms depict the distribution of distance to the closest vaccination point for the E convenience, F p-center G p-median, and H p-probability algorithms. The bottom-row histograms depict the corresponding workload distributions for the tents selected under the different algorithms I–L. The maps in panels A-D were created using R package ggmap version 4.0.0 (https://cran.r-project.org/web/packages/ggmap/).

Figure 4

Predicted vaccination campaign participation. Panel A shows tent locations (white triangles) used in the 2016 MDVC, while panel B shows the optimized placement of tents obtained using the p-probability method. Houses (colored dots) are shaded according to their probability of participating in the MDVC, which was determined using our mixed-effects Poisson regression function with the random-effects coefficient for 2016 that related participation probability to distance to the nearest vaccination tent. These maps were created using R package ggmap version 4.0.0 (https://cran.r-project.org/web/packages/ggmap/).