Exploring scenarios of chikungunya mitigation with a data-driven agent-based model of the 2014–2016 outbreak in Colombia

New epidemics of infectious diseases can emerge any time, as illustrated by the emergence of chikungunya virus (CHIKV) and Zika virus (ZIKV) in Latin America. During new epidemics, public health officials face difficult decisions regarding spatial targeting of interventions to optimally allocate limited resources. We used a large-scale, data-driven, agent-based simulation model (ABM) to explore CHIKV mitigation strategies, including strategies based on previous DENV outbreaks. Our model represents CHIKV transmission in a realistic population of Colombia with 45 million individuals in 10.6 million households, schools, and workplaces. Our model uses high-resolution probability maps for the occurrence of the Ae. aegypti mosquito vector to estimate mosquito density in Colombia. We found that vector control in all 521 municipalities with mosquito populations led to 402,940 fewer clinical cases of CHIKV compared to a baseline scenario without intervention. We also explored using data about previous dengue virus (DENV) epidemics to inform CHIKV mitigation strategies. Compared to the baseline scenario, 314,437 fewer cases occurred when we simulated vector control only in 301 municipalities that had previously reported DENV, illustrating the value of available data from previous outbreaks. When varying the implementation parameters for vector control, we found that faster implementation and scale-up of vector control led to the greatest proportionate reduction in cases. Using available data for epidemic simulations can strengthen decision making against new epidemic threats.


71
The IPU algorithm assigned demographic characteristics to households and individuals but did not assign 72 geographical locations. We created a "household locator" algorithm that assigned geographical locations 73 to households according to data on population density and land-use. We created a 1 x 1 km grid across the 74 country (789,116 grid cells that each contained at least one person). Then we assigned each household in 75 a municipality (from the IPU algorithm) to a random grid cell in the urban area within the geographical 76 boundaries of this municipality. As we placed households in grid cells, we ensured that the population 77 density of each cell did not exceed the observed values. For each household within a grid cell, we 78 randomly selected coordinates from available house coordinates in the OpenStreetMap dataset. If no 79 house coordinates were available from OpenStreetMap to assign to a synthetic household, we assigned 80 random coordinates within a 100-meter radius from a road within the grid cell. If a grid cell did not 81 contain any roads, we assigned random coordinates from anywhere within a grid cell. For 30 out of 33 82 departments the spatial correlation coefficient 10 between the population density of our synthetic 83 population and the WorldPop data was >80% (Fig. S7) . Population density in the remaining three 84 departments was very low and did not match as well (these low-density areas did not contribute much to 85 disease transmission).
Ministry of Education on the number of schools per municipality, school grade levels, and school sizes.

90
For each school in a municipality, we randomly selected coordinates from available school coordinates in 91 the OpenStreetMap dataset. If coordinates were not available for every school listed by the Ministry of 92 Education, we randomly assigned the remaining schools to one of five 10 x 10 km grid cells that had the 93 highest number of synthetic students and that were located within the municipality. We assigned random 94 coordinates from within this grid cell to each school for which no coordinates were available in OpenStreetMap dataset (a school served as a workplace for teachers and staff). If coordinates were not 104 available for every workplace, we randomly assigned the remaining workplaces to one of the five 1 x 1 105 km grid cells that had the highest population density and that was located within a municipality. We 106 assigned random coordinates from within these grid cells to each workplace for which no coordinates 107 were available in OpenStreetMap.

112
People designated to attend schools by the IPU algorithm were assigned to specific schools. Children 1-5 113 years were assigned to a pre-school, 6-10 to a primary school, 11-17 to a secondary school, and 17+ to a 114 university. We assigned students to a school within, or outside their municipality according to 115 information about this assigned by the IPU algorithm. We randomly assigned students going to a school 116 within their municipality to one of the five closest schools with availability for the student age and grade.

117
For students going to a school outside their municipality, we randomly assigned them to one of the five 118 closest schools in a neighboring municipality. This algorithm has been used previously to represent 119 student mobility 11 . Students >17 years were assigned randomly to one of the five closest universities 120 located within their department.

123
We assigned employees to workplaces based on commuting times assigned to them by the IPU algorithm, 124 based on census data. The IPU algorithm also assigned employees to be working within or outside of their 125 municipality, based on census data. For each employee working within her municipality, we used her 126 commuting time to determine the distance to her workplace, assuming an average travel speed of 30 127 km/hr. along Euclidean distance. We then drew a circle around her house with the commute distance as 128 radius and randomly assigned the employee to one of five workplaces located within the municipality and 129 closest to the circle. We assigned employees working outside of their municipality to a random workplace To enable travel between departments, we swapped the department of the school or workplace for 5% of 134 students and employees that lived close to a department border. i.e., for each department, we randomly 135 selected a 5% sample of all students and employees that lived within 10 km of a border. For each of these 136 students and employees, we also randomly selected a "partner student or employee" from the neighboring 137 department, also living within 10 km of the border. We then swapped the school or workplace between 138 these partners.

140
3. Mosquitos 141 142 The spatial occurrence of mosquitoes was determined by their probability of occurrence computed from   probability of transmission between mosquitos and humans also depended on the mosquito biting rate.
We assumed an average mosquito biting rate * of 0.5/day/mosquito 19,20 . An infectious human could infect a mosquito, when visiting a location with mosquitos and when bitten, with an infection rate * of 176 0.876/day (calibrated). We assumed that this infection rate is the same for symptomatic and asymptomatic 177 individuals as other similar models have assumed 21 . Similarly, infectious mosquitos would infect 178 susceptible humans who visited their location with an infection rate 2 of 0.196/day (calibrated) (Fig. 1).

179
The probability of a human being bitten by an infectious mosquito in a location depended on the density  municipalities that also fitted their observed data (Fig. S1). We instantiated two vector-control parameters

241
We conducted a sensitivity analysis to better understand how the number of pupae per person (ρ, default 242 1.02) and temperatures affected the simulated CHIKV epidemic curve. We simulated the CHIKV 243 epidemic for the cities of Santa Marta and Riohacha, for scenarios with and without vector control 244 intervention, while ranging the pupae per person from 50% to 150% of the default value in 10% 245 increments. We conducted 20 simulations for each scenario and reported the average epidemic curve for 246 each scenario (Fig. S3). Similarly, we simulated the epidemic using each of the monthly temperature 247 grids instead of the average annual temperature. Monthly temperatures varied from 95% to 103% of the 248 default average annual value. We also conducted 20 simulations for each temperature scenario and 249 reported the average curve for each (Fig. S4).

253
At the end of our study, additional data on the CHIKV epidemic became available. We used CHIKV case 254 counts reported until week three of 2016 to test the model fit for the entire epidemic period. We compared 255 the aggregate simulated case counts for each of the six regions in Colombia with the observed data and 256 found a good fit for every region except Caribe and Insular (Fig. 4)       Decided by authors