Quantifying transmission of emerging zoonoses: Using mathematical models to maximize the value of surveillance data

Monique R. Ambrose; Adam J. Kucharski; Pierre Formenty; Jean-Jacques Muyembe-Tamfum; Anne W. Rimoin; James O. Lloyd-Smith

We first validated the inference framework using a simulation study, then applied the 215 Across 125 simulations (25 simulations for each of five parameter sets), estimated values 245 clustered around the true parameter values. The true value for R was included in the 95% 246 credible interval (CI) 119 times (95.2%) and for λ z was included 121 times (96.8%) (Fig 3A) . On 247 average, the posterior mean estimate of R differed from the true value by 8.6%; the analogous 248 percent errors for λ z and σ estimates were 6.3% and 7.0%, respectively (S1 Table) . Inferences were performed A) when the true number of localities under surveillance was known, 254 B) when the true number was unknown and it was assumed that the number of observed 255 localities was the total number of localities, and C) when the true number of localities was 256 unknown and the corrected-denominator method was used to control for the locality observation 257 process. 258 259 However, this method assumes that the true number of localities under surveillance is 260 known. In real-world situations, 'silent' localities that experience zero cases often do not appear 261 in the dataset, resulting in an unknown true number of surveilled localities. We investigated 262 possible biases in parameter estimates that could arise from assuming that the number of 263 localities that reported one or more cases represents the total number of localities under 264 surveillance. To do so, we used the same set of simulated datasets as described above, but 265 removed knowledge about the number of silent localities. In these datasets, silent localities make 266 . CC-BY 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/677021 doi: bioRxiv preprint up between 21% and 85% of all localities under surveillance, with the proportion driven 267 primarily by the spillover rate. Estimates for the reproductive number R were not strongly 268 impacted (95.2% of the 95% CIs contained the true value with an average percent error of 8.4%), 269 but the spillover rate λ z was consistently overestimated (Fig 3B) . The true value for λ z was 270 contained in none of the simulations' 95% CIs and the posterior mean had an average percent 271 error of 153% (S1 Table) . 272