Inferring pandemic growth rates from sequence data

de Silva, Eric; Ferguson, Neil M.; Fraser, Christophe

, and then selected one representative simulation for each set of parameters which gave a realized epidemic growth rate r (measured using least-squares fit to log incidence) which matched that expected analytically in the infinite time (i.e. deterministic) limit. This selection step aimed to limit the impact of demographic stochasticity on incidence curves at the start of the simulated epidemic. The three resulting datasets had real-time growth rates of r % 0.05, r % 0.13 and r % 0.2, respectively. For each of these simulated datasets, sequences were randomly sampled per generation log-proportionally, and the BSPs were estimated from these sequences. This random sampling was then repeated four more times (on the same original simulated epidemics) so that five sampled sequence sets were generated for each simulation (table 3; we are not interested here in exploring between-simulation variation). We analysed these datasets with BEAST using the same non-parametric model for the growth of N e as earlier, and then estimated the growth rate by fitting an exponential curve to the portion of the BSP that showed near exponential growth (truncating the curve following the last coalescent event). The resulting non-parametric estimates of the growth rates along with CIs for the parametric estimates are shown in table 3. The same samples were also analysed using BEAST to estimate an exponential growth rate for a parametric coalescent model with assumed exponentially growing effective population size. Figure 5a and table 3 show the resulting estimates of growth rates against the true values for the growth rate for each of the sampled sets. Even after correcting for the artificial flattening of the BSP, the exponential parametric model appears to be more accurate at estimating growth rates. The larger spread in estimated growth rates for higher simulated growth rate is probably a result of the larger truncation of BSPs owing to earlier flattening of LTT corresponding plots, and the corresponding loss of information on coalescence in the phylogeny.