Inferring pandemic growth rates from sequence data

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

Finally, we comment on the issue of why the 95% credibility intervals (CIs) do not overlap the actual or effective samples sizes towards the end of the simulation. Strictly, the skyline plot is an estimate of the harmonic mean of the effective population size over an inter-coalescent interval [5] . When the inter-coalescent periods are large, as they are in the end of the simulation, CIs on the estimate of the harmonic mean of the effective population size need not overlap with the actual value over the whole interval, and non-overlap does not represent a technical problem with the estimation. Figure 4 shows the effects on the BSP of sampling increasingly more epidemiologically linked sequences from a simulated epidemic. Here, we simulate an epidemic with R ¼ 1.35 and k ¼ 1 over 36 generations, with the resulting epidemic incidence curve shown in black. The purple curve (0% randomly sampled) shows the BSP inferred by sampling (one sequence per generation) from just one lineage, that is one direct descendent generation-to-generation. These closely related samples then represent an epidemiologically linked cluster and the resulting BSP is very poor at capturing the changes in the effective population size. In comparison, the green curve (100% randomly sampled) is the BSP inferred by sampling the same simulated population but this time completely randomly (once again one sequence per generation). The intermediate curves represent intermediate degrees of randomly sampled epidemiologically linked sequences. The final effective population size is proportional to the percentage of random sampling and only in the 100 per cent randomly sampled BSP is there continued growth (although small) beyond the last coalescent event (for corresponding LTT plot, see the electronic supplementary material).