I think the issue is when it is appropriate to use the binomial approximation. You are approximating (1-p)^N = 1-Np for various Ns (p is p_i, the current infection frequency in the population). This only holds for |pN| << 1. Even for p_i = 0.002, N > 100 will be too big and the linear approximation will have large error. I think the probabilities in the tables need to also include an error bar to bound the errors passed on from this approximation.