rhysinward
commented
2 years ago [15:38] Rhys Inward Thanks for pointing out this typing error in the manuscript this will be rectified in the revision. In terms of what family to use either QP or NB there isn't a general consensus to which is better and is ultimately a model selection problem. The variance of a quasi-Poisson model is a linear function of the mean while the variance of a negative binomial model is a quadratic function of the mean. The key difference in using two different models is in how the variance relationships affect the weights in the iteratively weighted least-squares algorithm of fitting models, because variance is a function of the mean it gives different weights to large and small case counts. NB tends to give a higher weighting to smaller case counts than QP. As our data is predominately covers periods of high case counts we would prefer our adjustments to be dominated by those with higher case counts. We also followed the precedent set by Pellis et al 2021 who also used a QP to calculate rt. However, I think it would make the paper stronger if we look at the concordance between to two families so will re-do the national level analysis with a NB and add to the supp.
Seems to be a bit of a preprint code mismatch here with the preprint saying Poisson and the code appears to use a quasipoisson.
https://github.com/fojackson8/COVID19_mapping_epiparams/blob/8df2f04567d80743afd1b5b1ce82d1eb9fb9f2f9/main/double_time.R#L7
Aside from it being useful when code and paper agrees this shouldn't have a great deal of impact I would think. Might be worth justifying why QP vs the more standard NB when overdispersion is suspected