MDS vs BMDS

[trvrb]

The authors say, "If a single map is desired, then the MDS method may be preferred while the BMDS method provides flexibility, allowing other sources of information, such as genetic data, to be more easily incorporated." Is this the first part of this sentence true? The BMDS method introduced in this paper is such a clear improvement on the original method in the 2004 Smith et al Science paper that this new approach should always be preferred. I don't see any difficulties with summarizing the posterior of the BMDS as done in this paper. It might be possible to make some non-Bayesian (such as maximum-likelihood version) of the method in this paper that included genetic data / dates / virus+serum effects, and then argue that such a maximum-likelihood version could be preferred to the BMDS method here. But if the comparison is simply with the 2004 Smith et al Science approach (which only uses HI data without dates / phylogeny / virus+serum effects), then the current approach is clearly better due to the inclusion of additional data, regardless of one's philosophical preference for Bayesian versus maximum likelihood.

[trvrb]

We agree that we may have been too hesitant in promoting the advances included in the BMDS method. To clarify this point, we've separated the comparison of ML vs Bayes from the comparison of the 2004 Smith et al. cartographic model and the new model that includes virus effects, serum effects and evolutionary priors. MDS vs BMDS is in some ways a matter of preference; MDS will be faster and have the convenience of giving a single map to work with, while BMDS will be slower to produce a full posterior sample and this sample will be less convenient to work with than a single set of points. We did find, however, the model extensions to be quite useful in the analysis of this data.