Hi Suz! Thanks for the comments and questions. Regarding your question about the negative scale parameter: That upper bound is sensitive to the threshold choice in two ways. The first: if I choose a threshold that is too low and include many storms, the distribution fails to fit an extreme value and the upper bound tends to be far too conservative (meaning, it is too low for what the data suggest - it will provide too low return levels for a specified return period, when, using a strictly empirical model, the return level should be higher. Second, if I choose a threshold that is too high, the model fails to compute and provides me with very low xi levels (below -3) and completely out-of-the-question return levels (100 m/s every 5 years for example). I'm considering writing a paper that quantifies that very thought - how sensitive is the xi parameter to the threshold choice. For this and other papers, I have it set at the 50th percentile so the theory is comparable from grid to grid and I can consider the idea of "relative risk" at each location. But a paper assessing the sensitivity to xi could be very enlightening about this methodology. I do agree that a connection could be made between MPI and this statistical upper bound. Perhaps a limiting mechanism from the MPI could be used as the limiting physical bound on the statistical model. Interesting, and great idea! I think that you and I should explore this idea further for a potential future project! And for your second question, I truly do not believe the model is sensitive to the undercount bias. At least for the locations I have tested. I have run the model for various locations using both the entirety of the data as well as the satellite record (1980 on) and at the locations where I retained enough data to model, it did not present a noticeable bias. If you would like, I can send you my RP paper from JAMC for Florida cities. It has been a little while, but I'm pretty certain I have results in there for Miami for data beginning in 1900 vs. 1980.

Thanks for the reply, Jill. I actually meant subsetting the data in the more traditional, non-GPD sense of withholding half (or some proportion of years) for validation. On one I imagine that kind of analysis is tricky when you're trying to get at tail behavior and you just halved the amount of data available for even defining what the tail means! But on the other hand you did something like that to check sensitivity to undercount bias in the early record and saw little sensitivity-- which is a really nice feature of this kind of modeling for the hurricane winds problem. I suppose the really big storms with extreme “tail” winds probably weren’t the ones that went uncounted in the early record.

The issue you bring up about sensitivity of return levels to threshold choice definitely seems like a critical one to explore further, given the implied range in return levels you describe. I'm looking forward to the paper you have planned on this! I agree that connecting the statistical upper bound with MPI would make a pretty cool project-- let's be in touch about this!

Hi Angel! Thank you! SSTs are averages taken from the 2X2 degree grid points in each hexagon from 1854-2010 for the months of August, September, and October. Each hexagon has one SST value related to the latent parameters. By domain, do you mean the size of the hexagons themselves? It didn't make it on to the poster, but some of the work I have done on this dealt with the modifiable areal unit problem and spatial autocorrelation (both in this larger region, as well as focusing on the Gulf coast more specifically). MAUP has never shown to be a problem, though I cannot make the hexagons too small or I'll run out of data to model. I have tried them in multiple sizes and have never been presented with anything that suggests an aggregation bias with the way I have currently done it. The latent variables are autocorrelated (all except the scale parameter) which makes physical sense, since a hurricane passing through multiple grids will have similar intensity values from one grid to the next (autocorrelation decreases the most nearest the coast - again makes physical sense). The xi parameter is not autocorrelated spatially and this is due to the sensitivity to the most extreme events. All of that being said, the results are consistent even if I change the domain size of the hexagons. I hope this answered your question (I tend to ramble - let me know if it needs to have more direction).-Jill

Hi Vasu, Thanks for your comment. Yes, that naivety is reasonable, as both are approaches to understanding hurricane limiting intensity. That is a fantastic idea and one I will explore further. I simply have not used OHC as a predictor of the latent variables, but based on these papers you are referring to, it might be the more appropriate choice. I'll make sure to look up the papers. Thank you!

Hi Christian,

Thank you for your question. I have not assessed those other variables, but you may have noted that one of the other commenters also thought OHC would be a great idea. I appreciate your recommendations and I look forward to incorporating those into my model as I move forward with my research.