Widely respected hurricane expert Jim Elsner of FSU has posted a lengthy response to these posts over at his blog. I’d encourage interested readers to have a look. This exchange reminds me of a quote attributed to John von Neumann speaking on statistics, “With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.” It also serves as a good reminder that Dan Sarewitz’s notion of an “excess of objectivity” is alive and well even when one is dealing with 34 data points. Let me start by acknowledging that Jim and I are going to agree to disagree and interested readers will have to judge the merits of our arguments themselves.


Elsner argues that the statistics of loss data are best fit by using a “random sum” that combines the statistics of frequency of losses with those of intensity of losses. This approach was first applied to hurricane damage data by my former colleague at NCAR Rick Katz in his 2002 paper “Stochastic modeling of hurricane damage” (in PDF). In my critique of Elsner’s work, I accept that the “random sum” methodology is indeed useful for deconvolving components of a statistical relationship (see, e.g., the acknowledgements in Rick’s paper). As Katz writes in his 2002 paper, “By enabling the variations in total damage to be attributed to either variations in event occurrence or in event damage, the present modeling approach has an inherent advantage over previous analyses.” But such a methodology, or any sophisticated statistics, cannot create a strong relationship in the real world where one does not exist.

I have focused my critique on the intensity part of Elsner’s analysis. With Jim’s help I have successfully replicated this part of their work (Part 4) and I have found that their results are highly unstable — that is they do not hold for 1950-2004 or for 1950-2006. What they report on large losses has much more to do with one event in 2005 (Katrina) than statistical properties of the dataset that are stable over time. On his blog Elsner suggests that the period 1950-2005 “is not intended to stand by itself.” That is good, because it does not stand by itself. Based on the lack of a relationship between SSTs and damage in the subset of data that Elsner claims that there is a strong relationship, I have concluded that there is little reason to expect that Elsner’s model would allow for an accurate prediction of future damage amounts conditional on SST. A question for Jim — What, for instance, would it have predicted for 2006 before the season?

Let me reassert that reasonable people can disagree on such subjects, as I had stated in Part One. Elsner would in my view make a much better case for his arguments by focusing his replies on the substantive questions, such as the obvious lack of stability in his intensity model or what physical basis there exists between May-June SSTs and damage that occurs within the hurricane season (points which he does not address). He is representing his work as “sound science that will likely have a major impact in the reinsurance industry” and indeed he is selling services to these companies. Thus, he should probably expect that his methods will attract attention (and in my experience in academia, attention means that one’s views are worth considering, which should be a compliment, even if the attention is critical as is often the case in academic discussions). If Jim is confident in his approach then he should welcome such scrutiny and efforts to clarify his methods and their significance. Bluster and invective are not only weak means of argumentation, but also make for poor marketing tools.

Let me also once again acknowledge that I did make a mistake in an earlier post, which was corrected online immediately when Jim pointed it out. In response to Jim’s complaints about a lack of apology I posted the following on Jim’s blog:

Jim- Let me once again formally apologize for making a mistake. It happens from time-to-time ;-) It has been corrected, as you know. As I wrote immediately after you brought the data issue to my attention in a personal email to you, “Thanks Jim for following up. Thanks for catching the data sort mix up, apologies for that.”

I’ll follow up on the substance next. Thanks!

In closing it is worth remembering the old adage that if one tortures data sufficiently it will confess. In this case, simple and straightforward analyses of the relationship of SST and hurricane damage without deconvolving intensity or frequency indicates that there is no relationship. Elsner and Faust both show that if you segregate the data in various ways you can use the influence of 2005 to attain, at best, a very marginal relationship. We disagree on whether such a relationship is indeed marginal and also the importance of such a relationship. Fair enough. As 2006 provides an excellent example of, scientists have no ability to predict hurricane landfalls with accuracy, much less frequency or intensity at landfall before the season starts. Until such a capability has been demonstrated, efforts to predict damage with accuracy will in my view amount to little more than statistical data mining.