I think you are both partially correct, but neither analysis is compelling.

The important limitation is the failure to disaggregate the loss data. A large annual loss can be the result of many small losses or one big loss, and the number of loss events should be independent of the magnitude of the event.

We did a study that will be published soon looking at the predictability of insured losses from US hurricanes (Atlantic basin, only). It is available at http://garnet.fsu.edu/~jelsner/www under Research. The bottom line is that preseason Atlantic SST influences the expected annual loss given an event, but not the probability of an event itself. Moreover, extreme annual losses are conditional on preseason values of the NAO and ENSO, but not SST.

Although we are using preseason covariate information (as opposed to contemporaneous), I think our analysis is the more compelling one.

Best,
Jim

just being an amateur in climate science and statistics I’m interested in where the normalized data used in these graphs could be obtained. I’d say this data would be a requirement for any assessment.

Some questions remain (at least for me):

1) The handling of zero losses is somehow unsatisfactorily. Binning well below average losses as zero should correspond to binning well above average losses as a corresponding maximum as well. I feel this comes more close to the rank test you’ve applied.

2) How does regional restriction of Hadley SSTs (10N…20N…) influence the results?

3) How sensitive are the results with respect to the applied normalisation (-formula) of losses?

4) How sensitive are the results with respect to the acquisition of losses in earlier times?

If there was ever a case where Ockham’s razor applied, this would be it. Your analysis appears straight forward and reasonable. The Faust analysis seems almost tortured by comparison, as if he were trying to find some way to arrange the data to get the results he wanted.

Furthermore, the conclusion “…that if the increases in tropical Atlantic SSTs were to continue in the long term due to anthropogenic climate change, we would have to expect a shift towards hurricane loss distributions with ever increasing high-loss portions” is not supported. Such a statement falsely assumes that increasing GHGs would only impact tropical SSTs, leaving all other factors governing tropical development and storm tracks unchanged.

Finally, the evidence strongly suggests that SST changes have such a minor role (if any) on US damages, that the difference has already been exceeded by improvements in building design and materials, even without the implementation of stricter building codes. Or put another way, if the cost of switching from roofing nails to roofing screws was the same as the cost of preventing any additional rise in GHGs, switching to roofing screws would be the better investment for hurricane damage mitigation!

1. The red line in 5.2.8 shows the moving average over a moving 0.2 degree window. Note that the relationship appears to break down outside of -0.4 to +0.4. You are absolutely correct that there are very few degrees of freedom shown here.

2. Figure 5.2.8 presents the same analysis as 5.2.9, although with (a) untransformed damage data (it is not normally distributed), (b) a moving average rather than a simple regression, and (c) excluding some of the data points. In my view all of (a), (b) and (c) are statstically improper ways to treat the data in this situation.

Figure 5.2.9 presents results that are contrary to 5.2.8 (and no p-value is reported). This suggests that the conclusions drawn from 5.2.9 are a function of the data handling/binning, not a significant relationship of damage and SST.

Thanks!

Thanks. A few replies.

1. See the first paragraph of my section. One should always be cautious about relationships found when “binning” data — especially when the original (non-binned) relationships don’t exist!

2. What are the p-values (statistical significance) for such a “trend”? (Hint see figures 5.2.7a and b as discussed on p. 548). There is no statistically significance.

3. There is some arbitriness in such binning. For instance, 1926 the largest loss year falls right on the border of warm/cold phases. Rather than binning the data in this manner, which may or may not make sense, take a look at 5.2.7a and b. Where is the relationship?

Thanks!

A couple of tecnical questions remain unclear to me.
First, the red curve in Figure 5.2.8 is based on the count of damage losses in running 0.2C-windows in the range -0.4C to 0.4 C? Does it therefore contain just 4 degrees of freedom?

Other question that I think is relevant
here is that statistical significance is not sufficient. The amount of explained variance is the critical number, and from Figure 5.2.8 seems to be quite low.