“Your analyzes are flawed. I believe your mistake was to align your SST values in chronological order with your damage estimates in descending order.

When you correct this you find a positive correlation between your damage estimates and your Atlantic SSTs and marginally significant for the largest loss events. This is consistent with our paper despite the differences in SST values (we use the AMO) and damage estimates (we use Collins and Lowe).”

I apologize for the mix up, and have successfully replicated Jim’s analysis in Part 4. I apologize for any confusion to Jim and our readers.

Thanks!

You make an excellent point. I agree that there are real problems in trying to detect the role of anthropogenic climate change in loss data because, IMO, the losses seem to be much more sensitive to human adaptation and maladaptation – (if I may use the latter term to characterize the Hurricane Katrina snafu, and policies that subsidize insurance in areas with greater-than-normal risk, for example).

I am, however, not sure that it is valid to exclude events (or years, for that matter) based on thresholds for the lower limit or higher limit. Was that done so that the distribution met some criterion necessary to enable statistical analysis? If so, that would not be justification to toss data out. The real world is what it is, and there’s nothing that says events necessarily have to be normally distributed (or whatever) or make it easy to analyze.

I would view “adjusting” data to account for abnormal events as a different kettle of fish. I think it would be appropriate for a hurricane-related analysis to adjust costs of Katrina, for instance, downwards, because much of that was due to a flood caused by the failure of the levees, arguably, due to human failure. However, to make such an adjustment, one should be able to somehow or the other identify the magnitude of the problem that was, indeed, caused by the flood/failure. I don’t think it is appropriate to either toss out that point or to use it unadjusted.

Thanks for your comments.

Note that I am simply trying to replicate Elsner’s results using a normalized loss database that we have developed, and which is far simpler (and more transparent) than the dataset Elsner is using. I cannot replicate his results.

This is a separate question than as to whether his approach is valid for detecting the role of anthropogenic climate change in loss data (which i do have questions about). Given that I cannot even come close to reproducing his results, and he has (thus far) not shared his data, it doesn’t make sense to dig deeper into the theoretical basis for the approach.

Thanks!

Feel free to prove me a moron, but Elsner appears to have used a time series approach and applied OLS.

As any regular reader of Climate Audit will know this opens up a massive can of worms due to the autocrrelation involved in these statistics (both the temperature series and the damages most likely).

Net result is biased estimators and unreliable results.

Roger’s analysis, which is cross sectional, would not suffer from that problem (although you would need to be aware of potential heteroschadicity) and hence is not susceptible to the problems inherent in Elsner’s approach.

I would think (from a full 5 minutes study) that Roger’s results and conclusions are more robust.

Your analyzes are flawed. I believe your mistake was to align your SST values in chronological order with your damage estimates in descending order.

When you correct this you find a positive correlation between your damage estimates and your Atlantic SSTs and marginally significant for the largest loss events. This is consistent with our paper despite the differences in SST values (we use the AMO) and damage estimates (we use Collins and Lowe).

I repeat my original point. A more compelling argument is made when the analysis and modeling considers the number of loss events separate from the amount of loss as done in Jagger et al. (2007) and Katz (2002). In this way we find that the NAO is associated with the number of loss events (as you might expect through storms steered toward or away from the US) while the SST influences the amount of loss conditional on a loss event (as you might expect through storm intensity).

If you are interested in the data values we used in our paper, please email coauthor Mark Saunders.

Best,
Jim

I reran the analysis with a $250M/storm threshold and the results are the same — no relationship. A linear regression returns an r^2 of 0.015.

It is presumably the independent variable, after all, not the dependent variable