Slower decay of landfalling Hurricanes in a warmer world — really?

Reposted from Dr. Judith Curry’s Climate Etc.

Posted on November 17, 2020 by curryja

by Frank Bosse

A recent paper published in “Nature” made some excitement in the media, see here or here.

In the paper by Li & Chakraborty (L&C 2020 thereafter), the authors find a statistically significant increase of the decay time when a North Atlantic hurricane makes a landfall due to warmer SST in a warming environment. They also undertake some model-related research about the impact of this observations.

The key point of thepaper is the finding that warmer SSTs lengthen the decay time of hurricanes after landfalls.

In L&C 2020, this is shown by figure 1f:

Fig.1: The reproduction of Fig.1 f in L&C 2020. The ordinate reflects the decay time τ in hours.

In the legend the authors state: “We note that the τ time series echoes the SST time series with Pearson correlation r = 0.73.”

The authors describe the way they found the relation, which declares an increase of the decay time of more than 40 hours per 1K SST increase:

“We average τ for all the landfall events in a given year and apply a 3-year smoothing, twice in a row, to this time series.”

They made a regression with strongly smoothed time series, a procedure that is normally frowned on.

In the supplementary data (freely available) one can download an Excel sheet where the raw data used can be found.

For the deduction of the increasing τ with increasing SST (in the area 10°N…35°N ; 100°W…75°W , the authors take advantage of the data for 71 landfalling hurricanes during 1967 to 2018.

The SST for each event is determined as follows:

“We average the SST in time over the hurricane season, June–November, …”

The result “R=0.73”, see Fig.1, of the linear regression implies that 53% of the variation of τ comes from the variation of the SST.

I had a look at the raw data and a few questions arose:

*The use of the average SST of the whole hurricane season for a single event?*

The actual named hurricane develops over a few days in an actual environment, not the average SST of the actual whole season. It makes a difference if the landfall happens during July or November, the average SST difference is 1.8 K in this case, which is much more than the range of the abscissa in Fig.1 .

The use of the seasonal average SST for all hurricanes during that season, rather than the actual SST applicable to each hurricane, has the potential to produce highly misleading results. The average SST applicable to each hurricane might have little relationship with the average SST during the whole season

*The use of the average of all τ in a year if more than one hurricane is involved*?

Every hurricane event is a discrete event. In the raw data many years have only one hurricane per year, these events are not averaged of course.

*Applying a double 3 years smoothing before making the regression shown in Fig.1. ?*

The authors state:

“this approach lessens the effects of non-climatic factors and random noise”.

However, the whole research is about the point:”To what degree impact the actual SST the decay time of landfalling hurricanes?” There will be some other influences and it’s not appropriate to smooth over several years partly out to elicit a strong climate related signal. Applying a 3-year smoothing to both decay time and SST data twice in a row is unjustifiable.

I decided to recalculate the regression shown in Fig.1 but I used the actual SST for every hurricane from the monthly ERSSTv5 data for the described area. I included every hurricane because this is the physical approach: It’s not justified at all to use an average in some years and in some years not, as that gives radically different weightings to hurricanes depending on how many are included in the raw data in each year.

I also use the unsmoothed data to avoid spurious correlations due to the applied smoothing.

This is the result:

Fig.2: The regression of the decay time on the SST without the data-preconditioning in L&C 2020.

There is only a tiny non significant trend in the raw data- p=0.1, so the slope does not reach the standard 95% confidence level.

On twitter Ryan Maue questioned the raw data selection; that issue is beyond the scope of this post.

The outcome of L&C 2020 is very overconfident when it comes to the dependency of the decay time on the SST. The R²=0.53 found in LC 2020 vanishes to an insignificant 0.04 if one uses the physical data, without the applying of unjustified averages and smoothing actions prior to the regression.

This means:

The SST impact on the decay time is negligible, other influences accounting for almost all variability in the decay time.

The peer review process of “Nature” for L&C 2020 lasted more than 8 months, it makes wonder if there was no reviewer with some fundamental skills in statistics involved.

However, this must be the case unfortunately: In the “methods-statistical significance” section the authors mention a test for autocorrelation and there is written: “(which we test using the Dublin–Watson test)”. This must be a typo, the name of the test is “**Durbin**– Watson”.

One should hope that the peer review process of “Nature” would be improved soon to avoid overconfident, obviously flawed papers like L&C 2020.

*Related*

Superforest,Climate Change

via Watts Up With That? https://ift.tt/1Viafi3