However, when data is limited (in terms of amount and quality), an absence of statistical significance doesn’t mean we should ignore the relationship between SST and the number of tropical cyclones. Scientists use such information to point them in the direction of trying to figure out causal mechanisms between the two variables.

Maybe you’re using “absence of statistical significance” differently that I’ve seen it used elsewhere, but I thought the entire purpose of demanding statistical significance was to allow people to ignore relationships which may have occurred purely by chance. I suppose one could use near significance as a guide for further research, but surely not as evidence for a theory. And even then, I’d think there are plenty of other pointers which could be persued than looking to ones which lack statistical significance.

I think I see a bit of “bias” creeping in, as researchers comb through data searching for something which will butress their existing ideas. But this means that other pointers, which would/could point in the opposite direction are ipso facto being ignored. This is why the sceptics are so worried about claims of “concensus.” It’s the old story of “to a person with a hammer, everything looks like a nail.”

Spectral analysis of the tropical cyclone database (not shown) reveals distinct peaks at 2-3, 5, 9 year and multi-decadal time periods. As longer period variations and trends are the focus of this paper, we require the removal of short-term variations, such as associated with the El Nino-Southern Oscillation (ENSO) and quasi-biennial oscillation (QBO), which are both of order 2-3 years in period. We tested filters and averaging periods between 3 and 15 years. The averaging showed considerable sensitivity to smoothing period out to 5 years but relatively stable at longer averaging periods. All data are therefore smoothed by a simple 9-y running mean.
Standard statistical tests are used throughout. Unless otherwise indicated all variance and correlation coefficients are valid at the 95% level using the t-statistic with correction for serial correlation (Davis 1976).

But this statement, as written, is false. Non-linear time-series need not exhibit state-switching behavior. It is a peculiar feature of the highly non-linear Lorenz “butterfly” attractor.

Actually it’s quite common in non-linear systems. The atmosphere displays it too, for example the PDO shift in 1976, and, I’d contend, the global temperature as measured by the satellites, flat before the 1998 El Nino, then a step up afterwards. We probably need more time to see if it’s really a step up, but it looks like it. Needless to say, I don’t buy into the linear trendline.

1. The North Atlantic is the only basin where the number of tropical cyclones is increasing. Globally, the number of tropical cyclones remains the same (implying a slight decrease outside the North Atlantic). Bill Gray has been saying for decades that the Atlantic tropical cyclones show different behavior from those in other basins, so yes the Atlantic is “special” when it comes to tropical cyclones. The number of NATL tropical cyclones (when filtered for el nino and other short term variability) does show a correlation with SST. Webster has done the calculation for 5 year running means, and the correlation between NATL SST and TC counts from 1910-2005 gives a correlation of 0.78 that is significant using the Mann Kendall statistic (this is a curry/webster paper that is in press, to be published around June 07). If Steve M, Bender, or Willis can provide any further insights into the statistical significance of this relationship, I would certainly be interested in pondering it and would certainly forward to Holland/Webster. However, when data is limited (in terms of amount and quality), an absence of statistical significance doesn’t mean we should ignore the relationship between SST and the number of tropical cyclones. Scientists use such information to point them in the direction of trying to figure out causal mechanisms between the two variables. The highest resolution climate model simulation run by Oouchi et al. found an increase in the number of NATL TCs in a doubled CO2 scenario (although the rate of increase with SST was only 20% of that inferred from the historical NATL data (note Oouchi et al. found that the number of TCs outside the NATL to decrease slightly). So apparently there is something in the coupled atmosphere/ocean system that causes the number of NATL TCs to increase with SST. An important issue to understand, and several groups are trying to understand this relationship in the context of physical mechanisms, statistical analysis, and model simulations. While i agree that robust statistical analysis is important (and that our field is often too sloppy in this area), none of our arguments stand or fall solely based upon statistical analyses of imperfect data.

2. Re the European TC forecasts, a paper has been submitted (Palmer, Vitart et al.) that is under review (last time I heard). They do plan to go operational for the 2007 season

3. Re Holland’s statement in the ppt presentation:
The strong relationship between increases in storm and hurricane numbers and increases in SSTs leads to the inescapable conclusion that the majority of current hurricane activity is a direct result of greenhouse warming.
This is a short cut version of the argument presented in our BAMS article, which rested on the causal chain:
1. observation of increased hurricane activity (global intensity and NATL number)
2. statistical link of global intensity and NATL number with SST (see Hoyos et al. for global intensity link with SST)
3. attribution (through climate model simulations) of increased SST (at least since about 1970) to greenhouse warming

The issue of data quality is rather overwhelming outside the NATL, so recent attention has focused on the NATL where the data quality is more credible. The “inescapable” word in the ppt presentation is too strong, but .ppt presentations don’t tell the entire story of what was actually said (I wasn’t there so i don’t know what was actually said).

4. Re 2006 data. The Holland manuscript was submitted last summer and accepted for publication before the end of the TC season. The TC data is still under revision (note a 10th storm was recently added), but arguably good enough to be included at this point. However, the 2006 SST data (at least in a form that is consistent with the rest of the climatological data set) is not yet available.

Figure 3b is a time plot of hurricanes (the blue line) as a proportion of all storms. From 1900 forward the proportion looks trendless to me, but the pre-1900 proportion is noticeably higher. Why is that?

Well, in the text (Section 3.1) they seem to dismiss this with the words, “The higher proportion of hurricanes in the early record (Fig 3b) was probably due to a tendency to miss weaker systems in earlier years.” I certainly agree with that.

So, there is an acknowledgement that weaker earlier storms were likely missed. Nevertheless, they seem to go ahead and use that problemed historical data in their key Figure 1!

Suppose that we adjust the number of pre-1900 storms upwards, so as to make the hurricane proportion the same as post-1900. That adjustment would add about 20% to the pre-1900 storm count (eyeball estimate).

Now, use that +20% to adjust Figure 1, so that it is apples-to-apples. that would bring the average annual storm count to about 9.5, which looks very close to the 10 average of the period 1935-1995. Basically, no step change. Worse that that for their argument is that SST stepped upwards around 1920. So, SST rose while the storm count stayed constant.

Perhaps there is an explanation, which I’ll find when I read the entire report.

The graphic you’re wondering about was taken from the new Holland/Webster paper, which is still available on Holland’s site.

Here is the “paper”. It is not actually a published “paper”, but a submitted manuscript. And it has flaws, such as the one I pointed to in #73. On p 13:

SST anomalies explain over 60% of the tropical cyclone variance since 1905

No reviewer should let statements on statistical association pass without a statement of confidence (p value) and full disclosure of the methods of analysis. And no time-series correlations whould ever be reported without having done a sensitivity analysis on the time-points (to avoid cherry-picking dates that favor the hypothesis). Why 1905? What happens if you pick 1906, or 1910, and so on?

Figure 4 is comical. The hypothesis here is that a step function fits the hurricane proportion data better than a smooth oscillatory function. Where is the hypothesis test? How were the parameters of the step function chosen? Where are the methods? Malcolm Hughes would give this figure a failing C grade if it were the product of a graduate student.

In discussing Fig 4 I was glad to see the reference to Lorenz on p. 9:

This sharp transition from one climatic state to another is a feature of non-linear time series (Lorenz 1963)

But this statement, as written, is false. Non-linear time-series need not exhibit state-switching behavior. It is a peculiar feature of the highly non-linear Lorenz “butterfly” attractor.

In short, this paper has a ways to go before being acceptable for publication. What other mistakes/oversights does it contain?

Re #70: There are rumors that the Euro modelers will be going public this spring.

And there are rumors they won’t. Shows you what rumors are worth. I’ll believe it when I see it.