Saturday, November 29, 2014

Torture of Tornado Data in the United States is Increasing

In the October 17, 2014 issue of Science, Brooks et al publish the claim that tornadoes, while not increasing in frequency, are increasingly “concentrated” among fewer, more intense days. [Ref: H.E. Brooks, G.W. Carbin, P.T.Marsh. Science, 346, 349-352 (2014).] This claim is based on analysis of EF1 or greater tornadoes in the NOAA Severe Weather Database files [Ref: ] for the 60 years 1954 - 2013.

It is well known that increased population and improved doppler weather radar systems has allowed weaker and more remote tornadoes that previously would have gone undetected to be counted. It is also well known that strong EF3+ tornadoes have been declining over this period. So, I suspected that this report of increasingly “concentrated” tornado days, much like the reports of increased rainfall intensity, in the absence of increasing precipitation overall, was an exercise in data torture and failure to recognize biased data. As you will see, I was not disappointed.

Let's begin with the last 60 years of tornado data, collected by year and by intensity on the enhancedFujita scale of EF0-EF5
Figure 1: Annual count of US tornadoes by enhanced Fujita scale classification reveals a sharp rise in the detection of weak tornadoes after 1990.

It is immediately evident that the annual number of recorded US tornadoes is rising, and that this increase appears to be driven by EF0 tornadoes, after 1990. This is not surprising, because 1990-1997 were the years that the national NEXRAD doppler radar network was deployed. [Ref:] To the eye, the EF1 and EF2 numbers do not appear to be biased by a similar sharp increase. But, let's look a little closer. First, we know that weaker tornadoes occur more frequently than stronger tornadoes, but we assume that the ratio of weak to strong tornado frequencies should not change very much over time. This is consistent with the theoretical and observed power law, or 1/f law, governing the relative frequencies of storms of different energies. [Ref: J B Elsner et al 2014 Environ. Res. Lett. 9 024018 ] Therefore we can examine how much more frequently EF0, EF1, and EF2 tornadoes are observed relative to EF3+ tornadoes. While it is very difficult to quantify exactly how much detection bias contaminates the tornado frequency data, we can approximate it by assuming that EF3, 4 and 5 tornadoes are unlikely to have been missed in any of the last six decades.

Figure 2: Frequency of EF0-2 tornadoes in the US relative to EF3+ tornadoes, 1954 – 2013.

Viewed in this way, the contamination of both EF0 and EF1 frequency data by detection bias is clear. In the 1960s and 1970s, EF1 tornadoes were recorded 4-5 times as frequently as EF3+ tornadoes, but since 2000, they have been detected about 8 times as frequently. There is no physical basis for this to be the case, which violates the approximate power law behavior of tornadoes [Ref: J B Elsner et al 2014 Environ. Res. Lett. 9 024018]. It is also clear that the EF2 data is not similarly biased, maintaining a flat frequency that is about 3 times that of the EF3+ tornadoes for the last 60 years, which is consistent with value of 2.8 taken from the observations of Elsner et al.

Brooks et al produced this figure, showing the number of days with EF1+ tornadoes, along with the number of days with over 30 EF1+ tornadoes:

Figure 3: (Brooks et al 2014 Fig. 4) used to claim that EF1+ tornadoes are increasingly clustered.

Using the NOAA source data and the OpenOffice spreadsheet program, I was able to reproduce essentially the same figure:
Figure 4: The number of days with observed EF1+ tornadoes is trending down, but the number of days per year with more than 30 observed EF1+ tornadoes (right side scale) has increased.

If the method of Brooks et al is applied in the same way to EF2+ tornadoes, while accounting for the fact that EF2+ tornadoes are about one-third as frequent, and adjusting the clustering threshold to >10 EF2+ tornadoes per day, figure 5 is obtained:

Figure 5: The number of days per year with observed EF2+ tornadoes is decreasing, but the number of those days with more than 10 observed EF2+ tornadoes (right side scale) remains flat at about 2.7 days per year.

Finally, if we restrict ourselves to the most destructive tornadoes, those of EF3 or greater, it is obvious that the frequency of EF3+ tornadoes in the United States is decreasing, and that there is no increased “clustering” of days with more than three:
Figure 6: Declining frequency of EF3+ tornadoes 1954 – 2013, and number of days with more than three (right side scale).

I was disappointed to see this misleading abuse of statistics reported so uncritically in Science. It is unfortunate that the Science reviewers were not able to see what someone with an undergraduate understanding of statistics should readily see. As demonstrated in the above figures, there is no reason to believe the perceived increase in days with large numbers of tornadoes claimed by Brooks et al is anything other than another artifact of detection bias, because it only applies to weak tornadoes that could have escaped detection in the past, and disappears when restricted to EF2+ or EF3+ tornadoes. This seems to be part of a trend towards credulity by reviewers, and lack of rigor by authors of papers that support public concern about climate change. One never sees this in the medical literature on new cancer treatments, although 30 or 40 years ago, one did.

I understand the frustration of global warming activists who want to be able to claim that destructive weather phenomena, like tornadoes, are becoming worse in some way, when the data says the opposite. Rather than following the data wherever it leads, Brooks et al have chosen to torture the data until it confesses. As the Scottish poet, Andrew Lang said in 1910, "He uses statistics in the same way that a drunk uses a lamp post, for support rather than illumination."