By now you will have probably seen the reports about new research apparently showing that female named Atlantic Hurricanes are deadlier than male named storms because female names are ‘less scary’ than male names. The main investigation consisted of an analysis of 92 Hurricanes between 1950 and 2012 that made landfall in the USA (Katrina and Audrey, the two most deadly storms were excluded) examining fatalities, normalised damage, category, minimum pressure, gender of the storm name and year of occurrence.

There has been a fair amount of criticism on the paper from around the web including GRRLScientist in the Guardian, Future Tense in Slate, Not Exactly Rocket Science in National Geographic and on Mashable. The authors have responded to some of the criticism in these pieces.

The authors have also made their data and statistics from the model fit available for use.

The principal criticism has been on the approach to show that their finding is ‘significant’ which involves generalised linear regression on a negative binomial distribution and two-way interaction terms. Less fancy methods (simple correlation and a multi-linear regression) found no significant effect. B

Is there a difference in the number of people who die in Hurricanes based on the storm name’s gender?

You can’t just compare the average rates – Hurricane fatalities are a fat tail phenomenon, a small number of big storms contribute most of fatalities. Let’s forget modelling, regression algorithms and the fancier statistics. Let’s just look at the data. Below I’ve plotted the cumulative distributions of fatalities from male named storms and female named storms from the entire dataset.

You can see that for the high frequency, low fatality events the distributions are basically the same but there appears to be some divergence at the more severe end with a higher fatality rate in female named storms. But the difference isn’t that much – and there’s the data issue, with storms between 1953 and 1979 only receiving female names there’s way more of them than male named storms. So visually there’s not a huge difference and we can quantify this using a statistical measure called the Two-sample Kolmogorov-Smirnov test. This test allows comparisons of two samples of data and helps you decide between one of two hypotheses:

H0 – They’re from the same distribution.

H1 – They’re from different distributions.

For these data the KS-test gives D = 0.1065, p-value = 0.976. This means that we can’t reject H0 – i.e. we can’t tell if the two samples are different or there’s no detectable difference in the underlying pattern of fatalities based on storm name gender.

We can also look at only the storms during the time in which both male and female names were used:

Here we can see that what little difference we saw in the complete dataset disappears completely, further suggesting that the effect found by Jung et. al. is quite probably a statistical fluke. If you look at enough variables you’ll eventually find a statistically significant correlation. In fact if you look hard enough you can find all sorts of strange things that correlate with each other.

To explain the effect, the authors turn to additional experiments where they measured a difference based on storm name gender in perceived threat and willingness to evacuate. Implicit sexism plays a big, big role in our society. (I highly recommend Cordelia Fine’s book, Delusions of Gender, which goes into the detail very well) but it’s hardly the only thing at play – many factors cause people to (usually) downplay threats from and delay responses to Hurricanes and other disasters.

The scenarios presented (you can see examples here) were not anything like real Hurricane warnings and media broadcasts, which tend to be much more alarming and action oriented, especially for more severe Hurricanes which is where the authors claim there is an effect. A better experiment would be to mock up a TV news broadcast of Hurricane Alexander/Alexandria and show it to people in hurricane prone areas and see what the results are. I would be very surprised if there was still a detectable storm name effect once a heavy dose of reality is injected.

As for policy recommendations I think its definitely too soon to consider changing the naming system, but whilst we’re talking about communication the NHC could really overhaul it’s woeful public advisory messages.

Alternatively there’s always this idea: