Female hurricanes are not deadlier than male hurricanes

Jung et al. (1) assert that hurricanes that made landfall in the United States killed more people when they had female names rather than male names. The article has stirred much controversy. Criticisms range from the inclusion of hurricanes from the era before they were given male names (2), over the selective interpretation and the overstatement of their results from the archival study in favor of their hypothesis (3), to the external validity of their six behavioral experiments for at-risk populations in at-risk situations (4).

The criticism of this letter is a different one: the results of their archival study are a function of the selective inclusion of regressors. Using the same data, methodology, and variables, I show in Table 1 that their results are not robust to the inclusion of the one two-way interaction they omitted from their analysis. Model 1 reproduces the authors’ main results. Models 2–4 show that the null that female- and male-named hurricanes were equally deadly cannot be rejected once the interaction effect of a hurricane’s barometric pressure and its normalized damage toll is included. A more accurate title to this letter should have stated exactly that.

Table 1.

Results from negative binomial regressions (outcome variable: total deaths)

	1	2	3	4	5	6	7
Variables	1950–2012	1950–2012	1950–2012	1950–2012	1950–1978	1979–2012	1979–2012
MP	−0.0713***	−0.0377**	−0.0405***	−0.0512**	−0.0525***	−0.0399***	−0.0385***
	(0.0213)	(0.0166)	(0.00813)	(0.0210)	(0.0153)	(0.00938)	(0.00941)
ND	−4.78e-05	−0.00222***	−0.00203***	−0.00185***	−0.00283***	−0.000850	2.28e-05**
	(3.06e-05)	(0.000461)	(0.000481)	(0.000573)	(0.000835)	(0.000686)	(1.12e-05)
MP × ND		2.42e-06***	2.19e-06***	1.99e-06***	3.10e-06***	9.29e-07
		(4.90e-07)	(5.18e-07)	(6.32e-07)	(8.73e-07)	(7.32e-07)
MFI	−6.163**	0.676	−0.00777	−1.529
	(2.436)	(1.825)	(0.0443)	(2.759)
MFI × MP	0.00632**	−0.000684		0.00156
	(0.00251)	(0.00190)		(0.00284)
MFI × ND	1.69e-05***		3.39e-06	5.50e-06
	(4.40e-06)		(3.48e-06)	(5.18e-06)
Likelihood ratio χ2	60.57***	68.62***	69.46***	69.76***	35.48***	37.65***	36.29***
Observations	92	92	92	92	38	54	54

Estimated coefficients with their SEs in parentheses. The analysis used the same data, variables, and estimator as ref. 1. I find similar results in models that winsorize the dependent variable and ND to their 90th percentile if they exceed that percentile (deaths > 55; ND > $20,335 million) and ordinary least squares models on the log-transformed dependent variable. Each model was run without and with robust SEs with consistent results. The coefficient for ND in model 7 was insignificant in the negative binomial regressions that winsorized the dependent variable and ND, suggesting that this result rests on only a few high-damage/high-fatality hurricanes. The second line in the column headings of the numbered columns indicates the range of years in the data included in the regression. MFI, masculinity-femininity index; MP, minimum pressure; ND, normalized damage.

Models 2–4 show that hurricanes with lower barometric pressures had higher death tolls and that hurricanes with higher damage tolls had smaller death tolls when the hurricanes were strong (lower pressure), but higher death tolls when the hurricanes were weak (higher pressure). The latter result is driven by the pre-1978 sample (model 5). In the post-1978 sample, the interaction effect becomes insignificant and the damage toll has a positive and significant relationship with the death toll (models 6 and 7).

Like the death toll, the damage toll is a simultaneous outcome of the storm and hence not a good explanatory variable. It merely reflects other underlying characteristics that could range from the size of the hurricane or its area of effect over the assets at risk and the infrastructure (including its safety infrastructure) to the size of the population at risk in that area. That the death toll decreases in the damage toll for strong storms before 1978 is consistent with an interpretation that the damage toll was most strongly reflective of the safety infrastructure in the area of effect for strong storms during that time. This would be plausible because the safety infrastructure is one of the few characteristics that come to mind that might correlate positively with the damage toll but negatively with the death toll. However, if we accepted this interpretation, we would have to explain why the damage toll is relatively less reflective of differences in the safety infrastructures and more reflective of other characteristics for weaker storms and after 1978. Even though a lower importance of safety infrastructures during weaker storms and an overall improvement in or a convergence of safety infrastructures after 1978 might explain these results, the ambiguity of the death toll variable disallows any strong or definitive interpretation.