Abstract

Jung et al. (2014) report that hurricanes with more feminine names cause more deaths. I reran their main model on the 92 US hurricanes in the published data: a negative binomial regression of deaths on name femininity, minimum pressure, damage, and the interactions femininity × pressure and femininity × damage. All coefficients reproduce exactly (femininity × damage b = 0.705; femininity × pressure b = 0.395). My standard errors are somewhat smaller than the published ones, and the published predicted death counts do not reproduce, because the paper does not say at which damage and pressure values they were computed.

I then ran nine alternative specifications. The femininity effect holds when I use binary gender, control for year, use updated pressure values, or drop the most extreme storms. It disappears (estimates near zero, p ≥ .16) without the interaction terms, with log-transformed damage, and with only the storms from 1979 on, when male and female names were both in use. The finding depends on raw, highly skewed damage and on the years before 1979, so the data do not robustly support the headline claim.

All numbers come from analysis.R (run: Rscript analysis.R; needs the R packages MASS and glmmTMB). The script writes robustness.csv and robustness.png.

1. Main result

The paper’s archival result is Model 4: a negative binomial regression of deaths on the standardised masculinity–femininity index (MFI), minimum pressure, normalised damage (NDAM), and the interactions MFI × pressure and MFI × damage. The data exclude Katrina (2005) and Audrey (1957), so n = 92. I fitted this model with MASS::glm.nb on the standardised variables in the file (ZMasFem, ZMinPressure_A, ZNDAM).

2. Reproduced vs published estimates

Source for all published numbers: https://pmc.ncbi.nlm.nih.gov/articles/PMC4066510/ (full text, opened in this session). I did not use my memory of the paper for any number.

Quantity Published Reproduced
MFI × pressure, b 0.395 0.395
MFI × pressure, SE 0.157 0.152
MFI × pressure, p 0.012 0.009
MFI × damage, b 0.705 0.705
MFI × damage, SE 0.184 0.150
MFI × damage, p < 0.001 < 0.001
Predicted deaths, MFI = 3, high damage 15.15 10.09 (damage +1 SD, pressure at mean)
Predicted deaths, MFI = 9, high damage 41.84 51.54 (damage +1 SD, pressure at mean)

3. Analysis decisions the paper did not settle for me

  1. Which pressure variable. The file has two. ZMinPressure_A is standardised from MinPressure_before, not from Minpressure_Updated_2014 as the README says. I checked this in the data. Using ZMinPressure_A gives the published coefficients, so I used it for the reproduction.
  2. Software and SE estimator. I used glm.nb (maximum likelihood, model-based SEs). The published SEs are larger, so the paper did something else.
  3. Values for predicted deaths. “Severe” damage and the pressure level are not defined in the text I read. I used +1 SD and the mean.
  4. Estimand for robustness. With interactions, the MFI main effect is the effect at average damage. That is not the paper’s claim, because the claim is about damaging storms. I used the effect of +1 SD femininity on log deaths at damage +1 SD and mean pressure (MFI + MFI × damage). For models without interactions, it is the MFI main effect.
  5. Fitting engine for the robustness checks. glm.nb did not converge for the log-damage interaction model (its answer depended on the iteration limit: 0.22 or 0.43). I fitted all robustness models with glmmTMB (NB2, direct maximum likelihood). All converged. glmmTMB’s SEs are slightly wider than glm.nb’s because they include uncertainty in θ. Point estimates for converged models are the same.

4. Robustness

Alternatives specified before running

The research question stays fixed: do more feminine-named hurricanes cause more deaths, holding storm severity constant? Expected effect on the conclusion is in brackets.

  1. Main effects only (no interactions). This tests whether there is a femininity effect overall. [Expected to change: the MFI main effect is non-significant even in Model 4.]
  2. Binary gender instead of MFI. [Expected to hold: a coarser version of the same variable.]
  3. log(damage) instead of raw damage. Damage is highly right-skewed. On the raw scale, a few storms with extreme damage drive the MFI × damage slope. [Expected to change.]
  4. Only hurricanes from 1979 on. Before 1979 all hurricanes had female names, so name gender is confounded with era (warnings, building standards). [Expected to change.]
  5. Updated 2014 pressure values. [Expected to hold.]
  6. Control for year. Another way to deal with the era confound. [Uncertain; expected to weaken the result.]
  7. Drop the 3 deadliest hurricanes. [Expected to weaken the result but keep it.]
  8. Drop the 3 most damaging hurricanes. [Expected to change, because these points give the interaction its leverage.]
  9. log(damage), main effects only. Combines 1 and 3. [Expected to change.]

Results

Effect = change in log(deaths) for +1 SD femininity at damage +1 SD (interaction models) or overall (main-effects models). For specification 2, the effect is female vs male, so its size is on a different scale. Its sign and CI are comparable to the others.

Specification n Estimate 95% CI p
0 Published Model 4 92 0.88 0.46 to 1.29 < .001
1 Main effects only 92 0.13 −0.12 to 0.38 .31
2 Binary gender instead of MFI 92 1.82 0.86 to 2.77 < .001
3 log(damage) instead of raw damage 92 0.30 −0.12 to 0.73 .16
4 Only 1979 on 54 0.27 −0.12 to 0.66 .17
5 Updated 2014 pressure values 92 0.79 0.35 to 1.22 < .001
6 Control for year 92 0.88 0.43 to 1.34 < .001
7 Drop 3 deadliest 89 0.68 0.21 to 1.15 .005
8 Drop 3 most damaging 89 1.18 0.54 to 1.81 < .001
9 log(damage), main effects only 92 0.08 −0.14 to 0.30 .48
Robustness of the femininity effect

How the results compare with my predictions: - As predicted, the effect disappeared without the interaction (1, 9), with log damage (3), and when only 1979-on storms were used (4). - Not as predicted: the result held when the 3 most damaging storms were dropped (8). Controlling for year (6) did not weaken it. - Held as expected: 2, 5, 7.

5. Does the headline claim hold up?

The published coefficients reproduce exactly, but the finding depends on specific choices. The femininity effect is present only when damage enters on its raw, highly skewed scale and interacts with MFI. Without interactions, with log damage, or with only the years in which male and female names were both used, the estimates are near zero and not significant. The data therefore do not give robust support for the claim that feminine-named hurricanes are deadlier.