Abstract
Jung et al. (2014, PNAS) report that US hurricanes with more feminine names cause more deaths, and attribute this to name-based gender expectations reducing preparedness. I re-analyse the 92-storm dataset distributed with the DHARMa R package. The paper's negative binomial models reproduce closely: the name × damage interaction is positive and significant in both the raw (β = 0.000016) and standardised (β = 0.705) specifications, matching the published estimates, and the predicted death ratio between a feminine and a masculine severe storm is 2.82 versus the published 2.76.
The result is, however, fragile. Across ten defensible alternative specifications, three that a reasonable analyst could defend — restricting to the post-1979 era when names actually alternated by gender, using ordinary least squares on log deaths, and fitting main effects only — produce intervals that include no effect. The headline claim is therefore supported by the paper's own model but depends on its exact choices (negative binomial errors, the damage interaction, and the pre-1979 female-only naming era), not on a general relationship in the data.
1. What the paper did and what I reproduced
The archival study regresses hurricane deaths (alldeaths) on the
masculinity–femininity index of the name (MasFem), minimum pressure,
and normalised damage (NDAM), using negative binomial regression, and
reports two two-way interactions (name × pressure, name × damage). The headline
claim is that death tolls rise with name femininity for high-damage
storms.
I reproduce the paper's models 3 (raw) and 4 (standardised) with
MASS::glm.nb, the predicted-count model behind Figure 1, and the
interaction figure. Output is in output/comparison_table.csv and
output/figure1_predicted_deaths.png.
Reproduction vs published (selected quantities)
| Quantity | Published | Reproduced | Source |
|---|---|---|---|
| n analysed | 92 | 92 | PMC |
| Mean deaths | 20.652 | 20.652 | PMC |
| Variance of deaths | 1673.152 | 1673.152 | PMC |
| MFI × pressure, raw β (SE) | 0.006 (0.0025) | 0.00611 (0.00253) | PMC |
| MFI × pressure, raw p | 0.012 | 0.016 | PMC |
| MFI × damage, raw β (SE) | 0.00002 (0.00001) | 0.000016 (0.000004) | PMC |
| MFI × damage, raw p | <0.001 | 1.3e-5 | PMC |
| MFI × pressure, standardised β | 0.395 | 0.395 | PMC |
| MFI × pressure, standardised p | 0.012 | 0.0095 | PMC |
| MFI × damage, standardised β | 0.705 | 0.705 | PMC |
| MFI × damage, standardised p | <0.001 | 2.6e-6 | PMC |
| Pearson χ²/df, model 3 | 1.107 | 1.140 | PMC |
| Predicted deaths, high damage, MFI = 1 | 10.80 | 12.63 | PMC |
| Predicted deaths, high damage, MFI = 11 | 58.70 | 71.13 | PMC |
| Predicted deaths, low damage, MFI = 1 | 5.86 | 5.15 | PMC |
| Predicted deaths, low damage, MFI = 11 | 3.69 | 3.35 | PMC |
| Predicted deaths, high damage, MFI = 3 (Fig. 1) | 15.15 | 17.85 | PMC |
| Predicted deaths, high damage, MFI = 9 (Fig. 1) | 41.84 | 50.35 | PMC |
| Ratio of predicted deaths, high damage, MFI 9/3 | 2.76 | 2.82 | PMC |
Source. All published numbers above come from the PMC full text I opened in this session, https://pmc.ncbi.nlm.nih.gov/articles/PMC4066510/. The publisher page (doi) returned HTTP 403 and the supplementary PDF (Table S2) also returned 403, so no supplementary-table numbers are used. No published value is taken from memory.
The interaction coefficients and the predicted ratio match the paper closely.
The predicted levels differ by roughly 10–20 % because the paper never
states how it split NDAM into "high" and "low"; I used a median
split. The two-way interaction is the load-bearing result and it reproduces.
2. Analysis decisions the paper did not settle
- Outlier rule. The paper removes Katrina (1833 deaths) and Audrey (416), both feminine-named, "because retaining them leads to poor model fit", but gives no operational rule. The supplied CSV already excludes them, so the no-removal alternative cannot be tested with this file.
- The damage split. For Figure 1 and the predicted counts,
NDAMis "factored into two categories" without saying whether the cut is the median, the mean, or something else. I used the median. - Which pressure variable. The file carries both
MinPressure_beforeandMinpressure_Updated_2014; the paper uses NOAA pressure and refers to updated 2014 values, but does not say which entered the models. I used the updated values and test the alternative. - Interaction structure. Both interactions are included, but the paper never justifies keeping the non-significant main effect of damage or dropping the pressure × damage interaction. Models with and without interactions give different answers.
- Model family. Negative binomial is chosen on overdispersion grounds; the alternatives (generalised Poisson, NB-P, etc.) appear only in the inaccessible SI. The error model materially changes the interval width.
- Functional form of damage. Raw
NDAMis used; no log or per-capita damage is considered. - Era confound. Before 1979 every Atlantic storm had a female name, so name gender is confounded with era. The paper reports the split-sample analyses only as a robustness note.
- Continuous vs binary name predictor. The paper reports both; they are not equivalent (MFI carries graded information).
3. Robustness
Operational definition of "the conclusion". A specification supports the headline claim if the 95 % CI for the feminine-vs-masculine ratio of predicted deaths for a severe (90th-percentile damage) storm lies above 1. If the interval contains 1, the specification does not support it.
Alternatives and expectations (written before running). I expected S2 (post-1979), S6 (OLS on log deaths) and S8 (no interactions) to change the conclusion, and the rest to leave it intact.
| # | Alternative choice | Expected direction |
|---|---|---|
| S1 | Keep all 94 storms (retain the two outliers) | weaken (overdispersion); not runnable here |
| S2 | Post-1979 only (alternating names) | weaken, likely non-significant |
| S3 | Binary Gender_MF instead of continuous MasFem | similar |
| S4 | MinPressure_before instead of the 2014 update | similar |
| S5 | Poisson instead of negative binomial | same sign, too-narrow CI |
| S6 | OLS on log(deaths + 1) | weaken, possibly non-significant |
| S7 | Drop the name × pressure interaction | retain, smaller |
| S8 | Main effects only (no interactions) | weaken, likely non-significant |
| S9 | Zero-inflated negative binomial | similar |
| S10 | Median-split NDAM (paper's interpretation model) | similar |
Results (output/robustness_table.csv). The estimate
is the ratio of predicted deaths, feminine vs masculine name, at mean pressure and
90th-percentile damage. The p-value is for the name × damage interaction
(or the name main effect in S8).
| # | Specification | n | Ratio [95 % CI] | p |
|---|---|---|---|---|
| S0 | Paper replication (NB, MFI, both interactions) | 92 | 4.37 [2.16, 8.85] | 1.3e-5 |
| S1 | Keep all 94 storms | — | not runnable (data file is the 92-storm set) | |
| S2 | Post-1979 only | 54 | 1.30 [0.64, 2.64] | 0.137 |
| S3 | Binary gender instead of MFI | 92 | 5.11 [2.25, 11.61] | 5.7e-5 |
| S4 | Original min pressure | 92 | 5.18 [2.56, 10.49] | 2.6e-6 |
| S5 | Poisson | 92 | 2.07 [1.79, 2.41] | 2.3e-13 |
| S6 | OLS on log(deaths + 1) | 92 | 1.47 [0.78, 2.79] | 0.235 |
| S7 | Drop name × pressure interaction | 92 | 2.82 [1.50, 5.28] | 3.3e-4 |
| S8 | Main effects only | 92 | 1.23 [0.77, 1.98] | 0.386 |
| S9 | Zero-inflated negative binomial | 92 | 4.37 (point only) | 2.3e-10 |
| S10 | Median-split NDAM | 92 | 2.82 [1.40, 5.67] | 0.045 |
Three of the nine runnable specifications (S2, S6, S8) have intervals containing 1 and therefore fail to support the claim, exactly as anticipated. S5 keeps significance but understates uncertainty because Poisson assumes mean–variance equality, which the data violate badly (variance 1673 vs mean 20.7). The remaining specifications reproduce the direction and significance. The effect is thus driven by the choice of negative-binomial error model, the inclusion of the damage interaction, and the inclusion of the pre-1979 era.
4. Conclusion
Under the paper's own specification, the result reproduces: severe hurricanes with more feminine names are estimated to kill substantially more people, and the interaction coefficients match the published values to two significant figures. But the claim is fragile: it disappears in the post-1979 era when names actually alternated by gender, in an equal-weight model on log deaths, and in a main-effects model, so it rests on a few high-damage pre-1979 storms rather than a general relationship. The data support the paper's model-dependent estimate, not the strong causal reading in the title.