In this post I consider various alternative specifications, partly mentioned and suggested by other authors, and partly based on my own ideas. My R script on GitHub includes some specifications (e.g. using a dummy variable for gender, or log-based specifications) which I do not discuss here. I have to state at the outset that I consider the available data (compiled and provided by Jung et al) as incomplete. As mentioned in my post on endogeneity, the estimates obtained by Jung et al (2014) may also suffer from an omitted variable bias (e.g. because of missing data on population and land usage). I still continue to use the available data for illustration purposes, without claiming to obtain valid conclusions about a potential gender effect in storm-related fatalities.
In my post on endogeneity I conclude that dealing with the likely endogeneity of the damage variable NDAM requires an instrument which may be difficult to find. Instead of dealing with endogeneity, I consider replacing NDAM with the variable ‘Category’, representing the Saffir–Simpson hurricane wind scale (higher values indicate stronger wind). This variable is included in the original dataset but not used by Jung et al. The table below shows that there is a pretty strong (almost monotic) relation between these five categories and the means of alldeaths, NDAM and MP (inversely) for each level of Category.
Category alldeaths MFI MP NDAM
1 12.72222 6.547839 982.8889 2206.917
2 23.80952 6.674602 964.1429 8011.905
3 17.03571 7.238095 951.2500 7051.464
4 29.40000 6.800000 934.8000 26782.000
5 159.00000 5.638890 915.5000 44885.000
I argue that this variable is a meaningful, ex-ante, exogenous characteristic of hurricans. Despite being (highly) correlated with MP, it may provide complementary information about the consequences of a storm. I do not argue, however, that including ‘Category’ as a regressor is sufficient, and makes the specification complete. The following table shows the results from the negative binomial regression. Except for MP, none of the coefficients comes even close to being significant. This may be due to multicollinearity effects induced by the strong similarity of MP and Category.
Estimate Std. Error z value Pr(>|z|) (Intercept) 85.349350 36.673682 2.327 0.0200 * MP -0.086180 0.037301 -2.310 0.0209 * Category2 0.537107 1.279563 0.420 0.6747 Category3 -1.589822 1.517688 -1.048 0.2949 Category4 -2.178989 2.221983 -0.981 0.3268 Category5 -2.001862 2.891100 -0.692 0.4887 MFI -3.359860 4.608751 -0.729 0.4660 MP:MFI 0.003666 0.004690 0.782 0.4343 Category2:MFI -0.199943 0.159782 -1.251 0.2108 Category3:MFI 0.008971 0.189445 0.047 0.9622 Category4:MFI 0.028367 0.288490 0.098 0.9217 Category5:MFI 0.086288 0.401466 0.215 0.8298
When excluding the interactions between categories and MFI we obtain the results below. The coefficients of MP, MFI and their interaction are comparable to those in Table S2 for Model 3 in Jung et al, and wind categories have a negative effect (as expected). Note that there are only 5 and 2 observations for catogory 4 and 5. A simplified version which combines these two categories leads to very similar results.
Estimate Std. Error z value Pr(>|z|) (Intercept) 92.196546 20.801568 4.432 9.33e-06 *** MP -0.092878 0.021382 -4.344 1.40e-05 *** Category2 -0.758633 0.453214 -1.674 0.09415 . Category3 -1.575901 0.568464 -2.772 0.00557 ** Category4 -2.075793 0.934387 -2.222 0.02631 * Category5 -1.638690 1.357516 -1.207 0.22738 MFI -3.986598 2.095612 -1.902 0.05712 . MP:MFI 0.004266 0.002174 1.963 0.04969 *
The following figures show the effects of the gender index on expected fatalities associated with all levels of Category and for two different levels of MP already used in another post. The left figure shows results for a high level of MP (80% quantile), the right figure for a low level of MP (20% quantile). Expected fatality counts are clearly affected by the level of MP. We find that counts are monotonically increasing with the gender index, but the slope is weaker for a low MP level. Low levels of MP are typically associated with higher damage. Thus, this result is in conflict with Jung et al’s finding “that the change in hurricane fatalities as a function of MFI was marginal for hurricanes lower in normalized damage, indicating no effect of masculinity-femininity of name for less severe storms. For hurricanes higher in normalized damage, however, this change was substantial.”
[to be continued]
Alois Geyer 