Reply to Bakkensen and Larson: Population may matter but does not alter conclusions

Abstract

We report that for highly damaging hurricanes, not for less damaging hurricanes, name femininity predicts more fatalities (1). We suggest this may be because, for damaging storms, factors such as storm names that motivate protective action are more predictive of survival. Bakkensen and Larson (2) assert that our modeling suffers from endogeneity and lack of adjustment for population. The authors report reversed or no effect of hurricane names in models with completely different inputs. We show below that their approach and analyses are flawed, yielding incorrect conclusions.First, Bakkensen and Larson (2) provide no evidence for endogeneity bias in our models. Contrary to their assertion, a standard assessment for endogeneity as detailed in Hilbe (3) and Cameron and Trivedi (4) shows no bias at the level requiring adjustments (P > 0.10). (To test whether normalized damage is endogenous to fatalities, we built a simple original model in which fatality was regressed on normalized damage and gender index. Next, we regressed normalized damage on gender index and minimum pressure and obtained residuals. Finally, residuals were added as an additional regressor in the original model as well as in the count model. The added residuals were not statistically different from zero.) Dropping normalized damage as an "endogenous variable" is therefore unwarranted.Nevertheless, Bakkensen and Larson drop the damage predictor, the focal indicator of actual impact on population centers (1), such that their modeling does not address our hypothesis. Instead, the authors add population main effects and interactions without a conceptual rationale (2). A close look at these analyses reveals multiple flaws: Although they report no dispersion or model-fit statistics, we reproduced their modeling (model 2) using annual US population data and found it is not viable because of serious overdispersion and poor model-fit (Pearson's χ We do agree that, as population at risk increases, hurricane fatalities should increase ceteris paribus. However, adjusting for population density of just five coastal counties (models 3-6) is problematic: 150 counties are in the average hurricane's path and inland fatalities account for an increasing fraction of deaths, up to 80% (5). We also cannot abide Bakkensen and Larson's (2) use of population indicators both as predictors and adjustments to outcomes (model 5). Finally, normalizing count data to use ordinary least-squares regressions (models 5-6) is inappropriate (3, 4), as is their log-transformation of normalized deaths, which caused the 10 observations with zero values to simply disappear.Instead, to address population at risk, fatality counts can be adjusted for contemporaneous US population in our original model (1). [For example, Hurricane Edna's 20 deaths in 1954 (US population: 163,000,000) would be adjusted to 39 deaths (US population in 2012: 314,000,000); population data source: US Census Bureau.] Doing this yields virtually identical results, replicating the focal gender index × normalized da...

Which Data to Meta-Analyze, and How?

“…At the same time, the Jung et al (2014a) paper triggered a daisy chain of critical, published letters to the editor (Bakkensen & Larson, 2014; Christensen & Christensen, 2014; Maley, 2014; Malter, 2014), along with published author replies and rebuttals (Jung et al, 2014b, 2014c, 2014d), all offering discrepant, and seemingly irreconcilable, views on which hurricane data to include and how to analyze them. Simonsohn et al (2015) assembled all these views (or, alternative specifications for data analysis) combinatorially, showed that this yielded 1,728 different ways to analyze more or less the same underlying data set, and further showed that the specific finding of female-named hurricanes being deadlier, as reported in Jung et al (2014a), belonged to a small subset of analyses (37 out of a total of 1,728 specifications, or 2.1%) which yielded a nominally significant result.…”