Fig 1.
Hashtag attention map and usage rate time series.
For 1-grams matching the case-insensitive pattern “#hurricane*” for all four hurricanes reaching at least category 4 in the 2017 hurricane season. Markers along the hurricane trajectory indicate the National Oceanic and Atmospheric Administration (NOAA) reported position for every day at noon UTC. On the map, the smoothed rate of hashtag usage is wrapped in an envelope around the hurricane trajectory in panel A, showing the spatial dependence of attention on Twitter. In the lower two plots, panels B and C, we show the usage rates for hashtags and 2-grams matching hurricane* in English language tweets on linear and logarithmic scales. Usage rates within all tweets are indicated with a solid line, while usage rates in ‘organic’ tweets (tweets that are not retweets), are represented by a dashed line. The day of maximum attention on Twitter is marked with a star or a diamond for hashtags or 2-grams, respectively. Generally, hurricanes making landfall on the continental United States received greater attention than those not making landfall. The hashtag usage rate for hurricanes Harvey and Irma at their maximum were approximately an order of magnitude larger than the maximum hashtag usage corresponding to Hurricane Maria, and two orders of magnitude larger than Hurricane Jose.
Fig 2.
Radar plots comparing the eight most monetarily damaging hurricanes in the North Atlantic basin from 2009 to 2018.
For each plot, starting at the top position and rotating clockwise the measures are: the sum of usage rate of the hashtag, the number of days to reach 90% and 50% of the total attention received during that season, the total cost in dollars attributed to damage caused by the hurricane (in its year), the number of deaths attributed to the hurricane, and maximum usage rate of the hashtag during the year of interest. All measurements are normalized to the maximum value achieved by any hurricane. Hurricane Harvey was the most talked about hurricane, as well as the most damaging. Hurricane Irma was the most talked about on any single day. Hurricane Maria caused the most deaths, and had the longest attention half-life of all measured hurricanes. Raw values for this figure are shown in S8 Table in S1 File. Hashtag usage rate spark lines above each radar plot are normalized to show the common decay shape, and can not be compared to evaluate relative volume, and are shown on a log scale.
Fig 3.
Scatter plots for integrated hashtag usage rate versus the deaths and damages caused by each storm.
There is a clear positive association between the total attention represented by hashtags and the impacts of these storms. We reported Spearman’s rho, ρs, in the top left corner of each plot. While for some categories, there is little evidence for a positive association, for the entire dataset ρs ∼ 0.54. We perform a Bayesian linear regression for each category storm between the logI and log impacts. We show the mean model, along with the credible interval within a standard deviation of the mean model. We use hybrid axis with logarithmic scaling for most horizontal and vertical values and linear scaling near zero, in order to show storms that caused zero deaths or damages, as well as storms for which we measured a hashtag usage rate of zero. Changes in axis scaling occur at the blue dashed lines. Generally, more powerful storms received more attention, higher category storms received more attention even when causing minimal damage, and high category storms had a higher regression slope. These results suggest that for powerful storms, a given increase in impact was associated with a larger increase in attention. While for category one storms a 10-fold increase in deaths is associated with a four-fold increase in attention, for category five hurricanes, this same 10-fold increase in attention is associated with a 25-fold increase in attention.
Fig 4.
Posterior distributions of regression parameters.
For the model log10 I ∼ a0+a1 Xi, where Xi is either the log number of deaths (A and C) or log damages in dollars associated with the storm (B and D), and log10 I is the log integrated hashtag usage rate. The trend in regression coefficients for association between the log attention and log deaths suggests that higher category storms receive more attention per unit impact, while the trend of intercepts shows increasing baseline attention for a hypothetical minimally disruptive storm causing exactly $1 in damages or one death. For regression coefficients relating log attention to log damages, Category 4 and 5 storms receive more attention per unit increase in log damages than lower category storms. However, the coefficients are smaller in magnitude due to damages varying across 7 orders of magnitude, as compared to deaths varying over 4 orders of magnitude. There is a larger uncertainty for the category 5 intercept values, as only 6 storms of this intensity formed between 2009 and 2019 in the Atlantic basin. At the right of each plot, we show the coefficients for the model fit for all hurricanes (blue violin), excluding tropical storms. Above each category, we show the value of the mean posterior distribution for each parameter. For a Table of mean parameter values, see S1 Table in S1 File.
Fig 5.
Parameter distributions for models 1, 2 and 3.
Plots A–C show posterior distributions for regression 1, plots D–G show distributions for regression 2, which includes the addition of an interaction term, and plots H–O showing distribution for regression 3, which includes indicators variables for hurricane categories two through five. The addition of the interaction term, ad,D increases posterior variance for adeaths as well as reducing its mean from adeaths = 0.49 in regression 1 to adeaths = 0.05 in regression 2 and adeaths = 0.12 in regression 3, suggesting that while the number of deaths is associated with increased attention, attention response is primed by destruction. Additionally, the hurricane category indicator variables in regression 3 show the progressive increase in attention given to higher category storms compared to category 1 hurricanes.
Table 1.
For each model presented in the paper. SubTable A refers to the regressions by category, while subTable B refers to the later sequential regression models. Each impact variable is presented as the expected increase in attention associated with a 10-fold increase in the variable of interest. For categorical variables we report the expected multiplier for the given hurricane category over a Cat 1 storm. The mean of the fitted posterior regression parameters are provided for the reader in the Appendix in S3, S5 and S7 Tables in S1 File.