The time- and distance-decay effects of hurricane relevancy on social media: an empirical study of three hurricanes in the United States

Figures & data

Table 1. Query parameters used to retrieve tweets via twitter API.

Hurricane	Keywords	Timeframe*	Spatial extent
Harvey	“Hurricane Harvey” OR “Harvey”	08/25/17–11/16/17	United States
Florence	“Hurricane Florence”OR “Florence”	09/14/18–12/06/18	United States
Laura	“Hurricane Laura” OR “Laura”	08/27/20–11/18/20	United States

*Each timeframe starts when its respective hurricane first made landfall within the contiguous US.

Figure 1. Hurricane paths and the spatial distribution of hurricane-related tweets.

Table 2. Number of tweets for the three hurricanes after manual removal of false positives. Geotagged tweets are tweets associated with point coordinates or Twitter places (city, neighbourhood, poi).

Hurricane	Total	Geotagged
Harvey	91,084	74,327
Florence	18,484	14,029
Laura	10,070	8,165

Figure 2. Comparing ROC of Tweets over Time.

Table 3. Comparing ROC of tweets over time (monthly).

Month	Harvey		Florence		Laura		Avg ROC
	Tweets	ROC	Tweets	ROC	Tweets	ROC
1	78,719	-	16,611	-	6,621	-
2	8,592	−89.1%	1,270	−92.4%	1,756	−73.5%	−85.0%
3	3,773	−56.1%	603	−52.5%	1693	−3.6%	−37.4%
Avg ROC	-	−72.6%	-	−72.4%	-	−38.5%	−61.2%

Table 4. Comparing ROC of tweets over time (weekly).

Week	Harvey		Florence		Laura		Avg ROC
	Tweets	ROC	Tweets	ROC	Tweets	ROC
1	44,279	-	12,096	-	4,809	-
2	22,857	−48.4%	2,304	−81.0%	710	−85.2%	−71.5%
3	7,837	−65.7%	967	−58.0%	567	−20.1%	−48.0%
4	3,045	−61.1%	940	−2.8%	419	−26.1%	−30.0%
5	2,088	−31.4%	628	−33.2%	342	−18.4%	−27.7%
6	1,572	−24.7%	382	−39.2%	424	24.0%	−13.3%
7	2,538	61.5%	212	−44.5%	577	36.1%	17.7%
8	2,232	−12.1%	216	1.9%	323	−44.0%	−18.1%
9	1,387	−37.9%	210	−2.8%	332	2.8%	−12.6%
10	1,085	−21.8%	183	−12.9%	450	35.5%	0.3%
11	1,242	14.5%	157	−14.2%	631	40.2%	13.5%
12	819	−34.1%	163	3.8%	432	−31.5%	−20.6%
Avg ROC	-	−23.7%	-	−25.7%	-	−7.9%	−19.1%

Figure 3. Word clouds of Hurricane Harvey tweets in Weeks 6 (left) and 7 (right) (the words “hurricane” and “harvey” were removed).

Table 5. Correlation between weekly tweet totals and time (Weeks 1–12).

Hurricane	Correlation coefficient	p-Value
Harvey	−0.937	<.001
Florence	−0.986	<.001
Laura	−0.329	.297

Table 6. Correlation between weekly tweet totals and time (Weeks 2–12).

Hurricane	Correlation coefficient	p-Value
Harvey	−0.918	<.001
Florence	−0.982	<.001
Laura	−0.127	.709

Figure 4. Kernel density maps of hurricane Harvey tweets over time.

Figure 5. Kernel density maps of hurricane Florence tweets over time.

Figure 6. Kernel density maps of hurricane Laura tweets over time.

Figure 7. Tweets per 100,000 people by zone over time.

Table 7. Average ROC of tweets per 100,000 people (all hurricanes) over time by zone.

Month	Primary		Secondary		Tertiary		Far		Very Far
	Tweets	ROC	Tweets	ROC	Tweets	ROC	Tweets	ROC	Tweets	ROC
1	76.7	-	7.9	-	8.9	-	4.9	-	5.0	-
2	6.4	−91.7%	0.7	−91.1%	1.0	−88.8%	0.7	−85.9%	5.4	−84.1%
3	3.2	−49.6%	0.4	−45.6%	0.5	−51.8%	0.4	−43.1%	0.9	−46.5%
Avg ROC	-	−70.6%	-	−68.3%	-	−70.3%	-	−64.5%	-	−65.3%

Table 8. Correlation between the number of tweets (per 100,000 people) and distance.

Hurricane	Correlation coefficient	p-Value
Harvey	−0.366	.009
Florence	−0.59	<.001
Laura	−0.361	.011

Table 9. Comparison of the true positive rate (TPR) and its 95% confidence interval (CI) of hurricane tweets based on random samples of tweets (n = 200) taken from the original tweets, tweets after manual cleaning, and excluded tweets on each hurricane.

Tweets sample	Harvey		Florence		Laura
	TPR (%)	95% CI	TPR (%)	95% CI	TRP (%)	95% CI
Original	74.5	[68.5, 80.5]	69.0	[62.6, 72.4]	23.5	[17.6, 29.4]
Cleaned	81.5	[76.1, 86.9]	82.0	[76.7, 87.3]	49.0	[42.1, 55.9]
Excluded	7.5	[3.9, 11.2]	1.5	[0.0, 3.2]	0.5	[0.0, 1.5]

Figure 8. Word clouds of hurricane tweets (cleaned) over time. Names of the three hurricanes and the word “hurricane” were removed prior to generating the word clouds.