Using Twitter to Analyze the Effect of Hurricanes on Human Mobility Patterns

Round 1

Reviewer 1 Report

The paper is well written and addresses an important problem. Its contribution and the significance of the findings need be better presented before publication. E.g., it shows that the Levy walk model resembles human mobility pattern during hurricane. What is the implication of this finding?

In conclusion, “Analysis of change in daily trip rates shows that travel interruptions are short term if no major pertinent transportation infrastructure problems occur in affected areas, like in the analyzed Hurricanes.” What is the result that support this conclusion? How does the Twitter data show whether there is a major pertinent transportation infrastructure problem?

P2, ln.89: The sentence of “...but not longer.” is incomplete.

P7, ln.191: “[100”->”]100”; ln.201:”MSD]”->”MSD”, ln.207: “additional”->”addition”; ln.209: what is the range of “i” in the equation? Ln.208: “raduis”->”radius”.

Author Response

We rewrote most of the discussion section to better explain the significance of the results for future research and for practical implications. With regards to the Levy walk model, since our results demonstrate a significantly better fit of the truncated power law to observed displacement before, during, and after the hurricanes than the non-truncated power law for two study regions, we propose to use the truncated power-law model as a baseline model for human mobility for disaster related movement simulations and analyses in the future.
As for practical implications, we pointed out that emergency response teams need to be prepared to face limited mobility of residents in affected regions, and also, that, based on exploratory analysis, the frequency of hurricane related hashtags used in a study region could provide some insight into the severity of transportation disruption. Due to the limited number of geo-tagged tweets with precise location information (which was also mentioned by this reviewer further below), we could, however, not yet statistically prove this hypothesis, hence kept this aspect as to be addressed in more detail in future work.

Thank you for pointing our this limitation in our chain of arguments. We could indeed not verify this fact based on tweets and therefore removed this statement from the conclusions. Instead, we addressed the magnitude of road network disruption during the hurricane events. For this purpose, we added a new table (Table 5) which reports the number of flooded roads, and road and bridge closures for the different study regions, based on a variety of online resources. This information now helps to underline our assumption that reduced mobility during the hurricanes, as it could be shown from the analysis of our used set of geo-tagged tweets, is at least partially caused by road network disruptions.

The syntactical issues have been fixed.

Reviewer 2 Report

Paper presents a robust approach to analysing and extracting meaningful patterns from Twitter data. It is very well-structured and is easy to read and follow. While I enjoyed reading the manuscript and appreciate the authors' effort in tackling such an exciting domain, I still have the following concerns and would like to invite the authors to address them in the manuscript.  

The presented methodology is limited to geotagged tweets. The fact that only 1 to 2 per cent of all tweets are geotagged can prevent one from drawing a firm conclusion about the outcomes of the study. Please address and discuss how it impacts the results of your study.

Most of the discussion section reviews (and repeats) the results. While the authors briefly discussed the implications of their work (@lines 405 to 411), it is not clear how such remarks are drawn and how they relate to the results. I would suggest rewriting the discussion section and explaining the importance of the results compared with other studies and what are the implications in a broader context (e.g. how your study can directly impact the emergency response/ decision-making / etc.).

I wish the authors best of luck in their important work.

Author Response

In the revision, we elaborated in more detail on the effects of this data limitation on analysis results and methods. It can be found following Table 5 in the discussion section. We pointed out that the low percentage of tweets with precise location tags has the following effects:
- results do not necessarily represent the entire Twitter user population
- it renders certain tests for spatial movements patterns inapplicable due to the small sample size. We provided the example of Zipf’s law for re-visitation frequencies of locations which we tested but could not report on.
Furthermore, we pointed out that Twitter, as it appears from one of their recent tweets, will turn off precise location tagging for regular posts altogether but instead offer this function to tagging multimedia content only through their camera app. We explained what this means for the availability and the abundance of geo-tagged tweets, e.g. by reporting the percentage of tweets with multimedia content that we found in one of our recent research studies. We also discussed that cell phone based GPS tracking points are likely the more effective means for movement analysis, disaster management and decision-making in the aftermath of natural disasters due to their data abundance, but also addressed some of the limitations of this method. Overall we hope that our new statements add to the discussion on the usability of different geo-positioning methods in the context of natural crisis analysis and management. 

We rewrote most of the discussion section. We shifted the focus from primarily reporting our research results to elaborating on how our findings complement those from other related studies and added new references as well. While the direct practical implications for rescue operations and decision making obtained from our results are somewhat limited we still aimed to better ground provided recommendations on our found research results. We also provided more justifications as to what may have led to shorter displacements and smaller activity spaces during hurricanes, including a new analysis on tweeting behavior around supply facilities during the hurricanes through a series of chi-squares tests of independence (section 4.3), a summary of numbers on road and bridge closures in the affected areas that were added in Table 5, and a set of daily precipitation maps for the different analysis regions (figure 10). We are confident that this new information adds useful content to the discussion about changed travel behavior during hurricanes.

Round 2

Reviewer 1 Report

P.13, ln.354: “North/Carolina”->”North/South Carolina”.

P.18, ln.525: “part”->”parts”.

 

There are some other related issues and stakeholders in disaster management such as insurance companies, home owners and public utility companies. They has been discussed in Kesete et al. (2014), Peng et al. (2014),  Iannell and Dell’Acqua (2017), Shan et al. (2016, 2017). The authors are encouraged to cite and discuss them.

 

References:

Kesete, Y., J. Peng, X. G. Shan, Y. Gao, R. Davidson, L. K. Nozick, and J. Kruse. “Modeling Insurer-homeowner Interactions in Managing Natural Disaster Risk”, Risk Analysis, 34(6): 1040-1055, 2014.

Peng J., X. G. Shan, Y. Kesete, Y. Gao, R. Davidson, L. K. Nozick, and J. Kruse. “Modeling the Integrated Roles of Insurance and Retrofit in Managing Natural Disaster Risk: A Multi-stakeholder Perspective”, Natural Hazards, 74(2): 1043-1068, 2014.

Shan X., J. Peng, Y. Kesete, Y. Gao, R. Davidson, J. Kruse, and L. K. Nozick, “Market Insurance and Self-insurance through Retrofit: analysis for hurricane risk in North Carolina”, Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering, in press, 2016.

Iannelli, G.C.; Dell’Acqua, F. Extensive Exposure Mapping in Urban Areas through Deep Analysis of Street-Level Pictures for Floor Count Determination. Urban Sci. 2017, 1(2), 16.

Shan, X., F. A. Felder, and D. W. Coit, “Game-theoretic Model for Electric Distribution Resiliency/Reliability from a Multiple Stakeholder Perspective”,IISE Transactions, 49(2): 159-177, 2017.

Author Response

“North/Carolina”->”North/South Carolina” and “part”->”parts”. These issues have been fixed. We could not see a direct connection between the references listed by the reviewer (insurer’s role in catastrophe risk management, mapping floor numbers of buildings using street level pictures) and the topic of our paper (effect of hurricanes on travel patterns). Also, the reviewer did not specify which related issues specifically should be discussed. However, when glancing through the papers we noted that insurances claims could be useful to estimate the severity of a hurricane across the affected area, and that this could be potentially tied to changes in travel behavior in future work. Therefore, we added this information, together with two new references, at the end of the discussion section.

Author Response File: Author Response.docx