Quantifying the economic impact of disasters on businesses using mobility data

Yabe, Takahiro; Zhang, Yunchang; Ukkusuri, Satish V. Working Paper Quantifying the economic impact of disasters on businesses using mobility data ADBI Working Paper Series, No. 1182 Provided in Cooperation with: Asian Development Bank Institute (ADBI), Tokyo Suggested Citation: Yabe, Takahiro; Zhang, Yunchang; Ukkusuri, Satish V. (2020) : Quantifying the economic impact of disasters on businesses using mobility data, ADBI Working Paper Series, No. 1182, Asian Development Bank Institute (ADBI), Tokyo This Version is available at: https://hdl.handle.net/10419/238539 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/3.0/igo/ ADBI Working Paper Series QUANTIFYING THE ECONOMIC IMPACT OF DISASTERS ON BUSINESSES USING MOBILITY DATA Takahiro Yabe, Yunchang Zhang, and Satish V. Ukkusuri No. 1182 September 2020 Asian Development Bank Institute The Working Paper series is a continuation of the formerly named Discussion Paper series; the numbering of the papers continued without interruption or change. ADBI’s working papers reflect initial ideas on a topic and are posted online for discussion. Some working papers may develop into other forms of publication. Suggested citation: Yabe, T., Y. Zhang, and S. Ukkusuri. 2020. Quantifying the Economic Impact of Disasters on Businesses Using Mobility Data. ADBI Working Paper 1182. Tokyo: Asian Development Bank Institute. Available: https://www.adb.org/publications/quantifying-economic-impact-disastersbusinesses-using-mobility-data Please contact the authors for information about this paper. Emails: [email protected] Takahiro Yabe is a PhD candidate at Lyles School of Civil Engineering of Purdue University. Yunchang Zhang is a PhD student at Lyles School of Civil Engineering of Purdue University. Satish V. Ukkusuri is a professor at Lyles School of Civil Engineering of Purdue University. The views expressed in this paper are the views of the author and do not necessarily reflect the views or policies of ADBI, ADB, its Board of Directors, or the governments they represent. ADBI does not guarantee the accuracy of the data included in this paper and accepts no responsibility for any consequences of their use. Terminology used may not necessarily be consistent with ADB official terms. Working papers are subject to formal revision and correction before they are finalized and considered published. Asian Development Bank Institute Kasumigaseki Building, 8th Floor 3-2-5 Kasumigaseki, Chiyoda-ku Tokyo 100-6008, Japan Tel: +81-3-3593-5500 Fax: +81-3-3593-5571 URL: www.adbi.org E-mail: [email protected] © 2020 Asian Development Bank Institute ADBI Working Paper 1182 Yabe, Zhang, and Ukkusuri Abstract In recent years, extreme shocks (such as natural disasters) have increased in both frequency and intensity. Consequently, many cities have experienced significant economic losses. Quantifying the economic cost to local businesses after extreme shocks is important for postdisaster assessment and pre-disaster planning. Conventionally, surveys have been the primary source of data to quantify the damages that are inflicted on businesses by disasters. However, surveys often suffer from high cost and their implementation can take a long time. They also suffer from spatio-temporal sparsity in observations and limitations in scalability. Recently, large scale human mobility data (e.g., mobile phone GPS) have been used to observe and analyze human mobility patterns in an unprecedented spatio-temporal granularity and scale. In this work, we use location data that were collected from mobile phones to estimate and analyze the causal impact of hurricanes on business performance. To quantify the causal impact of the disaster, we use a Bayesian structural time series model to predict the counterfactual performance of affected businesses (what if the disaster did not occur?), which may use the performance of other businesses outside the disaster areas as covariates. We have tested our method by quantifying the resilience of 635 businesses across nine categories in Puerto Rico after Hurricane Maria. Furthermore, hierarchical Bayesian models are used to reveal the effect of business characteristics (e.g., location and category) on the long-term resilience of these businesses. This study presents a novel and more efficient method to quantify business resilience, which could assist policy makers in disaster preparation and relief processes. Keywords: disaster resilience, mobile phones, human mobility, causal inference JEL Classification: Q54, C54, J6 1 Introduction Natural hazards are currently increasing in both frequency and intensity in many parts of the world. The economic losses caused by these extreme events have exceeded a total of $2.5 trillion across the globe since 2000, and are rising each year due to rapid urbanization in many cities [ 1 ]. With the intensifying threat of significant economic damage, the question of how to improve the resilience of cities has attracted interest from a wide range of fields, including public policy, urban planning, complex systems, and economics [ 2 ]. Among the various dimensions of disaster resilience, the ability of businesses to bounce back afterwards is a critical component that significantly contributes to the economic recovery of cities after disasters. Previous studies have analyzed the post-disaster recovery of businesses through the means of surveys and interviews. These studies have identified factors such as the pre-disaster size of the business and category of business to partly explain the reopening and demise of businesses after disasters, including Hurricanes Katrina [ 3 , 4 ], Andrew [ 5 ], and (more recently) Harvey [ 6 ]. Although these studies provide a general understanding of the effect of various characteristics of businesses that affect the post-disaster recovery performance, they suffer from two critical drawbacks. First, observations are limited to discrete measurements at a few number of timings, which fails to give a quantifiable, continuous and longitudinal understanding of the recovery process of businesses. Second, the applied methods fail to model the causal effect of the disaster, which requires a statistical framework that predicts the performance of businesses if the disaster did not occur. With the emergence of novel and often large-scale data collected from mobile sensors and online social platforms, we are now capable of observing and analyzing the dynamics of people, goods, and information at an unprecedented spatio-temporal granularity [ 7 ]. In particular, location data collected from mobile phones (e.g. call detail records, GPS trajectories) have enabled us to observe individual mobility patterns at an unprecedented high spatio-temporal granularity [ 8 , 9 ]. These datasets are now utilized for a wide range of applications to solve urban challenges, including population density estimation [ 10 , 11 ], traffic estimation [ 12 , 13 ], predicting poverty [ 14 ], and modeling the spread of epidemics [ 15 ]. In the context of extreme events, several studies have used mobile phone data to analyze mobility patterns during and after disasters, such as earthquakes [ 16 , 17 , 18 ], cyclones [ 19 ], and other anomalous events [ 20 ]. Despite this progress, none of the previous studies have used large scale mobility data to analyze the recovery of businesses after disasters. Recent advances in statistical models, in particular Bayesian structural time series (BSTS) models, allow us to make flexible predictions of time series data, which can be used to estimate the causal impact [ 21 ]. BSTS models have several advantages over conventional difference in differences models [ 22 ], including their flexibility to model the causal impact over a longitudinal time horizon rather that across two time points. A recent study using website click-through data applied BSTS models to quantify the causal impact of an online advertisement [ 23 ]. We aim to take advantage of this recently proposed methodology to quantify the causal impact of hurricanes on businesses in Puerto Rico. This study makes several contributions to overcome the drawbacks in the previous studies on business recovery after disasters. First, this is the first work to utilize large scale mobility data collected from mobile phones to estimate the popularity of businesses before, during, and after a disaster. Second, a Bayesian structural time series model combined with an inter-city matching scheme is proposed to infer the causal impact of the disaster on businesses. Third, the proposed methodology is applied on mobile phone data collected from Puerto Rico to quantify the resilience of businesses after Hurricane Maria. Figure 1 gives an overview of this study. The causal inference procedure is composed of three steps. i) To measure the causal impact of the disaster on business i , we first identify a similar business j in another region that was not affected by the disaster. ii) We then predict the counterfactual (“what-if the disaster did not occur?”) visit count of i after the disaster timing using observed data from j , via a Bayesian structural time series model. iii) Finally, we can quantify the causal impact of the disaster by taking the difference between the predicted and observed visit counts in i. 2 Related Works 2.1 Resilience of businesses after disasters The economic impact of disasters on businesses have conventionally been studied through surveys that are performed after the disaster. Studies using surveys have identified various factors that affect the reopening and demise of businesses after disasters through econometric models (e.g. logistic regression) [ 3 , 4 ]. Important factors that affected the outcomes of businesses after Hurricane Katrina include the household size of the business owner, previous disaster experience, number of employees, business age, and the legal structure of the business [ 3 ]. The qualitative details in the collected data are a significant advantage of these surveys. However, these surveys suffer from several drawbacks, including the high cost and long implementation time, spatio-temporal sparsity in observations, and limitations in scalability. Due 2 Figure 1: Overview of study. Our causal inference procedure is composed of three steps. i) To measure the causal impact of the disaster on business i , we first identify a similar business j in another region that was not affected by the disaster. ii) We then predict the counterfactual (“what-if the disaster did not occur?”) visit count of i after the disaster timing using observed data from j . iii) Finally, we can quantify the causal impact of the disaster by taking the difference between the predicted and observed visit counts in i. to these limitations, it is difficult to obtain a quantifiable, continuous, and longitudinal understanding of the recovery process of these businesses. Moreover, the applied methods fail to model the causal effect of the disaster, which requires a statistical framework that predicts the performance of businesses if the disaster did not occur. 2.2 Mobility analysis using mobile phone data With the emergence of novel and often large-scale data collected from mobile sensors and online social platforms, we are now capable of observing and analyzing the dynamics of people, goods, and information at an unprecedented spatio-temporal granularity [7]. In particular, location data collected from mobile phones (e.g. call detail records, and GPS trajectories) have enabled us to observe individual mobility patterns at an unprecedented high spatio-temporal granularity [ 8 , 9 ]. These new datasets are becoming new standards for population level studies, and they are used to understand the population distribution in cities [ 10 ]. Furthermore, these datasets are now utilized for a wide range of applications to solve urban challenges, including population density estimation [ 11 ], estimation of dynamic traffic flows [ 12 , 13 ], predicting poverty in developing counties [ 14 ], and modeling the impact of human mobility patterns on the spread of epidemics [ 15 ]. In the context of extreme events and disasters, several studies have used mobile phone data to analyze mobility patterns during and after disasters [ 16 , 17 , 18 ]. Studies using this large scale data have revealed important insights on the evacuation and migration patterns of the affected people [ 16 , 19 ]. However, despite this progress, none of the previous studies have used large scale mobility data to analyze the recovery of businesses after disasters. A recent study using mobile phone GPS data (which is the same data as used in this study) has revealed the impact of the recent policy regarding the usage of bathrooms in Starbucks on the visit behavior of people to the cafe chain [ 24 ]. They validated that the spatio-temporal granularity of the mobile phone GPS data is of sufficient detail to analyze the store level visit behavior. In this study, we apply a similar approach and estimate the visit behavior of people to stores and businesses using mobile phone GPS data. 3 2.3 Statistical methods for causal inference 2.3.1 Difference in Differences The difference in differences (DiD) method is a statistical method that is used to estimate the treatment effects between the "treatment" group versus the "control" group. For a before-and-after study, DiD compares the average change over time between treatment and control groups. This provides us with a classical method to estimate the causal effects of natural experiments without strict randomization [ 22 , 25 ]. However, DiD has several limitations: First, it follows the parallel trends assumption, which requires that the differences between treatment and control group are invariant overtime in absence of the treatment [ 26 , 27 ]. In a before-and-after study, the parallel trends assumption necessitates that the means of two groups should be balanced overtime. Consequently, issues such as time-correlated responses will contaminate the causal inference with DiD [ 28 ]. Second, only two time steps (i.e., pre-treatment time and post-treatment time) are considered in the classical DiD, which merely captures the static causal effects for a specific before-and-after study. This can be implausible and useless if the outcome of interest dynamically changes over time, such as recovery patterns after disaster, radioactive decay and so on [23]. 2.3.2 Bayesian Structural Time Series Models Compared with the classical DiD model, a structural time series model promisingly relaxes the parallel trends assumption and captures the variations of time-varying local trends and seasonality for time-correlated response variables [ 21 , 29 ]. In addition, structural time series models encompass a flexible model structure, which enables us to analyze the dynamic effects of the outcome of interest during a time period [ 30 ]. Due to a large number of predictors in structural time series models, a Bayesian approach was introduced to sparse the estimation of coefficients. Scott and Varian [ 31 , 32 ] proposed a spike-and-slab prior to the regression coefficients in a Google search query study, which significantly reduces the size of the problem. Nakajima and West [ 33 ] elicited a dynamic spike-and-slab prior that sparsified the estimation of time-varying parameters for a Bayesian macroeconomic time series model. The most recent Google study for causal inference of a market intervention [ 23 ] slightly revised the dynamic version of pike-and-slab prior [ 33 ] with a weakly informative prior. In addition, Bayesian structural time series models (BSTS) have been constructed to strengthen causal inference for time series data. To address the fundamental problem in causal inference [ 34 ], pre-treatment observations are trained and tested via BSTS. Consequently, the fitted BSTS can simulate the counterfactual as the synthetic post-treatment controls via posterior predictive samples. This method is extensively applied in causal inference throughout a wide variety of fields, such as socio-economics [ 35 , 36 ], political science [ 37 , 38 ], and environmental studies [39, 40]. The causal inference methodology that was proposed in the previous section is applied on data collected from Puerto Rico before and after Hurricane Maria, which made landfall on September 20, 2019, and caused a long term devastating humanitarian and economic crisis. Although the number of fatalities as a consequence of Maria are still under investigation, recent estimates suggest that between 793 to 8,498 excess deaths occurred following the storm [ 41 ]. Heavy rainfall, flooding, storm surge, and high winds caused considerable damage to various infrastructure systems, causing power outages and water shortages for the entire island for months. The total economic losses to Puerto Rico and the US Virgin Islands are estimated to be $90 billion, with a 90% confidence range of ± $25.0 billion, which makes Maria the third costliest hurricane in US history, behind Katrina (2005) and Harvey (2017) [42]. 3 Data Three main data sources are used in this study: (1) business visit data collected from mobile phones, (2) spatial distribution of housing damages due to Hurricane Maria, and (3) socio-economic factors of census blocks in Puerto Rico. In this section, we describe how these datasets were collected, processed, and used to infer the causal impact of the hurricane on businesses. 3.1 Business visit data collected from mobile phones Establishment-level visit data are provided by Safegraph 1 , which is a company that aggregates anonymized location data collected from smartphone applications to provide insights about physical places. Safegraph’s location dataset covers around 10% of all smartphones in the United States, and each observation is consisted of a unique (but anonymized) user ID, longitude, latitude, and timestamp information. The longitude and latitude information are accurate to within a few meters. This allows us to analyze the visit counts to each establishment. To detect a user visiting an establishment, the location data are first cleaned by removing GPS signal drifts and jumpy observations using a spatial threshold. The 1https://www.safegraph.com/ 4 Table 1: Summary statistics of business visit time series data. Business Category Region Puerto Rico Downstate New York Upstate New York Building Material 62 36 584 Gasoline Stations 10 143 1,160 Grocery Stores 34 97 1,227 Hospitals 12 25 76 Hotels 8 81 692 Restaurants 322 585 6,352 Supermarkets 61 0 52 Telecommunication 101 9 67 Universities 25 34 199 Total 635 1,102 10,409 data are then clustered into a staypoint using a spatio-temporal DBSCAN algorithm. Then, the visited establishment is predicted from establishments nearby the clustered staypoint by using a machine learning algorithm that takes into account features such as distances from establishment to the cluster centroid, time of day, and the North American Industry Classification System (NAICS) code. Performing this procedure for all days in the dataset produces a time series data of daily visit counts for each establishment. We use daily visit data of establishments located in Puerto Rico and the State of New York between January 2017 and March 2018 to quantify the causal impact of the hurricane on business resilience. Daily visit data of businesses in New York are used because these businesses constitute a reasonable control group that was not affected by the disruptions caused by Hurricane Maria. We will describe how we use the visit data from the control group in the causal inference model in the methods section. We limit our analysis to business categories that sell products or services directly to the customers because we will approximate business performance from the number of visits per day, as observed from mobile phone data. We also limit the analysis to medium or large sized businesses with more than 100 customers per day on average (before the disaster) because we are unable to observe visit patterns below that level using mobile phone data. As summarized in Table 1, the daily visit data of a total of 635 businesses in Puerto Rico were analyzed, along with 1,102 and 10,409 businesses in Manhattan and Up-State New York, respectively. Figure 2A plots the locations of the different types of businesses in Puerto Rico, and Figure 2B shows the three regions which were used for modeling the spatial differences in disaster impacts. 3.2 Socio-economic data In this study, the population and income data of each county were used in the later analysis. Population data were obtained from the US National Census 2 , and median income data were obtained from the American Community Survey3. 3.3 Spatial distribution of housing damages due to Hurricane Maria The physical damage caused by the hurricane is measured by the housing damage rates in each county, which was provided through the “Housing Assistance Data” provided by the Federal Emergency Management Agency (FEMA). The raw data can be found through the following link 4 . We defined “housing damage rate” for each county as the total number of houses that were inspected to have had more than $ 10,000 worth of damage due to the target hurricane, which is divided by the number of households in that county. Many of the counties in Puerto Rico experienced high housing damage rates of between 20% and 60%. 2https://www.census.gov/ 3https://www.census.gov/programs-surveys/acs 4https://www.fema.gov/media-library/assets/documents/34758 5 Figure 2: Characteristics of businesses in Puerto Rico. (A) Business locations and categories in Puerto Rico. (B) 3 regions of Puerto Rico used in this study. 4 Methods 4.1 Bayesian structural time series model The basic structural time series model is defined as follows:    yt,i =µt,i +τt,i +βxt,i +t,i ∀t t,i ∼ N(0, σ2 y) σy∼Cauchy(0,2.5) (1) where yt,i is the observed daily visits to business i on day t in the target region (in our case, Puerto Rico). yt,it is predicted by state components µt,i , τt,i and βxt,i that capture critical features of the time-series data [ 23 ]. A weakly informative prior is elicited for each state component. A graphical representation of the model is shown in Figure 3. Local Level Trend: The local level model represents local variations of the time series data. To simplify the model structure, we assume that the mean of the trend is a random walk with the initialization of µ1:      µt+1,i =µt,i +η1,t,i ∀t > 1 µ1,i ∼ N(µ0, σ2 0) η1,t,i ∼ N(0, σ2 µ) σ0, µ0, σµ∼Cauchy(0,2.5) (2) Seasonality: Let S denote the total number of seasons. The sum of seasonal effects over S time periods is assumed to be zero. In this study, weekly seasonality is taken into account ( S= 7 ) with the initialization of τ1,i , τ2,i , τ3,i , τ4,i , τ5,i , and τ6,i:        τt+1,i =−PS−2 s=0 τt−s,i +η2,t ∀t > 1 τ1,i, τ2,i, τ3,i, τ4,i, τ5,i, τ6,i ∼ N(µτ0, σ2 τ0) η2,t ∼ N(0, σ2 τ) µτ0, στ0, στ∼Cauchy(0,2.5) (3) 4.1.1 Choice of Covariates Apart from the local level model and seasonality, there are other unobserved effects (e.g., impacts of holidays and sport events) that may contaminate the estimation of the yt,i . To capture the unobserved heterogeneity, xt,i in Equation (1) is 6 8 Conclusion Quantifying the economic impact of disasters to businesses is crucial for disaster relief and preparation. The availability of large scale human mobility data enables us to observe daily visit counts to businesses in an unprecedented spatiotemporal granularity. In this work, we presented a methodology to estimate the causal impact of disasters to businesses from mobile phone location data, using a Bayesian modeling framework. We used this methodology to quantify the causal impact of Hurricane Maria on businesses in Puerto Rico. The estimation results provide insights into what types of businesses (and their locations) are able to recovery quickly after the hurricane. These insights could assist policy makers during disaster preparation and relief processes. References [1] UNDRR. Global assessment report on disaster risk reduction 2013: From shared risk to shared value: the business case for disaster risk reduction, 2013. [2] Hallie Eakin, Luis A Bojórquez-Tapia, Marco A Janssen, Matei Georgescu, David Manuel-Navarrete, Enrique R Vivoni, Ana E Escalante, Andres Baeza-Castro, M Mazari-Hiriart, and Amy M Lerner. Opinion: urban resilience efforts must consider social and political forces. Proceedings of the National Academy of Sciences, 114(2):186–189, 2017. [3] Maria I Marshall, Linda S Niehm, Sandra B Sydnor, and Holly L Schrank. Predicting small business demise after a natural disaster: an analysis of pre-existing conditions. Natural Hazards, 79(1):331–354, 2015. [4] Sandra Sydnor, Linda Niehm, Yoon Lee, Maria Marshall, and Holly Schrank. Analysis of post-disaster damage and disruptive impacts on the operating status of small businesses after hurricane katrina. Natural Hazards, 85(3):1637–1663, 2017. [5] Gary R Webb, Kathleen J Tierney, and James M Dahlhamer. Predicting long-term business recovery from disaster: A comparison of the loma prieta earthquake and hurricane andrew. Global Environmental Change Part B: Environmental Hazards, 4(2):45–58, 2002. [6] Jim Lee. Business recovery from hurricane harvey. International Journal of Disaster Risk Reduction, 34:305–315, 2019. [7] Michael Batty. Big data, smart cities and city planning. Dialogues in Human Geography, 3(3):274–279, 2013. [8] Marta C Gonzalez, Cesar A Hidalgo, and Albert-Laszlo Barabasi. Understanding individual human mobility patterns. nature, 453(7196):779, 2008. [9] Vincent D Blondel, Adeline Decuyper, and Gautier Krings. A survey of results on mobile phone datasets analysis. EPJ data science, 4(1):10, 2015. [10] Pierre Deville, Catherine Linard, Samuel Martin, Marius Gilbert, Forrest R Stevens, Andrea E Gaughan, Vincent D Blondel, and Andrew J Tatem. Dynamic population mapping using mobile phone data. Proceedings of the National Academy of Sciences, 111(45):15888–15893, 2014. [11] NA Wardrop, WC Jochem, TJ Bird, HR Chamberlain, D Clarke, D Kerr, L Bengtsson, S Juran, V Seaman, and AJ Tatem. Spatially disaggregated population estimates in the absence of national population and housing census data. Proceedings of the National Academy of Sciences, 115(14):3529–3537, 2018. [12] Md Shahadat Iqbal, Charisma F Choudhury, Pu Wang, and Marta C González. Development of origin–destination matrices using mobile phone call data. Transportation Research Part C: Emerging Technologies, 40:63–74, 2014. [13] Francesco Calabrese, Giusy Di Lorenzo, Liang Liu, and Carlo Ratti. Estimating origin-destination flows using opportunistically collected mobile phone location data from one million users in boston metropolitan area. IEEE Pervasive Computing, 10(4):36–44, 2011. [14] Joshua Blumenstock, Gabriel Cadamuro, and Robert On. Predicting poverty and wealth from mobile phone metadata. Science, 350(6264):1073–1076, 2015. [15] Linus Bengtsson, Jean Gaudart, Xin Lu, Sandra Moore, Erik Wetter, Kankoe Sallah, Stanislas Rebaudet, and Renaud Piarroux. Using mobile phone data to predict the spatial spread of cholera. Scientific Reports, 5:8923, 2015. [16] Xin Lu, Linus Bengtsson, and Petter Holme. Predictability of population displacement after the 2010 haiti earthquake. Proceedings of the National Academy of Sciences, 109(29):11576–11581, 2012. 13 [17] Xuan Song, Quanshi Zhang, Yoshihide Sekimoto, and Ryosuke Shibasaki. Prediction of human emergency behavior and their mobility following large-scale disaster. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 5–14. ACM, 2014. [18] Robin Wilson, Elisabeth zu Erbach-Schoenberg, Maximilian Albert, Daniel Power, Simon Tudge, Miguel Gonzalez, Sam Guthrie, Heather Chamberlain, Christopher Brooks, Christopher Hughes, et al. Rapid and near real-time assessments of population displacement using mobile phone data following disasters: the 2015 nepal earthquake. PLoS currents, 8, 2016. [19] Xin Lu, David J Wrathall, Pål Roe Sundsøy, Md Nadiruzzaman, Erik Wetter, Asif Iqbal, Taimur Qureshi, Andrew Tatem, Geoffrey Canright, Kenth Engø-Monsen, et al. Unveiling hidden migration and mobility patterns in climate stressed regions: a longitudinal study of six million anonymous mobile phone users in bangladesh. Global Environmental Change, 38:1–7, 2016. [20] James P Bagrow, Dashun Wang, and Albert-Laszlo Barabasi. Collective response of human populations to large-scale emergencies. PloS one, 6(3):e17680, 2011. [21] Andrew Harvey and SJ Koopman. Structural time series models. Wiley StatsRef: Statistics Reference Online, 2014. [22] Richard Ernest Bellman and Kenneth L Cooke. Differential-difference equations. 1963. [23] Kay H Brodersen, Fabian Gallusser, Jim Koehler, Nicolas Remy, Steven L Scott, et al. Inferring causal impact using bayesian structural time-series models. The Annals of Applied Statistics, 9(1):247–274, 2015. [24] Umit Gurun and David Solomon. The perils of private provision of public goods, 2019. [25] Andrew Goodman-Bacon. Difference-in-differences with variation in treatment timing. Technical report, National Bureau of Economic Research, 2018. [26] Alberto Abadie. Semiparametric difference-in-differences estimators. The Review of Economic Studies, 72(1):1–19, 2005. [27] John Antonakis, Samuel Bendahan, Philippe Jacquart, and Rafael Lalive. On making causal claims: a review and recommendations. The leadership quarterly, 21(6):1086–1120, 2010. [28] Christian B Hansen. Asymptotic properties of a robust variance matrix estimator for panel data when t is large. Journal of Econometrics, 141(2):597–620, 2007. [29] Andrew C Harvey. Forecasting, structural time series models and the Kalman filter. Cambridge university press, 1990. [30] Peter SH Leeflang, Tammo HA Bijmolt, Jenny Van Doorn, Dominique M Hanssens, Harald J Van Heerde, Peter C Verhoef, and Jaap E Wieringa. Creating lift versus building the base: current trends in marketing dynamics. International Journal of Research in Marketing, 26(1):13–20, 2009. [31] Steven L Scott and Hal R Varian. Predicting the present with bayesian structural time series. Available at SSRN 2304426, 2013. [32] Steven L Scott and Hal R Varian. Bayesian variable selection for nowcasting economic time series. Technical report, National Bureau of Economic Research, 2013. [33] Jouchi Nakajima and Mike West. Bayesian analysis of latent threshold dynamic models. Journal of Business & Economic Statistics, 31(2):151–164, 2013. [34] Guido W Imbens and Donald B Rubin. Rubin causal model. The New Palgrave Dictionary of Economics, pages 1–10, 2017. [35] Andrew C Harvey, Thomas M Trimbur, and Herman K Van Dijk. Trends and cycles in economic time series: a bayesian approach. Journal of Econometrics, 140(2):618–649, 2007. [36] Obryan Poyser. Exploring the dynamics of bitcoin’s price: a bayesian structural time series approach. Eurasian Economic Review, 9(1):29–60, 2019. [37] Patrick T Brandt and John R Freeman. Advances in bayesian time series modeling and the study of politics: theory testing, forecasting, and policy analysis. Political Analysis, 14(1):1–36, 2006. [38] Patrick T Brandt, John R Freeman, and Philip A Schrodt. Real time, time series forecasting of inter-and intra-state political conflict. Conflict Management and Peace Science, 28(1):41–64, 2011. [39] Yu Jiang, Zhe Song, and Andrew Kusiak. Very short-term wind speed forecasting with bayesian structural break model. Renewable Energy, 50:637–647, 2013. 14 [40] Frank de Vocht. Inferring the 1985–2014 impact of mobile phone use on selected brain cancer subtypes using bayesian structural time series and synthetic controls. Environment international, 97:100–107, 2016. [41] Nishant Kishore, Domingo Marqués, Ayesha Mahmud, Mathew V Kiang, Irmary Rodriguez, Arlan Fuller, Peggy Ebner, Cecilia Sorensen, Fabio Racy, Jay Lemery, et al. Mortality in puerto rico after hurricane maria. New England journal of medicine, 2018. [42] Richard J Pasch, Andrew B Penny, and Robbie Berg. National hurricane center tropical cyclone report: Hurricane maria. TROPICAL CYCLONE REPORT AL152017, National Oceanic And Atmospheric Administration and the National Weather Service, pages 1–48, 2018. [43] Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin. Bayesian data analysis. Chapman and Hall/CRC, 2013. 15