USER CONTRIBUTION AND ITS SOCIAL - WELFARE VALUE IN A MOBILE NAVIGATION APP FOR REAL - TIME TRAFFIC INFORMATION AROUND URBAN AREAS By Tae Hun Kim A DISSERTATION Submitted to Michigan State University in partial fulfilment of the requirement s for the degree of Business Administration Business Information Systems Doctor of Philosophy 2018 ABSTRACT USER CONTRIBUTION AND ITS SOCIAL - WELFARE VALUE IN A MOBILE NAVIGATION APP FOR REAL - TIME TRAFFIC INFORMATION AROUND URBAN AREAS By Tae Hun Kim Today , users of mobile devices (smartphones and tablets) adopt a variety of apps, use social features, and engage in crowdsourcing content as a public good. Th is dissertation explains their user community, i.e., a mobile virtual community, in terms of user cont ribution and its social - welfare value around urban areas. Essay 1 conceptualizes a virtual and spatial factor, i.e., virtual crowdedness, and addresses its role in encouraging user contribution. The findings are theoretically explained by the tension betwe en prosocial behavior of and bystander effect on the mobile virtual community. Essay 2 theorizes whether and how user contribution, attributed to self - interest, supports social welfare for the whole citizenry . I found that user contribution improv es the mo bility of urban transportation and reduc es social and economic costs. As an exemplar of citizen - data science, this dissertation takes a spatial and panel data approach to analyze the large - scale data on mobile app users, traffic conditions, and their locat ions in the urban region. The empirical findings are visualized and discussed to explain practical implications for mobile app design and policy on urban transportation. Copyright by TAE HUN KIM 2018 iv ACKNOWLEDGEMENTS First, I would like to present my gratitude to the co - chairs of my dissertation committee, Dr. Vallabh Sambamurthy and Dr. Anjana Susarla , who have been good role models for me . Their excellent directions have guided my research and Ph.D. studies to be completed successfully. I also hearti ly appreciate the empirical and analytical guidance from the other committee members, Dr. Chenhui Guo, Dr. Ashton Shortridge, and Dr. Roger Calantone , whose suggestions have been helpful in improving my dissertation in terms of the various aspects of resea rch. It was lucky for me to stay with all the faculty and staff members in Department of Accounting & Information Systems as well as at Broad College of Business for the last five years at Michigan State University. Their enjoyable lectures and thoughtful services made my Ph.D. studies success ful . I wish to present my warm thanks to the IS scholars who provided critical comments and constructive suggestions on my research at the IS conferences and job talk presentations. I would like to send my regards to D r. Jae - Nam Lee and all the other faculty members at Korea University Business School. Their kind introduction to the academic world has encouraged me to devote myself to actively engaging in research projects. I also wish to thank my friends and colleagues who gave me beautiful memories at Michigan State University, University of British Columbia, and Korea University. Finally, my parents and my sister who always love me and pray for my academic career. Thanks to their unconditional support, I could, can, a nd will be able to keep going as an enthusiastic scholar and hopefully as a good citizen. v TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ ........................ vii LI ST OF FIGURES ................................ ................................ ................................ ....................... xi ESSAY 1. PROSOCIAL BEHAVIOR OR BYSTANDER EFFECT? THE ROLE OF VIRTUAL CROWDEDNESS IN ENCOURAGING USER CONTRIBUTION OF A MOBILE VIRTUAL COMMUNITY ................................ ................................ ................................ ............................... 1 ABSTRACT OF ESSAY 1 ................................ ................................ ................................ ......... 1 INTRODUCTION ................................ ................................ ................................ ...................... 2 RESEARCH CONTEXT ................................ ................................ ................................ ............ 6 Waze App ................................ ................................ ................................ ................................ 6 Mobile Virtual Community ................................ ................................ ................................ ..... 8 THEORETICAL BACKGROUND ................................ ................................ ............................ 9 User Contribution ................................ ................................ ................................ .................. 10 Virtual Crowdedness ................................ ................................ ................................ ............. 12 Development of Hypotheses ................................ ................................ ................................ . 13 DATA ................................ ................................ ................................ ................................ ....... 19 Data Collection Approach ................................ ................................ ................................ ..... 19 Measurement of Study Variables and Their Statistics ................................ .......................... 20 EMPIRICAL APPROACH ................................ ................................ ................................ ....... 25 Zero - Inflated Count Data Models ................................ ................................ ......................... 25 Heckman S election Models ................................ ................................ ................................ .. 27 Matching Method ................................ ................................ ................................ .................. 28 RESULTS ................................ ................................ ................................ ................................ . 30 Results of Test ing Hypothesis 1 ................................ ................................ ........................... 31 Results of Testing Hypothesis 2 ................................ ................................ ........................... 36 Results of Testing Hypothesis 3 ................................ ................................ ........................... 36 Robustness Checks ................................ ................................ ................................ ................ 39 DISCUSSION ................................ ................................ ................................ ........................... 41 Implications for Research ................................ ................................ ................................ ..... 42 Implications for Practice ................................ ................................ ................................ ....... 44 Limitation and Future Direction ................................ ................................ ........................... 46 CONCLUSION ................................ ................................ ................................ ......................... 47 ESSAY 2. DOES USER ENGAGEMENT ENHANCE SOCIAL WELFARE? EFFECTIVENESS OF USER - CROWDSOURCED CONTENT IN IMPROVING URBAN TRANSPORTATION ................................ ................................ ................................ ................... 48 ABSTRACT OF ESSAY 2 ................................ ................................ ................................ ....... 48 INTRODUCTION ................................ ................................ ................................ .................... 49 RESEARCH CONTEXT ................................ ................................ ................................ .......... 51 New York City and Its T raffic Congestion ................................ ................................ ........... 51 User - Crowdsourced Content and Its Value ................................ ................................ .......... 52 vi THEORETICAL BACKGROUND ................................ ................................ .......................... 54 SPATIAL PANEL DATA ................................ ................................ ................................ ........ 57 Temporal Nature of Traffic Congestion Duration and User Contribution ............................ 61 Spatial Nature of Traffic Congestion Duration and User Contribution ................................ 62 EMPIRICAL APPROACH ................................ ................................ ................................ ....... 65 Spatial Autocorrelation ................................ ................................ ................................ ......... 65 Spatial Panel Estimators ................................ ................................ ................................ ....... 67 EMPIRICAL RESULTS AND FINDINGS ................................ ................................ ............. 70 Effects of User Co ntribution on Traffic Congestion Duration ................................ ............. 71 Dynamic Effects of User Contribution on Traffic Congestion Duration .............................. 75 Quadratic Re lationship of Primary Information and Traffic Congestion Duration .............. 78 Interactions Between Different Types of User Contribution ................................ ................ 81 Welfare Analysis ................................ ................................ ................................ ................... 87 Sensitivity and Robustness Checks ................................ ................................ ....................... 91 DISCUSSION ................................ ................................ ................................ ........................... 92 Theoretical and Empirical Contributions ................................ ................................ .............. 93 Implications for Policy and Practice ................................ ................................ ..................... 94 Limitations and Future Directions ................................ ................................ ........................ 99 CONCLUSION ................................ ................................ ................................ ....................... 101 APPENDICES ................................ ................................ ................................ ............................ 102 APPENDIX A. Correlation Analysis and Distribution of Boroughs and Weather Conditions ................................ ................................ ................................ ................................ ................ 103 APPENDIX B. Technical Details for Models ................................ ................................ ........ 104 APPENDIX C. Propensity - Scor e Matching Tests with Ten Groups ................................ ...... 107 APPENDIX D. Robustness Checks ................................ ................................ ........................ 108 ity ................................ ... 112 APPENDIX F. Spatial Nature of Traffic Congestion Duration and User Contribution ......... 114 APPENDIX G. Matching Method in Quasi - Natural Experimental Setting ........................... 116 APPENDIX H. Sensitivity Checks: Summary of Empirical Results from Spatial Panel Data ................................ ................................ ............... 119 APPENDIX I. Summary of Sensitivity and Robustness Checks ................................ ............ 130 APPENDIX J. Summary of Additional Robustness Checks ................................ .................. 13 3 REFERENCES ................................ ................................ ................................ ........................... 135 vii LIST OF TABLES Table 1. Hybrid characteristics of mobile virtual communities ................................ ...................... 9 Table 2. Definitions, descriptive statistics, and data sources of study variables .......................... 24 Table 3. Summary of hypotheses testing results ................................ ................................ ........... 31 Table 4. Results of zero - inflated count data models for initial contribution ................................ 32 Table 5. Results of propensity - score matching tests with five groups for initial contr ibution ..... 34 Table 6. Rush hours vs. non - rush hours: Results of zero - inflated count data models for initial contribution ................................ ................................ ................................ ................................ ... 37 Table 7. Results of Heckman selection models for feedback contribution ................................ ... 39 Table 8. Results of Heckman selection models for confirmative contribution ............................. 40 Table 9. Descriptive statistics of study variables and their data sources ................................ ...... 59 Table 10. Correlation matrix of study variables ................................ ................................ ........... 60 Table 11. Distributions of observations across boroughs and weather conditions in New York City ................................ ................................ ................................ ................................ ................ 60 Table 12. Statistics of spatial weight matrix iguity ...................... 64 Table 13. Diagnostics of spatial autocorrelations of study variables with spatial weight matrix ................................ ................................ ................................ .. 66 Table 14. Effects of user contribution on traffic congestion duration: Results of panel data models ................................ ................................ ................................ ................................ ........... 71 Table 15. Effects of user contribution on traffic congestion duration: R esults of spatial panel data models with spatial weight matrix ................................ ..... 73 Table 16. Effects of user contribution on traffic congestion duration for rush and business hours: Results of spatial panel data models with spatial weig ht matrix contiguity ................................ ................................ ................................ ................................ ...... 74 Table 17. Dynamic effects of user contribution on traffic congestion duration: Results of dynamic spatial panel data models with spatial weight matrix contiguity ................................ ................................ ................................ ................................ ...... 77 Table 18. Quadratic relationship of primary information and traffic congestion duration: Results of spatial panel data models with spatial weight matrix .... 78 viii Table 19. Quadratic relationship of primary information and traffic congestion duration for rush and business hours: Results of spatial panel dat a models with spatial weight matrix based on ................................ ................................ ................................ ................. 80 Table 20. Interaction between primary information and follow - up feedback: Results of spatial panel data models with s patial weight matrix .................... 81 Table 21. Interaction between primary information and follow - up feedback for rush and business hours: Results of spatial panel data models with spatial weight matrix contiguity ................................ ................................ ................................ ................................ ...... 82 Table 22. Interaction between primary information and collective confirmation: Results of spatial panel data models wit h spatial weight matrix .................... 83 Table 23. Interaction between primary information and collective confirmation for rush and business hours: Results of spatial pan el data models with spatial weight matrix based on ................................ ................................ ................................ ................. 84 Table 24. Interaction between follow - up feedback and collective confirmation: Results of spatial panel data mod els with spatial weight matrix .................... 85 Table 25. Interaction between follow - up feedback and collective confirmation for rush and business hours: Results of spati al panel data models with spatial weight matrix based on ................................ ................................ ................................ ................. 86 Table A1. Pairwise correlations between study variables ................................ .......................... 103 Table A2. Distributions of observations across boroughs and weather conditions .................... 103 Table A3. Results of propensity - score matching tests with ten g roups for initial contribution . 107 Table A4. Results of panel negative binomial models for initial contribution ........................... 108 Table A5. Re sults of panel Poisson models for initial contribution ................................ ........... 109 Table A6. Winter vs. summer: Results of zero - inflated count data models for initial contribution ................................ ................................ ................................ ................................ ..................... 110 Table A7. Winter vs. summer: Results of Heckman selection models for feedback and confirmative contribution ................................ ................................ ................................ ............ 111 Table A8. Statistics of spatial weight matrix ...................... 113 Table A9. Diagnostics of spatial autocorrelations of study variables with spatial weight matrix ................................ ................................ ................................ .. 114 Table A10. Results of propensity - score matching tests ................................ .............................. 118 ix Table A11. Effects of user contribution on traffic congestion duration: Results of spati al panel data models with spatial weight matrix .............................. 119 Table A12. Effects of user contribution on traffic congestion duration for rush and business hours: Results of spatial panel data models with spatial wei ght matrix contiguity ................................ ................................ ................................ ................................ .... 120 Table A13. Dynamic effects of user contribution on traffic congestion duration: Results of dynamic spatial panel data models with spatial weig ht matrix ................................ ................................ ................................ ................................ ..................... 121 Table A14. Quadratic relationship of primary information and traffic congestion duration: Results of spatial panel data models with spatial wei ght matrix contiguity ................................ ................................ ................................ ................................ .... 122 Table A15. Quadratic relationship of primary information and traffic congestion duration for rush and business hours: Results of spatial panel data models with spatial weight matrix based ................................ ................................ ................................ ............ 123 Table A16. Interaction between primary information and follow - up feedback: Results of spatial panel data models w ith spatial weight matrix .................... 124 Table A17. Interaction between primary information and follow - up feedback for rush and business hours: Results of spatial panel data models with spatial weight matrix based on ................................ ................................ ................................ ................. 125 Table A18. Interaction between primary information and collective confirmation: Results of spatial panel data mod els with spatial weight matrix ......... 126 Table A19. Interaction between primary information and collective confirmation for rush and business hours: Results of spa tial panel data models with spatial weight matrix based on ................................ ................................ ................................ ................. 127 Table A20. Interaction between follow - up feedback and collective confirmation: Results of spatial panel data models with spatial weight matrix ......... 128 Table A21. Interaction between follow - up feedback and collective confirmation for rush and business hours: Result s of spatial panel data models with spatial weight matrix based on ................................ ................................ ................................ ................. 129 Table A22. Sensitivity and robustness checks: Results of spatial autoregressive models .......... 130 Table A23. Sensitivity and robustness checks: Results of spatial Durbin models ..................... 131 Table A24. Sensitivity and robustness checks: Res ults of spatial error models and spatial autocorrelation models ................................ ................................ ................................ ................ 132 Table A25. Robustness check: Results of Tobit models ................................ ............................. 133 x Tab le A26. Robustness check: Results of Tobit models with observations for rush and business hours ................................ ................................ ................................ ................................ ............ 134 xi LIST OF FIGURES Figure 1. Waze app: GPS navigation, map, and social traffic tool ................................ ................. 7 Figure 2. Role - based conceptualization of user contribution ................................ ....................... 11 Figure 3. NTA - by - NTA densities of total use r contribution and Waze users .............................. 21 Figure 4. Hour - by - hour average density levels of total user contribution and Waze users .......... 22 Figu re 5. Distributions of virtual crowdedness with five groups for propensity - score matching 29 Figure 6. Relationship of virtual crowdedness and initial contribution ................................ ........ 33 Figure 7. Visualized results of propensity - score matching tests with five groups ....................... 35 Figure 8. Rush hours vs. non - rush hours: Relationship of virtual crow dedness and initial contribution ................................ ................................ ................................ ................................ ... 38 Figure 9. User - crowdsourced content on Waze navigation map ................................ .................. 53 Figure 10. Neighborhood tabula tion areas over boroughs of New York City .............................. 58 Figure 11. Hour - by - hour averages of traffic congestion duration and user contribution in New York City ................................ ................................ ................................ ................................ ...... 61 ............................... 62 Figure 13. Connectivity of spatial weight matrix ............... 64 Figure 14. Visualized spatial autocorrelations of study variables with spatial weight matrix ................................ ................................ ................................ .. 66 Figure 15. Flow chart of spatial panel model specifications and their relationships .................... 68 Figure 16. Quadratic relationship of primary information and t raffic congestion duration ......... 80 Figure 17. Interaction between primary information and follow - up feedback ............................. 83 Figure 18. Interactio n between primary information and collective confirmation ....................... 85 Figure 19. Interaction between follow - up feedback and collective confirmation ........................ 87 Figure 20. Annual congestion cost per driver saved thanks to total user contribution ................. 91 Figure 21. Screenshots of Waze app platform design ................................ ................................ ... 97 xii Figure A1. Distributions of virtual crowdedness with ten groups for propensity - score matching ................................ ................................ ................................ ................................ ..................... 107 Figure A2. Context - specific example of ................................ ................................ ................................ ................................ ..................... 112 Figure A3. Connectivity of spatial weight matrix .............. 113 Figure A4. Visualized spatial autocorrelations of study variables with spatial weight matrix ................................ ................................ ................................ .. 115 1 ESSAY 1. PROSOCIAL BEHAVIOR OR BYSTANDER EFFECT? THE ROLE OF VIRTUAL CROWDEDNESS IN ENCOURAGING USER CONTRIBUTION OF A MOBILE VIRTUAL COMMUNITY ABSTRACT OF ESSAY 1 S martphone users adopt a variety of mobile apps, use social features, and eng age in generating content as a public good in real time. Their adoption and use of social mobile apps contribute to user - generated content that is crowdsourced by mobile virtual communities. Specifically, a mobile navigation app, Waze, enables users to: (1 ) initially generate traffic content (initial contribution via alert postings) in real time, (2) additionally amplify the content (feedback contribution via comments) , and (3) collectively verify the content (confirmative contribution via thumbs - up) over t ime. This research conceptualize s a virtual and spatial factor, virtual crowdedness ( defined as the density of nearby Waze users who are potential contributors from the viewpoint of a focal user ) and examine s its role in encouraging user contribution. T he tension between prosocial behavior and bystander effect is explore d to explain the impact of virtual crowdedness on the three types of user contribution. This research analyze s a panel dataset of user contribution from locations in New York City. The findi ngs suggest that the tension varies by degree of virtual crowdedness, time of day, and types of user contribution. This research discuss es practical implications on how app designers should visualize Waze users and improve the design of social features for encouraging user contribution to the mobile virtual community. 2 INTRODUCTION T he digital economy is i ncreasingly powered by social, mobile, analytic, and cloud (SMAC) applications. Digital mediation of social interactions continues to attract attention in new business as well as research ( Brynjolfsson et al. 2013 ; Ghose and Han 2014 ) . While prior research has explained what motivates user contribution in social and digital platforms (e.g., Burtch et al. 2014 ; Tang et al. 2012 ) , the emergence of mobile app platforms for social interactions such as Waze, Kamino, and WeatherSignal transforms the nature of civic engagement through mobile virtual communities. Specificall y, this research investigate s a unique mobile app, Waze, with which users socially crowdsource, refine, and verify traffic content as wisdom of crowds in real time and over time. Unlike Apple and Google Maps, Waze provides users with social features. Their us age of the social features generates initial alert postings on traffic events and conditions and subsequent comments and thumbs - up on the alerts, thus contributing to user - generated content (UGC) crowdsourced by the mobile virtual community. The sociall y - crowdsourced traffic content goes beyond just informing Waze users of slow - moving traffic status by generating real - time details on why the traffic is slow: e.g., due to accidents ahead, stalled vehicles, and police officers nearby ( Perez 2015 ) . Mobile virtual communities, such as those centered around Waze, generate a rich trail of data that allow s researchers to study their unique aspects. Such user communities focus on specific goals (e.g., for Waze users, the goal is to navigate traffic in real time to improve their driving); user interactions are likely to be of shorter duration but of higher velocity; user communities are large in size but volatile in individual engagement rates. A mobile virtual community starts with a few active users, and then spreads to a number of other users around social and virtual spaces ( Rheingold 2000 ; Rheingold 2002 ) . That is, the mobile virtual community initially consists of small groups of phys ically close associates. However, certain 3 circumstances facilitate its social engagement and mobilize its large groups of users virtually over time. This research is motivated by social engagement in user contribution to mobile virtual communities. In addi tion to spatial factors, this research conceptualize s a virtual and spatial factor, virtual crowdedness, in encouraging user contribution to a public good, UGC via a navigation app. This research define s virtual crowdedness as the density of Waze users nea rby , which is distinct from physical crowdedness (e.g., the density of traffic on the roads). A focal user observes nearby users on her/his smartphone screen virtually when driving and using Waze. This virtual density is different from the physical crowded ness of potential contributors whom the focal user observes outside of her/his automobile. While physical crowdedness causes traffic congestion around urban areas, virtual crowdedness informs Waze users of the traffic status and existence of other users ar ound their local areas. This research analyze s the role of virtual crowdedness in encouraging three types of user contribution : initial contribution via alert postings, feedback contribution via comments, and confirmative contribution via thumbs - up . These three types of user contribution address the following research questions : (1) does initial contribution via alert postings increase along with virtual crowdedness in a positive relationship? ; (2) how does the magnitude of the relationship change as virtua l crowdedness increases? ; (3) how does the relationship of virtual crowdedness and initial contribution vary across rush hours and non - rush hours? ; (4) do feedback contribution and confirmative contribution via comments and thumbs - up increase along with vi rtual crowdedness? The current literature on user engagement in mobile commerce has restricted its attention to the use of apps for mobile commerce, social games, and networking services . 1 This research 1 For exampl e, Ghose and Han (2014 ) showed different consumer preferences according to a set of specific app characteristics and positive vs. negative effects of in - app purchase opti ons on app demand. Han et al. (2016 ) 4 focus es on virtual crowdedness, by which a focal user observes the usage of others virtually when driving and using the app, and how that influences the decision of user contribution. Then, this research analyze s the relationships of virtual crowdedness and the three types of user contribution to the UGC cro wdsourced by Waze users in New York City (NYC). The spatial data approach measures the densities of Waze users (virtual crowdedness) , traffic and population (physical crowdedness) , and the total volumes of the three types of user contribution in NYC. The a ggregate data approach at a granular unit enables us to measure the virtual and spatial differences of collective Waze users and investigate its impact on the total volumes of the different types of user contribution at the geographic level . For empirical analyses, this research employ s zero - inflated count data estimators and Heckman selection models , given that only a small proportion of users actively engage in user contribution to the general UGC platforms ( a large number of zero - valued observations rega rding user contribution ) . I adopt a propensity - score matching (PSM) method that implements a quasi - experimental research design. PSM estimator provides consistent results, while mitigating endogeneity concerns caused by confounding effects of multifaceted dynamics of traffic status in the urban areas. S everal robustness checks are also conduct ed with the multiple alternative estimators and sub - datasets across different seasons. The findings highlight the role of virtual crowdedness in encouraging Waze users to use the social features and engage in the three types of user contribution. First, while a positive relationship is observed between virtual crowdedness and initial contribution, the magnitude of explained the app choice behavior of consumers with baseline utility and satiation. Kwon et al. (2016 ) described digital vulnerabilities (the habitual consumption of and excessive dependence on social apps) with rational and myopic addictions. Other studies inve stigate the effects of contextual factors on the use of apps, the contextual ( Ghose and Han 2011 ) , location ( Molitor et al. 2015 ) , location and time at which consumers receive mobile advertisements ( Luo et al. 2013 ) , climate - and weather - related factors ( Molitor et al. 2012 ) , and product characteristi cs ( Bart et al. 2014 ) . 5 the positive relationship decreases as virtual crowdedne ss increases (a concave - down, increasing relationship). Waze users tend to engage in user contribution by posting alerts as the number of other users increases nearby. However, the magnitude of the positive relationship decreases as virtual crowdedness rea ches a high level. These findings suggest a concave - down, increasing relationship of virtual crowdedness and initial contribution over the whole range of virtual crowdedness. As this research hypothesize s , the concave relationship may be explained by the c ompeting effects of prosocial behavior of and bystander effect on Waze users. Second, the relationship is more concave - down for rush hours than for non - rush hours. This suggests that the tension between prosocial behavior of active users and bystander effe ct on passive users intensifies during rush hours when the real - time UGC is critically valuable to Waze users and useful for their driving. Third, Waze users usually engage in feedback contribution and confirmative contribution by commenting on and verifyi ng the initial alerts when they observe a higher level of virtual crowdedness. This finding implies that the decision to contribute is influenced by virtual crowdedness differently according to the types of user contribution. That is, the decision may depe nd on the amount of effort required to contribute and ambiguity of the cancels out the bystander effect, which increases along with virtual crowdedness, when the users engage in sharing additional content (feedback contribution via comments and confirmative contribution via thumbs - up). Overall, this research examine s which influence ( the prosocial behavior of or bystander effect on Waze users ) prevail s under differ ent levels of virtual crowdedness. The se findings may help mobile app designers and developers understand how to encourage users to engage in prosocial behavior and be independent of bystander effect. Such 6 practical implications suggest how to virtually vi sualize user density and improve the design of social features for prosocial contribution to the mobile virtual communities. Understanding the role of virtual crowdedness in encouraging user contribution around urban areas promises new opportunities for bi g data and Internet of T hings (IoT) initiatives. This research is in the tradition of citizen - data science and provides practical guidelines for the design of smart cities. The rest of this research is organized as follows. This research first describe s th e research context, Waze and its user community. Then, I introduce the theoretical backgrounds and hypotheses of this research . Next , this research describe s and visualize s the large - scale, spatial data collected from multiple open sources of NYC, followed by the empirical approaches of this research . I report the empirical results and robustness checks. This research then summarize s the findings and discuss es the theoretical and practical implications. RESEARCH CONTEXT Waze App Waze is one of the most pop ular mobile apps on both iOS and Android as a GPS navigation and mapping tool for live , crowdsourced traffic information. Waze is elected as one of the Essentials by App Store and has been a top - ranked app for a long time. Mashable (2016 ) conducted a survey to select the be st mobile apps of all time since 2007 when Steve Jobs unveiled iPhone. Waze is ranked 22nd among the most groundbreaking apps that have a s . Digital Marketing Ramblings ( Smith 2017 ) estimates that more than 65 million users actively use Wa ze every month across 185 countries around the world. As depicted in Figure 1, Waze app provides a virtual space where users generate traffic content in real time by using social features that are simple and easy to use. Waze users 7 Figure 1 . Waze app: GPS navigation, map, and social traffic tool 2 themselves identify points on the map to create initial alert s on accidents, traffic jams, traffic speed, police traps, and road hazards, as well as to react to and verify the initial alerts by adding comments and tapping thumbs - up secondarily. (1) Alert posting is a social feature which allows Waze users to originally announce their traffic experiences, such as changes and emergencies in the traffic status and conditions, to other user s. By posting alerts, Waze users create original traffic content in real time. In addition, (2) commenting and (3) verifying features enable Waze users to generate additional content via comments on and thumbs - up on the posted alerts. By commenting on the initial alerts, Waze users react to the original traffic content and patch it up, thus providing feedback on the original traffic content over time. At the same time, the collective of Waze users verify the original traffic content by putting thumbs - up on the initial alerts. Using such social features, Waze users engage in user contribution in three distinct ways of 2 The screenshots of Waze Version 4.0 are sourced from App Store on October 18, 2015. 8 crowdsourcing the real - time UGC. The socially - crowdsourced user contribution offers real - time and location - specific traffic content. O rdinary c itizens using Waze continuously update UGC by generating a pool of traffic information as a public good. That is, a real - time collective crowd of citizens capture rapidly varying traffic status rather than the traffic information generated from algorithms or a few expert observers ( Howe 2008 ) . Such crowdsourced user contri bution provides fast and intuitive traffic content which is more helpful for drivers in a socialized and localized way than the top - down information from GPS apps as Apple Maps ( Alvarez 2015 ) . Mobile Virtual Community The emerging SMAC innovations generate a new kind of virtual community with the unique characteristics of mobile app users as a group . Such mobile virtual communities are identified not only by virtual characteristics but also mobile features. A mobile virtual community initially consists of small groups of close associates in geographic spaces and is accessible to unknown people who ar e located nearby. However, certain circumstances trigger community starts with a few active agents and then spreads to a large number of other users over time . Thanks to the mobile features, its members communicate with each other not only via formatted texts but also via graphics, animations, videos, sounds, and other ways. Mobile virtual communities are identified with the hybrid characteristics of virtual co mmunities and mobile communications ( Rheingo ld 2000 ; Rheingold 2002 ) , as summarized in Table 1. 9 Table 1 . Hybrid characteristics of mobile virtual communities THEORETICAL BACKGROUND A UGC platform refers to a system that enables users to contribute, consume, and evaluat e digital content ( Levina and Arriaga 2014 ) . A common phenomenon among UGC platforms is a relatively small proportion of users who actively contribute to content production ( Kane 2011 ; von Krogh et al. 2003 ) . Mo reover, UGC platforms do not survive without active users who are willing to invest their time and effort to generate useful and interesting content for other users ( Levina and Arriaga 2014 ) and contribution is important in providing specific and unique implications for the design decisions on Waze: how to encour age Waze users to engage in contribution to the useful yet free UGC for other users. Such implications may help the mobile app remain successful by encouraging users to actively engage in social production, refinement, evaluation, and consumption of the mo bile UGC on urban traffic. 10 User Contribution The wisdom of crowds has been explored by IS research in a variety of UGC platforms ( Levina and Arriaga 2014 ) . A previous finding is that reputation is a key motivator in explaining why users contribute their time and effort to virtual communities ( Butler 2001 ; Faraj et al. 2011 ; Faraj and Johnson 2011 ; Faraj et al. 2008 ; Moon and Sproull 2008 ; Wasko and Faraj 2005 ) . Members of a virtual community desire to help others by engaging in prosocial behavior with their content as a public g ood. Such prosocial contributors achieve distinction , social stratification, by generating helpful content for other users. A great variation has also been found in terms of the nature of contribution ( Levina and Arriaga 2014 ) . In the UGC platforms, users differentiate their contribution with roles based on what types of contribution they produce to other users ( Butler et al. 2002 ; Kraut et al. 2011 ; Preece and Shneiderman 2009 ; Smith and Kollock 1999 ; von Krogh et al. 2003 ; Welser et al. 2007 ) . In the related literatur e, the roles of users have been understood with their levels of participation in virtual communities ( Oestreicher - Singer and Zalma nson 2013 ) . Wenger (1998 ) posits that user par ticipation evolves increasingly over time from peripheral users (who do not participate in the community) to boundary users (who span boundaries and link communities) with inbound users (who start participating in the community) and insider users (who full y participate in the community) along t he way. Similarly, Kim (2000 ) in online community: from visitors through novices and regulars to leaders. More recently, Li and Bernoff (2011 ) differentiate levels of user participation in the Web 2.0 social community between spectators ( content consumption), collectors (content organization), critics (community involvement), and creators (continent creation). In a similar vein, Preece and Shneiderman (2009 ) framed the stages of community participation (reader contributor collaborator leader) to explain the different 11 levels of user participation in virtual community by characterizing the roles of passive and active users. Based on the levels of user participation, this research identif ies specific roles of users, who use social features to engage in a mobile virtual community, to conceptualize user contribution in the context of Waze app. Three types of user contribution include (1) initial contribution to original traffic content as well as (2) feedback contribution and (3) confirmative contributi on to additional traffic content. As described in Figure 2, social features of Waze enable users to stratify their contribution to the mobile virtual community in three different ways. As creators , Waze users initially post alerts to generate the real - time content and share their traffic experience with other users. Such content creators are active participants who initially contribute to the community of Waze users: the role of posters (who attract the attention of other users) rather than one of lurkers ( who pay attention to content produced by other users) in general UGC platforms. 3 At the same time, Waze users are content critics and assessors who passively participate in the content crowdsourcing by placing comments and thumbs - up on the initial alerts. These show feedback contribution (refining the initial alerts) and confirmative Figure 2 . Role - based conceptualization of user contribution 3 While posters refer to those users who actively post their experience on a UGC platform, lurkers refer to those ( Ridings et al. 2006 ; Schlosser 2005 ) . 12 contribution (verifying the alerts) , respectively. The critics and assessors perform b oth roles of posters and lurkers because they not only consume the initial UGC produced by other users , but also generate the secondary UGC by patching up and verifying the posted alerts. Thanks to the three types of user contribution, Waze users as conten t spectators are updated of the real - time traffic nearby by utilizing the initial and secondary UGC. Virtual Crowdedness When driving with Waze app, users observe potential contributors outside of their automobiles physically as well as other Waze users o n smartphone screens virtually . Such a ( the density of other users as potential contributors whom a focal user observes on the navigation map ) : virtual crowdedness refers to the number of Waze users per km 2 . The virtual and spatial factor ( other users ) influences the decision on whether to contribute via the social features. Virtual crowdedness is spatially derived yet virtually distinct from physical crowdedness ( Stokols 1972 ) . The degree of population density per unit area was examined by Andrews et al. (2016 ) in terms of its effects on consumer responses in the context of hyper - contextual mobile advertising. This research focuses on the virtual density of users to understand its role in encouraging their cont ribution. Waze users observe the virtual density of other users nearby thanks to mobile technology. Virtual crowdedness influences their contribution decision virtually and spatially. In the context of Waze, physical crowdedness is observed in the reality of actual traffic whereas virtual crowdedness is observed on the Waze map, i.e., on a smartphone screen. That is, while physical crowdedness usually indicates traffic congestion in urban areas, virtual crowdedness signals the degree of potential contributo 13 Development of Hypotheses Prosocial Behavior. Waze users of the mobile virtual community benefit from free contribution crowdsourced by themselves. In a UGC platform, user contribution is explained by pro social behavior . Contributors are motivated to perform prosocial behavior by social effects, such as reciprocity among peers in a user community ( Andreoni 1988 ) , the expected amount of potential contribution by others in a social reference group ( Andreoni and Scholz 1998 ) , prestige ( Harbaug h 1998 ) , the impact of contribution on social welfare ( Duncan 2004 ) , the number of beneficiaries ( Andreoni 2007 ) , and social images which the c ontributors project ( Andreoni and Bernheim 2009 ; Ellingsen and Johannesson 2008 ) . Such prosocial beh avior increases when members of a virtual community are in geographic proxim ity to each other and share common interests ( Blanchard and Horan 1998 ) . Waze users post al erts and generate initial traffic content, which is useful for other users nearby. Like users of virtual communities in general UGC platforms, Waze users also experience the social effects motivating prosocial behavior. In addition, Waze users share a comm on interest in sharing the real - time UGC on the traffic status and changes with other users nearby to improve their driving. C ommon interest s and close proximities increase their willingness to engage in initial contribution by posting initial alerts on th eir experience. These factors augment the prosocial behavior of Waze users, thus encouraging their initial contribution. This suggests that the more Waze users the more initial alerts ; compared to sparse users, crowded users are expected to engage in a gre at amount of prosocial behavior by posting initial alerts more. Thus, this research propose s a positive relationship of virtual crowdedness and initial contribution, as follows: Hypothesis 1A. As virtual crowdedness increases, initial contribution increas es. 14 Bystander Effect. Besides pure altruism, impure altruism also exists in the mobile virtual community and affects user contribution to the real - time traffic UGC, a public good ( Andreoni 1989 ; Andreoni 1990 ; Cornes and Sandler 1984 ; Steinberg 1987 ) . While pure altruism encourages prosocial behavior, impure altruism causes free - riding behavior in user communities. The free - riding behavior is explained by bystander effect : when more observers exist around an emergent situation, each individual is l ess likely to help the victim and more likely to be a passive bystander ( Darley and Latané 1968 ; Latané and Darley 1968 ; Latané and Darley 1970 ; Latané and Nida 1981 ) . 4 In addition to prosocial behavior incentives, a mobile virtual community could also e xhibit bystander effect if users can observe the existence of other users virtually. When a Waze user observes a number of other users as alternative contributors, her/his relative importance of prosocial behavior fades away whereas her/his own private ben efit crops up as a main factor in terms of the decision on user contribution. Exposed to a high level of virtual crowdedness nearby, Waze users tend to be passive bystanders rather than active contributors ( Andreoni 2006 ; Ribar and Wilhelm 2002 ) . Given the competing effects of prosocial behavior and bystander effect, the magnitude of the relationship between virtual crowdedness and initial contribution degreases as virtual crowdedness increases. Therefore , this research propose s a concave - down relationship of virtual crowdedness and initial contribution, as follows: Hypothesis 1B. As vir tual crowdedness increases, the rise of initial contribution decreases. 4 The bystander effect is caused by three driver s: diffusion of responsibility, evaluation apprehension, and pluralistic ignorance ( Latané and Darley 1970 ) . Diffusion of responsibili ty refers to the tendency to subjectively divide the personal responsibility for helping others by the number of bystanders. Evaluation apprehension is the fear of being judged by others when acting publicly due to the risks of making mistakes or acting in adequately. Pluralistic 15 Time - of - Day Effect. Based on the concept of task - technology fit, users perceive the effectiveness of technology for given tasks when individual abilities, technology functionalities, and task requirements fit each other ( Goodhue 1995 ) . Waze provides a location - specific mobile and soc ial platform in which users share their daily driving experience with other users (prosocial behavior). At the same time, Waze users also benefit from the initial UGC contributed by other users (free - riding behavior). Active users are able to generate orig inal content (individual ability) by using the alert posting feature (technology function) for initial contribution thanks to the posted alerts from active users (technology function) to improve their driving (task requirement). Urban drivers face more dynamic changes in traffic status (e.g., traffic jams and vehicle collisions) for rush hours 5 compared to non - rush hours. The time - of - day effect makes active u sers perceive that the alert posting feature is effective for their task because they are able to generate more initial alerts due to more traffic events for rush hours than for non - rush hours. The increased task - ten denc ies of prosocial behavior for rush hours. At the same time, passive users utilize the initial alerts because they are need ed to deal with more traffic events for rush hours than non - rush hours. Thus, they keep acting as bystanders, who only consume oth contribution. To sum up, serious and frequent traffic problems for rush hours induce active contributors to perceive a cognitive match between the alert posting features and the experience of sh aring traffic information with others. My explanation relies on the relatively higher perceived usefulness of Waze in rush hours compared to non - rush hours. An alternative 5 Rush hour may have multiple definitions according to different contexts. In this research, rush hour refers to the peak p eriods of the traffic volume of a day due to heavy commute around urban areas. 16 explanation is that during rush hours, the average speed of traffic is lower, so tha t the drivers have time to pay attention to their smartphone screens. However, the speed of traffic is control led for in the area, so this should not be a major bias in estimation. On the other hand, such rush - hour traffic dynamics keep bystanders benefiti rather than engaging in their own contribution. During rush hours, therefore, the both sides, contributors and bystanders, intensify the tension between prosocial behavior and bystander effect. T he time - of - day effe - technology fit for rush hours rather than non - rush hours in urban areas. The increase in the perceived task - technology fit intensifies the tension between prosocial behavior of and bystander effect on Waze users. That is, active users engage more in prosocial behavior due to their increased task - technology fit in contributing by using the alert posting feature for rush hours than for non - rush hours. On the other hand, passive users are seriou sly influenced by the bystander effect for rush hours because they perceive the task - technology fit in performing free - riding - of - day effect on the intensity of the tension, rush hours in urban areas lead to a steeper rise and more sluggish rise of initial contribution for the low and high levels of virtual crowdedness, respectively. This research therefore propose s that the relationship of virtual crowdedness and initial contribut ion become more concave - down for rush hours than for non - rush hours, as follows: Hypothesis 2. The time - of - day effect moderates the rise of initial contribution along with virtual crowdedness. A s virtual crowdedness increases at its low levels, the rise o f initial contribution increases further for rush hours than for non - rush hours; as virtual crowdedness increases at its 17 higher levels, the rise of initial contribution decreases further for rush hours than for non - rush hours. Reduced Bystander Effect. Th e bystander effect is reduced when an emergency is recognized unambiguous and severe ( Fischer et al. 2006 ) . Bystanders more clearly perceive the increased cost of not helping the potential victim (s) if they face situations which are less ambiguous and more severe. Such situations encourage bystanders to help others even when a number of alternative contributors exist nearby. This is because they are afraid to be blamed if they do not intervene in unambiguous and severe situations. They tend to perceive a higher level of danger to others (potential victims) and expec t more potential reward thanks to their interventions as the situation is less ambiguous and more dangerous. In this way, the cost - reward approach suggests that the bystander effect is reduce d when an emergency is not ambiguous yet potentially serious ( Fischer et al. 2011 ) . In the context of Waze users, feedback contribution via comments provides additional information by amplifying and patching up the initial traffic content . Waze users are more likely to understand the related traffic event clearly when adding comments on initial alerts than when posting alerts initially. This is because Waze users comment on the posted alerts a while later than when the traffic events of in terest occur red . Also, they already receive original information on the events from the initial alerts. Given such a time gap and extra information from initial alerts, Waze users keep engaging in feedback contribution, even when they observe alternative c ontributors because they recognize the traffic events less ambiguous. This suggests that a Waze user is more likely to perceive the elevated clarity of a severe emergency, which reduces the bystander effect, when they comment on a posted alert secondarily than when they post an alert 18 initially. That is, Waze users become independent of bystander effect when they engage in feedback contribution due to the timing of their comments and the extra information from initial contribution. This research therefore pr opose s a positive relationship of virtual crowdedness and feedback contribution, as follows: Hypothesis 3A. As virtual crowdedness increases, feedback contribution increases. Waze users rate and verify the posted alerts by tapping thumbs - up. Th is assessi ng feature also generates additional content, which signals the collective confirmability of initial contribution. A number of thumbs - up on an initial alert indicates that its traffic event is serious for other users and influential to their driving. That is, Waze users engage in confirmative contribution when they predict a high level of danger to others due to related traffic events as well as perceive a high level of responsibility for informing others of the initial alert. Such perceptions increase the expected cost of not contributing and increase the potential reward of contributing via thumbs - up, thus reducing the bystander effect on Waze users. Thus, Waze users are more likely be independent of bystander effect when engaging in confirmative contribut ion via thumbs - up than when engaging in initial contribution. Moreover, confirmative contribution can be easily done by tapping on the thumb - up button without any texting and multiple - time tapping, thus requiring less effort than the other types of user co ntribution. The reduced effort to contribute also encourages users to be independent of bystander effect when engaging in confirmative contribution via thumbs - up. These suggest that Waze users are willing to perform prosocial behavior for confirmative cont ribution even when they observe a high level of virtual crowdedness indicating many alternative contributors. T his research propose s a positive 19 relationship of virtual crowdedness and confirmative contribution, as follows: Hypothesis 3B. As virtual crowde dness increases, confirmative contribution increases. DATA Data Collection Approach This research analyze s a panel dataset for NYC, one of the most populous cities in the U.S. The large - scale dataset was gathered for 98 days in 2016 (two waves of data: 4 9 days from January 11 to February 28 in winter and the other 49 days from July 11 to August 28 in summer) from multiple open sources of urban data: mobile app APIs, U.S. Census Bureau, NYC Police Department (NYPD), DemographicsNow, and Weather Underground . A driver - level panel dataset is not available because the data do not provide a unique identifier for each driver. Instead, a zone - level panel dataset was collected due to the real - time dynamics of moving users and their activities in each zone. By e xplo iting the data collection capability, I captured Waze data hourly to take a snapshot per hour for the fourteen weeks. The dataset was aggregated into 2,352 - time points (24 hours × 98 days) over the urban areas. During the time windows for winter and summer in 2016, I tracked 5,005,924 pieces of user contribution (2,112,547 alert postings as well as 827,125 comments and 2,066,252 thumbs - up on the posted alerts) and 5,280,499 users as well as their locations as points (coordinates) on the map. Given the large number of zones in NYC, the raw point data were aggregated into the level of N eighborhood T abulation A rea ( NTA) by using QGIS software and its Python Console with their coordinates and NTA shapefiles on the map across the five NYC boroughs : Bronx (38 NTAs ), Brooklyn (51 NTAs), Manhattan (29 NTAs), Queens (58 NTAs), and Staten Island (19 20 NTAs). I Central Park in Manhattan). Hence, the final dataset includes 190 NTAs over t he five NYC boroughs. As a result, this research analyze s a set of balanced panel data with 446,880 observations (190 NTAs × 2,352 - time points). Figure 3 spatially visualizes the densities of total contribution and Waze users per km 2 in NYC, while Figure 4 visualizes the average levels of total user contribution and Waze users per km 2 over 24 hours of a day. The visuals identify two peaks of total user contribution and Waze users in a day as rush hours : rush hours in the morning (7:00 to 9:59) and rush hour s in the afternoon (16:00 to 18:59) on weekdays. T his definition is further utilized to employ a binary variable for testing Hypothesis 2. Measurement of Study Variables and Their Statistics User Contribution. This research examin es the three types of use r contribution. In particular, , , and , measure the number of alert postings, as well as the numbers of comments and thumbs - up on the alerts in NTA i at Time t . Based on that, the dependent variables are generated , including initial contribution measured by ln ( ) (i.e. , ln ( )) , feedback contribution measured by a ratio of to (i.e. , ), and confirmative contribution measured by a ratio of to (i.e. , ). User Community Activity. The independent variable of interest is ln ( ) , i.e., the log ged number of Waze users per km 2 in NTA i at Time t . This variable indicates a density of nearby users as p otential contributors whom a focal user can 21 Figure 3 . NTA - by - NTA densities of total user contribution and Waze users 6 6 The map backdrop cartography is sourced from stamen.com using OpenStreetMap data. 22 Figure 4 . Hour - by - hour average density levels of total user contribution and Wa ze users 23 observe on her/his smartphone screen virtually when driving and using Waze . 7 Since the driving speed may influence the decision on whether to use the Waze features to contribute or not, is also measured as a time - variant contro l variable. Physical Crowdedness Measures. Waze users are exposed not only to virtual crowdedness but also to physical crowdedness when us ing the social features to contribute. Physical crowdedness not only indicates traffic congestion but also reflects th e density of potential contributors whom a focal user can observe outside of her/his automobile physically. To control for the effects of such physical densities, physical crowdedness is measured by , , and as time - invariant control variables. Traffic Condition. Drivers use Waze app to generate and consume traffic information. When Waze users use the app for driving, they are exposed to traffic condition s aroun d the urban areas. Therefore, traffic condition s influence the decision to engage in user contribution. To measure traffic condition, this research employ s , , and as time - variant control variables . In addition, and are used as time - invariant control variables. Weather Condition. Driving and user contribution are influenced by weather condition s because Waze users drive and contribute not inside but outside. They are more likely to be able to afford to contribute while driving with calm weather than with severe weather. To control for weather effects, this research use s time - variant weather dummies, i.e. , (dropped as the reference category), , , , and . 7 that precisely measures the exact number of users on their smartphone screens. If it is supposed that the true number of users on the smartphone screens is proportional to conditional on the control varia bles, then with logarithm transformation, a constant is additive to ln ( ). 24 Table 2 . Definitions, descriptive statistics, and data sources of study variables Table 2 provides the detailed definitions, descriptive statistics, and data sources for all the study variables. Table A1 (Appendix A) shows the pairwise correlations between all the study variables. ln ( ) is positively correlated with the logarithm of initial contribution , i.e., ln ( ) , and feedback contribution , i.e., the ratio of 25 to , whereas it is negatively correlated with confirmative contribution , i.e., the ratio of to . Table A2 ( see Appendix A) summarizes the distributions of boroughs and weather conditions in NYC. EMPIRICAL APPROACH This research analyze s the large - scale data with both spatial and temporal di mensions. My strategy to analyze a diverse range of large - scale datasets of NYC combines a spatial data approach and econometric analysis. The spatial data analysis enables us to aggregate point data (coordinates) on the map into zone - level information (sp atial characteristics of polygons, i.e., NTAs). In addition, econometric models are utilize d to deal with a large proportion of zero - valued observations and endogeneity issues. Zero - Inflated Count Data Models C ount data models are employ ed to examine the relationship between virtual crowdedness and initial contribution. However, a common feature of UGC platforms is that a small proportion of users actively engage in contribution ( Levina and Arriaga 2014 ) . It is possible that there is no t any initial contribution via alert postings in a certain NTA i at Time t . The dataset has a large proportion of zero - valued observations for init ial contribution, which would lead to large estimation bias if it is not accounted for. Therefore, zero - inflated negative binomial (ZINB) and zero - inflated Poisson (ZIP) models are appl ied to estimate impact of virtual crowdedness on initial contribution. In brief, the zero - inflated models assume that zero outcomes are due to two different processes. In this research context, a certain NTA - time observation may not have any Waze users driving at that time. Therefore , the number of initial contribution via al ert postings is zero: the observation is zero - inflated. By contrast, if there are some Waze users for that 26 observation, then the number of initial contribution covariates follows a count data distribution, which takes nonnegative integer values (see Append ix B for details on the zero - inflated count data models). The zero - inflated count data model is estimated with two regression equations: while the main equation is a count data model for the dependent variable, the zero - inflation equation is a binary outco me model for observing non - zero observation. In the main equation, is include d as the main independent variable that captures the positive relationship ( Hypothesis 1 A ) between the logarithms of virtual crowdedness and initi al contribution. 8 In addition, the estimation include s a rich set of covariates that control for condition, and weather condition, which may have effects on initial co ntribution. It also include s time - variant controls ( ln ( ) , ln ( ) , ln ( ) , ln ( ) , and weather condition indicators) and time - invariant control variables ( ln ( ) , ln ( ) , ln ( ) , , and ln ( ) ). 9 Date and Hour are dummy indicators to control for time effects across the fourteen weeks. T he zero - inflation equation include s all the covariates except 8 In count data models, the conditional mean of dependent variable is an exponential function of covariates: E[ Y ] = exp( XB ). Therefore, given a log - transform ation for independent variable (use of log ( X ) as a regressor), the coefficient should be explained as elasticity : a relative (proportional) change of X will lead to a relative change of Y . Since the logged virtual crowdedness is not included into the ze ro - inflation equation, zero - inflated models have the same parameter interpretation, except that the effect is conditional on the observation not being excess zero. 9 Such traffic - and road condition - related control variables play important roles in the an alysis. For instance, the impact of virtual crowdedness should be stronger during rush hours. However, an alternative explanation is that since traffic conditions are poor (e.g. average speed of traffic is lower) for rush hours, drivers have more time to p ay attention to smartphone screens. This should not be a main factor because this research controls for traffic conditions in the area. Likewise, NTA - specific time - invariant covariates control for the locational nature of urban areas because the traffic fe atures in central business districts should be different from residential districts . For robustness checks, this research utilizes the NTA - fixed effects to fully wipe out NTA - specific unobserved heterogeneity in the estimation. 27 ln ( ) to address the propensity of excess zero values. I n this research , both ZINB and ZIP models are utilize d t o check robustness of the estimation . I further modify the main equation by including squared term of ln ( ) to test the decrease in magnitude of the relationship between virtua l crowdedness and initial contribution ( Hypothesis 1 B ). Notably, this research employ s robust standard errors on both NTA and hourly time point for all the estimations. Therefore, all the results with the robust standard errors ensure that the empirical re sults are free of potential problems regarding serial correlation or heteroscedasticity ( Wooldridge 2012 ) . Heckman Selection Models To test Hypothes e s 3 A and 3 B , this research define s the ratio s of feedback contribution and confirmative contribution to initial contribution as dependent variables. However, as mentioned before , it is possible that there are zero - valued initial contribution counts for a given NTA i at Time t . In this case, the measures of the ratio are not applicable, thus causing a sample selection issue. Clearly, if the sample selection is associated with the characteristics of the NTA at that time, the estimated effect would be biased. T herefore , the Heckman selection model, a two - stage estimator , is employ ed to deal with selection bias ( Heckman 1979 ) in terms of the ratio of feedback contribution or confirmativ e contribution to initial contribution, i.e., ratios of the numbers of comments or thumbs - up to the number of alert postings, conditional on that at least one alert is posted in NTA i at Time t (see Appendix B for details on the specification of Heckman se lection model). In the main equation, the j subscript of the dependent variable is 2 ( = ) or 3 ( = ). Like previous specification, the Heckman selection model include s time - variant control variables ( ln ( ) , 28 ln ( ) , ln ( ) , ln ( ) , and weather condition indicators). Moreover, Date and Hour are also included as dummy indicators to control for time effects. T he selection equation include s time - invariant control variables ( ln ( ) , ln ( ) , ln ( ) , , and ln ( ) ), in addition to time - variant control variables and date and hour - of - the - day dummies. Notably, time - invariant controls are presented in the selection equation but not in the main equatio n so that the exclusion restriction for Heckman selection model is satisfied. To ensure empirical robustness, two estimation methods are adopt ed - step estimation and maximum likelihood estimation with simultaneous equations. Matching Method In the previous section, I conduct regression analysis with control variables to examine the relationship between logarithms of virtual crowdedness and initial contribution. The conclusions will be valid only if ln ( ) is an exogenous variable; the included control variables absorb all the alternative explanations. Those alternative explanations may contain the tion does not hold, the relationship is association rather than causality. Therefore, I attempt to test the purely causal impact by controlling for the multifaceted confounding effects in a quasi - experimental setting. I therefore apply the PSM estimator to mitigate endogeneity due to alternative explanations on the relationship of virtual crowdedness and initial contribution. The PSM estimator has the advantages of dealing with high - dimensional matching using many variables and enables to derive A verage T re atment E ffect (ATE) from observational data ( Abadie and Imbens 2006 ) . In general, the PSM estimator approximates the counterfactual 29 outcome for each observation by using an a verage of the outcomes of similar observations that receive the other treatment level. The similarity between observations is measured based on the estimated treatment probabilities, i.e., propensity scores, which are derived from a binary regression model . Therefore, the ATE refers to the average of the difference between the observed and potential outcomes for each observation. In this research , PSM estimator enables us to match the pairs of observations, which are different from each other regarding virt ual crowdedness but are similar to each other regarding all the control variables. As a result, the estimated ATE should be attributed to the different values of virtual crowdedness. Unlike regression models, PSM estimator does not rely on any functional f orm assumptions regarding how control variables should influence the dependent variable. Figure 5 . Distributions of virtual crowdedness with five groups for propensity - score matching 30 O bservations in NTA i at Time t are assign ed into multiple groups as treatments according to the different levels of virtual crowdedness and test their ATEs on initial contribution. Specifically, I first discretize the continuous treatment variable, i.e., ln ( ) , into five groups by taking five equally - wide intervals along its distribution (after excluding zeros) ( e.g., Aral and Nicolaides 2017 ) . As a result, data observations are classified into five groups ranging from Group 1 , representing very low virtual crowdedness , to Group 5 , representing very high virtual crowdedness (see Figure 5). Using ln ( ) as the outcome variable, I employ the PSM estimator between every pair of adjace nt groups to mitigate endogeneity concerns. RESULTS This research combine s the evidence from regression analysis and matching method to report the findings which support all the hypotheses. Results of testing Hypothes e s 1 A and 1 B show a concave - down, incr easing relationship of virtual crowdedness on initial contribution. Though logged virtual crowdedness has a positive relationship with logged initial contribution, the magnitude of the relationship decreases as virtual crowdedness increases. The concave re lationship is additionally supported by results from PMS estimator in a quasi - experiment setting. Results of testing Hypothesis 2 indicate that the relationship is more concave - down for rush hours than for non - rush hours. Based on results of testing Hypoth es e s 3 A and 3 B , virtual crowdedness has a positive relationship with each of feedback contribution and confirmative contribution. T he complete findings suggest that the intensity of the tension between prosocial behavior of and bystander on Waze users vari es by the degree of virtual crowdedness, the time of day, and the types of user contribution. Table 3 summarizes the main findings and how they correspond to my hypotheses. 31 Table 3 . Summary of hypotheses testing results Note. The details for Equations ( 9 ), ( 11 ), and ( 12 ) are provided in Appendix B . Results of Testing Hypothesis 1 This research first report s the results of ZINB and ZIP models for initial c ontribution. Table 4 show s that the logged virtual crowdedness is positively associated with the logged initial contribution. Zero - inflated count data models consistently show that virtual crowdedness has a significant coefficient on initial contribution: = 0.6266 significant at p < 0.01 for the ZINB model (see Column 1); = 0.4696 at p < 0.01 for the ZIP model (see Column 3). 10 Therefore, ln ( ) has a positive relationship with the logarithm of initial contribution. The results s upport Hypothesis 1 A suggesting that as virtual crowdedness increase s , initial contribution also increases ; the crowded users engage in a greater amount of prosocial behavior by posting initial alerts than the sparse users. In Table 4, zero - inflated count data models also consistently show that ln ( ) has a concave - down, increasing relationship with the logged initial 10 This research also tests t he estimation specifications. Result of a Vuong test favors the ZINB model for this estimation rather than the negative binomial model. Result of a likelihood - ratio test suggests that the ZINB model is superior over the ZIP model for this estimation. I fur ther apply the two tests into the specifications for testing Hypothesis 1B and Hypothesis 2, and all the null hypotheses are rejected. 32 Table 4 . Results of zero - inflated count data models for initial contribution Variables Dependent variable: = (1) ZINB (2) ZI NB (3) ZIP (4) ZIP Main equation Zero inflation Main equation Zero inflation Main equation Zero inflation Main equation Zero inflation { ln ( )} 2 0.0853*** 0.0719*** (0.0024) (0.0023) ln ( ) 0.6266*** 0.9157*** 0.4696*** 0.7580*** (0.0036) (0.0095) (0.0036) (0.0096) Time - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Obs. 444,528 444,528 444,528 444,528 Non - zero obs. 303,469 303,469 303,469 3 03,469 Zero obs. 141,059 141,059 141,059 141,059 Log likelihood Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators ar e omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). ZINB: zero - inflated negative binomial; ZIP: zero - inflated Poisson. * p < 0.10; ** p < 0.05; *** p < 0.01 . 33 contribution: = 0.9157 and significant at p < 0.01 for the ZINB model (see Column 2); = 0.7580 and p < 0.01 for the ZIP model (see Column 4). The results support Hypothesis 1 B by indicating that as virtual crowdedness increases, the rise of initial contribution decreases. In other words, as virtual crowdedness increases, the bystander effect on Waze users may cancel out their tendency of prosocial behavior and decrease the magnitude of the relationship between virtual crowdedness and initial contribution. On average, the logged initial contribution increases along with ln ( ) , but at a much lower increasing rate when ln ( ) is higher. T he left panel of Figure 6 visualize s average values of ln ( ) in each small bin along the range of ln ( ) , fitted with a quadrat ic trend with 95% confidence intervals, as model - free evidence. Consistently, on the right panel, the linear and quadratic trends derived from the ZINB model verify the positive and concave - down relationships clearly. Notice that the model - free plot does n ot control for other covariates , exhibit ing a slightly different shape to the model - based plot. Figure 6 . Relationship of virtual crowdedness and initial contribution 34 Table 5 . Results of propensity - score matching tests with five groups for initial contribution ln ( ) ln ( ) ln ( ) ln ( ) This research adopts a matching estimator t o strengthen the conclusions towards causality. As shown in Table 5, the difference in ln ( ) between Group 2 vs. Group 1 makes a significant positive increase in the logged initial contribution (ATE = 0.5502 significant at p < 0.01), which supports Hypothesis 1 A . The matching estimator shows that the ATE remains significant but decreases as virtual crowdedness increases: ATE = 0.3310 at p < 0.01 for ln ( ) increasing from Group 2 to Group 3; ATE = 0.1829 at p < 0.01 for the increase from Group 3 to Group 4. More importantly, the increase from Group 4 to Group 5 does no t have a significant effect on the logged initial contribution: ATE = 0.1740 , ns . Clearly, the impact of ln ( ) on the logged initial contribution fades away as virtual crowdedness increases: in highly crowded NTAs, additional users do not lead to more initial contribution. This supports the concave - down, increasing relationship between the logarithms of virtual crowdedness and initial contribution, as suggested by Hypothesis 1 B , with a reduced risk of endogeneity problems. The result s of PSM to test Hypothesis 1 B are visualized in Figure 7. T he ATE decreases with a larger Abadie - Imbens robust standard error from the comparison between groups with less virtual crowdedness (Group 2 and Group 1) to groups with more virtual crowdedness (G roup 5 35 Figure 7 . Visualized results of propensity - score matching tests with five groups and Group 4). As virtual crowdedness increases, t he difference between the adjacent groups becomes smaller whereas the robust standard error becomes larger (see the red - dot line in Figure 7 ) . The trends of ATE not only support the concave - down, increasing relationship ( Hypothesis 1 B ) further but also mit igate endogeneity issues on the finding. As a robustness check, the same PSM tests are additionally conducted with ten groups evenly divided along the distribution of ln ( ) . The results consistently support the conclusion: the magnitude of impact decreases as virtual crowdedness increases (see Figure A1 and Table A3 of Appendix C for details). 36 Results of Testing Hypothesis 2 To test Hypothesis 2, the whole data are di vide d into rush - hour and non - rush - hour subsets . Z ero - inflated count data models are appl ied to both subsets in order to compare the relationships across peak hours and idle hours of weekdays . As summarized in Table 6, a c ompari son of Panel A (rush - hour obs ervations) and Panel B (non - rush - hour observations) suggests that the relationship is more concave for rush hours than for non - rush hours. The squared term of ln ( ) has a greater, negative coefficient on the logged initial contribution for rush hours ( significant at p < 0.01 for ZINB model; p < 0.01 for ZIP model) than that for non - rush hours ( p < 0.01 fo r the ZINB model; = p < 0.01 for the ZIP model). The results indicate a time - of - day effect that intensifies the tension between prosocial behavior and bystander effect. During rush hours in urban areas, there is a steeper rise of initial contribution (at low levels of virtual crowdedness) and a correspondingly steeper fall of initial contribution (at high levels of virtual crowdedness). Consistently across the model - free plot and the model - based plot estimated ZINB model, Figure 8 visualiz es that the concave - down relationship between logarithm of virtual crowdedness and initial contribution for rush hours is more evident compared with that for non - rush hours. Results of Testing Hypothesis 3 This research further turn s to the estimation res ults for testing Hypothes e s 3 A and 3 B . The simultaneous and two - step Heckman selection models provide consistent results. In Table 7, ln ( ) has a positive, statistically significant coefficient on feedback contribution ( ): = 0.0670 significant at p < 0.01 for the simultaneous Heckman model (see Column 1); = 0.0598 at p < 0.01 for the two - step Heckman model (see Column 2). These 37 Table 6 . Rush hours vs. non - rush hours: Results of zero - inflated count data models for initial contribution Panel A: Rush - hour observations Variables Dependent variable: = (1) ZINB (2) ZINB (3) ZIP (4) ZIP Main e quation Zero i nflation Main e quation Zero i nflation Main e qua tion Zero i nflation Main e quation Zero i nflation { ln ( )} 2 0.1474*** 0.1042*** (0.0042) (0.0043) ln ( ) 0.5409*** 1.0458*** 0.4070*** 0.8170*** (0.0059) (0.0165) (0.0060) (0.0170) T ime - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Obs. 77,112 77,112 77,112 77,112 Non - zero obs. 69,207 69,207 69,207 69,207 Zero obs. 7,905 7,905 7,905 7,905 Log likelihood 223,010.37 222,392.02 296,032.16 293,993.4 Panel B: Non - rush - hour observations Variables Dependent variable: = (1) ZINB (2) ZINB (3) ZIP (4) ZIP Main e quation Zero i nflation Main e quation Zero i nflation Main e quation Z ero i nflation Main e quation Zero i nflation { ln ( )} 2 0.0570*** 0.0509*** (0.0028) (0.0027) ln ( ) 0.6482*** 0.8403*** 0.5030*** 0.7077*** (0.0045) (0.0113) (0.0044) (0.0115) Time - va riant control variables Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Y es Yes Yes Yes Yes Yes Obs. 367,416 367,416 367,416 367,416 Non - zero obs. 234,262 234,262 234,262 234,262 Zero obs. 133,154 133,154 133,154 133,154 Log likelihood Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables i nclude ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). ZINB: zero - inflated negative binomial; ZIP: zero - inflated Poisson. * p < 0.10; ** p < 0.05; *** p < 0.01 . 38 Figure 8 . Rush hours vs. non - rush hours: Relationship of v irtual crowdedness and initial contribution results consistently support Hypothesis 3 A, suggesting that as virtual crowdedness increases, feedback contribution increases. The results imply that Waze users are willing to perform prosocial behavior for feed back contribution even when they observe many alternative contributors. 39 Table 7 . Results of Heckman selection models for feedback contribution Variables Dependent variable: = (1) Simultaneous Heckman (2) Two - step Heckman Main e quation Selection e quation Main e quation Selection e quation ln ( ) 0.0670*** 0.5775*** 0.0598*** 0.5771*** (0.0015) (0.0047) (0.0019) (0.0047) Time - variant control variables Yes Yes Yes Yes Time - invariant control variables No Yes No Yes Weather condition indictors Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Obs. 444,528 444,528 Censo red obs. 141,059 141,059 Uncensored obs. 303,469 303,469 Log - likelihood 409,308.52 Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). * p < 0.10; ** p < 0.05; *** p < 0.01 . In Table 8, ln ( ) also has a positive, significant coefficient on confirmative contribution ( ): = 0.1107 significant at p < 0.01 for the simultaneous Hec kman model (see Column 1); = 0.0798 at p < 0.01 for the two - step Heckman model (see Column 2). Hence, the positive relationship supports Hypothesis 3 B , suggesting that confirmative contribution increases along with virtual crowdedness. The positive rela tionship implies that Waze users become independent of bystander effect when they engage in confirmative contribution. Robustness Checks The empirical evidence from the zero - inflated count data models supports Hypothes i s 1 and Hypothes i s 2. However, zero - inflated count data models do not allow us to include NTA - fixed effects because of the incidental parameters problem . 11 If the NTA - fixed effects are 11 There is no way to get rid of fixed effects using the cancelling - out approach. 40 Table 8 . Results of Heckman selection models for confirmative contribution Variabl es Dependent variable: = (1) Simultaneous Heckman (2) Two - step Heckman Main e quation Selection e quation Main e quation Selection e quation ln ( ) 0.1107*** 0.5771*** 0.0798*** 0.5771*** (0.0055) (0.0047) (0.0089) (0.0047) Time - variant control variables Yes Yes Yes Yes Time - invariant control variables No Yes No Yes Weather condition indicators Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Obs. 444,528 444,528 Censored obs. 141,059 141,059 Uncensored obs. 303,469 303,469 Log - likelihood 874,142.75 Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). * p < 0.10; ** p < 0.05; *** p < 0.01 . include d , all the NTA - specific effects are fully controlled for and the estimation results will be free of time - invariant confounding factors. To confirm the evidence further, I test the hypotheses with st andard count data models without zero inflation (pooled and fixed - and random - effects negative binomial as well as Poisson models; see Tables A4 and A5 of Appendix D for details). Their estimated results are also consistent with the main results of the ZIN B and ZIP models, although the Vuong test supports zero - inflated models. T hese alternative models cannot handle zero - inflation issue and are not preferred. Furthermore, I test whether the results vary across different seasons. I expect that the relationshi p of virtual crowdedness and initial contribution is more concave - down in winter than in summer. This is because Waze users can afford to contribute when driving with a long period of daylight (even with rainy weather) for summer rather than with a short p eriod of daylight (besides snowy and icy roads) for winter. Specifically, I estimate the same zero - inflated count data models with two subsets of all the data, i.e., winter and summer observations. The results 41 are qualitatively consistent with the main res ults while having different effect sizes (see Table A6 of Appendix D for details). This confirms the seasonal difference: the effect is stronger in summer compared to winter, meaning that Waze users are more independent of bystander effect when contributin g for summer than for winter. T o check out the seasonal difference in the results of testing Hypothesis 3, likewise, I estimate the same simultaneous and two - step Heckman selection models with the two sub - datasets of winter and summer observations, respect ively. The results also support Hypothesis 3 (see Table A7 of Appendix D for details). In addition, I expect that Waze users are more proactive to contribute additional information in summer than in winter due to the differences in daylight and weather con ditions. That is, the positive relationships of virtual crowdedness with additional contribution might be greater in summer than in winter. These results indicate the seasonal difference from the results for feedback contribution: a stronger impact of virt ual crowdedness is observe d in summer than in winter . H owever , there is no seasonal difference from the results for confirmative contribution. This implies that Waze users in NYC are less influenced by seasonal effect when engaging in confirmative contribu tion. A possible explanation may be the regional characteristics of the urban areas. Unlike rural residents, the citizens need to consume real - time traffic information for their commutes across the different seasons. Their seasonal consumption of socially - crowdsourced information results in only confirmative contribution, which requires less effort than initial contribution and feedback contribution. DISCUSSION It is important to understand what encourages or discourages mobile app users to engage in mobil e virtual communities by contributing to UGC as a public good even without expecting any 42 monetary rewards. The overall findings suggest that virtual crowdedness plays a pivotal role in encouraging Waze users to contribute to their mobile virtual community in the urban areas. Effects of the virtual and spatial factor on social engagement vary by its degree and the types of user contribution. Implications for Research The se findings may fill the gap in the literature on social mobile technology: the limited and intensive focus on mobile commerce or a few limited apps in prior studies. Most recent studies have mainly focused on mobile app design and advertising characteristics or aspects of consumer behavior in the mobile - commerce app context (e.g., Ghose and Han 2014 ; Han et al. 2016 ; Kwon et al. 2016 ) . Moreover, co mobile promotions has been exhaustively explained by contextual factors (e.g., Ghose and Han 2011 ; Luo et al. 2013 ; Molitor et al. 2012 ; Molitor et al. 2015 ) and product characteristics ( e.g., Bart et al. 2014 ) . A recent paper ( Andrews et al. 2016 ) suggests a counterintuitive finding that physical crowdedness, i.e., the degree of population density or number of people per unit area, has a positive effect on consumer responses to hyper - contextual mobile advertising. The findings extend the current understanding of mobile app usage to a public good. T hat is, the socially - crowdsourced traffic information depends not only on spatial factors alone but also on virtual factors . The findings lead to the following theoretical discussions. F irst, a s hypothesized, virtual crowdedness has a positive and concave - down relationship with initial contribution. The concave - down , increasing relationships suggest that the reciprocity among Waze users varies by the degree of virtual crowdedness. The positive relationship depicts that prosocial behavior via user contribution increases as virtual crowdedness becomes denser. However, the additional user 43 density decreases user contribution beyond a certain threshold of virtual crowdedness. A s more - riders increases ( Olson 1965 ) . The findings are theoretically explained by the tension between prosocial behavior and bystander effect, which together influence the dec ision on user contribution. Second, the relationship of virtual crowdedness and initial contribution is more concave - down for rush hours than for non - rush hours. This finding suggests that the bystander effect, which causes free - riding behavior of Waze use rs, becomes much more dominant than their prosocial behavior is for rush hours when the UGC is critically valuable for their driving due to the increased task - technology fit . Serious traffic congestion and frequent automobile accidents in urban areas ampli - time crowdsourced traffic information to improve commutes during rush hours, thus intensifying the tension between prosocial behavior and free - riding behavior of Waze users. The stronger concave - down relationship for rush hours is c onsistent with a key feature of virtual civic engagement : users are likely to engage in information sharing with a virtual community when they share specific common interests with others nearby ( Blanchard and Horan 1998 ) . In this instance , Waze users tend to share a common interest of updating traffic information locally during rush hours rather than non - rush hours. Finally, the relationship of virtual crowdedness and u ser contribution varies by whether users engage in initial information sharing via alert postings or additional information sharing via comments and thumbs - up. The i mpact of virtual crowdedness varies by the types of user contribution. This may be because the bystander effect is reduced when emergencies are recognized as unambiguous and severe. Waze users perceive the clarity of a severe emergency when commenting on and verifying a posted alert secondarily rather than when posting an alert initially. 44 Impli cations for Practice The se findings provide mobile app designers and developers with practical insights on how to encourage app users to engage in user contribution to UGC further, as follows. First, the concave - down, increasing relationship of virtual cro wdedness and initial contribution may be understood as the tension between prosocial behavior and bystander effect as a social psychological phenomenon in my research context. That is, Waze users tend to perform civic engagement and help other users by sha ring traffic information. The willingness to help others reduces as the number of bystanders increases. A Waze user tends to feel less personal responsibility and is inclined to be a free - rider, who consumes information from other users without her/his own contribution, when she/he notices many alternative and potential contributors. This implies that display ing encourage or discourage their prosocial behavior. Based on the results, Waze app needs to show fewer users on the navigation map than the actual number when the users are crowded. A high density of users on the navigation map signal s that a Waze user has many alternative contributors. More users on the map make the focal user feel less responsible for con tributing and being more influenced by the bystander effect. In this sense, the increase on prosocial behavior of Waze users may be expected by reducing the number of visible users when they are crowded on the navigation map: e.g., for rush hours in urban areas. Second, the relationship between initial contribution and virtual crowdedness plateaus after a threshold. This rise in the rate of initial contribution with virtual crowdedness is steeper in the case of rush hours. A key challenge for a successful U GC platform is to encourage mobile app users to share their common interests with each other. Locally shared interests among nearby Waze users could maximize the virtual civic engagement to satisfy their information needs in terms of the main task, i.e., n avigating to improve driving quality around the urban areas, which 45 they complete together using the mobile platform and its social features. The se findings suggest that Waze app needs to provide additional social features which enable users to query each o ther and provide answer s in terms of traffic. Based on the results, such interactive question - and - answer features might encourage Waze users to share common interests by interacting with each other and focusing on specific traffic issues locally. Third, ad ditional contribution require s less effort to generate information on less ambiguous and more serious events than initial contribution. This condition for reducing the bystander effect explains the findings of the positive relationships of virtual crowdedn ess with both feedback contribution and confirmative contribution. That is, Waze users feel more compelled to contribut e when they engage in additional contribution rather than initial contribution. In other words, the bystander effect for initial contribu tion could be lowered if Waze users are informed of existing alerts on serious traffic issues nearby. For example, Waze app might allow users to be updated of the existing alerts on severe traffic conditions farther away from their locations than it allows currently. Such longer - distance announcements of serious traffic status es may encourage Waze users to engage in initial contribution further by making them feel more responsible and letting them know the nearby traffic conditions better from a wider view. Also, Waze users contribute to additional information more easily than to immunizing them against the bystander effect , even when they observe many alternative co ntributors. This may in turn enable Waze users to engage in further prosocial behavior. In this sense, I suggest that Waze provide speech recognition features, which enables users to post initial alerts with voice. Such a voice recognition feature might re contribute and help the users focus on driving. In addition, I recommend Waze app to provide 46 not only thumbs - up but also thumbs - down to let Waze users assess the initial alerts negatively. Such negative assessments would inform W aze users of traffic status es further by verifying initial contribution . Based on the findings, the added confirmative contribution will relieve Waze be independent of bystander effect and willing to contribute. Finally, smartphone users are citizens who actively use a variety of mobile apps around urban areas and interact with each other as mobile virtual communities. Citizen - data scientists are expected to generate predictive and prescriptive models which bridge the gap between the mainstream of business users and the advanced analytics techniques of data scientists ( Schlegel et al. 2016 ) . In this sense, the findings were supported by empirical analyses with citizen data (large - scale locational data on Waze users and their social engagem ent as well as traffic, demographic, and weather data around NYC). Limitation and Future Direction A limitation of this research may be that the analyses are exposed to potential endogeneity of virtual crowdedness. To deal with the concern, this research first hypothesize s the impact of virtual crowdedness on user contribution by drawing on theory supported by prior empirical studies. Second, this research test s all the hypotheses controlling for a variety of confounding factors including the community act ivities of Waze users, multiple physical crowdedness measures, and traffic and weather conditions, as well as space - and time - related dummy indicators. Third, the findings were supported consistently not only by the multiple rigorous estimators but also by the robustness checks for the seasonal effects. Finally, the additional PSM analyses, a quasi - experimental setting via the matching method, supported the main conclusion for Hypothesis 1. S uch theoretical and empirical efforts may be enough to mitigate th e 47 endogeneity - related concerns on the findings. I call for future research to explain the complex dynamics of UGC in the context of mobile apps. Prior studies mainly focused on understanding the motivators of user contribution ( Wasko and Faraj 2005 ) . Thus, follow - up research needs to study the consequences of user contribution ( Levina and Arriaga 2014 ) . In the context of Waze, for example, future studies may investigate the effectiveness of user contribution in relieving serious congestion and solving traffic problems around urba n areas as well as reducing the related social and economic costs. Such studies might address the social and economic value of socially - crowdsourced collaboration among mobile app users. CONCLUSION A s an exemplar of citizen - data science , this research may guide citizen - data scientists for their data analytics opportunities with big data technologies and even for smart cities driven by IoT initiatives. This research describe s how users engage in a mobile virtual community regarding the relationship of virtu al crowdedness and user contribution in the specific context of Waze. For empirical evidence, this research combines the spatial data approach and advanced econometric analyses to study a diverse range of large - scale, locational, and real - time citizen data about NYC. The findings were guided by theory and supported by empirical evidence, thereby resulting in theoretical contribution to IS research and practical implications for mobile business. The se findings advance the current understanding of mobile app usage by addressing the impact of virtual crowdedness, a unique (virtual and spatial) factor emerged thanks to mobile technology and its social features, on user contribution in the socially - crowdsourced UGC context. 48 ESSAY 2. DOES USER ENGAGEMENT ENHANCE SOCIAL WELFARE? EFFECTIVENESS OF USER - CROWDSOURCED CONTENT IN IMPROVING URBAN TRANSPORTATION ABSTRACT OF ESSAY 2 Waze app platform, a mobile social media app, attracts user engagement and provides joyful opportunities for users to connect with others whi le driving. E arlier research demonstrates that users participate in forums hosted by social media apps because they tend to tap into their prosocial aspirations. In a mobile virtual community of Waze users, user engagement generates the forms of primary in formation, follow - up feedback, and collective confirmation and encourages users to share social sentiments with others. The growth in adoption of Waze and usage of its features might occur due to such social characteristics of user engagement. However, it is not clear whether user contribution, i.e., the outcomes of user engagement, leads to social - welfare consequences. This research examines whether and how the contribution to user - crowdsourced content generates desirable effects on social welfare by relie ving traffic jams in congested urban areas . Utilizing large - scale data on behaviors and locations of Waze users in New York City, this research applies the spatial panel data approach to estimating the social - welfare value of user - crowdsourced content in t he mobile virtual community. The results provide empirical evidence on welfare: user contribution by Waze users reduces the duration of traffic congestion around the urban areas. A welfare analysis estimates that user contribution saves 2.28% of annual tot al congestion cost per driver, $2,533 , in New York City. This social - welfare value is achieved by user contribution that is effective in saving social and economic costs , as well as improving the mobility of urban transportation. This research discusses im plications for practice of mobile app design and policy on urban transportation. 49 INTRODUCTION In the past decade, mobile devices (smartphones and tablets) have become the appliance s of daily life. They host social media app platforms and facilitate ways of interacting with others in a variety of contexts: e.g., collaborative virtual projects, micro - blogs, content communities, social networking platforms, and virtual worlds ( Kaplan 2012 ) . The se mobile apps enable users to engage in communities according to their own interests by creating and exchanging information, i.e., user - crowdsourced content 12 relevant to specific locations at certain times. Users engage with social media apps for various reasons. Many mobile apps contain features that facilitate playfulness and cognitive absorption ( Agarwal and Karahanna 2000 ; Agarwal e t al. 1997 ) . User engagement might be motivated by such hedonic reasons. On the other hand, social media apps also provide opportunities for connections as users exchange information and share sentiments with each other. The extant research has demonstra ted that such aspects of hedonic and social connections have given rise to the rapid proliferation of social media apps in a wide variety of contexts (e.g., Hel ander et al. 2014 ; Hsiao et al. 2016 ; Kim et al. 2013 ; Wei and Lu 2014 ) . Waze is a unique mobile a pp platform that connects drivers with each other and hosts real - time forums for them to share information on traffic, expertise about navigation, and sentiments on helpfulness of posted information. This mobile app also facilitates active forums for socia l engagement, particularly as drivers not only crowdsource traffic information but also relieve the monotony of lonely driving. Waze app features enable drivers to gain real - time information from the wisdom of crowds and have fun by engaging with nearby st rangers who are also drivers. In such processes of social engagement, users also generate the crowdsourced content that could be 12 Such user - generated content is generic for the wisdom of crowds , e.g., Wikipedia. 50 helpful in navigating around traffic congestion or accidents. By using mobile app features, the motivated Waze users contribute to their user community by generating real - time traffic content in three ways : (1) primary information by posting alerts as well as additional information , (2) follow - up feedback by commenting on and (3) collective confirmation by verifying the initial al erts , pieces of primary information . I investigate effectiveness of user contribution in enhancing the value of social welfare for citizens around urban regions. In this research, the social - welfare value specifically refers to the reduced duration of traf fic congestion, indicating better traffic conditions with improved aggregate mobility of urban drivers. On one hand, individual users might learn information and feel a sense of relief from boredom. At the same time, their collective contribution could hel p some drivers avoid congested routes and receive early warnings about accidents or traffic delays. Such benefits could help in improving the mobility of urban transportation and creating gains in public welfare. This research examines whether and how user contribution does indeed create the social - welfare consequences. The specific research questions include : (1) does user contribution reduce the duration of traffic jams around urban areas? ; (2) how do different types of user contribution (primary informat ion, follow - up feedback, and collective confirmation) interact with each other in reducing traffic congestion duration? ; (3) are the effects of user contribution on the duration of traffic jams different during the peak hours of a day ( rush and business ho urs ) when traffic congestion is high relative to the other times when the congestion might not be as high ? Theories of user adoption of social media apps and public welfare are used to develop the conceptual foundations for this research. Spatial panel est imations are used to address the research questions. The remainder of this research is organized as follows. The research context 51 is described in the next section, followed by theoretical background of this research . Then, a large - scale dataset and describ e its spatial and panel characteristics are described and visualized. An empirical approach is also explained with spatial panel data models and their specifications and extensions . Next, the empirical results and findings are summarized to understand the effectiveness of user contribution in relieving urban traffic congestion. In addition, the social - welfare value of user contribution is estimated in term of annual opportunity cost per driver due to the serious traffic congestion in New York City ( NYC ) . Fi nally, this research discusses theoretical and empirical contributions , as well as interprets the findings to suggest implications for practice of mobile app design and policy on urban transportation. RESEARCH CONTEXT The research scope focuses on NYC, on e of the most traffic congested cit ies in the world , over its five boroughs Brooklyn, Bronx, Manhattan, Queens, and Staten Island to analyze the effects of user contribution from Waze users on the severity of traffic jams. This research scope might be appr opriate to study the value of user contribution in terms of u rban t ransportation because commuter s and drivers experience serious traffic congestion in NYC , and many of them are interested in adopting and using Waze to create a heightened awareness of traf fic events . New York City and Its Traffic Congestion Such great cities as Paris, London, and NYC have been attractive places to live thanks to their variety of facilities, social events, and choices of lifestyle s with intense atmosphere s . On the other han d, residents living around the urban areas like Manhattan ( the most densely populated borough of N YC ) are always exposed to large numbers of people, a high population density, and heterogeneity of population. They experience such urban circumstance s as ove rloads of roles, 52 norms, cognitive functions, and facilities ( Milgram 1970 ) . In terms of the urban traffic, life in NYC is even worse. According to the annual INRIX Global Traffic Scorecard ( Cookson 2018 ) , urban drivers were stuck in heavy traffic for 91 congestion - peak hours on average in NYC, 2017. Moreover, p rivate car commuters spent 13 % of their time sitting in id le automobiles due to NYC traffic congestion . NYC has been one of the most congested cities in the world . The serious congestion causes a large amount of social and economic costs . In 2017, for example, businesses have suffered the most from traffic conges tion with an average of 14 % of travel time on weekdays in gridlock around NYC , where drivers wasted daytime hours being stuck in traffic jams ( Cookson 2018 ) . User - C rowdsourced Content and Its Value Waze was originally e stablished by Waze Mobile, an Israeli company , in 2009 . It was acquired by Google Inc. i n 2013 . Waze still provides different features from Google Maps mainly in the social network and interface aspect s ( Pandian 2018 ) . Waze users take active role s by posting alerts on traffic events, such as traffic jams, automobile accidents, police traps, closed roads, and other hazards , to inform other users of what is going on nearby . T hey pas sively patch up or react to the reports by commenting on and verify ing the alerts by using a thumbs - up feature . O ver the recent several years , Waze has dramatically boosted its market share in the mobile navigation category of App Store and Google Play, es pecially for mobile app users around metropolitan area s ( Leon 2016 ; Rao 2011a ) U.S. became far more pronounced immediately after Tim Cook, Apple CEO, apologized for glitches of Apple Maps in September, 2012 ( Cutler 2012 ) . second next to Google Maps for all time since 2013. According to Fast Company (2018 ) , Waze ranked 28th overall and first in the transportation sector as the W M ost I nnovative 53 C o mpanies in 2018 . Waze has continued to grow its market share collaborating with local governments , including 34 states along with the U.S. Department of Transportation , to manage urban traffic and upgrade transportation systems for Connected Citizen s P rogr am ( McCorvey 2018 ) . Waze app platform displays and updates user - crowdsourced information in real time on the live navigation maps, as shown in Figure 9 . Waze users report a traffic jam every 4.2 second s and an accident every 44 seconds on average , thus generating a massive amount of real - time traffic information ( Rao 2011b ) . The mobile virtual community of Waze users post s alerts which provide primary information on the real - time traffic content . In addition , users comment on initial alerts to add follow - up feedback as well as verify the alerts by turning thumbs - up to conduct collective confirmation of the primary information . Those user - crowdsourced updates Figure 9 . User - crowdsourced content on Waze navigation map 54 are the most unique features of Waze app platform. Self - interest of drivers triggers user engagement in real - time traffic updates, which cannot be expected even by any local radio stations. As a result, Waze users are better aware of traffic events with location - and time - sensitive updates by simply using Waze when driving as more and more users download and use the app. At this moment, around 90 million Waze users are contributing real - time information on traffic conditions , such as a utomobile accidents and road closures ( McCorvey 2018 ) . Their user - crowdsourced traffic content informs and updates urban commuters who were estimated to waste 42 hours in traffic a year on average costing $1,445 in idling fuel and time away in 24 0 U . S . cities ( Cookson 2018 ) . 13 On the other hand, it is not always true that the wisdom of crowds provides reliable real - time information due to potential risks of deception behavior ( Hendrix 2016 ; Love 2015 ; Madrigal 2018 ) . It is also possible the user - crowdsourced content to include unintentional fake information with inaccuracy due to the limits of human cognition rather than machine algorithm ( Hendrix 2016 ; Weise 2017 ) . The intentional deception , selfishness , and unintentional fake postings question the desirable effects of user contribution on welfare cons equences for relieving the urban traffic congestion and improving the mobility of urban transportation. THEORETICAL BACKGROUND Two rival economic theories provide the initial conceptual foundations for this research. According to c lassical utilitarianism , self - interest refers to the rationale for self - directed and 13 According to a recent report based on transportation analytics ( Cookson 2018 ) , Los Angeles, CA, and NYC, NY, were the most congested urban areas not only in the U.S. but also in the world in 2017. Angelenos spent 102 hours in traffic jams on average during the peak commute hours, costing $2,828 for each driver and $19.2 bill ion in total. New Yorkers idled an average of 91 hours in congestion during the peak congestion hours costing $2,982 per capita and $33.7 billion in total. 55 goal - seeking behavior ( Heilbroner 1 967 ; Mencher 1967 ) . It is an influential motive of human actions because people act in their own rational self - interest ( Bowles and Gintis 2000 ) . This research specifies the self - interest as altruism ( Levina and Arriaga 2014 ) . Waze users contribute by posing alerts, commenting, and turning thumbs - up on the alerts. Their self - interest reflects in their enjoyment of using technology or desires for engaging in social connect ions. Complementing the self - interest theories, the theory of cognitive a bsorption suggests that people use technology purely for fun and enjoyment without any intention of helping others. Users start adopting and keep using a technology because they expec t and perceive usefulness of the technology. The concept of cognitive absorption ( Agarwal et al. 1997 ) suggests that the instrumentality of technology is perceived with three dimensions: i.e., flow experience ( Trevino and Webster 1992 ; Wild et al. 1995 ) , playfulness ( Webster and Martocchio 1992 ) , and usability ( Adams et al. 1992 ) . Users deeply attend to and engage in their interactions with a technology when the technology is playful, easy to use, and help ful for their personal innovation ( Agarwal and Karahanna 2000 ) . That is, users attempt at, keep engaging in, and are absorb ed by their intense use of technology when their own interests such as flow experience, p layfulness, intrinsic interest, and curiosity are satisfied by using technology. The c ognitive a bsorption is also applied to user engagement in the mobile virtual community of Waze users. Urban drivers need to be quickly and locally updated of the fluctuat ing traffic conditions to drive better and safely as well as avoid traffic congestions and obstruction s. Some of them are playful and social enough to be willing to enjoy coexistence and interactions with others even when driving alone inside of their auto mobiles. The ir self - interest encourages them to use Waze and engage in user 56 contribution by reporting alerts on their traffic experiences to actively share useful information with others or by commenting and turning thumbs - ively react to and express thanks to someone else s contribution. Waze users are exposed to serious information overload and omit a large amount of given information, thus becoming highly selective to an affordable amount of key information ( Eppler and Mengis 2004 ) . As shown in Figure 9, a navigation map on the Waze app screen of a mobile device shows a variety of dynamic content simultaneously . 14 The dynamics and amount of user - crowdsourced content in the Waze app platform causes information - and cognitive - overload problems , which are more overwhelming than those in other social media platforms. On top of that, the information overload increases if a user drives around the urban region. Driving around great cities becomes more dynamic and complex as urban transportation systems enlarge and diversify with increasing numbers of automobiles and population over time. An urban driver experiences serious information overload because she/he drives monitoring the urban traffic with to process information, Waze users t end to ignore a large amount of given information with selective attention ; they experience a confusion between the overall perspective and details ( Eppler and Mengis 2004 ) regarding spatial and temporal information on the Waze navigation map. In contrast, s ocial welfare is broadly defined as political, economic, and religious/patriotic 14 When a driver observes the navigation map on Waze app platform, many symbols of multiple - type aler ts on traffic jams, car accidents, police traps, hazard issues, closed roads, and others keep appearing and disappearing with their location marks on Waze map. When the driver approaches the location identified by an alert, the alert pops up on the Waze ma p with additional detailed information, i.e., follow - up comments and the number of thumbs - other lines also display with different colors indicati ng the speed of traffic flows. Moreover, the driver is exposed to a variety of numbers and texts, which identify the estimated time of arrival, distances, speed limits, driving speed, and so on, as well as many advertisements of nearby shops, restaurants, gas stations with information of fuel prices, and others. It is overwhelming for a driver to pay attention to all the information on the Waze app platform. 57 expressions of collective social concern for collectivities of population ( Smith 1965 ) . I n modern society, welfare can be ensured with social benefits to and well - being for citizens and their families ( Marshall 1950 ) . In general, social welfare refers to benefits to the whole society and its members who experien ce certain risks rather than to a few individuals or interest groups. In this research , the concept of social welfare is defined as social and economic benefits to urban drivers in a specific context of urban transportation. Social welfare refers to better traffic condition s with greater aggregate mobility in this research . Urban drivers can achieve social and economic benefits by enjoying a high quality of driving environment with better urban transportation , as well as saving the potential opportunity cos ts with l ess traffic congestion . The mitigated traffic congestion and improved traffic mobility save a large amount of social and economic costs of citizens around urban areas, given that the immeasurable opportunity costs are borne by urban drivers due to their waste of time and energy idl ed away in traffic jams. These contrasting theories frame this research as it examines whether one of byproducts of user engagement with Waze is social welfare in the form of reduction of traffic delays around urban areas . This frame work highlights the social - welfare value of the mobile technology that connects collective users with each other as an interest - oriented community. SPATIAL PANEL DATA This research analyze s a spatial panel dataset of NYC . The city is one of th e most traffic - jammed urban regions in the world ( Cookson 2018 ; Walker 2018a ) . New Yorkers suffer from serious traffic congestion which causes a large amount of social and economic cos ts, i.e., their opportunity costs while idling time and fuel away in automobiles stuck in traffic ( Cookson 2018 ; Walker 2018b ) . A large - scale dataset was collected from multiple open sources , such as Waz e app APIs, the U.S. Census Bureau, NYC Police Department (NYPD), DemographicsNow, and 58 Weather Underground. The spatial panel dataset is structured as follows: Temporally , the dataset consists of two - time windows in 2016 : the first wave of 49 days from Jan uary 11 to February 28 and the second wave of another 49 days from July 11 to August 28 , i.e., 98 days in total. For the fourteen weeks, t he data were captured at 2,352 - time points (24 hours × 98 days) hour by hour around NYC . The data include 2 , 881 , 065 tr affic jams and 4 , 509 , 091 pieces of Waze user - crowdsourced content. The real - time generated content includes 1 , 886 , 166 alerts as well as 729 , 782 comments and 1 , 893 , 143 thumbs - up , added on the alerts within the same hour , with their location al information, i .e., coordinates on the map. T he raw Figure 10 . Neighborhood tabulation areas over boroughs of New York City 15 15 The map backdrop cartography is sourced from stamen.com using OpenStreetMap data. 59 Table 9 . Descriptive statistics of study variables and their data sources data were aggregated into the zone level, i.e., N eighborhood T abulation A rea (NTA). Spatially , NYC consists of a total of 195 NTAs , as depicted in Figure 10, over the five boroug hs , i.e., Brooklyn, Bronx, Manhattan, Queens, and Staten Island . S even NTAs were excluded . 16 As a result , the raw data were aggregated into the remaining 188 NTAs over the 16 The seven excl uded NTAs are not appropriate to spatial weight matrices for the spatial data approach due to the following reasons. Five NTAs are the areas of parks or cemeteries in which the contributing activities of Waze users are not available. Another NTA is the are a of Rikers Island (the main jail complex in NYC) which is spatially isolated from the other NTAs. The other NTA includes both areas of LaGuardia Airport and John F. Kennedy 60 Table 10 . Correlation matrix of study variables Variables 1 2 3 4 5 6 7 8 1. ln ( ) 1.0000 2. ln ( ) 0.2729 1.0000 3. ln ( ) 0.1850 0.7456 1.0000 4. ln ( ) 0.1907 0.7591 0.6981 1.0000 5. ln ( ) 0.6726 0.4588 0.3023 0.3154 1.0000 6. ln ( ) 0.6052 0.3008 0.2509 0.2336 0.5669 1.0000 7. ln ( ) 0.3295 0.4889 0.3753 0.4357 0.4669 0.2222 1.0000 8. ln ( ) 0.0599 0.1470 0.1199 0.1734 0.0111 0.0900 0.2677 1.0000 9. ln ( ) 0.0626 0.1429 0.1076 0.11 86 0.1223 0.0646 0.0830 0.0013 10. ln ( ) 0.0225 0.0518 0.0399 0.0433 0.0415 0.0246 0.0256 0.0006 11. ln ( ) 0.0792 0.0337 0.0356 0.0330 0.0766 0.0536 0.4216 0.0615 12. ln ( ) 0.2367 0.0375 0.0129 0.0 249 0.3362 0.0569 0.4161 0.1492 13. ln ( ) 0.2374 0.1809 0.0952 0.1085 0.3580 0.0698 0.5203 0.1321 14. 0.2661 0.1618 0.0720 0.0851 0.3904 0.0917 0.4912 0.1307 15. ln ( ) 0.0477 0.2437 0. 1854 0.2153 0.0925 0.0196 0.3188 0.0061 16. ln ( ) 0.1123 0.0899 0.0830 0.1087 0.1211 0.0704 0.3617 0.1625 Variables 9 10 11 12 13 14 15 16 9. ln ( ) 1.0000 10. ln ( ) 0.4181 1.0000 11 . ln ( ) 0.0132 0.0102 1.0000 12. ln ( ) 0.0257 0.0101 0.6259 1.0000 13. ln ( ) 0.0605 0.0141 0.6374 0.8585 1.0000 14. 0.0598 0.0196 0.5502 0.8727 0.8885 1.0000 15. ln ( ) 0.0446 0.0024 0.2483 0.1199 0.4453 0.2298 1.0000 16. ln ( ) 0.0292 0.0144 0.8465 0.7897 0.7047 0.6610 0.1298 1.0000 Table 11 . Distributions of observations across boroughs and weather c onditions in New York City Borough Obs. % Weather condition Obs. % Brooklyn 117,600 26.6 0 Clear 178,614 40.39 Bronx 84,672 19.15 Cloudy 226,190 51.15 Manhattan 65,856 14.89 Foggy 3,862 0.87 Queens 131,712 29.79 Rainy 21,206 4.8 0 Staten Island 42,336 9 .57 Snowy 12,304 2.78 Total 442,176 100 .00 Total 442,176 100 .00 NYC boroughs : 50 NTAs in Brooklyn, 36 NTAs in Bronx, 28 NTAs in Manhattan, 56 NTAs in Queens, and 18 NTAs in Staten Island. This research analyze s a set of balanced spatial panel data with 442,176 observations (1 88 NTAs × 2,352 - time points). Table 9 summarizes the detailed definitions, descriptive statistics, International Airport, which are located far away from each other: this singular N TA, unlike the other NTAs, is split into two zones spatially distant from each other. 61 and data sources for all the study variables measured at the NTA level. Table 10 provides correlations among the study variables. Tabl e 11 describes d istributions of all the observations across boroughs and weather conditions in N YC. Temporal Nature of Traffic Congestion Duration and User Contribution Regarding the temporal nature , Figure 11 visualizes the trajector ies of hour - by - hour a verages of traffic congestion duration and user contribution per NTA for weekdays in all of Figure 11 . Hour - by - hour averages of traffic congestion duration and user contribution in New York City 62 NYC and its five boroughs. Based on the hour - by - hour averages, the peaks of traffic congestion duration and user contribution start around 8:00 in the morning and sustain until around 19:59 in the evening, thus indicating the period as busy hours (rush and business hours). Among the five boroughs, Manhattan has the most traffic congestion duration and user contribution per NTA whereas Staten Island has the least. S patial Nature of Traffic Congestion Duration and User Contribution I n a spatial analysis , the measure of contiguity ( adjacency ) needs to be decided for a spatial weight matrix due to the spatial nature of data. C ontiguity refers to how to set boundaries am ong spatial neighbors. The most commonly used contiguity measures are rook bishop and queen ( Lloyd 2010 ) . 17 As depicted in Figure 12 , t sets the spatial neighbors, which share their common boundaries along borders only, as adjacent l units, which are adjacent at points only, as Figure 12 17 The names of contiguity measures originate from chessmen. 63 case contiguity sets the spatial neighbors, which share either of common borders or common points , as neigh boring units. contiguity are chosen to generate spatial weight matrices due to the research context of traffic transportation and mobile app users , which dynamically move along the network s of roads across spatial zones over the urban areas. contiguity is selected to generate a spatial weight matrix for the main analyses in this research . In general, the n - by - n spatial weight matrix is expressed as follows: where = 1 if Zone i and Zone j are adjacent to each other by sharing a boundary ( bnd ( i ) bnd ( j ); = 0 if they are not adjacent and share any boundary ( bnd ( i ) bn d ( j ) = ); = 0 if i = j . 18 That is, each entry of the spatial weight matrix indicates is one if NTA i and NTA j are adjacent as neighboring or zero if NTA i and NTA j are not adjacent to each other ( LeSage and Pace 2009 ) . The diagonal elements are set to zero in order to inhib it all the NTAs from neighboring themselves. Table 12 and Figure 13 summarize s tatistics and visualizes connectivity of the 188 - by - 188 spatial weight matrix based on the q , 18 This defines the queen rook is defined to exclude the neighbors, Zone j point (a shared corner point on a grid of spatial units) with a focal spatial unit, Zone i , as follows: where indicates the length of shared boundary, bnd ( i ) bnd ( j ) , between Zone i and Zone j . 64 Table 12 . Statistics of spatial weight matrix (i) Summary (ii) Tabulation of links Matrix features Description Number of links Obs . Dimensions 188 x 188 1 3 Values 2 13 Min . 0 3 34 Min . > 0 0.09 1 4 47 Mean 0.005 5 37 Max . 1 6 30 Links 7 14 T otal 870 8 4 Min . 1 9 5 Mean 4.62 8 11 1 Max . 11 Sum 188 Figure 13 . Connectivity of spatial weight matrix respectively . A zone, i.e., N TA, is adjacent to 4.628 zones on average and the number of spatial neighbors ranges from one to eleven. In addition, the r ook s case contiguity is used for sensitivity and robustness checks of results from the main analyses. Appendix E provides a context - contiguity , as shown in Figure A2, as well as statistics and connectivity of the spatial weight matrix based on the rook contiguity , as summarized in Table A8 and visualized in Figure A3. 65 EMPIRICAL A PPROACH This research employs a set of multiple spatial panel data models to explore the effect of user contribution on traffic congestion duration i n NYC. For empirical analyses, a consideration is whether the s patial panel data collected from the geographically proximate entities , i.e., NTAs over NYC, are independent. Neighboring observations are more likely to be similar with each other than far - off observations across their remote locations ( Tobler 1970 ) . In this sense, the concept of spatial autocorrelation 19 calls for a spatial data approach on top of the panel data approach, the concept which suggests that spatially adja cent things are more similar with each other than spatially distant things ( Cliff and Ord 1970 ; Goodchild 1986 ; Sokal and Oden 1978 ; Upton and Fingleton 1985 ) . Spatial Autocorrelation The spatial autocorrelation s of traffic congestion duration ( ) and user contribution ( , , and ) are diagnose d to measure the dependence among NTAs over NYC with three different measures of spatial autocorrelation: i.e., I , Geary c G ( Lee and Wong 2001 ; 2010 ) . All t he measures of global spatial autocorrelation indicate that my spatial panel data around NYC are significantly exposed to spatial autocorrelation in terms of the dependent variable and independent variables of interest . That is, the correlat ions of the study variables between a focal 19 Spatial autocorrelation refers to a property of the mapped data which exhibits an organized pattern, i.e., systematic spatial vari ation in values across a map ( Upton and Fingleton 1985 ) . For example, if a high value at one locality is associated with high values at adjacent localities, the spatial autocorrelation is positive. On the other hand, if a high value and low value alternate betw een neighboring localities, the spatial autocorrelation is negative. The spatial similar to other objects/activities located nearby ( Goodchild 1986 ) . That is, the concept of spatial autocorrelation ed to everything else, but near things are more related ( Tobler 1970, p. 236 ) . 66 Table 13 . Diagnostics of spatial autocorrelations of study variables with spatial weight matrix Variables I E( I ) S.D. ( I ) z - score ln ( ) 0.569*** 0.005 0.047 12.170 0.315*** 0.005 0.047 6.780 0.292*** 0.005 0.047 6.308 0.282 *** 0.005 0.047 6.081 Variables c E( c ) S.D. ( c ) z - score ln ( ) 0.293*** 1.000 0.062 11.363 0.652*** 1.000 0.061 5.704 0.660*** 1.000 0.061 5.597 0.677*** 1.000 0. 057 5.679 Variables G E( G ) S.D. ( G ) z - score ln ( ) 0.027*** 0.025 0.000 7.130 0.028*** 0.025 0.001 4.269 0.029*** 0.025 0.001 3.173 0.028*** 0.025 0.001 3.27 2 Note. Two - tailed test s. * p < 0.10; ** p < 0.05; *** p < 0.01 . Figure 14 . Visualized spatial autocorrelations of study variables with spatial weight matrix 67 zone, NTA i , and their spatial lags, average values of the variables for the neighboring zones, NTA j s shown in Table 13 and Figure 14 , the study variables of interest are positively correlated with their spatial lags: for ln ( ) , I = 0.569 sign ificant at p < 0.01 ; for ln ( ) , I = 0.315 at p < 0.01 ; for ln ( ) , I = 0.292 at p < 0.01 ; for ln ( ) , I = 0.282 at p < 0.01 . The significant spatial autocorrelations are also diagno s ed with the spatial weight matrix based on the q contiguity (refer to Table A9 and Figure A4 in Appendix F for details). Spatial Panel Estimators Due to the significant spatial autocorrelation, this research employs spatial panel estimators, i. e., spatial autoregressive (SAR) model, spatial Durbin model (SDM), spatial error model (SEM), and spatial autocorrelation (SAC) model, to leverage both the temporal and the spatial nature of the unique and real - time data on Waze users and traffic conditio n s around NYC. The spatial panel data models are employed to deal with not only temporal but also spatial spillovers, which increase (or decrease) as distance between zones decreases (or increases) ( LeSage and Pace 2009 ) , based on the Quasi - Maximum Likelihood (QML) estimation for balanced spatial panel datase ts ( Lee and Yu 2010 ; Yu et al. 2008 ) . Figure 15 summarizes estimation specifications of all the spatial panel models and their relationships . Based on Belotti et al. ( 2016 ) explanation, each spatial panel data model can be expressed as described below. The equation of SAR model can be expressed, as follows: , t T , ( 1 ) 68 Note. While the estimation models in the dotted - line rectangle s are not employed, the models in the solid - line rectangles are employed for the spatial panel estimations in this research: OLS , SAR , SDM , S EM , SAC , DSAR , and DSDM are used whereas SLX , SDEM , and GNS are not used in this research. OLS: ordinary least squares model; SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SA C: spatial autocorrelation model; DSAR: dynamic spatial autoregressive model; DSDM: dynamic spatial Durbin model; SLX: spatial lag of X model; SDEM: spatial Durbin e rror model; GNS: general nesting spatial model. Figure 15 . Flow chart of spatial panel model specification s and their relationships 20 20 ( 2015, p. 343 ) comparison of spatial econometric model specifications. 69 where the spatial weight matrix 21 is a 180 × 180 matrix of spatial weights of NTAs over NYC refers to the 180 - by - one column vector of ln ( ) for each time period t T . is the 180 × k matrix of regressors, i.e., the logged independent and control variables, where k is the number of regressors. The estimation assum es that in the case of random - effects. The is a vector of parameters to be estimated in the fixed - effects variant. The SAR model assumes that and for i j and/or t s . Generalized from the SAR model, SDM also includes spatiall y weighted independent variables, as follows: . ( 2 ) The SDM is generalized by employing spatial weights which are different for the spatially lagged dependent variable and th e spatially weighted regressors . SAC model integrates the SAR model with a spatial autoregressive error, as follows: , ( 3 ) , ( 4 ) 21 In this research, the row - normalized weights are used for all the spatial panel analyses. That is, of the spatial weight matrix are rescaled as follows: . As the standard in spatial econometrics, the row - normalization constrains the parameter space and makes all the spatial panel estimations c omparable in this research. 70 The SAR is a special case of the SAC specification. Finally, SEM identifies spatial autocorrelation in the error term, as follows: , ( 5 ) . ( 6 ) The SEM is a special case of the SAC model as well as one of the SDM. In addition, this research estimate s not only the static but also the dynamic effects of user contribution on traffic congestion duration. The SAR model, SDM, SEM, and SAC model estimate the static effects of use r contribution on traffic congestion duration by including only the contemporaneous values of their dependent and independent variables ( Belotti et al. 2016 ) . To include even the dynamic effects, the SAR models and SDMs can be extended by adding the temporally lagged dependent variable and/or both the spatially and temporally dependent variable into the left side of their estimations, respectively, as follows: , ( 7 ) . ( 8 ) EMPIRICAL RESULTS AND FINDINGS This research explores the effects of user contribution on traffic congestion duration in NYC. The effects are analyzed with spatial panel estimators and their dynamic models. In addition, follow - up analyses focus on the quadratic relationship of primary information and 71 traffic congestion duration as well as potential interactions between the different types of user contribution . These analyses are also conducted with the observations during rush and business hours of weekdays. A welfare analysis is followed to show the effectiveness of user contribution with the currency value based on opportunity costs. The other additional ana lyses are used to check the s ensitivity and r obustness of the main empirical findings. Effects of User Contribution on Traffic Congestion Duration P anel data models pooled OLS and fixed - and random - effects estimators consistently show significant, negativ e coefficients of ln ( ) , a s shown in Table 14. The panel results suggest that all the types of user contribution mitigate traffic congestion by reducing its existing duration. In Column 2, for example , results from the fixed - effects estimator indicate: for Table 14 . Effects of user contribution on traffic congestion duration: Results of panel data models Variables Dependent variable: = ln ( ) (1) Pooled OLS a (2) Fixed effects b (3) Random effects b 0.0541*** 0.0504*** 0.0505*** (0.0022) (0.0051) (0.0051) 0.0207*** 0.0218*** 0.0218*** (0.0021) (0.0054) (0.0053) 0.0047*** 0.0116*** 0.0113*** (0.0016) (0.0038) (0.0037) Time - variant control variables Yes Yes Yes Time - invariant control variables Yes No Yes Weather condition indictors Yes Yes Yes Borough indicators Yes No Yes Date and hour - of - day indicators Yes Yes Yes Obs. 442,176 442,176 442,176 R 2 0.5411 0.4914 Number of NTAs 188 188 Note. a Robust standard errors in parentheses. b Robust standard errors clustered on NTA in parentheses. Estimates for constant, control variables, and dummy weather condition, borough, and date and hour - of - day indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 72 . The empirical re sults from the panel data estimations might be exposed to the spatial autocorrelations of the study variables. This research therefore employs spatial panel estimators the SAR model, SDM, SEM, and SAC model with the spatial weight matrix based on the qu to extend the spatial data approach into the panel data approach. When controlling for both spatial and temporal fixed effects, as summarized in Table 15, the spatial panel estimators consistently show significant, negative coefficien ts of ln ( ) even with the space - lagged dependent variable. In Column 1, results from the SAR estimator are shown: for These results indicate that primary information (via alert postings) and follow - up feedback (via comments) significantly reduce the duration of traffic jams, even with the significant space - lagged effects of neighboring zones on a focal zone: = 0.2185 at . This is consistent even with a spatial autoregressive error (see the SEM results in Column 3) and with the spatial autocorrelation in the error term (see the SAC results in Column 4). While has insignificant, negative coefficients ln ( ) from the SAR, SEM, and SAC estimators, it has a significant, negative coefficient from the SDM estimator which is extended with spatially weighted regressors as additional explainers. As shown in Table 15, C olumn 2 summarizes the SDM results: This indicates that c ollective confirmation (via thumbs - up) does not substantially mitigate traffic congestion without spatial spillover from a focal zone to adjacent zones. However, the SDM suggests that c ollective confirmation mitigates traffic 73 Table 15 . Effects of user contribution on traffic congestion duration: Results of spatial panel data models with spatial weight matri x Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0505*** 0.0422*** 0.0099 0.0473*** 0.0470*** (0.0050) (0.0049) (0.00 71) (0.0052) (0.0052) 0.0195*** 0.0281*** 0.0253*** 0.0222*** 0.0222*** (0.0050) (0.0047) (0.0076) (0.0051) (0.0051) 0.0028 0.0059** 0.0078 0.0042 0.0043 (0.0033) (0.0030) (0.0049) (0.0033) (0.003 3) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2185*** 0.2649*** 0.0121 (0.0064) (0.0071) (0.0 229) 0.2733*** 0.2839*** (0.0078) (0.0233) Obs. 442,176 442,176 442,176 442,176 R 2 0.5274 0.5299 0.5293 0.5290 Number of NTAs 188 188 188 188 Panel length 2,352 2,352 2,352 2,352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . congestion duration, given that its spillover exists. Regarding the spillove rs of user contribution, only follow - up feedback in a focal zone, NTA i , has a significant, positive coefficient on the durations of traffic congestion in adjacent zones, NTA j = 0.0253 significant at p < 0.01 (see Column 2). S patial panel estima tors are also employed with a sub - dataset of the observations which occurred between 8:00 and 19:59, i.e., rush and business hours , for weekdays only. This research conducts these spatial panel analyses with the sub - dataset to explore how the effects of us er contribution on traffic congestion duration differentiate during rush and business hours. As shown in Table 16, Column 1 summarizes results from the SAR model suggesting that the negative coefficients of ln ( ) are sustained significantly but their magnitude decreases during rush and business hours: for 74 Table 16 . Effects of user contribution on traffic congestion duration for rush an d business hours : Results of spatial panel data models with spatial weight matrix contiguity Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0293*** 0.0268*** 0.0006 0.0280*** 0.0277*** (0.0044) (0.0048) (0.0075) (0.0047) (0.0047) 0.0092** 0.0120*** 0.0154** 0.0099** 0.0099** (0.0040) (0.0042) (0.0065) (0.0041) (0.0041) 0.0012 0 .0019 0.0057 0.0015 0.0015 (0.0027) (0.0030) (0.0051) (0.0029) (0.0029) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Y es Yes 0.2298*** 0.2617*** 0.0352 (0.0081) (0.0072) (0.0430) 0.2628*** 0.2931*** (0.0073) (0.0373) Obs. 153,408 153,408 153,408 153,408 R 2 0.3398 0.3340 0.3379 0.3367 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . . At the same time, the magnitude of space - lagged effects (from neighboring zones to a focal zone) increase for the peak hours: = 0.2298 significant at That is, primary information and follow - up feedback are still effective in reducing the duration of traffic jams, but their effectiveness decreases during the peak hours when the traffic congestion is too heavy around NYC. This may be also related to the increase in the spatially lagged effects for the peak hours. On the other hand, collective confirmation is not sufficiently effective in reducing the duration of traffic jams for rush and business hours. These SAR results about the weakened effecti veness of user contribution during the peak hours are also consistent with results from the other spatial panel data models, i.e., the SDM, SEM, and SAC estimators, as shown in Columns 75 2 - 4. In Column 2, the SDM estimator indicates the spatial spillover of follow - up feedback is still significant during rush and business hours. However, the spatial spillover also weakens for the peak hours: = 0 .0154 significant at p < 0.0 5 . Based on the SDM results, spatial spillovers of the other user contribution types are not significant. The spatial panel data mo dels account for the unobserved effects not only over time but also across spac e, thus enabling robust inferences from the their results ( Elhorst 2014 ) . The spatial panel data models, i.e., QML estimators , provide consistent results with spatial - and temporal - fixed effects ( Lee and Yu 2010 ) . Even though the number of time periods T is large relative to the number of individual unit s n , like the spatial panel data of this research , the QML estimator s eliminate the asymptotic bias and provide a centered confidence interval ( Yu et al. 20 08 ) . On top of such advantages of the spatial panel estimators, this research additionally employs a matching method in a quasi - natural experimental setting . As described in Appendix G, results from the matching method also support the main findings from the spatial panel estimators empirically (see Table A10 for details). The consistent results might mitigate any potentials of endogeneity regarding the causation of interest. Dynamic Effects of User Contribution on Traffic Congestion Duration Dynamic spa tial panel data models are employed to control for not only short - and long - term effects of user contribution on traffic congestion duration but also its local and periphery effects in a single estimation. The d ynamic spatial autoregressive (DSAR) models a nd dynamic spatial Durbin models (DSDMs) are specified with either of a time - lagged dependent variable or a space - lagged dependent variable as well as with both of them. 76 As described in Table 17, DSAR and DSDM estimators consistently support that primary information and follow - up feedback traffic congestion duration in short and long term as well as temporally and spatially. Column 3 summarizes the DSAR results with the temporally and both the temporally and spatia lly lagged dependent variable. First of all, the space - lagged traffic congestion influences the duration of traffic congestion in a focal zone from its adjacent zones: = 0.2164 significant at . In addition, the time - lagged traffic congestion also has a significant, positive effect on the current traffic congestion over time: = 0.0179 at . Both the time - and space - lagged traffic congestion also influ ence the current traffic congestion in a focal zone not only across boundaries of the zone but also over time: = 0.0135 at . When controlling for the mixed locally and peripheral ly long - term effects, in addition, the local and short - term effects of ln ( ) : for the dependent variable lagged in time only (see Column 1) or the one lagged in both time and space only (see Column 2). On the other hand, DSDMs consistently show that the effectiveness of collective conf irmation is also significant given that its spatial spillover exists when controlling for the mixed locally and peripheral ly long - term effects. In Column 6, ln ( ) with the temporally and the both temporally and spatially lagged dependent variable: = 0.0059 significant at p < 0.0 5. The the dependent variable lagged in time (in Column 4) and lagged in both time and space (in Column 5). In addition, Column 6 indicates that the spatial spillover of follow - up feedback is significant and positive = 77 Table 17 . Dynamic effects of user contribution on traffic congestion duration: Results of dynamic spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) DSARs DSDMs (1) (2) (3) (4) (5) (6) Main Main Main Main Main Main 0.0200*** 0.0179*** 0.0208*** 0.0180*** (0.0023) (0.0023) (0.0022) (0.0023) 0.0220*** 0.0135*** 0.0275*** 0.0190*** (0.0 042) (0.0042) (0.0041) (0.0041) 0.0506*** 0.0504*** 0.0505*** 0.0423*** 0.0095 0.0421*** 0.0097 0.0423*** 0.0095 (0.0049) (0.0049) (0.0049) (0.0049) (0.0070) (0.0049) (0.0071) (0.0049) (0.0070) 0.0198*** 0.0196*** 0.0199*** 0.0284*** 0.0252*** 0.0282*** 0.0250*** 0.0284*** 0.0250*** (0.0049) (0.0049) (0.0049) (0.0046) (0.0075) (0.0046) (0.0076) (0.0046) (0.0075) 0.0029 0.0029 0.0029 0.0059** 0.0076 0.0059** 0.0 077 0.0059** 0.0076 (0.0033) (0.0033) (0.0033) (0.0029) (0.0048) (0.0030) (0.0049) (0.0029) (0.0048) Spatial fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Time - variant control variables Y es Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Yes 0.2172*** 0.2168*** 0.2164*** 0.2641*** 0.2637*** 0.2635*** (0.0064) (0.0063) (0.0063) (0.0071) (0.0071) (0.0071) Obs. 441,988 441,988 441,988 441,988 441,988 441,988 R 2 0.5289 0.5288 0.5295 0.5326 0.5337 0.5346 Number of NTAs 188 188 188 18 8 188 188 Panel length 2,352 2,352 2,352 2,352 2,352 2,352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted fro m the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). DSAR: dynamic spatial autoregressive model; DSDM: dynamic spatial Durbin model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 78 0.0250 at p < 0.0 1 whereas the other types of user contribution do not have their significant spatial spillovers. These results are consistent the temporally lagged dependent variable only (in Column 4) or both temporally an d spatially lagged one only (in Column 5). Quadratic R elationship of P rimary I nformation and T raffic C ongestion D uration This research explores whether primary information (via alerts) has a linear or quadratic relationship with the reduced duration of tr affic congestion. As shown in Table 18, the spatial panel estimators ( added the squared term of consistently support the quadratic Table 18 . Quadratic relationship of primary information and traffic congestio n duration: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0501*** 0.0549*** 0.0125* 0.0551*** 0.0552*** (0.0045) (0.0043) (0.0070) (0.0046) (0.0046) 0.0588*** 0.0725*** 0.0237 0.0700*** 0.0706*** (0.0094) (0.0086) (0.0149) (0.0095) (0.0095) 0.0142*** 0.00 76** 0.0192*** 0.0145*** 0.0145*** (0.0042) (0.0038) (0.0060) (0.0042) (0.0042) 0.0031 0.0003 0.0065 0.0022 0.0021 (0.0028) (0.0026) (0.0045) (0.0028) (0.0028) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Y es Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2164*** 0.2643*** 0.0155 (0.0063) (0.0069) (0.0218) 0.2726*** 0.2861*** (0.0075) (0.0219) Obs. 442,1 76 442,176 442,176 442,176 R 2 0.5293 0.5320 0.5320 0.5317 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitte d from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0. 10; ** p < 0.05; *** p < 0.01 . 79 relationship of primary information and traffic jam duration. In Column 1, for example, results of the SAR estimator indicate: for the squared term of ln ( ) , = 0.0501 significant at p < 0.0 1; for ln ( ) , = 0.0588 at p < 0.0 1. While ln ( ) has a significant, positive coefficient, its squared term has a significant, negative coefficient. Based on the results, the relationship between primary information and traffic jam duration is quadratic, i.e., a decreasing, concave - down relationship. This suggests that as the amount of primary information increases, the duration of traffic congestion decreases and the magnitude of decrease in the duration increases further. As shown in Table 19, all the spatial panel estimators indicate that the negative coefficient of the squared ln ( ) is significant yet its magnitude decreases during rush and business hours: e.g., = 0.0140 significant at p < 0.0 1 in the SAR estimation (see Column 1). At the same time, the positive coefficient of ln ( ) becomes insignificant or degreases for the peak hours: e.g., = 0.0076 ns according to the SAR results (see Column 1). That is, the spatial panel estimators with the squared term of the decreasing, concave - down relationship flattens during rush and business hours of a weekday. Based on the SAR results, Figure 16 visualizes that the quadratic relationship with predictive margins with 95% confidence intervals. As visualized, the quadratic relationship is decreasing, concave do wn for all time but becomes flat for the peak hours. 80 Table 19 . Quadratic relationship of primary information and traffic congestion duration for rush and business hours : Results of spatial panel data models with spatial weight m atrix based on Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0140*** 0.0211*** 0.0176*** 0.0187*** 0.0192*** (0.0031) (0. 0034) (0.0052) (0.0033) (0.0034) 0.0076 0.0256*** 0.0414*** 0.0201** 0.0218** (0.0088) (0.0092) (0.0148) (0.0092) (0.0092) 0.0024 0.0021 0.0072 0.0010 0.0007 (0.0038) (0.0040) (0.0066) (0.0040) (0.0040 ) 0.0025 0.0036 0.0071 0.0031 0.0031 (0.0027) (0.0030) (0.0051) (0.0029) (0.0029) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2296*** 0.2620*** 0.0397 (0.0081) (0.0072) (0.0430) 0.2633*** 0.2974*** (0.0073) (0.0371) Obs. 153,408 153,408 153,408 153,408 R 2 0.3408 0.3355 0.3394 0.3381 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . Figure 16 . Quadratic relationship of primary information and traffic congestion duration 81 I nteraction s B etween Different Types of User Contribution Potential interactions among the different user contribution types are explored in terms of their synergies in reducing traffic congestion duration. Based on the spatial panel data models, the intera ction between primary information and follow - up feedback is consistently significant: e.g., for the interaction term, ln ( ) × ln ( ) , = 0.0487 significant at p < 0.0 1 from the SAR model (see Column 1 of Table 20). Th i the effectiveness primary information (via alerts) in reducing the duration of traffic jams strengthens as follow - up feedback (via comments on the alerts) increases. Table 20 . Interaction between primary informa tion and follow - up feedback: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) ( 1 ) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0374*** 0.0292*** 0.0092 0.0336*** 0.0333*** (0.0048) (0.0048) (0.00 66) (0.0050) (0.0050) 0.0933*** 0.0994*** 0.0085 0.1030*** 0.1034*** (0.0108) (0.0098) (0.0191) (0.0107) (0.0107) 0.0013 0.0045 0.0078* 0.0027 0.0027 (0.0031) (0.0028) (0.0047) (0.0031) (0.0031) 0.0487*** 0.0558*** 0.0162* 0.0544*** 0.0546*** (0.0052) (0.0048) (0.0094) (0.0052) (0.0053) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control varia bles Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2175*** 0.2649*** 0.0157 (0.0064) (0.0070) (0.0224) 0.2734*** 0.2871*** (0.0077) (0.0226) Obs. 442,176 442,176 442,176 442,176 R 2 0.5285 0.5311 0 .5308 0.5305 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0. 10; ** p < 0.05; *** p < 0.01 . 82 As depicted in Table 21, however, the moderating effect weakens during the peak hours of a weekday compared to that during all time: e.g., for the interaction term, ln ( ) × ln ( ) , = 0.0153 significant at p < 0.0 1 in the SAR estimation (see Column 1). P redictive margins in Figure 17 visualizes the interaction between primary information and follow - up feedback for all time and their reduced synergy for rush and business hours. In Table 22, the multiple spatial panel data models show the significant interaction term of = 0.0351 significant at p < 0.0 1 from the SAR results (see Column 1) collective confirmation , the more strengthen ed effectiveness primary information in reducing the duration of traffic jams. T able 21 . Interaction between primary information and follow - up feedback for rush and business hours : Results of spatial panel data models with spatial weight matrix contiguity Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0231*** 0.0185*** 0.0082 0.0199*** 0.0193*** (0.0046) (0.0049) (0.0077) (0.0049) (0.0048) 0.0273*** 0.0459*** 0.0479*** 0.0401*** 0.0416*** (0.0083) (0.0090) (0.0166) (0.0087) (0.0089) 0.0017 0.0023 0.0063 0.0020 0.0020 (0.0027) (0.0030) (0.0051) (0.0029) (0.0029) 0.0153*** 0.0248*** 0.0 268*** 0.0212*** 0.0218*** (0.0034) (0.0038) (0.0066) (0.0037) (0.0037) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2297*** 0.2620*** 0.0397 (0.0081) (0.0072) (0.0431) 0.2633*** 0.2975*** (0.0073) (0.0372) Obs. 153,408 153,408 153,408 153,408 R 2 0.3404 0.3347 0.3388 0.3376 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 83 Figure 17 . Interaction between primary information and follow - up feedback Table 22 . Interaction between primary information and collective confirmation: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) ( 1 ) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0242*** 0.0154*** 0.0113 0.0195*** 0.0191*** (0.0053) (0.0054) (0.00 75) (0.0056) (0.0056) 0.0040 0.0033 0.0219*** 0.0032 0.0032 (0.0040) (0.0037) (0.0060) (0.0040) (0.0040) 0.0518*** 0.0535*** 0.0042 0.0554*** 0.0555*** (0.0067) (0.0066) (0.0113) (0.0069) (0.0069) 0.0351*** 0.0386*** 0.0088 0.0386*** 0.0387*** (0.0050) (0.0046) (0.0072) (0.0051) (0.0051) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variabl es Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2175*** 0.2648*** 0.0147 (0.0064) (0.0070) (0.0223) 0.2731*** 0.2859*** (0.0077) (0.0225) Obs. 442,176 442,176 442,176 442,176 R 2 0.5286 0.5314 0.5 310 0.5307 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control v ariables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0. 10; ** p < 0.05; *** p < 0.01 . 84 As summarized i n Table 23, however, t he moderating effect of collective confirmation on effectiveness of primary information also decreases during rush and business hours compared to other times: e.g., for the interaction t erm, ln ( ) × , = 0.0068 is significant at p < 0.0 5 from the SAR estimation (see Column 1). In Figure 18, the p lots of predictive margins depict the interaction between primary information and collective confirmation for all time and their reduced synergy for rush and business hours. Finally, this research analyzes mutual interactions between the two types of additional information, i.e., follow - up feedback and collective confirmation, in reducing the duration of traffic congestion . Table 24 indicates that the interaction term of Table 23 . Interaction between primary information and collective confirmation for rush and business hours : Results of spatial panel data models with spatial weight matrix based on Variables Dep endent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0231*** 0.0166*** 0.0163* 0.0187*** 0.0180*** (0.0050) (0.0053) (0.0087) (0.0053) (0.0053) 0 .0061 0.0065 0.0068 0.0052 0.0050 (0.0038) (0.0040) (0.0066) (0.0040) (0.0040) 0.0135** 0.0242*** 0.0411*** 0.0203*** 0.0211*** (0.0056) (0.0061) (0.0109) (0.0060) (0.0061) 0.0 068** 0.0127*** 0.0194*** 0.0105*** 0.0109*** (0.0028) (0.0031) (0.0052) (0.0030) (0.0031) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition in dictors Yes Yes Yes Yes Yes 0.2298*** 0.2619*** 0.0375 (0.0081) (0.0072) (0.0431) 0.2632*** 0.2954*** (0.0073) (0.0373) Obs. 153,408 153,408 153,408 153,408 R 2 0.3402 0.3344 0.3386 0.3374 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 85 Figure 18 . Interaction between primary information and collective confirmation Table 24 . Interaction between follow - up feedback and collec tive confirmation: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) ( 1 ) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0544*** 0.0472*** 0.0074 0.0519*** 0.0516*** (0.0051) (0.0050) (0.00 72) (0.0053) (0.0053) 0.0218*** 0.0214*** 0.0045 0.0254*** 0.0256*** (0.0065) (0.0065) (0.0114) (0.0067) (0.0067) 0.0109*** 0.0105*** 0.0011 0.0116*** 0.0116*** (0.0032) (0.0032) (0.0063) (0.0033) (0.0033 ) 0.0197*** 0.0240*** 0.0107* 0.0228*** 0.0230*** (0.0036) (0.0036) (0.0063) (0.0038) (0.0038) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant contr ol variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2183*** 0.2651*** 0.0139 (0.0064) (0.0071) (0.0228) 0.2735*** 0.2856*** (0.0078) (0.0232) Obs. 442,176 442,176 442,176 442,176 R 2 0.5278 0.5304 0.5299 0.5296 Number of NTAs 188 188 188 188 Panel length 2,352 2,352 2,352 2,352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0. 10; ** p < 0.05; *** p < 0.01 . 86 e.g., = 0.0197 at p < 0.0 1 based on the SAR results (see Column 1) he effectiveness of follow - up feedback increases as the amount of collective confirmation increases, or vi ce versa . In Table 25, all the spatial panel estimations indicate that the synergy between follow - up feedback and collective confirmation is less effective during the peak hours than for the other time of a weekday: e.g., for the interaction of , = 0.0050 is significant at p < 0.0 1 in the SAR model (see Column 1). Plots of predictive margins in Figure 19 visually compare the synergy for all time with that for the peak hours. Table 25 . Interaction between follow - up feedback and collective confirmation for rush and business hours : Results of spatial panel data models with spatial weight matrix based on Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0303*** 0.0289*** 0.0040 0.0296*** 0.0293*** (0.0044) (0.0049) (0.0077) (0.0047) (0.0047) 0.0014 0.0079 0.0195* 0.0069 0.0076 (0.0055 ) (0.0059) (0.0113) (0.0058) (0.0059) 0.0057* 0.0101*** 0.0204*** 0.0085** 0.0088** (0.0034) (0.0037) (0.0069) (0.0036) (0.0037) 0.0050** 0.0097*** 0.0167*** 0.0081*** 0.0084** * (0.0022) (0.0025) (0.0045) (0.0024) (0.0025) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2298** * 0.2618*** 0.0368 (0.0081) (0.0072) (0.0431) 0.2630*** 0.2948*** (0.0073) (0.0373) Obs. 153,408 153,408 153,408 153,408 R 2 0.3400 0.3340 0.3382 0.3370 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust stan dard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), l n ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 87 Figure 19 . Interaction between follow - up feedback and collective confirmation Welfare Analysis To avoid an overestimation, this conservatively estimates the social - welfare value of user contribution in relieving the urban traffic congestion. Urban drivers and commuters, stuck in the frequent and heavy traffic jams around NYC, suffer a large amount of social and economic costs, 88 i.e., opportunity costs of idling time and wasted fuel. As fou nd from the spatial panel data analyses, the duration of traffic congestion may be substantially reduced thanks to user contribution enabled by using the crowdsourcing features: posting alerts and adding comments and thumbs - up on the alerts. The real - time traffic content is shared with other users and processed by Waze users for their better driving mobilities. This research explains the welfare value of user engagement via a mobile app, i.e., the value of the improved mobilities of urban transportation, in the U.S. dollar currency . According to the most recent report from the U.S. Department of Transportation, the hour - by - hour value of travel time savings per person in the U.S. was $14.10 22 in 2015 ( White 2016 ) . The value, $14.10, is adjusted to $14.28 in 2016 due to the inflation in the U.S. (i.e., the inflation rate from 2015 to 2016 is 1.3%). First, this research estimates the welfare value realized by individual types of user contribution. In doing so, I rely on the log - log transformed results from the SAR estimation with spatial weight matrix based on q (see Column 1 of Table 15) and the hourly value of time per driver ($14 .28) in the U.S., 2016. This welfare analysis of individual user contribution types includes value estimations of primary information (via alerts) and follow - up feedback (via comments) only. Collective confirmation (via thumbs - up) is excluded because its e ffectiveness is not significant based on the SAR estimator. Regarding the saved congestion cost per driver thanks to primary information , the main analyses with the SAR estimator suggest that when increased double around its hourly average (4.27 alerts), the 3.56% duration of traffic congestion (0.37 minutes) reduced from the average duration (10.40 22 The hourly value travel time savings per person is an average weighted by distributions of vehicle travels in the U.S. by all the purposes: i.e., the distributions of local travels via vehicles are 95.4% for the personal purpo se and 4.6% for the business purpose. According to White (2016 ) , the figures were derived from the 2001 National Household Travel Su rvey using annual person - mile data available at: nhts.ornl.gov. 89 minutes). 23 A driver around NYC therefore saved $0.09 per hour on average. The commuter saved th e hourly congestion cost due to the traffic jams two times per day. In other words, she/he commutes twice a day from homeplace to workplace and from workplace to homeplace. In addition, the number of business days was 261 days in 2016. Therefore, the annua l congestion cost per driver was saved by $46.03 thanks to primary information crowdsourced by Waze users, as follows: (Annual congestion cost per driver saved thanks to primary information) = $0.09 × two times a day × 261 business days = $46.03. In a si milar way, this research estimates how much follow - up feedback , i.e., comments added on alerts, saved congestion cost per driver in NYC. Based on the results of the SAR estimation ( ), when increased double around its average (1.65 comments), the 1.36% duration of traffic congestion (0.14 minutes) reduced from the average duration (10.40 minutes). On average, a NYC commuter saved $0.03 per hour. Thus, the annual congestion cost per driver saved thanks to follow - up feedback, added by Waze users, was $17.58, as follows: (Annual congestion cost per driver saved thanks to follow - up feedback) = $0.03 × two times a day × 261 business days = $17.58. 23 This research interprets the meaning of a negative coefficient as effect size in the log - log transformed functional form, i.e., ln ( ) = f ( ln ( )) ( Lin et al. 2013 ) . According to Vittinghoff et al. (2011 ) , the coefficient can be interpreted into the effect size, as follows: for a 100% increase in , decreases on average by the percentage of 100 × ( 1) which is approximately equal to when | | < 0.1. In such a conservative way, the effect sizes are interpreted for all the welfare estimations in this research. 90 In addition, the welfare value of total user contribution is estimated including primary information, follow - up feedback, and collective confirmation altogether, based on the results of an additional SAR estimation. In doing so, I keep including ln( ) and all the control variables into the SAR esti mation while excluding ln( ) and ln( ) due to the existing correlations between the different types of user contribution. In the additional estimation, the coefficient of ln( ) indicates the total eff ectiveness of three user contribution types by including the mutual effects, omitted due to their correlations when ln( ) and ln( ) are excluded, in addition to their individual effects on the reduced duration of traffic c ongestion. The additional estimation indicates: = 0.0630 significant at p < 0.01. Thus, when increased twice around its hour - by - hour average and and also increased in proportion to their correl ations with , the 4.46% duration of traffic jams (0.46 minutes) reduced from the average duration, as shown in Figure 20 . This implies that a driver saved $0.11 per hour on average. In 2016, the annual total congestion cost per driver there fore was reduced by $57.67 thanks to all the type of user contribution by Waze users around NYC, as follows: (Annual congestion cost per driver saved thanks to total user contribution) = $0.11 × two times a day × 261 business days = $57.67. Realized than ks to total user contribution in 2016, the per - capita social - welfare value ($57.67) amounts to 2.28% of annual total congestion cost per driver ( $2,533 ) estimated by Cookson and Pishue (2017 ) . 24 In 2016, the total congestion cost per driver in NYC ranked first among cities in 24 To avoid over estimating the social - welfare value, I estimate the annual congestion cost per driver saved thanks to user contribution in a conservative way as Cookson and Pishue (2017 ) estimated the annual total congestion cost per 91 the world. Given a large number of drivers a nd commuters, the social - welfare value may be a substantial amount of saved opportunity costs around the urban areas. Figure 20 . Annual congestion cost per driver saved thanks to total user contribution Sensitivity and R obustn ess C hecks Based on all the aforementioned empirical results, first of all, the consistent results across the multiple spatial panel estimator s support findings on the effectiveness of user contribution in reducing the duration of traffic congestion. As sh own in Tables A11 - A21 (see Appendix H), second, all the results with the spatial weight matrix concern on the sensitivity of the findings. On top of the spatial panel data models with both spatial and temporal fixed effe cts, I additionally conduct spatial panel data estimators with driver in NYC, 2016. 92 random effects, spatially fixed effects only, and temporally fixed effects only, which consistently support robustness of the findings (see Tables A22 - A24 in Appendix I). Finally, Tobit models for additional robustness checks are also employed to deal with the left - censoring when ln ( ) = 0. When a zone NTA i has no traffic jams at Time t , the observation has a zero value for the dependent variable, i.e., traffic jams duration. As shown in Tables A25 and A26 (see Appendix J), both the cross - sectional and panel Tobit models 25 provide consistent results, thus supporting robustness of the findings further. All in all, the se consistent results suggest substantial robustness of the empirical findings. DISCUSSION The empirical findings suggest that user contribution substantial ly reduc es the duration of traffic jams and effectively removes opportunity costs of urban traffic congestion, thus showing mobilities in NYC. S erious traffic jams around urban areas cause a large amount of s ocial and economic costs due to not only the futile waste of time, which drivers spend in doing nothing inside of vehicles, but also the wasteful consumption of fuels for the unmoving vehicles stuck in traffic jams. Based on the empirical findings, this re search propose s that user contribution is effective in generating social and economic benefits by reducing costs by 2.28% of annual total traffic congestion cost per driver in 2016. 25 The panel Tobit model is a random - effects estimation using the adaptive Gauss - Hermite quadrature approximation. The numerical integration is to calculate a likelihood function based on the maximum - likelihood estimation. 93 Theoretical and Empirical Contributions In various research context, it h as been skeptical whether society benefits when individual citizen s act in their own interest s ( Cassidy 2009 ) . In this sense, this research theorizes the prosocial effects of user contribution, i.e., its unintended consequence for urban transportation. The empirical findings may confirm that the s elf - i nterest of mobile app users b enefit the whole s ociety of citizens around urban areas. That is, user contribution driven by the self - interest of Waze users benefits the urban drivers by reducing the duration of traffic congestion and improving their mobilities around NYC. The f indings indicate that a mobile app platform, Waze, enables not only individuals to act in their own interests (their desires to enjoy engaging in sharing and exchanging information with others and their need for real - time updates to avoid traffic jams) but also their actions, driven by self - interest, to eventually contribute to the social benefits of citizens in NYC (the reduced duration of traffic jams and the improved mobilities in driving around the urban areas). In this research , the social - welfare valu e is explained by the desirable traffic congestion around NYC in which the congested traffic causes a large amount of social and economic costs ( Cookson 2018 ; Walker 2018a ; Walker 2018b ) . Based on the empir contribution mitigates traffic congestion by generating and sharing traffic content through their active or passive engagement in the mobile virtual community. This research theorizes the effectiveness of user contribution in ens uring the social - welfare value not only for the mobile virtual community but also for urban drivers and their driving quality. This is also consistent with the concepts of c ognitive a bsorption ( Agarwal and Karahanna 2000 ; Agarwal et al. 1997 ) and i nformation p rocessing ( Eppler and Mengis 2004 ) . That is, the unintentional contribution of Waze users, who focus on their own interests by using an enjoyable mobile app , relieves the serious traffic congestion in NYC by generating real - t ime traffic content and sharing it with other 94 users, thus contributing to the whole driver community of NYC citizens . As a piece of citizen - data science, this research empirically contributes by adopting the spatial panel data models robustly. This researc h use s the SAR model for the main results because its estimation is based on the basic specification of spatial panel estimations among multiple estimators . In terms of the effects of spatial spillovers , this research focuses on the results from the SDM , w hich is extended from the SAR estimator by including spatially weighted independent variables as additional regressors. T he inclusion of such spatially weighted independent variables provides additional , important findings on how the effect s of user c ontribution on traffic congestion duration in a focal zone influence the duration s of traffic congestion in its adjacent zones: the effects of spatial spillovers . Consistent results of SAC and SEM estimators are also provided for robust empirical findings even with spatial autoregressive errors and spatial autocorrelation in the error term. In addition, the empirical findings from the dynamic spatial panel data models (DSARs and DSDMs) explain both the spatial and panel dynamics in terms of the effectivenes s of user contribution in relieving the urban traffic congestion. There are hidden or unobservable effects of user contribution which are not able to be captured by static specifications. The dynamic models specify both time - lagged and a space - lagged effec ts and , thus revealing the mixed effects of short - and long - term and local and periphery spatial panel effects . Implications for Policy and Practice First, the empirical findings support that user contribution mitigat es traffic jams by i mproving the mobilities of urban transportation. The effectiveness of user contribution, crowdsourced by ordinary citizens, may justify the effectiveness of the on - going collaborative projects of municipalities and high technology compan ies in eliminating traffic and support the 95 feasibility of civic navigation software as a system of traffic record . Such projects would enable urban drivers as active users to voluntarily contribute by generating content under supervision of the municipalit y. For example, Waz e , Esri ( a leading GIS software provider), and the municipalit y launched Connected Citizen s Program . They collaborate to allow drivers, i.e., ordinary citizens, in Johns Creek , GA, to report potholes and to enable the city municipalit y to patch their repor ts ( Hall - Geisler 2016 ; McCorvey 2018 ) . Se cond, the effectiveness of user contribution in reducing the duration of traffic congestion differentiates across the types of user contribution , i.e., primary information (via alerts), follow - up feedback (via comments), and collective confirmation (via th umbs - up). For example, while primary information and follow - up feedback are effective in reducing the duration of traffic jams individually, collective confirmation alone is not significantly effective. This may be explained by the novelty and newness of u ser - crowdsourced content. Alerts and comments generally provide the initial information and its real - time updates whereas thumbs - up verify the alerts and comments are reliable. That is, alerts and comments are generated when the content is fresh enough to provide new information. However, thumbs - up are typically added on alerts when Waze users recognize the alerts were useful in the past. When thumbs - up verify alerts, the alert content is not refresh ed anymore. This is enhanced further due to the rapidly ch anging traffic conditions around the urban areas. The value of real - time information decays without updates over time. Confirmation on the past may not be useful in making decisions for driving at the present and in the future. Such timing of user - crowdsou rced content may explain the different findings between the effectiveness of primary information/follow - up feedback and collective confirmation. This implies that user engagement in generating real - time content needs to encourage further promptness to impr ove the effectiveness of user contribution. The Waze app 96 platform may need to design and adopt an effective and prompt reward system in order to further motivate Waze users to engage in generating real - time content instantly. In addition, if the Waze app p latform provides more features, which are much easier for a user to engage in user contribution (e.g., voice recognition feature s), it would be expected that more user contri bu tion will generate more benefits to both Waze users and all urban drivers. Third , the quadratic relationship of primary information and traffic jam duration can be explained by a paradoxical effect of social information overload. In general, social media platforms for user - crowdsourced content provide too much information to be adopte d and utilized by a single user. The social information overload also causes redundant information, thus reducing the value of user - crowdsourced content. However, information overload operates in a different way when Waze users process primary information via alerts. It is difficult for a Waze user, a driver, to pay attention to all individual alerts. Instead, she/he recognizes a group of alerts around a location on the Waze navigation map. Waze users are more likely to cope with traffic uncertainties by pa ying broad attention to multiple alerts as a group of information rather than concentrating on individual alerts. This is because they cannot afford to drive around the urban areas and adopt all the information and its details simultaneously. Such a parado xical effect of social information overload may explain the quadratic relationship: as the number of alerts increases, their magnitude of effectiveness further increases in reducing traffic congestion duration. Drawing on the paradoxical effect, I would su ggest that the Waze app platform needs to visualize user - crowdsourced content on the small screen of mobile devices in a zoomed - out way to maximize effectiveness of user contribution. As shown in Figure 21 (i), the Waze app platform currently shows user - cr owdsourced content on the zoomed - in map. As a result, a user is more likely focus on a few local events with less individual alerts and miss information on the 97 other events which do not show up within the current zoomed - in map. When she/he approach es the t raffic events, a pop - up also shows up the detailed information on a singular traffic event. Alternatively, as described in Figure 21 (ii), if the Waze app platform displays more alerts on multiple events simultaneously on a zoomed - out map, a user is expose d to broad information at a glance . In addition, the pop - up needs to show detailed and sequential information on the multiple traffic events. These changes may help users better understand a singular event and its related traffic events, as well as acquire real - time information on the relationships and the sequence of - making for better outcomes with sequential information from a variety of content sources ( Moore and Whinston 1986 ) . As a result, the effectiveness of user contribution may reduce the duration of traffic congestion Figure 21 . Screenshots of Waze app platform design 98 further. Fourth, the significant interactions among the different user contribution types suggest that synergies exist and are effective for Waze use heterogeneous content. The empirical findings explain which pattern s of user contribution, i.e., the combination of different user contribution types , are more (or less) effective in mitigating traffic congestion than the othe r patterns . For example, when a user observes more primary information from alerts with more follow - up feedback from comments, their mutual effectiveness increases compared to situations with more primary information and less follow - up feedback, with less primary information and more follow - up feedback, and with less primary information with less follow - up feedback. Such an interpretation can be applied in the same way to the synergies between primary information and collective confirmation and between foll ow - up feedback and collective confirmation. The different types of user contribution complement information gain and the interplays augment their individual effects. I would therefore suggest that Waze designers need to consider mul tiple customized reward systems not only for encouraging the active contribution to primary information but also for encouraging users to engage in the passive contribution to follow - up feedback and collective confirmation. Finally, the empirical findings also show that t he effectiveness of user contribution differs across hours of a day . The quadratic relationship of primary information and traffic jam duration flattens and the interactions among the different types of user contribution weaken during rush and business hours of weekdays. That is, the more drivers during the peak hours , the less benefit per individual from additional information. The effectiveness of user contribution differentiated o gain new information in addition to their own driving routine around urban areas. Many drivers are less motivated to navigate new 99 routes and acquire new traffic information from other sources for the busy hours. This may be because the heavy traffic and frequent congestion always affect and disrupt their driving for rush hours and business hours. There is no additional space for urban drivers to attempt new route s in such a big city as NYC . In this sense, the findings can be discussed in terms of the NYC policy plan for congestion pricing. A traffic a dvisory p anel 26 recently proposed a plan to the state legislature for congestion pricing in the busiest traffic parts of Manhattan in order to ease the traffic in Manhattan ( Fix NYC A dvisory Panel 2018 ) . I would suggest that the congestion pricing policy will appropriately improve the effectiveness of user contribution, crowdsourced by Waze users, during the busy hours in NYC. The numbers of drivers and their vehicles may reduce due to the additional costs of driving around the jammed areas during the peak hours. The reduced traffic congestion may revive the effectiveness of user contribution even during the NYC rush and business hours. Limitations and Future Directions In NYC, traff ic conditions and mobile app users are exposed to the huge endogeneity due to their dynamics around the urban transportation system over its complex network of roads. This may cause endogeneity concerns on causalities of the empirical findings. However, I include a variety of control variables, capturing the multifaceted dynamics around NYC, available from open - data sources to mitigate such endogeneity - related concerns. In addition, this research employs the spatial panel models ( controlling for both spatia l and temporal 26 Fix NYC Advisory Panel was appointed by Andrew M. Cuomo, Governor of the State of New York, in October, 2017 ( Fix NYC Advisory Panel 2018 ) . Members of the traff ic advisory panel include community representatives, government officials, and business leaders, such as former city and state politicians, the chairman of the Regional Plan Association, and the mayor of Yonkers, NY. In January of 2018, their congestion pr icing proposal was suggested to improve the traffic congestion in Manhattan. 100 autocorrelations ) as well as their dynamic models ( controlling for the mixed spillovers across space over time ) to tease out alternative explanations on the causation of interest in this research . Such empirical benefits from the big - data pr ocessing techniques and advanced spatial panel analytics may be enough to mitigate any potential endogeneity issues. The spatial unit of analysis is NTA in this research . Such zones over NYC are demarcated by artificial boundaries designed for civic admini strative purposes. This may cause concerns on the potential risks of empirical sensitivity across different zone levels, e.g., blocks, census tracks, counties, cities, and others. However, the scope of this research focus es on the context of user engagemen t and urban transportation over a whole city. The research questions are about the causal relationships of user contribution, crowdsourced by collective Waze users, and its effects on traffic congestion, locally and temporally influenced by user contributi on. Based on the research scope and objectives, it would be true that NTA is the most appropriate unit of analysis to analyze the effects of user contribution on traffic congestion duration in NYC by providing a balanced view to the whole city across its b oroughs ; a block and census track are too narrow units whereas a county and city are too wide units to measure the contributing activities of collective mobile app users and local conditions of traffic congestion everywhere around the city. In my view, the re are promising research ideas on the usage of mobile technology in urban areas and its value for a society. For example, the findings from this research might be further extended by studying potential competing or moderating effects among multiple inform ation sources given to drivers around urban areas. This research focuses on the effectiveness of user - crowdsourced information on the real - time traffic from a singular mobile app. However, when driving in NYC, mobile - device users are exposed to a variety o f real - time information sources which include algorithm - generated information from other mobile navigation apps, traffic 101 information services from local radio stations, signals of p ublic t ransportation g uide s ystem s, their own traffic physical observations ( through the windshield and side windows ) , and so on around urban areas. Given such multiple information sources, future research might explain their potential interactions in terms of their mutual effects on traffic congestion. In addition, the effective ness of user contribution can explain not only the duration of traffic congestion but also many alternative outcomes, such as the reduced numbers of automobile accidents or driving offences, which might show the different aspect of social - welfare value of urban transportation. Future studies may extend my empirical findings by exploring the effectiveness of user contribution in leading to other social and economic benefits to citizens. CONCLUSION In a context of urban transportation, this research theorize s the technology - driven value of user - crowdsourced content in supporting social welfare for the whole citizens in urban areas. The empirical findings support that the self - interest of mobile app users eventually contribute to the social welfare. The welfar e contribution is enabled by ordinary citizens and their voluntary usage of a mobile app. In addition, this research utilizes large - scale data on urban areas and mobile phone users with spatial and panel analytics technique s. The advanced empirics might be an exemplar of big - data analytics for citizen - data science. 102 APPENDICES 103 APPENDIX A. Correlation Analysis and Distribution of Boroughs and Weather Conditions Table A 1 . Pairwise correlations between study variables Variable 1 2 3 4 5 6 7 8 1. : 1.0000 2. ln ( ) : ln ( ) 0.6648 1.0000 3. : 0.8461 0.5442 1.0000 4. : 0.5718 0.3842 0.5599 1.0000 5. 0.1023 0.2110 0.4020 0.1293 1.0000 6. 0.0018 0.0136 0.0521 0.4048 0.1633 1.0000 7. ln ( ) 0.3323 0.4995 0.2438 0.1773 0.0716 0.0049 1.0000 8. ln ( ) 0.0611 0 .1631 0.0779 0.0627 0.0911 0.0847 0.2838 1.0000 9. ln ( ) 0.0012 0.0573 0.0164 0.0142 0.0280 0.0147 0.3937 0.0564 10. ln ( ) 0.3819 0.1168 0.3478 0.2721 0.0330 0.0513 0.2469 0.1265 11. ln ( ) 0.05 09 0.1455 0.0296 0.0244 0.0403 0.0667 0.5023 0.1163 12. ln ( ) 0.3011 0.4281 0.1847 0.1219 0.0145 0.0841 0.5451 0.0176 13. ln ( ) 0.0760 0.1164 0.0434 0.0324 0.0085 0.0050 0.1101 0.0144 14. ln ( ) 0.0211 0.0385 0.0149 0.0090 0.0055 0.0014 0.0343 0.0054 15. 0.0354 0.1281 0.0319 0.0268 0.0487 0.0753 0.4733 0.1349 16. ln ( ) 0.2292 0.2517 0.1454 0.1123 0.0144 0.0041 0.3224 0.0090 Variable 9 10 11 12 13 14 15 16 9. ln ( ) 1.0000 10. ln ( ) 0.4450 1.0000 11. ln ( ) 0.5706 0.7298 1.0000 12. ln ( ) 0.2968 0.3328 0.5122 1.0000 13. ln ( ) 0.0688 0.0759 0.1231 0.1239 1.0000 14. ln ( ) 0.0227 0.0285 0.0367 0.0394 0.4068 1.0000 15. 0.5552 0.7207 0.8904 0.5246 0.1091 0.0369 1.0000 16. ln ( ) 0.2402 0.0403 0.4272 0.1806 0.0757 0.0086 0.2183 1.0000 Table A 2 . Distributions of observations across boroughs and weather conditions Borough Obs. % Weather condition Obs. % Bronx 87,024 19.47 Clear 179,857 40.25 Brooklyn 117,600 26.32 Cloudy 229,232 51.30 Manhattan 65,856 14.74 Foggy 3,869 0.87 Queens 134,064 30.00 Rainy 21,473 4.81 Staten Island 42,336 9.47 Snowy 12,449 2.79 Total 446,880 100.00 Total 446,880 100.00 104 APPENDIX B. Technical Details for Models Zero - Inflated Count Data Models. Conditio nal on the observation not being zero - inflated, it is assume d that the number of initial contribution via alert postings follows a negative binomial distribution. Let be a latent variable with negative binomial distribution conditional on a set of covariates . With the standard negative binomial model setting, the parameter is an exponential function of a vector of covariates: . This research specif ies the as a linear combination of observed variables including ln ( ) . Hence, the probability mass function of , conditional on and the dispersion parameter ca n be written as follows: . ( 9 ) In the vector of covariates in s pecification , i.e., E quation ( 9 ), ln ( ) is the main independent variable that captures the positive relationship betw een virtual crowdedness and initial contribution. In addition, I include a rich set of covariates that control for confounding factors, such as Waze weather condition, which may hav e effects on initial contribution. Specifically, Z 1, it is a set of time - variant road condition control variables ( ln ( ) , ln ( ) , ln ( ) , ln ( ) , and weather condition indicators) whereas Z 2, i is a set of time - invariant control variables ( ln ( ) , ln ( ) , ln ( ) , 105 , and ln ( ) ). Finally, Date and Hour are dummy indicators to control for time effects across the hour - by - hour 2,352 - time points for the fourteen weeks. Notice that initial contribution, the dependent variable , i.e. , , equal s to zero if the observation is zero - inflated, and otherwise , as follows: T he zero - inflation process is modeled usin g standard logit model with covariates . Therefore, this research define s the probability of being zero - inflated as and specif ies as a linear combination of observed variables , as follows : . ( 10 ) E xcept for ln ( ), a ll the covariates are included to the zero - inflation specification, i.e., E quation ( 10 ) , in order to model the propensity of havi ng excess zero values for initial contribution. Formally, the zero - inflated count data model can be estimated based on a two - stage conditional approach or full information maximum likelihood ( Vuong 1989 ) . T he full - observa tion log - likelihood function is derive d as follows : . 106 Notice that the first component is the log - likelihood of observing zeros, the second component is the log - likelihood of observing non - zero values. To check the robustness of the proposed model, I further utilize ZIP model. The o nly difference is the probability mass function of conditional on : I further modify the specification , i.e., E quation ( 9 ) , by including the squared term of in the zero - inflated count data model to test the decrease in magnitude of the relationship between virtual crowdedness and initial contribution (H ypothesis 1 B ), as follows: . ( 11 ) Specification of Heckman Selection Model s . The Heckman selection model s are s pecified as follows . Main e quation: , ( 12 ) Selection e quation: ( 13 ) 107 APPENDIX C . Propensity - Score Matching Tests with Ten Groups Note. The ten groups are eve nly divided along the distribution of ln ( ) where ln ( ) > 0: min. = 0.0343; max. = 5.2307; bin width = 0.5196. Figure A 1 . Distributions of virtual crowdedness with ten groups for p ropensity - score matching Table A 3 . Results of propensity - score matching tests with ten groups for initial contribution Treatment groups Obs. ATE AI robust standard error z - value Lower bound of 95% CI Upper bound of 95% CI Group 2 vs. Group 1 160,212 0.3251*** 0.0091 35.63 0.3072 0.3430 Group 3 vs. Group 2 206,264 0.3263*** 0.0043 75.10 0.3178 0.3349 Group 4 vs. Group 3 185,969 0.2483*** 0.0051 48.26 0.2382 0.2584 Group 5 vs. Group 4 116,799 0.1927*** 0.0070 27.68 0.1790 0.206 3 Group 6 vs. Group 5 59,965 0.1023*** 0.0097 10.56 0.0833 0.1213 Group 7 vs. Group 6 25,359 0.1150*** 0.0155 7.41 0.0845 0.1454 Group 8 vs. Group 7 7,904 0.0867*** 0.0273 3.18 0.0333 0.1402 Group 9 vs. Group 8 1,828 0.1476* 0.0879 1.68 0.3199 Group 10 vs. Group 9 N/A N/A N/A N/A N/A N/A Note. Dependent variable: ln ( ) = ln ( ); treatments: the five groups evenly divided along by the distribution of ln ( ) where ln ( ) > 0. The requested number of matches per observation is at least 1. In estimating the ATE s in population, the propensity - score matching tests include all the control variables as well as the following dummy indicators: weather condition indictor (clear = 1; oth erwise = 0), borough indicator (Manhattan = 1; the other boroughs = 0), date indicator (weekday = 1; weekend and holiday = 0), and hour - of - day indicator (daytime from sunrise to sunset = 1; night time from sunset to sunrise = 0; data source: timeanddate.co m). ATE: average treatment effect; AI: Abadie - Imbens; CI: confidence interval of treatment effect. * p < 0.10; ** p < 0.05; *** p < 0.01 . 108 APPENDIX D . Robustness Checks Table A 4 . Results of panel negative binomial models for initi al contribution Variables Dependent variable : = (1) Pooled negative binomial a (2) Pooled negative binomial a (3) Fixed - effects negative binomial b (4) Fixed - effects negative binomial b (5) Random - effects negative binomial b (6) Random - effects negative binomial b { ln ( )} 2 0.0865*** 0.0331*** 0.0467*** (0.0020) (0.0016) (0.0017) ln ( ) 0.5597*** 0.8889*** 0.2060*** 0.3292*** 0.2392*** 0.4154*** (0.0030) (0.0076) (0.0025) (0.0066) (0.0027) (0.0070) Time - variant control variables Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes No No Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes No No Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Y es Yes Yes Yes Yes Obs. 444,528 444,528 446,880 446,880 444,528 444,528 Log likelihood Number of NTAs 190 190 189 189 Note. a Robust standard errors in parentheses. b Standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). * p < 0.10; ** p < 0.05; *** p < 0.01 . 109 Table A 5 . Results of panel Poisson models for initial contribution Variables Dependent vari able : = (1) Pooled Poisson (2) Pooled Poisson (3) Fixed - effects Poisson (4) Fixed - effects Poisson (5) Random - effects Poisson (6) Random - effects Poisson { ln ( )} 2 ( 0.0023) (0.0106) (0.0155) ln ( ) 0.5655*** 1.0116*** 0.2250*** 0.3762*** 0.2251*** 0.3766*** (0.0036) (0.0094) (0.0093) (0.0470) (0.0365) (0.0938) Time - variant control variables Yes Yes Yes Yes Yes Yes Time - invariant control v ariables Yes Yes No No Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes No No Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Yes Yes Obs. 444,528 444,528 446,880 446,880 444 ,528 444,528 Log likelihood Number of NTAs 190 190 189 189 Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are o mitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). * p < 0.10; ** p < 0.05; *** p < 0.01 . 110 Table A 6 . Winter vs. summer: Results of zero - inflated count data models for ini tial contribution Panel C: Winter observations Variables Dependent variable: = (1) ZINB (2) ZINB (3) ZIP (4) ZIP Main e quation Zero i nflation Main e quation Zero i nflation Main e quation Zero i nflation Main e quation Zero i nfl ation { ln ( )} 2 0.0100*** 0.0149*** (0.0030) (0.0030) ln ( ) 0.5281*** 0.5652*** 0.3917*** 0.4559*** (0.0051) (0.0131) (0.0045) (0.0130) Time - variant control variables Yes Yes Yes Y es Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Date indicators Yes Yes Yes Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Obs. 222,264 222,264 222,264 222,264 Non - zero obs. 148,152 148,152 148,152 148,152 Zero obs. 74,112 74,112 74,112 74,112 Log likelihood 468,598.03 468,591.86 592,795.49 592,710.28 Panel D: Summer observations Variables Dependent variable: = (1) ZINB (2) ZINB (3) ZIP (4) ZIP Main e quation Zero i nflation Main e quation Zero i nflation Main e quation Zero i n flation Main e quation Zero i nflation { ln ( )} 2 0.1292*** 0.1418*** (0.0053) (0.0051) ln ( ) 0.8482*** 1.2101*** 0.6757*** 1.1359*** (0.0060) (0.0170) (0.0064) (0.0163) Time - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Date indicators Yes Yes Ye s Yes Yes Yes Yes Yes Hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Obs. 222,264 222,264 222,264 222,264 Non - zero obs. 155,317 155,317 155,317 155,317 Zero obs. 66,947 66,947 66,947 66,947 Log likelihood Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy b orough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables i nclude ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). ZINB: zero - inflated negative binomial; ZIP: zero - inflated Poisson. * p < 0.10; ** p < 0.05; *** p < 0.01 . 111 Table A 7 . Winter vs. summer: Results of Heckman selection models for feedback and confirmative contribution Panel C: Winter observations Variables Dependent variable: = Dependent variable: = (1) Simultane ous Heckman (2) Two - step Heckman (3) Simultaneous Heckman (4) Two - step Heckman Main e quation Selection e quation Main e quation Selection e quation Main e quation Selection e quation Main e quation Selection e quation ln ( ) 0.0515*** 0. 5070*** 0.0508*** 0.5070*** 0.1139*** 0.5070*** 0.0983*** 0.5070*** (0.0018) (0.0060) (0.0021) (0.0060) (0.0070) (0.0060) (0.0113) (0.0060) Time - variant control variables Y es Y es Y es Y es Y es Y es Y es Y es Time - invariant control variables N o Y es N o Y es N o Y es N o Y es W eather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Y es Y es Y es Y es Y es Y es Y es Y es Date indicators Y es Y es Y es Y es Y es Y es Y es Y es Hour - of - day indicators Y es Y es Y es Y es Y es Y es Y es Y es Obs. 222,264 222,264 222, 264 222,264 Censored obs. 74,112 74,112 74,112 74,112 Uncensored obs. 148,152 148,152 148,152 148,152 Log - likelihood 184,228.15 433,330.33 Panel D: Summer observations Variables Dependent variable: = Dependent variable: = (1) Simultaneous Heckman (2) Two - step Heckman (3) Simultaneous Heckman (4) Two - step Heckman Main e quation Selection e quation Main e quation Selection e quation Main e quation Selection e quation Main e quation Selection e quation ln ( ) 0.1034*** 0.7801*** 0.0854*** 0.7791*** 0.1142*** 0.7791*** 0.0774*** 0.7791*** (0.0028) (0.0078) (0.0037) (0.0079) (0.0100) (0.0078) (0.0153) (0.007 9) Time - variant control variables Y es Y es Y es Y es Y es Y es Y es Y es Time - invariant control variables N o Y es N o Y es N o Y es N o Y es W eather condition indicators Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Y es Y es Y es Y es Y es Y es Y es Y es Date indicat ors Y es Y es Y es Y es Y es Y es Y es Y es Hour - of - day indicators Y es Y es Y es Y es Y es Y es Y es Y es Obs. 222,264 222,264 222,264 222,264 Censored obs. 66,947 66,947 66,947 66,947 Uncensored obs. 155,317 155,317 155,317 155,317 Log - likelihood Note. Robust standard errors in parentheses. Estimates for constant, control variables, and dummy borough, date, and hour - of - day indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and weather condition indicators. Time - invariant control variables include ln ( ), ln ( ), ln ( ), , and ln ( ). * p < 0.10; ** p < 0.05; *** p < 0.01 . 112 APPENDIX ontiguity Note. The above map mainly shows a part of Manhattan. Based on the , the focal area, i.e., MN21, has four s patial neighbors , i.e., MN13, MN20, MN22, and MN50, which share their common boundar ies along borders only . On the other hand, the queen adds two more spatial neighbors , i.e., MN17 and MN23, which are adjacent to the focal area, i.e., MN2 1, at points . Figure A 2 . Context - 27 27 The map backdrop cartography is sourced from stamen.com using OpenStreetMap data. 113 Table A 8 . Statistics of spatial weight matrix (i) Summary (ii) Tabulation of links Matrix features Description Number of links Obs . Dimensions 188 x 188 1 6 Values 2 23 Min . 0 3 45 Min . > 0 0.125 4 52 Mean 0.005 5 37 Max . 1 6 15 Links 7 6 Total 744 8 4 Mi n . 1 Sum 188 Mean 3.957 Max . 8 Figure A 3 . Connectivity of spatial weight matrix 114 APPENDIX F. Spatial Nature of Traffic Congesti on Duration and User Contribution Table A 9 . Diagnostics of spatial autocorrelations of study variables with spatial weight matrix Variables Moran s I E( I ) S.D. ( I ) z - score ln ( ) 0.611 *** 0.005 0.051 12.054 0.339 *** 0.005 0.051 6.731 0.307 *** 0.005 0.051 6.112 0.312 *** 0.005 0.051 6.194 Variables Geary s c E( c ) S.D. ( c ) z - score ln ( ) 0.284 *** 1.000 0. 065 10.994 0.635 *** 1.000 0.064 5.702 0.653 *** 1.000 0.064 5.453 0.654 *** 1.000 0.060 5.753 Variables Getis and Ord s G E( G ) S.D. ( G ) z - score ln ( ) 0.023 *** 0.021 0.000 6.465 0.023 *** 0.021 0.001 3.376 0.024 *** 0.021 0.001 2.560 0.024 *** 0.021 0.001 2.651 Note. Two - tailed test s. * p < 0.10; ** p < 0.05; *** p < 0.01 . 115 Figure A 4 . Visualized spatial autocorrelations of study variables with spatial weight matrix 116 APPENDIX G. Matching Method in Q uasi - N atural E xperimental S etting This research employs a matching method, i.e. , propensity - score matching (PSM) tests , in a quasi - natural experimental setting to remove any potential e ndogeneity issues further. The PSM estimator approximates the counterfactual outcome for each observation based on an average of the outcomes of simil ar observations with other treatment level s ( Abadie and Imbens 2006 ) . The similarit ies between observations can be measured by estimat ing the probabilities of different - level treatments , i.e., propensity scores , derived from a binary regression model. A verage T reatment E ffect (ATE) refers to the difference between the observed and potential outcomes for each observation. In doing so, I place all the observations into multiple groups based on the degree of user contribution and compare their traffic congestion duration between the groups. To set a quasi - natural experiment , all the observations are assigned into seven groups as treatments along the number of alerts, posted in NTA i at Time t , as the degree of user contribution, as follows: Group 1 includes the observations with no alerts ( = 0 ); Group 2 with one or two alerts ( ); Group 3 with a few more alerts (3 5); Group 4 with several alerts (6 10); Group 5 with many alerts (11 20); Group 6 with m uch more alerts (21 35); and Group 7 with more - than - 35 alerts. The comparisons of traffic congestion duration are conducted with the first six groups only. Group 7 was excluded as outliers due to the lack of observations for a comparis on: the number of observations assigned to Group 7 is 4,019 (top 0.91%) out of the total observations. As summarized in Table A10, results of the PSM tests also support that the negative relationship of ln ( ) is consist ent ly significant along by the overall 117 distribution of . That is , ATE between Group 1 and Group 2 is 0.0228 which is negative and significant at p < 0.01 ; ATE between Group 2 and Group 3 is 0.0212 at p < 0.01 ; ATE between Group 3 and Grou p 4 is 0.0314 at p < 0.01 ; ATE between Group 4 and Group 5 is 0.0357 at p < 0.01 ; ATE between Group 5 and Group 6 is 0.0317 at p < 0.01 . T he PSM results suggest that traffic congestion duration decreases as user contribution to primary information incre ases. The PSM results show the consistent effectiveness of primary information in reducing traffic congestion duration over the whole range of . In addition to the robust findings from spatial panel estimators, the PSM results may further eliminate any potentials of endogeneity regarding the casual effects of user contribution, specifically primary information (via alerts), on the decreased duration of traffic jams. 118 Table A 10 . Results of propensity - score matching tests Note. Dependent variable: = ln ( ) ; treatments: the six groups divided along by the distribution of . The requested number of matches per observation is at least 1. In estimating the ATE s in population, the pr opensity - score matching tests include all the controls as well as the following dummy indicators: weather condition indictor s , borough indicator (Manhattan = 1; the other boroughs = 0), date indicator (weekday = 1; weekend and holiday = 0), and hour - of - day indicator (daytime from sunrise to sunset = 1; night time from sunset to sunrise = 0; data source: timeanddate.com) . ATE: average treatment effect; AI: Abadie - Imbens; CI: confidence interval of treatment effect . * p < 0.10; ** p < 0.05; *** p < 0.01 . 119 A PPENDIX H. Sensitivity C heck s : Summary of Empirical Results from Spatial Panel Data Models Based on r C ase C ontiguity Table A 11 . Effects of user contribution on traffic congestion duration: Results of spatial panel data model s with spatial weight matrix Variables Dependent variable: = ln ( ) (1) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0504*** 0.0423*** 0.0090 0.0476*** 0.0474*** (0.0049) (0.0049) (0.006 6) (0.0052) (0.0051) 0.0194*** 0.0264*** 0.0187*** 0.0214*** 0.0214*** (0.0050) (0.0046) (0.0072) (0.0051) (0.0051) 0.0027 0.0059** 0.0071 0.0042 0.0042 (0.0033) (0.0030) (0.0052) (0.0033) (0.0033 ) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2147*** 0.2597*** 0.0106 (0.0063) (0.0071) (0.02 22) 0.2669*** 0.2764*** (0.0078) (0.0232) Obs. 442,176 442,176 442,176 442,176 R 2 0.5273 0.5304 0.5293 0.5290 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates f or control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 120 Table A 12 . Effects of user contribution on traff ic congestion duration for rush and business hours : Results of spatial panel data models with spatial weight matrix contiguity Variables Dependent variable: = ln ( ) (1) SAR (1) SDM (1) SEM (1) SAC Main Main Main Main 0.0291*** 0.0258*** 0.0011 0.0277*** 0.0273*** (0.0044) (0.0049) (0.0075) (0.0047) (0.0047) 0.0093** 0.0115*** 0.0122** 0.0098** 0.0098** (0.0039) (0.0042) (0.0062) (0.0041) (0.0041 ) 0.0015 0.0019 0.0043 0.0016 0.0016 (0.0027) (0.0030) (0.0052) (0.0029) (0.0029) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2276*** 0.2583*** 0.0310 (0.0083) (0.0074) (0.0384) 0.2593*** 0.2866*** (0.0075) (0.0338) Obs. 153,408 153,408 153,408 153,408 R 2 0.3390 0.3342 0.3378 0.3369 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 121 Table A 13 . Dynamic effects of user contribution on traffic congestion duration: Results of dynamic spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) DSARs DSDMs (1) (2) (3) (4) (5) (6) Main Main Main Main Main Main 0.0200*** 0.0181*** 0.0207*** 0.0179*** (0.0023) (0.0023) (0.0022) (0.0023) 0.0202*** 0.011 8*** 0.0255*** 0.0172*** (0.0039) (0.0039) (0.0038) (0.0038) 0.0505*** 0.0503*** 0.0504*** 0.0424*** 0.0087 0.0421*** 0.0089 0.0423*** 0.0086 (0.0049) (0.0049) (0.0049) (0.0049) (0.0066) (0.0049) (0.0066) (0.0049 ) (0.0066) 0.0197*** 0.0195*** 0.0197*** 0.0267*** 0.0187*** 0.0264*** 0.0184** 0.0266*** 0.0184** (0.0049) (0.0049) (0.0049) (0.0045) (0.0072) (0.0046) (0.0072) (0.0045) (0.0072) 0.0028 0.0028 0.0 028 0.0059** 0.0069 0.0059** 0.0069 0.0059** 0.0069 (0.0033) (0.0033) (0.0033) (0.0030) (0.0052) (0.0030) (0.0052) (0.0029) (0.0052) Spatial fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Yes 0.2136*** 0.2133*** 0.2129*** 0.2589*** 0.2586*** 0.2584*** (0.0063) (0.0062) (0.0062) (0.0071) (0.0071) (0.0071) Obs. 441,988 441,988 441,988 441,988 441,988 441,988 R 2 0.5288 0.5286 0.5293 0.5330 0.5339 0.5348 Number of NTAs 188 188 188 18 8 188 188 Panel length 2,352 2,352 2,352 2,352 2,352 2,352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted fro m the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). DSAR: dynamic spatial autoregressive model; DSDM: dynamic spatial Durbin model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 122 Table A 14 . Quadratic relationship of primary information and traffic congestion duration: Results of spatial panel data models with spatial weight matrix contiguity Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0503*** 0.0550*** 0.0122* 0.0550*** 0.0551*** (0.0044) (0.0042) (0.0072) (0.0045) (0.0045) 0.0593*** 0.0722*** 0.0220 0.0695*** 0. 0699*** (0.0093) (0.0085) (0.0154) (0.0094) (0.0095) 0.0145*** 0.0090** 0.0138** 0.0150*** 0.0151*** (0.0042) (0.0038) (0.0054) (0.0042) (0.0042) 0.0032 0.0003 0.0061 0.0022 0.0022 (0.0028) (0.0026) (0. 0046) (0.0028) (0.0028) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2129*** 0.2590*** 0.0127 (0 .0062) (0.0068) (0.0211) 0.2662*** 0.2775*** (0.0075) (0.0218) Obs. 442,176 442,176 442,176 442,176 R 2 0.5293 0.5326 0.5319 0.5317 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in pa rentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 123 Table A 15 . Quadratic relat ionship of primary information and traffic congestion duration for rush and business hours : Results of spatial panel data models with spatial weight matrix based Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0143*** 0.0215*** 0.0177*** 0.0189*** 0.0194*** (0.0031) (0.0035) (0.0050) (0.0033) (0.0034) 0.0086 0.0272*** 0.0428*** 0.0210** 0.0224** (0.0 088) (0.0093) (0.0149) (0.0092) (0.0093) 0.0024 0.0014 0.0039 0.0008 0.0005 (0.0038) (0.0040) (0.0062) (0.0040) (0.0040) 0.0028 0.0037 0.0060 0.0032 0.0032 (0.0027) (0.0030) (0.0051) (0.0029) (0.0029 ) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2274*** 0.2587*** 0.0349 (0.0083) (0.0074) (0.03 85) 0.2598*** 0.2904*** (0.0074) (0.0337) Obs. 153,408 153,408 153,408 153,408 R 2 0.3401 0.3357 0.3393 0.3383 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for contr ol variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 124 Table A 16 . Interaction between primary information and f ollow - up feedback: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0372*** 0.0293*** 0.0085 0.0339*** 0.0336*** (0.0048) (0.0049) (0.0062) (0.0050) (0.0050) 0.0941*** 0.1014*** 0.0158 0.1035*** 0.1038*** (0.0107) (0.0096) (0.0182) (0.0105) (0.0105) 0.0012 0.0045 0.0073 0.0026 0.0027 (0.0031) (0.0029) (0.0049) (0.0031) (0.0031) 0.0489*** 0.0561*** 0.0169* 0.0543*** 0.0545*** (0.0052) (0.0048) (0.0091) (0.0052) (0.0052) Spatial fixed effects Yes Yes Ye s Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2139*** 0.2597*** 0.0132 (0.0063) (0.0070) (0.0217) 0.2670*** 0.2788** * (0.0076) (0.0225) Obs. 442,176 442,176 442,176 442,176 R 2 0.5284 0.5317 0.5308 0.5305 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weat her condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0. 10; ** p < 0.05; *** p < 0.01 . 125 Table A 17 . Interaction between primary information and follow - up feedback for rush and business hours : Results of spatial panel data models with spatial weight matrix based on iguity Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0227*** 0.0176*** 0.0097 0.0195*** 0.0190*** (0.0046) (0.0050) (0.0077) (0.0049) (0.0048) 0.0278*** 0.0475*** 0.0530*** 0.0405*** 0.0419*** (0.0083) (0.0091) (0.0157) (0.0088) (0.0089) 0.0019 0.0024 0.0049 0.0021 0.0021 (0.0027) (0.0030) (0.0051) (0.0029) (0.0029) 0.0156*** 0.0254*** 0.0276*** 0.0213*** 0.0219*** (0.0034) (0.0039) (0.0063) (0.0037) (0.0038) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes We ather condition indictors Yes Yes Yes Yes Yes 0.2275*** 0.2587*** 0.0349 (0.0083) (0.0074) (0.0385) 0.2598*** 0.2905*** (0.0074) (0.0338) Obs. 153,408 153,408 153,408 153,408 R 2 0.3397 0.3349 0.3388 0.3378 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; * * p < 0.05; *** p < 0.01 . 126 Table A 18 . Interaction between primary information and collective confirmation: Results of spatial panel data models with spatial weight matrix Variables Dependent v ariable: = ln ( ) ( 1 ) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0240*** 0.0155*** 0.0100 0.0198*** 0.0196*** (0.0053) (0.0054) (0.0074) (0.0056) (0.0056) 0.0042 0.0 018 0.0164*** 0.0038 0.0038 (0.0039) (0.0037) (0.0053) (0.0040) (0.0040) 0.0521*** 0.0537*** 0.0042 0.0553*** 0.0554*** (0.0066) (0.0062) (0.0108) (0.0067) (0.0067) 0.0353*** 0.0 388*** 0.0084 0.0386*** 0.0387*** (0.0050) (0.0045) (0.0074) (0.0050) (0.0050) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Y es Yes Yes Yes 0.2139*** 0.2595*** 0.0123 (0.0062) (0.0069) (0.0215) 0.2668*** 0.2777*** (0.0076) (0.0224) Obs. 442,176 442,176 442,176 442,176 R 2 0.5286 0.5319 0.5310 0.5307 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 35 2 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 127 Table A 19 . Interaction between primary information and collective confirmation for rush and business hours : Results of spatial panel data models with spatial weight matrix based on Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0226*** 0.0154*** 0.0174** 0.0182*** 0.0175*** (0.0050) (0.0053) (0.0089) (0.0053) (0.0053) 0.0061 0.0059 0.00 36 0.0050 0.0048 (0.0038) (0.0041) (0.0062) (0.0040) (0.0040) 0.0142** 0.0250*** 0.0396*** 0.0208*** 0.0215*** (0.0056) (0.0062) (0.0105) (0.0060) (0.0061) 0.0071** 0.0131*** 0 .0194*** 0.0108*** 0.0111*** (0.0028) (0.0031) (0.0051) (0.0030) (0.0031) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Ye s Yes Yes 0.2276*** 0.2586*** 0.0330 (0.0083) (0.0074) (0.0385) 0.2597*** 0.2887*** (0.0074) (0.0338) Obs. 153,408 153,408 153,408 153,408 R 2 0.3395 0.3345 0.3386 0.3376 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 128 Table A 20 . Interaction between follow - up feedback and collective confirmation: Results of spatial panel data models with spatial weight matrix Variables Dependent variable: = ln ( ) ( 1 ) SAR ( 2 ) SDM ( 3 ) SEM ( 4 ) SAC Main Main Main Main 0.0544*** 0.0474*** 0.0062 0.0522*** 0.0520*** (0.0051) (0.0050) (0.0068) (0.0053) (0.0052) 0.0224*** 0.0239*** 0.0037 0.0263*** 0.0265*** (0.0065) (0.0065) (0.0102) (0.0066) (0.0067) 0.0112*** 0.0108*** 0.0003 0.0117*** 0.0117*** (0.0032) (0.0032) (0.0061) (0.0033) (0.0033) 0.0199*** 0.0245*** 0.0116** 0.0229*** 0.0230*** (0.0036) (0.0035) (0.0058) (0.0037) (0.0038) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.2146*** 0.2599*** 0.0121 (0.0063) (0.0071) (0.0221) 0.2672*** 0.2780*** (0.0078) (0.0231) Obs. 442,176 442,176 442,176 442,176 R 2 0.5277 0.5308 0.5298 0.5296 Number of NTAs 188 188 188 188 Panel length 2,352 2,352 2,352 2,352 Note. Robu st standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 129 Table A 21 . Interaction between follow - up feedback and collective confirmation for rush and business hours : Results of spatial panel data models with spatial weight matrix based on Variables Dependent variable: = ln ( ) (1) SAR (2) SDM (3) SEM (4) SAC Main Main Main Main 0.0301*** 0.0280*** 0.0026 0.0293*** 0.0290*** (0.0044) (0.0049) (0.0076) (0.0047) (0.0047) 0.0016 0.0092 0.0240** 0.0073 0.0080 (0. 0055) (0.0060) (0.0107) (0.0058) (0.0059) 0.0061* 0.0104*** 0.0194*** 0.0087** 0.0090** (0.0034) (0.0037) (0.0066) (0.0036) (0.0037) 0.0052** 0.0102*** 0.0173*** 0.0082*** 0.00 85*** (0.0022) (0.0025) (0.0043) (0.0024) (0.0025) Spatial fixed effects Yes Yes Yes Yes Yes Temporal fixed effects Yes Yes Yes Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes 0.22 76*** 0.2585*** 0.0325 (0.0083) (0.0074) (0.0386) 0.2596*** 0.2881*** (0.0075) (0.0339) Obs. 153,408 153,408 153,408 153,408 R 2 0.3392 0.3342 0.3382 0.3372 Number of NTAs 188 188 188 188 Panel length 816 816 816 816 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). SAR: spatial autoregressive model; SDM: spatial Durbin model; SEM: spatial error model; SAC: spatial autocorrelation model. * p < 0.10; ** p < 0.05; *** p < 0.01 . 130 APPENDIX I. Summary of Sensitiv ity and Robustness Checks Table A 22 . Sensitivity and robustness checks: Results of spatial autoregressive models (a) With spatial weight matrix based on q Variables Dependent variable: = ln ( ) (1) R andom effects a ( 2 ) S patial fixed effects only b ( 3 ) T emporal fixed effects only b 0.0581*** 0.0581*** 0.0701*** (0.0047) (0.0046) (0.0071) 0.0208*** 0.0209*** 0.0133** (0.0046) (0.0046) (0.0061) 0.0107*** 0.0111*** 0.0005 (0.0034) (0.0034) (0.0045) Spatial fixed effects No Yes No Temporal fixed effects No No Yes Time - variant control variables Yes Yes Yes Time - invariant control variables Yes No No Weather condition indictors Yes Yes Yes 0.2622*** 0.2623*** 0.2219*** (0.0077) (0.0077) (0.0066) Obs. 442,176 442,176 442,176 R 2 0.5324 0.5278 0.5283 Number of NTAs 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 (b) With spatial weight matrix base d on r Variables Dependent variable: = ln ( ) (1) R andom effects a ( 2 ) S patial fixed effects only b ( 3 ) T emporal fixed effects only b 0.0573*** 0.0573*** 0.0699*** (0.0046) (0.0046) (0.007 2) 0.0209*** 0.0210*** 0.0133** (0.0046) (0.0046) (0.0061) 0.0106*** 0.0109*** 0.0008 (0.0034) (0.0034) (0.0045) Spatial fixed effects No Yes No Temporal fixed effects No No Yes Time - variant contr ol variables Yes Yes Yes Time - invariant control variables Yes No No Weather condition indictors Yes Yes Yes 0.2562*** 0.2563*** 0.2160*** (0.0076) (0.0076) (0.0066) Obs. 442,176 442,176 442,176 R 2 0.5323 0.5278 0.5284 Number of NTAs 188 1 88 188 Panel length 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. a Estimates for constant, control variables, and dummy weather condition indicators are omitted from the above results. b Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 131 Table A 23 . Sensitivity and robustness checks: Results of spatial Durbin models (a) With spatial weight matrix based on q iguity Variables Dependent variable: = ln ( ) (1) Random effects a (2) Spatial fixed effects only b (3) Temporal fixed effects only b Main Main Main 0.0403*** 0.0081 0.0401*** 0.0081 0.0718*** 0.0068 (0.0050) (0.0061) (0.0049) (0.0062) (0.0072) (0.0096) 0.0301*** 0.0127* 0.0302*** 0.0129* 0.0222*** 0.0192* (0.0046) (0.0071) (0.0046) (0.0071) (0.0060) (0.0103) 0.0073** 0.0024 0.0074** 0.0 025 0.0046 0.0051 (0.0031) (0.0047) (0.0031) (0.0047) (0.0044) (0.0065) Spatial fixed effects No No Yes Yes No No Temporal fixed effects No No No No Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes No No No No Weather condition indictors Yes Yes Yes Yes Yes Yes 0.3454*** 0.3455*** 0.2695*** (0.0073) (0.0073) (0.0073) Obs. 442,176 442,176 442,176 R 2 0.5441 0.5352 0.5361 Number of NTAs 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 (b) With spatial weight matrix based on r Varia bles Dependent variable: = ln ( ) (1) Random effects a (2) Spatial fixed effects only b (3) Temporal fixed effects only b Main Main Main 0.0400*** 0.0081 0.0399*** 0.0080 0.0708*** 0.0081 (0. 0049) (0.0058) (0.0049) (0.0058) (0.0070) (0.0092) 0.0287*** 0.0077 0.0288*** 0.0079 0.0203*** 0.0133 (0.0045) (0.0068) (0.0045) (0.0068) (0.0059) (0.0096) 0.0072** 0.0023 0.0074** 0.0023 0.0044 0.0 023 (0.0031) (0.0049) (0.0031) (0.0049) (0.0044) (0.0063) Spatial fixed effects No No Yes Yes No No Temporal fixed effects No No No No Yes Yes Time - variant control variables Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes No No No No Weather condition indictors Yes Yes Yes Yes Yes Yes 0.3328*** 0.3329*** 0.2632*** (0.0074) (0.0074) (0.0072) Obs. 442,176 442,176 442,176 R 2 0.5442 0.5355 0.5362 Number of NTAs 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. a Estimates for constant, control variables, and dummy weather condition indicators are omitted from the above results. b Estimates for control variables and dummy weather condition indicators are omitted from the above results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 132 Table A 24 . Sensitivity and robustness checks: Results of spatial error models and spatial autocorrelation models (a) With spatial weight matr ix based on q Variables Dependent variable: = ln ( ) Spatial error models Spatial autocorrelation models (1) Spatial fixed effects only (2) Temporal fixed effects only (3) Spatial fixed effects only (4) Tem poral fixed effects only 0.0479*** 0.0745*** 0.0452*** 0.0737*** (0.0052) (0.0075) (0.0055) (0.0074) 0.0276*** 0.0181*** 0.0282*** 0.0170*** (0.0051) (0.0062) (0.0051) (0.0062) 0.0088** 0.0041 0.0083** 0.0030 (0.0035) (0.0046) (0.0035) (0.0045) Spatial fixed effects Yes No Yes No Temporal fixed effects No Yes No Yes Time - variant control variables Yes Yes Yes Yes Time - invariant control variables No No No No Weather c ondition indictors Yes Yes Yes Yes 0.0420* 0.0633*** (0.0227) (0.0199) 0.3567*** 0.2746*** 0.3920*** 0.2170*** (0.0076) (0.0079) (0.0224) (0.0225) Obs. 442,176 442,176 442,176 442,176 R 2 0.5314 0.5319 0.5302 0.5323 Number of NTAs 188 188 188 188 Panel lengt h 2 , 352 2 , 352 2 , 352 2 , 352 (b) With spatial weight matrix based on r Variables Dependent variable: = ln ( ) Spatial error models Spatial autocorrelation models (1) Spatial fixed effects only (2) Temporal fix ed effects only (3) Spatial fixed effects only (4) Temporal fixed effects only 0.0484*** 0.0742*** 0.0459*** 0.0735*** (0.0052) (0.0074) (0.0054) (0.0074) 0.0269*** 0.0175*** 0.0274*** 0.0167*** (0. 0050) (0.0062) (0.0050) (0.0062) 0.0089** 0.0039 0.0085** 0.0028 (0.0034) (0.0046) (0.0035) (0.0045) Spatial fixed effects Yes No Yes No Temporal fixed effects No Yes No Yes Time - variant control variables Yes Yes Yes Yes T ime - invariant control variables No No No No Weather condition indictors Yes Yes Yes Yes 0.0380* 0.0551*** (0.0218) (0.0197) 0.3427*** 0.2679*** 0.3756*** 0.2173*** (0.0078) (0.0078) (0.0226) (0.0223) Obs. 442,176 442,176 44 2,176 442,176 R 2 0.5315 0.5319 0.5304 0.5322 Number of NTAs 188 188 188 188 Panel length 2 , 352 2 , 352 2 , 352 2 , 352 Note. Robust standard errors in parentheses. Estimates for control variables and dummy weather condition indicators are omitted from the ab ove results. Time - variant control variables include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 133 APPENDIX J. Summary of Additional Robustness Checks Table A 25 . Robustness check: Results of Tobit models Variables Dependent variable: = ln ( ) Cross - sectional Tobit models Panel Tobit models (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) 0.0746*** 0.0800*** (0.0015) (0.0015) 0.0761*** 0.0906*** 0.0563*** 0.0374*** 0.0831*** 0.0713*** 0.1070*** 0.0491*** 0.0276*** 0.0780*** (0.0027) (0.0043) (0.0028) (0.0029) (0.0028) (0.0028) (0.0044) (0.0029) (0.0030) (0.0028) 0.0293*** 0.0210*** 0.1410*** 0.0065** 0.0339*** 0.0297*** 0.02 39*** 0.1544*** 0.0085*** 0.0383*** (0.0026) (0.0028) (0.0054) (0.0028) (0.0044) (0.0027) (0.0029) (0.0056) (0.0029) (0.0045) 0.0006 0.0056*** 0.0001 0.0807*** 0.0197*** 0.0097*** 0.0004 0.0072*** 0.0818*** 0.0136*** (0.0020 ) (0.0020) (0.0020) (0.0030) (0.0023) (0.0021) (0.0021) (0.0021) (0.0032) (0.0024) 0.0724*** 0.0788*** (0.0020) (0.0021) 0.0527*** 0. 0577*** (0.0014) (0.0015) 0.0297*** 0.0322*** (0.0016) (0.0017) Time - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Ye s Yes Yes Yes Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Date and hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Obs. 442,176 442,17 6 442,176 442,176 442,176 442,176 442,176 442,176 442,176 442,176 Left - censored obs. 97,438 97,438 97,438 97,438 97,438 97,438 97,438 97,438 97,438 97,438 Uncensored obs. 344,738 344,738 344,738 344,738 344,738 344,738 344,738 344,738 344,738 344,738 Nu mber of NTAs 188 188 188 188 188 Note. Standard errors in parentheses. Estimates for constant, control variables, and dummy weather condition, borough, and date and hour - of - day indicators are omitted from the above results. Time - variant control varia bles include ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 134 Table A 26 . Robustness check: Results of Tobit models with observations for rush and business hours Variables Dependent variable: = ln ( ) Cross - sectional Tobit models Panel Tobit models (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) 0.0304*** 0.0247*** (0.0020) (0.0023) 0.0462*** 0.0348*** 0.0351*** 0.02 85*** 0.0500*** 0.0299*** 0.0362*** 0.0184*** 0.0136*** 0.0323*** (0.0040) (0.0067) (0.0041) (0.0043) (0.0040) (0.0042) (0.0074) (0.0043) (0.0047) (0.0042) 0.0116*** 0.0045 0.0590*** 0.0008 0.0178*** 0.0202*** 0.0082** 0. 0445*** 0.0122*** 0.0061 (0.0034) (0.0036) (0.0071) (0.0035) (0.0056) (0.0035) (0.0037) (0.0076) (0.0036) (0.0058) 0.0028 0.0012 0.0029 0.0340*** 0.0087*** 0.0108*** 0.0086*** 0.0100*** 0.0211*** 0.0004 (0.0028) (0.0028) (0.0028) (0.0045) (0.0033) (0.0029) (0.0029) (0.0029) (0.0050) (0.0035) 0.0289*** 0.0271*** (0.0026) (0.0028) 0.0206*** 0.0175*** (0.0020) (0.0023) 0.0136*** 0.0125*** (0.0021) (0.0022) Time - variant control variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time - invariant control variables Yes Yes Yes Yes Y es Yes Yes Yes Yes Yes Weather condition indictors Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Borough indicators Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Date and hour - of - day indicators Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Obs. 153,408 153,408 153,408 1 53,408 153,408 153,408 153,408 153,408 153,408 153,408 Left - censored obs. 12,600 12,600 12,600 12,600 12,600 12,600 12,600 12,600 12,600 12,600 Uncensored obs. 140,808 140,808 140,808 140,808 140,808 140,808 140,808 140,808 140,808 140,808 Number of NTA s 188 188 188 188 188 Note. Standard errors in parentheses. Estimates for constant, control variables, and dummy weather condition, borough, and date and hour - of - day indicators are omitted from the above results. Time - variant control variables includ e ln ( ), ln ( ), ln ( ), ln ( ), and ln ( ). Time - invariant control variables include ln ( ), ln ( ), ln ( ), , ln ( ), and ln ( ) . * p < 0.10; ** p < 0.05; *** p < 0.01 . 135 REFERENCES 136 REFERENCES Econom etrica (74:1), pp. 235 - 267. MIS Quarterly (16:2), pp. 227 - 247. MIS Quarterly (24:4), pp. 665 - 694. Procee dings of the Annual Meeting of the Academy of Management , Boston, MA. Available at: engadget.com/2015/10/19/waze - app - update - ios - android/, October 19, 2015. Andreoni, J. Journal of Public Economics (37:3), pp. 291 - 304. Journal of Political Economy ( 97:6), pp. 1447 - 1458. - Glow Economic Journal (100:401), pp. 464 - 477. Handbook of the Economics of Giving, Altruism and Recipro city, Vol. 2 , S. - C. Kolm and J. Mercier Ythier (eds.), Amsterdam, The Netherlands: North Holland, pp. 1201 - 1269. Journal of Public Economics (91:9), pp. 1731 - 174 9. 50 Norm: A Theoretical and Econometrica (77:5), pp. 1607 - 1636. with Economic Inquiry (36:3), pp. 410 - 428. - Marketing Science (35:2), pp. 218 - 233. Aral, S., and Nicola Nature Communications (8:14753), pp. 1 - 8. 137 Advertising? A Field Study of Mobile Display Advertising Effe cts on Consumer Journal of Marketing Research (51:3), pp. 270 - 285. CEIS Working Paper No. 373, Available at SSRN: ssrn.com/abstract=275 4703. Social Science Computer Review (16:3), pp. 293 - 307. - Nordic Journal of Political Economy (26), pp. 33 - 53. Sloan Management Review (54:4), pp. 23 - 29. Determinants o MIS Quarterly (38:3), pp. 773 - 794. - Information Systems Research (12:4), pp. 346 - 362. Butler, Leadership at a Distance: Research in Technologically - Supported Work , S. Weisband (ed.) New York, NY: Psychology Press, pp. 171 - 194. C assidy, J. 2009. How Markets Fail: The Logic of Economic Calamities , New York, NY: Farrar, Straus and Giroux. Economic Geography (46:Suppleme nt), pp. 269 - 292. 2017. Economic Journal (94:375), pp. 580 - 598. Cutler, K. - - apple - maps - onavo/, October 19, 2012. Journal of Personality and Social Psychology (8:4), pp. 377 - 383. 138 Journal of Public Economics (88:9 ), pp. 2159 - 2180. Handbook of Regional Science , M.M. Fischer and P. Nijkamp (eds.), Berlin, Germany: Springer, pp. 1637 - 1652. centive American Economic Review (98:3), pp. 990 - 1008. Literature from Organization Science, Accounting, Marketing, MIS, and Related Information Soc iety (20:5), pp. 325 - 344. Organization Science (22:5), pp. 1224 - 1239. Organization Science (22:6), pp. 1464 - 1480. Knowledge Management: An Evolutionary View of the Field , I. Becerra - Fernandez and D.E. Leidner (eds.), Armonk, NY: M.E. Sharpe, pp. 270 - 291. Available at: fastcompany.com/most - innovative - companies/2018/sectors/transportation, February19, 2018. Fischer, P., Greitemeyer, T., P European Journal of Social Psychology (36:2), pp. 267 - 278. Fischer, P., Krueger, J.I., Greitemeyer, T., Vogrincic, C., Kastenmüller, A., Frey, D., Heene, M., - Effect: A Meta - Analytic Review on Bystander Intervention in Dangerous and Non - Psychological Bulletin (137:4), pp. 517 - 537. Authority and HNTB Corporation, January 19, 2018. Management Science (57:9), pp. 1671 - 1691. Management Science (60:6), pp. 1470 - 1488. Goodchild, M.F. 1986. Spatial Autocorrelation , Norwich, UK: Geo Books. 139 Management Science (41:12), pp. 1827 - 1844. Hall - - to - at: techcrunch.com/2016/10/11/waze - and - ezri - make - app - to - infrastructure - possible/, October 11, 2016. Journal of Regional Science (55:3), pp. 339 - 363. - Continuous MIS Quarterly (40: 4), pp. 983 - 1008. American Economic Review (88:2), pp. 277 - 282. Econometrica (47:1), pp. 153 - 161. Heilbroner, R.L. 1967. The Worldly Philosophers: The Lives, Times, and Ideas of the Great Economic Thinkers, 3rd Edition , New York, NY: Simon & Schuster. Use of a Free Mobil e App for Dietary Self - Monitoring with Photography and Peer Journal of Medical Internet Research (16:4), Available at: jmir.org/2014/4/e109/. - Weary Homeowners and Waze Are at War, Again. Guess Who wapo.st/1UChBto?tid=ss_mail&utm_term=.72f12461f560, June 5, 2016. Howe, J. 2008. Crowdsourcing: How the Power of the Crowd is Driving the Future of Business , New York, NY: Random House. Hsiao, C. - H., Chang, J. - J ., and Tang, K. - Continuance Usage of Mobile Social Apps: Satisfaction, Habit, and Customer Value Telematics and Informatics (33:2), pp. 342 - 355. ACM Transactions on Management Information Systems (2:1), pp. 1 - 16. Business Horizons (55:2), pp. 129 - 139. Kim, A.J. 2000. Community Building on the Web: Secret Strategies for Successful Online Communities , Berkeley, CA: Peachpit Press. 140 Engagement Motivations, Perceived Value, Sat isfaction, and Continued Engagement Decision Support Systems (56), pp. 361 - 370. Kraut, R.E., Resnick, P., Kiesler, S., Burke, M., Chen, Y., Kittur, N., Konstan, J., Ren, Y., and Riedl, J. 2011. Building Successful Online Communities: Evidence - B ased Social Design , Cambridge, MA: MIT Press. Information Systems Research (27:4), pp. 919 - 939. Latané, B., and Darley, J.M. 1968. Journal of Personality and Social Psychology (10:3), pp. 215 - 221. Latané, B., and Darley, J.M. 1970. , New York, NY: Appleton - Century Crofts. Lata Psychological Bulletin (89:2), pp. 308 - 324. Lee, J., and Wong, D.W.S. 2001. Statistical Analysis with ArcView GIS ® , New York, NY: John Wiley & Sons. timation of Spatial Autoregressive Panel Data Models with Fixed Journal of Econometrics (154:2), pp. 165 - 185. News, Available at: nydailynews.com/autos/street - smarts/top - 5 - navigation - maps - apps - drivers - article - 1.2314257, November 15, 2016. LeSage, J., and Pace, R.K. 2009. Introduction to Spatial Econometrics , Boca Raton, FL: Taylor & Francis. n User - Generated Information Systems Research (25:3), pp. 468 - 488. Li, C., and Bernoff, J. 2011. Groundswell: Winning in a World Transformed by Social Technologies , Boston, MA: Harvard Business Press. Too Big to Fail: Large Samples and the p - Information Systems Research (24:4), pp. 906 - 917. Lloyd, C.D. 2010. S patial Data Analysis: An Introduction for GIS Users , Oxford, UK: Oxford University Press. 141 - dailydot.com/layer8/police - waze - fake - speed - traps/, February 12, 2015. Luo, X., A Management Science (60:7), pp. 1738 - 1756. theatlantic.com/technology/archive/2018/03/mapping - apps - and - the - pric e - of - anarchy/555551/, March 15, 2018. Citizenship and Social Class: And Other Essays , T.H. Marshall (ed.) Cambridge, UK: Cambridge University Press pp. 1 - 85. mashable.com/2015/12/08/100 - best - iphone - apps/#4.kFAKTcKPqa, June 25, 2016. Available at: fastcompany.com/40525518/how - waze - works - with - local - gover nments - to - clear - the - roads, March 5, 2018. Mencher, S. 1967. Poor Law to Poverty Program: Economic Security Policy in Britain and the United States , Pittsburgh, PA: University of Pittsburgh Press. Scie nce (167:3924), pp. 1461 - 1468. - Based Advertising: Measuring the Impact of Context - Available at SSRN: ssrn.com/abstract=2116359. Molito Location - Available at SSRN: ssrn.com/abstract=2645281. of Feedback in Managing the Internet - Based Information Systems Research (19:4), pp. 494 - 515. - Making with Sequential Information - Decision Support Syste ms (2:4), pp. 285 - 307. Geographic Information Analysis, 2nd Edition , Hoboken, NJ: John Wiley & Sons. Oestreicher - Strategy for Content Provid MIS Quarterly (37:2), pp. 591 - 616. Olson, M. 1965. The Logic of Collective Action , Cambridge, MA: Harvard University Press. 142 Available at: digita ltrends.com/mobile/waze - vs - google - maps/, February 18, 2018. Now Less Cluttered, but Still Needs - app - waze - gets - a - huge - redesign - n ow - less - cluttered - but - still - needs - improvement/, October 19, 2015. - to - Leader Framework: Motivating Technology - AIS Transactions on Human - Computer Interaction (1:1), pp. 13 - 32. - to - Speech Voice - app - waze - drives - past - 5 - million - users - adds - text - to - speech - voice - guidance/, July 26, 2011a. Rao, L. TechCrunch, Available at: techcrunch.com/2011/12/08/waze - brings - crowdsourced - social - traffic - data - to - local - broadcast - stations/, December 8, 2011b. Rheingold, H. 2000. The Virtual Co mmunity: Homesteading on the Electronic Frontier, Revised Edition , Cambridge, MA: MIT Press. Rheingold, H. 2002. Smart Mobs: The Next Social Revolution , Cambridge, MA: Basic Books. - of - Giving Motivati ons in Charitable Journal of Political Economy (110:2), pp. 425 - 457. Communications of the Association for Information Systems (18:1), pp. 329 - 354. Schlegel, K., Linden, A., Sallam, R.L., Howson, C., Sicular, S., Hare, J., Krensky, P., Gartner, Inc., Available at: gartner.com /doc/3531618, November 30, 2016. Journal of Consumer Research (32:2), pp. 260 - 265. s, Available at: expandedramblings.com/index.php/waze - statistics - facts/, March 20, 2017. Smith, E.A. 1965. Social Welfare: Principles and Concepts , New York, NY: Association Press. Smith, M.A., and Kollock, P. 1999. Communities in Cyberspace , New York, NY: Psychology Press. 143 Biological Journal of the Linnean Society (10:2), pp. 199 - 228. American Economic Review (77:1), pp. 24 - 36. Psychological Review (79:3), pp. 275 - 277. bution for Revenue Sharing and Journal of Management Information Systems (29:2), pp. 41 - 76. Economic Geography (46 ), pp. 234 - 240. - Mediated Communication: Electronic Communication Research (19:5), pp. 539 - 573. Upton, G., and Fingleton, B. 1985. Spatial Data Analysis by E xample, Volume 1: Point Pattern and Quantitative Data , Chichester, UK: John Wiley & Sons. Vittinghoff, E., Glidden, D.V., Shiboski, S.C., and McCulloch, C.E. 2011. Regression Methods in Biostatistics: Linear, Logistic, Survival, and Repeated Measures Model s, 2nd Edition , New York, NY: Springer. Research Policy (32:7), pp. 1217 - 1241. Tests for Model Selection and Non - Econometrica (57:2), pp. 307 - 333. ny.curbed.com/2018/2/6/16979696/new - york - city - traffic - congestion - second - w orst, February 6, 2018a. ny.curbed.com/2018/1/18/16903152/nyc - traffic - congestion - cost - analysis, January 18, 2018b. ining Social Capital and MIS Quarterly (29:1), pp. 35 - 57. MIS Quarterly ( 16:2), pp. 201 - 226. 144 Wei, P. - S., and Lu, H. - Internet Research (24:3), pp. 313 - 331. Today, Available at: usatoday.com/story/tech/news/2017/03/06/mapping - software - routing - waze - google - traffic - calming - algorithmsi/98588980/, March 6, 2017. izing the Signatures of Social Journal of Social Structure (8:2), pp. 1 - 32. Wenger, E. 1998. Communities of Practice: Learning, Meaning, and Identity , Cambridge, UK: Cambridge University Press. ental Guidance on Valuation of Travel Time in Economic Available at: transportation.gov/sites/dot.gov/files/docs/2016%20Revised%20Value%20of%20Travel %20Time%20Guidance .pdf, September 27, 2016. Journal of Personality and Social Psychology (69:3), pp. 569 - 579. Wooldridge, J.M. 2012. Introductory Econometrics: A Modern Approach , Mason, OH: Cengage Learning. - Maximum Likelihood Estimators for Spatial Dynamic Panel Data with Fixed Effects When Both n and T Journal of Econometrics (146:1), pp. 118 - 134.