ESSAYS ON THE IMPACT OF SOCIAL MEDIA IN THE AUTOMOBILE INDUSTRY
By
Yen-Yao Wang

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Business Administration — Business Information Systems — Doctor of Philosophy
2017

ABSTRACT
ESSAYS ON THE IMPACT OF SOCIAL MEDIA IN THE AUTOMOBILE INDUSTRY
By
Yen-Yao Wang
The U.S. automobile industry is increasing reliance on social media marketing and is
ranked the highest one in traditional and digital advertising spending. Despite a significant body
of research in social media, the effectiveness of social media in this competitive marketplace has
not received the detailed examination. My dissertation seeks to fill in this gap by conducting two
studies to examine the impact of social media on customers’ engagement behaviors as well as
firms’ sale performance.
My first essay examines the dynamic interactions between firm-generated content (FGC),
user-generated content (UGC), and offline sales (light vehicles) in the setting of the firm’s
Facebook fan page in the U.S. automobile industry. The findings suggest that (1) FGC is more
effective in influencing offline car sales than UGC, (2) offline car sales would trigger more FGC
and UGC activities, and (3) there is a positive feedback effect between FGC and UGC. These
findings vary across different forms of format presentation and content of post, suggesting that
firms need to fully customize their social media strategy to reach their goals more efficiently.
Furthermore, customers in different groups (luxury versus non-luxury) demonstrate dramatically
different patterns.

My second essay explores the dynamics of online word-of-mouth (WOM) and its
spillover effects by considering the relative effects at the stages of customer awareness and
consideration. The findings indicate that (1) online WOM at the stage of consideration has the
stronger effect on offline car sales than online WOM at the stage of awareness, (2) spillover
effects exist across both stages of awareness and consideration, though effects are heterogeneous
in direction: positive spillover effects at the stage of awareness while negative spillover effects at
the stage of consideration, and (3) at the stage of awareness, online WOM initiated by firms is
more effective in influencing offline car sales than online WOM initiated by users. Furthermore,
not every mechanism at Facebook (i.e., post, like, comment, and share) has the equal impact on
offline car sales and these different mechanisms also influence how customers appreciate online
WOM at the stage of consideration. Finally, the results vary significantly across origin of brand,
market structure, and price factor.
In summary, my dissertation offers valuable insights for firms on how to better develop
their social media strategy to engage with their customers and boost offline car sales in this
economically important industry. Furthermore, these two studies would also advance the
literature by understanding further the dynamics of social media on offline sales of the durable or
high-involvement products. Finally, the unique and rich data also allows me to test the
underlying mechanism at work that will shed light on our theoretical understanding of the impact
of social media from different perspectives.

Copyright By
YEN-YAO WANG
2017

The dissertation is dedicated to my family.

v

ACKNOWLEDGEMENTS

This dissertation would not have been finished without the generous support and
encouragement of faculty, colleagues, friends, and family members. I would like to express my
sincere gratitude to them.
First, I would like to express my deepest gratitude to my dissertation committee co-chairs,
Dr. Vallabh Sambamurthy and Dr. Anjana Susarla, for their invaluable guidance, unstinting
support, and never-ending engorgement throughout my entire PhD journey. I am fortunate to be
guided by them and have benefited from their experience in shaping my dissertation and other
research projects. Your enthusiasm, patience, and supervision have not only made me to build a
solid foundation for my academic career as a researcher, a scholar, and a teacher but also made a
tremendous impact on my personal development as a colleague. Thank you! I would not have
gotten where I am today without your support.
I am heartily thankful to my committee members, Dr. Ranjani Krishnan and Dr. Roger
Calantone, for providing insightful comments and suggestions on my dissertation. I am truly
appreciated your input and encouragement during this journey. I would also like to thank my
former advisor at Iowa State University, Dr. Brian Mennecke, who guided me into academia and
provided continuous support from Iowa State University to Michigan State University.
Thank you my Ph.D. fellows at MSU, Dr. Eun Ju Jung, Dr. Kangkang Qi, Yu Huang, Tae

vi

Hun Kim, Yimo Liu, and Ink Yu Kim. I am indebted to them for providing a warm environment
that motivated me to get my Ph.D. degree.
This journey would not have been completed without the support and love of my family.
I would like to thank my father, Wen-Yu Wang, my mother, Jui-Wan Lin, my sister, Chieh-Yun
Wang, and my aunt, Jui-Sui Lin for your unconditional support and love, which helped me go
through all difficult moments in the adventure of pursuing my Ph.D. degree. I am forever
grateful to them for the opportunity they provided to me. I also would like to thank my parentsin-law, Chung-Yi Shen and Yu-Hsiu Hsu, and my brother-in-law, Yu-Chuan Shen, for your
continues care and prayer. I love you all.
Finally, my special thanks go to my beautiful and loving wife, Yu-Hsien Shen, for being
a listener and soul mate throughout my Ph.D. program and life. Without your love, I could not
make this journey happen. Thank you and I love you!

vii

TABLE OF CONTENTS

LIST OF TABLES .............................................................................................................. x
LIST OF FIGURES .......................................................................................................... xv
CHAPTER 1. OVERVIEW OF DISSERTATION ............................................................ 1
CHAPTER2. ESSAY 1: ONLINE TO OFFLINE: THE IMPACT OF SOCIAL MEDIA
ON OFFLINE SALES IN THE AUTOMOBILE INDUSTRY ......................................... 3
2.1 INTRODUCTION .................................................................................................. 3
2.2 LITERATURE REVIEW ..................................................................................... 10
2.2.1 Effect of UGC on Sales .............................................................................. 10
2.2.2 Effect of FGC on Sales ............................................................................... 11
2.2.3 Relative Effects of UGC and FGC on Sales............................................... 12
2.3 RESEARCH MODEL AND HYPOTHESES ...................................................... 17
2.4 DATA AND EMPIRICAL METHODOLOGY ................................................... 25
2.4.1 Research Context ................................................................................................ 25
2.4.2 Data ............................................................................................................ 25
2.4.3 PVAR Model Specification ........................................................................ 28
2.4.4 PVAR Model Estimation Procedure .......................................................... 31
2.5 RESULTS ............................................................................................................. 36
2.5.1 Main PVAR Analysis Results .................................................................... 36
2.5.1.1 Effect of Overall Posts ...................................................................... 36
2.5.1.2 Effect of Format Presentation............................................................ 37
2.5.1.3 Effect of Content ............................................................................... 40
2.5.2 Sample Split PVAR Analysis Results ........................................................ 42
2.5.2.1 Effect of Overall Posts ...................................................................... 43
2.5.2.2 Effect of Format Presentation............................................................ 44
2.5.2.3 Effect of Content ............................................................................... 47
2.5.3 Impulse Response Functions (IRFs) Results .............................................. 51
2.5.3.1 Sample Split Impulse Response Functions (IRFs) Results ............... 58
2.6 DISCUSSION ....................................................................................................... 68
2.6.1 Theoretical Implications ............................................................................. 68
2.6.2 Managerial Implications ............................................................................. 72
2.6.3 Limitations and Future Research ................................................................ 73

viii

2.7 CONCLUDING REMARKS ................................................................................ 73
CHAPTER3. ESSAY 2: DYNAMICS OF ONLINE WORD OF MOUTH (WOM)
SPILLOVER EFFECTS ................................................................................................... 75
3.1 INTRODUCTION ................................................................................................ 75
3.2 LITERATURE REVIEW ..................................................................................... 81
3.2.1 Effect of Online WOM on Sales ................................................................ 81
3.2.2 Spillover Effects of Online WOM.............................................................. 84
3.3 RESEARCH MODEL AND HYPOTHESES ...................................................... 86
3.4 DATA AND EMPIRICAL METHODOLOGY ................................................... 94
3.4.1 Research Context ........................................................................................ 94
3.4.2 Data ............................................................................................................ 95
3.4.3 Bayesian Model Specification and Estimation Procedure........................ 102
3.5 RESULTS ........................................................................................................... 105
3.5.1 Main Bayesian Analysis Results .............................................................. 105
3.5.2 Sample Split Bayesian Analysis Results on the Origin of Brands ........... 126
3.5.3 Sample Split Bayesian Analysis Results on Market Structure ................. 181
3.5.4 Sample Split Bayesian Analysis Results on Luxury and Non-Luxury Brands
................................................................................................................................. 236
3.6 DISCUSSION ..................................................................................................... 273
3.6.1 Theoretical Implications ........................................................................... 274
3.6.2 Managerial Implications ........................................................................... 277
3.6.3 Limitations and Future Research .............................................................. 278
3.7 CONCLUDING REMARKS .............................................................................. 278
REFERENCES………………………………………………………………………….280

ix

LIST OF TABLES

Table 1. Summary of Selected Literature ......................................................................... 15	
Table 2. Variable Definition ............................................................................................. 27	
Table 3. Descriptive Statistics........................................................................................... 28	
Table 4. Fisher-type Unit-Root Tests................................................................................ 31	
Table 5. Im-Pesaran-Shin Unit-Root Tests ....................................................................... 32	
Table 6. Lag Selection of Overall Post Model .................................................................. 33	
Table 7. Lag Selection of Link Post Model ...................................................................... 33	
Table 8. Lag Selection of Photo Post Model .................................................................... 34	
Table 9. Lag Selection of Status Post Model .................................................................... 34	
Table 10. Lag Selection of Video Post Model .................................................................. 34	
Table 11. Lag Selection of Informative Post Model ......................................................... 34	
Table 12. Lag Selection of Sentiment Post Model ........................................................... 34	
Table 13. Overall Post PVAR Regression Results ........................................................... 37	
Table 14. Format Presentation: Link PVAR Regression Results ..................................... 38	
Table 15. Format Presentation: Photo PVAR Regression Results ................................... 38	
Table 16. Format Presentation: Status PVAR Regression Results ................................... 39	
Table 17. Format Presentation: Video PVAR Regression Results ................................... 39	

x

Table 18. Examples of Content……………… ................................................................. 40	
Table 19. Informative Post PVAR Regression Results .................................................... 42	
Table 20. Sentiment PVAR Regression Results ............................................................... 42	
Table 21. Descriptive Statistics for Luxury Car Makes .................................................... 42	
Table 22. Descriptive Statistics for non-Luxury Car Makes ............................................ 43	
Table 23. Overall PVAR Regression Results (Luxury) .................................................... 44	
Table 24. Overall PVAR Regression Results (Non-Luxury)............................................ 44	
Table 25. Link PVAR Regression Results (Luxury) ........................................................ 44	
Table 26. Link PVAR Regression Results (Non-Luxury) ................................................ 44	
Table 27. Photo PVAR Regression Results (Luxury) ...................................................... 45	
Table 28. Photo PVAR Regression Results (Non-Luxury) .............................................. 46	
Table 29. Status PVAR Regression Results (Luxury) ...................................................... 46	
Table 30. Status PVAR Regression Results (Non-Luxury) .............................................. 46	
Table 31. Video PVAR Regression Results (Luxury) ...................................................... 47	
Table 32. Video PVAR Regression Results (Non-Luxury) .............................................. 47	
Table 33. Informative Post PVAR Regression Results (Luxury) ..................................... 48	
Table 34. Informative Post PVAR Regression Results (Non-Luxury) ............................. 48	
Table 35. Sentiment PVAR Regression Results (Luxury)................................................ 49	
Table 36. Sentiment PVAR Regression Results (Non-Luxury) ....................................... 50	

xi

Table 37. Summary of Results .......................................................................................... 50	
Table 38. Full Content Example of Test Drive Experience ............................................ 101	
Table 39. Variable Definition ......................................................................................... 101	
Table 40. Descriptive Statistics....................................................................................... 102	
Table 41. Bayesian Estimation Results for Posts............................................................ 106	
Table 42. Bayesian Estimation Results for Likes ........................................................... 111	
Table 43. Bayesian Estimation Results for Comments ................................................... 116	
Table 44. Bayesian Estimation Results for Shares ......................................................... 122	
Table 45. Descriptive Statistics for Asian-Based Brands ............................................... 126	
Table 46. Descriptive Statistics for European-Based Brands ......................................... 127	
Table 47. Descriptive Statistics for US-Based Brands ................................................... 127	
Table 48. Bayesian Estimation Results for Posts (Asian-based) .................................... 129	
Table 49. Bayesian Estimation Results for Posts (European-based) .............................. 129	
Table 50. Bayesian Estimation Results for Posts (US-based) ........................................ 130	
Table 51. Bayesian Estimation Results for Likes (Asian-based) .................................... 142	
Table 52. Bayesian Estimation Results for Likes (European-based) .............................. 143	
Table 53. Bayesian Estimation Results for Likes (US-based) ........................................ 143	
Table 54. Bayesian Estimation Results for Comments (Asian-based) ........................... 156	
Table 55. Bayesian Estimation Results for Comments (European-based) ..................... 156	

xii

Table 56. Bayesian Estimation Results for Comments (US-based) ............................... 156	
Table 57. Bayesian Estimation Results for Shares (Asian-based) .................................. 169	
Table 58. Bayesian Estimation Results for Shares (European-based) ............................ 169	
Table 59. Bayesian Estimation Results for Shares (US-based) ...................................... 169	
Table 60. Descriptive Statistics for Group 1 ................................................................... 182	
Table 61. Descriptive Statistics for Group 2 ................................................................... 182	
Table 62. Descriptive Statistics for Group 3 ................................................................... 183	
Table 63. Bayesian Estimation Results for Posts (Group 1) ........................................... 184	
Table 64. Bayesian Estimation Results for Posts (Group 2) ........................................... 184	
Table 65. Bayesian Estimation Results for Posts (Group 3) ........................................... 185	
Table 66. Bayesian Estimation Results for Likes (Group 1) .......................................... 197	
Table 67. Bayesian Estimation Results for Likes (Group 2) .......................................... 198	
Table 68. Bayesian Estimation Results for Likes (Group 3) .......................................... 198	
Table 69. Bayesian Estimation Results for Comments (Group 1) .................................. 211	
Table 70. Bayesian Estimation Results for Comments (Group 2) .................................. 211	
Table 71. Bayesian Estimation Results for Comments (Group 3) .................................. 211	
Table 72. Bayesian Estimation Results for Shares (Group 1) ........................................ 224	
Table 73. Bayesian Estimation Results for Shares (Group 2) ........................................ 224	
Table 74. Bayesian Estimation Results for Shares (Group 3) ........................................ 224	

xiii

Table 75. Descriptive Statistics for Luxury brands ........................................................ 236	
Table 76. Descriptive Statistics for Non-Luxury brands ................................................ 237	
Table 77. Bayesian Estimation Results for Posts (Luxury) ............................................ 238	
Table 78. Bayesian Estimation Results for Posts (Non-Luxury) .................................... 238	
Table 79. Bayesian Estimation Results for Likes (Luxury) ............................................ 247	
Table 80. Bayesian Estimation Results for Likes (Non-Luxury) ................................... 247	
Table 81. Bayesian Estimation Results for Comments (Luxury) ................................... 255	
Table 82. Bayesian Estimation Results for Comments (Non-Luxury) ........................... 256	
Table 83. Bayesian Estimation Results for Shares (Luxury) .......................................... 264	
Table 84. Bayesian Estimation Results for Shares (Non-Luxury) .................................. 265	
Table 85. Summary of Results ........................................................................................ 273	

xiv

LIST OF FIGURES

Figure 1. Conceptual Framework ..................................................................................... 19	
Figure 2. Example of FGC (left) and UGC (right) ........................................................... 19	
Figure 3. Example of FGC in the form of Link, Photo, Status, and Video ...................... 20	
Figure 4. Analysis Procedure ............................................................................................ 35	
Figure 5. Impulse Response Functions for Overall Post................................................... 52	
Figure 6. Impulse Response Functions for Link ............................................................... 53	
Figure 7. Impulse Response Functions for Photo ............................................................. 54	
Figure 8. Impulse Response Functions for Status ............................................................. 55	
Figure 9. Impulse Response Functions for Video ............................................................. 56	
Figure 10. Impulse Response Functions for Informative Posts ........................................ 57	
Figure 11. Impulse Response Functions for Sentiment .................................................... 58	
Figure 12. Impulse Response Functions for Overall Post (Luxury) ................................. 59	
Figure 13. Impulse Response Functions for Overall Post (Non-Luxury) ......................... 59	
Figure 14. Impulse Response Functions for Link (Luxury) ............................................. 60	
Figure 15. Impulse Response Functions for Link (Non-Luxury) ..................................... 61	
Figure 16. Impulse Response Functions for Photo (Luxury) ............................................ 62	
Figure 17. Impulse Response Functions for Photo (Non-Luxury) ................................... 62	

xv

Figure 18. Impulse Response Functions for Status (Luxury) ........................................... 63	
Figure 19. Impulse Response Functions for Status (Non-Luxury) ................................... 63	
Figure 20. Impulse Response Functions for Video (Luxury) ........................................... 64	
Figure 21. Impulse Response Functions for Video (Non-Luxury) ................................... 65	
Figure 22. Impulse Response Functions for Informative Posts (Luxury) ......................... 66	
Figure 23. Impulse Response Functions for Informative Posts (Non-Luxury) ................ 66	
Figure 24. Impulse Response Functions for Sentiment (Luxury) ..................................... 67	
Figure 25. Impulse Response Functions for Sentiment (Non-Luxury)............................. 67	
Figure 26. Conceptual Model ........................................................................................... 89	
Figure 27. Example of Post at the Stage of Awareness .................................................. 100	
Figure 28. Example of Post at the Stage of Consideration ............................................. 100	
Figure 29. Assessment of Model Convergence for F-Post (a) ........................................ 106	
Figure 30. Assessment of Model Convergence for C-F-Post (a) .................................... 107	
Figure 31. Assessment of Model Convergence for U-Post (a) ....................................... 107	
Figure 32. Assessment of Model Convergence for C-U-Post (a) ................................... 107	
Figure 33. Assessment of Model Convergence for TD-Post (c) ..................................... 108	
Figure 34. Assessment of Model Convergence for C-TD-Post (c) ................................. 108	
Figure 35. Assessment of Model Convergence for TMS ............................................... 108	
Figure 36. Assessment of Model Convergence for Price ............................................... 109	

xvi

Figure 37. Assessment of Model Convergence for GT .................................................. 109	
Figure 38. Assessment of Model Convergence for GPI ................................................. 109	
Figure 39. Assessment of Model Convergence for CCI ................................................. 110	
Figure 40. Assessment of Model Convergence for F-Like (a) ....................................... 111	
Figure 41. Assessment of Model Convergence for C-F-Like (a) ................................... 112	
Figure 42. Assessment of Model Convergence for U-Like (a) ....................................... 112	
Figure 43. Assessment of Model Convergence for C-U-Like (a) ................................... 112	
Figure 44. Assessment of Model Convergence for TD-Post (c) ..................................... 113	
Figure 45. Assessment of Model Convergence for C-TD-Post (c) ................................. 113	
Figure 46. Assessment of Model Convergence for TMS ............................................... 113	
Figure 47. Assessment of Model Convergence for Price ............................................... 114	
Figure 48. Assessment of Model Convergence for GT .................................................. 114	
Figure 49. Assessment of Model Convergence for GPI ................................................. 114	
Figure 50. Assessment of Model Convergence for CCI ................................................. 115	
Figure 51. Assessment of Model Convergence for F-Comment (a) ............................... 116	
Figure 52. Assessment of Model Convergence for C-F-Comment (a) ........................... 117	
Figure 53. Assessment of Model Convergence for U-Comment (a) .............................. 117	
Figure 54. Assessment of Model Convergence for C-U-Comment (a) .......................... 117	
Figure 55. Assessment of Model Convergence for TD-Post (c) ..................................... 118	

xvii

Figure 56. Assessment of Model Convergence for C-TD-Post (c) ................................. 118	
Figure 57. Assessment of Model Convergence for TMS ............................................... 118	
Figure 58. Assessment of Model Convergence for Price ............................................... 119	
Figure 59. Assessment of Model Convergence for GT .................................................. 119	
Figure 60. Assessment of Model Convergence for GPI ................................................. 119	
Figure 61. Assessment of Model Convergence for CCI ................................................. 120	
Figure 62. Assessment of Model Convergence for F-Share (a)...................................... 122	
Figure 63. Assessment of Model Convergence for C-F-Share (a).................................. 123	
Figure 64. Assessment of Model Convergence for U-Share (a) ..................................... 123	
Figure 65. Assessment of Model Convergence for C-U-Share (a) ................................. 123	
Figure 66. Assessment of Model Convergence for TD-Post (c) ..................................... 124	
Figure 67. Assessment of Model Convergence for C-TD-Post (c) ................................. 124	
Figure 68. Assessment of Model Convergence for TMS ............................................... 124	
Figure 69. Assessment of Model Convergence for Price ............................................... 125	
Figure 70. Assessment of Model Convergence for GT .................................................. 125	
Figure 71. Assessment of Model Convergence for GPI ................................................. 125	
Figure 72. Assessment of Model Convergence for CCI ................................................. 126	
Figure 73. Assessment of Model Convergence for F-Post (a) (Asian-based) ................ 130	
Figure 74. Assessment of Model Convergence for C-F-Post (a) (Asian-based) ............ 131	

xviii

Figure 75. Assessment of Model Convergence for U-Post (a) (Asian-based) ................ 131	
Figure 76. Assessment of Model Convergence for C-U-Post (a) (Asian-based) ............ 131	
Figure 77. Assessment of Model Convergence for TD-Post (c) (Asian-based) ............. 132	
Figure 78. Assessment of Model Convergence for C-TD-Post (c) (Asian-based) ......... 132	
Figure 79. Assessment of Model Convergence for TMS (Asian-based) ........................ 132	
Figure 80. Assessment of Model Convergence for Price (Asian-based) ........................ 133	
Figure 81. Assessment of Model Convergence for GT (Asian-based) ........................... 133	
Figure 82. Assessment of Model Convergence for GPI (Asian-based) .......................... 133	
Figure 83. Assessment of Model Convergence for CCI (Asian-based).......................... 134	
Figure 84. Assessment of Model Convergence for F-Post (a) (European-based) .......... 134	
Figure 85. Assessment of Model Convergence for C-F-Post (a) (European-based) ...... 134	
Figure 86. Assessment of Model Convergence for U-Post (a) (European-based) .......... 135	
Figure 87. Assessment of Model Convergence for C-U-Post (a) (European-based) ...... 135	
Figure 88. Assessment of Model Convergence for TD-Post (c) (European-based) ....... 135	
Figure 89. Assessment of Model Convergence for C-TD-Post (c) (European-based) ... 136	
Figure 90. Assessment of Model Convergence for TMS (European-based) .................. 136	
Figure 91. Assessment of Model Convergence for Price (European-based) .................. 136	
Figure 92. Assessment of Model Convergence for GT (European-based) ..................... 137	
Figure 93. Assessment of Model Convergence for GPI (European-based) .................... 137	

xix

Figure 94. Assessment of Model Convergence for CCI (European-based).................... 137	
Figure 95. Assessment of Model Convergence for F-Post (a) (US-based)..................... 138	
Figure 96. Assessment of Model Convergence for C-F-Post (a) (US-based)................. 138	
Figure 97. Assessment of Model Convergence for U-Post (a) (US-based) .................... 138	
Figure 98. Assessment of Model Convergence for C-U-Post (a) (US-based) ................ 139	
Figure 99. Assessment of Model Convergence for TD-Post (c) (US-based) ................. 139	
Figure 100. Assessment of Model Convergence for C-TD-Post (c) (US-based) ........... 139	
Figure 101. Assessment of Model Convergence for TMS (US-based) .......................... 140	
Figure 102. Assessment of Model Convergence for Price (US-based) .......................... 140	
Figure 103. Assessment of Model Convergence for GT (US-based) ............................. 140	
Figure 104. Assessment of Model Convergence for GPI (US-based) ............................ 141	
Figure 105. Assessment of Model Convergence for CCI (US-based) ............................ 141	
Figure 106. Assessment of Model Convergence for F-Like (a) (Asian-based) .............. 143	
Figure 107. Assessment of Model Convergence for C-F-Like (a) (Asian-based) .......... 144	
Figure 108. Assessment of Model Convergence for U-Like (a) (Asian-based) ............. 144	
Figure 109. Assessment of Model Convergence for C-U-Like (a) (Asian-based) ......... 144	
Figure 110. Assessment of Model Convergence for TD-Post (c) (Asian-based) ........... 145	
Figure 111. Assessment of Model Convergence for C-TD-Post (c) (Asian-based) ....... 145	
Figure 112. Assessment of Model Convergence for TMS (Asian-based) ...................... 145	

xx

Figure 113. Assessment of Model Convergence for Price (Asian-based) ...................... 146	
Figure 114. Assessment of Model Convergence for GT (Asian-based) ......................... 146	
Figure 115. Assessment of Model Convergence for GPI (Asian-based) ........................ 146	
Figure 116. Assessment of Model Convergence for CCI (Asian-based)........................ 147	
Figure 117. Assessment of Model Convergence for F-Like (a) (European-based) ........ 147	
Figure 118. Assessment of Model Convergence for C-F-Like (a) (European-based) .... 147	
Figure 119. Assessment of Model Convergence for U-Like (a) (European-based) ....... 148	
Figure 120. Assessment of Model Convergence for C-U-Like (a) (European-based) ... 148	
Figure 121. Assessment of Model Convergence for TD-Post (c) (European-based) ..... 148	
Figure 122. Assessment of Model Convergence for C-TD-Post (c) (European-based) . 149	
Figure 123. Assessment of Model Convergence for TMS (European-based) ................ 149	
Figure 124. Assessment of Model Convergence for Price (European-based) ................ 149	
Figure 125. Assessment of Model Convergence for GT (European-based) ................... 150	
Figure 126. Assessment of Model Convergence for GPI (European-based) .................. 150	
Figure 127. Assessment of Model Convergence for CCI (European-based).................. 150	
Figure 128. Assessment of Model Convergence for F-Like (a) (US-based) .................. 151	
Figure 129. Assessment of Model Convergence for C-F-Like (a) (US-based) .............. 151	
Figure 130. Assessment of Model Convergence for U-Like (a) (US-based) ................. 151	
Figure 131. Assessment of Model Convergence for C-U-Like (a) (US-based) ............. 152	

xxi

Figure 132. Assessment of Model Convergence for TD-Post (c) (US-based) ............... 152	
Figure 133. Assessment of Model Convergence for C-TD-Post (c) (US-based) ........... 152	
Figure 134. Assessment of Model Convergence for TMS (US-based) .......................... 153	
Figure 135. Assessment of Model Convergence for Price (US-based) .......................... 153	
Figure 136. Assessment of Model Convergence for GT (US-based) ............................. 153	
Figure 137. Assessment of Model Convergence for GPI (US-based) ............................ 154	
Figure 138. Assessment of Model Convergence for CCI (US-based) ............................ 154	
Figure 139. Assessment of Model Convergence for F-Comment (a) (Asian-based) ..... 157	
Figure 140. Assessment of Model Convergence for C-F-Comment (a) (Asian-based) . 157	
Figure 141. Assessment of Model Convergence for U-Comment (a) (Asian-based) ..... 157	
Figure 142. Assessment of Model Convergence for C-U-Comment (a) (Asian-based) . 158	
Figure 143. Assessment of Model Convergence for TD-Post (c) (Asian-based) ........... 158	
Figure 144. Assessment of Model Convergence for C-TD-Post (c) (Asian-based) ....... 158	
Figure 145. Assessment of Model Convergence for TMS (Asian-based) ...................... 159	
Figure 146. Assessment of Model Convergence for Price (Asian-based) ...................... 159	
Figure 147. Assessment of Model Convergence for GT (Asian-based) ......................... 159	
Figure 148. Assessment of Model Convergence for GPI (Asian-based) ........................ 160	
Figure 149. Assessment of Model Convergence for CCI (Asian-based)........................ 160	
Figure 150. Assessment of Model Convergence for F-Comment (a) (European-based).160	

xxii

Figure 151. Assessment of Model Convergence for C-F-Comment (a) (European-based)161	
Figure 152. Assessment of Model Convergence for U-Comment (a) (European-based)161	
Figure 153. Assessment of Model Convergence for C-U-Comment (a) (European-based)161	
Figure 154. Assessment of Model Convergence for TD-Post (c) (European-based) ..... 162	
Figure 155. Assessment of Model Convergence for C-TD-Post (c) (European-based) . 162	
Figure 156. Assessment of Model Convergence for TMS (European-based) ................ 162	
Figure 157. Assessment of Model Convergence for Price (European-based) ................ 163	
Figure 158. Assessment of Model Convergence for GT (European-based) ................... 163	
Figure 159. Assessment of Model Convergence for GPI (European-based) .................. 163	
Figure 160. Assessment of Model Convergence for CCI (European-based).................. 164	
Figure 161. Assessment of Model Convergence for F-Comment (a) (US-based) .......... 164	
Figure 162. Assessment of Model Convergence for C-F-Comment (a) (US-based) ...... 164	
Figure 163. Assessment of Model Convergence for U-Comment (a) (US-based) ......... 165	
Figure 164. Assessment of Model Convergence for C-U-Comment (a) (US-based) ..... 165	
Figure 165. Assessment of Model Convergence for TD-Post (c) (US-based) ............... 165	
Figure 166. Assessment of Model Convergence for C-TD-Post (c) (US-based) ........... 166	
Figure 167. Assessment of Model Convergence for TMS (US-based) .......................... 166	
Figure 168. Assessment of Model Convergence for Price (US-based) .......................... 166	
Figure 169. Assessment of Model Convergence for GT (US-based) ............................. 167	

xxiii

Figure 170. Assessment of Model Convergence for GPI (US-based) ............................ 167	
Figure 171. Assessment of Model Convergence for CCI (US-based) ............................ 167	
Figure 172. Assessment of Model Convergence for F-Share (a) (Asian-based) ............ 170	
Figure 173. Assessment of Model Convergence for C-F-Share (a) (Asian-based) ........ 170	
Figure 174. Assessment of Model Convergence for U-Share (a) (Asian-based) ........... 171	
Figure 175. Assessment of Model Convergence for C-U-Share (a) (Asian-based) ....... 171	
Figure 176. Assessment of Model Convergence for TD-Post (c) (Asian-based) ........... 171	
Figure 177. Assessment of Model Convergence for C-TD-Post (c) (Asian-based) ....... 172	
Figure 178. Assessment of Model Convergence for TMS (Asian-based) ...................... 172	
Figure 179. Assessment of Model Convergence for Price (Asian-based) ...................... 172	
Figure 180. Assessment of Model Convergence for GT (Asian-based) ......................... 173	
Figure 181. Assessment of Model Convergence for GPI (Asian-based) ........................ 173	
Figure 182. Assessment of Model Convergence for CCI (Asian-based)........................ 173	
Figure 183. Assessment of Model Convergence for F-Share (a) (European-based) ...... 174	
Figure 184. Assessment of Model Convergence for C-F-Share (a) (European-based) .. 174	
Figure 185. Assessment of Model Convergence for U-Share (a) (European-based) ..... 174	
Figure 186. Assessment of Model Convergence for C-U-Share (a) (European-based) . 175	
Figure 187. Assessment of Model Convergence for TD-Post (c) (European-based) ..... 175	
Figure 188. Assessment of Model Convergence for C-TD-Post (c) (European-based) . 175	

xxiv

Figure 189. Assessment of Model Convergence for TMS (European-based) ................ 175	
Figure 190. Assessment of Model Convergence for Price (European-based) ................ 176	
Figure 191. Assessment of Model Convergence for GT (European-based) ................... 176	
Figure 192. Assessment of Model Convergence for GPI (European-based) .................. 176	
Figure 193. Assessment of Model Convergence for CCI (European-based).................. 177	
Figure 194. Assessment of Model Convergence for F-Share (a) (US-based) ................ 177	
Figure 195. Assessment of Model Convergence for C-F-Share (a) (US-based) ............ 177	
Figure 196. Assessment of Model Convergence for U-Share (a) (US-based) ................ 178	
Figure 197. Assessment of Model Convergence for C-U-Share (a) (US-based) ............ 178	
Figure 198. Assessment of Model Convergence for TD-Post (c) (US-based) ............... 178	
Figure 199. Assessment of Model Convergence for C-TD-Post (c) (US-based) ........... 179	
Figure 200. Assessment of Model Convergence for TMS (US-based) .......................... 179	
Figure 201. Assessment of Model Convergence for Price (US-based) .......................... 179	
Figure 202. Assessment of Model Convergence for GT (US-based) ............................. 180	
Figure 203. Assessment of Model Convergence for GPI (US-based) ............................ 180	
Figure 204. Assessment of Model Convergence for CCI (US-based) ............................ 180	
Figure 205. Netzer et al.’s (2012) Market Structures ..................................................... 182	
Figure 206. Assessment of Model Convergence for F-Post (a) (Group 1) ..................... 185	
Figure 207. Assessment of Model Convergence for C-F-Post (a) (Group 1) ................. 186	

xxv

Figure 208. Assessment of Model Convergence for U-Post (a) (Group 1) .................... 186	
Figure 209. Assessment of Model Convergence for C-U-Post (a) (Group 1) ................ 186	
Figure 210. Assessment of Model Convergence for TD-Post (c) (Group 1) .................. 187	
Figure 211. Assessment of Model Convergence for C-TD-Post (c) (Group 1) .............. 187	
Figure 212. Assessment of Model Convergence for TMS (Group 1)............................. 187	
Figure 213. Assessment of Model Convergence for Price (Group 1)............................. 188	
Figure 214. Assessment of Model Convergence for GT (Group 1) ............................... 188	
Figure 215. Assessment of Model Convergence for GPI (Group 1) .............................. 188	
Figure 216. Assessment of Model Convergence for CCI (Group 1) .............................. 189	
Figure 217. Assessment of Model Convergence for F-Post (a) (Group 2) ..................... 189	
Figure 218. Assessment of Model Convergence for C-F-Post (a) (Group 2) ................. 189	
Figure 219. Assessment of Model Convergence for U-Post (a) (Group 2) .................... 190	
Figure 220. Assessment of Model Convergence for C-U-Post (a) (Group 2) ................ 190	
Figure 221. Assessment of Model Convergence for TD-Post (c) (Group 2) .................. 190	
Figure 222. Assessment of Model Convergence for C-TD-Post (c) (Group 2) .............. 191	
Figure 223. Assessment of Model Convergence for TMS (Group 2)............................. 191	
Figure 224. Assessment of Model Convergence for Price (Group 2)............................. 191	
Figure 225. Assessment of Model Convergence for GT (Group 2) ............................... 192	
Figure 226. Assessment of Model Convergence for GPI (Group 2) .............................. 192	

xxvi

Figure 227. Assessment of Model Convergence for CCI (Group 2) .............................. 192	
Figure 228. Assessment of Model Convergence for F-Post (a) (Group 3) ..................... 193	
Figure 229. Assessment of Model Convergence for C-F-Post (a) (Group 3) ................. 193	
Figure 230. Assessment of Model Convergence for U-Post (a) (Group 3) .................... 193	
Figure 231. Assessment of Model Convergence for C-U-Post (a) (Group 3) ................ 194	
Figure 232. Assessment of Model Convergence for TD-Post (c) (Group 3) .................. 194	
Figure 233. Assessment of Model Convergence for C-TD-Post (c) (Group 3) .............. 194	
Figure 234. Assessment of Model Convergence for TMS (Group 3)............................. 195	
Figure 235. Assessment of Model Convergence for Price (Group 3)............................. 195	
Figure 236. Assessment of Model Convergence for GT (Group 3) ............................... 195	
Figure 237. Assessment of Model Convergence for GPI (Group 3) .............................. 196	
Figure 238. Assessment of Model Convergence for CCI (Group 3) .............................. 196	
Figure 239. Assessment of Model Convergence for F-Like (a) (Group 1) .................... 198	
Figure 240. Assessment of Model Convergence for C-F-Like (a) (Group 1) ................ 199	
Figure 241. Assessment of Model Convergence for U-Like (a) (Group 1) .................... 199	
Figure 242. Assessment of Model Convergence for C-U-Like (a) (Group 1) ................ 199	
Figure 243. Assessment of Model Convergence for TD-Post (c) (Group 1) .................. 200	
Figure 244. Assessment of Model Convergence for C-TD-Post (c) (Group 1) .............. 200	
Figure 245. Assessment of Model Convergence for TMS (Group 1)............................. 200	

xxvii

Figure 246. Assessment of Model Convergence for Price (Group 1)............................. 201	
Figure 247. Assessment of Model Convergence for GT (Group 1) ............................... 201	
Figure 248. Assessment of Model Convergence for GPI (Group 1) .............................. 201	
Figure 249. Assessment of Model Convergence for CCI (Group 1) .............................. 202	
Figure 250. Assessment of Model Convergence for F-Like (a) (Group 2) .................... 202	
Figure 251. Assessment of Model Convergence for C-F-Like (a) (Group 2) ................ 202	
Figure 252. Assessment of Model Convergence for U-Like (a) (Group 2) .................... 203	
Figure 253. Assessment of Model Convergence for C-U-Like (a) (Group 2) ................ 203	
Figure 254. Assessment of Model Convergence for TD-Post (c) (Group 2) .................. 203	
Figure 255. Assessment of Model Convergence for C-TD-Post (c) (Group 2) .............. 204	
Figure 256. Assessment of Model Convergence for TMS (Group 2)............................. 204	
Figure 257. Assessment of Model Convergence for Price (Group 2)............................. 204	
Figure 258. Assessment of Model Convergence for GT (Group 2) ............................... 205	
Figure 259. Assessment of Model Convergence for GPI (Group 2) .............................. 205	
Figure 260. Assessment of Model Convergence for CCI (Group 2) .............................. 205	
Figure 261. Assessment of Model Convergence for F-Like (a) (Group 3) .................... 206	
Figure 262. Assessment of Model Convergence for C-F-Like (a) (Group 3) ................ 206	
Figure 263. Assessment of Model Convergence for U-Like (a) (Group 3) .................... 206	
Figure 264. Assessment of Model Convergence for C-U-Like (a) (Group 3) ................ 207	

xxviii

Figure 265. Assessment of Model Convergence for TD-Post (c) (Group 3) .................. 207	
Figure 266. Assessment of Model Convergence for C-TD-Post (c) (Group 3) .............. 207	
Figure 267. Assessment of Model Convergence for TMS (Group 3)............................. 208	
Figure 268. Assessment of Model Convergence for Price (Group 3)............................. 208	
Figure 269. Assessment of Model Convergence for GT (Group 3) ............................... 208	
Figure 270. Assessment of Model Convergence for GPI (Group 3) .............................. 209	
Figure 271. Assessment of Model Convergence for CCI (Group 3) .............................. 209	
Figure 272. Assessment of Model Convergence for F-Comment (a) (Group 1) ............ 212	
Figure 273. Assessment of Model Convergence for C-F-Comment (a) (Group 1) ........ 212	
Figure 274. Assessment of Model Convergence for U-Comment (a) (Group 1) ........... 212	
Figure 275. Assessment of Model Convergence for C-U-Comment (a) (Group 1) ....... 213	
Figure 276. Assessment of Model Convergence for TD-Post (c) (Group 1) .................. 213	
Figure 277. Assessment of Model Convergence for C-TD-Post (c) (Group 1) .............. 213	
Figure 278. Assessment of Model Convergence for TMS (Group 1)............................. 214	
Figure 279. Assessment of Model Convergence for Price (Group 1)............................. 214	
Figure 280. Assessment of Model Convergence for GT (Group 1) ............................... 214	
Figure 281. Assessment of Model Convergence for GPI (Group 1) .............................. 215	
Figure 282. Assessment of Model Convergence for CCI (Group 1) .............................. 215	
Figure 283. Assessment of Model Convergence for F-Comment (a) (Group 2) ............ 215	

xxix

Figure 284. Assessment of Model Convergence for C-F-Comment (a) (Group 2) ........ 216	
Figure 285. Assessment of Model Convergence for U-Comment (a) (Group 2) ........... 216	
Figure 286. Assessment of Model Convergence for C-U-Comment (a) (Group 2) ....... 216	
Figure 287. Assessment of Model Convergence for TD-Post (c) (Group 2) .................. 217	
Figure 288. Assessment of Model Convergence for C-TD-Post (c) (Group 2) .............. 217	
Figure 289. Assessment of Model Convergence for TMS (Group 2)............................. 217	
Figure 290. Assessment of Model Convergence for Price (Group 2)............................. 218	
Figure 291. Assessment of Model Convergence for GT (Group 2) ............................... 218	
Figure 292. Assessment of Model Convergence for GPI (Group 2) .............................. 218	
Figure 293. Assessment of Model Convergence for CCI (Group 2) .............................. 219	
Figure 294. Assessment of Model Convergence for F-Comment (a) (Group 3) ............ 219	
Figure 295. Assessment of Model Convergence for C-F-Comment (a) (Group 3) ........ 219	
Figure 296. Assessment of Model Convergence for U-Comment (a) (Group 3) ........... 220	
Figure 297. Assessment of Model Convergence for C-U-Comment (a) (Group 3) ....... 220	
Figure 298. Assessment of Model Convergence for TD-Post (c) (Group 3) .................. 220	
Figure 299. Assessment of Model Convergence for C-TD-Post (c) (Group 3) .............. 221	
Figure 300. Assessment of Model Convergence for TMS (Group 3)............................. 221	
Figure 301. Assessment of Model Convergence for Price (Group 3)............................. 221	
Figure 302. Assessment of Model Convergence for GT (Group 3) ............................... 222	

xxx

Figure 303. Assessment of Model Convergence for GPI (Group 3) .............................. 222	
Figure 304. Assessment of Model Convergence for CCI (Group 3) .............................. 222	
Figure 305. Assessment of Model Convergence for F-Share (a) (Group 1) ................... 225	
Figure 306. Assessment of Model Convergence for C-F-Share (a) (Group 1) ............... 225	
Figure 307. Assessment of Model Convergence for U-Share (a) (Group 1) .................. 225	
Figure 308. Assessment of Model Convergence for C-U-Share (a) (Group 1) .............. 226	
Figure 309. Assessment of Model Convergence for TD-Post (c) (Group 1) .................. 226	
Figure 310. Assessment of Model Convergence for C-TD-Post (c) (Group 1) .............. 226	
Figure 311. Assessment of Model Convergence for TMS (Group 1)............................. 227	
Figure 312. Assessment of Model Convergence for Price (Group 1)............................. 227	
Figure 313. Assessment of Model Convergence for GT (Group 1) ............................... 227	
Figure 314. Assessment of Model Convergence for GPI (Group 1) .............................. 228	
Figure 315. Assessment of Model Convergence for CCI (Group 1) .............................. 228	
Figure 316. Assessment of Model Convergence for F-Share (a) (Group 2) ................... 228	
Figure 317. Assessment of Model Convergence for C-F-Share (a) (Group 2) ............... 229	
Figure 318. Assessment of Model Convergence for U-Share (a) (Group 2) .................. 229	
Figure 319. Assessment of Model Convergence for C-U-Share (a) (Group 2) .............. 229	
Figure 320. Assessment of Model Convergence for TD-Post (c) (Group 2) .................. 230	
Figure 321. Assessment of Model Convergence for C-TD-Post (c) (Group 2) .............. 230	

xxxi

Figure 322. Assessment of Model Convergence for TMS (Group 2)............................. 230	
Figure 323. Assessment of Model Convergence for Price (Group 2)............................. 231	
Figure 324. Assessment of Model Convergence for GT (Group 2) ............................... 231	
Figure 325. Assessment of Model Convergence for GPI (Group 2) .............................. 231	
Figure 326. Assessment of Model Convergence for CCI (Group 2) .............................. 232	
Figure 327. Assessment of Model Convergence for F-Share (a) (Group 3) ................... 232	
Figure 328. Assessment of Model Convergence for C-F-Share (a) (Group 3) ............... 232	
Figure 329. Assessment of Model Convergence for U-Share (a) (Group 3) .................. 233	
Figure 330. Assessment of Model Convergence for C-U-Share (a) (Group 3) .............. 233	
Figure 331. Assessment of Model Convergence for TD-Post (c) (Group 3) .................. 233	
Figure 332. Assessment of Model Convergence for C-TD-Post (c) (Group 3) .............. 234	
Figure 333. Assessment of Model Convergence for TMS (Group 3)............................. 234	
Figure 334. Assessment of Model Convergence for Price (Group 3)............................. 234	
Figure 335. Assessment of Model Convergence for GT (Group 3) ............................... 235	
Figure 336. Assessment of Model Convergence for GPI (Group 3) .............................. 235	
Figure 337. Assessment of Model Convergence for CCI (Group 3) .............................. 235	
Figure 338. Assessment of Model Convergence for F-Post (a) (Luxury) ...................... 239	
Figure 339. Assessment of Model Convergence for C-F-Post (a) (Luxury) .................. 239	
Figure 340. Assessment of Model Convergence for U-Post (a) (Luxury) ...................... 239	

xxxii

Figure 341. Assessment of Model Convergence for C-U-Post (a) (Luxury) .................. 240	
Figure 342. Assessment of Model Convergence for TD-Post (c) (Luxury) ................... 240	
Figure 343. Assessment of Model Convergence for C-TD-Post (c) (Luxury) ............... 240	
Figure 344. Assessment of Model Convergence for TMS (Luxury) .............................. 241	
Figure 345. Assessment of Model Convergence for Price (Luxury) .............................. 241	
Figure 346. Assessment of Model Convergence for GT (Luxury) ................................. 241	
Figure 347. Assessment of Model Convergence for GPI (Luxury) ................................ 242	
Figure 348. Assessment of Model Convergence for CCI (Luxury)................................ 242	
Figure 349. Assessment of Model Convergence for F-Post (a) (Non-Luxury) .............. 242	
Figure 350. Assessment of Model Convergence for C-F-Post (a) (Non-Luxury) .......... 243	
Figure 351. Assessment of Model Convergence for U-Post (a) (Non-Luxury) ............. 243	
Figure 352. Assessment of Model Convergence for C-U-Post (a) (Non-Luxury) ......... 243	
Figure 353. Assessment of Model Convergence for TD-Post (c) (Non-Luxury) ........... 244	
Figure 354. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury) ....... 244	
Figure 355. Assessment of Model Convergence for TMS (Non-Luxury) ...................... 244	
Figure 356. Assessment of Model Convergence for Price (Non-Luxury) ...................... 245	
Figure 357. Assessment of Model Convergence for GT (Non-Luxury) ......................... 245	
Figure 358. Assessment of Model Convergence for GPI (Non-Luxury)........................ 245	
Figure 359. Assessment of Model Convergence for CCI (Non-Luxury) ....................... 246	

xxxiii

Figure 360. Assessment of Model Convergence for F-Like (a) (Luxury) ...................... 247	
Figure 361. Assessment of Model Convergence for C-F-Like (a) (Luxury) .................. 248	
Figure 362. Assessment of Model Convergence for U-Like (a) (Luxury) ..................... 248	
Figure 363. Assessment of Model Convergence for C-U-Like (a) (Luxury) ................. 248	
Figure 364. Assessment of Model Convergence for TD-Post (c) (Luxury) ................... 249	
Figure 365. Assessment of Model Convergence for C-TD-Post (c) (Luxury) ............... 249	
Figure 366. Assessment of Model Convergence for TMS (Luxury) .............................. 249	
Figure 367. Assessment of Model Convergence for Price (Luxury) .............................. 250	
Figure 368. Assessment of Model Convergence for GT (Luxury) ................................. 250	
Figure 369. Assessment of Model Convergence for GPI (Luxury) ................................ 250	
Figure 370. Assessment of Model Convergence for CCI (Luxury)................................ 251	
Figure 371. Assessment of Model Convergence for F-Like (a) (Non-Luxury) .............. 251	
Figure 372. Assessment of Model Convergence for C-F-Like (a) (Non-Luxury) .......... 251	
Figure 373. Assessment of Model Convergence for U-Like (a) (Non-Luxury) ............. 252	
Figure 374. Assessment of Model Convergence for C-U-Like (a) (Non-Luxury) ......... 252	
Figure 375. Assessment of Model Convergence for TD-Post (c) (Non-Luxury) ........... 252	
Figure 376. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury) ....... 253	
Figure 377. Assessment of Model Convergence for TMS (Non-Luxury) ...................... 253	
Figure 378. Assessment of Model Convergence for Price (Non-Luxury) ...................... 253	

xxxiv

Figure 379. Assessment of Model Convergence for GT (Non-Luxury) ......................... 254	
Figure 380. Assessment of Model Convergence for GPI (Non-Luxury)........................ 254	
Figure 381. Assessment of Model Convergence for CCI (Non-Luxury) ....................... 254	
Figure 382. Assessment of Model Convergence for F-Comment (a) (Luxury) ............. 256	
Figure 383. Assessment of Model Convergence for C-F-Comment (a) (Luxury) ......... 256	
Figure 384. Assessment of Model Convergence for U-Comment (a) (Luxury) ............. 257	
Figure 385. Assessment of Model Convergence for C-U-Comment (a) (Luxury) ......... 257	
Figure 386. Assessment of Model Convergence for TD-Post (c) (Luxury) ................... 257	
Figure 387. Assessment of Model Convergence for C-TD-Post (c) (Luxury) ............... 258	
Figure 388. Assessment of Model Convergence for TMS (Luxury) .............................. 258	
Figure 389. Assessment of Model Convergence for Price (Luxury) .............................. 258	
Figure 390. Assessment of Model Convergence for GT (Luxury) ................................. 259	
Figure 391. Assessment of Model Convergence for GPI (Luxury) ................................ 259	
Figure 392. Assessment of Model Convergence for CCI (Luxury)................................ 259	
Figure 393. Assessment of Model Convergence for F-Comment (a) (Non-Luxury) ..... 260	
Figure 394. Assessment of Model Convergence for C-F-Comment (a) (Non-Luxury) . 260	
Figure 395. Assessment of Model Convergence for U-Comment (a) (Non-Luxury)..... 260	
Figure 396. Assessment of Model Convergence for C-U-Comment (a) (Non-Luxury). 261	
Figure 397. Assessment of Model Convergence for TD-Post (c) (Non-Luxury) ........... 261	

xxxv

Figure 398. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury) ....... 261	
Figure 399. Assessment of Model Convergence for TMS (Non-Luxury) ...................... 262	
Figure 400. Assessment of Model Convergence for Price (Non-Luxury) ...................... 262	
Figure 401. Assessment of Model Convergence for GT (Non-Luxury) ......................... 262	
Figure 402. Assessment of Model Convergence for GPI (Non-Luxury)........................ 263	
Figure 403. Assessment of Model Convergence for CCI (Non-Luxury) ....................... 263	
Figure 404. Assessment of Model Convergence for F-Share (a) (Luxury) .................... 265	
Figure 405. Assessment of Model Convergence for C-F-Share (a) (Luxury) ................ 265	
Figure 406. Assessment of Model Convergence for U-Share (a) (Luxury) ................... 266	
Figure 407. Assessment of Model Convergence for C-U-Share (a) (Luxury) ............... 266	
Figure 408. Assessment of Model Convergence for TD-Post (c) (Luxury) ................... 266	
Figure 409. Assessment of Model Convergence for C-TD-Post (c) (Luxury) ............... 267	
Figure 410. Assessment of Model Convergence for TMS (Luxury) .............................. 267	
Figure 411. Assessment of Model Convergence for Price (Luxury) .............................. 267	
Figure 412. Assessment of Model Convergence for GT (Luxury) ................................. 268	
Figure 413. Assessment of Model Convergence for GPI (Luxury) ................................ 268	
Figure 414. Assessment of Model Convergence for CCI (Luxury)................................ 268	
Figure 415. Assessment of Model Convergence for F-Share (a) (Non-Luxury) ............ 269	
Figure 416. Assessment of Model Convergence for C-F-Share (a) (Non-Luxury) ........ 269	

xxxvi

Figure 417. Assessment of Model Convergence for U-Share (a) (Non-Luxury) ........... 269	
Figure 418. Assessment of Model Convergence for C-U-Share (a) (Non-Luxury) ....... 270	
Figure 419. Assessment of Model Convergence for TD-Post (c) (Non-Luxury) ........... 270	
Figure 420. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury) ....... 270	
Figure 421. Assessment of Model Convergence for TMS (Non-Luxury) ...................... 271	
Figure 422. Assessment of Model Convergence for Price (Non-Luxury) ...................... 271	
Figure 423. Assessment of Model Convergence for GT (Non-Luxury) ......................... 271	
Figure 424. Assessment of Model Convergence for GPI (Non-Luxury)........................ 272	
Figure 425. Assessment of Model Convergence for CCI (Non-Luxury) ....................... 272	

xxxvii

CHAPTER 1. OVERVIEW OF DISSERTATION

The automobile is one typical example of durable goods. When purchasing durable
goods, customers are more likely to experience high-involvement decisions, meaning that
customers may engage in an extensive search for information or a comprehensive evaluation of
the choice alternatives because the wrong decision would force customers to deal with the poor
product for long periods of time. Currently, the power of digitization is challenging the business
model of the automobile industry. In particular, social media has dramatically changed the way
of how firms maintain the relationships with their customers and how customers engage an
extensive search for their vehicle purchase behaviors.
The U.S. automobile industry is increasing reliance on social media marketing and is
ranked the highest one in traditional and digital advertising spending. Despite the significant
amount of spending in digital advertising and anecdotal evidence of social media in driving up
automobile sales, the effectiveness of social media in this competitive marketplace has not
received the detailed examination.
Furthermore, in the setting of customer purchase decision-making, customers may
experience a multi-stage decision process, which typically includes the stage of awareness,
interest, and final decision. However, our understanding about the role of social media across
these different stages of decision process and its corresponding impact is still very limited.
The goal of this dissertation is to integrate the IS and marketing literature and to study the
1

mechanisms by which social media impacts customer demands in the U.S. automobile industry.
Specifically, essay one examines the dynamic relationships between firm-generated content
(FGC), user-generated content (UGC), and offline car sale in the setting of the firm-initiated
Facebook fan page and studies how different forms of format presentation, content of post, and
firm characteristics would vary these dynamic relationships. Essay two explores the dynamics of
online word-of-mouth (WOM) and its spillover effects by considering the relative effects at the
stages of customer awareness and consideration and also explores how different mechanisms at
the stage of customer awareness would change customers’ appreciation about the role of social
media at the stage of consideration and how firm characteristics would vary this dynamics.
In summary, my dissertation offers valuable insights for firms on how to better develop
their social media strategy to engage with their customers and boost offline car sales in this
economically important industry. Furthermore, these two studies would also advance the
literature by understanding further the dynamics of social media on offline sales of the durable
goods. Finally, the unique and rich data also allows me to test the underlying mechanisms at
work that will shed light on our theoretical understanding of the impact of social media from
different perspectives.

2

CHAPTER 2.
ESSAY 1: ONLINE TO OFFLINE: THE IMPACT OF SOCIAL MEDIA ON OFFLINE
SALES IN THE AUTOMOBILE INDUSTRY

2.1 INTRODUCTION
With the explosion of social media, firms have been exploring how to harness social
media and promote consumer-firm relationships, learn about customers and the marketplace, and
improve their market performance (Chen, De, & Hu, 2015; Dewan & Ramaprasad, 2014).
Among the different forms of social media, Facebook strongly leads the social networking space,
with 87% of social networkers accessing the platform regularly, followed by 43% of Instagram
users (eMarketer, 2015c). In particular, Deloitte’s recent report indicates that Facebook added
$100 billion to the U.S. economy and helped support 1 million jobs in the U.S. in 2014 (Deloitte,
2015). Due to its significant impact, Facebook has become a leading avenue for more than 54
million businesses to set up their online brand communities (i.e., fan page) for marketing
purposes (Facebook, 2015a). At a firm’s Facebook page, both firms and users can generate
contents to interact with firms or other users (hereafter termed as firm-generated content (FGC)
and user-generated content (UGC)). To influence customers’ purchase decisions, firms
deliberately engage in advertising, products/services, deals, and customer relationship (Goh,
Heng, & Lin, 2013). Simultaneously, customers are also allowed and encouraged to voice their
opinions, express their personal experiences about the firm, and form relationships with other
3

customers as well as the focal firm, thereby creating social interactions (Agarwal, Animesh, &
Prasad, 2009). Because of the simultaneous engagement of consumers and marketers on social
media, consumers’ purchase decisions are often influenced by both FGC and UGC (Goh et al.,
2013).
Despite the prevalent use of social media and extensive research in this domain, the
relationship between FGC, UGC, and firm performance has not received systematic scrutiny.
Anecdotal evidence suggests that such a relationship exists, but to the best of our knowledge, no
study directly tests for this relationship or explains the direction and magnitude of its effect.
First, previous research has documented that firm’ efforts on social media can affect marketing
outcomes (e.g., Luo, Zhang, & Duan, 2013) and that UGC can affect sales (e.g., Ghose &
Ipeirotis, 2011; Tirunillai & Tellis, 2012). However, this same literature typically examines the
isolated impact of either FGC or UGC on sales or does not distinguish between these two social
media contents. This approach neglects any effects that FGC and UGC might have on one
another, and indirectly, through one another on sales. More importantly, this isolated approach
fails to explain how firm’s media channels should operate as a system and thus scholars call for
more systematic research on the firm’s integrated media channels (e.g., Dewan & Ramaprasad,
2014; Luo et al., 2013; Smith, Gopalakrishna, & Chatterjee, 2006; Stephen & Galak, 2012).
Second, most studies in this stream focus on the impact of FGC or UGC on online sales of nondurable and media goods such as movie, DVD, music, book, or clothing (e.g., Chen et al., 2015;

4

Goh et al., 2013). Yet, as of the 4th quarter 2015, online sales only account for 7.5% of all retail
purchases (U.S. Department of Commerce, 2016). Third, previous studies focus considerably on
UGC in the form of online consumer review (e.g., Chevalier & Mayzlin, 2006; Gu, Park, &
Konana, 2012), forum (e.g., Stephen & Galak, 2012), or blog (e.g., Dewan & Ramaprasad, 2014;
Luo et al., 2013). However, FGC or UGC in the setting of the firm’s Facebook fan page
represents different aspects of goals (e.g., advertising, customer engagement, or social
networking) and could reach a variety of audiences. It, therefore, demonstrates a very different
context. Finally, the extant literature examines the impact of social media exclusively either over
a short duration of time (e.g., Chen et al., 2015; Dewan & Ramaprasad, 2014) or in a single firm
(e.g., Goh et al., 2013; Rishika, Kumar, Janakiraman, & Bezawada, 2013; Stephen & Galak,
2012). As a result, to properly understand the total impact of both FGC and UGC on offline sales
of the durable goods and vice versa, it is necessary to have an integrated perspective.
The objective of this essay is to assess the dynamic interactions between FGC, UGC, and
offline sales of light vehicles (durable goods) in the setting of the firm-initiated Facebook fan
page. The study context involves the U.S. automobile industry. The U.S. automobile industry is
ranked as the second highest in digital advertising spending (eMarketer, 2015b) and the highest
one in both traditional and digital advertising spending (Nielsen, 2015). Despite the large amount
of spending in digital advertising and the abundant anecdotal evidence indicating that social
media could drive up automobile sales (e.g., eMarketer, 2014a; MacArthur, 2015), no empirical

5

research examines the relative effectiveness of FGC and UGC on offline car sales in the setting
of the firm-initiated Facebook page. Therefore, this study aims to answer the following research
questions:
(1) Do FGC and UGC have an effect on offline car sales, after controlling for other influential
factors such as traditional media spending?
(2) What are the dynamic relationships between FGC, UGC, and offline car sales?
(3) How do these dynamic relationships vary across different forms of format presentation (link,
photo, status, and video) and content of post (informative post and sentiment of post)?
(4) How do these dynamic relationships vary for luxury versus non-luxury car brands?
To answer the above questions, I collected detailed FGC and UGC data from the official
Facebook pages of 30 car companies in the U.S. and matched these data with their offline car
sales from 2009 to 2014 in the monthly level. I also supplemented the data from these firms’
traditional media spending, Google Trends, gasoline index, consumer confidence index, and
S&P 500. This approach allows me to avoid potential bias in the estimation of the effect of FGC
and UGC on offline car sales or inferring a spurious relationship. My empirical analysis is
conducted using the panel vector autoregression (PVAR) model at three different dimensions:
overall post (volume), format presentation (link, photo, status, and video), and content of posts
(informative posts and sentiment of posts). The PVAR is suitable in our setting for several
reasons. First, UGC or FGC is generated continuously over time and it is not a discrete event

6

(Srinivasan & Hanssens, 2009). The PVAR model allows me to examine the immediate and
lagged-term effects of FGC and UGC on offline product sales (Love & Zicchino, 2006). Second,
it allows me to treat all of the major variables (i.e., FGC, UGC, and offline car sales) as jointly
endogenous, and assess the nature of bidirectional causality between all pairs of variables, in
addition to controlling for a variety of factors that can affect offline car sale. Finally, I
supplement the PVAR results with impulse response functions (IRFs) to investigate the
evolutionary pattern of the PVAR model.
My results suggest that (1) FGC is more effective in influencing offline car sales than
UGC, (2) offline car sales would attract more customers’ and firms’ attentions by disseminating
information or voicing opinions more actively to strengthen customer/customer or customer/firm
relationships and, (3) there is a positive feedback effect between FGC and UGC. These findings
vary across different forms of format presentation and content of post. Furthermore, customers in
two different groups (luxury versus non-luxury car brands) demonstrate dramatically different
patterns.
This paper makes a number of contributions to the IS and marketing literature. First, to the
best of our knowledge, this is the first academic study that rigorously examines the intricate and
distinct roles of FGC and UGC and quantifies their economic impact on consumers’ offline
commerce activities on the durable product (vehicle) at the industry level. I uncover that
customers hold different perspectives on two different forms of social media content and show

7

how customers’ offline commerce activities would influence their engagement behaviors online
and firms’ social media strategy. Thus, this work echoes and promotes the importance of taking a
systematic view of the firm’s media channels and its firm performance (e.g., Dewan &
Ramaprasad, 2014; Luo et al., 2013; Smith et al., 2006; Stephen & Galak, 2012). Second, I add
to the literature by showing that different forms of social media channels have their own
characteristics and firms should pay special attention to their resource allocations if they have a
significant presence at other social media channels. Prior research has focused exclusively on
online reviews (e.g., Gu et al., 2012), online forum (e.g., Gopinath, Chintagunta, &
Venkataraman, 2013), or blog (e.g., Dewan & Ramaprasad, 2014). I provide evidence that UGC
in our setting is not very effective in driving offline car sales. My study, in fact, also reinforces
the importance of context-specific theorizing in IS research (Hong, Chan, Thong, Chasalow, &
Dhillon, 2013).
Third, my work sheds light on customers’ evaluations of content in different formats and
shows how firms can utilize the combinations of different formats and contents to reach their
goals more efficiently. Currently, retailers are increasingly focusing on driving consumers into
their physical store locations while maintaining their online presence (i.e., online to offline
setting) (eMarketer, 2014b). Particularly in the automobile industry, companies often feature
their online automobile models online, trying to attract potential buyers into the showroom where
a test drive is often followed by a purchase (Maloney, 2000). Thus, the design of an effective

8

online format presentation becomes an extremely important issue to maintain companies’
competitive advantage. In light of my findings, companies need to reconsider the format and
content of their marketing communications in social media and strive to develop the most
appropriate conversations to engage customers and boost offline sales. My study, therefore,
responds to the recent call on online content formats in customer preferences (Berger, Matt,
Steininger, & Hess, 2015; Yi, Jiang, & Benbasat, 2015), the importance of social media contents
(e.g., Chen et al., 2015; Gopinath, Thomas, & Krishnamurthi, 2014), and customers’ social
media participation (e.g., Rishika et al., 2013). Finally, this essay serves as the first attempt to
measure how customers in two different markets (luxury versus non-luxury car brands) would
hold different views on firms’ social media strategy and opinions of other customers. This
approach provides insights for managers in two different markets to fully customize their social
media strategy. Together, my study yields interesting managerial implications for companies
interested in leveraging social media to boost their offline sales and customer engagements
online and in understating the true needs of their targeted audiences.
In the remainder of this paper, I first provide an overview of the relevant prior literature
and develop hypotheses. After elaborating on my data and their sources, I detail our PVAR
empirical specification and the estimation procedure. Then I present my results before
concluding with a discussion of the implications of my findings.

9

2.2 LITERATURE REVIEW
The current study is related to the literature that examines the effect of online contents on
marketing outcomes. The advent of social media has burgeoned understanding how the effect of
online FGC and UGC can influence sales and other marketing performance. In this section, I
briefly review the current streams of research and discuss how my study contributes to the extant
literature.
2.2.1 Effect of UGC on Sales

User generated content (UGC) on social media refers to content created by users to share
information and/or opinions with others (Tang, Fang, & Wang, 2014). Prior research focuses
considerably on the impact of online reviews, online ratings, blogs, and forums on online sales of
media and non-durable goods such as movie, book, DVD, or music. For example, Chevalier and
Mayzlin (2006) examined the effects of online reviews on the relative sales of books at
Amazon.com and Barnesandnoble.com and found a positive relationship between online book
reviews and online book sales. Focusing on movie ratings from professional and amateur
communities, Moon et al. (2010) suggest that high early movie revenues enhance subsequent
movie ratings and high advertising spending on movies supported by high ratings maximizes the
movie’s revenues. Furthermore, online video game reviews have a greater influence on the sales
of less popular games (Zhu & Zhang, 2010).
Some scholars have taken more nuanced approaches to examine the relationship between

10

online reviews and online sales. For instance, Forman et al. (2008) examine attributes of
reviewers and posit that reputable reviews have greater impact on online book sales. Ghose and
Ipeirotis (2011) find that the extent of subjectivity, informativeness, readability, and linguistic
correctness in reviews matters in Amazon’s sales. Scholars also examined other forms of UGC to
investigate their economic impacts on sales. For example, consumer blogs are considered to be a
prominent source for future Amazon album sales (Dhar & Chang, 2009). Rui et al. (2013)
suggest that the valence of the tweet, influence level of the tweeter, and the intentions expressed
by the tweeter influence movie sales. More recently, Dewan and Ramaprasad (2014) examined
the relationship between consumer blogs, traditional media, and music sales. They found that
radio play is positively related to future music sales at both the song and album levels.
2.2.2 Effect of FGC on Sales

On social media, firm generated content (FGC) refers to messages or posts made by
company representatives on an official company blog, forum, Twitter, Facebook, or other forms
of social media channels (Goh et al., 2013). Compared to UGC, firms own more control of FGC
because they can decide when, how, and what they want to distribute and communicate with
their customers (Miller & Tucker, 2013). Consistent with the findings on the impact of UGC on
sales, the extant literature generally suggests that FGC on social media positively influence their
marketing outcomes. For example, Goh et al. (2013) find that FGC in a firm’s Facebook page
influences consumers’ apparel purchase expenditures. Furthermore, Chen et al. (2015) study the

11

effect of artists’ broadcasting activities on MySpace and suggest that broadcasting in social
media has a significant effect on music sales and the effect mainly comes from personal
messages (one form of FGC on social media) rather than automated messages. In essence,
current research concludes that FGC and UGC mostly in the form of online review, blog, or
forum positively influence online sales of experience, information, and non-durable goods.
However, this literature either does not distinguish FGC from UGC (e.g., Dewan & Ramaprasad,
2014), or it tends to examine these effects independently (e.g., H. Chen et al., 2015; Gu et al.,
2012), thereby making it difficult to compare the relative effects of FGC and UGC on marketing
outcomes such as sales.
2.2.3 Relative Effects of UGC and FGC on Sales

As of yet, there is little academic research directly studying the relative effectiveness of
FGC and UGC on offline sales of the durable goods (vehicle in our case). FGC and UGC on the
firm’s Facebook pages represent a different form of online content. First, compared to online
reviews, FGC and UGC on Facebook pages may embed a variety of topics range from new
product/service announcements, deal information, after-sales services, customer loyalty,
customer complaints to customer reviews (Goh et al., 2013). Second, FGC and UGC on
Facebook may reach more audiences than online reviews (eMarketer, 2015a; Goh et al., 2013).
For example, online reviews tend to reach people who are interested in buying a certain product
or service, whereas FGC and UGC on Facebook may go beyond this audience by including the

12

firm itself, other consumers with common interests, and one’s Facebook friends and
acquaintances. Third, FGC and UGC on Facebook represent richer content than online reviews
in terms of format presentation. In the context of online reviews, reviewers usually use the text to
evaluate a certain product or service. However, in the setting of the firm’s Facebook pages, both
firms and customers usually use a mixture of photo, video, link, and status to distribute
information or express opinions. Previous research indicates that different forms of format
presentation (e.g., text, graphic, or video) do matter in influencing people’s decision-makings
(Berger et al., 2015; Lim & Benbasat, 2000; Yi et al., 2015). However, the assessment regarding
consumers’ appreciation of different formats is still under-explored in the setting of the firm’s
Facebook page (Berger et al., 2015). Finally, FGC and UGC tend to have a higher degree of
social interactivity than blogs or forums (Stephen & Galak, 2012). To conclude, due to the
uniqueness of each form of FGC and UGC, the current findings from the existing literature need
extension (see Table 1 for a summary of selected research). In line with a recent call for more
comprehensive and multifaceted research on social media (Chen et al., 2015; Stephen & Galak,
2012), I examine the joint effects of FGC and UGC on offline car sales in the setting of the
firm’s Facebook fan page.
There are some notable studies of the relative effects of FGC versus UGC. For example,
Chen and Xie (2008) used analytic models to argue that online reviews can serve as a newer
element in the marketing communication mix. They used these models to answer when and how

13

the sellers should adjust their own marketing communication in response to consumer reviews.
However, their study does not provide the relative comparison on the profit impact of consumer
reviews and traditional marketing communications. Stephen and Galak (2012) examined how
traditional media (publicity and press mentions) and social media (blog and online forum posts)
affect sales and activity in each other. However, their research setting is a nonprofit
organizational and focuses on blog and forum. Accordingly, they call for more research to
examine how firms’ media channels should operate as a system. Finally, Goh et al.’s study
(2013) is highly relevant to my research context. They studied the relative impacts of FGC and
UGC on sales in the setting of a causal wear apparel retailer’s Facebook page. However, our
study (1) explains how firms can leverage social media to affect offline sales of the durable
product, (2) focuses on the impact at the industry level, (3) captures the lagged effects when the
effects of marketing activities on sales are studied, and (4) takes the role of format presentation
and firm characteristics into account. By contract, I explicitly look at all of these different
aspects and aim to provide a comprehensive view on the impact of social media as a marketing
channel.

14

Table 1. Summary of Selected Literature
Research
Godes and Mayzlin (2004)

FGC
X

UGC
Online forum activities

Sale Metric
TV show ratings

Key Findings
UGC positively influences TV show ratings.

Chevalier and Mayzlin (2006)

X

Online reviews

Online book sales

UGC positively influences sales.

Duan et al. (2008)

X

Online reviews

Movie sales

UGC positively influences sales.

Forman et al. (2008)

X

Online reviews

Online book sales

Reviewer disclosure of identity-descriptive
information matters in sales.

Trusov et al. (2009)

X

Online referrals

Sign-ups for the online

UGC has longer carryover effects than

social networking site

traditional marketing actions.
UGC positively influences sales.

Dhar and Chang (2009)

X

Blog posts

Online music sales

Zhu and Zhang (2010)

X

Online reviews

Console and video game UGC is more influential for sales of less
sales

popular games and games whose players
have greater Internet experience.

Feng and Papatla (2011)

X

Online reviews

Car sales

UGC positively influences sales.

Ghose and Ipeirotis (2011)

X

Online reviews

Online audio player,

The extent of subjectivity, informativeness,

video player, digital

readability, and linguistic correctness in

camera, and DVD sales reviews matters in sales and perceived
usefulness.
Moe and Trusov (2011)

X

Gu et al. (2012)

X

Online ratings
Online reviews

Bath, fragrance, and

UGC is influenced by previously posted

beauty products

ratings and has the direct impact on sales.

Online digital camera

Internal UGC (Amazon) has a limited

sales

impact on sales, while external UGC
(CNET, Gpreview, and Epinions) has
a significant impact on sales.

Stephen and Galak (2012)

Firm’s blog Blog posts from Google Financial services
post

blog search

(microloans)

15

Both UGC and FGC influence sales.

Table 1 (cont’d)
Goh et al. (2013)

Firm’s posts Customers’ posts
and

Appeal sales

and comments

UGC exhibits a stronger impact than FGC
on sales.

comments
Gopinath et al. (2013)

X

Blog posts

Movie sales

UGC metrics matter in sales and the effects
Vary across pre-and post-release movie days.

Lu et al. (2013)

Firms’

Online reviews

Restaurant revenue

Both UGC and FGC influence sales.

Online reviews

Online blue-ray and

UGC positively influences sales of models of

DVD player sales

weak brands, while it does not have the impact

Online
coupons
No-Dac et al. (2013)

X

on sales of the model of strong brands.
Rui et al. (2013)

X

Twitter

Movie sales

UGC positively influences sales and
the celebrity effect does matter.

Dewan and Ramaprasad (2014)
Gopinath et al. (2014)
Chen et al. (2015)

X
X

Artists’

Blog posts from Google Music sales

UGC is not related to album sales but

blog search

negatively related to song sales.

Online forum activities

Cell phone sales

UGC positively influences sales.

Online reviews

Online music sales

Both UGC and FGC positively influence

personal

sales but there is a much weaker or no

messages

association between UGC and FGC.

16

2.3 RESEARCH MODEL AND HYPOTHESES
In contrast to the existing literature, my study is unique in the following ways. First, I
distinguish two different social media contents, FGC and UGC, and examine the dynamics of a
system of interdependent variables. This approach allows me to understand how firm’s media
channels could operate as a system (e.g., Dewan & Ramaprasad, 2014; Stephen & Galak, 2012).
Furthermore, I examine the dynamics of FGC, UGC, and offline car sales in the setting of the
firm’s initiated Facebook page. Compared to online reviews, online forums, or blog, the nature
of Facebook is dramatically different because the content represents more aspects of goals from
advertising, customer engagement, to social networking and could reach a variety of audiences
(eMarketer, 2015a; Goh et al., 2013). Thus, focusing on UGC and FGC at Facebook could shed
light on the effectiveness of social media in this stream of the literature. Finally, I focus on what
firms and consumers do online versus consumers’ commerce activities that occur in offline
settings. Particularly, the difference between non-durable products (e.g., music, movie) and
durable products (i.e., vehicle in our setting) allows me to explore the impacts of social media
from the different perspectives.
My conceptual model is shown in Figure 1. I examine these relationships at three
different dimensions: overall post (volume), format presentation, and content of post. Figure 2
shows one example of FGC and UGC. Previous research indicates that the content format (e.g.,
text, graphic, or video) does matter in influencing people’s decision-makings (Berger et al.,
17

2015; Nah, Eschenbrenner, & DeWester, 2011; Yi et al., 2015). In the online setting, the content
format is critical for firms that want to leverage online content to generate more revenues
(Berger et al., 2015; Yi et al., 2015). Therefore, I emphasize four common types of content
formats at Facebook: link, photo, status, and video (see Figure 3 for the example of FGC in these
four different forms). For the content of the post, I focus on two aspects: informative posts and
sentiment of posts. Informative posts should include details about products, promotions,
availability, price, comparisons with alternatives, and product related aspects that could be used
in optimizing the purchase decision (Abernethy & Franke, 1996; Resnik & Stern, 1977). Thus,
the relevance of this consideration reflects the common advertising practice of rational appeals
that tends to highlight product attributes (Gopinath et al., 2014). With respect to sentiment of
posts, I focus on positive and negative posts. Delineating between informative posts and
sentiment of posts is indeed grounded in the information processing domain of the existing
literature (e.g., Gopinath et al., 2014; Shiv & Fedorikhin, 1999). Finally, I am also interested in
how the interactions shown in Figure 1 vary by firm characteristics, that is, luxury versus nonluxury car brands.

18

Figure 1. Conceptual Framework

Figure 2. Example of FGC (left) and UGC (right)

19

Figure 3. Example of FGC in the form of Link, Photo, Status, and Video

The effect of FGC on sales has been well documented in prior research. For instance, to
explicitly look at online content created by artists as a means of marketing communication, Chen
et al. (2015) compared personal messages created by artists with automated messages created by
MySpace and find that personal messages are predictive of music sales. I posit a positive
relationship between FGC created on the firm-initiated Facebook page and offline car sales. In
the journey of purchasing a product, as consumers are typically averse to losses (Narayanan,
Chintagunta, & Miravete, 2007; Nelson, 1970), they may seek more product-related information
to reduce their uncertainties (Goh et al., 2013). In turn, when uncertainties regarding a product
are reduced during this purchasing journey, consumers bear more confidence in making purchase
decisions and therefore are more likely to purchase that specific product (Goh et al., 2013;
Schubert & Ginsburg, 2000). In the setting of the firm-initiated Facebook page, firms explicitly

20

and actively promote products/services and engage customer relationship to attract more visitors
and influence customers’ purchase decisions (Goh et al., 2013; Miller & Tucker, 2013). Thus, I
hypothesize:
Hypothesis 1: FGC is positively associated with offline car sales.
Ample evidence has indicated a significant impact of UGC on various marketing
outcomes. For example, the earlier studies show that activities on the online discussion forum
affect TV show ratings (Godes & Mayzlin, 2004), online reviews influence sales such as book,
music, digital camera, and DVD (Chevalier & Mayzlin, 2006; Forman et al., 2008; Gu et al.,
2012), blog posts influence music sales (Dhar & Chang, 2009). More recently, UGC is also
considered as a leading indicator of firm equity value (Lu et al., 2013; Luo & Zhang, 2013). I
expect that there is a positive association between UGC on the firm-initiated Facebook page and
offline car sales. Consumers often love to share and relate their product experiences with
members of a brand community and their participations in the brand community are critical to
brand-related purchase behavior (Algesheimer, Dholakia, & Herrmann, 2005; Rishika et al.,
2013). The firm-initiated Facebook page as one form of the online brand community provides a
platform for consumers to share their product experiences or voice personal opinions (e.g.,
customer satisfaction with a product) with other consumers and the focal firm. Therefore, the
positive effect of UGC on offline car sales is due to the WOM (word of mouth) effect, whereby
social interactions and influence between consumers affect their purchasing decision making

21

(Dewan & Ramaprasad, 2014). Accordingly, I hypothesize:
Hypothesis 2: UGC is positively associated with offline car sales.
Previous research has demonstrated that sales may trigger more FGC and UGC activities.
For example, Stephen and Galak (2012) indicate that repeat loan sales are positively associated
with the number of blog posts. Dewan and Ramaprasad (2014) suggest that album sales
positively influence album blog buzz, whereas song sales negatively influence song blog buzz. I
posit that there is a positive relationship from offline car sales to both UGC and FGC in the
setting of the firm-initiated Facebook page. The underlying principle is intuitive: for firms that
extensively leverage the Facebook pages as one of their marketing channels, an increase in
firms’ sales can raise their brand recognition and appreciation on the Facebook pages so that
both firms and customers are more likely to disseminate their products related information and
brand experiences through this channel. An increase in firms’ sales implies that firms need to
maintain or enhance their current customer-firm relationships to sustain their competitive
advantage. Therefore, it is reasonable to expect that an increase in sales will trigger firms to have
a more active and vibrant Facebook community by posting a variety of topics regularly and
actively. Such an active approach also allows customers to infer the level of a firm’s relationship
commitment and therefore strengthens customers’ bond with the firm (Rishika et al., 2013).
From the customers’ perspectives, given that they often love to relate their product experiences
with members of a brand community (Algesheimer et al., 2005), an increase in firms’ sales

22

expand the firms’ existing customer pools and therefore trigger more new and old customers to
share their experiences with other customers in a brand community. In addition, because
individuals also wish to validate their choices in an environment that further strengthens their
own affinity with the firm (Zhu, Benbasat, & Jiang, 2010), the firm-initiated Facebook page
provides a great environment to form this relationship. Thus, it is intuitively logical that an
increase in offline sales will drive more UGC. As a result, I hypothesize:
Hypothesis 3: Offline car sales are positively associated with FGC.
Hypothesis 4: Offline car sales are positively associated with UGC.
The firm-initiated Facebook page creates a channel for communicating with the firm’s
customers and building relationships with them (Miller & Tucker, 2013). An active social media
page with regular new messages/postings can help customers form more positive attitudes
toward the firm and increase the interactions between firms and customers (Rishika et al., 2013).
These increased interactions are essential to form customer identification with the firm and can
create greater trust to the firm and customer loyalty (Algesheimer et al., 2005). Prior research in
relationship marketing claims that commitment and trust are two critical components for a
party’s intentions to continue the relationship with the other part (Morgan & Hunt, 1994).
Therefore, in the setting of the firm-initiated Facebook page, if firms can interact their customers
actively by providing more products/services information, customer are more likely to get
involve in the interaction with firms as well as other customers. In addition, Miller and Tucker’s

23

study (2013) on social media management provides empirical evidence that actively managing a
Facebook page (i.e., updating status more frequently) increases UGC activities. Therefore, I
hypothesize:
Hypothesis 5: FGC is positively associated with UGC.
Previous research suggests that UGC changes sellers’ communication strategy to best
respond to those user-based reviews (Chen & Xie, 2008). I posit that UGC positively influences
FGC in the setting of the firm-initiated Facebook page. UGC is a direct expression of consumers’
personal experiences toward firms’ products or services (Goh et al., 2013) and allows firms to
better realize their customers’ preferences. In the context of the firm-initiated Facebook page, if a
large number of consumers exhibit favorable attitudes and sentiments toward a product or
service, firms are more likely to disseminate more information related to that product or service
to enhance customer relationships, advertise that product or service, and therefore boost sales.
The extant literature has recognized the importance of active social media management and
suggested that firm’s posts need to be specifically targeted to clients’ interests (Miller & Tucker,
2013). In this regard, one of the key approaches for firms to generate targeted content is through
the learning of UGC at the firm-initiated Facebook page. Thus, I hypothesize
Hypothesis 6: UGC is positively associated with FGC.

24

2.4 DATA AND EMPIRICAL METHODOLOGY

2.4.1 Research Context

My major research context is Facebook because it is the most visited social media site in
the US (eMarketer, 2015c). As a result, more than 54 million businesses have set up their online
brand communities (i.e., fan page) for marketing purposes (Facebook, 2015a). Thus, the firminitiated Facebook pages are representative sources of FGC and UGC. I selected the U.S.
automobile industry because of its economic importance and increasing reliance on social media
marketing (eMarketer, 2015b; Tang et al., 2014). The U.S. automobile industry generates sales
representing 3%-3.5% of U.S. gross domestic sales and remains one of the most important
segments in the U.S. economy (Hill, Menk, & Cooper, 2010). In response to the global economic
recession, automobile companies have turned their attention to social media marketing to
enhance customer relationships, disseminate a variety of information, engage customers, and
boost sales (Tang et al., 2014). Accordingly, the U.S. automobile industry represents a great
setting to examine the dynamics of relationships between FGC, UGC, and offline car sales.
2.4.2 Data

My samples consist of 30 major car brands in the U.S. automobile industry1. Each one of
these brands has its official Facebook fan page. I used three main data sources for our panel,

1

These car brands include Acura, Audi, BMW, Buick, Cadillac, Chevrolet, Chrysler, Dodge, Ford, Honda, Hyundai,
Infiniti, Jaguar, KIA, Lexus, Mazda, Mercedes-Benz, Mitsubishi, Nissan, Porsche, Toyota, Volkswagen, Volvo,
Lincoln, Subaru, Saab, FIAT, Jeep, Land Rover, and Scion.
25

namely, the Facebook graph API (Application Programming Interface)2 for FGC and UGC,
monthly offline car sales from the WardsAuto Premium database, and traditional media
advertising expenditure3 from Kantar Media. I also collected the consumer search volume index
in the U.S. from Google Trends to control for the popularity effects of each car make (Luo et al.,
2013). Finally, I collected some macroeconomic indicators to control for their potential impacts
on offline car sales. These indicators include the monthly gasoline price index from U.S. Bureau
of Labor Statistics, the conference board’s consumer confidence index, and S&P 500 monthly
return.
To collect FGC and UGC, I built a software tool in PHP to connect with the Facebook
graph API. I started our data collection on FGC and UGC on November 2014. It took around
three months to collect detailed information on all activities from these 30 car manufacturers’
official Facebook pages over a period of 66 months (2009.5 to 2014.10)4. Specifically, my data
on FGC and UGC included post id, post time, post type, post source (user or firm), post content,
and post link. I aggregated my data at the monthly level for two reasons: (1) the variation of firm’
social media activities at both levels was relatively small, and more importantly (2) offline car
sale data from the WardsAuto Premium database is only available at the monthly level.

2

Facebook Graph API: https://developers.facebook.com/docs/graph-api

3

Traditional media categories included network TV, Spanish-language network TV, cable TV, syndication, sport
TV, magazines, Sunday magazines, local magazines, Hispanic magazines, B-to-B magazines, national newspapers,
newspapers, Hispanic newspapers, network radio, national sport radio, local radio summary, local radio historical,
and outdoor displays.
4
My goal was to provide a historical trend of FGC and UGC. Thus, I collected data from the beginning of each
Facebook page. Most car brands initiated their Facebook pages on May 2009. Among these 30 car brands, 23 firms
started their pages in 2009, 5 firms started their pages in 2010, and 2 firms started their pages in 2011.
26

My raw dataset included 58,158 firm-generated posts and 806,705 user-generated posts. I
cleaned user-generated posts to reflect the following criteria5: keeping only messages written in
English and removing spam6. At some firms’ pages, users started to post at pages few months
after firms had initiated their pages. Given my research objective, we also removed the months
that had zero activity on user-generated posts. For example, in my dataset, one car brand initiated
its Facebook page on May 2009. However, users started to post on that page on August 2009.
Thus, the starting point for this car brand was August 2009. After clearing, my final dataset
included 58,158 firm-generated posts and 706,527 user-generated posts with a total of 1,764
firm-month observations. To conduct content analysis, I relied on the Harvard General Inquirer7,
which classifies words into 182 categories such as positive, negative, weak, etc. It has been
widely used in the literature to extract the tone and sentiment of many textual contents such as
financial reports (e.g., Loughran & McDonald, 2011), news stories (e.g., Tetlock, Saa‐
Tsechansky, & Macskassy, 2008), and review text (e.g., Shen, Hu, & Rees, 2015). Tables 2 and
3 present the definition of variables and summary statistics, respectively.
Variable
Sales
FGC-Post
FGC-Link
FGC-Photo
FGC-Status
FGC-Video
Informative-FGC

Table 2. Variable Definition

Definition
Total number of offline car sales made by one car make in month t
Total number of posts by one car make in month t
Total number of link posts by one car make in month t
Total number of photo posts by one car make in month t
Total number of status posts by one car make in month t
Total number of video posts by one car make in month t
Total number of informative posts by one car make in month t

5

I only applied the criteria to user-generated posts because firm-generated posts were always written in English and
relevant to the individual car make.
6

Users may use the firm-initiated Facebook pages to advertise products, services, or Web sites that are not relevant
to car brands.
7

Harvard General Inquirer: http://www.wjh.harvard.edu/~inquirer/
27

Pos-FGC
Neg-FGC
UGC-Post
UGC-Link
UGC-Photo
UGC-Status
UGC-Video
Informative-UGC
Pos-UGC
Neg-UGC

Table 2 (cont’d)

Total number of positive posts by one car make in month t
Total number of negative posts by one car make in month t
Total number of posts by users at one car make’s Facebook page in month t
Total number of link posts by users at one car make’s Facebook page in month t
Total number of photo posts by users at one car make’s Facebook page in month t
Total number of status posts by users at one car make’s Facebook page in month t
Total number of video posts by users at one car make’s Facebook page in month t
Total number of informative posts by users at one car make’s Facebook page in
month t
Total number of positive posts by users at one car make’s Facebook page in
month t
Total number of negative posts by users at one car make’s Facebook page in
month t
Total amount of money spent by one car make on traditional media in month t

Traditional Media
Spending (TMS)
Google Trends (GT)
The Google Trends search interest index for one car make in month t in the U.S.
Gasoline Price
The U.S. gasoline price index in month t
Index (Cao, Gedajlovic, & Zhang)
Consumer Confidence The conference board consumer confidence index in month t
Index
S&P 500 Monthly
The S&P 500 return index in month t
Return (S&P500)

Table 3. Descriptive Statistics

Variable
No. of obs.
Sale
1764
FGC-Post
1764
FGC-Link
1764
FGC-Photo
1764
FGC-Status
1764
FGC-Video
1764
Informative-FGC 1764
Pos-FGC
1764
Neg-FGC
1764
UGC-Post
1764
UGC-Link
1764
UGC-Photo
1764
UGC-Status
1764
UGC-Video
1764
Informative-UGC 1764
Pos-UGC
1764
Neg-UGC
1764
TMS
1764
GT
1764
GPI
1764
CCI
1764
S&P 500
1764

Mean
38643.63
33.53
6.61
20.76
2.54
3.19
25.61
20.31
0.92
400.53
40.36
137.92
207.12
15.21
257.66
165.52
59.65
34357912
61.33
3.41
66.34
1440.06

Std. dev.
48343.08
34.74
8.64
33.05
5.99
3.35
15.42
12.68
1.08
426.34
54.19
208.51
248.37
20.87
273.12
191.91
70.79
34516073.9
18.72
.39
12.84
288.94

Min
97
0
0
0
0
0
0
0
0
4
0
0
0
0
0
0
0
14,800
16
2.31
40.9
919.14

Max
245,239
1,042
62
973
74
20
218
170
6
3,543
487
1,910
2,512
380
2,439
3,212
1,023
197,000,000
100
3.98
94.1
2018.05

2.4.3 PVAR Model Specification

I used a PVAR model to examine the dynamic interactions between FGC, UGC, and
offline car sales. The PVAR model or the VAR (Vector Autoregression) model is suitable for
studying the relationships between a system of interdependent variables without imposing ad hoc

28

model restrictions (Adomavicius, Bockstedt, & Gupta, 2012). It has been proven to be especially
useful for examining the dynamic behavior of firms’ efforts on social media and marketing
outcomes (e.g., Chen et al., 2015; Dewan & Ramaprasad, 2014; Luo et al., 2013). Particularly in
our setting, because UGC or FGC is generated continuously over time and is not a discrete event
(Srinivasan & Hanssens, 2009), the PVAR model allows us to examine the immediate and
lagged-term effects of FGC and UGC on sales and the system of interdependent relationships.
The strengths of the PVAR model come from the benefits of the VAR model and the structure of
panel data. In a PVAR model, main variables are assumed to be endogenous, so no prior
information is required (Chen et al., 2015). It also allows the inference of bidirectional
relationships between endogenous variables and ensures robustness of the model to issues of
non-stationarity, spurious causality, endogeneity, serial correlation, and reverse causality
(Granger & Newbold, 1986). Furthermore, the long-term dynamics between endogenous
variables can be assessed through impulse response functions (H. Chen et al., 2015). The
availability of panel data allows to control for unobserved individual heterogeneity and utilizes
instruments within the model such as lagged dependent variables in the GMM (generalized
method of moments) estimation to obtain consistent estimates (Chen et al., 2015; Love &
Zicchino, 2006). It is also worth noting the limitation of PVAR. One of key limitations is the
number of endogenous variables or the number of lags. Given the structure of PVAR, as the
number of endogenous variables or the number of lags increase, the number of parameters

29

increase rapidly, leading to inefficiencies in the estimation approach (Chen et al., 2015).
Therefore, studies should proceed with caution when applying PVAR to examine the dynamic
relationships.
My PVAR model is specified (for all different dimensions) as follows:

𝑦!,! =

𝑆𝑎𝑙𝑒!,!
𝐹𝐺𝐶!,!
𝑈𝐺𝐶!,!

=

!
!!! ÎŚ!

∙

𝑆𝑎𝑙𝑒!,!!!
𝐹𝐺𝐶!,!!! + 𝛽! ∙ 𝑇𝑟𝑎𝑑!,!!! + 𝛽! ∙ 𝐺𝑜𝑜𝑔𝑙𝑒𝑇𝑟𝑒𝑛𝑑𝑠 !,!!! + 𝛽! ∙ 𝐺𝑎𝑠𝑜𝑙𝑖𝑛𝑒!!! +
𝑈𝐺𝐶!,!!!

𝛽! ∙ 𝐶𝑜𝑛𝑠𝑢𝑚𝑒𝑟!!! + 𝛽! ∙ 𝑆&𝑃500!!! + 𝜉! + 𝑓! + 𝜀!,! (1)

where 𝑦!,! = (𝑆𝑎𝑙𝑒!,! , 𝐹𝐺𝐶!,! , 𝑈𝐺𝐶!,! )′' is a three-element column vector for each car brand 𝑖 iat
time tt, containing the log and Helmert transformation of the dependent variable (see 3.4 for the
rationale of log and Helmert transformation); Φ′! are 3x3 matrices of slope coefficients for
endogenous variables; 𝑝 is the number of lags; 𝑇𝑟𝑎𝑑!,!!! is the log and Helmert transformation
of the monthly expenditure on traditional channel promotions by car make 𝑖 at time 𝑡 − 1 ;
𝐺𝑜𝑜𝑔𝑙𝑒𝑇𝑟𝑒𝑛𝑑𝑠 !,!!! GoogleTrends!,!-! is the log and Helmert transformation of the monthly
Google search index in the U.S. for car brand 𝑖 at time 𝑡 − 1; 𝐺𝑎𝑠𝑜𝑙𝑖𝑛𝑒!!! is the log and Helmert
transformation of the monthly gasoline price index at time 𝑡 − 1; 𝐶𝑜𝑛𝑠𝑢𝑚𝑒𝑟!!! is the log and
Helmert transformation of the monthly consumer confidence index at time 𝑡 − 1; 𝑆&𝑃500!!! is
the log and Helmert transformation of the monthly S&P 500 return index at time 𝑡 − 1;
𝜉! = (𝜉!,! , 𝜉!,!, 𝜉!,! ) ′ is a column vector of monthly time dummies that control for any time
!

effect such as seasonality; 𝑓! = 𝑓!,! , 𝑓!,! , 𝑓!,! is a column vector of unobserved individual
effects, characterizing car brands’ time-invariant attributes; and 𝜀!,! = (𝜀!,!,! , 𝜀!,!,! , 𝜀!,!,! ) ′
30

ε!,!,! , ε!,!,! , ε!,!,! 'is a three-element vector of errors, satisfying the assumption that Ε 𝜀!,!,! =
Ε 𝜀!,!,! 𝜀!,!,! = 0 for 𝑚 = 1, 2, 3 and 𝑡 ≠ 𝑠. To conclude, the endogenous variables include
offline sales, FGC, and UGC. I also control for exogenous variables such as traditional media
expenditure, Google Trends, gasoline price index, consumer confidence index, S&P 500 monthly
return index, and time effects.

2.4.4 PVAR Model Estimation Procedure

I followed the standard approaches to conduct our PVAR analysis: helmert
transformation, unit root tests, lag length selection, PVAR model analysis, impulse response
functions, and sample split PVAR analysis. Figure 4 shows my analysis procedure. First, to
conduct PVAR related analysis, the variables must be stationary. Given the unbalanced panel, I
conducted Fisher-Type (Choi, 2001) and Im-Pesaran-Shin (Im, Pesaran, & Shin, 2003) root unit
tests to verify the absence of unit roots. The results for Fisher-Type and Im-Pesaran-Shin tests
are shown in Tables 4 and 5, respectively. All p-values are smaller than 0.01, except for S&P
500; so, I conclude that there is no unit root in our panel except for S&P 500. I thus used the first
differences for S&P 500 as suggested by the current literature (e.g., Luo et al., 2013; Tirunillai &
Tellis, 2012) and no unit root remained in ΔS&P 500.
Table 4. Fisher-type Unit-Root Tests
Variable
Inverse chi-squared
Inverse normal
Inverse logit t
Modified inverse
swdwidjwijdiowjdiojwoidiojwiodjiowwdjiwjdiojwdiowidjiowjdijwiodjiwjdchi-squared
Sale
483.62 (0.000)
-18.24 (0.000)
-24.41 (0.000)
38.67 (0.000)
FGC-Post
380.85 (0.000)
-15.56 (0.000)
-19.19 (0.000)
29.29 (0.000)
FGC-Link
304.06 (0.000)
-13.16 (0.000)
-15.27 (0.000)
22.28 (0.000)
FGC-Photo
123.47 (0.000)
-5.08 (0.000)
-5.09 (0.000)
5.52 (0.000)
FGC-Status
417.54 (0.000)
-16.24 (0.000)
-21.03 (0.000)
32.63 (0.000)
FGC-Video
509.65 (0.000)
-18.94 (0.000)
-25.73 (0.000)
41.04 (0.000)
Informative237.45 (0.000)
-10.1 (0.000)
-11.55 (0.000) 16.19 (0.000)
31

Table 4 (cont’d)

FGC
Pos-FGC
263.62 (0.000)
-10.95 (0.000)
-12.89 (0.000)
18.59 (0.000)
Neg-FGC
725.75 (0.000)
-23.56 (0.000)
-36.63 (0.000)
60.77 (0.000)
UGC-Post
306.82 (0.000)
-12.65 (0.000)
-15.31 (0.000)
22.53 (0.000)
UGC-Link
149.47 (0.000)
-6.59 (0.000)
-6.88 (0.000)
8.17 (0.000)
UGC-Photo
342.01 (0.000)
-12.37 (0.000)
-16.89 (0.000)
25.74 (0.000)
UGC-Status
349.46 (0.000)
-13.83 (0.000)
-17.48 (0.000)
26.42 (0.000)
UGC-Video
148.92 (0.000)
-6.68 (0.000)
-6.89 (0.000)
8.12 (0.000)
Informative324.34 (0.000)
-12.88 (0.000)
-16.16 (0.000)
24.13 (0.000)
UGC
Post-UGC
389.34 (0.000)
-14.09 (0.000)
-19.48 (0.000)
30.06 (0.000)
Neg-UGC
341.64 (0.000)
-13.45 (0.000)
-17.08 (0.000)
25.71 (0.000)
TMS
597.52 (0.000)
-20.97 (0.000)
-30.16 (0.000)
49.06 (0.000)
GT
430.44 (0.000)
-16.93 (0.000)
-21.72 (0.000)
33.82 (0.000)
GPI
337.55 (0.000)
-14.21 (0.000)
-17.02 (0.000)
25.33 (0.000)
CCI
384.56 (0.000)
-16.05 (0.0000)
-19.42 (0.000)
29.62 (0.000)
ΔS&P 500
916.76 (0.000)
-27.5 (0.000)
-46.29 (0.000)
78.21(0.000)
Notes: The tests here are conducted on logged, Helmert transformed variables as suggested
by Dewan and Ramaprasad (2014). Numbers in parentheses are p-value. The null hypothesis
that all panels contain unit roots is rejected for all variables. These criteria also apply for the
Table 5.

Table 5. Im-Pesaran-Shin Unit-Root Tests
Variable
Sale
Sale
FGC-Post
FGC-Link
FGC-Photo
FGC-Status
FGC-Video
InformativeFGC
Pos-FGC
Neg-FGC
UGC-Post
UGC-Link
UGC-Photo
UGC-Status
UGC-Video
InformativeUGC
Post-UGC
Neg-UGC
TMS
GT
GPI
CCI
ΔS&P 500

Z-t-tilde-bar
-16.49 (0.000)
-16.49 (0.000)
-13.23 (0.000)
-10.78 (0.000)
-10.76 (0.000)
-14.61 (0.000)
-18.09 (0.000)
-13.81 (0.000)
-14.6 (0.000)
-24.2 (0.000)
-13.43 (0.000)
-7.72 (0.000)
-12.31 (0.000)
-14.69 (0.000)
-9.57 (0.000)
-13.79 (0.000)
-14.78 (0.000)
-14.35 (0.000)
-17.2 (0.000)
-9.22 (0.000)
-6.3 (0.001)
-10.22 (0.000)
-27.98 (0.000)

Second, I determined the appropriate lag length using a set of criteria: Akaike’s
information criterion (Zaichkowsky) (Akaike, 1969) and moment and model selection criteria
(MMSC) (Andrews & Lu, 2001). As the first step, I followed the standard approach in the VAR

32

literature using AIC (see Holtz-Eakin, Newey, & Rosen, 1988; Love & Zicchino, 2006).
Specifically, following Dewan and Ramaprasad (2014), I calculated AIC for each cross section
and took the modal value of the optimal lag length among all cross sections. The results indicate
that first-order panel VAR (lag 1) is the preferred model. Subsequently, to double check the
validity of my selection, I followed Abrigo and Love (2015) to apply the consistent MMSC for
GMM models. Andrews and Lu (2001) proposed consistent MMSC for GMM models based on
Hansen’s (1982) statistic of over-identifying restrictions. Their proposed MMSC are analogous
to various commonly used maximum likelihood-based model selection such as AIC. Tables 6 to
12 report the results of MMSC for overall posts, format presentation (link, photo, status, and
video), and content of posts (informative and sentiment), respectively. The criterion is to select
the test with the smallest MMSC-Bayesian information criterion (MBIC), MMSC-Akaike’s
information criterion, and MMSC-Hannan and Quinn information criterion (MQIC). The results
indicate that the length of lag 1 is the preferred number for each corresponding model, providing
further evidence on the legitimacy of the selection of 1 in my research setting.
Table 6. Lag Selection of Overall Post Model

Lag
CD
J
J p-value MBIC
MAIC
MQIC
1
0.999
50.05 0.059
-213.08 -21.95
-93.17
2
0.999
36.04 0.114
-161.3
-17.95
-71.37
3
0.998
22.79 0.198
-108.7
-13.21
-48.82
4
0.984
5.59 0.779
-60.18
-12.4
-30.20
Notes: CD stands for the overall coefficient of determination; J
stands for Hansen’s J chi-squared statistic.

Lag
1
2
3
4

Table 7. Lag Selection of Link Post Model
CD
0.999
0.997
0.994
-0.001

J
70.31
35.75
21.52
7.97

J p-value
0.0005
0.12
0.253
0.537

MBIC
-192.76
-161.6
-110.03
-57.81

MAIC
-1.62
-18.25
-14.47
-10.02

MQIC
-72.84
-71.66
-50.07
-27.83

33

Lag
1
2
3
4

Lag
1
2
3
4

Lag
1
2
3
4

Lag
1
2
3
4

Lag
1
2
3
4

Table 8. Lag Selection of Photo Post Model
CD
0.999
0.998
0.998
0.991

J
50.73
25.28
15.67
4.99

J p-value
0.053
0.558
0.615
0.834

MBIC
-212.4
-172.07
-115.89
-60.78

MAIC
-21.26
-28.71
-20.32
-13.00

Table 9. Lag Selection of Status Post Model
CD
0.997
0.998
0.995
0.983

J
45.33
25.39
16.11
4.19

J p-value
0.13
0.55
0.58
0.99

MBIC
-217.79
-171.95
-115.44
-61.67

MAIC
-26.66
-28.6
-19.88
-13.89

MQIC
-92.48
-82.12
-55.93
-30.80

MQIC
-97.88
-82.02
-55.48
-31.69

Table 10. Lag Selection of Video Post Model
CD
0.995
0.981
0.970
0.895

J
71.98
29.04
13.05
5.43

J p-value
0.0003
0.359
0.788
0.795

MBIC
-191.14
-168.31
-118.51
-60.35

MAIC
-0.015
-24.96
-22.94
-12.56

MQIC
-71.23
-78.37
-58.55
-30.37

Table 11. Lag Selection of Informative Post Model
CD
0.999
0.999
0.998
0.996

J
42.37
26.17
14.2
8.94

J p-value
0.215
0.509
0.72
0.442

MBIC
-220.75
-171.18
-117.36
-56.83

MAIC
-29.62
-27.83
-21.79
-9.05

MQIC
-100.84
-81.24
-57.40
-26.86

Table 12. Lag Selection of Sentiment Post Model
CD
0.999
0.998
0.996
0.976

J
J p-value
119.76 0.08
60.42 0.89
37.28 0.91
14.92 0.94

MBIC
-611.16
-487.76
-328.17
-167.8

MAIC
-80.23
-89.75
-62.71
-35.07

MQIC
-278.06
-237.94
-161.63
-84.53

To conduct PVAR analysis, I performed two transformations on all variables: natural log8
and Helmert transformation. I took the natural log to remove the scaling effects to improve the
model fit. Second, although introducing the fixed effects allows us to ensure the underlying
structure the same by allowing for individual heterogeneity, these fixed effects are correlated
with the regressors due to the lags of the dependent variables (Arellano & Bover, 1995; Dewan

8

Since some FGC and UGC have zero observations, I added 1 to them before the log transformation.

34

& Ramaprasad, 2014; Love & Zicchino, 2006). To avoid this problem, I used forward-mean
differencing, also referred to as the Helmert transformation (see Arellano & Bover, 1995). This
transformation removes the mean of all the future observations available for each and thus
ensures orthogonality between the forward-differenced variables and their lagged values (Love
& Zicchino, 2006). Accordingly, to address the issue of simultaneity, I can use the lagged
regressors as instruments and estimate the coefficient by the system GMM estimator (Arellano &
Bover, 1995; Love & Zicchino, 2006). Also, the use of forward orthogonal deviations does not
induce autocorrelation in the error terms and frees me from serial correlation (Drakos &
Konstantinou, 2014). Finally, I supplemented our PVAR analysis with the impulse response
functions (IRFs) to obtain the evolutionary patterns of our interested relationships. Together
PVAR and IRFs allow me to gain a comprehensive understanding of the relationships between
FGC, UGC, and offline car sales. I then applied the same PVAR analysis for my sample split
analysis.

Figure 4. Analysis Procedure

35

2.5 RESULTS
2.5.1 Main PVAR Analysis Results
2.5.1.1 Effect of Overall Posts

Table 13 shows the estimation results for overall posts (volume). In the Sales equation,
the coefficient estimate on F-Post at lag 1 is positive (0.032) and statistically significant at the
0.1% level, indicating that FGC at month t-1 positively influences offline car sales at month t and
therefore H1 is supported. On the contrary, the coefficient estimate on U-Post is negative (0.018) and insignificant at the 5% level, thereby rejecting H2. The results from FGC and UGC at
the overall post level are the important findings of this study and can be explained by the nature
of the durable product and Facebook.
Compared to media and non-durable products (e.g., book, DVD, or music), the vehicle is
a durable product. As a result, exposure to UGC may not have a direct impact on offline sales of
the durable product because customers may need to invest additional effort before making their
purchase decisions. In addition, the nature of Facebook may also matter. Although online review
sites and Facebook are all considered social media channels, the purpose of these two platforms
is dramatically different. The review site such as Amazon’s online review provides an
environment for current and potential customers to share their usage experience and help other
consumers to make better purchase decisions (Forman et al., 2008). On the other hand, the firm’s
Facebook fan page provide an environment for customers to express their loyalty, receive the

36

latest information from the firm, and interact with other customers, namely, social networking.
Therefore, the various topics on UGC may result in the insignificant effect of UGC on offline car
sales. The results from FGC and UGC also imply that firm-initiated Facebook pages play
different roles in terms of predicting offline car sales from firms’ and customers’ perspectives. In
addition, the results also appear that offline car sales are positively related to FGC, whereas
offline car sales do not have any discernible association with UGC, thereby supporting H3 and
rejecting H4. Finally, in the U-Post equation, the coefficient estimate on F-Post is positive (0.01)
and insignificant at the 5% level, rejecting H5. Furthermore, I also cannot find the positive effect
of UGC on FGC, thus rejecting H6.
Table 13. Overall Post PVAR Regression Results
IV
Salesi, t-1
F-Posti, t-1
U-Posti, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Posti, t
0.574*** (0.059) 0.727*** (0.188)
0.032***
(0.008) 0.701***
(0.032)
-0.018+ (0.01)
0.049+ (0.028)
0.053***
(0.015) 0.091** (0.034)
-0.11+ (0.065)
0.059 (0.177)
0.145** (0.049) 0.038 (0.155)
0.149** (0.038) -0.247*+ (0.102)
0.248* (0.1)
-0.576 (0.3)

U-Posti, t
-0.62 (0.177)
0.01 (0.022)
0.787*** (0.032)
0.108** (0.036)
0.048 (0.193)
0.21 (0.154)
-0.181+ (0.106)
-0.245 (0.289)

Notes: IV stands for the independent variable. Numbers in parentheses are standard errors.
Time fixed effects are included in the estimation, but the coefficient estimates are not
shown to conserve space. This applies to Tables 7 to 19. + p<0.1; * p<0.05; ** p<0.01; ***
p<0.001

2.5.1.2 Effect of Format Presentation

To further explore consumers’ appreciation on content in different formats and the
individual assessment of each format, I differentiated total numbers of posts into four major
categories of format presentation at Facebook (link, photo, status, and video), reported in Tables
14 to 17. Table 14 shows the results for posts in the form of link. The results suggest that both

37

FGC and UGC in the form of link are not effective in driving offline car sales, rejecting both H1
and H2. Furthermore, results support H3 while H4 has to be rejected, suggesting that offline car
sales positively influence FGC in the form of link. Finally, I find the support for both H5 and H6,
suggesting that there is a positive feedback effect between FGC and UGC in the form of link.
Table 14. Format Presentation: Link PVAR Regression Results
IV
Salesi, t-1
F-Linki, t-1
U-Linki, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Linki, t
0.558*** (0.122) 1.61* (0.82)
0.008 (0.006)
0.712*** (0.038)
-0.004 (0.007) 0.124**
(0.042)
0.084*** (0.02) -0.178+ (0.104)
-0.107 (0.128) 1.497 (0.916)
0.171* (0.088) -1.161* (0.627)
0.249*** (0.07) -1.677** (0.482)
0.449*** (0.126) -2.29** (0.823)

U-Linki, t
1.127 (0.89)
0.09** (0.035)
0.82*** (0.042)
0.234*
(0.117)
1.65+ (0.969)
-0.234 (0.573)
0.064 (0.456)
-0.951 (0.805)

Tables 15 and 16 show the results for posts in the form of photo and status, respectively.
These results indicate similar patterns even though they show different trends compared to the
results in Table 14, suggesting that different forms of format presentation have different impacts
in influencing offline car sales and customer engagements. For example, FGC in the form of
photo and status is effective in predicting offline car sales with the stronger effect via FGC in the
form of status, in support of H1, while H2 is rejected. I also find evidence that offline car sales
are positively associated with FGC, confirming H3, while offline car sales do not positively
influence UGC. Finally, there is no positive feedback effect between FGC and UGC in the form
of photo and status, rejecting H5 and H6.
Table 15. Format Presentation: Photo PVAR Regression Results
IV
Salesi, t-1
F-Photoi, t-1
U-Photoi, t-1
TMSi, t-1

Dependent variable
Salesi, t
F-Photoi, t
0.523*** (0.081) 1.48** (0.49)
0.01** (0.005) 0.77***
(0.03)
-0.003 (0.006) -0.06+ (0.038)
0.051** (0.015) 0.084 (0.071)
38

U-Photoi, t
0.381 (0.327)
-0.002 (0.018)
0.79*** (0.029)
0.164** (0.049)

Table 15 (cont’d)
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

-0.13+ (0.076)
1.2* (0.543)
0.18*** (0.053) 0.807* (0.37)
0.166*** (0.048) -0.207 (0.265)
0.301** (0.104) -0.601 (0.651)

0.482+(0.297)
0.396 (0.228)
-0.182 (0.18)
0.647 (0.438)

Table 16. Format Presentation: Status PVAR Regression Results
IV
Salesi, t-1
F-Statusi, t-1
U-Statusi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Statusi, t
0.448** (0.113) 1.26* (0.51)
0.022** (0.007) 0.66*** (0.033)
-0.014 (0.011) -0.009 (0.041)
0.09***
(0.019) 0.007 (0.06)
-0.198+ (0.118) 0.103 (0.43)
0.222** (0.075) -0.74* (0.33)
0.317*** (0.074) -1.48*** (0.336)
0.427** (0.131) -2.38*** (0.59)

U-Statusi, t
0.48 (0.454)
-0.01 (0.031)
0.7*** (0.052)
0.064 (0.06)
1.19* (0.467)
0.34 (0.31)
-0.67* (0.291)
-0.63 (0.479)

Table 17 shows the results for posts in the form of video. In this relationship, I only
observe that that user-generated videos positively influence the number of firm-generated videos,
supporting my H6. Together, we see that the dynamic relationships shown in Figure 1 vary
across different forms of format presentation, implying that firms need to leverage different
combinations of formats to maximize the utility of each post. Specifically, FGC in the form of
photo and status drives consumptions. However, UGC in any form of format presentation does
not influences offline car sales, contrary to the general belief that UGC is effective in influencing
sales. In addition, the positive feedback between UGC and FGC only exits in the form of link. I
also observe that UGC in the form of video positively influences FGC in the form of video.
Finally, offline car sales positively influence FGC in the form of link, photo, and status.
Table 17. Format Presentation: Video PVAR Regression Results
IV
Salesi, t-1
F-Videoi, t-1
U-Videoi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Videoi, t
0.6*** (0.07)
0.303 (0.34)
0.008 (0.006)
0.466*** (0.02)
-0.002 (0.006) 0.09** (0.03)
0.06**+ (0.022) 0.065 (0.06)
-0.186 (0.101) -0.637 (0.46)
0.159* (0.062) -0.032 (0.35)
0.161*** (0.043) -0.46* (0.213)
0.442** (0.11) -0.175 (0.53)

39

U-Videoi, t
0.46 (0.37)
0.013 (0.027)
0.832*** (0.038)
0.098 (0.072)
-0.53 (0.48)
0.22 (0.42)
-0.02 (0.218)
-0.725 (0.598)

2.5.1.3 Effect of Content

To investigate what kinds of content matters in influencing offline car sales and customer
engagements, I followed prior research to use the keyword approach to conduct our content
analysis (e.g., Chen et al., 2015; Shen et al., 2015). Specifically, I used the frequency of
keywords to measure if a post belongs to our interested categories. The underlying rationale is
that when a certain keyword appears, I can use that keyword to classify the topics discussed in
posts. I relied on nine categories of the Harvard General Inquirer to identify informative posts:
“exch”, “place”, “goal”, “know”, “quality”, “quan”, “numb”, “space”, and “compare”9. For
sentiment analysis, I leveraged “positive” and “negative” categories from the Harvard General
Inquirer. Table 18 shows the examples of these posts. Tables 19 and 20 show the results of these
two aspects respectively. The results from informative posts (see Table 19) only support our H1
and H3, suggesting that firm’s attribute-based posts positively influence customers’ offline
commerce activities and vice versa. I now turn my attention to the sentiment of posts. Given that
I have more than one variable in each category (e.g., positive FGC and negative FGC), I only
present our findings here instead of examining each hypothesis described above.
Table 18. Examples of Content
Category
Informative FGC

Sample Messages
Our Project Driveway program featuring Equinox fuel cell EVs powered
by hydrogen passed 1 million miles this week. No other automaker comes
close to the number of miles we’ve driven in real world conditions with real
people driving these vehicles. Check out this Good Morning America
segment from today’s show. (Chevrolet, 2009.9)

Informative UGC

Would be nice if Ford offered this to other C-max as a retrofit. Shouldn’t be
to hard fasten on top and Run a wire to the batteries. Charge a few hundred bucks

9

Categories of Harvard General Inquirer: http://www.wjh.harvard.edu/~inquirer/homecat.htm
40

Table 18 (cont’d)

to do it (Ford’s user, 2014.1).
Positive FGC

The Mustang can inspire powerful moments and create memories that are as
unique as you are. Capture your moment in a photo and post it to Instagram using
the #MustangInspires hashtag bit.ly/1eokGYK. (Ford, 2013.10)

Negative FGC

We want to let you know that we are conducting a voluntary safety recall on the
2010 model year Lexus HS 250h vehicles to update software in the anti-lock brake
system. For more information, please visit Our Newsroom
at http://bit.ly/ds2BRy. (Lexus, 2010.2)s

Positive UGC

Got this yesterday. I LOVE YOU HONDA. (Honda’s user, 2013.5)

Negative UGC

I have a 2010 White Toyota Sienna. The Van has dulled out spots all over it. The
problem was noticed about 6 months after we bought the van brand in 2010.
What can I do to get these dull spots out. It looks awful. Never had this problem
with other vehicles. (Toyota’s user, 2014.6)

In the Sales equation, the results appear that positive FGC positively affects offline car
sales and negative UGC negatively affects offline car sales. In the Pos-F equation, I find that
both offline car sales and negative FGC positively influence positive FGC. Furthermore, in the
Neg-F equation, offline car sales, positive FGC, and negative UGC positively influence negative
FGC, while positive UGC negatively affects the number of negative FGC. Finally, I observe that
the more positive UGC at time t-1, the more negative UGC at time t. This is a very interesting
finding and suggests the importance of the firm’s effective control and management on social
media (Miller & Tucker, 2013). Particularly, given my finding, firms could benefit of more
positive UGC, on one hand, and these positive UGC could backfire by triggering more negative
UGC, on the other hand. Although firms’ Facebook pages provide a good channel for customers
to voice opinions, firms should pay special attention to what customers really talk about or
address customers’ complaints in a timely manner. For example, firms should contact users with
negative experience to figure out the issues of their products/services more effectively or provide
some incentives for users who show concerns on firms’ products/services. By these approaches,
41

firms could fully leverage social media to trigger more positive UGC and reduce negative UGC
simultaneously.
Table 19. Informative Post PVAR Regression Results
Dependent variable
Salesi, t
Info-Fi, t
0.42*** (0.084) 1.0*** (0.25)
0.027* (0.01)
0.68*** (0.03)
-0.014 (0.013) 0.02 (0.03)
0.079*** (0.017) 0.085* (0.038)
-0.17 (0.112)
0.28 (0.263)
0.179* (0.073) 0.005 (0.193)
0.267*** (0.051) -0.4* (0.11)
0.51*** (0.126) -0.31 (0.32)

IV
Salesi, t-1
Info-Fi, t-1
Info-Ui, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Info-Ui, t
-0.03 (0.241)
0.005 (0.028)
0.69*** (0.042)
0.087* (0.04)
0.35 (0.29)
0.63** (0.192)
-0.15 (0.13)
-0.03 (0.321)

Table 20. Sentiment PVAR Regression Results
IV
Salesi, t-1

Salesi, t
0.527*** (0.052)

Dependent variable
Pos-Fi, t
Neg-Fi, t
0.416** (0.142) 0.371* (0.418)

Pos-Ui, t
0.31+ (0.166)

Neg-Ui, t
-0.008 (0.169)

0.022**
(0.008)
-0.01++ (0.006)
0.012 (0.012)
-0.02* (0.011)
0.078*** (0.013)
-0.17* (0.07)
0.11* (0.048)
0.234*** (0.034)
0.525*** (0.096)

0.689*** (0.025)
0.036* (0.017)
0.012 (0.03)
-0.007 (0.028)
0.18*** (0.031)
0.088 (0.183)
0.171 (0.135)
-0.156+ (0.091)
0.09 (0.269)

-0.028+ (0.024)
0.032 (0.019)
0.777*** (0.037)
0.005 (0.032)
0.17*** (0.042)
0.371 (0.238)
0.091 (0.157)
-0.304** (0.107)
-0.038 (0.293)

-0.001 (0.025)
0.014 (0.019)
0.183*** (0.037)
0.603*** (0.63
0.115** (0.039
0.08+(0.227)
0.26 (0.15)
0.024 (0.105)
-0.25 (0.293)

Table 20 (cont’d)

Pos-Fi, t-1
Neg-Fi, t-1
Pos-Ui, t-1
Neg-Ui, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

0.081** (0.026)
0.117*** (0.02)
-0.082* (0.036)
0.121***
(0.036)
-0.067+ (0.043)
0.247 (0.248)
-0.5** (0.178)
-0.171 (0.133)
-0.384 (0.342)

2.5.2 Sample Split PVAR Analysis Results

I next explore how the relationships in Figure 1 vary for firm characteristics (luxury
brands versus non-luxury brands)10. Descriptive statistics for luxury and non-luxury are shown in
Tables 21 and 22, respectively. The PVAR results for luxury versus non-luxury are reported in
Tables 23 to 36.
Table 21. Descriptive Statistics for Luxury Car Makes

Variable
Sale
FGC-Post
FGC-Link
FGC-Photo

No. of obs.
787
787
787
787

Mean
11429.27
30.78
5.48
19.36

Std. dev.
8620.54
24.78
7.33
24.69

10

Min
97
0
0
0

Max
43981
436
51
422

Luxury: Acura, Audi, BMW, Buick, Cadillac, Infiniti, Jaguar, Lexus, Mercedes-Benz, Porsche, Volvo, Lincoln,
Saab, and Land Rover. Non-Luxury: Chevrolet, Chrysler, Dodge, Ford, Honda, Hyundai, KIA, Mazda, Mitsubishi,
Nissan, Toyota, Volkswagen, Subaru, FIAT, Jeep, and Scion.
42

Table 21 (cont’d)

FGC-Status
FGC-Video
Informative-FGC
Pos-FGC
Neg-FGC
UGC-Post
UGC-Link
UGC-Photo
UGC-Status
UGC-Video
Informative-UGC
Pos-UGC
Neg-UGC
TMS
GT
GPI
CCI
S&P 500

787
787
787
787
787
787
787
787
787
787
787
787
787
787
787
787
787
787

2.14
3.58
3.35
3.30
23.86
13.14
19.1
10.53
0.93
1.06
246.52
212.78
31.91
46.23
80.64
105.41
122.79
113.77
11.32
19.51
141.93
108.30
97.39
135.36
33.77
51.17
17803196.95 12804924.54
65.47
17.28
3.43
0.38
66.47
12.82
1444.49
286.04

0
37
0
17
0
81
0
74
0
5
4
1557
0
353
0
1422
0
977
0
380
6
879
3
3212
0
879
14800
75289900
23
100
2.59
3.98
40.9
94.1
919.32
2018.05

Table 22. Descriptive Statistics for non-Luxury Car Makes

Variable

No. of obs.

Sale
FGC-Post
FGC-Link
FGC-Photo
FGC-Status
FGC-Video
Informative-FGC
Pos-FGC
Neg-FGC
UGC-Post
UGC-Link
UGC-Photo
UGC-Status
UGC-Video
Informative-UGC
Pos-UGC
Neg-UGC
TMS
GT
GPI
CCI
S&P 500

977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977
977

Mean

60565.53
35.75
7.53
21.89
2.85
3.05
27.02
21.28
0.91
524.58
47.16
184.06
275.04
18.35
350.87
220.39
80.49
47693184.08
57.98
3.40
66.25
1436.49

Std. dev.

Min

55528.65
40.92
9.48
38.47
7.38
3.38
16.9
14.11
1.09
507.28
58.99
24558
301.08
21.41
325.28
212.17
77.27
40253712.69
19.18
0.4
12.87
291.35

500
0
0
0
0
0
0
0
0
4
0
0
0
0
0
0
0
238700
16
2.31
40.9
919.14

Max
245239
1042
62
973
74
20
218
170
6
3543
487
1910
2512
182
2439
1512
654
197000000
100
3.98
94.1
2018.05

2.5.2.1 Effect of Overall Posts

Tables 23 and 24 shows results for overall posts for the luxury and non-luxury groups,
respectively. Consistent with my main results (see Table 4), FGC shows a positive impact on
offline cars sales for both luxury and non-luxury groups, although the magnitude of the
coefficient for the luxury group is larger than the corresponding coefficient for the non-luxury
group. I also observe that offline car sales also positively influence FGC for both groups and that

43

the magnitude of the coefficient for the non-luxury group is larger than the corresponding
coefficient for the luxury group. Finally, FGC positively influence the volume of UGC for the
luxury group, whereas it does not have any impact on UGC for the non-luxury group.
Table 23. Overall PVAR Regression Results (Luxury)
IV
Salesi, t-1
F-Posti, t-1
U-Posti, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Posti, t
0.444*** (0.073) 0.453*** (0.183)
0.04** (0.013) 0.67*** (0.045)
-0.008 (0.0014) 0.012 (0.037)
0.096*** (0.018) 0.089* (0.039)
-0.304 (0.209) 1.152*
(0.516)
0.008 (0.073)
0.371++ (0.203)
0.285*** (0.05) -0.24 (0.132)
0.653*** (0.155) 0.078 (0.423)

U-Posti, t
-0.054 (0.214)
0.109** (0.032)
0.689*** (0.044)
0.035 (0.041)
-0.636 (0.531)
0.143+ (0.197)
-0.24 (0.138)
-0.527 (0.402)

Table 24. Overall PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
F-Posti, t-1
U-Posti, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Posti, t
0.583*** (0.07) 0.801** (0.238)
0.024* (0.009) 0.75*** (0.044)
-0.017 (0.0012) 0.059 (0.04)
0.047** (0.021) 0.09 (0.081)
-0.08 (0.068)
0.35 (0.216)
0.187* (0.077) -0.148 (0.279)
0.186*** (0.041) -0.33* (0.14)
0.413** (0.139) -1.17* (0.442)

U-Posti, t
-0.359 (0.22)
-0.007 (0.029)
0.791*** (0.04)
0.205* (0.085)
0.201 (0.193)
0.329 (0.215)
0.124 (0.126)
0.036 (0.366)

2.5.2.2 Effect of Format Presentation

Tables 25 and 26 shows the results of the sample split for posts in the form of link for
two groups, respectively. The findings for the non-luxury group are consistent with our main
results shown in Table 14 while there are some differences for the luxury group.
Table 25. Link PVAR Regression Results (Luxury)
IV
Salesi, t-1
F-Linki, t-1
U-Linki, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Linki, t
0.596*** (0.062) 0.285 (0.246)
-0.002 (0.007) 0.815*** (0.026)
0.001 (0.007)
0.053* (0.025)
0.093*** (0.015) 0.041 (0.051)
-0.022 (0.142) 0.057 (0.515)
0.11 (0.071)
-0.791** (0.286)
0.16** (0.051) -0.708** (0.215)
0.612*** (0.139) -2.779*** (0.536)

44

U-Linki, t
0.714** (0.252)
0.064* (0.029)
0.908*** (0.03)
-0.086 (0.053)
-0.635 (0.598)
0.729* (0.348)
-0.039 (0.218)
-0.811 (0.586)

Table 26. Link PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
F-Linki, t-1
U-Linki, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Linki, t
0.602*** (0.086) 1.14* (0.469)
0.008 (0.006)
0.71*** (0.038)
-0.004 (0.006) 0.166***
(0.042)
0.04 (0.026)
-0.296+ (0.155)
0.013 (0.066)
-0.372 (0.42)
0.217* (0.096) -1.67** (0.55)
0.212*** (0.06) -1.55*** (0.379)
0.524*** (0.138) -2.217* (0.96)

U-Linki, t
0.277 (0.417)
0.059* (0.028)
0.933*** (0.03)
0.263* (0.141)
0.024 (0.341)
-0.879+ (0.493)
0.32 (0.296)
0.482 (0.711)

Tables 27 and 28 present the results of the sample split for posts in the form of photo for
two groups. The results for the non-luxury group are consistent with major results (Table 15).
Interestingly, for the luxury group, the findings indicate that both FGC and UGC positively
influence offline car sales and the impact of UGC on sales is stronger than the impact of FGC on
sales, suggesting that for customers in the luxury group, they rely more on other customers’
opinions in the form of photo to make purchase decisions than firms’ activities. However, for the
non-luxury group, we cannot observe this relationship, implying that customers in different
groups focus on different aspects when making their purchase decisions. Furthermore, I find that
although offline car sales are positively associated with FGC for both groups, the magnitude of
the coefficient for the non-luxury group is larger than the corresponding coefficient for the
luxury group.
Table 27. Photo PVAR Regression Results (Luxury)
IV
Salesi, t-1
F-Photoi, t-1
U-Photoi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Photoi, t
0.549*** (0.052) 0.398*** (0.17)
0.01* (0.005)
0.818*** (0.029)
0.017** (0.006) -0.015 (0.027)
0.102***
(0.013) -0.004 (0.037)
-0.152+ (0.092) 0.527 (0.349)
-0.065 (0.047) 0.658** (0.235)
0.169*** (0.033) 0.134 (0.14)
0.231* (0.114) 0.319 (0.511)

45

U-Photoi, t
0.198 (0.185)
-0.007 (0.023)
0.823*** (0.03)
0.211*** (0.034)
-0.116 (0.292)
0.589** (0.212)
-0.201 (0.14)
1.937*** (0.465)

Table 28. Photo PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
F-Photoi, t-1
U-Photoi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Photoi, t
0.624*** (0.078) 1.00* (0.365)
0.013* (0.005) 0.835*** (0.037)
-0.003 (0.009) -0.001 (0.043)
0.001 (0.019)
0.049 (0.109)
-0.163* (0.078) 1.05** (0.393)
0.178* (0.073) 0.657 (0.416)
0.093* (0.044) -0.275 (0.232)
0.21 (0.141)
-0.177 (0.722)

U-Photoi, t
0.2 (0.343)
0.028 (0.023)
0.78*** (0.05)
0.011 (0.114)
0.674* (0.343)
0.474 (0.331)
-0.225 (0.189)
0.36 (0.559)

The results for posts in the form of status for two groups, shown in Tables 29 and 30, also
indicate interesting patterns. For example, I find that FGC positively influence offline sales of
the non-luxury group, whereas FGC is not effective in driving offline car sales of the luxury
group. I also observe that offline car sales positively influence FGC for both groups, although the
magnitude of the coefficient for the non-luxury group is stronger than the corresponding
coefficient for the luxury group. Finally, FGC positively influence UGC for the luxury group,
whereas FGC does not have any impact on UGC for the non-luxury group.
Table 29. Status PVAR Regression Results (Luxury)
IV
Salesi, t-1
F-Statusi, t-1
U-Statusi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Statusi, t
0.591*** (0.073) 1.001*** (0.278)
0.003 (0.007)
0.74*** (0.032)
-0.007 (0.001) 0.034 (0.039)
0.097*** (0.018) -0.059 (0.05)
-0.241 (0.179) -0.771 (0.68)
0.084 (0.061)
-0.806** (0.253)
0.243** (0.055) -1.101*** (0.212)
0.52*** (0.152) 0.743 (0.551)

U-Statusi, t
-0.014 (0.195)
0.051* (0.025)
0.741*** (0.043)
0.029 (0.042)
1.253**
(0.496)
0.342+ (0.176)
-0.359* (0.157)
-0.231 (0.442)

Table 30. Status PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
F-Statusi, t-1
U-Statusi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Salesi, t
0.55*** (0.101)
0.016*
(0.007)
-0.02+ (0.011)
0.06* (0.03)
0.011 (0.069)
0.269** (0.088)
0.2*** (0.065)
0.45*** (0.144)

Dependent variable
F-Statusi, t
1.76* (0.85)
0.58*** (0.066)
0.1 (0.096)
0.145 (0.256)
0.202 (0.574)
-1.31 (0.851)
-1.87*** (0.66)
-4.44** (1.64)

46

U-Statusi, t
0.337 (0.47)
0.007 (0.035)
0.73*** (0.06)
0.116 (0.137)
0.491 (0.426)
0.422 (0.426)
-0.568+ + (0.31)
-1.11 (0.65)

I find another interesting patterns for posts in the form of video, as shown in Tables 31
and 32. For example, the results from the non-luxury group are consistent with the main results,
supporting H6. However, for the luxury group, I observe that there is a positive feedback effect
between UGC and FGC, supporting both H5 and H6. In addition, offline car sales also positively
influence UGC, supporting H4.
Table 31. Video PVAR Regression Results (Luxury)
IV
Salesi, t-1
F-Videoi, t-1
U-Videoi, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
F-Videoi, t
0.529*** (0.068) 0.269 (0.324)
0.004 (0.007)
0.452*** (0.037)
-0.014 (0.009) 0.103* (0.041)
0.095*** (0.017) 0.076 (0.059)
-0.008 (0.125) 0.049 (0.604)
0.191* (0.008) -0.252 (0.407)
0.247*** (0.049) -0.618** (0.227)
0.532*** (0.154) -0.948 (0.644)

U-Videoi, t
0.766* (0.341)
0.081* (0.037)
0.798*** (0.043)
-0.062 (0.061)
-0.773 (0.625)
0.562 (0.422)
-0.118+(0.235)
-1.268 (0.656)

Table 32. Video PVAR Regression Results (Non-Luxury)
Dependent variable
IV
Salesi, t
F-Videoi, t
U-Videoi, t
Salesi, t-1
0.714*** (0.069) 0.26 (0.369)
0.27 (0.355)
F-Videoi, t-1
0.008 (0.005)
0.535*** (0.033) 0.042 (0.034)
U-Videoi, t-1
-0.002 (0.007) 0.085* (0.042)
0.84*** (0.047)
TMSi, t-1
0.017 (0.02)
0.151 (0.098)
0.315* (0.129)
GTi, t-1
-0.008 (0.06)
-0.345 (0.338)
-0.088 (0.373)
GPSt-1
0.08 (0.008)
-0.033 (0.491)
0.074 (0.55)
CCIt-1
0.106* (0.042) -0.281 (0.255)
0.093+ (0.247)
ΔS&P500t-1
0.349** (0.121) 0.372 (0.652)
1.46 (0.79)

2.5.2.3 Effect of Content

Tables 33 and 34 shows the results of the sample split for informative posts. Consistent
with major results (Table 19), there is a positive feedback effect between informative FGC and
offline car sales (H1 and H3). In particular, in the Sales equation, the magnitude of the
coefficient from Info-F for the luxury group is larger than the corresponding coefficient for the
non-luxury group. On the other hand, in the Info-F equation, the magnitude of the coefficient
47

from Sales for the non-luxury group is larger than the corresponding coefficient for the luxury
group.
Table 33. Informative Post PVAR Regression Results (Luxury)
IV
Salesi, t-1
Info-Fi, t
Info-Ui, t
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Dependent variable
Salesi, t
Info-Fi, t
0.41*** (0.072) 0.45* (0.211)
0.043* (0.016) 0.71*** (0.039)
-0.021 (0.017) -0.043 (0.045)
0.121*** (0.018) 0.092* (0.039)
0.012 (0.266)
0.097 (0.75)
0.097 (0.085)
0.6* (0.24)
0.285*** (0.053) -0.33* (0.149)
0.6*** (0.164) 0.171 (0.42)

Info-Ui, t
-0.21 (0.18)
0.035 (0.035)
0.68***
(0.053)
0.069+ (0.037)
-0.5 (0.65)
0.756*** (0.21)
-0.19 (0.136)
-0.592 (0.4)

Table 34. Informative Post PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
Info-Fi, t
Info-Ui, t
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Salesi, t
0.51*** (0.07)
0.027** (0.009)
-0.019 (0.012)
0.042* (0.02)
-0.1 (0.72)
0.2** (0.078)
0.19*** (0.042)
0.45** (0.138)

Dependent variable
Info-Fi, t
0.72** (0.21)
0.72** (0.035)
0.037 (0.035)
0.08 (0.06)
0.33* (0.159)
-0.077+ (0.239)
-0.21 (0.11)
-0.67* (0.34)

Info-Ui, t
-0.12 (0.23)
-0.02(0.03)
0.78***
(0.047)
0.172+ (0.091)
0.231 (0.2)
0.31 (0.237)
0.08 (0.12)
0.13 (0.38)

Tables 35 and 36 shows the results of the sample split for sentiment of posts. First, for the
luxury group, the results appear that negative FGC and positive UGC positively influence car
sales. Surprisingly, I cannot find the positive impact of positive FGC on offline car sales. This
seems to suggest that for customers who purchase luxury cars, they already have their strong
mindsets regarding their purchase decisions. That is, they may not pay special attentions to
positive contents that firms try to deliver to them to influence their purchase decisions. On the
other hand, for the same group of customers, they may pay attention to negative FGC and these
negative FGC may influence their purchase decisions. The fact that negative FGC positively
influences offline car sales is also a very interesting finding because it contradicts the general
belief that negative posts may have the negative impact on sales. However, my detailed
48

examination on those negative FGC shows that these negative posts are exclusively the
company’s voluntary announcements on vehicle recalls11. The current literature suggests that
negative posts can still increase sales by increasing product awareness (Berger, Sorensen, &
Rasmussen, 2010).
In addition, prior research also indicates that automobile companies can leverage social
media to discover vehicle defect efficiently and manage their reputation before it becomes crisis
(Abrahams, Jiao, Wang, & Fan, 2012; Yan Liu & Shankar, 2015). In fact, practitioners start to
utilize social media to mitigate the impact of the negative events such as vehicle recalls12.
Therefore, the results for the luxury group represent one example of how firms can leverage
social media to increase customers’ awareness and mitigate the impact of the negative events. In
contrast to the luxury group, positive FGC positively influences offline car sales of the nonluxury group, while negative UGC negatively influences offline car sales of the non-luxury. The
rest of comparisons between these groups provide further evidence that customers’ evaluations
of sentiment in two groups demonstrate different patterns.
Table 35. Sentiment PVAR Regression Results (Luxury)
IV
Salesi, t-1
Pos-Fi, t-1
Neg-Fi, t-1
Pos-Ui, t-1
Neg-Ui, t-1
TMSi, t-1
GTi, t-1
GPSt-1

11

Salesi, t
0.535*** (0.037)
0.005 (0.008)
0.026** (0.007)
0.022* (0.011)
-0.015 (0.01)
0.11*** (0.011)
-0.455* (0.12)
0.049 (0.045)

Dependent variable
Pos-Fi, t
Neg-Fi, t
+
0.178 (0.095)
0.176 (0.136)
0.71*** (0.025) 0.121*** (0.026)
0.095*** (0.018) 0.198*** (0.026)
0.055* (0.027)
-0.139*** (0.032)
0.014 (0.027)
0.09** (0.033)
0.1*** (0.022)
-0.039 (0.032)
-0.43 (0.28)
-1.72*** (0.41)
0.13 (0.12)
-0.029 (0.14)

Pos-Ui, t
0.204* (0.098)
0.12*** (0.023)
0.01 (0.01)
0.846*** (0.025)
-0.094*** (0.02)
0.059* (0.024)
-0.78* (0.3)
0.09 (0.11)

Neg-Ui, t
-0.15 (0.11)
0.163*** (0.027)
0.005 (0.02)
0.1** (0.036)
0.62*** (0.03)
-0.02 (0.03)
-1.3*** (0.37)
0.19 (0.13)

One example of negative FGC from Lexus: https://www.facebook.com/90671958533/posts/296236784524

12

http://www.nytimes.com/2014/03/24/business/after-huge-recall-gm-speaks-to-customers-through-socialmedia.html?_r=0
49

Table 35 (cont’d)

CCIt-1
0.227*** (0.029)
ΔS&P500t-1 1.04*** (0.1)

-0.189 (0.075)
1.09*** (0.252)

-0.038 (0.109)
-0.178 (0.352)

-0.35*** (0.292) -0.08 (0.098)
0.15 (0.272)
-0.11 (0.303)

Table 36. Sentiment PVAR Regression Results (Non-Luxury)
IV
Salesi, t-1
Pos-Fi, t-1
Neg-Fi, t-1
Pos-Ui, t-1
Neg-Ui, t-1
TMSi, t-1
GTi, t-1
GPSt-1
CCIt-1
ΔS&P500t-1

Salesi, t
0.517*** (0.068)
0.03***
(0.009)
-0.013+ (0.007)
0.009 (0.016)
-0.03* (0.015)
0.06** (0.02)
-0.09 (0.06)
0.267*** (0.075)
0.211*** (0.043)
0.429** (0.124)

Dependent variable
Pos-Fi, t
Neg-Fi, t
0.91*** (0.193) 0.621* (0.3)
0.681*** (0.035) 0.019 (0.037)
0.032 (0.023)
0.08* (0.033)
0.013 (0.049)
0.122* (0.06)
0.045 (0.046)
-0.023 (0.056)
0.017 (0.063)
-0.0001
(0.11)
0.416** (0.16)
0.489+ (0.251)
-0.169 (0.243)
-1.16*** (0.178)
-0.337** (0.129) 0.075 (0.192)
-1.95*** (0.373) -0.29 (0.519)

Pos-Ui, t
-0.142 (0.235)
-0.004 (0.029)
0.03 (0.027)
0.6*** (0.06)
0.117** (0.043)
0.193* (0.08)
0.266 (0.175)
0.391 (0.243)
-0.111 (0.135)
-0.714* (0.363)

Neg-Ui, t
-0.4 (0.3)
-0.042 (0.037)
0.017 (0.028)
0.296*** (0.063)
0.45*** (0.06)
0.359** (0.1)
0297 (0.241)
0.609* (0.3)
0.442** (0.168)
0.2 (0.464)

Taken together, the luxury/non-luxury sample split suggests again that different forms of
format presentation and contents of post show different patterns on customers’ online
engagements and offline commerce activities. In addition, each car brand, depending on its
belonging group, would take the dramatically different approaches to boost offline sales and
encourage customer engagements. The summary of PVAR analysis results is shown in Table 37.
Analysis level
Whole sample
Overall Post
Link
Photo
Status
Video
Informative Post
Sentiment
Sample split
Overall Post
Link
Photo
Status
Video
Informative Post
Sentiment

Table 37. Summary of Results
Results

H1 and H3
H3, H5, and H6
H1 and H3
H1 and H3
H6
H1 and H3
PF->S, NU (-)-> S, S ->PF, NF->PF, S-> NF, PF->NF, PU (-)-> NF, NU->NF,
PU-> NU
Luxury: H1, H3, and H5; Non-Luxury: H1 and H3
Luxury: H4, H5, and H6; Non-Luxury: H3, H5, and H6
Luxury: H1, H2, and H3; Non-Luxury: H1 and H3
Luxury: H3 and H5; Non-Luxury: H1 and H3
Luxury: H4, H5, and H6; Non-Luxury: H6
Luxury: H1 and H3; Non-Luxury: H1 and H3
Luxury: NF->S, PU->S, NF->PF, PU->PF, PF->NF, PU (-)->NF, NU->NF, S->PU,
PF->PU, NU (-)-> PU, PF->NU, PU->NU
Non-Luxury: PF->S, NU (-)->S, S->PF, S->NF, PU->NF, NU->PU, PU->NU

Notes: PF refers to positive FGC; NF refers to negative FGC; PU refers to positive UGC; NU refers to negative
UGC; S refers to offline car sales. The sign of “-“ stands for a negative relationship for the corresponding
relationship; otherwise, it is a positive relationship.

50

2.5.3 Impulse Response Functions (IRFs) Results

I supplement the PVAR regression estimates with the analysis of the corresponding IRFs.
IRFs allow us to learn whether a shock to one variable (e.g., FGC -> Sales) will have a
permanent or transitory effect on any of the dependent variables, and if the effect is transitory,
how long it will take to dissipate, namely, the dynamics of interested variables over time. Thus,
IRFs provide an alternative means of measuring the trajectory of the effectiveness of one
variable in the PVAR system on another variable. Specifically, IRFs plot the response of current
and future values of the variables in the PVAR model to one-unit increase in the current value of
one of the PVAR error terms (Enders, 2008; Stock & Watson, 2001). The assumption here is that
the error return to zero in subsequent periods and all of the other errors are zero. I conducted
IRFs along with 95% confidence intervals generated from Monte Carlo simulations with 1,000
repetitions. Figures 5 to 11 shows IRFs results for the whole sample.
Figure 5 illustrates the results of IRFs for the complete sample for overall posts.
Particularly, Figures 5.3 and 5.6 allow me to observe how offline car sales respond to a shock to
UGC and FGC over time, respectively. For example, the results show that an unexpected oneunit increase in the variable FGC (or UGC) is associated with 3.2% increase (or 1.8% decrease)
in the logarithm of offline car sales at t=1. The effect of FGC to sales is significantly different
from zero; however, the effect of UGC is not significantly different from zero. I also observe that
the effect of FGC to offline car sales reaches a peak at t=2 with around a 4% increase in the

51

logarithm of offline car sales and this effect attenuates gradually over time. Figures 5.7 and 5.8
show how UGC and FGC respond to a shock to offline car sales, respectively. Finally, I observe
that the effect of sales to UGC is not significantly different from zero (see Figure 5.7). However,
offline car sales have the strong immediate and significant impact on FGC in the subsequent
month and this impact starts to attenuate at around month 3 (see Figure 5.8).
Figure 5. Impulse Response Functions for Overall Post

Notes: X-axis is the forecast horizon (in months), and y-axis is the forecasted response of the dependent variable to
a unit shock in the corresponding error term. F stands for “firm-generated” and U stands for “user-generated”. This
applies to Figures 6 to 25.

Figure 6 presents the results of IRFs for posts in the form of link for the complete sample.
In contrast to Figure 5, there is a positive feedback effect between FGC and UGC in the form of
link. Figures 6.2 and 6.4 demonstrate how firm-generated links respond to a shock to usergenerated links and how user-generated links respond to a shock to firm-generated links,
respectively. Specifically, in Figure 6.2, a one-

unit increase in user-generated links at time
52

zero results in around 12.5% increase in the logarithm of firm-generated links at month 1 and
this effect reaches the equilibrium levels at month 4. Similarly, the magnitude of the effect of
FGC on UGC (both in the form of link) in the first time period is around 0.1 (see Figure 6.4),
implying that a one-unit increase in firm-generated links at time zero results in 10% increase in
the logarithm of user-generated links at month 1. Furthermore, this effect also reaches the
equilibrium level at month 4.
Figure 6. Impulse Response Functions for Link

Figure 7 illustrates the results of IRFs for posts in the form of photo for the complete
sample. Figure 7.6 allows me to observe how offline car sales respond to a shock to firmgenerated photos over time. Specifically, the result shows that a one-unit increase in firmgenerated photos is associated with 1 % increase in the logarithm of offline car sales at t=1 and
this effect reaches a peak at t=2 with around 1.7% increase in the logarithm of offline car sales.
53

Additionally, Figure 7.8 also shows that offline car sales have the strong immediate and
significant impact on firm-generated photos in the subsequent month and this impact starts to
attenuate at around month 3 (see Figure 7.8).

Figure 7. Impulse Response Functions for Photo

Figure 8 illustrates the results of IRFs for posts in the form of status for the complete
sample and shows the similar patterns with Figure 7. For example, similar to Figure 7.6, Figure
8.6 shows that a one-unit increase in firm-generated statuses is associated with 2.2 % increase in
the logarithm of offline car sales at t=1, suggesting that firm-generated statuses are more
effective in influencing offline car sales than firm-generated photos. Figure 8.8 also shows that
offline car sales have the strong immediate and significant impact on firm-generated photos in
the subsequent month and this impact starts to diminish at around month 3 (see Figure 8.8).

54

Figure 8. Impulse Response Functions for Status

Figure 9 illustrates the results of IRFs for posts in the form of video for the complete
sample. Figure 9.2 shows that a one-unit increase in user-generated videos is associated with
around 9% increase in the logarithm of firm-generated videos.

55

Figure 9. Impulse Response Functions for Video

Figure 10 shows the results of IRFs for informative posts for the complete sample. In
particular, Figure 10.6 illustrates that a one-unit shock of informative FGC has an increasing
impact on the logarithm of offline car sales, reaching a peak at month 2. The results also suggest
that offline car sales also have an immediately positive impact on informative FGC (see Figures
10.8), suggesting a positive feedback effect between informative FGC and offline car sales.

56

Figure 10. Impulse Response Functions for Informative Posts

Finally, Figure 11 illustrates the results of IRFs for sentiment for the complete sample.
Figure 11.20 shows that a one-unit increase in positive FGC is associated with 2% increase in the
logarithm of offline car sales at t=1 and this effect reaches a peak at t=2 with around 2.2%
increase in the logarithm of offline car sales. I also observe that in Figure 4.5 a one-unit increase
in negative UGC is associated with 2% decrease in the logarithm of offline car sales at t=1 and
this effect reaches the lowest point at t=2. Figures 11.14 and 11.24 illustrate how positive FGC
responds to a shock to negative FGC and sales over time, respectively. I observe that a one-unit
increase in negative FGC is associated with 3.6% increase in the logarithm of positive FGC and
this effect reaches a peak at month 1. Furthermore, a one-unit increase in offline sales is
associated with around 42% increase in the logarithm of positive FGC and this effect reaches a

57

peak at month 2 (see Figure 11.24). Figures 11.3, 11.8, 11.18, and 11.23 show how negative
FGC responds to a shock to negative UGC, positive UGC, positive FGC, and offline car sales
over time, respectively. Finally, Figure 11.6 shows how negative UGC responds to a shock to
positive UGC over time. The result indicates that a one-unit increase in positive UGC is
associated with around 18% increase in the logarithm of negative UGC and this effect reaches a
peak at t=2.
Figure 11. Impulse Response Functions for Sentiment

2.5.3.1 Sample Split Impulse Response Functions (IRFs) Results

I now turn my attention to the IRFs for my sample split analysis. Figures 12 and 13 show
the results of IRFs for overall posts for the luxury and non-luxury groups, respectively. Figures
12.6 and 13.6 show how offline car sales respond to FGC over time, suggesting that for the
luxury group an unexpected one-unit increase in the variable FGC is associated with 4% increase

58

in the logarithm of offline car sales at t=1 while for the non-luxury group it has around 2.5%
increase at t=1. For both groups, these effects reach the peak at around month 2. Particularly for
the luxury group, I also observe that an unexpected one-unit increase in the variable FGC is
associated with 10% increase in the logarithm of UGC at t=1 and this effect continuous to reach
the peak at around month 3 (see Figure 12.4), suggesting again that customers in two different
groups demonstrate different patterns in their engagements.
Figure 12. Impulse Response Functions for Overall Post (Luxury)

Figure 13. Impulse Response Functions for Overall Post (Non-Luxury)

59

Figures 14 and 15 show the results of IRFs for posts in the form of link for the luxury and
non-luxury groups, respectively. Consistent with the main results, there is a positive feedback
relationship between user-generated links and firm-generated links for both groups (see Figures
14.2, 14.4, 15.2, and 15.4) and there effects last for around three months. For the luxury group, I
also observe that an unexpected one-unit increase in the variable offline car sale is associated
with 70% increase in the logarithm of UGC at t=1 and this effect continuous to reach the peak at
around month 4 (see Figure 14.7). On the other hand, for the non-luxury group, I find that an
unexpected one-unit increase in the variable offline car sale is associated with 120% increase in
the logarithm of FGC at t=1 and this effect starts to attenuate at month 4 (see Figure 15.8).
Figure 14. Impulse Response Functions for Link (Luxury)

60

Figure 15. Impulse Response Functions for Link (Non-Luxury)

Figures 16 and 17 show the results of IRFs for posts in the form of photo for the luxury
and non-luxury groups, respectively. Consistent with the main results, I find that there is a
positive feedback relationship between offline car sales and firm-generated photo for two groups
and these effects all reach the peak at around month 4 (see Figures 16.6, 16.8, 17.6, and 17.8).
Interestingly, for the luxury group, I find that an unexpected one-unit increase in user-generated
photo is associated with 2% increase in the logarithm of offline car sales at t=1 (see Figure 16.3)
and this effect lasts for around three months.

61

Figure 16. Impulse Response Functions for Photo (Luxury)

Figure 17. Impulse Response Functions for Photo (Non-Luxury)

Figures 18 and 19 show the results of IRFs for posts in the form of status for the luxury
and non-luxury groups, respectively. These two groups demonstrate different patterns in terms of
the dynamics between user-generated status, firm-generated status, and offline car sales. For
62

example, Figure 18.8 shows how firm-generated status responds to offline car sales over time,
suggesting that an unexpected one-unit increase in offline car sales has the immediately
positively impact on firm-generated status and this effect reaches the peak at month 4, while this
relationship cannot find in the non-luxury group (see Figure 19.8).

Figure 18. Impulse Response Functions for Status (Luxury)

Figure 19. Impulse Response Functions for Status (Non-Luxury)

63

Figures 20 and 21 show the results of IRFs for posts in the form of video for the luxury
and non-luxury groups, respectively. The results demonstrate again that customers in two
different groups show dramatically different patterns in terms of the interactions between firmgenerated video, user-generated video, and offline car sales. For example, fort the luxury group,
there is a positive feedback between firm-generated video and user-generated video (see Figures
20.2 and 20.4) and the effects reach the peak at month two. On the other hand, for the non-luxury
group I only find that an unexpected one-unit increase in user-generated video is associated with
8% increase in the logarithm of firm-generated video at t=1 (see Figure 21.2) and this effect
starts to attenuate at month four.

Figure 20. Impulse Response Functions for Video (Luxury)

64

Figure 21. Impulse Response Functions for Video (Non-Luxury)

Finally, Figures 22 to 25 show the results of IRFs for informative posts and sentiment for
two different groups. For informative posts, these two groups share similar patterns with the
positive feedback relationship between firm-generated informative posts and offline car sales
(see Figures 22.6, 22.8, 23.6, and 23.8). Finally, the results of IRFs for sentiment appear again
that how customers appreciate sentiment from these two different groups are dramatically
different.

65

Figure 22. Impulse Response Functions for Informative Posts (Luxury)

Figure 23. Impulse Response Functions for Informative Posts (Non-Luxury)

66

Figure 24. Impulse Response Functions for Sentiment (Luxury)

Figure 25. Impulse Response Functions for Sentiment (Non-Luxury)

Together, IRFs provide a graphical representation of the evolutionary patterns of our
interested relationships, and their short-term results at t=1 are consistent with the interpretation
67

of the coefficient estimates shown above in PVAR analysis results.
2.6 DISCUSSION
Motivated by the fact that most of the current literature focus exclusively on the isolated
impact of either FGC or UGC on online sales of non-durable or media products, the current
study explores the dynamic relationships among FGC, UGC, and offline car sales in the U.S.
automobile industry in the setting of the firm’s Facebook fan page. Overall, the results suggest
that (1) FGC is more effective in triggering offline car sales than UGC, (2) offline car sales
would attract more customers’ and firms’ attentions by disseminating information or voicing
opinions more actively to strengthen customer/customer or customer/firm relationships, and (3)
there is a positive feedback effect between FGC and UGC. Furthermore, different forms of
format presentation and contents of posts do play important roles in these dynamic relationships.
Finally, the results vary significantly across luxury versus non-luxury car brands. These varied
results also provide further evidence that depending upon their belonging groups firms would
leverage different mechanisms to strategically drive sales and encourage customer engagements.
2.6.1 Theoretical Implications

There are several key contributions from this research. First and foremost, the
predominant emphasis of prior social media research focuses on the isolated impact of either
FGC or UGC on online sales of media and non-durable goods such as movie, DVD, or music,
thereby overlooking any effects that FGC and UGC might have on one another, and indirectly,

68

through on another on offline sales of other types of products. The current study is the first
academic study that rigorously examines and quantifies the relative effectiveness of FGC and
UGC on offline sales of the durable product (i.e., vehicle) and their feedback effects on each
other over a long period of time at the industry level. Therefore, my work echoes the call for
more research on how firms’ media channels should operate as a system (e.g., Dewan &
Ramaprasad, 2014; Luo et al., 2013; Smith et al., 2006; Trusov et al., 2009) and the impact of
social media on different types of products (e.g., Goh et al., 2013; Stephen & Galak, 2012).
Interestingly, I find that FGC consistently has a positive impact on offline car sales throughout a
set of analysis, whereas UGC is not very effective in driving offline car sales in the setting of the
firm-initiated Facebook page. The results of UGC on sales contradict the general belief that UGC
is effective in driving sales (e.g., Ghose & Ipeirotis, 2011) and could be explained by the nature
of the durable product and Facebook. Compared to media and non-durable products, customers
may need to involve more efforts before making purchase decisions on durable products. Thus,
UGC may not have a direct impact on offline car sales. In fact, buying a car is a high
involvement activity. Therefore, customers may need to take a series of test-drives before
purchasing a car. The differences between the durable product and non-durable products may
constitute a possible alternative explanation for the ineffectiveness of UGC in my setting. The
nature of Facebook may also contribute to why UGC is less effective in driving sales. The major
purpose of voicing opinions at Facebook is to build social intractability (Stephen & Galak,

69

2012). In particular, the firm’s Facebook fan page provides an environment for customers to
express their loyalty, receive the latest information from the firm, and interact with other
customers, namely, social networking. On the other hand, UGC in the form of online review
tends to help potential customers to make better purchase decisions (Forman et al., 2008).
Indeed, in a majority of UGC in our setting, I find that customers tend to use firm’s Facebook
page to build the relationships with the focal firm and other customers. Therefore, the difference
between online review sites and firm’s Facebook pages may explain why most of UGC in our
setting do not have a direct impact on sales.
Second, not all format presentations are similar. My research suggests that some format
presentations are more effective in influencing offline car sales, whereas other aspects of format
presentations are more effective in triggering customer engagements. My study, therefore,
responds to Berger et al.’s (2015) call for further examination on consumers’ appreciation of
different content formats and the individual assessment of each content format. Depending on the
purpose of each post, firms or customers use different combinations to maximize the utility of
each post. Interestingly, we find that to effectively boost sales, firms could rely more on contents
in the form of photo and status. On the other hand, if firms want to trigger more customer
engagements, they should rely more on contents in the form of link. Thus, my study sheds light
on customers’ evaluations of content in different formats and contributes to both decisionmaking and marketing literature by demonstrating how firms can utilize the combinations of

70

format presentation to achieve their goals effectively.
Third, my content analysis on informative posts and sentiment also contributes to the
current literature by showing how firms can develop and control of their posts to trigger
consumptions and customer engagements. My study shows that if firms can provide enough
product information for their customers, the likelihood of customers’ consumptions would
increase accordingly. Further, prior research has suggested that firms can leverage social media
to mitigate the impact of the negative events by increasing customer awareness (Abrahams et al.,
2012; Berger & Milkman, 2012; Yan Liu & Shankar, 2015). I agree and extend this stream of
research by showing that luxury groups’ voluntary announcements on the negative events (i.e.,
vehicle recalls) can increase customer awareness and, in fact, have the positive impact on sales.
Thus, given the popularity of social media, firms can equip with more potent tools to avoid a
negative event before it becomes a crisis.
Finally, my split sample analysis offers very interesting insights regarding customers’
behaviors in two different groups. For example, for customers who belong to the luxury group,
their exposure to overall FGC, FGC in the form of photo, negative FGC, positive UGC, and
informative FGC would influence their purchase behaviors. On the other hand, customers who
belong to the non-luxury group demonstrate dramatically different patterns. Given the limited
resources, firms do need to customize their social media strategy based on their targeted
audiences. Therefore, my study contributes to the current literature by showing the patterns of

71

how firms with different focus and market can develop their social media strategy more
effectively.
2.6.2 Managerial Implications

The study also provides important managerial implications. Due to the popularity of
Facebook, a number of firms have set up their fan page for marketing purposes. My results
demonstrate that maintaining online communities in the form of Facebook fan page has
marketing implications for companies to reach their customers and then boost sales. This is
important for firms with the limited resources to reach large amounts of customers. However, it
should leverage social media with caution that not every form of format presentation has the
equal impact on sales and customer engagements. Therefore, firms should pay special attention
to the content of their marketing communications and develop the most effective conversations
to engage their customers. Second, although UGC is not effective in influencing offline car sales
in our setting, it does not mean that firms can get rid of this part in their fan page. Instead, firms
need to look at these contents seriously. For example, firms may try to identify if there are any
opinion leaders in their online communities. Identifying these opinion leaders would be very
important for firms if they want to initiate some seed marketing campaign. Third, my split
sample analysis also implies that firms need to be aware of their own identity. That is, depending
on the targeted market, firms need to carefully choose their contents to match their customers’
preferences. Finally, the insight from the positive impact of negative FGC on sales also provide

72

the lesson to managers that they can leverage social media to mitigate the negative events to
maintain their reputation.
2.6.3 Limitations and Future Research

This study does have some limitations. First, I only focus on Facebook. However, it is
likely that firms have a significant presence at other social media sites, such as Twitter and
Instagram, as well. Activities on these sites could affect offline car sales. Due to data limitations,
I am not able to study the interactions between offline car sales and the overall marketing
intensity across different social media sites. Second, I only examine the dynamic interactions
between FGC, UGC, and offline car sales in one industry. Finally, the current study only
explores the volume of FGC and UGC and some aspects of content. It may be interesting for
future research to develop systematics categories on contents by applying some techniques such
as topic modeling, so the abstract topics can be distinguished from one another. For example,
latent dirichlet allocation (LDA) could be utilized to take a step further to examine the impact of
social media.

2.7 CONCLUDING REMARKS
While prior studies examine the impacts of social media from different perspectives, they
focus exclusively on the isolated impacts of either FGC or UGC on online sales of nondurable
and media products. In addition, how FGC and UGC may influence offline sales of the durable

73

product and the relationships between FGC and UGC is under explored. Therefore, this study
provides an initial step on exploring the dynamic relationships between FGC, UGC, and offline
car sales in the setting of the firm-initiated Facebook fan page and how these relationships vary
across format presentation, content of post, and firm characteristics.

74

CHAPTER 3.
ESSAY 2: DYNAMICS OF ONLINE WORD OF MOUTH (WOM) SPILLOVER
EFFECTS

3.1 INTRODUCTION
A significant body of research has established that online word-of-mouth (WOM) play an
important aspect in consumers’ purchasing decisions (e.g., Chen et al., 2015; Dewan &
Ramaprasad, 2014) and provide a means for firms to learn about customers and the marketplace
(Borah & Tellis, 2016). However, online WOM regarding a specific product or brand may also
remind the consumers of options that would not have been salient otherwise. For example, a
customer’s experience of Samsung Galaxy note 7 shared on social media channels may influence
other potential or existing customers’ perception or confidence of Samsung’s smartphones in
general, and thereby influence the latter’s decision to buy a non-Samsung smartphone13. This is
one way by which online WOM about the focal brand’s competitors may influence performance
of the focal product through the spillover effects. Spillover effects occur when information and
existing perceptions influence beliefs that are not directly addressed by or related to the original
information source or perception object (Ahluwalia, Unnava, & Burnkrant, 2001). The presence
and strength of spillover effects in consumer decision-making is extremely important (Chae,
Stephen, Bart, & Yao, 2017; Dong & Chintagunta, 2015; Sabnis & Grewal, 2015). Particularly

13

Some Samsung Galaxy Note 7 Buyers Are Defecting to the iPhone: http://fortune.com/2016/09/27/note-7defecting-iphone/
75

in the setting of online WOM, because firms use online WOM to compete, studying online
WOM spillover effects is extremely important to inform a focal firm’s own performance (Sabnis
& Grewal, 2015).
Despite the richness of research on online WOM, prior research on the dynamics of
online WOM spillover effects is very scarce with few exceptions (e.g., Borah & Tellis, 2016;
Chae et al., 2017; Sabnis & Grewal, 2015). Anecdotal evidence suggests such relationships exist,
but our current knowledge is still very limited on how online WOM about the focal brand or
product would influence the performance of the focal brand (i.e., static view of online WOM).
First of all, the current literature focuses considerably on the impact of online WOM on online
sales of non-durable, media goods such as movie, DVD, music, book, or clothing (e.g., Chen et
al., 2015; Goh et al., 2013). For this type of product, customers are more likely to experience
low-involvement decisions, meaning that a consumer rarely engage in an “extensive search for
information or a comprehensive evaluation of the choice alternatives” (Zaichkowsky, 1985,
p.341). On the other hand, customers who purchase durable products such as cars are more likely
to engage in extensive search because the wrong decision force the consumer to deal with the
poor product for long periods of time (Laurent & Kapferer, 1985). Due to these two extremely
different search processes, the findings from the existing literature may or may not apply to nondurable products. Besides, while online WOM is increasing in importance, as of the 4th quarter
2015, online sales only account for 7.5% of all retail purchases (Commerce, 2016). Therefore,

76

there is a significant knowledge gap on the phenomenon of online WOM to offline sales of
durable products.
Second, the same stream of the literature focuses considerably on the immediate effects
of online WOM about the focal brand on the performance of the focal brand without considering
the competition nature in the marketplace (e.g., Goh et al., 2013). However, it is very likely that
in the competitive marketplace, customers will be influenced not only by online WOM of
interest to them but also by online WOM of other competing or related products before making
purchases. Unfortunately, the broader consequences of the effects of online WOM (i.e., the
competition nature of online WOM) have received scant attention in the current literature.
Third, prior research on consumer choice suggests that purchase decisions are considered
to be the result of a multi-stage decision process (Bettman, 1979; Hauser & Wernerfelt, 1990;
Shocker, Ben-Akiva, Boccara, & Nedungadi, 1991), which typically includes the stages of
awareness, interest, and final decision (De Bruyn & Lilien, 2008). Depending on certain
circumstances, this multi-stage decision process does not necessarily imply sequencing and it
may occur simultaneously (De Bruyn & Lilien, 2008; Shocker et al., 1991). However, most of
the current literature in online WOM does not distinguish the stage of awareness from the stage
of interest or just examines exclusively the mixed effect from these two stages on firm
performance (e.g., Borah & Tellis, 2016; Chevalier & Mayzlin, 2006; Duan et al., 2008; Goh et
al., 2013; Sabnis & Grewal, 2015). For example, although Sabnis and Grewal (2015) provide one

77

of initial evidence about online WOM spillover effects in the cable industry, they still do not
explain the relative effects from the stages of awareness and interest. Similarly, although Borah
and Tellis (2016) and Chae et al. (2017) provide further evidence of online WOM spillover
effects, their studies tend to examine the mixed effects from the stages of awareness and interest
in essence. Separate analyses on each stage of the purchase decision process would provide
deeper insight into the questions of how firms could allocate their resources and efforts in each
stage and convert potential customers into actual sales. However, our knowledge on the relative
degree to which online WOM effects and their spillover effects from these two stages jointly
influence firm performance is still not clear.
The objective of this paper is to examine the dynamics of online WOM and its spillover
effects by considering the relative effects at the stages of customer awareness and consideration
in the U.S. automobile industry. I define online WOM spillover effect as the positive or negative
influence of online WOM about competitors on offline car sales of the focal brand. I select the
U.S. automobile industry because of its considerable economic significant and its increasing
reliance social media marketing (eMarketer, 2015b; Tang et al., 2014). Specifically, I am
interested in the following research questions:
(1) What are the relative effects of online WOM at the stage of awareness and online WOM at
the stage of consideration on offline car sales?
(2) Do the spillover effects exist in online WOM at the stages of awareness and consideration?

78

That is, does online WOM about competitors at the stages of awareness and consideration spill
over into offline car sales of the focal brand?
(3) What are the patterns of spillover effects across two stages and how do these patterns vary if
different mechanisms at the stage of awareness are considered?
(4) How do dynamics of online WOM vary across firm characteristics (i.e., origin of brand,
market structure, and price factor)?
To measure online WOM and its spillover effects across two stages of customer purchase
processes, I suggest combining data from firms’ Facebook fan pages and customers’ test drive
experience shared on different social media channels. Among different forms of social media,
Facebook strongly leads the social networking space (eMarketer, 2015c). Due to its popularity,
Facebook has become a leading avenue for more than 54 million businesses to set up their online
brand communities (i.e., fan page) to enable direct communication with prospective customers,
increase their awareness, and make profits (Facebook, 2015b; Goh et al., 2013). Particularly, it is
noteworthy the newsfeed function at Facebook. The newsfeed function at Facebook allows users
to be exposed to customized content posted on the network based on each their own using
behaviors (e.g., the frequency to interact with the friends, pages, or public figures) (Backstrom,
2013). Due to this unique feature, firms’ Facebook fan pages is particularly effective to raise
awareness for their products (Debatin, Lovejoy, Horn, & Hughes, 2009; Yong Liu, 2006) by
appearing in the newsfeeds of all fans (Goh et al., 2013) and by increasing fan page engagement

79

and usage intensity (Verma, Jahn, & Kunz, 2012). Therefore, activities at firms’ Facebook fan
pages provide a good proxy to measure online WOM at the stage of awareness. I leverage data
specifically about customers’ test drive experience shared on different social media channels as a
proxy for online WOM at the stage of interest (i.e., consideration). Because automobile purchase
is a high-involvement decision for most consumers, the test drive is one of the critical aspects of
the pre-purchase production evaluation. Therefore, by focusing on the test-drive as a measure for
online WOM at the stage of interest where more specific aspects of the product, such as part
particular product attributes or functionalities, are provided, I can understand deeply about the
effects of online WOM and its spillover effects occurred at the stage of interest.
The results indicate that (1) online WOM at the stage of consideration has the stronger
effect on offline car sales than online WOM at the stage of awareness, (2) spillover effects exist
across both stages of awareness and consideration, though effects are heterogeneous in direction:
positive spillover effects at the stage of awareness while negative spillover effects at the stage of
consideration, and (3) at the stage of awareness, online WOM initiated by firms is more effective
in influencing offline car sales than online WOM initiated by users. Furthermore, not every
mechanism at Facebook (i.e., post, like, comment, and share) has the equal impact on offline car
sales and these different mechanisms also influence how customers appreciate online WOM at
the stage of consideration. Finally, the results vary significantly across origin of brand, market
structure, and price factor.

80

The rest of this paper is organized as follows. The second section presents the current
stream of the literature and the gaps, the third section introduces research model and hypotheses,
the fourth section describes the method, and the last two sections present the results and
discussion.

3.2 LITERATURE REVIEW
The current study is related to the literature that examines the effect of online WOM on
marketing outcomes. In this section, I briefly review the current streams of research and discuss
how my study contributes to the extant literature.
3.2.1 Effect of Online WOM on Sales

A number of studies have examined the impact of online WOM on online sales of media
or non-durable products (i.e., products that customers usually experience low-involvement
decisions). For example, Chevalier and Mayzlin (2006) examine the effect of online reviews on
relative sales of books at Amazon.com and Barnesandnoble.com and find the positive
relationship between online book reviews and online book sales. Forman et al. (2008) posit that
reviews posted by reputable reviewers have greater impact on product sales than those by less
reputable reviewers. Ghose and Ipeirotis (2011) find that reviews that have a mixture of
objective, and highly subjective sentences are negatively associated with product sales,
compared to reviews that tend to include only subjective or only objective information.

81

Furthermore, Goh et al. (2013) study the relative impacts of firm-generated content and UGC on
sales in the setting of a casual wear apparel retailer’s Facebook page and find that FGC in a
firm’s Facebook page influences consumers’ apparel purchase expenditures. More recently,
Chen et al. (2015) study the effect of artists’ broadcasting activities on MySpace and suggest that
broadcasting in social media has a significant effect on music sales and the effect mainly comes
from personal messages (one form of FGC on social media) rather than automated messages.
Although the current literature helps us understand the impact of online WOM on firm
performance, our understanding regarding the impact of online WOM on offline sales of the
high-involvement products is still very limited. Prior research indicates that information-search
behaviors have been shown to vary for different levels of product involvement (Geva,
Oestreicher-Singer, Efron, & Shimshoni, 2017). Product involvement refers to consumers’
interest in a product and their perceptions regarding its importance (Richins & Bloch, 1986;
Traylor, 1981). One major difference between high-involvement products and low-involvement
products is the effort for searching for product information to make the right decision (Gu et al.,
2012). For high-involvement products, customers often spend a significant amount of time
conducting information searches before purchasing these products (Kuruzovich, Viswanathan,
Agarwal, Gosain, & Weitzman, 2008). On the other hand, for low-involvement products such as
music, because the consequences of making the wrong purchase decisions are limited (Gu et al.,
2012), consumers are less likely to conduct extensive information searches before making the

82

purchase (Mathwick & Rigdon, 2004). Therefore, it can be simply stated that the more important
the product is to a consumer, the more motivated the consumer is to search and be involved in
the decision (Geva et al., 2017). Gu et al.’s study (2012) is one of remarkable examples on the
impact of online WOM on products that would trigger high-involvement decisions. They find
that a retailer’s internal WOM has a limited influence on its sales of high-involvement products,
while external WOM sources have a significant impact on the retailer’ sales. However, they still
only focus on online sales in the setting of Amazon. Particularly, in the setting of automobile
industry, Tang et al. (2014) and Geva et al. (2017) are two remarkable examples to examine the
effect of online WOM on offline sales of durable products (vehicles). Given that online sales
only account for 7.5% of all retail purchases (U.S. Department of Commerce, 2016) and
customers experience two dramatically different decision processes for two types of products,
our knowledge on the impact of online WOM on the offline sales of durable product is still very
limited.
Another noteworthy gap in this stream of literature is that we still know very little about
the role of online WOM played in different stages of customer decision process. Prior research
suggests that in the setting of purchase decision process, customers may experience the stages of
awareness, interest (consideration), and final decision (Bettman, 1979; De Bruyn & Lilien, 2008;
Hauser & Wernerfelt, 1990; Shocker et al., 1991). Depending on certain circumstances, this
multi-stage decision process does not necessarily imply sequencing and it may occur

83

simultaneously (De Bruyn & Lilien, 2008; Shocker et al., 1991). At the stage of awareness,
consumers know the alternative exists, but may not have either interest in it or sufficient
information to understand its possible benefits, while at the stage of interest, customers are aware
of available alternatives, and hence decide to learn more about the product (De Bruyn & Lilien,
2008). It is generally believed that at the stage of awareness, customers tend to expose more
general and abstract information given that their product selection set is relatively large (Shocker
et al., 1991). On the other hand, at the stage of interest stage, customers tend to expose to more
concrete or detailed information given that they have decided to pay more attention on the subset
of their selection set (Shocker et al., 1991). However, a considerable literature does not examine
the relative role played by online WOM across these two stages or tend to examine the mixed
effect from these two stages on firm performance (e.g., Borah & Tellis, 2016; Chevalier &
Mayzlin, 2006; Duan et al., 2008; Geva et al., 2017; Sabnis & Grewal, 2015). Therefore, our
knowledge on the role played by online WOM in different stages of decision process is still very
limited.
3.2.2 Spillover Effects of Online WOM

Spillover effects occur when information and existing perceptions influence beliefs that
are not directly addressed by or related to the original information source or perception object
(Ahluwalia et al., 2001). Prior research has provided evidence of spillover effects in various
setting such as IT investment (Tambe & Hitt, 2013), advertising (Sabnis & Grewal, 2015), brand

84

scandal (Roehm & Tybout, 2006), brand portfolio (Lei, Dawar, & Lemmink, 2008), or customer
satisfactory (Dong & Chintagunta, 2015). The accessibility-diagnosticity theory proposed by
Feldman and Lynch (1988) provides an appropriate framework to study and explain spillover
effects. They suggest that if a consumer thinks that information for brand A is accessible and
diagnostic of brand B (i.e., informative about), the consumer will use perceptions of brand A’s
quality to infer quality of brand B (e.g., Ahluwalia & GĂźrhan-Canli, 2000; Roehm & Tybout,
2006). Accessibility is such that concepts, such as brand, firm characteristics, and product
attributes, reside in a network and can activate one another when having strong links (Anderson,
2013; Collins & Loftus, 1975). On the other hand, diagnosticity is a function of consumers’
implicit theories about how things relate in the world (Broniarczyk & Alba, 1994a, 1994b).
Therefore, it is intuitive that in the competitive marketplace, firm’s performance depends not
only on its own marketing strategy and effort but also of its competitors (DubĂŠ & Manchanda,
2005; Naik, Raman, & Winer, 2005).
In the context of online WOM, some scholars have started to examine spillover effects of
online WOM in different settings. For example, Sabnis and Grewal (2015) study the dyanimcs of
online WOM in the cable industry and find evidecne of a statistically significant relationship
between competitors’ user-generated content and focal firm’s viewership. Particually in the
automobile industry, Borah and Tellis (2016) stduy spillover effects of online WOM in the
setting of automobile recalls and find that negative online chatter about one nameplate increase

85

negative chatter for another nameplate. More recently, Chae et al. (2017) investigate the effects
of Seeded marketing campaigns (SMCs) that extend beyond the generation of WOM for a
campaign’s focal product by considering how seeding can affect WOM spillover effects at the
brand and category levels. Their study provides further evidence of online WOM spillover
effects and suggests that marketers can use SMCs to focus online WOM on a particular product
by drawing consumers away from talking about other related, but off-topic, products.
Althoguh the extant literature has provided the initial understanding about online WOM
spillover effects, these studis, again, ingore the relative role of online WOM played during
different stages of consumer decision process (i.e., the stage of awareness and interest
(consideration). Distinct from these stuides of online WOM and spillover effects, I examine
whetehr online WOM spillover effects occur across the stage of awareness and interst
(consideration) and how firm characteristics may vary these spillover effects.

3.3 RESEARCH MODEL AND HYPOTHESES
In contrast to the existing literature, the current study is unique in the following ways.
First, I consider the role played by online WOM across the stage of awareness and consideration
in the consumer decision process and examine their relative effect on offline car sales. This
approach allows me to better understand the impact of online WOM and expects to provide a
more insightful lesson to practitioners in terms of their resource allocation to leverage online

86

WOM. Second, by considering online WOM spillover effects across two stages, the current
study could shed light on the effectiveness of online WOM and better realize the competition
nature in the competitive marketplace. Finally, I focus on what firms and consumers do online
versus consumers’ commerce activities that occur in offline settings. Particularly, the difference
between non-durable products (e.g., music, movie) and durable products (i.e., vehicle in our
setting) allows me to explore the impacts of online WOM from the different perspective.
My conceptual model is shown in Figure 26. Regarding online WOM at the stage of
awareness, I leverage data from firm’s official Facebook page. Currently, due to the popularity
of Facebook, most of firms set up their Facebook fan pages to disseminate information to their
customers more effectively. In general, firms tend to leverage their Facebook pages to catch their
customers’ attention first and then refer their customers to some specific sites or representatives
for further details. In other words, at Facebook pages, information tends to be more abstract and
more general, which coincides with the nature of the stage of awareness (Shocker et al., 1991).
Therefore, leveraging data from firm’s Facebook pages provides a good proxy for online WOM
at the stage of awareness. I emphasize four major mechanisms at the firm’s Facebook page and
examine how these four mechanisms may vary the relationships shown in Figure 26. These
mechanisms include post, “Like” associated posts (i.e., how many likes a single post received),
“Comment” associated with posts (i.e., how many comments a single post received), and “Share”
associated with posts (i.e., how many shares a single post received). Since at the firm’s

87

Facebook, both firms and users can initiate contents to interact with firms or other users, I also
consider these two sources of online WOM at the stage of awareness together to shed light on
how firm’s media channels could operate as a system (Dewan & Ramaprasad, 2014; Stephen &
Galak, 2012). For online WOM at the stage of consideration, I focus specifically on test drive
experience shared from different social media channels. Compared to posts at firm’s Facebook,
test drive experience posts shared on social media channels tend to be very detailed and concern
information, which provides sufficient information for customers to narrow down their selection
set. The nature of detailed and concern information also coincides with the characteristics of the
stage of consideration (Shocker et al., 1991).
I am also interested in how the interactions shown in Figure 26 vary by firm
characteristics. Specifically, I focus on three different aspects of firm characteristics: origin of
brand, market structure, and price factor. Prior research suggests that country of origin would
moderate the spillover effects (Borah & Tellis, 2016) because consumers might use the origin as
an attribute and make similar inferences for brands that belong to the same origin (Hong &
Wyer, 1990). Thus, in the category of origin of brand, I focus on Asian-based, European-based,
and US-based brands to investigate how origin of brand may vary dynamics of online WOM in
the current setting. Furthermore, I also focus on market structure. Netzer, Feldman, Goldenberg,
and Fresko (2012) applied the text-mining approach on user-generated content in the U.S
automobile industry to identify market structure of the US automobile industry. I, therefore,

88

leverage their three sub market structures to examine how market structure may vary the
relationships shown in Figure 26. Finally, prior research on WOM has indicated that consumers’
motivation to engage in online WOM differs across brands and that consumers may be
particularly inclined to converse about highly regarded or high-quality brands (Lovett, Peres, &
Shachar, 2013). Therefore, I consider the price factor by examining how the category of luxury
versus non-luxury may vary dynamics of online WOM examined in the current study.

Figure 26. Conceptual Model

It has been widely acknowledged that customer awareness on firms have a positive
impact on firm performance (e.g., Goh et al., 2013; Homburg, Klarmann, & Schmitt, 2010;
Hoyer & Brown, 1990). Customer awareness is defined as the capacity of decision-makers to
distinguish or recall a brand (Homburg et al., 2010; Hoyer & Brown, 1990). Typically, in the
journey of consumer choice, customers are uncertain about product quality and therefore

89

perceive their decisions as risky because the consequences of a purchase cannot be entirely
anticipated (Homburg et al., 2010). Prior research suggests that customer awareness is effective
in reducing customer uncertainty by reducing buyer information costs and buyer-perceived risk
(Erdem & Swait, 1998; Erdem, Swait, & Valenzuela, 2006; Homburg et al., 2010).
In the setting of firms’ Facebook pages, to increase customer awareness with the hope of
converting customers into profits, firms deliberately engage in advertising, products/services,
deals, and customer relationship (Goh et al., 2013; Miller & Tucker, 2013). The intensity of these
marketing efforts has been shown to have a positive impact on firm performance (Goh et al.,
2013). Particularly, the newsfeed function at Facebook allows firm’s efforts (e.g., new postings)
to automatically appear in the personal newsfeed of all Facebook users directly connected to the
fan page in addition to appearing on the fan page itself (Debatin et al., 2009). This mechanism is
very effective for awareness creation (Debatin et al., 2009) because it increases the likelihood for
and the speed of viral information distribution through network effects (Trusov et al., 2009).
Thus, I hypothesize:
H1: Online WOM about the focal brand at the stage of awareness is positively associated
with offline car sales of the focal brand.
I posit that online WOM about the focal brand’s competitors at the stage of awareness
will have a positive impact on offline car sales of the focal brand (i.e., positive online WOM
spillover effects). Prior research on marketing spillover effects find positive spillover effects

90

such that a marketing-related action has an effect in the same direction on a focal product and
associated products (Chae et al., 2017). For example, in the context of online advertising, Lewis
and Nguyen (2015) find positive spillover effects where display ads for Samsung tablets increase
search volume for that product and Apple iPads. Similarly, Borah and Tellis (2016) also identify
a positive spillover effect (they refer it as the “perverse halo effect). They find that in the setting
of U.S. automobile recalls negative consumer sentiment an automotive brand increases negative
sentiment toward other automotive brands in the same category.
The positive spillover effects could be explained by the fact that firm’s marketing efforts
for the focal brand can cue consumers to think about associated but broader concepts related to
non-focal brands (Chae et al., 2017). Namely, thinking about a focal brand or product could
trigger thoughts about higher-level concepts, which in turn open up the possibility of thinking
about other brands or products (Berger & Schwartz, 2011). Construal level theory also further
supports this positive nature (Trope & Liberman, 2010; Trope, Liberman, & Wakslak, 2007).
This perspective also implies that in the context of customer purchase decisions if consumers, in
response to exposure to a stimulus, adopt an abstract and broad perspective (i.e., higher-level
construal), a positive spillover effect could occur (Chae et al., 2017).
In the stage of awareness, customers know the alternative exists, but may not have either
interest in it or sufficient information to understand its possible benefits (De Bruyn & Lilien,
2008). Particularly, compared to test drive experience shared on different social media channels,

91

firms’ Facebook fan pages tends to provide more general information to catch customers’
attention and let them aware information in a timely manner. For more specific information,
firms tend to ask their customers refer to some specialists for further information. Thus, it is
plausible that relative to online WOM at the stage of interest, measured by test drive experience,
customers tend to receive abstract and broad information in the stage of awareness, which is
measured by activities at firm’s Facebook page. Thus, I hypothesize:
H2: Online WOM about the focal brand’s competitors at the stage of awareness is
positively associated with offline car sales of the focal brand.
Extant research has well recognized the positive impact of online WOM on various
marketing outcomes (Chevalier & Mayzlin, 2006; Forman et al., 2008; Sabnis & Grewal, 2015).
I posit that online WOM about the focal brand at the consideration stage, measured by
customers’ test drive experience, has the positive impact on offline car sales of the focal brand.
Compared to low-involvement products such as music, customers who experience highinvolvement decisions (i.e., purchasing a car in my setting) would engage in a comprehensive
evaluation of the choice alternatives before purchasing (Zaichkowsky, 1985). Online WOM,
measured by test drive experience, provides a good outlet for potential customers to
comprehensively and seriously evaluate the vehicles and their alternatives. Usually, customers
would provide very detailed information about their test drive experience range from exterior
looks, interior designs, price information, car features (e.g., horsepower), to control conditions in

92

the road, namely, a very narrow and concrete aspects of the vehicle and test drive experience.
This detailed information allows customers to deeply understand the features of each evaluated
vehicle and therefore make a better purchase decision. Thus, I hypothesize:
H3: Online WOM about the focal brand at the stage of consideration is positively
associated with offline car sales of the focal brand.
I posit the opposite direction for online WOM spillover effects at the stage of interest
(i.e., negative online WOM spillover effects). At the stage of consideration, customers already
know the alternatives existed, start to develop some interest, and hence decide to learn more
about the brand or product (De Bruyn & Lilien, 2008). In other words, at this stage customers
would seek for narrower, detailed, and concrete information to help them refine their selection
set (Shocker et al., 1991). As described above, online WOM, measured by test drive experience,
provides a very concrete and narrow description about the functionality of each evaluated
vehicles. Therefore, compared to online WOM from firms’ Facebook fan page that usually
provides more abstract and broader information, test drive experience provides more information
about functionality and feasibility rather than only desirability. These types of information have
been shown to represent lower-level construal (R. Dhar & Kim, 2007; Liberman & Trope, 1998),
which will show the opposite effect as higher-level construal would have. Thus, I hypothesize:
H4: Online WOM about the focal brand’s competitors at the stage of consideration is
negatively associated with offline car sales of the focal brand.

93

3.4 DATA AND EMPIRICAL METHODOLOGY
3.4.1 Research Context

I select the U.S. automobile industry to analyze the relative effects of online WOM of the
stages of awareness and interest and their spillover effects for several reasons. First, the U.S.
automobile industry is of considerable economic significance. It generates sales representing 3%3.5% of U.S. gross domestic sales and accounts for 1 in 7 jobs in the U.S. economy (Hill et al.,
2010; Kalaignanam, Kushwaha, & Eilert, 2013; Pauwels & Srinivasan, 2004). In response to the
global economic recession, automobile companies have turned their attention to social media
marketing to enhance customer relationships, disseminate a variety of information, engage
customers, and boost sales (Pauwels & Srinivasan, 2004). This suggests that the U.S. automobile
industry provides an appropriate setting to study online WOM related issues. Second, the highinvolvement nature of the automobile industry leads consumers to discuss and gather information
more frequently than other industries (Borah & Tellis, 2016). Given this unique nature, the
automobile industry provides a considerable amount of online WOM as consumers actively and
frequently participate in numerous social media sites (Borah & Tellis, 2016).
I focus on firm’s Facebook fan page as a proxy for online WOM at the stage of
awareness because Facebook is the most visited social media site in the US (eMarketer, 2015c).
Due to this popularity, more than 54 million businesses have set up their online brand
communities (i.e., fan page) for marketing purposes (Facebook, 2015b). More importantly, in the

94

setting of firm’s Facebook page, firms usually use general information to catch their customer
attention first. Once customers are aware of this information, firms would ask those customers to
refer to some specialists or some web sites for more detailed information. The nature of general
information at firm’s Facebook fan page matches the characteristics of the stage of awareness.
Namely, customers know the alternative exists, but may note have either interest in it or
sufficient information to understand its possible benefits. Thus, firm’s Facebook fan page
provides a representative sources to measure online WOM at the stage of awareness.
For online WOM at the stage of consideration, I focus on customers’ test drive
experience shared on different social media channels. As discussed, the automobile industry is a
high-involvement industry. Thus, customers actively and frequently share their information such
as test drive experience in different social media sites. Because test drive experience is one of the
critical factors before the final purchase decision, I, therefore, believe that customer test drive
experience shared on different social media sites represent an appropriate source as a proxy for
online WOM at the stage of consideration.
3.4.2 Data

My samples consist of 30 major car brands in the U.S. automobile industry14. My dataset
on online WOM covers two parts from May 2009 to October 2014: (1) firm’s Facebook data for

14

These car brands are: Acura, Audi, BMW, Buick, Cadillac, Chevrolet, Chrysler, Dodge, FIAT, Ford, Honda,
Hyundai, Infiniti, Jaguar, Jeep, KIA, Land Rover, Lexus, Lincoln, Mazda, Mercedes-Benz, Mitsubishi, Nissan,
Porsche, Saab, Scion, Subaru, Toyota, Volkswagen, and Volvo.

95

the stage of awareness and (2) customers’ test drive experience shared on different social media
channels for the stage of consideration (interest). To collect online WOM at the stage of
awareness from firm’s Facebook fan pages, I relied on the Facebook graph API (Application
Programming Interface)15. Each of these thirty brands has its official Facebook fan page. Thus, I
built a software tool in PHP to connect with the Facebook graph API to collect data at the stage
of awareness. My data starting point was May 2009 because most car brands initiated their
Facebook pages on May 200916.
I considered four major mechanisms at firm’s Facebook page as the metrics for online
WOM at the stage of awareness, trying to understand how these mechanisms may vary the
relationships shown in my research model. These four metrics include: (1) total number of posts,
(2) total number of likes associated with posts, (3) total number of comments associated with
posts, and (4) total number of shares associated with posts. For each single post, my data covered
detailed information, including post id, post time, post type, post source, post content, and post
link. In addition, each single comment associated with each single was also collected. Figure 27
shows one example of the firm’s post at its Facebook fan page and its associated metrics. In this
example, there are 271 “Likes”, 16 “Comments”, and 30 “Shares” associated with this post.
Besides, as discussed above, firms tend to use their Facebook fan page to let their customers

15

Facebook Graph API: https://developers.facebook.com/docs/graph-api

16

Among these 30 car brands, 23 firms started their pages in 2009, 5 firms started their pages in 2010, and 2 firms
started their pages in 2011.

96

aware some information available. Namely, at firm’s Facebook page, relative to customers’ testdrive experience, firms use more abstract and general information to catch their attention without
providing too many details. For those customers who have paid attention to some particular
information, firms would ask them to refer to some specialists for detailed information.
Therefore, this approach provides further evidence of why firm’s Facebook fan page could be a
good proxy for online WOM at the stage of awareness. In other words, through firm’s Facebook
page, customers only know the alternative exists, but they do not have sufficient information to
understand its possible benefits. To conclude, my data on online WOM at the stage of awareness
from firm’s Facebook fan page included 59,405 firm-generated posts, 146,556,793 likes
associated with those firm-generated posts, 4,209,129 comments associated with those firmgenerated posts, 11,257,103 shares associated with those firm-generated posts, 811,387 usergenerated posts, 19,395,881 likes associated with those user-generated posts, 1,634,363
comments associated with those user-generated posts, and 155,343 shares associated with those
user-generated posts.
To collect customers’ test drive experience for online WOM at the stage of consideration
(interest), I used a combination of web crawler and artificial intelligence based text reducer
provided by a commercial third party analyst. This type of the approach has been used recently
to collect more comprehensive online WOM to better understand its impact in various settings
(e.g., Borah & Tellis, 2016). There were three major steps to collect customers test-drive

97

experience data shared from different social media channels. First, I collected the completed list
of nameplates for each single brand (e.g., Toyota Corolla, Toyota Camry, Honda Civic) from the
WardsAuto Premium database. Then, for each single nameplate, I used the combination of each
nameplate and test drive keyword as a search query to collect test drive data by using the server
provided by this third party analyst at the monthly level17. Third, for each single brand (e.g.,
Toyota), I summed up all test drive data in the nameplate level. The approach of collecting test
drive data in the nameplate level and then summing them up for the brand level allows me have
the most comprehensive set of online WOM data for the stage of consideration (interest). The
dataset included customers’ test drive experience on different platforms of social media (e.g.,
Facebook), various forums such as Automotiveforum.com, and review sites such as
Edmunds.com. Overall, approximately 1,000 different social media sites were included in my
dataset. Specifically, this third-party data provider scraped these sites to obtain any online WOM
related to customers’ test drive experience that mentioned nameplate across the time frame of the
current study. For each single post, this third-party data provider also reported published time,
the source of the post, the link of the post, and the author of the post. Figure 28 shows the screen
shot of one test drive experience post and Table 38 shows the full content examples of test-drive
experience from my dataset. As these sample examples show, relative to posts at firm’s
17

Example of the search query for collecting test drive posts used in the current study: (Chevrolet OR #Chevrolet
OR #ChevroletMalibu OR #Malibu OR @Chevrolet OR @ChevyCustCare OR @ChevyLife) NEAR/3 ("Malibu")
AND (test drive)

98

Facebook pages, these posts represent very detailed and narrow descriptions for the functionality
of each evaluated vehicle rather than just desirability, which provide more sufficient and
concrete information at the stage of consideration (interest). To conclude, there were 444,035 test
drive related posts in my final dataset.
I collected monthly offline car sales and list prices from the WardsAuto Premium
database, and traditional media advertising expenditure from Kantar Media. I also collected the
consumer search volume index in the U.S. from Google Trends to control for the popularity
effect of each car brand. Finally, I collected the monthly gasoline price index from the U.S.
Bureau of Labor Statistics and the conference board’s consumer confidence index. I focused on
monthly data because offline car sale data is only available at the monthly level. My final panel
contains total 1,789 firm-month observations. Tables 39 and 40 present the definition of
variables and summary statistics, respectively.

99

Figure 27. Example of Post at the Stage of Awareness

Figure 28. Example of Post at the Stage of Consideration

100

Table 38. Full Content Example of Test Drive Experience
Link

Variable
Sales
F-Post (a)
C-F-Post (a)
F-Like (a)
C-F-Like (a)
F-Comment (a)
C-F-Comment (a)
F-Share (a)
C-F-Share (a)
U-Post (a)
C-U-Post (a)
U-Like (a)
C-U-Like (a)
U-Comment (a)
C-U-Comment (a)
U-Share (a)
C-U-Share (a)
TD-Post (c)
C-TD-Post (c)
Traditional Media
Spending (TMS)
Price
Google Trends (GT)
Gasoline Price
Index
Consumer Confidence
Index

Note

https://goo.gl/C25cxg

Test drive for BMW

http://goo.gl/10iH7S

Test drive for Chevrolet

http://goo.gl/tYX91M

Test drive for Nissan

Table 39. Variable Definition

Definition
Total number of offline car sales made by one car make in month t
Total number of posts by one car make at its Facebook page in month t
Total number of posts by competitors at their Facebook pages in month t
Total number of likes associated with one car make’s posts at its Facebook page
in month t
Total number of likes associated with competitors’ posts at their Facebook pages
in month t
Total number of comments associated with one car make’s posts at its Facebook
page in month t
Total number of comments associated with competitors’ posts at their Facebook
pages in month t
Total number of shares associated with one car make’s posts at its Facebook page
in month t
Total number of comments associated with competitors’ posts at their Facebook
pages in month t
Total number of posts by one car make’s users at its Facebook page in month t
Total number of posts by competitors’ users at their Facebook pages in month t
Total number of likes associated with one car make’s users posts at its Facebook
page in month t
Total number of likes associated with competitors’ users posts at their Facebook
pages in month t
Total number of comments associated with one car make’s users posts at its
Facebook page in month t
Total number of comments associated with competitors’ users posts at their
Facebook pages in month t
Total number of shares associated with one car make’s users posts at its Facebook
page in month t
Total number of shares associated with competitors’ users posts at their Facebook
pages in month t
Total number of test drive posts associated with one car make from different
Social media channels in month t
Total number of test drive posts associated with competitors from different
Social media channels in month t
Total amount of money spent by one car make on traditional media in month t
Average price of one car make in month t
The Google Trends search interest index for one car make in month t in the U.S.
The U.S. gasoline price index in month t
The conference board consumer confidence index in month t

Notes: (a) refers to the metrics at the stage of awareness (i.e., activities from firm’s Facebook page); (c) refers to the
metrics at the stage of consideration (i.e., test drive experience).

101

Table 40. Descriptive Statistics

Variable
No. of obs.
Sale
1789
F-Post (a)
1789
C-F-Post (a)
1789
F-Like (a)
1789
C-F-Like (a)
1789
F-Comment (a)
1789
C-F-Comment (a) 1789
F-Share (a)
1789
C-F-Share (a)
1789
U-Post (a)
1789
C-U-Post (a)
1789
U-Like (a)
1789
C-U-Like (a)
1789
U-Comment (a)
1789
C-U-Comment (a) 1789
U-Share (a)
1789
C-U-Share (a)
1789
TD-Post (c)
1789
C-TD-Post (c)
1789
TMS
1789
Price
1789
GT
1789
GPI
1789
CCI
1789

Mean
38331.55
33.21
916.79
81914.72
2,291,620
2352.78
65377.8
6292.4
176240.8
453.54
12439.43
10841.74
303,613
913.56
25309.42
86.83
2434.79
248.2
6972.81
34332752.43
35357.37
54.86
3.39
66.11

Std. dev.
48224.67
34.62
407.61
149010.7
2,133,963
2849.67
38698.07
12465.24
174943.3
537.73
4422.59
30856.97
386555.5
1840.7
16117.62
515.17
3915.27
334.9
4910.05
34561846.7
15608.98
19.59
0.4
12.89

Min
97
0
48
0
3,093
0
420
0
27
0
0
0
254
0
480
0
0
5
1,862
65,200
14192.5
7
2.31
40.9

Max
244,501
1,042
2,605
1,128,101
5,997,541
26,648
153,749
115,372
580,559
9,624
22,539
560,219
2,498,196
36,118
109,163
16,400
21,905
3,771
35,106
200,000,000
90,775
100
3.98
94.1

3.4.3 Bayesian Model Specification and Estimation Procedure

I employ a Bayesian framework to examine dynamics of online WOM and estimate it
using Markov Chain Monte Carlo (MCMC) method for several reasons. First, there have always
been some arguments among frequentists regarding the appropriate sample size for utilizing the
asymptotic inference (Rossi & Allenby, 2003). The Bayesian framework is compelling in the
sense that it provides a unified approach to modeling, incorporation of prior information, and
inference (Rossi & Allenby, 2003). Inference in the setting of the Bayesian framework refers to
making a posterior statement about all unobservable variables, including both parameters and, as
yet unrealized, data (prediction) (Rossi & Allenby, 2003). Thus, Bayesian inference adheres to
the likelihood principle and is conducted using formal rules of probability theory, meaning that
under mind condition, Bayes estimators are consistent, asymptotically efficient, and admissible
(Rossi & Allenby, 2003). The nature of Bayesian inference also allows this framework to
102

naturally get rid of the asymptotic assumption and deliver exact and finite sample inference
(Rossi & Allenby, 2003). Second, MCMC method is flexible and robust to estimate any
functions of parameters without the “plug-in” method due to its nature of simulation process
(Rossi & Allenby, 2003). Given the unique nature of the Bayesian framework, a considerable
literature in online WOM has utilized the Bayesian framework to study the effect of online
WOM (e.g., Ghose, Goldfarb, & Han, 2012; Sabnis & Grewal, 2015; Zhou & Duan, 2016).
My Bayesian model is specified as follows:
𝑆𝑎𝑙𝑒!,! =
𝛽! + 𝛽! 𝐴𝑤𝑎𝑟𝑒𝑛𝑒𝑠𝑠!,!!! + 𝛽!

!!! 𝐴𝑤𝑎𝑟𝑒𝑛𝑒𝑠𝑠!,!!!

+ 𝛽! 𝐶𝑜𝑛𝑠𝑖𝑑𝑒𝑟𝑎𝑡𝑖𝑜𝑛!,!!! + 𝛽!

!!! 𝐶𝑜𝑛𝑠𝑖𝑑𝑒𝑟𝑎𝑡𝑖𝑜𝑛!,!!!

+

𝛽! 𝑇𝑟𝑎𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙𝑀𝑒𝑑𝑖𝑎!,!!! + 𝛽! 𝑃𝑟𝑖𝑐𝑒!,!!! + 𝛽! 𝐺𝑜𝑜𝑔𝑙𝑒𝑇𝑟𝑒𝑛𝑑𝑠!,!!! + 𝛽! 𝐺𝑎𝑠𝑜𝑙𝑖𝑛𝑒𝑃𝑖𝑟𝑐𝑒𝐼𝑛𝑑𝑒𝑥!!! +
𝛽! 𝐶𝑜𝑛𝑠𝑢𝑚𝑒𝑟𝐶𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒𝐼𝑛𝑑𝑒𝑥!!! + 𝜏! + 𝜉! + 𝜀!,! (1)
!
𝜀!,! ~ i. i. d. 𝑁 0, 𝜎!! , 𝜏! ~ i. i. d. 𝑁 0, 𝜎!"#$
, 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 ,

𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 , 𝛽! ~ 𝑁 0, 100 ,
!
𝛽! ~ 𝑁 0, 100 , 𝜉! ~ 1 flat , 𝜎!! ~ InvGamma 0.001, 0.001 , 𝜎!"#$
~ InvGamma 0.001, 0.001

, where 𝑆𝑎𝑙𝑒!,! are offline car sales for the focal brand at time t; 𝐴𝑤𝑎𝑟𝑒𝑛𝑒𝑠𝑠!,!!! are online
WOM at the stage of awareness, measured by activities at the focal brand’s official Facebook
page (i.e., post , like, comment, or share), for the focal brand at time t-1; 𝐴𝑤𝑎𝑟𝑒𝑛𝑒𝑠𝑠!,!!! are
online WOM at the stage of awareness for the focal brand’s competitors at time t-1 (i.e., online
WOM spillover effects at the stage of awareness); 𝐶𝑜𝑛𝑠𝑖𝑑𝑒𝑟𝑎𝑡𝑖𝑜𝑛!,!!! are online WOM at the
stage of consideration, measured by test drive experience, for the focal brand at time t-1;
𝐶𝑜𝑛𝑠𝑖𝑑𝑒𝑟𝑎𝑡𝑖𝑜𝑛!,!!! are online WOM at the stage of consideration for the focal brand’s

103

competitors at time t-1 (i.e., online WOM spillover effects at the stage of consideration);
𝑇𝑟𝑎𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙𝑀𝑒𝑑𝑖𝑎!,!!! is the traditional media spending for the focal brand at time t-1;
𝑃𝑟𝑖𝑐𝑒!,!!! is the average price for the focal brand at time t-1; 𝐺𝑜𝑜𝑔𝑙𝑒𝑇𝑟𝑒𝑛𝑑𝑠!,!!! is the Google
search index in the U.S. for the focal brand at time t-1 𝐺𝑎𝑠𝑜𝑙𝑖𝑛𝑒𝑃𝑟𝑖𝑐𝑒𝐼𝑛𝑑𝑒𝑥 !!! is the U.S.
gasoline price index at time t-1; 𝐶𝑜𝑛𝑠𝑢𝑚𝑒𝑟𝐶𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒𝐼𝑛𝑑𝑒𝑥 !!! is the conference board
consumer confidence index at time t-1; 𝜏! are random effects to control for individual
heterogeneity; 𝜉! are monthly time dummies ; 𝜀!,! is the error term.
To estimate my model, I used MCMC methods (e.g., Rossi & Allenby, 2003).
Specifically, I used Metropolis-Hastings (MH) algorithm and implemented the estimation
procedures in Stata. The estimation gives the parameters for the effects of own online WOM,
competitive effects of competitors’ online WOM (i.e., spillover effects), and unobserved
heterogeneity. I took the natural log on all variables to remove the scaling effect. I used
conjugate and noninformative priors for all parameters. Specifically, 𝛽! to 𝛽! follow the normal
!
distribution with mean 0 and variance 100. 𝜎!! and 𝜎!"#$
follow the inverse gamma distribution

with parameters 0.001 and 0.001. 𝜏! follows the normal distribution. Finally, 𝜉! follows the
uniform distribution. To assess model convergence, I checked trace plots, autocorrelation plots,
histogram plots, and kernel density plots, which showed that the model specification converged.
To test the significance of parameter values, I checked whether the 95% posterior intervals
contained 0 (the norm in Bayesian estimation; e.g., Rossi & Allenby, 2003) to verify if the

104

estimated parameter is different from 0.

3.5 RESULTS
3.5.1 Main Bayesian Analysis Results

Table 41 shows my Bayesian estimation results at the post level (i.e., Facebook post and
test drive post). In this model, I ran the MCMC chain for 999,961 iterations and, I discarded the
first 200,000 iterations to ensure convergence (i.e., the burn-in time period). Because of the high
auto-correlation of parameter posterior draws, I thinned the sampling chain and kept one from
every fortieth iterations for the remaining iterations.
First of all, at the stage of awareness, I find that firm-generated posts by the focal firm (FPost (a)) have the positive impact on offline car sales of the focal firm, supporting my H1. On the
other hand, firm-generated posts by competitors (C-F-Post (a)) also have the positive impact on
offline car sales of the focal firm (i.e., positive spillover effect), thereby supporting my H2.
However, regarding user-generated posts by the focal firm (U-Post (a)) and competitors (C-UPost (a)) at the stage of awareness, I cannot find any evidence to support my H1 and H2.
The results also suggest that online WOM for the focal brand at the stage of consideration
(TD-Post (c)) positively influences offline car sales of the focal brand, supporting my H3. In
addition, the negative spillover effect is observed at the stage of consideration. Namely, test
drive posts about the competitors (C-TD-Post (c)) have the negative impact on offline car sales

105

of the focal firm (i.e., negative spillover effect), supporting my H4. The assessment of model
convergence (see Figures 29 to 39) showed that the model specification converged.
Table 41. Bayesian Estimation Results for Posts
Sales A, t
Posterior Mean
95% Credible Level
0.02 (0.009)
(0.003, 0.037)
F-Post (a) A, t-1
0.163 (0.019)
(0.126, 0.19)
C-F-Post (a) J, t-1
-0.007 (0.006)
(-0.019, 0.006)
U-Post (a) A, t-1
-0.044 (0.016)
(-0.075, -0.013)
C-U-Post (a) J, t-1
0.029 (0.012)
(0.006, 0.052)
TD-Post (c) A, t-1
-0.06 (0.015)
(-0.09, -0.029)
C-TD-Post (c) J, t-1
0.12
(0.01)
(0.1, 0.14)
TMS A, t-1
0.031 (0.062)
(-0.089, 0.153)
Price A, t-1
0.11 (0.034)
(0.041, 0.174)
GT A, t-1
0.15 (0.069)
(0.018, 0.286)
GPI A, t-1
0.237 (0.039)
(0.158, 0.313)
CCI A, t-1
Notes: Posterior means and posterior standard deviations (in parentheses) are reported, and estimates that are significant at 95%.
C- refers to competitors (i.e., spillover effects), (a) refers to the metrics at the stage of awareness, (c) refers to the metrics at the
stage of consideration (i.e., test drive experience), TMS refers traditional media spending, GPI refers to the gas price index, and
CCI refers to the consumer confidence index. Please also note that I took the natural log transformations on all of these variables
to remove the scaling effect. These criteria apply for the rest of results.
Parameters

Figure 29. Assessment of Model Convergence for F-Post (a)

106

Figure 30. Assessment of Model Convergence for C-F-Post (a)

Figure 31. Assessment of Model Convergence for U-Post (a)

Figure 32. Assessment of Model Convergence for C-U-Post (a)

107

Figure 33. Assessment of Model Convergence for TD-Post (c)

Figure 34. Assessment of Model Convergence for C-TD-Post (c)

Figure 35. Assessment of Model Convergence for TMS

108

Figure 36. Assessment of Model Convergence for Price

Figure 37. Assessment of Model Convergence for GT

Figure 38. Assessment of Model Convergence for GPI

109

Figure 39. Assessment of Model Convergence for CCI

I then turn my attention to three different mechanisms associated with posts at the stage
of awareness: Like, Comment, and Share. Table 42 shows my Bayesian estimation results at the
like level (i.e., “Like” associated with posts at Facebook and test drive post). In this model, I ran
the MCMC chain for 999,961 iterations and, I discarded the first 200,000 iterations to ensure
convergence (i.e., the burn-in time period). I also thinned the sampling chain and kept one from
every fortieth iterations for the remaining iterations in this relationship. The assessment of model
convergence (Figures 40 to 50) also suggested the model specification converged.
In this relationship, at the stage of awareness, I find that the volume of like associated
with the focal firm’s posts (F-Like (a)) does not have the impact on offline car sales of the focal
firm, suggesting that “Like” may not be an effective mechanism for the focal firm to enhance
their profit. On the other hand, the volume of like associated with competitors’ posts (C-U-Like
(a)) have the positive impact on offline car sales of the focal firm (i.e., positive spillover effects),

110

in support of H2. For likes associated with the focal firm’s user posts (U-Like (a)) and
competitors’ user posts (C-U-Like (a)), I cannot find evidence of supporting my H1 and H2,
implying that customers pay different attentions on information generated by firms and
customers at the stage of awareness. At the stage of consideration, both H3 and H4 are
supported.
Table 42. Bayesian Estimation Results for Likes
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.009 (0.005)
0.075 (0.01)
0.0007 (0.003)
0.017 (0.009)
0.025 (0.011)
-0.039 (0.014)
0.114 (0.009)
-0.011 (0.06)
0.151 (0.034)
-0.056 (0.062)
0.083 (0.04)

Sales A, t
95% Credible Level
(-0.018, 0.0009)
(0.055, 0.095)
(-0.006, 0.008)
(-0.001, 0.036)
(0.003, 0.047)
(-0.07, -0.011)
(0.095, 0.133)
(-0.129, 0.108)
(0.084, 0.218)
(-0.179, 0.067)
(0.005, 0.161)

Figure 40. Assessment of Model Convergence for F-Like
(a)

111

Figure 41. Assessment of Model Convergence for C-F-Like
(a)

Figure 42. Assessment of Model Convergence for U-Like
(a)

Figure 43. Assessment of Model Convergence for C-U-Like (a)

112

Figure 44. Assessment of Model Convergence for TD-Post (c)

Figure 45. Assessment of Model Convergence for C-TD-Post (c)

Figure 46. Assessment of Model Convergence for TMS

113

Figure 47. Assessment of Model Convergence for Price

Figure 48. Assessment of Model Convergence for GT

Figure 49. Assessment of Model Convergence for GPI

114

Figure 50. Assessment of Model Convergence for CCI

Table 43 shows my Bayesian estimation results at the comment level (i.e., “Comment”
associated with posts at Facebook and test drive post). In this model, I ran the MCMC chain for
999,961 iterations and, I discarded the first 200,000 iterations to ensure convergence (i.e., the
burn-in time period). I also thinned the sampling chain and kept one from every fortieth
iterations for the remaining iterations in this relationship. The assessment of model convergence
(Figures 51 to 61) also suggested the model specification converged.
In this relationship, a very interesting pattern is observed. First, similar to results at the
like level, at the stage of awareness, I only find that the volume of comment associated with
competitors’ posts (C-F-Comment (a)) has the positive impact on offline car sales of the focal
firm (i.e., positive spillover effects), in support of H2. The volume of comment associated with
the focal firm’s posts again does not have the impact on offline car sales of the focal firm,
rejecting H1. The mechanism of comment also changed how customers appreciate online WOM
at the stage of consideration. For example, in

this relationship, test drive posts related to the
115

focal brand (TD-Post (c)) do not have any impact on offline car sales of the focal brand, rejecting
my H3. On the other hand, I find that negative spillover effects still exist in this relationship, in
support of H4.
Table 43. Bayesian Estimation Results for Comments
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.002 (0.005)
0.107 (0.012)
-0.003 (0.004)
0.009 (0.011)
0.02 (0.011)
-0.057 (0.014)
0.113 (0.009)
0.007 (0.059)
0.163 (0.02)
0.138 (0.052)
0.206 (0.038)

Sales A, t
95% Credible Level
(-0.012, 0.008)
(0.084, 0.13)
(-0.011, 0.006)
(-0.012, 0.031)
(-0.001, 0.043)
(-0.085, -0.029)
(0.093, 0.132)
(-0.109, 0.126)
(0.125, 0.204)
(0.037, 0.239)
(0.132, 0.282)

Figure 51. Assessment of Model Convergence for F-Comment (a)

116

Figure 52. Assessment of Model Convergence for C-F-Comment
(a)

Figure 53. Assessment of Model Convergence for U-Comment (a)

.

Figure 54. Assessment of Model Convergence for C-U-Comment (a)

117

Figure 55. Assessment of Model Convergence for TD-Post (c)

Figure 56. Assessment of Model Convergence for C-TD-Post (c)

Figure 57. Assessment of Model Convergence for TMS

118

Figure 58. Assessment of Model Convergence for Price

Figure 59. Assessment of Model Convergence for GT

Figure 60. Assessment of Model Convergence for GPI

119

Figure 61. Assessment of Model Convergence for CCI

Table 44 shows my Bayesian estimation results at the share level (i.e., “Share” associated
with posts at Facebook and test drive post). In this model, I ran the MCMC chain for 999,961
iterations and, I discarded the first 200,000 iterations to ensure convergence (i.e., the burn-in
time period). I also thinned the sampling chain and kept one from every fortieth iterations for the
remaining iterations in this relationship. The assessment of model convergence (Figures 62 to
72) also suggested the model specification converged.
The results at the share level show another interesting patterns, suggesting that not very
mechanism at the firm’s Facebook page has the equal impact on firm performance and different
mechanism also changes the effect of online WOM across the stages of awareness and
consideration. For example, I observe that at the stage of awareness, the more shares associated
with competitors’ posts (C-F-Share (a)), the most offline sales of the focal brand (i.e., positive
spillover effects), in support of H2. Regarding the volume of share associated with the focal

120

brand’s user posts (U-Share (a)) and competitors’ user posts (C-U-Share (a)), I, again, cannot
find any evidence to support my H1 and H2. Finally, in this relationship, at the stage of
consideration, I find that test drive posts associated with the focal brand have the positive impact
on offline car sales of the focal brand, supporting my H3. On the other hand, the mechanism of
share does change the way of how competitors’ customers appreciate the effect of online WOM
at the stage of consideration. I cannot find evidence to support spillover effects in this
relationship, rejecting my H4.
The results from the whole sample analysis could be summarized as the following. First,
at the stage of awareness, online WOM regarding the focal brand and competitors has the
positive impact on offline car sales of the focal firm. This implies that in the U.S. automobile
industry customers think that abstract and general information (i.e., characteristics of information
for the stage of awareness) for brand A is accessible and diagnostic of brand B and customers
will use perception of brand A’s quality to infer quality of brand B. This linking mechanism
therefore has the positive impact on offline car sales of the focal brand. Second, once customers
decide to receive more concrete or detailed information (i.e., characteristics of information for
the stage of consideration), online WOM regarding the focal brand and competitors demonstrates
the opposite effect on offline car sales of the focal firm with the positive effect from the focal
brand and the negative effect from the competitors (i.e., negative spillover effect). Third, at the
stage of awareness, information initiated by firms (e.g., F-Post (a), F-Share (a)) and users (e.g.,

121

U-Post (a), U-Share (a)) shows dramatically different impacts on firm performance, suggesting
that customers place different weight on these two different sources of online WOM when
making their purchase decisions. Finally, the results suggest that different mechanisms showed at
the stage of awareness (i.e., post, like, comment, and share) do not have the equal impact on firm
performance and these varied mechanisms also change how customers appreciate the effect of
online WOM at the stage of consideration. For example, the results suggest that the volume of
posts by firms is the most effective mechanism to influence customer purchase decisions and the
“secondary” mechanisms (like, comment, share) are not very effective.
Table 44. Bayesian Estimation Results for Shares
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.001 (0.003)
0.049 (0.007)
0.006 (0.003)
-0.017 (0.007)
0.034 (0.011)
-0.024 (0.016)
0.118 (0.009)
0.011 (0.061)
0.108 (0.033)
0.268 (0.072)
0.224 (0.038)

Sales A, t
95% Credible Level
(-0.007, 0.005)
(0.035, 0.062)
(-0.0007, 0.013)
(-0.031, -0.004)
(0.011, 0.056)
(-0.054, 0.007)
(0.098, 0.138)
(-0.108, 0.13)
(0.042, 0.172)
(0.128, 0.408)
(0.149, 0.297)

Figure 62. Assessment of Model Convergence for F-Share
(a)

122

Figure 63. Assessment of Model Convergence for C-F-Share (a)

Figure 64. Assessment of Model Convergence for U-Share
(a)

Figure 65. Assessment of Model Convergence for C-U-Share (a)

123

Figure 66. Assessment of Model Convergence for TD-Post (c)

Figure 67. Assessment of Model Convergence for C-TD-Post (c)

Figure 68. Assessment of Model Convergence for TMS

124

Figure 69. Assessment of Model Convergence for Price

Figure 70. Assessment of Model Convergence for GT

Figure 71. Assessment of Model Convergence for GPI

125

Figure 72. Assessment of Model Convergence for CCI

3.5.2 Sample Split Bayesian Analysis Results on the Origin of Brands

Prior research suggests that country of origin effects would moderate the spillover effects
(Borah & Tellis, 2016; Maheswaran & Chen, 2006) because consumers might use the origin as
an attribute and make similar inferences for brands that belong to the same origin (Hong &
Wyer, 1990). Therefore, in this sample split analysis, I examine how the origin of brand may
moderate dynamics of online WOM by considering three origins of brand: Asian-based,
European-based, and US-based brands18. Tables 45 to 47 show descriptive statistics for Asianbased, European-based, and US-based brands, respectively.
Table 45. Descriptive Statistics for Asian-Based Brands

Variable
No. of obs.
Sale
745
F-Post (a)
745
C-F-Post (a)
745
F-Like (a)
745
C-F-Like (a)
745
F-Comment (a)
745
C-F-Comment (a) 745
F-Share (a)
745
C-F-Share (a)
745
U-Post (a)
745

Mean
43368.12
32.8
913.54
55067.66
2,302,998
1884.11
65603.57
3418.65
177862.8
364.48

Std. dev.
43,517
23.69
404.5
96237.15
2,153,474
2533.42
38718.41
543,1
177686.9
387.56

18

Min
2,923
0
53
0
4,583
0
863
0
84
0

Max
210,134
344
2,605
722,146
6,000,000
26,648
153,749
41,089
580,559
4,625

Asian-based: Acura, Honda, Hyundai, Infiniti, KIA, Lexus, Mazda, Mitsubishi, Nissan, Scion, Subaru, and
Toyota; European-based: Audi, BMW, Jaguar, Land Rover, Mercedes-Benz, Porsche, Saab, Volkswagen, and
Volvo; US-based: Buick, Cadillac, Chevrolet, Chrysler, Dodge, FIAT, Ford, Jeep, and Lincoln.
126

Table 45 (cont’d)
C-U-Post (a)
U-Like (a)
C-U-Like (a)
U-Comment (a)
C-U-Comment (a)
U-Share (a)
C-U-Share (a)
TD-Post (c)
C-TD-Post (c)
TMS
Price
GT
GPI
CCI

745
745
745
745
745
745
745
745
745
745
745
745
745
745

12523.47
6866.44
305631.8
666.67
25504.07
69.16
2432.74
270.98
7023.92
38013691.28
28531.08
52.74
3.39
66.06

4339.73
22293.83
391109.5
1593.09
16093.13
616.79
3904.39
369.72
4989.19
32083898.34
9927.16
20.2
0.4
12.87

971
0
808
0
923
0
0
11
1,862
1,500,000
14192.5
10
2.31
40.9

Table 46. Descriptive Statistics for European-Based Brands

Variable
No. of obs.
Sale
489
F-Post (a)
489
C-F-Post (a)
489
F-Like (a)
489
C-F-Like (a)
489
F-Comment (a)
489
C-F-Comment (a) 489
F-Share (a)
489
C-F-Share (a)
489
U-Post (a)
489
C-U-Post (a)
489
U-Like (a)
489
C-U-Like (a)
489
U-Comment (a)
489
C-U-Comment (a) 489
U-Share (a)
489
C-U-Share (a)
489
TD-Post (c)
489
C-TD-Post (c)
489
TMS
489
Price
489
GT
489
GPI
489
CCI
489

Mean
12255.04
32.17
934.83
150018.5
2,252,274
2752.16
65926.9
11349.78
173222.3
376.35
12735.96
11050.07
307736.4
499.99
26123.09
89.04
2469.14
188.09
7050.71
14894591.82
51335.55
57.08
3.42
66.27

Std. dev.
11221.09
28.73
405.91
231163.8
2,074,290
3217.59
37798.81
19854.05
170249.1
524.82
4423.51
36830.84
384425.1
887.49
16149.87
472.43
3940.1
194.82
4932.11
13042812.82
16294.94
14.42
0.4
12.89

Min
97
0
48
0
4,226
0
826
0
84
0
647
0
254
0
480
0
0
5
1,983
65,200
23798.6
26
2.31
40.9

Table 47. Descriptive Statistics for US-Based Brands

Variable
No. of obs.
Sale
555
F-Post (a)
555
C-F-Post (a)
555
F-Like (a)
555
C-F-Like (a)
555
F-Comment (a)
555
C-F-Comment (a) 555
F-Share (a)
555
C-F-Share (a)
555
U-Post (a)
555
C-U-Post (a)
555
U-Like (a)
555
C-U-Like (a)
555
U-Comment (a)
555
C-U-Comment (a) 555
U-Share (a)
555
C-U-Share (a)
555
TD-Post (c)
555
C-TD-Post (c)
555
TMS
555
Price
555

Mean
54546.26
34.66
905.28
57968.14
2,311,868
2630.02
64590.92
5693.99
176722.9
641.12
12065.34
15994.46
297731.1
1609.36
24331.22
108.61
2407.29
270.59
6835.58
46518267.92
30442.51

Std. dev.
62880.57
48.83
413.41
81127.72
2,163,995
2820.77
39502.81
8978.39
175590.4
661.59
4513.38
34099.1
385749.9
2473.05
16101.68
384.94
3914.77
374.96
4787.97
42715942.85
10386.73
127

Min
500
0
50
0
3,093
0
420
0
27
0
857
0
613
0
690
0
0
10
1,885
238,700
16,000

22,539
463,332
2,500,000
36,118
109,059
16,400
21,905
3,771
35,106
200,000,000
50928.1
100
3.98
94.1
Max
44,005
436
2,579
1,100,000
5,900,000
20,577
151,722
115,372
572,859
9,624
22,520
560,219
2,500,000
10,539
109,163
5,283
21,904
1,595
35,072
75,000,000
90,775
91
3.98
94.1

Max
244,501
1,042
2,584
490,822
6,000,000
25,734
152,256
62,457
578,089
4,918
22,459
252,535
2,500,000
28,081
108,643
6,362
21,905
3,348
34,817
190,000,000
55828.8

Table 47 (cont’d)
GT
GPI
CCI

555
555
555

55.75
3.39
66.04

22.29
0.4
12.93

7
2.31
40.9

100
3.98
94.1

Tables 48 to 50 show my sample split Bayesian estimation results at the post level (i.e.,
Facebook post and test drive post) for Asian-based, European-based, and US-based brands,
respectively. In these models, I ran the MCMC chain for 999,961 iterations and, I discarded the
first 200,000 iterations to ensure convergence. Because of the high auto-correlation of parameter
posterior draws, I thinned the sampling chain and kept one from every fortieth iterations for the
remaining iterations. The assessment of model convergence (Figures 73 to 105) also suggested
the model specification converged for each relationship.
The results in the post level indicate heterogeneity across three origins of brands
regarding the effects of online WOM on offline car sales of the focal brand at the stages of
awareness and consideration. First, at the stage of awareness, I observe that online WOM
initiated by the focal firm (F-Post (a)) has the positive impact on offline car sales for the
European-based group only, supporting H1. Regarding spillover effects at the stage of
awareness, all three groups share similar patterns, namely, the more posts by competitors within
the given group (C-F-Post (a)), the more offline car sales of the focal brand in the given group,
with the strongest magnitude of the coefficient from the Asian-based group. Thus, H2 is
supported for three different groups. Regarding posts generated by users at the stage of
awareness (U-Post (a) or C-U-Post (a)), I find that the significant effect of U-Post (a) on offline
car sales of the US-based group, although the direction of this effect is opposite as hypothesized,
128

thereby rejecting my H1. Interestingly, for the Asian-based and European-based group, I also
observe negative spillover effects from posts by competitors’ user posts (C-U-Post (a)), which
show the opposite direction as hypothesized. Therefore, these results provide further evidence
that when considering online WOM to make better purchase decisions, customers place different
weights on those WOM by firms and WOM by other customers.
Finally, with respect to online WOM at the stage of consideration, H3 and H4 are
supported for the Asian-based group only. For European-based and US-based group, customers
seek detailed or other sufficient information from other sources rather than test drive posts to
help them make better purchase decisions at the stage of consideration.
Table 48. Bayesian Estimation Results for Posts (Asian-based)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.008 (0.012)
0.172 (0.027)
0.003 (0.012)
-0.083 (0.024)
0.038 (0.017)
-0.053 (0.022)
0.072 (0.018)
0.214 (0.092)
0.118 (0.044)
0.069 (0.099)
0.218 (0.055)

Sales A, t
95% Credible Level
(-0.014, 0.03)
(0.12, 0.23)
(-0.019, 0.026)
(-0.13, -0.035)
(0.004, 0.072)
(-0.095, -0.011)
(0.035, 0.11)
(0.034, 0.397)
(0.03, 0.203)
(-0.126, 0.263)
(0.108, 0.328)

Table 49. Bayesian Estimation Results for Posts (European-based)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1

Posterior Mean
0.055 (0.02)
0.132 (0.043)
-0.006 (0.01)
-0.067 (0.034)
0.035 (0.023)
-0.059 (0.03)
0.12 (0.016)
-0.02 (0.124)
0.33 (0.1)

129

Sales A, t
95% Credible Level
(0.013, 0.096)
(0.048, 0.216)
(-0.027, 0.016)
(-0.133, -0.0006)
(-0.01, 0.08)
(-0.12, 0.008)
(0.09, 0.15)
(-0.266, 0.227)
(0.133, 0.526)

Table 49 (cont’d)
GPI A, t-1
CCI A, t-1

0.44 (0.16)
0.27 (0.084)

(0.135, 0.749)
(0.107, 0.438)

Table 50. Bayesian Estimation Results for Posts (US-based)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.011 (0.015)
(-0.02, 0.04)
0.164 (0.029)
(0.106, 0.221)
-0.041 (0.013)
(-0.067, -0.015)
0.029 (0.027)
(-0.025, 0.083)
0.013 (0.02)
(-0.027, 0.054)
-0.046 (0.028)
(-0.101, 0.01)
0.167 (0.02)
(0.1, 0.14)
0.15 (0.113)
(-0.089, 0.153)
-0.054 (0.061)
(0.041, 0.174)
0.17 (0.115)
(0.018, 0.286)
0.279 (0.065)
(0.158, 0.313)

Figure 73. Assessment of Model Convergence for F-Post (a) (Asian-based)

130

Figure 74. Assessment of Model Convergence for C-F-Post (a) (Asian-based)

Figure 75. Assessment of Model Convergence for U-Post (a) (Asian-based)

Figure 76. Assessment of Model Convergence for C-U-Post (a) (Asian-based)

131

Figure 77. Assessment of Model Convergence for TD-Post (c) (Asian-based)

Figure 78. Assessment of Model Convergence for C-TD-Post (c) (Asian-based)

Figure 79. Assessment of Model Convergence for TMS (Asian-based)

132

Figure 80. Assessment of Model Convergence for Price (Asian-based)

Figure 81. Assessment of Model Convergence for GT (Asian-based)

Figure 82. Assessment of Model Convergence for GPI (Asian-based)

133

Figure 83. Assessment of Model Convergence for CCI (Asian-based)

Figure 84. Assessment of Model Convergence for F-Post (a) (European-based)

Figure 85. Assessment of Model Convergence for C-F-Post (a) (European-based)

134

Figure 86. Assessment of Model Convergence for U-Post (a) (European-based)

Figure 87. Assessment of Model Convergence for C-U-Post (a) (European-based)

Figure 88. Assessment of Model Convergence for TD-Post (c) (European-based)

135

Figure 89. Assessment of Model Convergence for C-TD-Post (c) (European-based)

Figure 90. Assessment of Model Convergence for TMS (European-based)

Figure 91. Assessment of Model Convergence for Price (European-based)

136

Figure 92. Assessment of Model Convergence for GT (European-based)

Figure 93. Assessment of Model Convergence for GPI (European-based)

Figure 94. Assessment of Model Convergence for CCI (European-based)

137

Figure 95. Assessment of Model Convergence for F-Post (a) (US-based)

Figure 96. Assessment of Model Convergence for C-F-Post (a) (US-based)

Figure 97. Assessment of Model Convergence for U-Post (a) (US-based)

138

Figure 98. Assessment of Model Convergence for C-U-Post (a) (US-based)

Figure 99. Assessment of Model Convergence for TD-Post (c) (US-based)

Figure 100. Assessment of Model Convergence for C-TD-Post (c) (US-based)

139

Figure 101. Assessment of Model Convergence for TMS (US-based)

Figure 102. Assessment of Model Convergence for Price (US-based)

Figure 103. Assessment of Model Convergence for GT (US-based)

140

Figure 104. Assessment of Model Convergence for GPI (US-based)

Figure 105. Assessment of Model Convergence for CCI (US-based)

I then turn my attention to examine how these three sub categories may vary the
relationships for three different mechanisms at firms’ Facebook pages: Like, Comment, and
Share. Tables 51 to 53 show my sample split Bayesian estimation results at the like level (i.e.,
“Like” associated with posts at Facebook and test drive post) for Asian-based, European-based,
and US-based brands, respectively. In these models, I also ran the MCMC chain for 999,961
iterations and, I discarded the first 200,000

iterations to ensure convergence. Because of
141

the high auto-correlation of parameter posterior draws, I thinned the sampling chain and kept one
from every fortieth iterations for the remaining iterations. The assessment of model convergence
(see Figures 106 to 138) suggested that the model specification converged for each relationship.
In this set of analysis, the results provide further evidence on heterogeneity across origins
of brands regarding the effects of online WOM in the form of like at the stage of awareness and
test drive posts at the stage of consideration. Consistent with the main results (see Table 42), the
volume of like associated with competitors’ posts (C-F-Like (a)) show positive spillover effects
on offline car sales of the focal firm in each given group, supporting H2. However, the volume of
like associated with the focal brand’s posts (F-Like (a)) is still not very effective mechanism to
influence focal firm performance, thereby rejecting H1. I also find that for the US-based group,
the volume of like associated with competitors’ user posts (C-U-Like (a)) has positive spillover
effects on offline car sales of the US-based focal brand, supporting H2. Finally, regarding online
at the stage of consideration, none of relationships (H3 and H4) are supported in this particular
relationship.
Table 51. Bayesian Estimation Results for Likes (Asian-based)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.003 (0.008)
(-0.012, 0.0181)
0.063 (0.015)
(0.034, 0.093)
-0.002 (0.005)
(-0.013, 0.008)
0.01 (0.014)
(-0.018, 0.038)
0.032 (0.017)
(-0.001, 0.066)
-0.039 (0.02)
(-0.079, 0.001)
0.062 (0.018)
(0.027, 0.098)
0.159 (0.089)
(-0.017, 0.335)
0.142 (0.044)
(0.054, 0.229)
-0.149 (0.089)
(-0.322, 0.025)
0.103 (0.057)
(-0.009, 0.215)

142

Table 52. Bayesian Estimation Results for Likes (European-based)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.011 (0.009)
0.116 (0.022)
0.009 (0.006)
-0.025 (0.019)
0.041 (0.022)
-0.005 (0.028)
0.114 (0.016)
-0.014 (0.119)
0.418 (0.098)
0.201 (0.134)
0.085 (0.086)

Sales A, t
95% Credible Level
(-0.031, 0.008)
(0.073, 0.159)
(-0.002, 0.021)
(-0.063, 0.013)
(-0.003, 0.085)
(-0.104, 0.004)
(0.082, 0.145)
(-0.248, 0.218)
(0.224, 0.611)
(-0.065, 0.46)
(-0.083, 0.253)

Table 53. Bayesian Estimation Results for Likes (US-based)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.009 (0.008)
0.052 (0.015)
-0.023 (0.007)
0.068 (0.016)
0.017 (0.019)
-0.025 (0.026)
0.166 (0.018)
0.087 (0.107)
-0.052 (0.058)
-0.016 (0.11)
0.087 (0.066)

Sales A, t
95% Credible Level
(-0.026, 0.007)
(0.022, 0.082)
(-0.036, -0.009)
(0.038, 0.099)
(-0.021, 0.055)
(-0.076, 0.027)
(0.13, 0.202)
(-0.122, 0.296)
(-0.167, 0.061)
(-0.218, 0.181)
(-0.042, 0.217)

Figure 106. Assessment of Model Convergence for F-Like (a) (Asian-based)

143

Figure 107. Assessment of Model Convergence for C-F-Like (a) (Asian-based)

Figure 108. Assessment of Model Convergence for U-Like (a) (Asian-based)

Figure 109. Assessment of Model Convergence for C-U-Like (a) (Asian-based)

144

Figure 110. Assessment of Model Convergence for TD-Post (c) (Asian-based)

Figure 111. Assessment of Model Convergence for C-TD-Post (c) (Asian-based)

Figure 112. Assessment of Model Convergence for TMS (Asian-based)

145

Figure 113. Assessment of Model Convergence for Price (Asian-based)

Figure 114. Assessment of Model Convergence for GT (Asian-based)

Figure 115. Assessment of Model Convergence for GPI (Asian-based)

146

Figure 116. Assessment of Model Convergence for CCI (Asian-based)

Figure 117. Assessment of Model Convergence for F-Like (a) (European-based)

Figure 118. Assessment of Model Convergence for C-F-Like (a) (European-based)

147

Figure 119. Assessment of Model Convergence for U-Like (a) (European-based)

Figure 120. Assessment of Model Convergence for C-U-Like (a) (European-based)

Figure 121. Assessment of Model Convergence for TD-Post (c) (European-based)

148

Figure 122. Assessment of Model Convergence for C-TD-Post (c) (European-based)

Figure 123. Assessment of Model Convergence for TMS (European-based)

Figure 124. Assessment of Model Convergence for Price (European-based)

149

Figure 125. Assessment of Model Convergence for GT (European-based)

Figure 126. Assessment of Model Convergence for GPI (European-based)

Figure 127. Assessment of Model Convergence for CCI (European-based)

150

Figure 128. Assessment of Model Convergence for F-Like (a) (US-based)

Figure 129. Assessment of Model Convergence for C-F-Like (a) (US-based)

Figure 130. Assessment of Model Convergence for U-Like (a) (US-based)

151

Figure 131. Assessment of Model Convergence for C-U-Like (a) (US-based)

Figure 132. Assessment of Model Convergence for TD-Post (c) (US-based)

Figure 133. Assessment of Model Convergence for C-TD-Post (c) (US-based)

152

Figure 134. Assessment of Model Convergence for TMS (US-based)

Figure 135. Assessment of Model Convergence for Price (US-based)

Figure 136. Assessment of Model Convergence for GT (US-based)

153

Figure 137. Assessment of Model Convergence for GPI (US-based)

Figure 138. Assessment of Model Convergence for CCI (US-based)

Tables 54 to 56 show my sample split Bayesian estimation results at the comment level
(i.e., “Comment” associated with posts at Facebook and test drive post) for Asian-based,
European-based, and US-based brands, respectively. In these models, I ran the MCMC chain for
1,849,921 iterations and, I discarded the first 250,000 iterations to ensure convergence. Because
of the high auto-correlation of parameter posterior draws, I thinned the sampling chain and kept
one from every eightieth iterations for the remaining iterations. The assessment of model

154

convergence (see Figures 139 to 171) suggested that the model specification converged for each
relationship.
First, I find that consistent with main results (see Table 43), the volume of comments
related to competitors’ posts (C-F-Comment (a)) has the positive spillover effects across three
different groups, supporting H2. On the other hand, the volume of comment associated with the
focal brand’s posts (F-Comment (a)) does not have any impact on offline car sales of the focal
brand, rejecting H1. The mechanism of comment related to the focal brand’s user posts (UComment (a)) and competitors’ user posts (C-U-Comment (a)) shows another interesting
patterns. For example, I find that for the European-based group, the volume of comment
associated with the focal brand’s user posts has the positive impact on offline car sales of the
focal brand, supporting H1. For the US-based group, I observe that the effect of comment
associated with the focal brand’s user posts has the negative impact on offline car sales of the
focal firm, which contradicts with my H1 and positive spillover effects (C-U-Comment (a)) are
observed, thereby supporting H2. Finally, consistent with main results, at the stage of
consideration, test drive posts regarding competitors have negative spillover effects for the
Asian-based and European-based group, supporting H4. However, for the US-based group, at the
stage of consideration, online WOM for the focal brand and online WOM for competitors do not
have any impact on offline car sales of the focal brand.

155

Table 54. Bayesian Estimation Results for Comments (Asian-based)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.011 (0.008)
(-0.003, 0.026)
0.089 (0.017)
(0.055, 0.124)
0.0001 (0.007)
(-0.014, 0.014)
-0.011 (0.017)
(-0.044, 0.022)
0.032 (0.017)
(-0.002, 0.066)
-0.057 (0.021)
(-0.098, -0.012)
0.067 (0.019)
(0.031, 0.104)
0.189 (0.093)
(0.007, 0.369)
0.126 (0.046)
(0.037, 0.215)
0.043 (0.082)
(-0.119, 0.206)
0.212 (0.055)
(0.105, 0.318)

Table 55. Bayesian Estimation Results for Comments (European-based)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.021 (0.012)
0.155 (0.027)
0.02 (0.009)
-0.035 (0.024)
0.03 (0.022)
-0.077 (0.028)
0.118 (0.016)
-0.048 (0.119)
0.33 (0.098)
0.303 (0.125)
0.207 (0.08)

Sales A, t
95% Credible Level
(-0.046, 0.003)
(0.102, 0.208)
(0.003, 0.037)
(-0.081, 0.012)
(-0.014, 0.075)
(-0.123, -0.012)
(0.086, 0.14)
(-0.28, 0.191)
(0.141, 0.526)
(0.058, 0.545)
(0.052, 0.366)

Table 56. Bayesian Estimation Results for Comments (US-based)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.003 (0.008)
0.083 (0.018)
-0.043 (0.008)
0.089 (0.019)
0.016 (0.019)
-0.038 (0.027)
0.162 (0.019)
0.142 (0.106)
-0.049 (0.05)
0.237 (0.092)
0.233 (0.062)

156

Sales A, t
95% Credible Level
(-0.021, 0.014)
(0.049, 0.12)
(-0.059, -0.027)
(0.053, 0.127)
(-0.023, 0.055)
(-0.09, 0.014)
(0.125, 0.199)
(-0.066, 0.349)
(-0.147, 0.053)
(0.055, 0.418)
(0.11, 0.35)

Figure 139. Assessment of Model Convergence for F-Comment (a) (Asian-based)

Figure 140. Assessment of Model Convergence for C-F-Comment (a) (Asian-based)

Figure 141. Assessment of Model Convergence for U-Comment (a) (Asian-based)

157

Figure 142. Assessment of Model Convergence for C-U-Comment (a) (Asian-based)

Figure 143. Assessment of Model Convergence for TD-Post (c) (Asian-based)

Figure 144. Assessment of Model Convergence for C-TD-Post (c) (Asian-based)

158

Figure 145. Assessment of Model Convergence for TMS (Asian-based)

Figure 146. Assessment of Model Convergence for Price (Asian-based)

Figure 147. Assessment of Model Convergence for GT (Asian-based)

159

Figure 148. Assessment of Model Convergence for GPI (Asian-based)

Figure 149. Assessment of Model Convergence for CCI (Asian-based)

Figure 150. Assessment of Model Convergence for F-Comment (a) (European-based)

160

Figure 151. Assessment of Model Convergence for C-F-Comment (a) (European-based)

Figure 152. Assessment of Model Convergence for U-Comment (a) (European-based)

Figure 153. Assessment of Model Convergence for C-U-Comment (a) (European-based)

161

Figure 154. Assessment of Model Convergence for TD-Post (c) (European-based)

Figure 155. Assessment of Model Convergence for C-TD-Post (c) (European-based)

Figure 156. Assessment of Model Convergence for TMS (European-based)

162

Figure 157. Assessment of Model Convergence for Price (European-based)

Figure 158. Assessment of Model Convergence for GT (European-based)

Figure 159. Assessment of Model Convergence for GPI (European-based)

163

Figure 160. Assessment of Model Convergence for CCI (European-based)

Figure 161. Assessment of Model Convergence for F-Comment (a) (US-based)

Figure 162. Assessment of Model Convergence for C-F-Comment (a) (US-based)

164

Figure 163. Assessment of Model Convergence for U-Comment (a) (US-based)

Figure 164. Assessment of Model Convergence for C-U-Comment (a) (US-based)

Figure 165. Assessment of Model Convergence for TD-Post (c) (US-based)

165

Figure 166. Assessment of Model Convergence for C-TD-Post (c) (US-based)

Figure 167. Assessment of Model Convergence for TMS (US-based)

Figure 168. Assessment of Model Convergence for Price (US-based)

166

Figure 169. Assessment of Model Convergence for GT (US-based)

Figure 170. Assessment of Model Convergence for GPI (US-based)

Figure 171. Assessment of Model Convergence for CCI (US-based)

167

Tables 57 to 59 show my sample split Bayesian estimation results at the share level (i.e.,
“Share” associated with posts at Facebook and test drive post) for Asian-based, European-based,
and US-based brands, respectively. In these models, I ran the MCMC chain for 999,961
iterations and, I discarded the first 200,000 iterations to ensure convergence. Because of the high
auto-correlation of parameter posterior draws, I thinned the sampling chain and kept one from
every fortieth iterations for the remaining iterations. The assessment of model convergence (see
Figures 172 to 204) suggested that the model specification converged for each relationship.
First, I find that consistent with main results (see Table 41), the mechanism of share
associated with the focal brand’s posts (F-Share (a)) is not effective in influencing offline car
sales of the focal brand across three groups, rejecting H1. However, the volume of share
associated with competitors’ posts has positive spillover effects on offline car sales of the focal
brand, thereby supporting H2. Regarding the effect related to user posts (U-Share (a) and C-UShared (a)), I cannot find any support for my H1 and H2. The results also suggest that share
associated with the focal brand’s user posts (U-Share (a)) and competitors’ user posts (C-UShare (a)) is also not very effective in influencing offline car sales of the focal brand, thereby
rejecting both H1 and H2. Finally, I find that at the stage of consideration, test drive posts about
the focal brand have the positive impact on offline car sales of the focal brand for the Asianbased and European-based group only, supporting H3.
To summarize, this series of analysis suggests that in the U.S. automobile industry,

168

customers do use the origin of brands as an attribute and make similar inferences for brands that
belong to the same origin. Besides, customers from different groups do appreciate different
mechanism significantly about how these mechanisms may influence their purchase decisions.
Table 57. Bayesian Estimation Results for Shares (Asian-based)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.000003 (0.004)
0.043 (0.009)
0.001 (0.005)
-0.007 (0.009)
0.039 (0.017)
-0.029 (0.022)
0.053 (0.018)
0.187 (0.088)
0.125 (0.042)
0.031 (0.1)
0.217 (0.052)

Sales A, t
95% Credible Level
(-0.008, 0.009)
(0.024, 0.062)
(-0.008, 0.011)
(-0.025, 0.012)
(0.006, 0.073)
(-0.073, 0.015)
(0.017, 0.089)
(0.013, 0.36)
(0.042, 0.207)
(-0.17, 0.231)
(0.114, 0.318)

Table 58. Bayesian Estimation Results for Shares (European-based)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.009 (0.008)
0.035 (0.015)
0.007 (0.007)
-0.11 (0.016)
0.049 (0.023)
-0.041 (0.032)
0.118 (0.016)
-0.027 (0.12)
0.37 (0.097)
0.42 (0.158)
0.27 (0.082)

Sales A, t
95% Credible Level
(-0.006, 0.025)
(0.004, 0.065)
(-0.008, 0.021)
(-0.04, 0.02)
(0.003, 0.095)
(-0.105, 0.022)
(0.087, 0.149)
(-0.27, 0.21)
(0.179, 0.563)
(0.107, 0.73)
(0.11, 0.43)

Table 59. Bayesian Estimation Results for Shares (US-based)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1

Posterior Mean
-0.008 (0.005)
0.06 (0.011)
0.011 (0.006)
-0.028 (0.011)
0.022 (0.02)
-0.009 (0.028)
0.172 (0.019)
0.13 (0.11)
-0.129 (0.055)

169

Sales A, t
95% Credible Level
(-0.018, 0.003)
(0.04, 0.082)
(-0.0005, 0.023)
(-0.05, -0.007)
(-0.018, 0.062)
(-0.065, 0.046)
(0.136, 0.208)
(-0.087, 0.35)
(-0.233, -0.019)

Table 59 (cont’d)
GPI A, t-1
CCI A, t-1

0.4 (0.12)
0.22 (0.062)

(0.171, 0.635)
(0.098, 0.34)

Figure 172. Assessment of Model Convergence for F-Share (a) (Asian-based)

Figure 173. Assessment of Model Convergence for C-F-Share (a) (Asian-based)

170

Figure 174. Assessment of Model Convergence for U-Share (a) (Asian-based)

Figure 175. Assessment of Model Convergence for C-U-Share (a) (Asian-based)

Figure 176. Assessment of Model Convergence for TD-Post (c) (Asian-based)

171

Figure 177. Assessment of Model Convergence for C-TD-Post (c) (Asian-based)

Figure 178. Assessment of Model Convergence for TMS (Asian-based)

Figure 179. Assessment of Model Convergence for Price (Asian-based)

172

Figure 180. Assessment of Model Convergence for GT (Asian-based)

Figure 181. Assessment of Model Convergence for GPI (Asian-based)

Figure 182. Assessment of Model Convergence for CCI (Asian-based)

173

Figure 183. Assessment of Model Convergence for F-Share (a) (European-based)

Figure 184. Assessment of Model Convergence for C-F-Share (a) (European-based)

Figure 185. Assessment of Model Convergence for U-Share (a) (European-based)

174

Figure 186. Assessment of Model Convergence for C-U-Share (a) (European-based)

Figure 187. Assessment of Model Convergence for TD-Post (c) (European-based)

Figure 188. Assessment of Model Convergence for C-TD-Post (c) (European-based)

175

Figure 189. Assessment of Model Convergence for TMS (European-based)

Figure 190. Assessment of Model Convergence for Price (European-based)

Figure 191. Assessment of Model Convergence for GT (European-based)

176

Figure 192. Assessment of Model Convergence for GPI (European-based)

Figure 193. Assessment of Model Convergence for CCI (European-based)

Figure 194. Assessment of Model Convergence for F-Share (a) (US-based)

177

Figure 195. Assessment of Model Convergence for C-F-Share (a) (US-based)

Figure 196. Assessment of Model Convergence for U-Share (a) (US-based)

Figure 197. Assessment of Model Convergence for C-U-Share (a) (US-based)

178

Figure 198. Assessment of Model Convergence for TD-Post (c) (US-based)

Figure 199. Assessment of Model Convergence for C-TD-Post (c) (US-based)

Figure 200. Assessment of Model Convergence for TMS (US-based)

179

Figure 201. Assessment of Model Convergence for Price (US-based)

Figure 202. Assessment of Model Convergence for GT (US-based)

Figure 203. Assessment of Model Convergence for GPI (US-based)

180

Figure 204. Assessment of Model Convergence for CCI (US-based)

3.5.3 Sample Split Bayesian Analysis Results on Market Structure

I then turn my attention to market structure. Netzer et al. (2012) applied the text-mining
approach on user-generated content in the U.S automobile industry to better understand market
structure and they identified three different sub market structures (see Figure 205 for their
identified market structure). I therefore, categorize my whole sample into these three sub market
structures19 to examine how these market structures may vary dynamics of online WOM in the
current study. Tables 60 to 62 show descriptive statistics for these three groups, respectively.

19

Group 1: Acura, Audi, BMW, Cadillac, Infiniti, Jaguar, Land Rover, Lexus, Mercedes-Benz, Porsche, Saab, and
Volvo; Group 2: Buick, Chevrolet, Chrysler, Dodge, FIAT, Ford, Jeep, and Lincoln; Group 3: Honda, Hyundai,
KIA, Mazda, Mitsubishi, Nissan, Scion, Subaru, Toyota, and Volkswagen. Please note that some brands (e.g., Land
Rover) covered in the current did not include in Netzer et al.’s study (2012). I assigned those brands into the
corresponding groups based on their existing characteristics. For example, Land Rover could be considered into the
group 1 based on the price factor. Scion could be considered into the group 3 based on the origin of the brand.
181

Figure 205. Netzer et al.’s (2012) Market Structures

Table 60. Descriptive Statistics for Group 1

Variable
No. of obs.
Sale
679
F-Post (a)
679
C-F-Post (a)
679
F-Like (a)
679
C-F-Like (a)
679
F-Comment (a)
679
C-F-Comment (a) 679
F-Share (a)
679
C-F-Share (a)
679
U-Post (a)
679
C-U-Post (a)
679
U-Like (a)
679
C-U-Like (a)
679
U-Comment (a)
679
C-U-Comment (a) 679
U-Share (a)
679
C-U-Share (a)
679
TD-Post (c)
679
C-TD-Post (c)
679
TMS
679
Price
679
GT
679
GPI
679
CCI
679

Mean
11208.15
30.67
926.96
133154.7
2,243,497
2347.04
65674.85
9337.56
173316.9
303.73
12698.2
9959.97
305400.4
474.42
25904.62
65.58
2463.07
148.87
7093.65
16830862.15
51759.78
63.44
3.40
66.13

Std. dev.
8387.419
23.19
408.36
211997.3
2,086,161
2573.44
38301.75
17383.42
172271.3
436.88
4436.63
32970.16
384998.7
961.21
16147.87
388.85
3958.59
157.30
4971.08
12915655.9
12193.72
16.04
0.4
12.89

Min
97
0
64
0
4,230
0
861
0
83
0
853
0
561
0
670
0
0
5
1,983
65,200
33,175
26
2.31
40.9

Table 61. Descriptive Statistics for Group 2

Variable
No. of obs.
Sale
489
F-Post (a)
489
C-F-Post (a)
489
F-Like (a)
489
C-F-Like (a)
489
F-Comment (a)
489
C-F-Comment (a) 489
F-Share (a)
489
C-F-Share (a)
489
U-Post (a)
489
C-U-Post (a)
489
U-Like (a)
489
C-U-Like (a)
489
U-Comment (a)
489
C-U-Comment (a) 489
U-Share (a)
489
C-U-Share (a)
489

Mean
60169.34
35.28
910.04
53748.09
2,336,259
2648.58
65037.49
5613.42
178403.5
685.83
12076.33
16323.49
299962.7
1703.87
24395.63
114.95
2422.84

Std. dev.
64971.91
51.74
410.38
75189.46
2,169,217
2817.23
39476.79
9153.28
175696.5
691.16
4470.32
34578.98
387392.4
2539.45
16118.24
408.40
3915.44

182

Min
500
0
50
0
3,093
0
420
0
27
0
857
0
613
0
690
0
0

Max
37,399
436
2,579
1,100,000
5,900,000
15,943
152,429
115,372
577,231
9,624
22,520
560,219
2,500,000
11,461
109,163
5,283
21,904
1,595
35,072
75,000,000
90,775
100
3.98
94.1
Max
244,501
1,042
2,584
490,822
6,000,000
25,734
152,256
62,457
578,089
4,918
22,459
252,535
2,500,000
28,081
108,643
6,362
21,905

Table 61 (cont’d)
TD-Post (c)
C-TD-Post (c)
TMS
Price
GT
GPI
CCI

489
489
489
489
489
489
489

290.91
6811.77
48687195.71
27558.14
52.12
3.39
66.14

393.26
4761.93
44841280.14
7208.62
21.13
0.4
12.93

10
1,885
238,700
16,000
7
2.31
40.9

Table 62. Descriptive Statistics for Group 3

Variable
No. of obs.
Sale
621
F-Post (a)
621
C-F-Post (a)
621
F-Like (a)
621
C-F-Like (a)
621
F-Comment (a)
621
C-F-Comment (a) 621
F-Share (a)
621
C-F-Share (a)
621
U-Post (a)
621
C-U-Post (a)
621
U-Like (a)
621
C-U-Like (a)
621
U-Comment (a)
621
C-U-Comment (a) 621
U-Share (a)
621
C-U-Share (a)
621
TD-Post (c)
621
C-TD-Post (c)
621
TMS
621
Price
621
GT
621
GPI
621
CCI
621

Mean
50792.28
34.33
911
48086.88
2,309,850
2126.13
65320.97
3497.47
177734.7
434.42
12442.41
7489.32
304945.3
771.39
25378.18
87.91
2413.28
323.17
6967.49
42166022.54
23564.44
47.62
3.39
66.06

Std. dev.
44383.67
27.37
405
76472.79
2,161,101
3131.96
38569.89
5775.98
177469.3
426.21
4356.02
24087.75
390721.6
1720.94
16077.35
683.83
3873.41
395.96
4961.65
33500246.34
4149.23
18.31
0.4
12.87

Min
2,923
1
48
1
4,226
7
826
0
84
0
647
0
254
0
480
0
0
13
1,862
1,500,000
14192.5
10
2.31
40.9

3,348
34,817
190,000,000
45,142
100
3.98
94.1
Max
210,134
344
2,605
722,146
6,000,000
26,648
153,749
41,089
580,559
4,625
22,539
463,332
2,500,000
36,118
108,721
16,400
21,905
3,771
35,106
200,000,000
33783.9
100
3.98
94.1

Tables 63 to 65 show my sample split Bayesian estimation results at the post level (i.e.,
Facebook post and test drive post) for group 1 of market structure, group 2 of market structure,
and group 3 of market structure, respectively. In these models, I ran the MCMC chain for
999,961 iterations and, I discarded the first 200,000 iterations to ensure convergence. Because of
the high auto-correlation of parameter posterior draws, I thinned the sampling chain and kept one
from every fortieth iterations for the remaining iterations. The assessment of model convergence
(Figures 206 to 238) also suggested the model specification converged for each relationship.
The results in the post level demonstrate that customers from these three different groups
show dramatically different patterns in terms of how online WOM across two different stages
183

would influence their purchase decision. First, the results from group 1 (Table 63) are consistent
with main results (see Table 41) except for the effect of the focal brand’s posts on offline car
sales of the focal firm. The comparison of group 2 and group 3 also indicates interesting patterns.
For example, for group 3, at the stage of awareness, both posts from focal brand and its
competitors positively influence offline car sales of the focal brand, supporting both H1 and H2.
However, for group 3, I only observe positive spillover effects (C-F-Post (a)) at the stage of
awareness, supporting H2 only. Group 2 also shows the opposite effect of posts by the focal
brand’s users (U-Post (a)), rejecting H1. Regarding the effect of online WOM at the stage of
consideration, H3 and H4 are supported for group 1 only. I cannot find any effect of online
WOM at the stage of consideration for group 2 and group 3.
Table 63. Bayesian Estimation Results for Posts (Group 1)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.025 (0.017)
0.157 (0.035)
0.002 (0.009)
-0.065 (0.028)
0.047 (0.019)
-0.06 (0.026)
0.113 (0.014)
-0.056 (0.11)
0.124 (0.083)
0.045 (0.124)
0.287 (0.069)

Sales A, t
95% Credible Level
(-0.008, 0.06)
(0.07, 0.226)
(-0.017, 0.02)
(-0.12, -0.009)
(0.008, 0.086)
(-0.118, -0.016)
(0.084, 0.14)
(-0.271, 0.155)
(-0.038, 0.288)
(-0.201, 0.286)
(0.149, 0.424)

Table 64. Bayesian Estimation Results for Posts (Group 2)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.003 (0.015)
(-0.026, 0.033)
0.168 (0.031)
(0.109, 0.229)
-0.052 (0.013)
(-0.079, -0.025)
0.041 (0.028)
(-0.015, 0.097)
0.003 (0.021)
(-0.038, 0.045)
-0.047 (0.03)
(-0.107, 0.012)
0.166 (0.02)
(0.125, 0.205)

184

Table 64 (cont’d)
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

0.132 (0.114)
-0.021 (0.06)
0.25 (0.121)
0.26 (0.068)

(-0.09, 0.361)
(-0.143, 0.098)
(0.02, 0.497)
(0.125, 0.391)

Table 65. Bayesian Estimation Results for Posts (Group 3)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.032 (0.011)
0.162 (0.027)
-0.002 (0.01)
-0.068 (0.023)
0.011 (0.019)
-0.031 (0.022)
0.093 (0.021)
0.006 (0.097)
0.179 (0.043)
0.28 (0.11)
0.188 (0.057)

Sales A, t
95% Credible Level
(0.009, 0.054)
(0.108, 0.216)
(-0.022, 0.019)
(-0.114, -0.023)
(-0.027, 0.048)
(-0.075, 0.013)
(0.051, 0.135)
(-0.184, 0.198)
(0.095, 0.266)
(0.081, 0.479)
(0.073, 0.301)

Figure 206. Assessment of Model Convergence for F-Post (a) (Group 1)

185

Figure 207. Assessment of Model Convergence for C-F-Post (a)
(Group 1)

Figure 208. Assessment of Model Convergence for U-Post (a) (Group 1)

Figure 209. Assessment of Model Convergence for C-U-Post (a)
(Group 1)

186

Figure 210. Assessment of Model Convergence for TD-Post (c) (Group 1)

Figure 211. Assessment of Model Convergence for C-TD-Post (c) (Group 1)

Figure 212. Assessment of Model Convergence for TMS (Group 1)

187

Figure 213. Assessment of Model Convergence for Price (Group 1)

Figure 214. Assessment of Model Convergence for GT (Group 1)

Figure 215. Assessment of Model Convergence for GPI (Group 1)

188

Figure 216. Assessment of Model Convergence for CCI (Group 1)

Figure 217. Assessment of Model Convergence for F-Post (a) (Group 2)

Figure 218. Assessment of Model Convergence for C-F-Post (a) (Group 2)

189

Figure 219. Assessment of Model Convergence for U-Post (a) (Group 2)

Figure 220. Assessment of Model Convergence for C-U-Post (a) (Group 2)

Figure 221. Assessment of Model Convergence for TD-Post (c) (Group 2)

190

Figure 222. Assessment of Model Convergence for C-TD-Post (c) (Group 2)

Figure 223. Assessment of Model Convergence for TMS (Group 2)

Figure 224. Assessment of Model Convergence for Price (Group 2)

191

Figure 225. Assessment of Model Convergence for GT (Group 2)

Figure 226. Assessment of Model Convergence for GPI (Group 2)

Figure 227. Assessment of Model Convergence for CCI (Group 2)

192

Figure 228. Assessment of Model Convergence for F-Post (a) (Group 3)

Figure 229. Assessment of Model Convergence for C-F-Post (a) (Group 3)

Figure 230. Assessment of Model Convergence for U-Post (a) (Group 3)

193

Figure 231. Assessment of Model Convergence for C-U-Post (a) (Group
3)

Figure 232. Assessment of Model Convergence for TD-Post (c) (Group 3)

Figure 233. Assessment of Model Convergence for C-TD-Post (c)
(Group 3)

194

Figure 234. Assessment of Model Convergence for TMS (Group 3)

Figure 235. Assessment of Model Convergence for Price (Group 3)

Figure 236. Assessment of Model Convergence for GT (Group 3)

195

Figure 237. Assessment of Model Convergence for GPI (Group 3)

Figure 238. Assessment of Model Convergence for CCI (Group 3)

Tables 66 to 68 show my sample split Bayesian estimation results at the like level (i.e.,
“Like” associated with posts at Facebook and test drive post) for group 1 of market structure,
group 2 of market structure, and group 3 of market structure, respectively. In these models, I also
ran the MCMC chain for 999,961 iterations and, I discarded the first 200,000 iterations to ensure
convergence. Because of the high auto-correlation of parameter posterior draws, I thinned the

196

sampling chain and kept one from every fortieth iterations for the remaining iterations. The
assessment of model convergence (see Figures 239 to 271) suggested that the model
specification converged for each relationship.
First, this sub group analysis in the like level, consistent with main results (see Table 42),
suggests that the volume of like associated with the focal brand’s posts (F-Like (a)) is not
effective in influencing offline car sales of the focal brand across three different groups, rejecting
H1. On the other hand, the volume of like associated with competitors’ user posts (C-F-Like (a))
has positive spillover effects on offline car sales of the focal brand across there different groups
with the strongest the magnitude of the coefficient for group 1, supporting H2. Interestingly, I
also find that for group 1 the volume of like related to the focal brand’s user posts (U-Like (a))
positively influences offline car sales of the focal brand, which is the relationship that I cannot
find in previous analysis, supporting H1. Furthermore, for online WOM at the stage of
consideration, I only find H3 supported for group 1, implying again that market structure does
influence how customers leverage online WOM to make their purchase decisions.
Table 66. Bayesian Estimation Results for Likes (Group 1)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.008 (0.008)
0.099 (0.018)
0.002 (0.005)
-0.023 (0.016)
0.038 (0.018)
-0.044 (0.023)
0.107 (0.013)
-0.028 (0.105)
0.215 (0.082)
-0.165 (0.11)
0.113 (0.07)

197

Sales A, t
95% Credible Level
(-0.02, 0.008)
(0.064, 0.134)
(0.01, 0.024)
(-0.055, 0.008)
(0.002, 0.075)
(-0.089, 0.001)
(0.08, 0.133)
(-0.233, 0.177)
(0.056, 0.377)
(-0.38, 0.053)
(-0.023, 0.252)

Table 67. Bayesian Estimation Results for Likes (Group 2)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.014 (0.009)
0.05 (0.017)
-0.028 (0.007)
0.078 (0.017)
0.012 (0.02)
-0.029 (0.027)
0.16 (0.019)
0.062 (0.11)
-0.018 (0.059)
0.065 (0.11)
0.06 (0.067)

Sales A, t
95% Credible Level
(-0.031, 0.004)
(0.017, 0.083)
(-0.042, -0.014)
(0.046, 0.11)
(-0.03, 0.052)
(-0.083, 0.024)
(0.12, 0.19)
(-0.148, 0.271)
(-0.135, 0.096)
(-0.145, 0.273)
(-0.075, 0.196)

Table 68. Bayesian Estimation Results for Likes (Group 3)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.012 (0.009)
(-0.005, 0.028)
0.045 (0.016)
(0.015, 0.077)
-0.003 (0.005)
(-0.013, 0.008)
0.02 (0.014)
(-0.004, 0.05)
0.011 (0.018)
(-0.025, 0.048)
-0.021 (0.021)
(-0.063, 0.02)
0.085 (0.021)
(0.044, 0.127)
-0.061 (0.094)
(-0.25, 0.124)
0.21 (0.043)
(0.123, 0.296)
0.086 (0.09)
(-0.092, 0.263)
0.082 (0.058)
(-0.034, 0.196)

Figure 239. Assessment of Model Convergence for F-Like (a) (Group 1)

198

Figure 240. Assessment of Model Convergence for C-F-Like (a) (Group 1)

Figure 241. Assessment of Model Convergence for U-Like (a) (Group 1)

Figure 242. Assessment of Model Convergence for C-U-Like (a) (Group 1)

199

Figure 243. Assessment of Model Convergence for TD-Post (c) (Group 1)

Figure 244. Assessment of Model Convergence for C-TD-Post (c) (Group 1)

Figure 245. Assessment of Model Convergence for TMS (Group 1)

200

Figure 246. Assessment of Model Convergence for Price (Group 1)

Figure 247. Assessment of Model Convergence for GT (Group 1)

Figure 248. Assessment of Model Convergence for GPI (Group 1)

201

Figure 249. Assessment of Model Convergence for CCI (Group 1)

Figure 250. Assessment of Model Convergence for F-Like (a) (Group 2)

Figure 251. Assessment of Model Convergence for C-F-Like (a) (Group 2)

202

Figure 252. Assessment of Model Convergence for U-Like (a) (Group 2)

Figure 253. Assessment of Model Convergence for C-U-Like (a) (Group 2)

Figure 254. Assessment of Model Convergence for TD-Post (c) (Group 2)

203

Figure 255. Assessment of Model Convergence for C-TD-Post (c) (Group 2)

Figure 256. Assessment of Model Convergence for TMS (Group 2)

Figure 257. Assessment of Model Convergence for Price (Group 2)

204

Figure 258. Assessment of Model Convergence for GT (Group 2)

Figure 259. Assessment of Model Convergence for GPI (Group 2)

Figure 260. Assessment of Model Convergence for CCI (Group 2)

205

Figure 261. Assessment of Model Convergence for F-Like (a) (Group 3)

Figure 262. Assessment of Model Convergence for C-F-Like (a) (Group 3)

Figure 263. Assessment of Model Convergence for U-Like (a) (Group 3)

206

Figure 264. Assessment of Model Convergence for C-U-Like (a) (Group
3)

Figure 265. Assessment of Model Convergence for TD-Post (c) (Group
3)

Figure 266. Assessment of Model Convergence for C-TD-Post (c) (Group 3)

207

Figure 267. Assessment of Model Convergence for TMS (Group 3)

Figure 268. Assessment of Model Convergence for Price (Group 3)

Figure 269. Assessment of Model Convergence for GT (Group 3)

208

Figure 270. Assessment of Model Convergence for GPI (Group 3)

Figure 271. Assessment of Model Convergence for CCI (Group 3)

Tables 69 to 71 show my sample split Bayesian estimation results at the comment level
(i.e., “Comment” associated with posts at Facebook and test drive post) for group 1 of market
structure, group 2 of market structure, and group 3 of market structure, respectively. I ran the
MCMC chain for 999,961 iterations and, I discarded the first 200,000 iterations to ensure
convergence. Because of the high auto-correlation of parameter posterior draws, I thinned the

209

sampling chain and kept one from every fortieth iterations for the remaining iterations. The
assessment of model convergence (see Figures 272 to 304) suggested that the model
specification converged for each relationship.
This split sample analysis in the comment level shows many interesting patterns that I
cannot obverse in my earlier analysis. For example, for group 3, I find that the volume of
comment associated with the focal brand’s posts (F-Comment (a)) and its competitors’ posts (CF-Comment (a)) have the positive impact on offline car sales of the focal brand, supporting both
H1 and H2. For group 2, I only find that C-F-Comment (a) has positive spillover effects on
offline car sales of the focal brand, supporting H2. However, for group 1, I only observe negative
spillover effects from the volume of comments received from competitors’ posts (C-F-Comment
(a)), thereby rejecting H2. Regarding the effects associated with user posts at the stage of
awareness (U-Comment (a) and C-U-Comment (a)), I only find H2 supported for group 2.
Namely, C-U-Comment (a) has positive spillover effects on offline car sales of the focal brand.
Finally, for online WOM at the stage of consideration, I find that C-TD-Post (c) negatively
influences offline car sales of the focal brand for group 1 only, supporting my H4.

210

Table 69. Bayesian Estimation Results for Comments (Group 1)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.028 (0.019)
-0.063 (0.023)
0.014 (0.008)
-0.03 (0.02)
0.028 (0.019)
-0.062 (0.024)
0.11 (0.014)
-0.052 (0.11)
0.177 (0.081)
-0.061 (0.11)
0.22 (0.161)

Sales A, t
95% Credible Level
(-0.003, 0.026)
(0.055, 0.124)
(-0.014, 0.014)
(-0.044, 0.022)
(-0.009, 0.065)
(-0.109, -0.016)
(0.008, 0.135)
(-0.265, 0.16)
(0.015, 0.33)
(-0.25, 0.137)
(0.084, 0.35)

Table 70. Bayesian Estimation Results for Comments (Group 2)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.011 (0.009)
0.081 (0.018)
-0.047 (0.008)
0.097 (0.019)
0.008 (0.02)
-0.037 (0.027)
0.157 (0.019)
0.096 (0.11)
-0.003 (0.058)
0.35 (0.097)
0.217 (0.063)

Sales A, t
95% Credible Level
(-0.029, 0.007)
(0.045, 0.12)
(-0.063, -0.031)
(0.059, 0.135)
(-0.031, 0.048)
(-0.092, 0.017)
(0.119, 0.195)
(-0.11, 0.31)
(-0.12, 0.11)
(0.157, 0.538)
(0.092, 0.342)

Table 71. Bayesian Estimation Results for Comments (Group 3)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.028 (0.009)
(0.01, 0.045)
0.066 (0.018)
(0.031, 0.1)
0.004 (0.007)
(-0.01, 0.018)
-0.003 (0.016)
(-0.035, 0.03)
0.014 (0.019)
(-0.022, 0.051)
-0.04 (0.022)
(-0.084, 0.002)
0.092 (0.022)
(0.05, 0.135)
-0.038 (0.096)
(-0.23, 0.149)
0.176 (0.045)
(0.085, 0.267)
0.286 (0.083)
(0.124, 0.45)
0.19 (0.056)
(0.085, 0.3)

211

Figure 272. Assessment of Model Convergence for F-Comment (a) (Group 1)

Figure 273. Assessment of Model Convergence for C-F-Comment (a) (Group 1)

Figure 274. Assessment of Model Convergence for U-Comment (a) (Group 1)

212

Figure 275. Assessment of Model Convergence for C-U-Comment (a) (Group 1)

Figure 276. Assessment of Model Convergence for TD-Post (c) (Group 1)

Figure 277. Assessment of Model Convergence for C-TD-Post (c) (Group 1)

213

Figure 278. Assessment of Model Convergence for TMS (Group 1)

Figure 279. Assessment of Model Convergence for Price (Group 1)

Figure 280. Assessment of Model Convergence for GT (Group 1)

214

Figure 281. Assessment of Model Convergence for GPI (Group 1)

Figure 282. Assessment of Model Convergence for CCI (Group 1)

Figure 283. Assessment of Model Convergence for F-Comment (a) (Group 2)

215

Figure 284. Assessment of Model Convergence for C-F-Comment (a) (Group 2)

Figure 285. Assessment of Model Convergence for U-Comment (a) (Group 2)

Figure 286. Assessment of Model Convergence for C-U-Comment (a) (Group 2)

216

Figure 287. Assessment of Model Convergence for TD-Post (c) (Group 2)

Figure 288. Assessment of Model Convergence for C-TD-Post (c) (Group 2)

Figure 289. Assessment of Model Convergence for TMS (Group 2)

217

Figure 290. Assessment of Model Convergence for Price (Group 2)

Figure 291. Assessment of Model Convergence for GT (Group 2)

Figure 292. Assessment of Model Convergence for GPI (Group 2)

218

Figure 293. Assessment of Model Convergence for CCI (Group 2)

Figure 294. Assessment of Model Convergence for F-Comment (a) (Group 3)

Figure 295. Assessment of Model Convergence for C-F-Comment (a) (Group 3)

219

Figure 296. Assessment of Model Convergence for U-Comment (a) (Group 3)

Figure 297. Assessment of Model Convergence for C-U-Comment (a) (Group 3)

Figure 298. Assessment of Model Convergence for TD-Post (c) (Group 3)

220

Figure 299. Assessment of Model Convergence for C-TD-Post (c) (Group 3)

Figure 300. Assessment of Model Convergence for TMS (Group 3)

Figure 301. Assessment of Model Convergence for Price (Group 3)

221

Figure 302. Assessment of Model Convergence for GT (Group 3)

Figure 303. Assessment of Model Convergence for GPI (Group 3)

Figure 304. Assessment of Model Convergence for CCI (Group 3)

222

Tables 72 to 74 show my sample split Bayesian estimation results at the share level (i.e.,
“Share” associated with posts at Facebook and test drive post) for group 1 of market structure,
group 2 of market structure, and group 3 of market structure, respectively. In these models, I ran
the MCMC chain for 999,961 iterations and, I discarded the first 200,000 iterations to ensure
convergence. Because of the high auto-correlation of parameter posterior draws, I thinned the
sampling chain and kept one from every fortieth iterations for the remaining iterations. The
assessment of model convergence (see Figures 305 to 337) suggested that the model
specification converged for each relationship.
Consistent with main results (see Table 44), the results suggest that at the stage of
awareness F-Share (a) is not very effective in influencing offline car sales of the focal brand
across three groups, rejecting H1. On the other hand, C-F-Share (a) shows positive spillover
effects on offline car sales of the focal brand across three different groups, supporting H2.
Regarding effects associated with user posts at the stage of awareness, I only find that U-Share
(a) has the positive impact on offline car sales, supporting my H1. Finally, for online WOM at
the stage of consideration, I find that TD-Post (c) has the positive impact on offline car sales of
the focal brand for group 1 only, in support of H3. However, no online WOM spillover effects
observed at the stage of consideration in this set of analysis.

223

Table 72. Bayesian Estimation Results for Shares (Group 1)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.004 (0.006)
0.053 (0.012)
-0.0001 (0.006)
-0.016 (0.013)
0.056 (0.019)
-0.042 (0.027)
0.11 (0.014)
-0.029 (0.106)
0.2 (0.082)
0.169 (0.129)
0.274 (0.067)

Sales A, t
95% Credible Level
(-0.016, 0.007)
(0.029, 0.076)
(-0.012, 0.012)
(-0.04, 0.008)
(0.018, 0.094)
(-0.095, 0.011)
(0.083, 0.138)
(-0.238, 0.179)
(0.037, 0.363)
(-0.086, 0.424)
(0.144, 0.404)

Table 73. Bayesian Estimation Results for Shares (Group 2)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.006 (0.006)
0.057 (0.011)
0.016 (0.006)
-0.025 (0.012)
0.008 (0.02)
-0.014 (0.029)
0.17 (0.019)
-0.12 (0.11)
-0.119 (0.056)
0.381 (0.123)
0.19 (0.064)

Sales A, t
95% Credible Level
(-0.017, 0.005)
(0.035, 0.078)
(0.004, 0.029)
(-0.048, -0.002)
(-0.033, 0.048)
(-0.07, 0.043)
(0.132, 0.209)
(-0.1, 0.342)
(-0.228, -0.008)
(0.138, 0.621)
(0.063, 0.316)

Table 74. Bayesian Estimation Results for Shares (Group 3)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.008 (0.005)
0.032 (0.01)
0.006 (0.005)
-0.009 (0.009)
0.014 (0.018)
0.002 (0.023)
0.079 (0.021)
-0.026 (0.093)
0.18 (0.042)
0.303 (0.106)
0.211 (0.054)

224

Sales A, t
95% Credible Level
(-0.0007, 0.018)
(0.012, 0.052)
(-0.004, 0.015)
(-0.029, 0.009)
(-0.022, 0.05)
(-0.044, 0.048)
(0.037, 0.12)
(-0.208, 0.158)
(0.097, 0.262)
(0.097, 0.51)
(0.104, 0.32)

Figure 305. Assessment of Model Convergence for F-Share (a) (Group 1)

Figure 306. Assessment of Model Convergence for C-F-Share (a) (Group 1)

Figure 307. Assessment of Model Convergence for U-Share (a) (Group 1)

225

Figure 308. Assessment of Model Convergence for C-U-Share (a) (Group 1)

Figure 309. Assessment of Model Convergence for TD-Post (c) (Group 1)

Figure 310. Assessment of Model Convergence for C-TD-Post (c) (Group 1)

226

Figure 311. Assessment of Model Convergence for TMS (Group 1)

Figure 312. Assessment of Model Convergence for Price (Group 1)

Figure 313. Assessment of Model Convergence for GT (Group 1)

227

Figure 314. Assessment of Model Convergence for GPI (Group 1)

Figure 315. Assessment of Model Convergence for CCI (Group 1)

Figure 316. Assessment of Model Convergence for F-Share (a) (Group 2)

228

Figure 317. Assessment of Model Convergence for C-F-Share (a)
(Group 2)

Figure 318. Assessment of Model Convergence for U-Share (a) (Group 2)

Figure 319. Assessment of Model Convergence for C-U-Share (a) (Group 2)

229

Figure 320. Assessment of Model Convergence for TD-Post (c) (Group 2)

Figure 321. Assessment of Model Convergence for C-TD-Post (c) (Group 2)

Figure 322. Assessment of Model Convergence for TMS (Group 2)

230

Figure 323. Assessment of Model Convergence for Price (Group 2)

Figure 324. Assessment of Model Convergence for GT (Group 2)

Figure 325. Assessment of Model Convergence for GPI (Group 2)

231

Figure 326. Assessment of Model Convergence for CCI (Group 2)

Figure 327. Assessment of Model Convergence for F-Share (a) (Group 3)

Figure 328. Assessment of Model Convergence for C-F-Share (a) (Group 3)

232

Figure 329. Assessment of Model Convergence for U-Share (a) (Group 3)

Figure 330. Assessment of Model Convergence for C-U-Share (a) (Group 3)

Figure 331. Assessment of Model Convergence for TD-Post (c) (Group 3)

233

Figure 332. Assessment of Model Convergence for C-TD-Post (c) (Group 3)

Figure 333. Assessment of Model Convergence for TMS (Group 3)

Figure 334. Assessment of Model Convergence for Price (Group 3)

234

Figure 335. Assessment of Model Convergence for GT (Group 3)

Figure 336. Assessment of Model Convergence for GPI (Group 3)

Figure 337. Assessment of Model Convergence for CCI (Group 3)

235

3.5.4 Sample Split Bayesian Analysis Results on Luxury and Non-Luxury Brands

Previous research on WOM has provided evidence that consumers’ motivation to engage
in online WOM differs across brands, and that consumers may be particularly inclined to
converse about high-regarded or high-quality brands, about luxury goods, or about brands with a
high degree of differentiation (Lovett et al., 2013). Such differences in online WOM activity may
naturally affect the informativeness of social media data for perdition (Geva et al., 2017). I,
therefore, examine whether and how the luxury versus non-luxury category20 may vary dynamics
of online WOM in the current study. Tables 75 and 76 show descriptive statistics for luxury and
non-luxury group, respectively.
Table 75. Descriptive Statistics for Luxury brands

Variable
No. of obs.
Sale
800
F-Post (a)
800
C-F-Post (a)
800
F-Like (a)
800
C-F-Like (a)
800
F-Comment (a)
800
C-F-Comment (a) 800
F-Share (a)
800
C-F-Share (a)
800
U-Post (a)
800
C-U-Post (a)
800
U-Like (a)
800
C-U-Like (a)
800
U-Comment (a)
800
C-U-Comment (a) 800
U-Share (a)
800
C-U-Share (a)
800
TD-Post (c)
800
C-TD-Post (c)
800
TMS
800
Price
800
GT
800
GPI
800
CCI
800

Mean
11221.53
30.50
927.75
115851.5
2,266,860
2146.45
66019.67
8182.67
174978.8
281.24
12722.51
8745.22
307365.1
442.67
25968.61
56.15
2478.32
139.54
7119.08
17976444.25
49494.1
61.51
3.40
66.17

Std. dev.
7948.68
24.7
406.43
199947.3
2,101,807
2445.04
38378.31
16276.19
173358.8
409.66
4424.98
30559.04
387251.3
895.59
16172.53
358.91
3971.74
152.18
4998.62
12746535.05
12755.88
16.05
0.39
12.88

20

Min
97
0
64
0
4,230
0
861
0
83
0
853
0
561
0
670
0
0
5
1,983
65,200
28,800
26
2.31
40.9

Max
37,399
436
2,584
1,100,000
6,000,000
15,943
152,429
115,372
578,089
9,624
22,520
560,219
2,500,000
11,461
109,163
5,283
21,905
1,595
35,072
75,000,000
90,775
100
3.98
94.1

Luxury: Acura, Audi, BMW, Buick, Cadillac, Infiniti, Jaguar, Land Rover, Lexus, Lincoln, Mercedes-Benz,
Porsche, Saab, and Volvo; Non-Luxury: Chevrolet, Chrysler, Dodge, FIAT, Ford, Honda, Hyundai, Jeep, KIA,
Mazda, Mitsubishi, Nissan, Scion, Subaru, Toyota, and Volkswagen
236

Table 76. Descriptive Statistics for Non-Luxury brands

Variable
No. of obs.
Sale
989
F-Post (a)
989
C-F-Post (a)
989
F-Like (a)
989
C-F-Like (a)
989
F-Comment (a)
989
C-F-Comment (a) 989
F-Share (a)
989
C-F-Share (a)
989
U-Post (a)
989
C-U-Post (a)
989
U-Like (a)
989
C-U-Like (a)
989
U-Comment (a)
989
C-U-Comment (a) 989
U-Share (a)
989
C-U-Share (a)
989
TD-Post (c)
989
C-TD-Post (c)
989
TMS
989
Price
989
GT
989
GPI
989
CCI
989

Mean
60260.79
35.39
907.93
54474.84
2,312,127
2519.67
64858.59
4763.35
177261.6
592.91
12210.44
12537.61
300836.7
1294.46
24776.21
111.64
2399.58
336.09
6854.49
47563335.39
23922.2
49.47
3.39
66.06

Std. dev.
55507.96
40.8
408.54
78756.19
2,161,215
3130.35
38966.45
7856.64
176295.6
586.58
4409.6
31007.68
387,759
2271.28
16061.45
612.16
3870.65
408.59
4836.52
40493361.94
4076.61
20.53
0.4
12.89

Min
500
0
48
0
3,093
0
420
0
27
0
647
0
254
0
480
0
0
13
1,862
238,700
14192.5
7
2.31
40.9

Max
244,501
1,042
2,605
722,146
6,000,000
26,648
153,749
62,457
580,559
4,918
22,539
463,332
2,500,000
36,118
108,721
16,400
21,905
3,771
35,106
200,000,000
33783.9
100
3.98
94.1

Tables 77 to 78 show my sample split Bayesian estimation results at the post level (i.e.,
Facebook post and test drive post) for luxury and non-luxury group, respectively. In these
models, I ran the MCMC chain for 999,961 iterations and, I discarded the first 200,000 iterations
to ensure convergence. Because of the high auto-correlation of parameter posterior draws, I
thinned the sampling chain and kept one from every fortieth iterations for the remaining
iterations. The assessment of model convergence (Figures 338 to 359) also suggested the model
specification converged for each relationship.
First, at the stage of awareness, I observe that for the luxury group, F-Post (a) is not
effective in influencing office line car sales of the focal brand, rejecting H1, while C-F-Post (a)
has the positive impact on offline car sales of the focal brand, supporting H2. On the other hand,
both H1 and H2 are supported for posts by the focal firm and its competitors. I also observe
237

another interesting patterns for posts associated with the focal brand’s user posts and its
competitors’ user posts (U-Post (a)) and (C-U-Post (a)) for the non-luxury group. Although I find
the statistically significant effects for these two mechanisms, they show the opposite effects as
hypothesized. Namely, both U-Post (a) and C-U-Post (a) have the negative impact on offline car
sales of the focal brand, thereby rejecting both H1 and H2. Finally, for online WOM at the stage
of consideration, I only find H3 and H4 supported for the luxury-group. It provides further
evidence that customers from these two different groups appreciate test drive experience posts
dramatically different to help them make purchase decisions.
Table 77. Bayesian Estimation Results for Posts (Luxury)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.007 (0.009)
0.165 (0.03)
0.007 (0.009)
-0.04 (0.026)
0.035 (0.017)
-0.061 (0.023)
0.121 (0.013)
-0.075 (0.1)
-0.033 (0.066)
-0.046 (0.11)
0.285 (0.153)

Sales A, t
95% Credible Level
(-0.024, 0.038)
(0.11, 0.225)
(-0.01, 0.025)
(-0.091, 0.011)
(0.002, 0.068)
(-0.1, -0.016)
(0.094, 0.147)
(-0.28, 0.127)
(-0.161, 0.097)
(-0.265, 0.173)
(0.164, 0.41)

Table 78. Bayesian Estimation Results for Posts (Non-Luxury)
Parameters
F-Post (a) A, t-1
C-F-Post (a) J, t-1
U-Post (a) A, t-1
C-U-Post (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.028 (0.009)
0.167 (0.022)
-0.021 (0.008)
-0.039 (0.019)
-0.001 (0.015)
-0.034 (0.019)
0.12 (0.02)
0.032 (0.074)
0.19 (0.037)
0.303 (0.082)
0.194 (0.047)

238

Sales A, t
95% Credible Level
(0.009, 0.047)
(0.124, 0.21)
(-0.038, -0.004)
(-0.076, -0.002)
(-0.032, 0.03)
(-0.073, 0.004)
(0.091, 0.15)
(-0.11, 0.176)
(0.12, 0.264)
(0.14, 0.462)
(0.1, 0.286)

Figure 338. Assessment of Model Convergence for F-Post (a) (Luxury)

Figure 339. Assessment of Model Convergence for C-F-Post (a) (Luxury)

Figure 340. Assessment of Model Convergence for U-Post (a) (Luxury)

239

Figure 341. Assessment of Model Convergence for C-U-Post (a)
(Luxury)

Figure 342. Assessment of Model Convergence for TD-Post (c) (Luxury)

Figure 343. Assessment of Model Convergence for C-TD-Post (c) (Luxury)

240

Figure 344. Assessment of Model Convergence for TMS (Luxury)

Figure 345. Assessment of Model Convergence for Price (Luxury)

Figure 346. Assessment of Model Convergence for GT (Luxury)

241

Figure 347. Assessment of Model Convergence for GPI (Luxury)

Figure 348. Assessment of Model Convergence for CCI (Luxury)

Figure 349. Assessment of Model Convergence for F-Post (a) (Non-Luxury)

242

Figure 350. Assessment of Model Convergence for C-F-Post (a) (Non-Luxury)

Figure 351. Assessment of Model Convergence for U-Post (a) (Non-Luxury)

Figure 352. Assessment of Model Convergence for C-U-Post (a) (Non-Luxury)

243

Figure 353. Assessment of Model Convergence for TD-Post (c) (Non-Luxury)

Figure 354. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury)

Figure 355. Assessment of Model Convergence for TMS (Non-Luxury)

244

Figure 356. Assessment of Model Convergence for Price (Non-Luxury)

Figure 357. Assessment of Model Convergence for GT (Non-Luxury)

Figure 358. Assessment of Model Convergence for GPI (Non-Luxury)

245

Figure 359. Assessment of Model Convergence for CCI (Non-Luxury)

Tables 79 to 80 show my sample split Bayesian estimation results at the like level (i.e.,
“Like” associated with posts at Facebook and test drive post) for luxury and non-luxury groups,
respectively. In these models, I also ran the MCMC chain for 999,961 iterations and, I discarded
the first 200,000 iterations to ensure convergence. Because of the high auto-correlation of
parameter posterior draws, I thinned the sampling chain and kept one from every fortieth
iterations for the remaining iterations. The assessment of model convergence (see Figures 360 to
381) suggested that the model specification converged for each relationship.
In this set of analysis, I find that C-F-Like (a) has positive spillover effects on offline car
sales of the focal brand for two groups, supporting my H2. For the luxury group, U-Like (a) also
has the positive impact on offline car sales of the focal brand, supporting my H1. On the other
hand, C-U-Like (a) also shows positive spillover effects on offline car sales of the focal brand,
supporting H2. Finally, online WOM at the stage of consideration is not very effective in this
particular relationship.
246

Table 79. Bayesian Estimation Results for Likes (Luxury)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.011 (0.007)
0.077 (0.016)
0.016 (0.005)
-0.006 (0.015)
0.024 (0.016)
-0.028 (0.021)
0.11 (0.013)
-0.045 (0.1)
-0.004 (0.064)
-0.172 (0.1)
0.137 (0.064)

Sales A, t
95% Credible Level
(-0.025, 0.003)
(0.046, 0.1)
(0.006, 0.026)
(-0.035, 0.023)
(-0.008, 0.055)
(-0.069, 0.013)
(0.09, 0.141)
(-0.242, 0.153)
(-0.131, 0.122)
(-0.37, 0.023)
(0.012, 0.263)

Table 80. Bayesian Estimation Results for Likes (Non-Luxury)
Parameters
F-Like (a) A, t-1
C-F-Like (a) J, t-1
U-Like (a) A, t-1
C-U-Like (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
-0.0007 (0.006)
(-0.013, 0.011)
0.061 (0.012)
(0.037, 0.084)
-0.012 (0.004)
(-0.021, -0.003)
0.04 (0.011)
(0.017, 0.062)
0.005 (0.02)
(-0.02, 0.035)
-0.029 (0.018)
(-0.065, 0.007)
0.11 (0.014)
(0.085, 0.141)
-0.038 (0.072)
(-0.177, 0.1)
0.239 (0.035)
(0.169, 0.31)
0.05 (0.072)
(-0.09, 0.19)
0.045 (0.048)
(-0.049, 0.139)

Figure 360. Assessment of Model Convergence for F-Like (a)
(Luxury)

247

Figure 361. Assessment of Model Convergence for C-F-Like (a)
(Luxury)

Figure 362. Assessment of Model Convergence for U-Like (a)
(Luxury)

Figure 363. Assessment of Model Convergence for C-U-Like (a)
(Luxury)

248

Figure 364. Assessment of Model Convergence for TD-Post (c) (Luxury)

Figure 365. Assessment of Model Convergence for C-TD-Post (c)
(Luxury)

Figure 366. Assessment of Model Convergence for TMS (Luxury)

249

Figure 367. Assessment of Model Convergence for Price (Luxury)

Figure 368. Assessment of Model Convergence for GT (Luxury)

Figure 369. Assessment of Model Convergence for GPI (Luxury)

250

Figure 370. Assessment of Model Convergence for CCI (Luxury)

Figure 371. Assessment of Model Convergence for F-Like (a) (Non-Luxury)

Figure 372. Assessment of Model Convergence for C-F-Like (a) (Non-Luxury)

251

Figure 373. Assessment of Model Convergence for U-Like (a) (Non-Luxury)

Figure 374. Assessment of Model Convergence for C-U-Like (a) (Non-Luxury)

Figure 375. Assessment of Model Convergence for TD-Post (c) (Non-Luxury)

252

Figure 376. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury)

Figure 377. Assessment of Model Convergence for TMS (Non-Luxury)

Figure 378. Assessment of Model Convergence for Price (Non-Luxury)

253

Figure 379. Assessment of Model Convergence for GT (Non-Luxury)

Figure 380. Assessment of Model Convergence for GPI (Non-Luxury)

Figure 381. Assessment of Model Convergence for CCI (Non-Luxury)

254

Tables 81 to 82 show my sample split Bayesian estimation results at the comment level
(i.e., “Comment” associated with posts at Facebook and test drive post) for luxury and nonluxury group, respectively. I ran the MCMC chain for 1,849,921 iterations and, I discarded the
first 250,000 iterations to ensure convergence. Because of the high auto-correlation of parameter
posterior draws, I thinned the sampling chain and kept one from every eightieth iterations for the
remaining iterations. The assessment of model convergence (see Figures 382 to 403) suggested
that the model specification converged for each relationship.
The results in the comment level are consistent with main results (see Table 43) with
some differences. For example, I find that for the luxury group, the volume of comment
associated with competitors’ user posts (U-Comment (a)) has the positive impact on offline car
sales of the focal brand, supporting H1. Furthermore, I also observe that C-U-Comment (a) has
positive spillover effects on offline car sales of the focal brand for the non-luxury group only,
supporting my H2. Regarding online WOM at the stage of consideration, similar to main results,
I find negative spillover effects (C-TD-Post (c) for these two different groups.
Table 81. Bayesian Estimation Results for Comments (Luxury)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.013 (0.008)
0.124 (0.019)
0.018 (0.007)
-0.009 (0.02)
0.018 (0.016)
-0.048 (0.021)
0.115 (0.013)
-0.077 (0.1)
0.001 (0.064)
-0.089 (0.091)
0.22 (0.059)

255

Sales A, t
95% Credible Level
(-0.03, 0.002)
(0.088, 0.161)
(-0.014, 0.014)
(0.004, 0.032)
(-0.013, 0.049)
(-0.09, -0.007)
(0.089, 0.141)
(-0.28, 0.11)
(-0.12, 0.13)
(-0.26, 0.09)
(0.1, 0.336)

Table 82. Bayesian Estimation Results for Comments (Non-Luxury)
Parameters
F-Comment (a) A, t-1
C-F-Comment (a) J, t-1
U-Comment (a) A, t-1
C-U-Comment (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
0.008 (0.007)
0.09 (0.014)
-0.019 (0.006)
0.027 (0.014)
0.0009 (0.015)
-0.043 (0.027)
0.114 (0.015)
0.031 (0.073)
0.23 (0.038)
0.31 (0.068)
0.189 (0.046)

Sales A, t
95% Credible Level
(-0.005, 0.021)
(0.063, 0.12)
(-0.03, -0.008)
(0.0002, 0.053)
(-0.029, 0.032)
(-0.08, -0.006)
(0.085, 0.143)
(-0.11, 0.174)
(0.157, 0.306)
(0.173, 0.443)
(0.097, 0.279)

Figure 382. Assessment of Model Convergence for F-Comment (a) (Luxury)

Figure 383. Assessment of Model Convergence for C-F-Comment (a) (Luxury)

256

Figure 384. Assessment of Model Convergence for U-Comment (a) (Luxury)

Figure 385. Assessment of Model Convergence for C-U-Comment (a) (Luxury)

Figure 386. Assessment of Model Convergence for TD-Post (c) (Luxury)

257

Figure 387. Assessment of Model Convergence for C-TD-Post (c) (Luxury)

Figure 388. Assessment of Model Convergence for TMS (Luxury)

Figure 389. Assessment of Model Convergence for Price (Luxury)

258

Figure 390. Assessment of Model Convergence for GT (Luxury)

Figure 391. Assessment of Model Convergence for GPI (Luxury)

Figure 392. Assessment of Model Convergence for CCI (Luxury)

259

Figure 393. Assessment of Model Convergence for F-Comment (a) (Non-Luxury)

Figure 394. Assessment of Model Convergence for C-F-Comment (a) (Non-Luxury)

Figure 395. Assessment of Model Convergence for U-Comment (a) (Non-Luxury)

260

Figure 396. Assessment of Model Convergence for C-U-Comment (a) (Non-Luxury)

Figure 397. Assessment of Model Convergence for TD-Post (c) (Non-Luxury)

Figure 398. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury)

261

Figure 399. Assessment of Model Convergence for TMS (Non-Luxury)

Figure 400. Assessment of Model Convergence for Price (Non-Luxury)

Figure 401. Assessment of Model Convergence for GT (Non-Luxury)

262

Figure 402. Assessment of Model Convergence for GPI (Non-Luxury)

Figure 403. Assessment of Model Convergence for CCI (Non-Luxury)

Tables 83 to 84 show my sample split Bayesian estimation results at the share level (i.e.,
“Share” associated with posts at Facebook and test drive post) for luxury and non-luxury group,
respectively. In these models, I ran the MCMC chain for 999,961 iterations and, I discarded the
first 200,000 iterations to ensure convergence. Because of the high auto-correlation of parameter
posterior draws, I thinned the sampling chain and kept one from every fortieth iterations for the
remaining iterations. The assessment of model convergence (see Figures 404 to 425) suggested
263

that the model specification converged for each relationship.
First, I find that the results for the luxury group are consistent with main results (see
Table 44). That is, C-F-Share (a) has positive spillover effects on offline car sales of the focal
brand and TD-Post (c) also positively influences offline car sales of the focal brand, thereby
supporting both H2 and H3. However, for the non-luxury group, I only observe that C-F-Share
(a) has positive spillover effects on offline car sales of the focal brand, supporting H3.
To summarize, a set of split analysis suggests that when leveraging online WOM to help
purchase decisions, origin of brand, market structure of brand, and price factor of brand do vary
dynamics between online WOM about the focal brand, online WOM about competitors, and
offline car sales of the focal brand. Furthermore, the results provide further evidence that no
every mechanism at Facebook (i.e., post, like, comment, and share) has the equal impact on
offline car sales and these different mechanisms also influence how customers appreciate online
WOM at the stage of consideration. Table 85 shows the summarized results of this study.
Table 83. Bayesian Estimation Results for Shares (Luxury)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Posterior Mean
-0.006 (0.006)
0.052 (0.011)
0.0009 (0.006)
-0.02 (0.011)
0.039 (0.016)
-0.023 (0.024)
0.12 (0.014)
-0.052 (0.105)
0.003 (0.067)
0.243 (0.118)
0.283 (0.06)

264

Sales A, t
95% Credible Level
(-0.017, 0.004)
(0.03, 0.073)
(-0.011, 0.013)
(-0.043, 0.002)
(0.005, 0.071)
(-0.007, 0.025)
(0.095, 0.148)
(-0.257, 0.155)
(-0.136, 0.137
(0.01, 0.472)
(0.166, 0.401)

Table 84. Bayesian Estimation Results for Shares (Non-Luxury)
Parameters
F-Share (a) A, t-1
C-F-Share (a) J, t-1
U-Share (a) A, t-1
C-U-Share (a) J, t-1
TD-Post (c) A, t-1
C-TD-Post (c) J, t-1
TMS A, t-1
Price A, t-1
GT A, t-1
GPI A, t-1
CCI A, t-1

Sales A, t
Posterior Mean
95% Credible Level
0.004 (0.004)
(-0.003, 0.011)
0.045 (0.008)
(0.029, 0.059)
0.007 (0.006)
(-0.0009, 0.015)
-0.013 (0.008)
(-0.028, 0.002)
0.003 (0.015)
(-0.026, 0.033)
-0.002 (0.019)
(-0.039, 0.036)
0.11 (0.015)
(0.084, 0.141)
-0.024 (0.072)
(-0.165, 0.118)
0.178 (0.036)
(0.11, 0.248)
0.311 (0.083)
(0.144, 0.474)
0.179 (0.044)
(0.091, 0.265)

Figure 404. Assessment of Model Convergence for F-Share (a) (Luxury)

Figure 405. Assessment of Model Convergence for C-F-Share (a) (Luxury)

265

Figure 406. Assessment of Model Convergence for U-Share (a) (Luxury)

Figure 407. Assessment of Model Convergence for C-U-Share (a) (Luxury)

Figure 408. Assessment of Model Convergence for TD-Post (c) (Luxury)

266

Figure 409. Assessment of Model Convergence for C-TD-Post (c) (Luxury)

Figure 410. Assessment of Model Convergence for TMS (Luxury)

Figure 411. Assessment of Model Convergence for Price (Luxury)

267

Figure 412. Assessment of Model Convergence for GT (Luxury)

Figure 413. Assessment of Model Convergence for GPI (Luxury)

Figure 414. Assessment of Model Convergence for CCI (Luxury)

268

Figure 415. Assessment of Model Convergence for F-Share (a) (Non-Luxury)

Figure 416. Assessment of Model Convergence for C-F-Share (a) (Non-Luxury)

Figure 417. Assessment of Model Convergence for U-Share (a) (Non-Luxury)

269

Figure 418. Assessment of Model Convergence for C-U-Share (a) (Non-Luxury)

Figure 419. Assessment of Model Convergence for TD-Post (c) (Non-Luxury)

Figure 420. Assessment of Model Convergence for C-TD-Post (c) (Non-Luxury)

270

Figure 421. Assessment of Model Convergence for TMS (Non-Luxury)

Figure 422. Assessment of Model Convergence for Price (Non-Luxury)

Figure 423. Assessment of Model Convergence for GT (Non-Luxury)

271

Figure 424. Assessment of Model Convergence for GPI (Non-Luxury)

Figure 425. Assessment of Model Convergence for CCI (Non-Luxury)

272

Analysis level
Whole Sample
Post
Like
Comment
Share

Table 85. Summary of Results
Results

H1 (F), H2 (F), H3, and H4
H2 (F), H3, and H4
H2 (F) and H4
H2 (F) and H3

Sample Split on Origin of Brand
Post
Like
Comment
Share

Asian: H2 (F), H3, and H4; European: H1 (F) and H2 (F); US: H2 (F)
Asian: H2 (F); European: H2 (F); US: H2 (F) and H2 (U)
Asian: H2 (F) and H4; European: H1 (U), H2 (F), and H4; US: H2 (F) and H2 (U)
Asian: H2 (F) and H3; European: H2 (F) and H3; US: H2 (F)

Sample Split on Market Structure
Post
Like
Comment
Share

Group 1: H2 (F), H3, and H4; Group 2: H2 (F); Group 3: H1 (F) and H2 (F)
Group 1: H1 (U), H2 (F), and H3; Group 2: H2 (F) and H2 (U); Group 3: H2 (F)
Group 1: H4; Group 2: H2 (F) and H2 (F); Group 3: H1 (F) and H2 (F)
Group 1: H2 (F) and H3; Group 2: H1 (U) and H2 (F); Group 3: H2 (F)

Sample Split on Price Factor
Post
Like
Comment
Share

Luxury: H2 (F), H3, and H4; Non-Luxury: H1 (F) and H2 (F)
Luxury: H1 (U) and H2 (F); Non-Luxury: H2 (F) and H2 (U)
Luxury: H1 (U), H2 (F), and H4; Non-Luxury: H2 (F), H2 (U), and H4
Luxury: H2 (F) and H3; Non-Luxury: H2 (F)

Notes: (F) refers to metrics initiated by firms at the stage of awareness (e.g., F-Post (a), F-Share (a), or C-F-Post
(a)); (U) refers to metrics initiated by users at the stage of awareness (e.g., U-Post (a), U-Share (a), or C-U-Post (a)).

3.6 DISCUSSION
Online WOM plays an important aspect in consumers’ purchase decisions (Chen et al.,
2015) and provide a means for firms to learn about customers and the marketplace (Borah &
Tellis, 2016). Despite the richness of research on online WOM, prior research on the dynamics
of online WOM and its spillover effects is very scarce with few exceptions (e.g., Borah & Tellis,
2016; Chae et al., 2017; Sabnis & Grewal, 2015). More importantly, although prior research
indicates that customers may experience a multi-stage process before making purchase decisions
(i.e., awareness, consideration, and final decision) (Bettman, 1979; De Bruyn & Lilien, 2008;
Hauser & Wernerfelt, 1990; Shocker et al., 1991), our knowledge on the relative effects played
by online WOM at the stage of awareness and consideration is still not clear. Therefore, the
273

objective of this paper is to examine the dynamics of online WOM and its spillover effects by
considering their relative effects at the stages of customer awareness and consideration in the
U.S. automobile industry.
Generally, the results indicate that (1) online WOM at the stage of consideration has the
stronger effect on offline car sales than online WOM at the stage of awareness, (2) spillover
effects exist across both stages of awareness and consideration, though effects are heterogeneous
in direction: positive spillover effects at the stage of awareness while negative spillover effects at
the stage of consideration, and (3) at the stage of awareness, online WOM initiated by firms is
more effective in influencing offline car sales than online WOM initiated by users. Furthermore,
not every mechanism at Facebook (i.e., post, like, comment, and share) has the equal impact on
offline car sales and these different mechanisms also influence how customers appreciate online
WOM at the stage of consideration. Finally, the results vary significantly across origin of brand,
market structure, and price factor. These varied results imply that (1) customers might use these
pre-existed attributes to make similar inferences for brands that belong to the same pre-existed
group and (2) depending upon their belonging groups, firms should leverage different
mechanism to strategically drive offline sales.
3.6.1 Theoretical Implications

There are several key contributions from this research. First, the current study is the first
study that examines the relative effects played by online WOM and its spillover effects across

274

two stages of customer awareness and consideration on offline car sales. Recent research has
noted the need to account for competition as means to extend our understanding on different
aspects of online WOM (Dou, Niculescu, & Wu, 2013) and has provided evidence of online
WOM spillover effects in some settings (e.g., Borah & Tellis, 2016; Chae et al., 2017; Sabnis &
Grewal, 2015). Examining online WOM and its spillover effects across two stages of decision
process, I find a variation in the sign of the results that though the effects of online WOM about
competitors is statistically significant, the precise nature of the effect depends on whether online
WOM provides sufficient or concrete information to help customers make purchase decisions.
Therefore, the current study contributes to the consumer choice literature by showing different
role of online WOM played across stages of awareness and consideration. In addition, this study
sheds new lights on the effectiveness of online WOM and the competition nature of online
WOM (e.g., Borah & Tellis, 2016; Chae et al., 2017; Sabnis & Grewal, 2015) and shows how the
competition nature may vary across different stages of consumer decision process. Furthermore,
my examination on the effect of online WOM on high-involvement products also responds to the
call for more research regarding the impact of online WOM on different types of products (e.g.,
Goh et al., 2013; Stephen & Galak, 2012).
Second, the approach of quantifying the relative effectiveness of online WOM initiated
by firms and online WOM by users at the stage of awareness (i.e., firm’s Facebook page) also
echoes the call for more research on how firm’s media channels should operate as a system (e.g.,

275

Dewan & Ramaprasad, 2014; Luo et al., 2013). The predominant emphasis of prior research
focuses on the isolated impact of either online WOM initiated by firms or online WOM initiated
on online sales of media or low-involvement products such as movie or music. The current
contributes to this stream of the literature by showing the relative effectiveness of these two
sources of online WOM at the stage of awareness. Furthermore, our findings also suggest that
not every mechanism at Facebook has the equal impact on offline car sales and these different
mechanisms influence how customers appreciate online WOM at the stage of consideration and
its corresponding impact. Thus, my study sheds lights on decision-making and marketing
literature by demonstrating how firms could utilize the combinations of different mechanisms to
strategically enhance firm performance.
Finally, my split sample analysis on origin of brand, market structure, and price factor
offers very interesting insights regarding how customers may attribute to their pre-existed
attributes to make similar inferences for brands that belong to the same pre-existed group. For
example, at the post level (posts at Facebook pages and test drive posts), online WOM at the
stage of consideration has the positive impact on offline car sales of the focal brand and online
WOM has negative spillover effects on offline car sales of the focal brand for the Asian-based
group. However, these relationships cannot observe for the European-based group. Therefore,
this study contributes to the current literature by showing different patterns of how firms from
different pre-existed groups could develop their marketing efforts more effectively to leverage

276

online WOM.
3.6.2 Managerial Implications

The study also provides important managerial implications. Practitioners have considered
online WOM as one of important tools to influence customer decision-making. The current study
has shown that the need to consider online WOM across different stages of customer decision
process and that the effect of online WOM and its spillover effects vary differently across stages.
Therefore, in the competitive marketplace, practitioners not only need to observe online WOM
regarding their own brand but also need to monitor online WOM regarding their competitors.
More importantly, they need to realize that online WOM has different aspects and that depending
on the sufficient degree that online WOM may provide (i.e., abstract information versus concern
information) firms need to place different weights to better understand consumer perceptions and
predict performance.
Second, when leverage their Facebook pages, firms should pay special attention on
different mechanisms provided by Facebook. My results indicate that the volume of post at
Facebook is more effective in influencing offline car sales than the volume of like or the volume
of comment does. Thus, given the limited resource each firm may have, firms could leverage my
findings to better allocate their resource and effort to leverage online WOM. Finally, my
examination on the relative effect of online WOM initiated by firms and online WOM initiated
by users at the stage of awareness also provides insights to managers to better understand how

277

they could operate their marketing channel as a whole system and the individual effect within the
system.
3.6.3 Limitations and Future Research

This study does have some limitations. First, when considering online WOM at the stage
of awareness, I only focus on firms Facebook fan pages. It is very likely that firms also leverage
other social media channels to enhance their customer awareness and then make profits. Due to
data limitations, I am not able to study these cross-channel effects at the stage of awareness.
Second, the current study only focuses on the U.S. automobile industry. Although this setting
allows me to echo the call for more research on the impact of online WOM on high-involvement
products, scholars should be still cautious of the unique characteristics of the automobile industry.
Thus, it would be worthwhile to investigate the generalizability of the results to other highinvolvement product categories.
Finally, the current study only explores the volume of online WOM across two stages of
decision process. It may be interesting for future research to apply some machine learning
approaches to extract aspects such as emotions embedded in online WOM and examine their
corresponding impact.

3.7 CONCLUDING REMARKS
Despite the richness of research on online WOM, prior research on the dynamics of

278

online WOM and its spillover effects is very scarce. Particularly, our knowledge on the role
played by online WOM at the stages of customer awareness and consideration is very limited.
This paper attempts to examine the dynamics of online WOM and its spillover effects by
considering their relative effects at the stages of customer awareness and consideration in the
U.S. automobile industry. The findings show that spillover effects are heterogeneous in direction
across two stages and that different mechanisms at the stage of awareness would influence how
customers appreciate online WOM at the stage of consideration. Furthermore, the dynamics of
online WOM vary across origin of brand, market structure, and price factor.

279

REFERENCES

280

REFERENCES

Abernethy, A. M., & Franke, G. R. (1996). The information content of advertising: A metaanalysis. Journal of Advertising, 25(2), 1-17.
Abrahams, A. S., Jiao, J., Wang, G. A., & Fan, W. (2012). Vehicle defect discovery from social
media. Decision Support Systems, 54(1), 87-97.
Abrigo, M. R., & Love, I. (2015). Estimation of Panel Vector Autoregression in Stata: a Package
of Programs. Retrieved from http://paneldataconference2015.ceu.hu/Program/MichaelAbrigo.pdf
Adomavicius, G., Bockstedt, J., & Gupta, A. (2012). Modeling supply-side dynamics of IT
components, products, and infrastructure: An empirical analysis using vector
autoregression. Information Systems Research, 23(2), 397-417.
Agarwal, R., Animesh, A., & Prasad, K. (2009). Research Note-Social Interactions and the
“Digital Divide”: Explaining Variations in Internet Use. Information Systems Research,
20(2), 277-294.
Ahluwalia, R., & GĂźrhan-Canli, Z. (2000). The effects of extensions on the family brand name:
An accessibility-diagnosticity perspective. Journal of Consumer Research, 27(3), 371381.
Ahluwalia, R., Unnava, H. R., & Burnkrant, R. E. (2001). The moderating role of commitment
on the spillover effect of marketing communications. Journal of Marketing Research,
38(4), 458-470.
Akaike, H. (1969). Fitting autoregressive models for prediction. Annals of the Institute of
Statistical Mathematics, 21(1), 243-247.
Algesheimer, R., Dholakia, U. M., & Herrmann, A. (2005). The social influence of brand
community: Evidence from European car clubs. Journal of Marketing, 69(3), 19-34.
Anderson, J. R. (2013). The architecture of cognition: Psychology Press.

281

Andrews, D. W., & Lu, B. (2001). Consistent model and moment selection procedures for GMM
estimation with application to dynamic panel data models. Journal of Econometrics,
101(1), 123-164.
Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of errorcomponents models. Journal of Econometrics, 68(1), 29-51.
Backstrom, L. (2013). News Feed FYI: A Window Into News Feed. Retrieved from
https://www.facebook.com/business/news/News-Feed-FYI-A-Window-Into-News-Feed
Berger, B., Matt, C., Steininger, D. M., & Hess, T. (2015). It Is Not Just About Competition with
“Free”: Differences Between Content Formats in Consumer Preferences and Willingness
to Pay. Journal of Management Information Systems, 32(3), 105-128.
Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of Marketing
Research, 49(2), 192-205.
Berger, J., & Schwartz, E. M. (2011). What drives immediate and ongoing word of mouth?
Journal of Marketing Research, 48(5), 869-880.
Berger, J., Sorensen, A. T., & Rasmussen, S. J. (2010). Positive effects of negative publicity:
When negative reviews increase sales. Marketing Science, 29(5), 815-827.
Bettman, J. R. (1979). Information processing theory of consumer choice: Addison-Wesley Pub.
Co.
Borah, A., & Tellis, G. J. (2016). Halo Effects in Social Media: Do Product Recalls Hurt or Help
Rival Brands. Journal of Marketing Research, 53(2), 143-160.
Broniarczyk, S. M., & Alba, J. W. (1994a). The importance of the brand in brand extension.
Journal of Marketing Research, 31(2), 214-228.
Broniarczyk, S. M., & Alba, J. W. (1994b). The role of consumers' intuitions in inference
making. Journal of Consumer Research, 21(3), 393-407.
Cao, Q., Gedajlovic, E., & Zhang, H. (2009). Unpacking organizational ambidexterity:
Dimensions, contingencies, and synergistic effects. Organization Science, 20(4), 781796.

282

Chae, I., Stephen, A. T., Bart, Y., & Yao, D. (2017). Spillover effects in seeded word-of-mouth
marketing campaigns. Marketing Science, 31(1), 89–104.
Chen, H., De, P., & Hu, Y. J. (2015). IT-enabled broadcasting in social media: An empirical
study of artists’ activities and music sales. Information Systems Research, 26(3), 513–
531.
Chen, Y., & Xie, J. (2008). Online consumer review: Word-of-mouth as a new element of
marketing communication mix. Management Science, 54(3), 477-491.
Chevalier, J. A., & Mayzlin, D. (2006). The effect of word of mouth on sales: Online book
reviews. Journal of Marketing Research, 43(3), 345-354.
Choi, I. (2001). Unit root tests for panel data. Journal of International Money and Finance,
20(2), 249-272.
Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory of semantic processing.
Psychological Review, 82(6), 407.
De Bruyn, A., & Lilien, G. L. (2008). A multi-stage model of word-of-mouth influence through
viral marketing. International Journal of Research in Marketing, 25(3), 151-163.
Debatin, B., Lovejoy, J. P., Horn, A. K., & Hughes, B. N. (2009). Facebook and online privacy:
Attitudes, behaviors, and unintended consequences. Journal of Computer‐Mediated
Communication, 15(1), 83-108.
Deloitte. (2015). Facebook’s global economic impact. Retrieved from
http://newsroom.fb.com/news/2015/01/new-deloitte-report-looks-at-facebooks-impacton-global-economy-jobs/
Dewan, S., & Ramaprasad, J. (2014). Social Media, Traditional Media, and Music Sales. MIS
Quarterly, 38(1), 101-121.
Dhar, R., & Kim, E. Y. (2007). Seeing the forest or the trees: Implications of construal level
theory for consumer choice. Journal of Consumer Psychology, 17(2), 96-100.
Dhar, V., & Chang, E. A. (2009). Does chatter matter? The impact of user-generated content on
music sales. Journal of Interactive Marketing, 23(4), 300-307.

283

Dong, X., & Chintagunta, P. K. (2015). Satisfaction Spillovers Across Categories. Marketing
Science, 35(2), 275–283.
Dou, Y., Niculescu, M. F., & Wu, D. (2013). Engineering optimal network effects via social
media features and seeding in markets for digital goods and services. Information
Systems Research, 24(1), 164-185.
Drakos, K., & Konstantinou, P. T. (2014). Terrorism, crime and public spending: Panel VAR
evidence from europe. Defence and Peace Economics, 25(4), 349-361.
Duan, W., Gu, B., & Whinston, A. B. (2008). The dynamics of online word-of-mouth and
product sales—An empirical investigation of the movie industry. Journal of Retailing,
84(2), 233-242.
DubĂŠ, J.-P., & Manchanda, P. (2005). Differences in dynamic brand competition across markets:
An empirical analysis. Marketing Science, 24(1), 81-95.
eMarketer. (2014a). Digital Ad Spending in US Auto Industry Racing Ahead. Retrieved from
https://www.emarketer.com/Article/Digital-Ad-Spending-US-Auto-Industry-RacingAhead/1010872
eMarketer. (2014b). Retailers look to merge offline and online shopping experiences in 2014.
Retrieved from https://www.emarketer.com/Article/Retailers-Look-Merge-OfflineOnline-Shopping-Experiences-2014/1010812
eMarketer. (2015a). The US Automotive Industry 2015: Digital Ad Spending Forecast and
Trends. Retrieved from https://www.emarketer.com/Report/US-Automotive-Industry2015-Digital-Ad-Spending-Forecast-Trends/2001569
eMarketer. (2015b). US Digital Ad Spending, by Industry, 2014-2019. Retrieved from
eMarketer. (2015c). US Digital Users: Q1 2015 Complete Forecast. Retrieved from
Enders, W. (2008). Applied Econometric Time Series: John Wiley & Sons.
Erdem, T., & Swait, J. (1998). Brand equity as a signaling phenomenon. Journal of Consumer
Psychology, 7(2), 131-157.
Erdem, T., Swait, J., & Valenzuela, A. (2006). Brands as signals: A cross-country validation
284

study. Journal of Marketing, 70(1), 34-49.
Facebook. (2015a). Facebook Statistics. Retrieved from
http://www.statisticbrain.com/facebook-statistics/
Facebook. (2015b). Facebook Statistics. Retrieved from http://www.statisticbrain.com/facebookstatistics/
Feldman, J. M., & Lynch, J. G. (1988). Self-generated validity and other effects of measurement
on belief, attitude, intention, and behavior. Journal of Applied Psychology, 73(3), 421.
Feng, J., & Papatla, P. (2011). Advertising: stimulant or suppressant of online word of mouth?
Journal of Interactive Marketing, 25(2), 75-84.
Forman, C., Ghose, A., & Wiesenfeld, B. (2008). Examining the relationship between reviews
and sales: The role of reviewer identity disclosure in electronic markets. Information
Systems Research, 19(3), 291-313.
Geva, T., Oestreicher-Singer, G., Efron, N., & Shimshoni, Y. (2017). Using Forum and Search
Data for Sales Prediction of High-Involvement Products. MIS Quarterly, Forthcoming.
Ghose, A., Goldfarb, A., & Han, S. P. (2012). How is the mobile Internet different? Search costs
and local activities. Information Systems Research, 24(3), 613-631.
Ghose, A., & Ipeirotis, P. G. (2011). Estimating the helpfulness and economic impact of product
reviews: Mining text and reviewer characteristics. IEEE Transactions on Knowledge and
Data Engineering, 23(10), 1498-1512.
Godes, D., & Mayzlin, D. (2004). Using online conversations to study word-of-mouth
communication. Marketing Science, 23(4), 545-560.
Goh, K.-Y., Heng, C.-S., & Lin, Z. (2013). Social media brand community and consumer
behavior: Quantifying the relative impact of user-and marketer-generated content.
Information Systems Research, 24(1), 88-107.
Gopinath, S., Chintagunta, P. K., & Venkataraman, S. (2013). Blogs, advertising, and localmarket movie box office performance. Management Science, 59(12), 2635-2654.
Gopinath, S., Thomas, J. S., & Krishnamurthi, L. (2014). Investigating the relationship between
285

the content of online word of mouth, advertising, and brand performance. Marketing
Science, 33(2), 241-258.
Granger, C. W. J., & Newbold, P. (1986). Forecasting economic time series (2nd ed.). Orlando,
FL: Academic Press.
Gu, B., Park, J., & Konana, P. (2012). The impact of external word-of-mouth sources on retailer
sales of high-involvement products. Information Systems Research, 23(1), 182-196.
Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica: Journal of the Econometric Society, 50(4), 1029-1054.
Hauser, J. R., & Wernerfelt, B. (1990). An evaluation cost model of consideration sets. Journal
of Consumer Research, 16(4), 393-408.
Hill, K., Menk, D., & Cooper, A. (2010). Contribution of the Automotive Industry to the
Economies of all Fifty States and the United States. Retrieved from
http://www.cargroup.org/?module=Publications&event=View&pubID=16
Ho-Dac, N. N., Carson, S. J., & Moore, W. L. (2013). The effects of positive and negative online
customer reviews: do brand strength and category maturity matter? Journal of Marketing,
77(6), 37-53.
Holtz-Eakin, D., Newey, W., & Rosen, H. S. (1988). Estimating vector autoregressions with
panel data. Econometrica: Journal of the Econometric Society, 56(6), 1371-1395.
Homburg, C., Klarmann, M., & Schmitt, J. (2010). Brand awareness in business markets: When
is it related to firm performance? International Journal of Research in Marketing, 27(3),
201-212.
Hong, S.-T., & Wyer, R. S. (1990). Determinants of product evaluation: Effects of the time
interval between knowledge of a product's country of origin and information about its
specific attributes. Journal of Consumer Research, 17(3), 277-288.
Hong, W., Chan, F. K., Thong, J. Y., Chasalow, L. C., & Dhillon, G. (2013). A framework and
guidelines for context-specific theorizing in information systems research. Information
Systems Research, 25(1), 111-136.
Hoyer, W. D., & Brown, S. P. (1990). Effects of brand awareness on choice for a common,
286

repeat-purchase product. Journal of Consumer Research, 17(2), 141-148.
Im, K. S., Pesaran, M. H., & Shin, Y. (2003). Testing for unit roots in heterogeneous panels.
Journal of Econometrics, 115(1), 53-74.
Kalaignanam, K., Kushwaha, T., & Eilert, M. (2013). The impact of product recalls on future
product reliability and future accidents: Evidence from the automobile industry. Journal
of Marketing, 77(2), 41-57.
Kuruzovich, J., Viswanathan, S., Agarwal, R., Gosain, S., & Weitzman, S. (2008). Marketspace
or marketplace? Online information search and channel outcomes in auto retailing.
Information Systems Research, 19(2), 182-201.
Laurent, G., & Kapferer, J.-N. (1985). Measuring consumer involvement profiles. Journal of
Marketing Research, 22(1), 41-53.
Lei, J., Dawar, N., & Lemmink, J. (2008). Negative spillover in brand portfolios: exploring the
antecedents of asymmetric effects. Journal of Marketing, 72(3), 111-123.
Lewis, R., & Nguyen, D. (2015). Display advertising’s competitive spillovers to consumer
search. Quantitative Marketing and Economics, 13(2), 93-115.
Liberman, N., & Trope, Y. (1998). The role of feasibility and desirability considerations in near
and distant future decisions: A test of temporal construal theory. Journal of Personality
and Social Psychology, 75(1), 5.
Lim, K. H., & Benbasat, I. (2000). The effect of multimedia on perceived equivocality and
perceived usefulness of information systems. MIS Quarterly, 24(3), 449-471.
Liu, Y. (2006). Word of mouth for movies: Its dynamics and impact on box office revenue.
Journal of Marketing, 70(3), 74-89.
Liu, Y., & Shankar, V. (2015). The Dynamic Impact of Product-Harm Crises on Brand
Preference and Advertising Effectiveness: An Empirical Analysis of the Automobile
Industry. Management Science, 61(10), 2514-2535.
Loughran, T., & McDonald, B. (2011). When is a liability not a liability? Textual analysis,
dictionaries, and 10‐Ks. Journal of Finance, 66(1), 35-65.

287

Love, I., & Zicchino, L. (2006). Financial development and dynamic investment behavior:
Evidence from panel VAR. The Quarterly Review of Economics and Finance, 46(2), 190210.
Lovett, M. J., Peres, R., & Shachar, R. (2013). On brands and word of mouth. Journal of
Marketing Research, 50(4), 427-444.
Lu, X., Ba, S., Huang, L., & Feng, Y. (2013). Promotional marketing or word-of-mouth?
Evidence from online restaurant reviews. Information Systems Research, 24(3), 596-612.
Luo, X., & Zhang, J. (2013). How Do Consumer Buzz and Traffic in Social Media Marketing
Predict the Value of the Firm? Journal of Management Information Systems, 30(2), 213238.
Luo, X., Zhang, J., & Duan, W. (2013). Social media and firm equity value. Information Systems
Research, 24(1), 146-163.
MacArthur, A. (2015). Car buyers test-drive dealers on social media. USA Today.
Maheswaran, D., & Chen, C. Y. (2006). Nation equity: Incidental emotions in country-of-origin
effects. Journal of Consumer Research, 33(3), 370-376.
Maloney, J. C. (2000). Curiosity versus disbelief in advertising. Journal of Advertising Research,
40(06), 7-13.
Mathwick, C., & Rigdon, E. (2004). Play, flow, and the online search experience. Journal of
Consumer Research, 31(2), 324-332.
Miller, A. R., & Tucker, C. (2013). Active social media management: the case of health care.
Information Systems Research, 24(1), 52-70.
Moe, W. W., & Trusov, M. (2011). The value of social dynamics in online product ratings
forums. Journal of Marketing Research, 48(3), 444-456.
Moon, S., Bergey, P. K., & Iacobucci, D. (2010). Dynamic effects among movie ratings, movie
revenues, and viewer satisfaction. Journal of marketing, 74(1), 108-121.
Morgan, R. M., & Hunt, S. D. (1994). The commitment-trust theory of relationship marketing.
Journal of Marketing, 58(3), 20-38.
288

Nah, F. F.-H., Eschenbrenner, B., & DeWester, D. (2011). Enhancing brand equity through flow
and telepresence: A comparison of 2D and 3D virtual worlds. MIS Quarterly, 35(3), 731747.
Naik, P. A., Raman, K., & Winer, R. S. (2005). Planning marketing-mix strategies in the
presence of interaction effects. Marketing Science, 24(1), 25-34.
Narayanan, S., Chintagunta, P. K., & Miravete, E. J. (2007). The role of self selection, usage
uncertainty and learning in the demand for local telephone service. Quantitative
Marketing and Economics, 5(1), 1-34.
Nelson, P. (1970). Information and consumer behavior. Journal of Political Economy, 78(2),
311-329.
Netzer, O., Feldman, R., Goldenberg, J., & Fresko, M. (2012). Mine your own business: Marketstructure surveillance through text mining. Marketing Science, 31(3), 521-543.
Nielsen. (2015). Top 10 US Industries, Ranked by Ad Spending, 2014. Retrieved from
Pauwels, K., & Srinivasan, S. (2004). Who benefits from store brand entry? Marketing Science,
23(3), 364-390.
Resnik, A., & Stern, B. L. (1977). An analysis of information content in television advertising.
Journal of Marketing, 41(1), 50-53.
Richins, M. L., & Bloch, P. H. (1986). After the new wears off: The temporal context of product
involvement. Journal of Consumer Research, 13(2), 280-285.
Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013). The effect of customers' social
media participation on customer visit frequency and profitability: an empirical
investigation. Information Systems Research, 24(1), 108-127.
Roehm, M. L., & Tybout, A. M. (2006). When will a brand scandal spill over, and how should
competitors respond? Journal of Marketing Research, 43(3), 366-373.
Rossi, P. E., & Allenby, G. M. (2003). Bayesian statistics and marketing. Marketing Science,
22(3), 304-328.

289

Rui, H., Liu, Y., & Whinston, A. (2013). Whose and what chatter matters? The effect of tweets
on movie sales. Decision Support Systems, 55(4), 863-870.
Sabnis, G., & Grewal, R. (2015). Cable News Wars on the Internet: Competition and UserGenerated Content. Information Systems Research, 26(2), 301-319.
Schubert, P., & Ginsburg, M. (2000). Virtual communities of transaction: The role of
personalization in electronic commerce. Electronic Markets, 10(1), 45-55.
Shen, W., Hu, Y. J., & Rees, J. (2015). Competing for attention: An empirical study of online
reviewers’ strategic behaviors. MIS Quarterly, 39(3), 683-696.
Shiv, B., & Fedorikhin, A. (1999). Heart and mind in conflict: The interplay of affect and
cognition in consumer decision making. Journal of Consumer Research, 26(3), 278-292.
Shocker, A. D., Ben-Akiva, M., Boccara, B., & Nedungadi, P. (1991). Consideration set
influences on consumer decision-making and choice: Issues, models, and suggestions.
Marketing Letters, 2(3), 181-197.
Smith, T. M., Gopalakrishna, S., & Chatterjee, R. (2006). A three-stage model of integrated
marketing communications at the marketing-sales interface. Journal of Marketing
Research, 43(4), 564-579.
Srinivasan, S., & Hanssens, D. M. (2009). Marketing and firm value: Metrics, methods, findings,
and future directions. Journal of Marketing Research, 46(3), 293-312.
Stephen, A. T., & Galak, J. (2012). The effects of traditional and social earned media on sales: A
study of a microlending marketplace. Journal of Marketing Research, 49(5), 624-639.
Stock, J. H., & Watson, M. W. (2001). Vector autoregressions. Journal of Economic
Perspectives, 15(4), 101-115.
Tambe, P., & Hitt, L. M. (2013). Measuring information technology spillovers. Information
Systems Research, 25(1), 53-71.
Tang, T., Fang, E., & Wang, F. (2014). Is neutral really neutral? The effects of neutral usergenerated content on product sales. Journal of Marketing, 78(4), 41-58.
Tetlock, P. C., Saa‐Tsechansky, M., & Macskassy, S. (2008). More than words: Quantifying
290

language to measure firms' fundamentals. Journal of Finance, 63(3), 1437-1467.
Tirunillai, S., & Tellis, G. J. (2012). Does chatter really matter? Dynamics of user-generated
content and stock performance. Marketing Science, 31(2), 198-215.
Traylor, M. B. (1981). Product involvement and brand commitment. Journal of Advertising
Research, 21(6), 51-56.
Trope, Y., & Liberman, N. (2010). Construal-level theory of psychological distance.
Psychological Review, 117(2), 440.
Trope, Y., Liberman, N., & Wakslak, C. (2007). Construal levels and psychological distance:
Effects on representation, prediction, evaluation, and behavior. Journal of Consumer
Psychology, 17(2), 83-95.
Trusov, M., Bucklin, R. E., & Pauwels, K. (2009). Effects of word-of-mouth versus traditional
marketing: findings from an internet social networking site. Journal of Marketing, 73(5),
90-102.
U.S. Department of Commerce (2016). The 4th Quarter 2015 Retail E-Commerce Sales Report
Retrieved from https://www2.census.gov/retail/releases/historical/ecomm/15q4.pdf
Verma, R., Jahn, B., & Kunz, W. (2012). How to transform consumers into fans of your brand.
Journal of Service Management, 23(3), 344-361.
Yi, C., Jiang, Z., & Benbasat, I. (2015). Enticing and Engaging Consumers via Online Product
Presentations: The Effects of Restricted Interaction Design. Journal of Management
Information Systems, 31(4), 213-242.
Zaichkowsky, J. L. (1985). Measuring the involvement construct. Journal of Consumer
Research, 12(3), 341-352.
Zhou, W., & Duan, W. (2016). Do Professional Reviews Affect Online User Choices Through
User Reviews? An Empirical Study. Journal of Management Information Systems, 33(1),
202-228.
Zhu, F., & Zhang, X. (2010). Impact of online consumer reviews on sales: The moderating role
of product and consumer characteristics. Journal of marketing, 74(2), 133-148.

291

Zhu, L., Benbasat, I., & Jiang, Z. (2010). Let's shop online together: An empirical investigation
of collaborative online shopping support. Information Systems Research, 21(4), 872-891.

292