๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
2๏ธโƒฃ Study/๐Ÿ“‘ ๋…ผ๋ฌธ์ฝ๊ธฐ

[DiD, Matching] Popularity or Proximity

by isdawell 2023. 7. 2.
728x90

 

 

 

 

 

๐Ÿ‘€ Keyword


 

โ—ฏ  Quasi-experimental research design 

 

•  DiD  

•  DDD  

•  Probit model  

•  Hazard model   

 

 

 

โ—ฏ  Matching 

 

๊ทธ๋ฃน์˜ ํŠน์„ฑ์„ ์œ ์‚ฌํ•˜๊ฒŒ ๋งŒ๋“ค์–ด Treatment ์˜ ํšจ๊ณผ๋งŒ์„ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ๋„๋ก ํ•จ

 

 

•  CEM : ๋‹จ์ˆœํ•˜๊ฒŒ ํ†ต์ œ๋ณ€์ˆ˜๋“ค์ด ๋น„์Šทํ•œ ๊ด€์ธก์น˜๋ผ๋ฆฌ ๋งค์นญํ•˜๋Š” ๋ฐฉ๋ฒ• 

 

 

 

 

 

•  PSM : ํ†ต์ œ ๋ณ€์ˆ˜๊ฐ€ ์ฃผ์–ด์ง„ ์ƒํƒœ์—์„œ Treatment ๋ฅผ ๋ฐ›์„ ํ™•๋ฅ  (Propensity score) ์ด ๋น„์Šทํ•œ ๊ด€์ธก์น˜๋ผ๋ฆฌ ๋งค์นญ 

 

 

 

•  EDM : ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋งค์นญ 

 

 

 

 

 

 

๐Ÿ‘€ ๋ฐ์ดํ„ฐ ํ•ด์„์„ ์œ„ํ•œ ๋„๋ฉ”์ธ ์ง€์‹


 

โ—ฏ  The hype machine 

 

•  Largest MP3 blog aggregator ๋กœ ๋ธ”๋กœ๊ทธ์— ํฌ์ŠคํŒ… ๋œ ์Œ์•…/ํŠธ๋ž™ ๋ฆฌ์ŠคํŠธ๋“ค์„ ์ˆ˜์ง‘ํ•˜์—ฌ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•œ๋‹ค. ์œ ์ €๋“ค์€ ์Œ์•…์„ ์ŠคํŠธ๋ฆฌ๋ฐ ํ•  ์ˆ˜ ์ž‡์œผ๋ฉฐ, ์Œ์•… ๋‹ค์šด๋กœ๋“œ๋Š” ๋ถˆ๊ฐ€๋Šฅํ•˜๋‹ค. 

 

•  ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•œ ์‹œ์ ์—์„œ 15~20% ์— ํ•ด๋‹นํ•˜๋Š” ์œ ์ €๋“ค๋งŒ social networking ๊ธฐ๋Šฅ๋“ค์„ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์—ˆ๊ณ  ๋‚˜๋จธ์ง€ ์œ ์ €๋“ค์€ isolate (๋‹ค๋ฅธ ์œ ์ €๋“ค์„ ํŒ”๋กœ์šฐ ํ•˜์ง€ ์•Š์Œ) ๋˜์–ด ์žˆ์—ˆ๋‹ค. 

 

•  Popularity information ์€ ์›น์‚ฌ์ดํŠธ์˜ ๋ชจ๋“  ์œ ์ €๋“ค์—๊ฒŒ 10์›” 1์ผ์„ ๊ธฐ์ ์œผ๋กœ ๋ชจ๋‘์—๊ฒŒ ๋ณด์—ฌ์กŒ๋‹ค. 10์›” 1์ผ ์‹œ์  ์ด์ „์—๋„ ์œ ์ €๋“ค์€ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅผ ์ˆ˜ ์žˆ์—ˆ๋‹ค. (favorites ์ˆซ์ž๋Š” ๋ณผ ์ˆ˜ ์—†์—ˆ์Œ)

 

•  ์œ ์ €๋“ค์—๊ฒŒ social network ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐœ๋ณ„์ ์ธ ๋Œ€์‹œ๋ณด๋“œ๋ฅผ ์ œ๊ณตํ•˜๋Š”๋ฐ, ์ข‹์•„ํ•˜๋Š” ํŠธ๋ž™์ด๋‚˜ ์œ ์ €๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค. (add favorite tracks and favorite users) ์œ ์ €๋ฅผ favorite ํ•˜๋Š” ๊ฒƒ์€ following ๊ณผ ๋น„์Šทํ•œ ํ–‰๋™์ด๋‹ค. (์ผ๋ฐฉํ–ฅ ํŒ”๋กœ์šฐ๋„ ๊ฐ€๋Šฅ) 

 

 

 

 

โ‘   Abstract 


 

โ—ฏ  Research Topic 

 

•  ์˜จ๋ผ์ธ ์Œ์•… ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์Œ์•… ์†Œ๋น„์— ์žˆ์–ด Popularity influence (๋Œ€์ค‘์„ฑ) ์™€ Proximity influence (๊ทผ์ ‘์„ฑ) ๊ฐ€ ๋ฏธ์น˜๋Š” ์˜ํ–ฅ 

 

 

 

โ—ฏ  Research Question 

 

•  RQ1. How does popularity influence affect music consumption choices?

•  RQ2. Is popularity influence more important for mainstream or niche music?

•  RQ3. How important is proximity influence on music consumption?

•  RQ4. What is the nature of the interaction between the two types of influence? Are they complements or substitutes?

 

 

 

โ—ฏ  Research Method 

 

•  quasi-experimental research design 

•  highly granular data from an online music community

 

 

 

โ—ฏ  Research Results 

 

•  Popularity influence ์™€ Proximity influence ๋ชจ๋‘ ์Œ์•…์†Œ๋น„์— Positive ํ•œ ์˜ํ–ฅ์„ ๋ผ์นจ์„ ์ฆ๋ช… 

•  ๋‘ ์˜ํ–ฅ์€ ์„œ๋กœ substitute ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, Proximity ์˜ํ–ฅ์ด ์žˆ๋Š” ๊ฒฝ์šฐ Popularity ์˜ํ–ฅ์„ ๋›ฐ์–ด ๋„˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. 

 

 

 

โ—ฏ  Research Contribution 

 

•   Design and marketing strategies for online communities

 

 

 

 

 

 

โ‘ก  Introduction 


 

โ—ฏ  Preliminary 

 

A ์œ ์ €๋ฅผ ๊ธฐ์ค€์œผ๋กœํ•˜์—ฌ ํ•ด์„

 

 

•  Total favorites → Popularity influence,  Friend favorites → Proximity influence 

 

•  ex. A ์œ ์ €๊ฐ€ ์ค‘์‹ฌ ์œ ์ €์— ํ•ด๋‹นํ•  ๋•Œ, Song1 ์˜ ๊ฒฝ์šฐ ์ด 5๋ช…์˜ ์œ ์ €๋“ค์ด ์ข‹์•„ํ•˜๊ณ , A์œ ์ €์™€ ์นœ๊ตฌ๋ฅผ ๋งบ๊ณ ์žˆ๋Š” ์œ ์ € ์ค‘์—๋Š” C๋งŒ ์ข‹์•„์š”๋ฅผ ๋ˆŒ๋ €๋‹ค. Song2 ์˜ ๊ฒฝ์šฐ ์ด 2๋ช…์˜ ์œ ์ €๊ฐ€ ์ข‹์•„ํ•˜๊ณ , A์œ ์ €์™€ ์นœ๊ตฌ์ธ ์œ ์ €๋“ค ์ค‘์—๋Š” ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์‚ฌ๋žŒ์ด ์—†๋‹ค. 

 

 

 

โ—ฏ  Research Design 

 

(์ง์ ‘ ์ •๋ฆฌํ•œ ํ‘œ)

 

•  Popularity influence  →  ์œ ์ €์˜ ์ข‹์•„์š” ํ–‰๋™ ์ง‘๊ณ„๋ฅผ "๋ณด์ด๊ฒŒ ํ•œ ๊ธฐ๋Šฅ" (์ด ์ข‹์•„์š” ์ˆ˜๊ฐ€ ๋ณด์ด๋Š”์ง€ ์—ฌ๋ถ€) ์„ ๋Œ€์ƒ์œผ๋กœ ํ•˜์—ฌ ๋ถ„์„์„ ์ง„ํ–‰ : CEM matching and DiD 

 

•  Proximity influence  →  ํŒ”๋กœ์šฐํ•œ ์นœ๊ตฌ์˜ ํŠน์ • ๋…ธ๋ž˜์— ๋Œ€ํ•œ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ํ–‰๋™ (focal user ๊ฐ€ ์ข‹์•„ํ•˜๋Š” ๋…ธ๋ž˜๋ฅผ ์ฃผ๋ณ€ ์นœ๊ตฌ๋“ค๋„ ์ข‹์•„ํ•˜๊ณ  ์žˆ๋Š”์ง€ ์—ฌ๋ถ€) ์„ ๋Œ€์ƒ์œผ๋กœ ๋ถ„์„์„ ์ง„ํ–‰ : PSM or EDM matching and Probit model, Hazard model 

 

 

 

 

โ‘ข  Literature Review 


 

โ—ฏ  Popularity influence 

 

•   Online Word of mouth (Mizerski 1982, Chevalier and Mayzlin 2006, Liu 2006)

•   Observational learning (Sorenson 2007, Duan et al. 2009, Salganik et al. 2006)

โ‡จ  In study, the number of favorites for a song is hybrid of WOM and OL

 

 

โ—ฏ  Proximity influence 

 

•  Influence in social networks (Brown et al. 1987, Valente 1995, Katz et.al 1955, Granovetter 1973)

•  Separating social influence and homophily : Aral et al.2009 (Dynamic matched sample estimation) โญ

 

 

 

 

 

โ‘ฃ  Methodology 


 

โ—ฏ  Data

 

•   The hype machine ์ด๋ผ๋Š” ์Œ์•… ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ 

 

 

•   Popularity influence : Feature implementation (์ข‹์•„์š” ์ˆ˜๋ฅผ ๋ณด์ด๊ฒŒ ํ•œ ๊ธฐ๋Šฅ) ์„ ๋„์ž…ํ•œ ์‹œ์  ์ „ํ›„๋ฅผ ๊ธฐ์ค€์œผ๋กœ, Song level ์—์„œ ํŠน์ • ์Œ์•…์— ๋Œ€ํ•œ ์ด ์ฒญ์ทจ ํšŸ์ˆ˜๋ฅผ ์ถ”์  

 

•   Proximity influence : User-song level ์—์„œ ํŠน์ • ์œ ์ €๋ฅผ ๊ธฐ์ค€์œผ๋กœ ํŠน์ • ์Œ์•…์— ๋Œ€ํ•ด, ์ฃผ๋ณ€ ์นœ๊ตฌ๋“ค๋„ ์ข‹์•„์š”๋ฅผ ๋ˆŒ๋ €๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์Œ์•… ์ฒญ์ทจ ์—ฌ๋ถ€๋ฅผ ์ถ”์  

 

 

•   Descriptive statistics 

 

song level

 

 

user-song level

 

 

 

 

โ—ฏ  Variable

 

 

•  Popularity influence → DiD (Song level)

โ†ช  Listens_jt : song j ์— ๋Œ€ํ•œ time t ์‹œ์ ์˜ ์ด ์ฒญ์ทจ ํšŸ์ˆ˜ 

โ†ช  PopTreatment_j : song j ์— ๋Œ€ํ•œ ์ด ์ข‹์•„์š” ๊ฐœ์ˆ˜๊ฐ€ ๋ณด์ด๋Š”์ง€์— ๋Œ€ํ•œ ์—ฌ๋ถ€ 

โ†ช  After_t : popularity treatment ์ดํ›„ ์‹œ์ ์ธ์ง€์— ๋Œ€ํ•œ ์—ฌ๋ถ€ (์ด ์ข‹์•„์š” ๊ฐœ์ˆ˜๊ฐ€ ๋ณด์ด๋Š” ๊ธฐ๋Šฅ ๋„์ž… ์‹œ์  ์ดํ›„)

 

 

•  Proximity influence  (User-song level)

โ†ช  Listen_ij : ์œ ์ € i ์ด song j ๋ฅผ ๋“ค์—ˆ๋Š”์ง€ ์—ฌ๋ถ€ 

โ†ช  ProxTreatment_ij : ์œ ์ € i ์˜ ํŒ”๋กœ์šฐ ์นœ๊ตฌ ์ค‘์— song j ์— ๋Œ€ํ•ด ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€ (0์ด๋ฉด ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ํ•œ ๋ช…๋„ ์—†๋‹ค๋Š” ๊ฒƒ์„ ๋œปํ•จ) 

โ†ช  Friends_i : ์œ ์ € i ๊ฐ€ ํŒ”๋กœ์šฐํ•˜๊ณ  ์žˆ๋Š” ์ด ์œ ์ €์˜ ์ˆ˜ 

 

 

•  Jointly model DDD (User-song granularity level) 

โ†ช  Listen_gjt : group g (์ฃผ๋ณ€์— ํ•ด๋‹น ์Œ์•…์„ ์ข‹์•„ํ•˜๋Š” ์นœ๊ตฌ๊ฐ€ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ง‘๋‹จ์„ ๊ตฌ๋ถ„) ์— ๋Œ€ํ•ด t ์‹œ์ ์— song j ๋ฅผ ์ฒญ์ทจํ•œ ์ด ํšŸ์ˆ˜  ๐Ÿ‘‰  Popularity ์™€ Proximity ๋ฅผ ๋™์‹œ์— ๊ณ ๋ ค 

 

โ€ป  g : index of proximity treatment (0 or 1) 

โ‡จ  g = 1 ์ด๋ฉด ProxTreatment = 1, g=0 ์ด๋ฉด ProxTreatment = 0 

 

โ†ช  PopTreatment_j : song j ์— ๋Œ€ํ•œ popularity influence treated ์—ฌ๋ถ€ (์ข‹์•„์š” ์ˆ˜ ๋ณด์ด๋Š” ๊ธฐ๋Šฅ ์ ์šฉ์—ฌ๋ถ€)

โ†ช  After_t : 10์›” 1์ผ ์ดํ›„๋ฉด 1, ์•„๋‹ˆ๋ฉด 0 

โ†ช  ProxTreatment_gj : group g ์— ์†ํ•˜๋Š” ์œ ์ € i ๊ฐ€ song j ์— ๋Œ€ํ•ด ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์žˆ๋‹ค๋ฉด 1, ์•„๋‹ˆ๋ฉด 0

 

 

•  Control variable 

โ†ช  PreFavorites_j :  song j ์— ๋Œ€ํ•ด ์—ฐ๊ตฌ ๊ด€์ธก ์‹œ์  ์ด์ „์— ๋ฐ›์•˜๋˜ ์ด ์ข‹์•„์š” ๊ฐœ์ˆ˜ 

โ†ช  SalesRank_j : ์•„๋งˆ์กด์— ์˜ฌ๋ผ์˜จ song j ์˜ ์ด ํŒ๋งค ์ˆœ์œ„ 

โ†ช  Genre_j : song j ์˜ ์žฅ๋ฅด 

 

 

 

โ—ฏ  Model 

 

โ‘ด  Popularity influence 

 

โ–ข  CEM

 

•   Match the two groups of songs  (๋งค์นญ ๋‹จ์œ„ : Song, e.g. Song A - Song B) 

•   ๋งค์นญ ๊ธฐ์ค€ : ์žฅ๋ฅด, feature implementation ์ด์ „ ์‹œ์ ์˜ ์ด ์ข‹์•„์š” ๊ฐœ์ˆ˜, ์•„๋งˆ์กด ์Œ๋ฐ˜ ํŒ๋งค ์ˆœ์œ„ ๋“ฑ ๊ด€์ธก ๊ฐ€๋Šฅํ•œ ํŠน์„ฑ๋“ค์„ ๊ธฐ์ค€์œผ๋กœ ๋งค์นญ์„ ์ง„ํ–‰ 

•   ์žฅ๋ฅด๋Š” ์ •ํ™•ํžˆ ๋งค์นญ์‹œํ‚ค๋„๋ก ํ–ˆ๊ณ , ์—ฐ์†๋ณ€์ˆ˜์— ํ•ด๋‹นํ•˜๋Š” ๋ณ€์ˆ˜๋“ค์€ ์ตœ๋Œ€ํ•œ ๊ทผ์ ‘ํ•œ ๊ฐ’์„ ๊ฐ€์ง€๋„๋ก ๋งค์นญํ–ˆ๋‹ค. (Stata ํˆด ์‚ฌ์šฉ) 

•   ๋งค์นญ์„ ํ†ตํ•ด treatment group ๊ณผ control group ์— ์žˆ๋Š” imbalance ๊ฐ€ ๊ฐ์†Œํ•จ์„ ํ™•์ธ 

 

 

 

โ–ข   DiD setting 

 

•   Assumption : ์ข‹์•„์š” ์ˆ˜๋ฅผ ๋ณด์ด๊ฒŒ ํ•œ popularity influence ๊ฐ€, ์Œ์•… ์†Œ๋น„์— Positive ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์ด๋ผ๊ณ  ๊ฐ€์ • (ATE) 

 

 

 

 โ—‹   PopTreatment_j 

 

•   Treatment group : ์ข‹์•„์š” ์ˆ˜๋ฅผ "๋ณด์ผ ์ˆ˜ ์žˆ๊ฒŒ ํ•˜๋Š”" ๊ธฐ๋Šฅ์„ ๋„์ž…ํ•œ ์‹œ์ ์ด 10์›” 1์ผ์ด๊ธฐ ๋•Œ๋ฌธ์—, ์ข‹์•„์š” ์ˆ˜๊ฐ€ ์–ด๋Š์ •๋„ ์Œ“์ผ ์ˆ˜ ์žˆ๋„๋ก 2008๋…„ 9์›” 29์ผ์— ํฌ์ŠคํŒ…๋œ song ๋“ค์„ ๊ธฐ์ค€์œผ๋กœ data set ์„ ๊ตฌ์„ฑ 

•   Control group : ์ข‹์•„์š” ์ˆ˜๋ฅผ "๋ณด์ผ ์ˆ˜ ์žˆ๊ฒŒ ํ•˜๋Š”" ๊ธฐ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š” 2008๋…„ 9์›” 22์ผ์— ํฌ์ŠคํŒ…๋œ song ๋“ค์„ ๊ธฐ์ค€์œผ๋กœ set ์„ ๊ตฌ์„ฑ

 

 

 โ—‹  After_t 

 

•   Pre-treatment : T1 - 1 (for treated group) , T0 - 1 (for control group) 

•   Post-treatment : T1 + 1, T0 + 1

 

 

 

 

•   ์œ ์ €์˜ listening behavior ๊ฐ€ ๋ผ์น˜๋Š” ์˜ํ–ฅ์„ ๋ฐฐ์ œํ•˜๊ธฐ ์œ„ํ•ด Short estimation window ๋ฅผ ์„ค์ • (± 1 day)

•   DiD ์ ์šฉ ๊ธฐ์ค€ ์‹œ์ ์ด ๋‹ค๋ฅธ ๊ฒƒ์ด ํŠน์ง• : T0, T1  ๐Ÿ‘‰  Confounding ์„ ๋ฐœ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ์š”์†Œ 

 

 

 

 

โ–ข   DiD Time separation ์„ ํ•œ ๊ฒƒ์ด ํฐ ์šฐ๋ ค์ง€์ ์ด ์•„๋‹Œ ์ด์œ  โญ

 

1) Pre-treatment ๊ธฐ๊ฐ„ ๋™์•ˆ ๊ฐ group ์—์„œ ๊ฐ ์‹œ๊ฐ„๋Œ€ ๋ณ„ ์ด ์ฒญ์ทจ ํšŸ์ˆ˜๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๋น„์Šทํ•œ ํŒจํ„ด์„ ๋ณด์ž„

 

 

•  2) ๋…ธ๋ž˜๊ฐ€ Posting ๋˜๋Š” time ์€ exogeneous ํ•˜๋‹ค. The hype machine ์— ์˜ํ•ด์„œ ๊ฒฐ์ •๋˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ Original MP3 blog (์—ฌ๋Ÿฌ ์Œ์•…๊ณผ ๊ด€๋ จ๋œ ๋ธ”๋กœ๊ทธ๋“ค) ๋“ค์„ ํ†ตํ•ด posting ๋œ๋‹ค. 

 

•  3) ์žฅ๋ฅด๋‚˜ ์ธ๊ธฐ๋„ ๊ด€์ ์—์„œ ๊ฐ ๊ทธ๋ฃน์˜ song ๋“ค์˜ ํŠน์„ฑ์ด ๋น„์Šทํ•˜๋‹ค. ๋˜ํ•œ ๋ถ„์„์—์„œ sample ๋“ค ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด CEM ๋งค์นญ ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•œ๋‹ค. 

 

•  4)  DiD ์˜ ํ•„์ˆ˜์ ์ธ ๊ฐ€์ •์ธ Common trend ๋ฅผ ๋งŒ์กฑํ•˜๊ณ  ์žˆ๋‹ค. 

 

 

 

 

| Treatment and control groups have a common trend in the absence of treatment =assumption of a common pretreatment trend for the treatment and control samples

 

์Œ์•…์ฒญ์ทจ ํšŸ์ˆ˜๊ฐ€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ๊ฐ์†Œํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์ด๋Š”๋ฐ, ๋…ธ๋ž˜๋งˆ๋‹ค ํฌ์ŠคํŒ… ์‹œ๊ฐ„์ด ๋‹ค๋ฅด๋ฏ€๋กœ, ์ฒ˜์น˜์ง‘๋‹จ๊ณผ ํ†ต์ œ์ง‘๋‹จ ๊ฐ„ pretreatment trend ๋ฅผ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด, posting ์ดํ›„์˜ ์ฒ˜์Œ 12์‹œ๊ฐ„ window ์™€ ๋‘๋ฒˆ์งธ 12์‹œ๊ฐ„ window ๊ฐ๊ฐ์˜ ํ‰๊ท ์ ์ธ ์ฒญ์ทจ ์ˆ˜์˜ ์ฐจ์ด (Difference of Listens) ๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค. (์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ฒญ์ทจ ์ˆ˜๋Š” ๊ฐ์†Œํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์ด๋ฏ€๋กœ Difference of Listens ๊ฐ€ ๋Œ€๋ถ€๋ถ„ ์Œ์ˆ˜๊ฐ’์„ ๊ฐ€์ง) ์ด๋•Œ ๋‘ ๊ทธ๋ฃน ๊ฐ„ ๋ถ„ํฌ๊ฐ€ ์œ ์‚ฌํ•˜๊ณ , t-test ๊ฒฐ๊ณผ๋„ mean difference ๊ฐ€ insignificant ํ•œ ๊ฒฐ๋ก ์„ ๋‚ด๋ ธ๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ€์ •์„ ๋งŒ์กฑํ•œ๋‹ค๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

 

 

•  5)  Robustness check : ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ๊ธฐ์ค€ ์‹œ์  (9/29,22 ๋ง๊ณ ) ์„ ๋‹ฌ๋ฆฌ ํ•˜๋ฉด์„œ ๋ถ„์„์„ ์ง„ํ–‰ํ•จ → Not sensitive to exactly when the songs are posted 

 

 

 

โ–ข  Model specification 

 

 

โ†ช  song j on day t : ๋ถ„์„ ๋‹จ์œ„๋Š” Song level 

โ†ช  t = {T0-1, T0+1, T1-1, T1+1} 

โ†ช  PopTreatment_j : 1์ด๋ฉด treated group, 0์ด๋ฉด control group 

โ†ช  After_t : 1์ด๋ฉด post-treatment, 0์ด๋ฉด pre-treatment period 

โ†ช  β4 : PopTreatment_j X After_t : magnitude of popularity influence โญ

 

 

 

 

 

โ‘ต  Proximity influence 

 

โ–ข  ๊ธฐ๋ณธ ๊ฐ€์ •

 

•   Assumption: ์ฃผ๋ณ€ ์นœ๊ตฌ๊ฐ€ ํŠน์ • song ์— ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ํ–‰๋™์ด, ์œ ์ €๊ฐ€ ํ•ด๋‹น song ์„ ๋“ฃ๋Š” ๊ฒƒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ : How favoriting a song by a focal user impacts the listening behavior of her social ties 

 

•  User-song ๋‹จ์œ„ ๋ถ„์„ 

 

•  Treatment group : ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ์ข‹์•„์š” ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์ ์–ด๋„ 1๋ช… ์ด์ƒ ์žˆ๋Š” ์œ ์ €๋“ค๋กœ ๊ตฌ์„ฑ

•  Control group : ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ์ข‹์•„์š” ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ๋‹จ ํ•œ๋ช…๋„ ์—†๋Š” ์œ ์ €๋“ค๋กœ ๊ตฌ์„ฑ 

 

 

 

 

โ–ข  Propensity score matching 

 

•  Control for potential homophily 

 

•  ๋งค์นญ ๋‹จ์œ„ : User ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋งค์นญ (e.g. User A - User B) 

 

•  ๋งค์นญ ๋ชฉํ‘œ : Control group ์˜ ์œ ์ €๋“ค์„ ์Œ์•…์ทจํ–ฅ, ์œ ์ €์— ๋Œ€ํ•œ ๊ด€์ธก ๊ฐ€๋Šฅํ•œ ํŠน์ง•, ์นœ๊ตฌ ์ˆ˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ Treatment group ์—์„œ ๊ฐ€์žฅ ์œ ์‚ฌํ•œ ์œ ์ €์™€ ๋งค์นญ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ → ์œ ์ €์— ๋Œ€ํ•œ demographic ํ•œ ์ •๋ณด๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— ์ถ”์ฒœ ์‹œ์Šคํ…œ๊ณผ ๋น„์Šทํ•˜๊ฒŒ song listening behavior ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋งค์นญ์„ ์ง„ํ–‰ํ•จ 

 

•  ๋งค์นญ ๊ธฐ์ค€ : Relative song listening profiles of users (taste

โ†ช  ์œ ์ €์˜ ์ฒญ์ทจ ๊ธฐ๋ก ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ, 3์ฃผ ๊ธฐ๊ฐ„๋™์•ˆ ๊ฐ ์œ ์ €์— ๋Œ€ํ•ด profile ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ 

โ†ช  ์•ฝ 8๋งŒ๊ฐœ์˜ ๋…ธ๋ž˜์— ๋Œ€ํ•ด ์žฅ๋ฅด, ์•„ํ‹ฐ์ŠคํŠธ ์ธ๊ธฐ๋„ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ๋“ค์„ ์ˆ˜์ง‘ํ•ด ๋…ธ๋ž˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๋ณด์™„ํ•œ๋‹ค. 

โ†ช  ๊ฐ User-Song pair ์— ๋Œ€ํ•ด, ์œ ์ €๊ฐ€ ํŠน์ • ๋…ธ๋ž˜๋ฅผ ๋“ค์€ ํšŸ์ˆ˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ ํ•˜์—ฌ ํ•ด๋‹น ์œ ์ €์— ๋Œ€ํ•œ ์ „์ฒด ์ฒญ์ทจ ํšŸ์ˆ˜์˜ ๋ฐฑ๋ถ„์œจ๋กœ Weight ๋ฅผ ๊ณ„์‚ฐํ•˜์˜€๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์œ ์ €๊ฐ€ ์ฒญ์ทจํ•œ ๋ชจ๋“  ๋…ธ๋ž˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ  28๊ฐœ์˜ ๊ฐ€์ค‘ํ‰๊ท ๋œ ๋…ธ๋ž˜ ํŠน์„ฑ์„ ๋ฐ˜์˜ํ•œ ํ•˜๋‚˜์˜ ๋ฒกํ„ฐ๋กœ ๋งŒ๋“ค์—ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด ์œ ์ €์˜ Profile ์„ ์ผ๋ จ์˜ ์ˆซ์ž (๋…ธ๋ž˜ ํŠน์„ฑ์˜ ๊ฐ€์ค‘ ํ‰๊ท )๋กœ ์š”์•ฝํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ ์ˆซ์ž๋Š” ํŠน์ • ์Œ์•… ํŠน์„ฑ์— ๋Œ€ํ•œ ์‚ฌ์šฉ์ž์˜ ์ทจํ–ฅ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. 

โ†ช  ์ƒ์„ฑ๋œ Profile ์€ PSM ์—์„œ ๋‘ ๊ทธ๋ฃน์˜ ์œ ์ €๋ฅผ ๋งค์นญ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋œ๋‹ค. 

 

โ†ช  ๋‘ ๊ทธ๋ฃน์— ์†ํ•œ ๊ฐ Song ์— ๋Œ€ํ•ด์„œ, ์œ ์ €์˜ ์ฒญ์ทจ๊ธฐ๋ก, ์นœ๊ตฌ ์ˆ˜, ํŠน์„ฑ์„ ๋ฐ˜์˜ํ•œ logit model ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Treatment group ์— ์†ํ•  ํ™•๋ฅ ์„ ํ• ๋‹นํ•œ๋‹ค. 

 

•  Propensity score 0.1 ์ด๋‚ด์— ์žˆ๋Š” ๊ด€์ธก์น˜๋“ค์„ ๊ธฐ์ค€์œผ๋กœ ๋งค์นญ (Caliper = 0.1) 

•  PSM ์ดํ›„์— ๊ฐ ๊ทธ๋ฃน์˜ Propensity score ๋ถ„ํฌ๋ฅผ ๋‚˜ํƒ€๋‚ด ์œ ์‚ฌ๋„๋ฅผ ํ™•์ธ 

 

๋งค์นญ ๊ณผ์ •์ด ์ž˜ ์ด๋ฃจ์–ด์ง!

 

 

 

โ€ป  ๋ณด์ถฉ ์„ค๋ช… 

 

•  Ti : ์ฒ˜์น˜์ง‘๋‹จ์— ์žˆ๋Š” User :  Song j ๋ฅผ ๋“ค์–ด๋ณธ ์ ์€ ์—†์ง€๋งŒ ํŒ”๋กœ์ž‰ํ•œ ์นœ๊ตฌ๊ฐ€ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ๊ฒฝ์šฐ์— ์†ํ•œ ์ง‘๋‹จ์˜ User

•  Ci : ํ†ต์ œ์ง‘๋‹จ์— ์žˆ๋Š” User : Ti ์˜ ์œ ์ €์™€ ๋น„์Šทํ•œ taste ๋ฅผ ๊ฐ€์ง€๊ณ  (based on listen profile), song j ๋ฅผ ๋“ค์–ด๋ณธ ์ ์ด ์—†์œผ๋ฉฐ ํ•ด๋‹น ๋…ธ๋ž˜์— ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ํ•œ ๋ช…๋„ ์—†๋Š” ๊ฒฝ์šฐ์— ์†ํ•œ ์ง‘๋‹จ์˜ user 

 

0. 3์ฃผ ๊ธฐ๊ฐ„๋™์•ˆ์˜ ์œ ์ €์˜ listening behavior ๋ฅผ profile ํ•จ (28๊ฐœ์˜ ์Œ์•… ํŠน์„ฑ์„ ํฌํ•จํ•˜๋Š” vector ์™€ ํŒ”๋กœ์ž‰ ์ˆ˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ) 

1. Treatment group (T) ์„ ์ •์˜ (Ti ์„ค๋ช… ์ฐธ๊ณ ) 

2. Potential control group PC ๋ฅผ ์ •์˜ : T ์— ์†ํ•˜์ง€ ์•Š๋Š” user 

3. Control group (C) ๋ฅผ ์ •์˜ํ•œ๋‹ค. ์œ ์ €์˜ taste profile ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ, treated ๋ฐ›์„ propensity ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” logit model ์„ ํ™œ์šฉํ•ด,  ์ฃผ์–ด์ง„ song ์„ ์ฒญ์ทจํ•  Propensity ๋ฅผ ๋งค์นญ (Matching the propensity of listening to any given song)   ํ•œ๋‹ค. T์— ์†ํ•œ ๊ฐ user Ti ๋ฅผ PC ์— ์†ํ•œ PCi ์œ ์ €์™€ ๋งค์นญํ•˜๊ณ , ๋งค์นญ๋œ ์œ ์ €๋“ค์„ group C ๋กœ ์ •์˜ํ•œ๋‹ค. 

4. T์™€ C ์—์„œ user-song observation ๋“ค์„ ๋ณต๊ตฌ (Recover) ํ•œ๋‹ค. ๊ฐ song j ์— ๋Œ€ํ•ด ๋งŒ์•ฝ user Ti ๊ฐ€ song j ๋ฅผ ์ข‹์•„์š” ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์žˆ๋‹ค๋ฉด user-song j pair ๋ฅผ ์žฌ๊ตฌ์„ฑํ•œ๋‹ค. ์ดํ›„ song j ์— ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์—†๋Š” user Ci ์™€ ๋งค์นญ์„ ์ง„ํ–‰ํ•œ๋‹ค. 

 

 

 

 

 

โ–ข  Euclidean Distance Matching 

 

•  Robustness check ์„ ์œ„ํ•ด ์‚ฌ์šฉ 

 

•  ๋งค์นญ ๋‹จ์œ„ : User-song level → ๊ฐ song ์— ๋Œ€ํ•ด > ๊ฐ user ๋ฅผ ๋งค์นญ : ์œ ์ €๊ฐ€ ์ฃผ์–ด์ง„ ๋…ธ๋ž˜๋ฅผ ์ฒญ์ทจํ•  likelihood ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ์ข‹์•„ํ•˜๋Š” ์นœ๊ตฌ๋ฅผ ๊ฐ€์งˆ likelihood (the likelihood of being treated) ๊นŒ์ง€ ๊ณ ๋ คํ•˜์—ฌ ๋งค์นญ์„ ์ง„ํ–‰ํ•œ๋‹ค. 

 

•  ๋งค์นญ ๊ธฐ์ค€ : ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ๋“ค์–ด๋ณธ์ ์€ ์—†์œผ๋‚˜ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์žˆ๋Š” ์œ ์ € Ti ์™€, (Ti์™€ ๋น„์Šทํ•œ ๋…ธ๋ž˜ ์ทจํ–ฅ์„ ๊ฐ€์ง€๊ณ , ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ๋“ค์–ด๋ณธ ์ ์ด ์—†๊ณ , ๊ทธ๋Ÿฌ๋‚˜ ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ์ข‹์•„ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ (์•„์ง ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ๊ฒƒ์€ ์•„๋‹˜) ์นœ๊ตฌ๋ฅผ ๊ฐ€์ง„) ์œ ์ € Ci ์™€ ๋งค์นญ์„ ์ง„ํ–‰ํ•œ๋‹ค. 238 ๊ฐœ์˜ treatment user-song pair ๊ฐ€ ์žˆ๋‹ค. ์ด์ฒ˜๋Ÿผ user-song level ๋กœ ๋ถ„์„ํ•˜๋ฉด treated observation ๊ฐœ์ˆ˜๊ฐ€ ๋งค์šฐ ์ ์–ด์ง€๋Š”๋ฐ, ์ด๋Ÿฐ ๊ฒฝ์šฐ ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€๋ถ„์„์„ ์ ์šฉํ•˜๊ฒŒ ๋˜๋ฉด ๋‹ค๋ฃจ๊ธฐ ์–ด๋ ค์›Œ์ง„๋‹ค. ์ด๋ฅผ ์™„ํ™”์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋งค์นญ ๊ณผ์ •์—์„œ ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. 

 

 

โ€ป  ๋ณด์ถฉ ์„ค๋ช… 

 

0.  PSM ๊ณผ์ •๊ณผ ๋™์ผํ•˜๊ฒŒ user ์˜ ์ฒญ์ทจํ–‰๋™์— ๋Œ€ํ•œ Profile vector ๋ฅผ ์ƒ์„ฑํ•˜๊ณ 

1. ๊ฐ song j ์— ๋Œ€ํ•ด 

   โˆ˜  1-(a). Tj (treatment group) ์„ ๊ฒฐ์ •ํ•œ๋‹ค : song j ๋ฅผ ๋“ค์–ด๋ณธ์ ์€ ์—†์ง€๋งŒ ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๊ฐ€ ์žˆ๋Š” ์œ ์ €

   โˆ˜  1-(b). PCj (potential control group) ์„ ๊ฒฐ์ •ํ•œ๋‹ค : song j ๋ฅผ ๋“ค์–ด๋ณธ์ ์ด ์—†๊ณ , ํ•ด๋‹น ๋…ธ๋ž˜์— ์ข‹์•„์š”๋ฅผ ๋ˆ„๋ฅธ ์นœ๊ตฌ๋„ ์—†๋Š” ์œ ์ € 

   โˆ˜  1-(c). 0๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ๊ตฌํ•œ ๋ฒกํ„ฐ ๊ฐ’์„ ๊ธฐ๋ฐ˜์œผ๋กœ Tj ์™€ PCj ์— ์žˆ๋Š” ์œ ์ €๋“ค ๊ฐ„์˜ ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. Treatment group ์— ์†ํ•œ ๊ฐ ์œ ์ €์— ๋Œ€ํ•ด, ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๊ฐ€ ๊ฐ€์žฅ ์งง์€ PC group ์— ์†ํ•œ 3๋ช…์˜ candidate ์œ ์ €๋“ค์„ ์„ ํƒํ•œ๋‹ค. 

   โˆ˜  1-(d). Cj ๋ฅผ ๊ฒฐ์ •ํ•œ๋‹ค : 1-c ์—์„œ ์ •ํ•œ candidate user ๋“ค์˜ ๊ฐ profile ๊ณผ song j ์˜ profile ์˜ ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๋ฅผ ๋น„๊ตํ•˜์—ฌ ๊ฐ€์žฅ ์งง์€ ๊ฑฐ๋ฆฌ๋ฅผ ๊ฐ€์ง€๋Š” user ๋ฅผ ์„ ํƒํ•œ๋‹ค. 

   โˆ˜  1-(e). Recover user-song j pair for users in Tj, Cj 

 

 

 

 

 

โ–ข  Probit model 

 

 

 

•  Examine the impact of proximity influence on the likelihood of listening to a song → implement a binary probit model (ํ•ด๋‹น ๋…ธ๋ž˜๋ฅผ ๋“ฃ๊ฒŒ ๋  ํ™•๋ฅ ์„ ์˜ˆ์ธก) 

•  9์›” 22์ผ์— posting ๋œ ๋…ธ๋ž˜๋ฅผ ๋Œ€์ƒ์œผ๋กœ, ๋…ธ๋ž˜๊ฐ€ posting ๋œ ์ดํ›„ 48์‹œ๊ฐ„ ๋™์•ˆ burn-in period ๋ฅผ allow ํ•˜์—ฌ favorites ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๊ฒŒ ํ•จ. burn-in ๊ธฐ๊ฐ„ ์ดํ›„์— ์œ ์ €์˜ listening choice ๋ฅผ 7์ผ๊ฐ„ following ํ•จ

•  β1 : coefficient of interest on a focal user’s listen decision, captures the impact of proximity influence on a focal user’s listen decision

 

 

 

 

โ–ข  Hazard model (Weibull model

 

 

•   proximity influence by looking at the time to a user’s first listen to a song

•   Hazard model ์„ ์‚ฌ์šฉํ•˜์—ฌ user i ๊ฐ€ song j ๋ฅผ ์ฒ˜์Œ์œผ๋กœ ๋“ฃ๊ธฐ ์ „ ๊ธฐ๊ฐ„์— proximity influence ๊ฐ€ ๋ผ์นœ ์˜ํ–ฅ์— ๋Œ€ํ•ด ์ถ”์ •

•   hazard rate : λ_ij → defined by whether and when user i listened to song j

•   β1 : coefficient of interest

 

 

 

โ‘ถ  Combined 

 

 

•   ๋‘ ์˜ํ–ฅ์„ ๋™์‹œ์— ๊ณ ๋ คํ•œ DDD ๋ชจํ˜• : popularity treatment  + After_t + proximity influence treatment

•   Song j ๋Š” DD ๋ถ„์„๊ณผ ๋™์ผํ•˜๊ฒŒ data set ์„ ์„ค์ • 

•   ์ดํ›„ PSM ์„ ํ†ตํ•ด ๋‘ Proximity group ์— ์†ํ•œ ์œ ์ €๋“ค์„ ๋งค์นญํ•จ 

 

•   Listen_gjt : total number of listens of proximity treatment type g of song j at time t , where t = {T0-1, T0+1, T1-1, T1+1}  โ‚ song j ์˜ ์ด ์ฒญ์ทจํšŸ์ˆ˜ 

•   After_t : pretreatment or posttreatment for popularity

 

•   β3 : represents the magnitude of proximity influence

•   β5 : represents the magnitude of popularity impact

•   β7 : three-way interaction term : nature of interaction between popularity and proximity influence (β7 > 0 : complementary, β7 < 0 : substitute)

 

 

 

โ‘ค  Results 


 

โ—ฏ  Popularity influence

 

PopTreatmentj X Aftert ๊ณ„์ˆ˜ ์ถ”์ •์— ์ฃผ๋ชฉ

 

 

 

โ–ข  ํ•ด์„ 

 

•  (Matching ํ›„) N ์— ์ฃผ๋ชฉํ•˜๊ธฐ 

•  PopTreatment x After : captures the average effect of the treatment on the number of listens after the availability of song popularity information on the website (๋…ธ๋ž˜ ์ธ๊ธฐ ์ •๋ณด ๊ณต๊ฐœ์™€ ์‚ฌ์šฉ์ž ์ฒญ์ทจ ํšŸ์ˆ˜ ์‚ฌ์ด์˜ ์ธ๊ณผ๊ด€๊ณ„)

•  After feature implementation, increases the total listens of average song by 19.6% (exp(0.18)) in model2

•  magnitude and significance are highest in model 3 and 4 = Popularity influence is strongest for isolates (users with no friends) and for the first listen of a song (↔ repeat listens)

 

 

 

โ–ข  Robustness check 

 

 

•   Validating DD research design with the treatment and control samples drawn from neighboring weeks 

•   ํŠน์ •ํ•œ ์š”์ผ์„ ์„ ํƒํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‚˜ํƒ€๋‚œ ๊ฒฐ๊ณผ๊ฐ€ ์•„๋‹˜์„ (for generalize) ๋ฐํžˆ๊ธฐ ์œ„ํ•ด ๋‹ค๋ฅธ ๋Œ€์•ˆ์ ์ธ ์š”์ธ๋“ค์„ ์„ ํƒํ•ด์„œ ๋ชจ๋ธ๋ง์„ ์ง„ํ–‰ ⇒ Robust and there no “secular” effects in different weeks 

 

 

 

โ–ข   Narrow-appeal music VS broad appeal music

 

 

•  ์•„๋งˆ์กด ์Œ๋ฐ˜ ํŒ๋งค ์ˆœ์œ„ ๋ฐ ์ฃผ๋ฅ˜/๋น„์ฃผ๋ฅ˜ ์žฅ๋ฅด๋ฅผ ๊ธฐ์ค€์œผ๋กœ Song group ์„ ๊ตฌ๋ถ„ 

•  ์„ธ ๋ชจ๋ธ์—์„œ ๋ชจ๋‘ interaction term ์˜ ๊ณ„์ˆ˜์—์„œ positive in sign, but significant only for narrow appeal song samples

 

 

 

 

โ—ฏ  Proximity influence

 

 

 

•  ProxTreatment → probit, hazard ๋ชจ๋ธ์—์„œ ๋ชจ๋‘ positive & significant ํ•˜๊ณ , random matching ๋ณด๋‹ค PSM or EDM ์ผ ๋•Œ magnitude of coefficient ๊ฐ€ goes down ํ•จ (to be able to isolate it from homophily)

 

 

 

 

โ—ฏ  Combined model of Popularity and Proximity influence

 

•  user-song granularity : Listen_gjt ⇒ g : user proximity group, j : song, time t 

 

๐Ÿค” ๋…ผ๋ฌธ์—์„œ dependent variable ์„ log(Listen_git) ๋กœ ์„ค์ •ํ•˜๊ณ , Treatment ์™€ Control group ์„ PSM ๋งค์นญํ•˜๋Š” ๊ณผ์ •์— ๋Œ€ํ•ด ์ถ”๊ฐ€์ ์ธ ์„ค๋ช…์ด ์—†์Œ (Supplement ์ž๋ฃŒ๋„ ์—†์Œ) 

 

 

•  Model1. Populairty influence 

โ†ช  PopTreatment x After

 

โˆ˜  Full sample ์ผ ๋• not significant โ‡จ  due to sparseness of total listens at the user-song granularity : User-song ๋‹จ์œ„๋กœ matching ๋œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์ง„ํ–‰๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ (N=160) โ‡จ g index (๊ธฐ๊ฐ„ T0, T1 ์— ์†ํ•˜๋Š” Song j ์— ๋Œ€ํ•ด, Proximity ์กฐ๊ฑด์— ๋งž๋Š” user ๋“ค์„ ์ถ”๋ฆฌ๋‹ค ๋ณด๋‹ˆ N์˜ ๊ฐœ์ˆ˜๊ฐ€ ๋งค์šฐ ์ค„์–ด๋“ฆ)

 

โˆ˜  subsample analysis ์—์„œ ProxTreatment = 0 ์ผ ๋•Œ๋งŒ (=user does not have a friend who has previously favorited the song) PSM ์—์„œ ๊ณ„์ˆ˜๊ฐ€ significant ํ•จ ⇒ substitutes ๊ด€๊ณ„ 

 

 

•  Model2. Proximity influence 

โ†ช  PopTreatment_gj 

 

โˆ˜  full model → positive and significant & magnitude declines under PSM

โˆ˜  subsample analysis ์—์„œ PopTreatment ๊ฐ€ 0์ผ ๋•Œ (absence ํ•  ๋•Œ) ProxTreatment ๋ณ€์ˆ˜๊ฐ€ greater sign and significance ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค ⇒ substitutes ๊ด€๊ณ„ 

 

 

•  Model3. Popularity & Proximity influence 

โ†ช  PopTreatment_j x After_t x ProxTreatment_gj 

 

โˆ˜  coefficient is negative and significant → popularity and proximity influence are substitutes

โˆ˜  specifically, popularity influence is less important in the presence of proximity influence

 

 

 

 

 

 

โ‘ฅ  Conclusion 


 

 

 

 

 

728x90

๋Œ“๊ธ€