๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

1๏ธโƒฃ AI•DS/๐ŸฅŽ Casual inference47

์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ํ†ต์ œํ•จ์ˆ˜์™€ ์„ ํƒ๋ชจํ˜• ์ฐธ๊ณ ์˜์ƒ : Bootcamp 4-4. ํ†ต์ œํ•จ์ˆ˜์™€ ์„ ํƒ๋ชจํ˜• โ–ธ Control function: selection bias correction method , LATE, 2SLS ์™€ ๋น„์Šทํ•˜๊ฒŒ Instrumental variable ์„ ํ™œ์šฉํ•œ๋‹ค. โ–ธ Heckman selection model : control function ์˜ special case 1. Causal inference = How to address Endogeneity โ—ฏ Causal inference • Selection model ๊ณผ Causal Graph ๋Š” ์ธ๊ณผ์ถ”๋ก ์˜ Researcg design ๊ณผ ๋‹ค๋ฅธ ๊ด€์ ์„ ์ทจํ•˜๊ณ  ์žˆ๋‹ค. • ๊ทธ๋Ÿฌ๋‚˜ LATE ์™€ selection model ์€ ๋ชจ๋‘ IV ๋ฅผ ํ™œ์šฉํ•œ๋‹ค. โ—ฏ Second approac.. 2023. 4. 26.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ํšŒ๊ท€๋ถˆ์—ฐ์† ์ฐธ๊ณ ์˜์ƒ : Bootcamp 4-3. ํšŒ๊ท€ ๋ถˆ์—ฐ์† 1. RD โ—ฏ Regression Discontinuity • Discontinuous ๊ฐ€ ๋ฐœ์ƒํ•˜๋ฉด ๊ทธ๊ฒƒ์„ ๊ธฐ์ ์œผ๋กœ ์ธ๊ณผ์ถ”๋ก ์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ• • Running variable = assignment variable = Forcing variable : Discontinuity ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๋ณ€์ˆ˜ • RD ์—์„œ์˜ counterfactual : running variable ์ด ์—†์—ˆ์„ ๋•Œ๋ฅผ ๊ฐ€์ •ํ•œ ์ถ”์ด (์ ์„ ) • counterfactual (์ ์„ )๊ณผ treatment ๋ฅผ ๋ฐ›์•„์„œ ๋‚˜์˜จ (์‹ค์„ ) ์ฐจ์ด๊ฐ€ causal effect ์ด๋‹ค. โ—ฏ Example of discontinuity • ์Œ์ฃผ์™€ ๊ฑด๊ฐ•/์‚ฌ๋ง ์‚ฌ์ด์˜ ์ธ๊ณผํšจ๊ณผ • ๋ฏธ๊ตญ์—์„œ๋Š” ๋ฒ•์ ์œผ๋กœ 21์„ธ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์Œ.. 2023. 4. 26.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ์ธ๊ณผ์ถ”๋ก  ๊ด€์ ์—์„œ์˜ ๋„๊ตฌ๋ณ€์ˆ˜ ์ฐธ๊ณ ์˜์ƒ : Bootcamp 4-2. Local Average treatment effect (LATE) 1. IV from perspective of potential outcome โ—ฏ LATE • IV ๊ฐ€ casual effect ์—์„œ ์–ด๋–ค ๋ถ€๋ถ„์„ ์ธก์ •ํ•˜๋Š” ๊ฒƒ์ธ์ง€ (์ธ๊ณผ๊ด€๊ณ„์—์„œ ์–ด๋–ค ํ•ด์„์„ ๊ฐ–๋Š” ๊ฒƒ์ธ์ง€) ์ง๊ด€์ ์ธ ์ดํ•ด๊ฐ€ ์กฐ๊ธˆ ์–ด๋ ต๋‹ค. • ๋„๊ตฌ๋ณ€์ˆ˜ ๋ถ„์„์„ potential outcome framework ์— ํ†ตํ•ฉํ•˜๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด LATE • LATE ๋ฅผ ๊ฐ€์ง€๊ณ , ๋„๊ตฌ๋ณ€์ˆ˜ ๋ถ„์„์„ ํ†ตํ•ด์„œ ์šฐ๋ฆฌ๊ฐ€ ์ถ”์ •ํ•˜๋Š” causal effect ๊ฐ€ ์–ด๋–ค ๊ฒƒ์ธ์ง€ ๋ถ„๋ช…ํ•˜๊ฒŒ ํ•ด์„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค. 2. IV as a treatment assignment mechanism • Research design : ์–ด๋–ป๊ฒŒ treatm.. 2023. 4. 25.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ๋„๊ตฌ๋ณ€์ˆ˜ ์ฐธ๊ณ ์˜์ƒ : Bootcamp 4-1. Instrumental variable and regression discontinuity 1. ๋„๊ตฌ๋ณ€์ˆ˜ โ—ฏ Causal Hierarchy • quasi-experiment design ์ด ๋ถˆ๊ฐ€๋Šฅํ•˜์ง€๋งŒ treatment๋ฅผ ํ•  ์ˆ˜ ์žˆ๋Š” ์™ธ์ƒ๋ณ€์ˆ˜๊ฐ€ ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ โ‡จ instrumental variable ๋ฅผ ์‚ฌ์šฉ • quasi-experiment design ์ด ๊ฐ€๋Šฅํ•˜๊ณ , ์ฒ˜์น˜์ง‘๋‹จ๊ณผ ํ†ต์ œ์ง‘๋‹จ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, longitudinal data (=panel data)๋ฅผ ๊ด€์ธกํ•  ์ˆ˜ ์—†์–ด (์ฆ‰, panel data ๊ฐ€ ์•„๋‹˜) treatment ๊ฐ€ arbitrary ํ•œ threshold๋กœ ์ ์šฉ๋˜๋Š” ๊ฒฝ์šฐ โ‡จ Regression discontinuity์„ ์‚ฌ์šฉ โ—ฏ Endogen.. 2023. 4. 25.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ๊ฐ€์ƒ์˜ ํ†ต์ œ์ง‘๋‹จ ์ฐธ๊ณ ์˜์ƒ : Bootcamp 3-4. ๊ฐ€์ƒ์˜ ํ†ต์ œ์ง‘๋‹จ 1. synthetic control vs DID โ—ฏ Synthetic control • synthetic control ์€ DID ์˜ ํ™•์žฅ๋ฒ„์ „์ด๋‚˜ ์ข€ ๋” ์œ ์—ฐํ•œ ๋ฐฉ๋ฒ•์ด๋‹ค. • ๋งค์นญ์ด ์„ฑ๋ฆฝ๋˜์ง€ ์•Š๊ณ , parallel trend ๊ฐ€์ •์ด ์„ฑ๋ฆฝํ•˜์ง€ ์•Š๋”๋ผ๋„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค. • ์ตœ๊ทผ ๊ฐ€์žฅ ์ฃผ๋ชฉ๋ฐ›๋Š” ๋ฐฉ๋ฒ•๋ก ์ด๊ธฐ๋„ ํ•˜๋‹ค. • control group ์„ ์กฐํ•ฉํ•จ์œผ๋กœ์„œ ๊ฐ€์ƒ์˜ ๋น„๊ต๊ฐ€๋Šฅํ•œ ํ†ต์ œ์ง‘๋‹จ์„ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. 2. Example โ—ฏ ์บ˜๋ฆฌํฌ๋‹ˆ์•„์˜ ๋‹ด๋ฐฐ ๊ทœ์ œ์˜ ๋‹ด๋ฐฐ ํŒ๋งค๋Ÿ‰์— ๋ฏธ์นœ ํšจ๊ณผ • ์บ˜๋ฆฌํฌ๋‹ˆ์•„์—์„œ๋งŒ 1988๋…„์— ๋„์ž…๋จ, ๊ทœ์ œ๊ฐ€ ๋„์ž…๋˜์ง€ ์•Š์€ 49๊ฐœ์˜ ๋‹ค๋ฅธ ์ฃผ์™€ ๋น„๊ตํ•˜๊ณ ์ž ํ•˜์ง€๋งŒ, ์œ„์˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™์ด parallel trend ๋ฅผ ๋”ฐ๋ฅด์ง€ ์•Š์Œ โ‡จ syn.. 2023. 4. 25.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ์ด์ค‘์ฐจ๋ถ„๋ฒ• ์ฐธ๊ณ ์˜์ƒ : Bootcamp 3-3. ์ด์ค‘์ฐจ๋ถ„๋ฒ• DID(Difference-in-Differences)๋Š” quasi-experimental design ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. Quasi-experimental design์€ randomized controlled trial(RCT)๊ณผ ๊ฐ™์€ ์—„๊ฒฉํ•œ ์‹คํ—˜์  ์„ค๊ณ„๊ฐ€ ์ ์šฉ๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ๋ฅผ ๋งํ•ฉ๋‹ˆ๋‹ค. DID๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‹ค์ œ ์‹คํ—˜์  ์กฐ์ž‘์„ ๊ฐ€ํ•˜์ง€ ์•Š๋Š” ์ž์—ฐ์ ์ธ ์‚ฌ๊ฑด, ์˜ˆ๋ฅผ ๋“ค์–ด ์ •์ฑ… ๋ณ€ํ™”, ์ž์—ฐ ์žฌํ•ด ๋“ฑ์ด ๋ฐœ์ƒํ•œ ๊ฒฝ์šฐ๋ฅผ ์ด์šฉํ•˜์—ฌ ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. DID๋Š” ๋‘ ๊ฐœ ์ด์ƒ์˜ ๊ทธ๋ฃน์„ ๋น„๊ตํ•˜์—ฌ ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. DID๋Š” ์‹œ๊ฐ„์ ์œผ๋กœ ์„œ๋กœ ๋‹ค๋ฅธ ๋‘ ๊ทธ๋ฃน์ด ์žˆ๋Š” ๊ฒฝ์šฐ์— ์ฃผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ทธ๋ฃน๊ณผ ๋Œ€์กฐ ๊ทธ๋ฃน์˜ ํŠน์„ฑ์ด ์„œ๋กœ ๋‹ค๋ฅด์ง€ ์•Š์€ ๊ฒƒ์ด ์ „์ œ๋˜์–ด์•ผ ํ•˜๋ฉฐ, ์ด๋ฅผ ์œ„ํ•ด DI.. 2023. 4. 24.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ์ค€์‹คํ—˜ ๋ถ„์„ ๋ฐฉ๋ฒ•๋ก  ์ฐธ๊ณ ์˜์ƒ : Bootcamp 3-2. ์ค€์‹คํ—˜ ๋ถ„์„ ๋ฐฉ๋ฒ•๋ก  1. Counterfactual and methods โ—ฏ Counterfactual revisited • ์šฐ๋ฆฌ์—๊ฒŒ ํ•„์š”ํ•œ๊ฑด, treatment group ์—์„œ treat ๋ฅผ ๋ฐ›์ง€ ์•Š์•˜๋”๋ผ๋ฉด ์–ด๋–ค ๊ฒฐ๊ณผ๋ฅผ ๋ณด์˜€๋Š”์ง€์— ๋Œ€ํ•œ counterfactual โ‡จ ํ˜„์‹ค์—์„œ ๊ด€์ฐฐ ๋ถˆ๊ฐ€๋Šฅํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ€์žฅ ๋น„์Šทํ•œ control group ์„ ์ฐพ๋Š” ๊ฒƒ์ด ๊ด€๊ฑด์ด๋‹ค. • Causal experiment method: control group ์„ ํ™œ์šฉํ•ด ์–ด๋–ป๊ฒŒ counterfactual ์„ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ์„์ง€์— ๋Œ€ํ•œ ๋ฐฉ๋ฒ• โ‡จ Research design and apply method • ์šฐ๋ฆฌ๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์€, treatment group ์— ๋Œ€ํ•œ ATE ์ธ .. 2023. 4. 24.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ๋””์ž์ธ ๊ธฐ๋ฐ˜์˜ ์ธ๊ณผ์ถ”๋ก  ์ฐธ๊ณ ์˜์ƒ : Bootcamp 3-1. ๋””์ž์ธ ๊ธฐ๋ฐ˜์˜ ์ธ๊ณผ์ถ”๋ก  1. Quasi-experimental designs โ—ฏ Quasi-experiment designs • Research designs without random assignment : ๋ฌด์ž‘์œ„ ์‹คํ—˜์„ ํ•  ์ˆ˜ ์—†์„ ๋•Œ, ์ด์™€ ์œ ์‚ฌํ•˜๊ฒŒ ์‹คํ–‰๋  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ์‹คํ—˜ • Quasi-experiment ๋‚˜ Instrumental variable์˜ ๊ฒฝ์šฐ, ๊ด€์ฐฐ๋˜์ง€ ์•Š์€ ๊ฐ’๋“ค๊นŒ์ง€ (selection bias) ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด research design์„ ๋„์ž…ํ•œ๋‹ค. โ—ฏ Random assignment is not always feasible • Quasi experiment ๋Š” RCT์™€ ๊ฑฐ์˜ ๋™์ผํ•˜๋‹ค๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๋‹ค๋งŒ, selection bias๋ฅผ ์—†์• ๊ธฐ.. 2023. 4. 24.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ๋งค์นญ๊ณผ ์—ญํ™•๋ฅ  ๊ฐ€์ค‘์น˜ ์ฐธ๊ณ ์˜์ƒ : Bootcamp 2-4. ๋งค์นญ๊ณผ ์—ญํ™•๋ฅ  ๊ฐ€์ค‘์น˜ 1. Matching • Regression ์€ control variable๊ณผ selection bias์— ๋Œ€ํ•ด linear form์œผ๋กœ ๊ฐ€์ •ํ•˜๊ณ , control variable์„ conditioning ํ•จ์œผ๋กœ์จ ํŠน์„ฑ์„ ์œ ์‚ฌํ•˜๊ฒŒ ๋งŒ๋“ค์—ˆ๋‹ค๋ฉด matching ์€ functional form ์—†์ด ๋‹จ์ˆœํžˆ control variable์—์„œ ํŠน์„ฑ์ด ์œ ์‚ฌํ•œ ๊ฒƒ๋“ค์„ ์ง์ ‘์ ์œผ๋กœ ๋งค์นญํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. ํ›จ์”ฌ ๋” ์ง๊ด€์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ ๋ณผ ์ˆ˜ ์žˆ๋‹ค. Flexible ํ•œ ํ˜•ํƒœ์ด๋‹ค. • ์ฆ‰, ํšŒ๊ท€๋ถ„์„๊ณผ Matching ์€ functional form ์„ ๊ฐ€์ •ํ•˜๋Š๋ƒ ์•ˆ ํ•˜๋Š๋ƒ์˜ ์ฐจ์ด๋งŒ ์กด์žฌํ•œ๋‹ค. • Matching ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์—ฌ๋Ÿฌ ๊ฐ€์ง€๊ฐ€ ์žˆ๋Š”๋ฐ, ๊ฐ•์˜์—์„œ๋Š” 2๊ฐ€์ง€๋งŒ ์†Œ๊ฐœ.. 2023. 4. 21.
์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ์ธ๊ณผ์ถ”๋ก  ๊ด€์ ์—์„œ์˜ ํšŒ๊ท€๋ถ„์„ ์ฐธ๊ณ ์˜์ƒ : Bootcamp 2-3. ์ธ๊ณผ์ถ”๋ก  ๊ด€์ ์—์„œ์˜ ํšŒ๊ท€๋ถ„์„ 1. Casual Hierarchy • ์–ด๋–ค ์ข…๋ฅ˜์˜ selection bias๋ฅผ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ๊ธฐ์ค€ • Selection on Observables strategies : ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๋ณ€์ˆ˜๋“ค์— ์˜ํ•ด์„œ๋งŒ treatment์™€ control ์ด ์„ ํƒ๋œ๋‹ค๋Š” ๊ฐ€์ •์„ ๊ฐ€์ง€๊ณ  ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๋ณ€์ˆ˜๋“ค๋งŒ์„ ๊ฐ€์ง€๊ณ  selection bias๋ฅผ ์„ค๋ช…ํ•˜๋ ค๋Š” ๊ฒฝํ–ฅ • Selection on Unobservables strategies : ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์€ ๊ต๋ž€ ์š”์ธ๋“ค๋„ ์ ์ ˆํ•œ ์‹คํ—˜๋””์ž์ธ์„ ํ†ตํ•ด ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ์ „๋žต โ‡จ ์ข€ ๋” powerful ํ•œ ์ „๋žต 2. How to balance between treatment and control groups โ—ฏ Re.. 2023. 4. 21.
728x90