Upgrade to Pro — share decks privately, control downloads, hide ads and more …

統計的因果推論とデータ解析 / causal-inference-and-data-analysis

統計的因果推論とデータ解析 / causal-inference-and-data-analysis

統計的因果推論の解釈と、実際の運用における注意点をまとめた資料です。
先日、closedな勉強会で発表したものを一部改変したものです。
twitter : @tomoshige_n
mail : [email protected]

注)一部、わかりやすさを厳密性よりも優先した部分があります。厳密性などを求める方は、文献等をあたってください。

Tomoshige Nakamura

May 31, 2019
Tweet

More Decks by Tomoshige Nakamura

Other Decks in Science

Transcript

  1. ໨࣍ ‣ ܏޲είΞ͸ສೳༀͳͷ͔ʂʁ ‣ ϥϯμϜԽൺֱ࣮ݧʢ3BOEPNJ[FE$POUSPM5SJBMʣ ‣ ݹయతճؼϞσϧͷഁ୼ ‣ ೣ΋उࢠ΋܏޲είΞɻ ‣

    ҼՌਪ࿦ͷຊ࣭ͱ࣮σʔλͷղੳ ‣ ܏޲είΞͷٯ਺ͦ͜ຊ࣭ʂʢ1SPQFOTJUZ4DPSFʣ ‣ σʔλ͕ͳ͍ͳΒɺ͹Β͖ͭΛߟ͑Δʢ#PPUTUSBQ.FUIPEʣ ‣ ฏۉॲஔޮՌͰ͸ͳ͘ɺ0WFSMBQΛߟ͑Δʢ0WFSMBQ8FJHIUTʣ ‣ ౷ܭతҼՌਪ࿦ͷ࠷ۙͷൃల ‣ ͋ͷਓͷҼՌޮՌ͸Ͳ͏ͩΖ͏͔ʁʢ$BVTBM'PSFTU3-FBSOFSʣ ‣ ·ͱΊ !3
  2. ϥϯμϜԽൺֱ࣮ݧ* ‣ ೥୅͔ΒɺσʔλΛ༻͍ͨޮՌͷଌఆ͸ߦΘΕ͖ͯͨɻ ‣ ༗໊ͳͷ͸ɺϑΟογϟʔ͕ߦͬͨ೶ༀ࣮ݧͰ͋Δɻ !4 ංྉ ͋Γ ංྉ ͳ͠

    μ0 μ1 μ1 − μ0 = 0?? ɾ೔౰ͨΓҧ͏ ɾਫ͸͚΋ҧ͏ ࡞෺ͷ੒௕ʹؔ࿈͢Δ ม਺͕ංྉΛ͋Γͳ͠Ͱ ҧ͑͹ग़ͯ͘Δ݁Ռ ɾංྉͷޮՌ ɾ౔৕ͷಛ࣭ ͭΛࠞͥͨޮՌ ංྉ ͋Γ ϥϯμϜʹ ංྉ͋Γ ංྉͳ͠ ഑ஔͯ͠ΈΑ͏ ྫ͑͹ίΠϯτεͰʂ
  3. ϥϯμϜԽൺֱ࣮ݧ** ‣ ϥϯμϜԽൺֱ࣮ݧ͸ɺௐ΂͍ͨӨڹʹରͯ͠ɺ ଞͷม਺͕Өڹ͢ΔͷΛ཈͑ΔͨΊͷख๏ɻ ‣ ྫ͑͹ɺ೔র͕࣌ؒɺ౦ʹߦ͚͹ߦ͘΄Ͳ௕͍ ৔߹ɺͦΕ͕ۮવͰ͋ͬͯ΋࡞෺ͷ੒௕΁ͷӨ ڹ͸ආ͚ΒΕͳ͍ɻ ‣ ϥϯμϜԽൺֱ࣮ݧ͸ɺ݁ՌʹӨڹ͕͋Δม਺

    ͕ɺංྉ͋Γɾͳ͠ͷ྆ํͰಉ͡෼෍Λ࣋ͭΑ ͏ʹ͢Δͱ͍͏ࢹ఺Ͱઃܭ͞ΕΔɻ ‣ ͭ·ΓɺॲஔͷޮՌΛਪఆ͢Δࡍʹ஫ҙΛ෷͏ ΂͖͜ͱ͸ʮॲஔ܈ʯͱʮରর܈ʯ͕ɺॲஔม ਺Ҏ֎ͷม਺ʹ͍ͭͯۉ࣭ͳूஂͰ͋Δ͔ͱ͍ ͏͜ͱΛߟ͑Δඞཁ͕͋ΔʢΊͪΌͪ͘Όେ ࣄʣɻ !5 ංྉ ͋Γ ංྉ ͳ͠ ๺ ౦ ೆ ੢ ೔র࣌ؒ ౦ 5ʢॲஔม਺ʣcc9ʢڞมྔʣ
  4. ϥϯμϜԽൺֱ࣮ݧ*** ‣ :Λ݁Ռม਺ɺ8Λංྉͷ༗ແɺ9Λ೔র࣌ؒͱ͢Δɻ ‣ ͜ͷͱ͖ɺӈͷਤͷΑ͏ͳؔ܎͕͋Δɻ ‣ ͜͜Ͱɺ4675"ΛԾఆ͢Δɻ ‣ ͜ͷͱ͖ɺ஌Γ͍ͨ݁Ռ͸ ‣

    ࣮ࡍʹܭࢉ͢Δฏۉͷࠩ͸ɺ ‣ !6 X W Y = WY1 + (1 − W)Y0 E[Y1 ] − E[Y0 ] E[Y|W = 1] − E[Y|W = 0] E[Y|W = 1] = ∫ × yf(y, x|w = 1)dydx = ∫ × y1 f(y1 , x|w = 1)dy1 dx = ∫ × y1 f(y1 , x)dy1 dx = E[Y1 ] X W 3BOEPNJ[BUJPO 3$5ͳͷͰ ˎڞมྔ͸ॲஔʹ͸Өڹ͠ͳ͍ ˎॲஔ͸જࡏ݁Ռม਺ʹӨڹ͠ͳ͍ Y0 , Y1 , X ⊥ ⊥ W Y0 , Y1 Y0 , Y1
  5. ϥϯμϜԽൺֱ࣮ݧ*7 ‣ ϥϯμϜԽ͸෺ཧతʹɺ݁Ռม਺ʹӨڹ͢Δม਺9͕ɺॲஔม਺8ʹӨڹΛ ༩͑ͳ͍Α͏ʹ͢ΔͨΊͷํ๏ɻ ‣ ղऍΛม͑Ε͹ɺॲஔΛड͚ͨ܈ͱɺड͚ͳ͔ͬͨ܈Ͱɺڞมྔͷ෼෍ͷࠩ ͕ͳ͘ͳΔΑ͏ʹ͍ͯ͠Δɻ ‣ ҼՌਪ࿦ͱ͸ڀۃతʹɺʮ݁Ռม਺ʹӨڹ͢Δม਺Λ͢΂ͯʯॲஔ܈ͱରর ܈Ͱɺ3BOEPNJ[FͰ͖Ε͹ɺͦΕͰ͢΂ͯࣄ଍ΓΔɻ

    ‣ Ͱ͸ɺ3$5͕Ͱ͖ͳ͍৔߹ͲͷΑ͏ʹ͢Ε͹͍͍ͩΖ͏ʁ ‣ ॲஔม਺8͕ɺ9ͷӨڹΛड͚͍ͯΔ৔߹ɺͲͷΑ͏ʹͯ͠ௐ੔͢Ε͹͍͍ ͩΖ͏ʁ ‣ ٸʹɺݱ୅తͳ࿩ʹߦ͘લʹɺ΋͏গ͠ྺ࢙తͳ࿩Λ͍ͨ͠ɻ ‣ ͜͜Ͱɺొ৔͢Δͷ͕ݹయతͳճؼϞσϧʢ"/07"ʣͰ͋Δɻ !7
  6. ݹయతճؼϞσϧͷഁ୼* ‣ ೶ༀ͕͋Δ২෺ͷ੒௕ʹ༩͑ΔӨڹΛݟΔͨΊɺ ‣ :Λ২෺ͷ੒௕౓߹͍ʢDNʣ ‣ 8Λ೶ༀͷ༗ແ ‣ 9Λ೔ࣹྔʢL8N?ʣ ‣

    ͱ͍͏ม਺͕؍ଌ͞Εͨͱ͠·͢ɻ ‣ ೶ༀͷࢄ෍͸೔ࣹྔ͕ଟ͍৔ॴͷํʹଟ͘ࢃ͍ͨͱ͠·͢ɻ ‣ ͜ͷͱ͖ɺ೶ༀΛ·͍ͨ৔߹ͷ੒௕ͱɺࢃ͔ͳ͔ͬͨ৔߹ͷ੒௕ͷજࡏ݁Ռ ม਺͸ɺҎԼͷΑ͏ʹද͞ΕΔͱ͠·͢ʢЋఆ਺F͸ޡࠩʣɻ ‣ ͜ͷͱ͖ɺ:Λ9ͱ8Ͱճؼͨ͠ͱ͖ͷ8ͷճؼ܎਺͸ɺ೶ༀޮՌͰ͠ΐ͏ ͔ʁ !8 Y1 = Xβ + α + e Y0 = Xβ + e
  7. ݹయతճؼϞσϧͷഁ୼** ‣ ͜ͷ৔߹ɺ݁Ռม਺ʹର͢ΔϞσϧ͸ɺ ‣ :9Ќ 8Ћ ޡ߲ࠩ ‣ ͱද͞ΕΔɻ ‣

    ઢܗճؼϞσϧΛ౰ͯ͸Ίͨ৔߹ͷɺ8ͷճؼ܎਺ͷਪఆྔ͸ɺ ‣ ౴͑ɿճؼ܎਺͸ҼՌޮՌʹର͢ΔҰகਪఆྔͰ͋Δɻ ‣ ཧ༝͸ɾɾɾ ‣ 8͔Β:΁ͷӨڹ͸ɺજࡏ݁Ռม਺ͷ͕ࠩ9ʹΑΒͣҰఆͷͱ ͖ɺҼՌޮՌ͸ճؼ܎਺ʹҰக͢Δ͔Βɻ !9 argmin ̂ β, ̂ α (Y − Xβ − Wα)2 ⇔ ̂ α = 1 N1 WT(Y − Xβ) → E[Y − Xβ|W = 1] (N → ∞) = E[Wα + ε|W = 1] = α
  8. ݹయతճؼϞσϧͷഁ୼*** ‣ Ͱ͸ɺ࣍ͷΑ͏ͳ৔߹͸Ͳ͏Ͱ͠ΐ͏͔ʁ ‣ ͜ͷ৔߹͸ɺ͏·͘ਪఆͰ͖Δɻ ‣ ॲஔม਺ʹӨڹ͢Δม਺͕ɺҼՌతޮՌ ʹؚ·Εͳ͍ɻ ‣ ͜ͷ৔߹ɺަབྷ͸ଘࡏ͠ͳ͍ɻ

    !10 Y1 = X1 + X2 + X3 + e Y0 = X1 + X2 + e W ∼ Bernoulli(p) logit(p) = (X1 + X2 )/2 9ͱ9͕ॲஔʹӨڹ͢Δ 9ͱ9͕྆ํͷજࡏ݁Ռม਺ʹӨڹ͠ 9͕:ͷΈʹؚ·ΕΔ
  9. ݹయతճؼϞσϧͷഁ୼*7 ‣ Ͱ͸ɺ࣍ͷΑ͏ͳ৔߹͸Ͳ͏Ͱ͠ΐ͏͔ʁ ‣ જࡏ݁Ռม਺ͷࠩ͸ɺ9ʹґଘɻ ‣ ॲஔม਺΋9ʹґଘɻ ‣ ͜ͷͱ͖ɺճؼ܎਺ͷ஋͸ɺҼՌతޮՌʹର ͯ͠όΠΞεͷ͋ΔਪఆྔͰ͋Δɻ

    !11 Y1 = X1 + X2 + X3 + e Y0 = X1 + X2 + e W ∼ Bernoulli(p) logit(p) = (X1 + X2 + X3 )/3 9ͱ9ͱ9͕ॲஔʹӨڹ͢Δ 9ͱ9͕྆ํͷજࡏ݁Ռม਺ʹӨڹ͠ 9͕ҼՌޮՌʹӨڹ͢Δ
  10. ݹయతճؼϞσϧͷഁ୼7 ‣ ʲ͜͜·Ͱͷ·ͱΊͱ௥هʳ ‣ ݹయతճؼϞσϧ͸ɺҼՌతޮՌʢજࡏ݁Ռม਺ͷࠩʣͱɺॲஔม਺ͷ྆ ํʹӨڹΛ༩͑Δม਺͕ଘࡏ͢Δ৔߹ɺഁ୼͢Δɻ ‣ ೶ༀͷྫͰઃఆͨ͠ϞσϧͰ͸ɺ೶ༀΛ༩͑ΒΕͯ΋༩͑ΒΕͳͯ͘ ΋ɺ੒௕཰͸ҰఆͰ͋Γɺ೶ༀͰͦΕ͕͋Δఔ౓৳͹ͤΔͱ͍͏Ϟσϧɻ ‣

    ͜ͷϞσϧ͸ɺݱ৅ʹରͯ͠ૉ௚ʹϞσϦϯά͍ͯ͠ΔͷͰɺ໰୊͕͋Δ Θ͚Ͱ͸ͳ͍ɻ ‣ Ͱ͸ɺͦ΋ͦ΋ʮҼՌతޮՌʯͱʮॲஔม਺ʯͷ྆ํʹӨڹΛ༩͑Δڞมྔ ͕ଘࡏ͢Δͷ͸ͲͷΑ͏ͳ৔߹͔ɻ ‣ ʲ޿ࠂʳޮՌ͕͋Γͦ͏ͳਓʹ޿ࠂΛ͏ͪɺޮՌͷେ͖͕͞ɺʮॲஔม਺ʯ ʹӨڹ͢Δڞมྔʹґଘ͢Δɻ ‣ ʲӸֶʳλόίΛٵ͏͔ٵΘͳ͍͔͸ɺͦͷਓͷݸਓಛੑʹґଘ͠ɺͨ͹ ͜ʹΑΔ࣬පൃ঱཰΋ݸਓಛੑʹΑܾͬͯఆ͞ΕΔɻ !12
  11. ݹయతճؼϞσϧͷഁ୼7* ‣ ݹయతͳճؼϞσϧ͸ɺಛʹ݁Ռม਺ʹରͯ͠ม਺ͷӨڹ͕ઢܗͰ͋Δ͜ͱ Λظ଴͢Δ͕ɺ࣮ࡍͷσʔλͷղੳʹ͓͍ͯ͸ɺ݁Ռม਺͸ਖ਼ن෼෍ͱ͸ݶ Βͳ͍ɻ ‣ $7ͷ৔߹\ ^CJOBSZ ‣ ਓ਺ͷ৔߹\

      ^EJTDSFUF ‣ ͭ·ΓɺҰൠతʹߟ͑Ε͹ɺજࡏ݁Ռม਺ʹର͢Δظ଴஋ΛϞσϦϯά͠ ͯɺͦͷࠩΛܭࢉ͢Δͱ͍͏͜ͱ͕ඞཁʹͳΔɻ ‣ ͜͏ͳͬͯ͘Δͱɺ͍Α͍Αखଓ͖͕൥ࡶͰ͋Γɺؔ਺HͱПͷઃఆΛ͢Δͳ Μͯ͜ͱ͸΍Γͨ͘ͳ͍ʢͨͩɺۙ೥͸ػցֶशͰ͜ΕΛղফ͢ΔΞϓϩʔ ν΋͋Δʣɻ !13 g1 (E[Y1 |X]) = ϕ1 (X) g0 (E[Y0 |X]) = ϕ0 (X) E[Y1 − Y0 ] = E[g−1 1 (ϕ1 (X))] − E[g−1 0 (ϕ0 (X))]
  12. ೣ΋उࢠ΋܏޲είΞʂʂ* ‣ ͦΜͳͱ͖ɺࢲͨͪ͸ʮ܏޲είΞʯͱ͍͏΋ͷΛݟ͚ͭ·͢ɻ ‣ ʮͲ͏΍Βɺॲஔม਺ʹର͢ΔϩδεςΟοΫճؼϞσϧ͔ɺΛ༧ଌ͢ Δػցֶश͔Βਪఆ͞ΕΔॲஔ֬཰ʯΛ༻͍ͯɺॏΈ෇͚͢Ε͹ҼՌతޮՌ ΛਪఆͰ͖ΔΒ͍͠ʂʂʂ ‣ ŠŠŠŠŠ܏޲είΞΛ༻͍ͨղੳQSPDFEVSFŠŠŠŠŠ ‣

    ॲஔม਺8Λɺڞมྔ9ʢ؍ଌ͞Ε͍ͯΔ΋ͷʣͰճؼ͢Δɻ ‣ -PHJTUJDSFHSFTTJPO3'9HCPPTU -1FOBMUZΛ࢖͏ ‣ "6$͕Ҏ্͘Β͍ͳΒɺͱΓ͋͑ͣ0,ɻ ‣ ٯ֬཰ॏΈ෇͚Λ͢Δɻ ‣ ͜ͷਪఆ݁ՌΛϨϙʔτ͓ͯ͠͠·͍ ‣ ŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠŠ !14 ̂ E[Y1 − Y0 ] = N ∑ i=1 Wi Yi π(Xi ) / N ∑ i=1 Wi π(Xi ) − N ∑ i=1 (1 − Wi )Yi 1 − π(Xi ) / N ∑ i=1 1 − Wi 1 − π(Xi ) Y X W
  13. ໨࣍ ‣ ܏޲είΞ͸ສೳༀͳͷ͔ʂʁ ‣ ϥϯμϜԽൺֱ࣮ݧʢ3BOEPNJ[FE$POUSPM5SJBMʣ ‣ ݹయతճؼϞσϧͷഁ୼ ‣ ೣ΋उࢠ΋܏޲είΞɻ ‣

    ҼՌਪ࿦ͷຊ࣭ͱ࣮σʔλͷղੳ ‣ ܏޲είΞͷٯ਺ͦ͜ຊ࣭ʂʢ1SPQFOTJUZ4DPSFʣ ‣ σʔλ͕ͳ͍ͳΒɺ͹Β͖ͭΛߟ͑Δʢ#PPUTUSBQ.FUIPEʣ ‣ ฏۉॲஔޮՌͰ͸ͳ͘ɺ0WFSMBQΛߟ͑Δʢ0WFSMBQ8FJHIUTʣ ‣ ౷ܭతҼՌਪ࿦ͷ࠷ۙͷൃల ‣ ͋ͷਓͷҼՌޮՌ͸Ͳ͏ͩΖ͏͔ʁʢ$BVTBM'PSFTU3-FBSOFSʣ ‣ ·ͱΊ !15
  14. #BMBODJOH$PWBSJBUFT** ‣ ۩ମతͳ࿩ʹೖΔલʹɺ6ODPOGPVOEFEOFTTʢڧ͘ແࢹՄೳͳׂΓ෇͚ʣʹ ͍ͭͯઆ໌Λ͓͖ͯ͠·͢ɻ ‣ ͜Ε͸ɺજࡏ݁Ռม਺ͱॲஔม਺ͷ྆ํʹӨڹ͢Δม਺Λɺࢲͨͪ͸؍ଌͰ ͖͍ͯΔͱ͍͏ԾఆͰ͢ɻ ‣ ͜ͷԾఆ͕ͳ͍ͱɺ࿩͕࢝·Γ·ͤΜ͕ɺݱ࣮ͷ໰୊Ͱ͢΂ͯ؍ଌͰ͖͍ͯ Δͱ͍͏ঢ়گ͸ߟ͑ʹ͘͘ɺ੒Γཱͨͳ͍͜ͱ΋ଟ্͍ʹɺνΣοΫͰ͖ͳ

    ͍ͱ͍͏ͳ͔ͳ͔ո͍͠ԾఆͰ͸͋Γ·͢ɻ ‣ ͨͩɺ੒Γཱͭͱͯ͠ߟ͓͔͑ͯͳ͍ͱ࢝·Βͳ͍ͱ͍͏ͷ΋ࣄ࣮Ͱ͢ɻ ‣ ͔͠͠ɺ͜ͷԾఆ͕੒Γཱ͍ͬͯͳ͍ͱҼՌతޮՌΛਪఆͨ݁͠Ռʹҙຯ͕ ͳ͍͔ͱݴΘΕΔͱͦ͏Ͱ͸͋Γ·ͤΜɻ ‣ ࠓ͔Βͷ࿩ͰͦΕ͕Θ͔Γ·͢ɻ !18 (Y1 , Y0 ) ⊥ ⊥ W X
  15. #BMBODJOH$PWBSJBUFT*7ʢ਺ֶతͳٞ࿦͕ඞཁʣ ‣ ॲஔ܈ͷ݁Ռม਺ͷฏۉΛɺ͢΂ͯͷඪຊ͕ॲஔΛड͚ͨ৔߹ͷ݁Ռม਺ͷ ฏۉ΁मਖ਼͢Δɻ ‣ ͜Ε͕ɺॲஔ܈Ͱͷ݁Ռม਺ͷฏۉΛॻ͖Լͨ͠৔߹Ͱ͢ɻ͜͜Ͱɺॏཁͳ ͷ͸ੵ෼ͷதʹG YcX ͕͋Δ͜ͱͰ͢ɻ ‣

    ͜Ε͸ɺॲஔ܈ͷڞมྔͷ෼෍Ͱ͋Γɺ͜Ε͸ʮॲஔΛड͚ͨ৔߹ͷજࡏ݁ Ռม਺Λɺڞมྔͷ෼෍શମͰ͸ͳ͘ɺॲஔ܈ͷڞมྔ෼෍Ͱੵ෼͍ͯ͠ ΔʯͷͰ࿪Έ͕ੜ͍ͯ͡Δͱ͍͏;͏ʹղऍ͠·͢ɻ ‣ Ͱ͸ɺͦͷ࿪ΈΛ΋ͱʹ໭͠·͠ΐ͏ɻ !20 E[WY] = ∫ w ⋅ y ⋅ f(y, w, x)dydxdw = ∫ y1 ⋅ f(y1 , w = 1,x)dydx = P(W = 1) ∫ y1 ⋅ f(y1 |w = 1,x)f(x|w = 1)dydx = P(W = 1) ∫ y1 ⋅ f(y1 |x)f(x|w = 1)dydx
  16. #BMBODJOH$PWBSJBUFT7ʢ਺ֶతͳٞ࿦ʣ ‣ ͍·ɺ ‣ ͱ͠·͠ΐ͏ɻ͜ΕͰॏΈ෇͚ΒΕͨ8:ͷظ଴஋͸ɺॲஔ܈ͰͷҼՌޮՌ ͷਪఆྔʹͳΓ·͢ɻ ‣ ͱ͜ΖͰɺ͜ͷॏΈͷٯ਺͸ɺ ‣ ͱͳΔͷͰɺ͜Ε͸܏޲είΞͰ͢ɻ

    !21 w1 (x) = f(x) f(x|w = 1)P(W = 1) E [w1 (x) ⋅ WY] = P(W = 1) ∫ y1 ⋅ f(y1 |x)f(x|w = 1) ⋅ w1 (x)dy1 dx = ∫ y1 ⋅ f(y1 |x)f(x)dy1 dx = E[Y1 ] 1 w1 (x) = f(x|w = 1)P(W = 1) f(x) = P(W = 1|X = x)
  17. #BMBODJOH$PWBSJBUFT7*ڞมྔ෼෍Λൺֱ͢Δ܈Ͱ౳͘͢͠Δɻ ‣ ΑͬͯɺҼՌਪ࿦ʹ͓͍ͯ܏޲είΞ͕ຊ࣭తͳͷͰ͸ͳ͘ɺͦͷٯ਺͕࣋ͭ ҙຯͦ͜ຊ࣭తͰ͋Δ͜ͱ͸ݴ͏·Ͱ΋͋Γ·ͤΜɻ ‣ ͜ͷΑ͏ͳΛ༻͍Δͱɺ෼෍͕ॲஔ܈ͱରর܈Ͱ࿪ΜͰ͍ͯ΋ɺॏΈ෇͚ʹ Αͬͯमਖ਼͕ߦΘΕɺ෼෍͕౳͘͠ͳΔ͜ͱ͕Θ͔Γ·͢ɻ !22 Covariate Density

    −3 −2 −1 0 1 2 3 0.00 0.10 0.20 0.30 Covariate Density −3 −2 −1 0 1 2 3 0.00 0.10 0.20 0.30 Covariate Density −3 −2 −1 0 1 2 3 0.00 0.05 0.10 0.15 Covariate Density −3 −2 −1 0 1 2 3 0.00 0.05 0.10 0.15 0.20 ॲஔ܈ ରর܈ ௐ੔લ ௐ੔લ
  18. #BMBODJOH$PWBSJBUFT7**ʢຊ࣭తͳ࿩ʣ ‣ ͜Ε͸ɺʮٙࣅతͳϥϯμϜԽʯͱΈͳ͢͜ͱ͕Ͱ͖·͢ɻ ‣ ͭͷ܈ͷ෼෍Λ௼Γ߹Θ͍ͤͯΔʢόϥϯγϯά͍ͯ͠Δʣͱ͍͏ͷ͸ɺॲ ஔม਺ʹӨڹΛ༩͑Δม਺જࡏ݁Ռม਺ʹӨڹΛ༩͑ΔӨڹ͕ɺͭͷ܈ Ͱಉ͡෼෍Λ࣋ͭΑ͏ʹ͢Δͱ͍͏͜ͱʹରԠ͠·͢ɻ ‣ ͜ͷ࿩ͷ݁Ռɺʮ"6$͕Ҏ্͕Ͳ͏ͷ͜͏ͷʯΈ͍ͨͳ࿩͸ʮφϯηϯ εʯͩͱ͍͏͜ͱ͕Θ͔Γ·͢ɻ

    ‣ ܏޲είΞΛ༻͍ͨҼՌతޮՌͷਪ࿦ʹ͓͍ͯ͸ ‣ ॲஔͱજࡏ݁Ռม਺ʹӨڹΛ༩͑Δม਺͸Կ͔Λߟ͑Δ ‣ ͦͷม਺͕ɺॏΈ෇͚ʹΑͬͯॲஔ܈ରর܈Ͱಉ͡෼෍ʹै͏͔Λ֬ೝ ͢Δ ‣ ͱ͍͏͜ͱΛ֬ೝ͠ͳͯ͘͸ͳΓ·ͤΜɻ"6$΍ଞͷ౷ܭతͳج४ྔ͸ͳΜ ͷࢀߟʹ΋ͳΒͳ͍ͷͰ͢ɻ !23
  19. #BMBODJOH$PWBSJBUFT*9 ‣ ڧ͘ແࢹՄೳͳׂΓ෇͚͕੒Γཱ͍ͬͯͳ͍ͷʹɺ܏޲είΞʹΑͬͯॏΈ ෇͚ͯ͠ਪఆ͢Δҙຯ͸͋Δͷʁͱ͍͏͜ͱΛݴ͍·ͨ͠ɻ ‣ ͔͠͠ɺ͍·ͳΒʮҙຯ͕͋Δʯͱݴ͑Δ͸ͣͰ͢ɻ ‣ ͳͥͳΒʮ܏޲είΞʹΑΔॏΈ෇͚ʯͱ͸ɺ ‣ ൺֱ͢Δ܈Λʮ౤ೖͨ͠มྔʯʹ͓͍ͯۉ࣭Խ͢Δߦҝ͔ͩΒɻ

    ‣ ͭ·Γɺগͳ͘ͱ΋ௐ੔Λߦͬͨ΄͏͕ɺ୯७ͳൺֱΛ͢ΔΑΓ΋ɺϚγͳ ݁ՌΛಘ͍ͯΔͱߟ͑͸ࣗવͰ͢ɻ ‣ ʲ݁࿦ʳڧ͘ແࢹՄೳͳׂΓ෇͚͕੒Γཱͭͷʹे෼ͳڞมྔ͕؍ଌ͞Εͯ ͍ͳ͔ͬͨͱͯ͠΋ɺॲஔͷޮՌΛਪఆ͢Δʹ͋ͨͬͯʮӨڹΛ༩͍͑ͯ Δʯͱߟ͑ΒΕΔม਺Λʮௐ੔ͯ͠ʯɺӨڹΛਪఆ͢Δ͜ͱ͸ҙຯ͕͋Δɻ !25
  20. ͹Β͖ͭΛߟ͑Α͏** ‣ ҼՌޮՌͷਪఆͱ͸͍͏ͳΕ͹ɺ͜ͷܽଌͷ෦෼͸ ‣ ʮ͜Ε͘Β͍ͩΑʔʋ ʔʆ ϊʯ ‣ ͱ͍͏༧૝Λͯ͠ɺਪఆ͍ͯ͠ΔΘ͚͔ͩΒɺ ‣

    ਓؒͰݴ͏ͱ͜Ζͷ ‣ ʮͨͿΜɺɺʓʓ͘Β͍ͩΑͶʯ ‣ ͱ͍͏͜ͱΛɺσʔλͰ΋͖ͪΜͱߟ͑ͳ͍ͱɺ ‣ ͦΜͳʹڧ͍͜ͱݴ͑ͳ͍ͷʹɺ ‣ ʮޮՌ͋ΔͰʂʂʢؔ੢หʣʯ ‣ Έ͍ͨͳ͜ͱΛݴͬͪΌͬͯɺ΍ͬͯΈͨΒޮՌͳ͍͡ΌΜʂʢඪ४ޠʣ ‣ ͱ͍͏ࣄʹͳΓ͔Ͷ·ͤΜɻ ‣ σʔλΛղੳ͢Δ͜ͱ͸ɺʮ͹Β͖ͭʯΛߟ͑Δͷͱಉ͡ ‣ ͱ͍͏͙Β͍େࣄͳ͜ͱͰ͢ 㱼ʆ !28
  21. ͹Β͖ͭΛߟ͑Α͏*7 !30 ඪຊ 1 N N ∑ i=1 Wi Yi

    π(Xi ) ਪఆྔ Կ౓΋ σʔλΛ ൃੜ {Xi , Wi , Yi }N i=1 ∼ F σʔλ 9 8 : ࣮ݱ஋ 1 N N ∑ i=1 wi yi π(xi ) σʔλͰ ࡞ͬͨਪఆ஋ σʔλ 9 8 : σʔλ# 9 8 : ʜ σʔλͰ ࡞ͬͨਪఆ஋ 1 N N ∑ i=1 wi yi π(xi ) σʔλ#Ͱ ࡞ͬͨਪఆ஋ 1 N N ∑ i=1 wi yi π(xi ) ਪఆ஋ͷώετάϥϜ ʢݱ࣮ʹ͸ॻ͚ͳ͍ʣ ಘΒΕ͍ͯΔ σʔλ͔Βܭࢉ͞Εͨ஋͸ ͔͜͜΋ ͠Εͳ͍ ͔͜͜΋ ͠Εͳ͍ ͔͜͜΋ ͠Εͳ͍
  22. ͹Β͖ͭΛߟ͑Α͏7*ʢ#PPUTUSBQ๏ʣ !32 1 N N ∑ i=1 Wi Yi π(Xi

    ) ਪఆྔ σʔλ {Xi , Wi , Yi }N i=1 ∼ F 9 8 : ࣮ݱ஋ 1 N N ∑ i=1 wi yi π(xi ) σʔλͰ ࡞ͬͨਪఆ஋ ϒʔτετϥοϓαϯϓϧͷ ώετάϥϜʢݱ࣮Ͱॻ͚Δʣ σʔλ ϒʔτετϥοϓ αϯϓϧ σʔλ͔Β Ϧαϯϓϧ ʜ 1 N N ∑ i=1 w(1) i y(1) i π(x(1) i ) ϒʔτετϥοϓαϯϓϧͰ ࡞ͬͨਪఆ஋ αΠζ͸/ ճ਺Λ# ϒʔτετϥοϓ αϯϓϧ# X(1) W(1) Y(1) X(B) W(B) Y(B) 1 N N ∑ i=1 w(B) i y(B) i π(x(B) i ) ʜ ώετάϥϜʹ͢Δ ਪఆ஋ ਪఆ஋
  23. ͹Β͖ͭΛߟ͑Α͏7** ‣ ͜ͷ"d'ͷͭͷ஋ʹͭ ͍ͯ͸ɺҼՌޮՌͷਪఆ Λͨ͠ղੳͰ͸ɺઈରʹ ࣔͨ͠΄͏͕͍͍ɻ ‣ ·ͨɺӈͷΑ͏ͳਤ΋߹ Θͤͯࣔ͢͜ͱͰɺͲͷ ఔ౓ɺਪఆ͕͹Βͭ͘ͷ

    ͔͕Θ͔Γ΍͘͢ͳΔɻ ‣ ͜Ε͸ͥͻ΍ͬͯ΄͍͠ Ͱ͢ʂʂ !33 σʔλͷ ਪఆ஋ #PPUTUSBQ αϯϓϧͷฏۉ ਪఆ஋ ϒʔτετϥοϓα ϯϓϧͷฏۉ ਪఆ෼ࢄ ϒʔτετϥοϓ αϯϓϧͷ෼ࢄ ৴པ۠ؒ ϒʔτετϥοϓ ৴པ۠ؒ " # $ % & ' #PPUTUSBQ $POpEFODF*OUFSWBM "TZNQUPUJD *OUFSWBM
  24. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ* ‣ 8FJHIUFE"WFSBHF5SFBUNFOU&⒎FDU 8"5& ॏΈ෇͚ฏۉॲஔޮՌ ͸ɺ ࣍ͷΑ͏ʹఆٛ͞ΕΔɻ ‣ ͜͜ͰɺI Y

    ʹ͢ΔͱɺɹɹɹɹɹɹʹͳΔͷͰɺ͜Ε͸"5&ΛؚΉਪ ఆྔͰ͋Δɻ ‣ G Y I Y ͸ɺI Y ʹΑͬͯG Y ʹରͯ͠ॏΈΛ෇͚ͯɺG Y ͷͲͷ෦෼ʹண໨͢ Δ͔Λද͢ɻ ‣ ྫ͑͹ɺI Y   FYQ Y ͷॏΈΛ͚ͭΔ৔߹Λߟ͑Δͱɺ9͕େ͖ ͘ͳΔʹ࿈ΕͯI Y ͕େ͖͘ͳΔͷͰɺG Y ͷ෼෍ͷӈଆʹେ͖ͳॏΈ͕ͭ͘ɻ !35 τ(x) = E[Y1 − Y0 |X = x] τh = ∫ τ(dx)f(x)h(x)μ(dx) ∫ f(x)h(x)μ(dx) ͜͜Ͱɺ τ1 = E[Y1 − Y0 ] Y Y Y͕େ͖͍΄Ͳ େ͖ͳॏΈ
  25. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ** ‣ I Y F Y F Y ͱ͍͏ܗΛԾఆ͢ΔͱɺG Y

    I Y ͸ͲͷΑ͏ʹͳΔͩΖ͏ ͔ʁ ‣ ܏޲είΞ͕ͷ෦෼ʢͭ·ΓɺॲஔΛड͚Δ֬཰͕ʣͷ෦෼ʹେ ͖ͳॏΈ͕͔͔Γɺ܏޲είΞ͕খ͍͞େ͖͍෦෼ʹ͸ॏΈ͕͔ͭͳ ͍ɻ ‣ ࠷ॳͷํͷ܏޲είΞͰͷॏΈ෇͚͸ɺॲஔ܈ͰͷฏۉΛɺॲஔΛड͚ͨ৔ ߹ͷજࡏ݁Ռม਺ͷظ଴஋ʹɺิਖ਼͢Δ΋ͷͰͨ͠ɻ ‣ ͜͜ͰٻΊͨॏΈ͸ɺ࣮͸࣍ͷ໰୊ͰɺI Y ͷ৔߹ͷղʹରԠ͠·͢ɻ ‣ ͭ·Γɺॲஔ܈ͱରর܈ͷڞมྔ෼෍Λ΋ͱͷڞมྔ෼෍ʹରͯ͠௼Γ߹Θ ͤΔͨΊͷॏΈͰ͋Δͱߟ͑ΒΕ·͢ʢ#BMBODJOH8FJHIU -JFUBM  !36 E [w1 (x) ⋅ WY] = E[Y1 ] f(x|w = 1)w1 (x) = f(x|w = 0)w0 (x) = f(x)h(x)
  26. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ*** ‣ σʔλ͸/Ͱɺڞมྔ9͸ /   ͔Βੜ੒ͨ͠ɻ ‣ ܏޲είΞ͸MPHJUF Y

    Yͱ͠ ͨɻ·ͨॲஔม਺8d#FS F Y  ‣ ͦͯ͠ɺɹɹɹɹɹɹɹɹɹɹʹ ରͯ͠ɺVOXFJHIUFEPWFSMBQ XFJHIUJOWFSTFQSPCBCJMJUZ XFJHIUΛߦͬͨ৔߹ͷɺ෼෍ͷ௼ Γ߹͍Λܭࢉ͢Δɻ ‣ ͜͜Ͱ࢖༻ͨ͠܏޲είΞ͸ɺਪ ఆ஋Ͱ͸ͳ͘ਅͷ஋Λར༻ͨ͠ɻ ‣ ݁ՌɺPWFSMBQͷ৔߹ͱɺ*18ͷ ৔߹Ͱ෼෍ͷ௼Γ߹͍Ͱҧ͍͕ݟ ΒΕͨɻ !37 f(x):original distribution x Density −4 −2 0 2 4 0.0 0.1 0.2 0.3 0.4 0.5 0.6 f(x)e(x)(1−e(x)): overlap distribution x Density −4 −2 0 2 4 0.0 0.1 0.2 0.3 0.4 0.5 0.6 f(X|T = 0), f(X|T = 1) Unweighted x Density −4 −2 0 2 4 0.0 0.1 0.2 0.3 0.4 0.5 0.6 using overlap weight x Density −4 −2 0 2 4 0.0 0.1 0.2 0.3 0.4 0.5 0.6 using inverse probability weights x Density −4 −2 0 2 4 0.0 0.1 0.2 0.3 0.4 0.5 0.6
  27. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ*7 ‣ ࣍ͷ݁Ռ͕ɺ#BMBODJOH8FJHIUͷॏཁੑΛอূ͢Δɻ !38 5IFPSFN -J .PSHBOBOE;BTMBWTLZ  5\ ^ͷ৔߹ɺ#BMBODJOH8FJHIU͔Βߏ੒͞ΕΔਪఆྔɹɹ͸XFJHIUFE

    BWFSBHFUSFBUNFOUF⒎FDU 8"5&ɺ ʹର͢ΔҰகਪఆྔͰ͋Δɻ ͜͜Ͱɺ͸ɺɹɹɹɹɹɹɹɹɹɹɹɹɹͱ͢Δͱ͖ɺ Ͱఆٛ͞ΕΔɻ ·ͨɺ5\ ^ͷ৔߹ͷ#BMBODJOHXFJHIUͱ͸ҎԼΛຬͨ͢ɺॏΈͰ͋Δɻ ̂ τh = ∑ i w1 (xi )Zi Yi ∑ i w1 (xi )Zi − ∑ i w0 (xi )(1 − Zi )Yi ∑ i w0 (xi )(1 − Zi ) → τh (n → ∞) ̂ τh fX|T=1 (x) × w1 (x) = fX|T=0 (x) × w0 (x) = f(x)h(x) τ(x) = E[Y(1) − Y(0)|X = x] τh τh ≡ ∫ τ(dx)f(x)h(x)μ(dx) ∫ f(x)h(x)μ(dx) τh
  28. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ7 ‣ ͜͜ͰɺɹɹɹɹɹɹɹɹɹΛ༻͍ͨ৔߹ͷॏΈΛPWFSMBQXFJHIUͱ͍͏ɻ ‣ ͜ͷΑ͏ͳॏΈ෇͚͸ɺӈԼਤͷΑ͏ͳ෼෍ͷॏΈΛߏ੒͢Δɻ ‣ ͜Ε͸ɺॲஔ܈ͱରর܈Ͱ͔ͿΓ͕େ͖͍෦෼ʹେ͖ͳॏΈ͕͔͔Γɺαϯ ϓϧ͕ଘࡏ͠ͳ͍෦෼ʹ͸ॏΈΛ͔͚ͳ͍ɻ !39 h(x)

    = e(x){1 − e(x)} fX|T=1 (x) × w1 (x) = fX|T=0 (x) × w0 (x) = f(x)e(x){1 − e(x)} ‣ ͜ͷͱ͖ɺ8"5&͸ɺॲஔΛड͚Δ֬཰͕ ͷαϯϓϧʹରͯ͠࠷΋େ͖ͳॏΈ͕ ͔͔ͬͨҼՌޮՌͷਪఆྔͱͳ͍ͬͯΔɻ ‣ ͜ͷ஋͕े෼େ͖͍͜ͱ͸ɺ޿ࠂΛݟΔՄ ೳੑͷ͋Δਓʹɺ޿ࠂΛଧͭ͜ͱͰɺߪങ ߦಈͷଅਐ͕Ͱ͖Δ͜ͱΛද͢ɻ
  29. Original Covariate Density −2 −1 0 1 2 0.0 0.1

    0.2 0.3 0.4 0.5 Covariate Density −2 −1 0 1 2 0.00 0.10 0.20 0.30 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 ATE:h(x)=1 Covariate Density −2 −1 0 1 2 0.00 0.10 0.20 0.30 Covariate Density −2 −1 0 1 2 0.00 0.10 0.20 0.30 Covariate Density −2 −1 0 1 2 0.00 0.10 0.20 0.30 ATT:h(x)=e(x) Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 ATU:h(x) = 1−e(x) Covariate Density −2 −1 0 1 2 0.0 0.2 0.4 0.6 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 Overlap:h(x) = e(x)(1−e(x)) Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 Matching:h(x) = min{e(x),1−e(x)} Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ7** !41
  30. ಛఆͷੑ࣭ͷ܈ʹண໨͢ΔҼՌతޮՌ7*** !42 ATE Covariate Density −2 −1 0 1 2

    0.00 0.10 0.20 0.30 ATT Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 ATU Covariate Density −2 −1 0 1 2 0.0 0.1 0.2 0.3 0.4 0.5 Overlap Covariate Density −2 −1 0 1 2 0.00 0.10 0.20 0.30 "5&"WFSBHF5SFBUNFOU&GGFDU ͢΂ͯͷαϯϓϧʹॲஔΛߦͬͨ৔߹ͷ݁Ռͱ ͢΂ͯͷαϯϓϧʹॲஔΛߦΘͳ͔ͬͨ৔߹ͷ݁Ռͷൺֱ "55"WFSBHF5SFBUNFOU&GGFDUPOUIF5SFBUFE ॲஔΛड͚ͨਓͷूஂʹண໨ͯ͠ɺ ॲஔΛड͚ͨ৔߹ͱɺॲஔΛड͚ͳ͔ͬͨ৔߹ͷ݁ՌΛൺֱ͢Δ "56"WFSBHF5SFBUNFOU&GGFDUPOUIF6OUSFBUFE ॲஔΛड͚ͳ͔ͬͨਓͷूஂʹண໨ͯ͠ɺ ॲஔΛड͚ͨ৔߹ͱɺॲஔΛड͚ͳ͔ͬͨ৔߹ͷ݁ՌΛൺֱ͢Δ "50"WFSBHF5SFBUNFOU&GGFDUPOUIF0WFSMBQ ॲஔΛड͚Δ֬཰ͱɺॲஔΛड͚ͳ͍֬཰͕͋Δఔ౓͋Δूஂʹண໨ͯ͠ɺ ॲஔΛड͚ͨ৔߹ͱɺॲஔΛड͚ͳ͔ͬͨ৔߹ͷ݁ՌΛൺֱ͢Δ
  31. ໨࣍ ‣ ܏޲είΞ͸ສೳༀͳͷ͔ʂʁ ‣ ϥϯμϜԽൺֱ࣮ݧʢ3BOEPNJ[FE$POUSPM5SJBMʣ ‣ ݹయతճؼϞσϧͷഁ୼ ‣ ೣ΋उࢠ΋܏޲είΞɻ ‣

    ҼՌਪ࿦ͷຊ࣭ͱ࣮σʔλͷղੳ ‣ ܏޲είΞͷٯ਺ͦ͜ຊ࣭ʂʢ1SPQFOTJUZ4DPSFʣ ‣ σʔλ͕ͳ͍ͳΒɺ͹Β͖ͭΛߟ͑Δʢ#PPUTUSBQ.FUIPEʣ ‣ ฏۉॲஔޮՌͰ͸ͳ͘ɺ0WFSMBQΛߟ͑Δʢ0WFSMBQ8FJHIUTʣ ‣ ౷ܭతҼՌਪ࿦ͷ࠷ۙͷൃల ‣ ͋ͷਓͷҼՌޮՌ͸Ͳ͏ͩΖ͏͔ʁʢ$BVTBM'PSFTU3-FBSOFSʣ ‣ ·ͱΊ !43
  32. ݸਓʹର͢ΔҼՌޮՌͷਪఆ* .-º$BVTBMJUZʣ ‣ ۙ೥ɼ4UBOGPSEେֶʹ͓͍ͯػցֶशΛ༻͍ͨ)FUFSPHFOFPVT5SFBUNFOU &⒎FDUͷਪఆ͕׆ൃԽ͍ͯ͠Δɽ ‣ ػցֶशΛ༻͍Α͏ͱ͍͏ྲྀΕ͸ɼҎલ͔Β΋͕͋ͬͨɼ൴Βͷ DPOUSJCVUJPOͷଟ͘͸ʮಘΒΕΔਪఆྔͷ઴ۙతੑ࣭ʯΛ໌Β͔ʹͨ͠ͱ ͍͏఺Ͱ͋Δɽ ‣

    ͭ·Γɼ౷ܭతਪ࿦ԾઆݕఆͳͲ͕ՄೳʹͳΔΑ͏ͳ'SBNFXPSLΛ࡞ Γग़ͨ͠ɽ ‣ Ϟνϕʔγϣϯ͸ʮ1FSTPOBMJ[FE.FEJDJOFʯʢݸผԽҩྍɿͲΜͳਓʹ ͲΜͳॲஔΛ͢Ε͹ɼޮՌ͕࠷େʹͳΔͷ͔Λ஌Γ͍ͨʣ ‣ େ͖͘෼͚Δͱɼ̎ͭͷΞϓϩʔν͕औΒΕ͍ͯΔ ‣ 3BOEPN'PSFTUΛԠ༻ͨ͠ख๏ʂ ‣ ݁Ռม਺܏޲είΞΛ.-3FHSFTTJPOͰਪఆ͔ͯ͠Βɼਪఆͨ͠஋Λ ༻͍ͯճؼͷ໰୊Λ.-Ͱղ͘ํ๏ !44
  33. ݸਓʹର͢ΔҼՌޮՌͷਪఆ** 1SPQFOTJUZ5SFF !45 ॲஔม਺8ͱڞมྔ9ʹ3FHSFTTJPO5SFFΛ౰ͯ͸ ΊΔͱ͍͏໰୊Λߟ͑ΔʢTQMJUDSJUFSJPO͸4- MPHMJLFMJIPPEʣͳͲɽ ͜ͷͱ͖ɼ-FBGͷதʹೖ͍ͬͯΔඪຊ͸ɼ͍ۙ͠ڞ มྔΛ࣋ͭʢͭ·ΓҰछͷ૚ผԽͰ͋Δʣɽ Αͬͯɼ-FBGΛݻఆͨ͠৔߹ʹ͸ɼ-FBGͷதͰ͸ɼ ڧ͘ແࢹͰ͖ΔׂΓ෇͚͕ۙࣅతʹ੒Γཱͭͱߟ͑

    Δ͜ͱ͕Ͱ͖Δɽ ·ͨɼ೚ҙͷ-FBG- 9 ͸઴ۙతʹEJBN - 9 ˠ Ͱ͋Δ͔Βɼ઴ۙతʹ͸9Λݻఆͨ͠΋ͱͰɼҼՌ తޮՌΛਪఆ͍ͯ͠Δ͜ͱʹͳΔɽ 9  9 0          ֤-FBGͷ਺ࣈ͸ɼͦͷ-FBG಺ʹ͋Δαϯϓϧͷ ฏۉͰܭࢉ͞ΕΔॲஔ֬཰Ͱ͋Δɽ E[Y(1)|X ∈ L(X)] − E[Y(1)|X ∈ L(X)] ≈ E[Y(1)|W = 1,X ∈ L] − E[Y(0)|W = 0,X ∈ L] 1SPQFOTJUZUSFF
  34. ݸਓʹର͢ΔҼՌޮՌͷਪఆ*** $BVTBM'PSFTUʣ ‣ 1SPQFOTJUZ5SFFΛCBTFMFBSOFSͱ͢Δ3BOEPN'PSFTUΛʮ$BVTBM'PSFTU 8BHFSBOE"UIFZ  ʯͱ͍͏ɽ ‣ <8BHFSBOE"UIFZ >Ͱ͸ɼ3BOEPN'PSFTUͷ઴ۙਖ਼نੑʹର͢Δূ

    ໌Λ༩͍͑ͯΔʢ਺஋తʹܭࢉՄೳͳਪఆྔ΋ʣɽ ‣ ͜ͷख๏ͷDPOUSJCVUJPO͸େ͖ͭ͘ ‣ ʮ܏޲είΞͷਪఆʯ͕ඞཁͳ͍ʢඞཁͳڞมྔΛಛఆ͢Ε͹ྑ͍ʣɽ ‣ ਪఆ͞ΕΔҼՌతޮՌ͸ʮ)FUFSPHFOFPVTʯͰ͋Γɼڞมྔ͕༩͑ΒΕ ͨݩͰͷҼՌతޮՌͷਪఆ͕Մೳ ‣ ͲΜͳਓʹରͯ͠ɼͲΕ͘Β͍ͷޮՌ͕͋Δ͔͕Θ͔Δʂ ‣ ࿦จʹ͸ɼݟࣄͳγϛϡϨʔγϣϯ݁Ռ͕ࡌ͍ͬͯΔɽ ‣ ߹Θͤͯɼศརͳख๏ʹ͸ɼ૬Ԡͷܽ఺͕ଘࡏ͢ΔͷͰɼQBDLBHF\HSG^Ͱݕ ূͯ͠΄͍͠ɽ !46 E[Y(1) − Y(0)|X = x]
  35. ݸਓʹର͢ΔҼՌޮՌͷਪఆ*7 QBDLBHF\HSG^ʣ ‣ 4BNQMFTJ[F/ ‣ /VNCFSPGGFBUVSFTQ ‣ ҼՌޮՌͷؔ਺͸ɼ͕ΑΓେ͖͍৔߹ʹͷΈଘࡏ͠ɼͦͷӨڹ͸ઢܕత Ͱ͋ΔΑ͏ͳઃఆͰ͋Δɽ !47

    Xj ∼ N(0,1) j = 1,2,...,p W ∼ Bernoulli (0.4 + 0.2 × I(X1 > 0)) Y = W × max(X1 ,0) + X[,2] + min(X3 ,0) + N(0,1) X1 $BVTBMF⒎FDU X1 / &TUJNBUFE 5SVF&⒎FDU $*
  36. ݸਓʹର͢ΔҼՌޮՌͷਪఆ7 3-FBSOFSʣ ‣ 3-FBSOFS͸΋ͱ΋ͱʢ/JFBOE8BHFS ʣͰ̎஋ͷॲஔʹରͯ͠ఏҊ ͞Εͨख๏ɽ ‣ &<:c9>ͱ&<8c9>ΛͦΕͧΕɼద౰ͳճؼϞσϧͰਪఆͨ͠ޙͰɼ࣍ͷํఔ ࣜͷղΛ-PDBM-JOFBS'PSFTU 'SJFECFSH

    FUBM  Ͱղ͘ɽ ‣ 9͕༩͑ΒΕͨ΋ͱͰͷɼඇઢܗͳҼՌޮՌ͕ਪఆՄೳͱͳΔɽ ‣ ࠷ۙͷ3BOEPN'PSFTUपΓͷൃల͸͍͢͝ʂ ‣ (FOFSBMJ[FE3BOEPN'PSFTU "UIFZ 5JCTIJSBOJ BOE8BHFS  ͸ 'PSFTUͰํఔࣜΛղ͘͜ͱ͕Ͱ͖Δʢ*OTUSVNFOUBMWBSJBCMF΍ɼRVBOUJMF SFHSFTTJPOͳͲ΋Ͱ͖Δɽʣɽ !48 Y − ̂ E[Y|X] = (W − ̂ E[W|X])τ(X) + ε
  37. ݸਓʹର͢ΔҼՌޮՌͷਪఆ7*ʢͦͷଞͷ࿩୊ʣ ‣ $BVTBMJUZº5SBOTGFS-FBSOJOH͕࢝·ͬͨɽ ‣ ,VO[FMFUBM  5SBOTGFS-FBSOJOHGPS&TUJNBUJOH$BVTBM &⒎FDUTVTJOH/FVSBM/FUXPSLT BS9JWW ‣

    ҼՌਪ࿦ʹసҠֶशΛద༻͢Δͱ͍͏ϑϨʔϜϫʔΫ͸ɼ͜Ε͔ΒͷτϨ ϯυʹͳΔՄೳੑ͕େ͖͍ɽ ‣ ·ͣ͸ɼ%PNBJO"EBQUJPO͔Β࢝·ΔͱࢥΘΕ·͢ɽ !49
  38. ໨࣍ ‣ ܏޲είΞ͸ສೳༀͳͷ͔ʂʁ ‣ ϥϯμϜԽൺֱ࣮ݧʢ3BOEPNJ[FE$POUSPM5SJBMʣ ‣ ݹయతճؼϞσϧͷഁ୼ ‣ ೣ΋उࢠ΋܏޲είΞɻ ‣

    ҼՌਪ࿦ͷຊ࣭ͱ࣮σʔλͷղੳ ‣ ܏޲είΞͷٯ਺ͦ͜ຊ࣭ʂʢ1SPQFOTJUZ4DPSFʣ ‣ σʔλ͕ͳ͍ͳΒɺ͹Β͖ͭΛߟ͑Δʢ#PPUTUSBQ.FUIPEʣ ‣ ฏۉॲஔޮՌͰ͸ͳ͘ɺ0WFSMBQΛߟ͑Δʢ0WFSMBQ8FJHIUTʣ ‣ ౷ܭతҼՌਪ࿦ͷ࠷ۙͷൃల ‣ ͋ͷਓͷҼՌޮՌ͸Ͳ͏ͩΖ͏͔ʁʢ$BVTBM'PSFTU3-FBSOFSʣ ‣ ·ͱΊ !50
  39. ·ͱΊ ‣ ਪఆͨ͠܏޲είΞͷଥ౰ੑͷνΣοΫ͸ʮڞมྔͷ௼Γ߹͍ΛݟΔ͜ͱʯ Ͱߦ͏͜ͱ͕Ͱ͖Δɻ ‣ "6$ͳͲͷ౷ܭతج४͸ɺ͜Εʹ୅ସͰ͖ͳ͍ɻ ‣ *18Λ࢝Ίͱ͢ΔҼՌޮՌʹର͢Δਪఆྔ͸ɺʮܽଌσʔλʹର͢Δ౷ܭత ਪ࿦ʯͰ͋Γɺඇৗʹෆ҆ఆͳ৔߹͕͋ΔͷͰɺͦͷਪఆྔͷ͹Β͖ͭ͸ඞ ͣซه͢΂͖Ͱ͋Δɻ

    ‣ #PPUTUSBQΛ༻͍ͯܭࢉػతʹܭࢉͰ͖Δ͠ɺਤ΋ॻ͚ΔͷͰ௚ײతʹΘ ͔Γ΍͍͢ɻ ‣ ҼՌਪ࿦Λߦ͏৔߹ʹ͸ʮͲΜͳҼՌޮՌΛ஌Γ͍͔ͨʯΛߟ͑Δ΄͏͕ɺ ਪఆ݁ՌΛԠ༻͢Δࡍʹɺ༗༻Ͱ͋Δɻ ‣ #BMBODJOH$PWBSJBUFͷߟ͑ํΛ༻͍Δͱ͜ΕΛ࣮ݱͰ͖ɺ0WFSMBQʹͭ ͍ͯߟ͑Δ͜ͱ͸ɺ޿ࠂͳͲͰ͸໾ཱͪͦ͏Ͱ͋Δʂ !51
  40. ࢀߟจݙ ‣ &GSPO # BOE)BTUJF 5  $PNQVUFS"HF4UBUJTUJDBM*OGFSFODF $BNCSJEHF 6OJWFSTJUZ1SFTT

    ‣ *NCFOT ( BOE3VCJO %  $BVTBM*OGFSFODFGPS4UBUJTUJDT 4PDJBM BOE #JPNFEJDBM4DJFODFT $BNCSJEHF6OJWFSTJUZ1SFTT ‣ -J ' .PSHBO , BOE;BTMBWTLZ "  #BMBODJOH$PWBSJBUFTWJB1SPQFOTJUZ 4DPSF8FJHIUJOH +"4" ‣ *NBJ , BOE3BULPWJD .  $PWBSJBUF#BMBODJOH1SPQFOTJUZ4DPSF+344# ‣ "UIFZ 4 5JCTIJSBOJ + BOE8BHFS 4  (FOFSBMJ[FE3BOEPN'PSFTU 5IF "OOBMTPG4UBUJTUJDT    ‣ /JF 9 BOE8BHFS 4  2VBTJ0SBDMF&TUJNBUJPOPG)FUFSPHFOFPVT5SFBUNFOU &⒎FDUT BS9JWW ‣ 8BHFS 4 BOE"UIFZ 4  &TUJNBUJPOBOE*OGFSFODFPG)FUFSPHFOFPVT 5SFBUNFOU&⒎FDUTVTJOH3BOEPN'PSFTU +"4"    ‣ ,VO[FMFUBM   .FUB-FBSOFSTGPSFTUJNBUJOH)FUFSPHFOFPVT5SFBUNFOU &⒎FDUTVTJOH.BDIJOF-FBSOJOHBS9JWW !52