Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
セミパラメトリック推論の基礎の復習
Search
Daisuke Yoneoka
November 14, 2023
Research
0
19
セミパラメトリック推論の基礎の復習
Daisuke Yoneoka
November 14, 2023
Tweet
Share
More Decks by Daisuke Yoneoka
See All by Daisuke Yoneoka
感染症の数理モデル4
kingqwert
0
27
感染症の数理モデル3
kingqwert
0
43
感染症の数理モデル2
kingqwert
0
51
感染症の数理モデル1
kingqwert
0
79
媒介分析と疫学
kingqwert
0
140
時系列解析と疫学
kingqwert
2
950
Supervised PCAとその周辺
kingqwert
1
25
Sparse linear models
kingqwert
0
16
SKAT for rare common variants
kingqwert
0
15
Other Decks in Research
See All in Research
【論文紹介】Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions
ichiroex
0
110
機械学習と最適化の融合動的ロットサイズ決定問題を例として
mickey_kubo
2
300
LLMマルチエージェントを俯瞰する
masatoto
26
17k
[2023 CCSE] ZOZOTOWN検索における 研究開発の取り組みについて
tomoyayama
0
130
継続的な研究費獲得のための考え方
moda0
2
470
Engineering LaCAM∗: Towards Real-Time, Large-Scale, and Near-Optimal Multi-Agent Pathfinding
kei18
0
110
Deep State Space Models 101 / Mamba
kurita
9
3.8k
第12回全日本コンピュータビジョン勉強会:画像の自己教師あり学習における大規模データセット
naok615
0
540
新入生向けチュートリアル:文献のサーベイv2
a1da4
9
6.6k
Alexander Mielke Hellinger--Kantorovich (a.k.a. Wasserstein-Fisher-Rao) Spaces and Gradient Flows
jjzhu
3
210
Generative AI - practice and theory
gpeyre
1
620
Rの機械学習フレームワークの紹介〜tidymodelsを中心に〜 / machine_learning_with_r2024
s_uryu
0
260
Featured
See All Featured
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
188
16k
Fantastic passwords and where to find them - at NoRuKo
philnash
39
2.5k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
660
120k
Designing for Performance
lara
601
67k
Infographics Made Easy
chrislema
238
18k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
358
22k
How STYLIGHT went responsive
nonsquared
92
4.9k
We Have a Design System, Now What?
morganepeng
43
6.8k
Learning to Love Humans: Emotional Interface Design
aarron
267
39k
Navigating Team Friction
lara
179
13k
VelocityConf: Rendering Performance Case Studies
addyosmani
321
23k
Reflections from 52 weeks, 52 projects
jeffersonlam
345
19k
Transcript
ηϛύϥϝτϦοΫਪͷجૅͷ෮श Daisuke Yoneoka September 29, 2014
Notations جຊతʹ Tsiatis,2006 ʹै͏. Θ͔Μͳ͔ͬͨΒࣗͰௐͯͶ! ϕΫτϧߦྻଠࣈʹͯ͠ͳ͍͚Ͳ, ͦࣗ͜Ͱิ͍ͬͯͩ͘͞. σʔλ i.i.d Ͱ
Zi = (Zi1, . . . , Zim) ∈ Rm αϯϓϧαΠζ n ਓ. i.e., Z1, . . . , Zn φ(Z) Өڹؔ u(Zi, θ) ਪఆؔ Լ͖ࣈͷ eff (ۙ) ༗ޮ (efficient) ͱ͍͏ҙຯ
ηϛύϥϝτϦοΫਪͱʁ Zi ͷີ͕ؔηϛύϥϝτϦοΫϞσϧʹै͏ͱ S = {p(z : θ, η)|θ ∈
Θ ⊂ Rr, η ∈ H} θ ༗ݶ࣍ݩͷڵຯ͋ΔύϥϝλͰ, η ແݶ࣍ݩͷͲ͏Ͱ͍͍ύ ϥϝλ (ہ֎ (nuisance) ύϥϝʔλʔ). ηϛύϥϝτϦοΫਪ: ͜ͷͱͰ θ ͷ࠷ྑͷਪఆྔ (RAL ਪఆ ྔ) ΛͱΊΔ͜ͱ
Өڹؔ θ ͳΜͰ͍͍͔Β࠷ྑΛݟ͚ͭΔͱ͍͏ͷແཧήʔ → Ϋϥε Λݶఆͯͦ͜͠Ͱݟ͚ͭΔ! (౷ܭͰΑ͘ΔΑͶ) Өڹؔ: ਪఆྔ ˆ
θ ͷӨڹؔͱ, (Ϟʔϝϯτʹ੍͕͋Δ) √ n(ˆ θ − θ) = 1 √ n n i=1 φ(Zi, θ, η) + op(1) Λຬͨ͢ϕΫτϧؔ. ˆ θ ۙઢܗਪఆྔͱݺͼ n → ∞ ͰҰகੑ ͱۙਖ਼نੑ͕͋Δ √ n(ˆ θ − θ) → N 0, E[φ(Zi, θ, η)φ(Zi, θ, η)T ] Πϝʔδతʹ͋Δσʔλ͕ͲΕ͚ͩਪఆʹӨڹΛ༩͍͑ͯΔ͔Λ දݱͨ͠ͷ
ਪఆؔͱ M ਪఆ ਪఆํఔࣜ n i=1 u(Zi, θ) ਪఆؔ =
0 ͷղͱͯ͠ಘΒΕΔͷΛ M ਪఆྔ ͱݺͿ. Α͘ݟΔ score ؔͳΜ͔ίϨ. ͨͩ͠, E[φ(Zi, θ)] = 0 ظ 0 , E[∥φ(Zi, θ)∥2] < ∞ ࢄతͳͷൃࢄ͠ͳ͍ . ͋ͱ͏গ͚ͩ݅͋͠Δ. Ұகੑͱۙਖ਼نੑΛ࣋ͭ √ n(ˆ θ − θ) = 1 √ n n i=1 E[ ∂u(Zi, θ) ∂θ ] −1 u(Zi, θ) ͕͜͜Өڹؔʹͳ͍ͬͯΔ +op(1) → N 0, E[ ∂u(Zi, θ) ∂θ ] −1 E[u(Zi, θ)u(Zi, θ)T ] E[ ∂u(Zi, θ) ∂θ ] −T ] ͜ͷۙࢄͷਪఆྔΛαϯυΠονਪఆྔͱݺΜͩΓ͢Δ
RAL ਪఆྔ ۙઢܥਪఆྔͳΜ͔ྑͦ͞͏ʂͰ super efficiency ͷ (Hodges) ͕Δʂ Super efficiency:
ۙతʹ Cramer-Rao ͷԼݶΑΓྑ͍ͷ͕Ͱ͖ Δͷ͜ͱ ͜ͷΛղܾͨ͠ͷ͕ RAL (Regular asymptotic linear) ਪఆྔ. ͦͷਖ਼ଇ݅ۃݶ͕ LDGP (local data generating process) ʹґ ଘ͠ͳ͍͜ͱ (ৄ͘͠ Tsiatis, 2006) ηϛύϥਪ͜ͷ RAL ਪఆྔͷӨڹؔΛٻΊΔ͜ͱΛߟ͑Δ
Parametric submodel ηϛύϥϝτϦοΫϞσϧ S ͷ֤ʹର͠ p(z; θ, η) ∈ Ssub
⊂ S Λຬͨ͢ύϥϝτϦοΫϞσϧ Ssub = {p(z; θ, γ)|θ ∈ Θ ⊂ Rr, γ ∈ Γ ⊂ Rs, s ∈ N} ΛύϥϝτϦοΫαϒϞσϧͱݺͿ.
Nuisance tangent space (ہ֎ۭؒ) ηϛύϥϝτϦοΫϞσϧ S ͷ֤ʹର͠, ύϥϝτϦοΫαϒϞσϧ Ssub ͷہ֎ۭؒΛ
TN θ,γ (Ssub) = {BT sγ(z, θ, γ)|B ∈ Rs} ͱ͢Δ. γ p(z; θ, η) ʹରԠ͢ΔͷͰ sγ(z, θ, γ) = ∂ ∂γ log p(z; θ, γ) Ͱ ද͞ΕΔ nuisance score ؔ. ͜ͷઢܗۭؒ͜ͷ nuisance score vector ʹ ΑͬͯுΒΕ͍ͯΔ. ͜ͷͱ͖ TN θ,η (S) = Ssub TN θ,γ (Ssub) Λ S ্ͷ p(z; θ, η) ʹ͓͚Δہ֎ۭؒͱΑͿ. ͪͳΈʹ, ଆͷू ߹ʹؔͯ͠ closure ΛͱΔԋࢉࢠ. Note:͜ͷۭؒେͰޙʹ, RAL ਪఆྔͷӨڹؔ͜ͷۭؒʹަۭͨؒ͠ʹ ଐ͢Δ͜ͱ͕ॏཁʹͳͬͯ͘Δʂ
ઢܗ෦ۭؒͷࣹӨͷزԿͱϐλΰϥεͷఆཧ
RAL ਪఆྔͷӨڹؔͷॏཁͳఆཧ ηϛύϥϝτϦοΫ RAL ਪఆྔ β ͷӨڹؔ φ(Z) ҎԼͷ݅Λຬ ͢Δ.
Corollary1 E[φ(Z)sβ] = E[φ(Z)sT efficient (Z, β0, η0)] = I. ͨͩ͠, s είΞؔͰ, sT efficient ༗ޮείΞؔ Corollary2 φ(Z) ہ֎ۭؒʹަ͍ͯ͠Δ. ༗ޮӨڹ্ؔͷ 2 ͭͷ݅Λຬͨ͠, ͦͷࢄߦྻ, ޮݶքΛୡ ͦ͠Ε φeffi(Z, β0, η0) = E[seff (Z, β0, η0)sT eff (Z, β0, η0)] −1 seff (Z, β0, η0)
ηϛύϥۭؒͷఆཧ ύϥϝτϦοΫαϒϞσϧͷ߹ͷ RAL ਪఆྔͷӨڹؔͱۭؒͱͷؔ Tsiatis, 2006 ͷ Ch4.3 ͋ͨΓΛݟͯͶʂ ఆཧ
1 RAL ਪఆྔͷӨڹؔ {φ(Z) + TN θ,η (S)⊥} ͱ͍͏ۭؒʹؚ·ΕΔ. ͨͩ͠, φ(Z) ҙͷ RAL ਪఆྔͷӨڹؔͰ, TN θ,η (S)⊥ ηϛύϥϝτϦο Ϋۭؒͷަิۭؒ ఆཧ 2 ηϛύϥϝτϦοΫ༗ޮͳਪఆྔ, ͦͷӨڹ͕ؔҰҙʹ well-defined Ͱܾఆ͞ Ε,φefficient = φ(Z) − {φ(Z)|TN θ,η (S)⊥} ͷཁૉ. ͪͳΈʹ, (h|U) projection of h ∈ H(ੵΛಋೖͨ͠ώϧϕϧτۭؒ) onto the space U (ઢܗۭؒ)
GEE ʹ͍ͭͯͷ Remarks Liang-Zeger ͷ GEE ͷηϛύϥϝτϦοΫϞσϧ (੍ϞʔϝϯτϞσϧ: 1 ࣍ͱ
2 ࣍ͷϞʔϝϯτʹ੍͚ͩΛஔ͍ͨϞσϧ) ҎԼͷಛΛͭ. ہॴ (ۙ༗) ޮਪఆྔ: ࢄؔͷԾఆ͕ਖ਼͚͠Ε, ༗ޮਪఆྔ Robustness: ແݶ࣍ݩͷύϥϝʔλਪఆ͕ඞཁ͕ͩ, ࢄؔΛ misspecify ͨ͠ͱͯ͠Ұகੑͱۙਖ਼نੑอ࣋ GEE ͷຊΛಡΊΘ͔Δ͚Ͳ, Working covariance matrix Λؒҧ͑ͯ ༗ޮੑࣦΘΕΔ͕, ͦͷଞͷ·͍͠ੑ࣭ (ۙਖ਼نੑͱҰகੑ) อ࣋Ͱ͖Δͬͯ͜ͱ