Unboundedly

統計的因果推論・疫学についてのお話

【点と矢印で因果関係を考える】因果関係がないときにデータから関連が生じるパターンとその対策まとめ:因果ダイアグラム(DAG)によるバイアスの視覚的整理


DAG*1DAG

DAG使DAG

f:id:KRSK_phs:20180829050845p:plain



1 



 DAG



1. 

2. 使
 




2



1調loss to follow up

2missing data

3self-selection



IPW

Multiple Imputation)
 




3

4








使

XY(Association)




(Causal Relationship)

*2*3



AssociationX)56Y)



*4

1 


confoundingconfounderXY

Z
1. X(Association between X and Z)
2. Y(Association between Y and Z)
3. X (Not a downstream consequence of X)*5

使DAG

 DAG


DAGBackdoor path)2backdoor pathAssociation使

DAG

f:id:KRSK_phs:20170320012009p:plainf:id:KRSK_phs:20170320013301p:plainf:id:KRSK_phs:20170320015346p:plain
1ZXYDAGXYbackdoor pathXYAssociationXYZbackdoor path4ZZXYXZ


backdoor pathDAGXYZZ-XZ-YXYY-Z-Xbackdoor path

23UUXYbackdoor pathXYAssociationUZbackdoor pathZZ

使DAG使

f:id:KRSK_phs:20170320021810p:plain
ZXYZU1U2collider)backdoor pathZ3

ZXU1backdoor path)YU2backdoor pathXZ調

ZZU1U2colliderZbackdoor path4ZXYDAGBackdoor path


Association)*6

使23
1. 

XY

subject-matter knowledgeXY

Conditional ExchangeabilityAssociation)


XYmagnitudeAssociationAssociation

1.1 

Association)backdoor pathDAGbackdoor path4Associationbackdoor path

f:id:KRSK_phs:20170123141447p:plain
調RegressionStratification)Limitation)Propensity Scoreg-estimationstructural nested model

oror

Conditional Exchangeabilityconditional on confoundersmarginal)



使調heterogeneous effect)ADA調A

1.2 marginal

DAGDAGbackdoor path調backdoor pathassociation = causal effectDAG

Standardization, Inverse Probability Weighting: IPWMarginal Structural Model

conditionalmarginal調*7


調Marginal
2. 使

*8

使Instrumental variable: IV)Regression Discontinuity Design

DAG

1
使X)使
2

3
Local Average Causal Effectpopulation)(


使使
使




2


Selection bias)*9Association

DAGDAGXYZColliderXYAssociationXYAssociationDAGZpathXYAssociation4Association

f:id:KRSK_phs:20170125023018p:plainf:id:KRSK_phs:20170125023033p:plain
*10


20

1調loss to follow up


調調調調調調調censoringDAGCXY

f:id:KRSK_phs:20170322152306p:plain
LUXXCLC)UL, UY)

調CensoringCensoringCensoring

CXLCpathXYbackdoor pathDAGXYAssociation

2missing data


BMI)X尿Y

f:id:KRSK_phs:20170322154316p:plain
LU尿調XCLC尿UL尿UY)

1XYC1DAGXYBackdoor path

3self-selection


f:id:KRSK_phs:20170322155510p:plain
調C調CYU調UC)UY)

C



IPW

調IPWMarginalDAG

調LLIPW使調
Multiple Imputation)

使

L使使

3


X)Y)Association)

XY)

YX)

f:id:KRSK_phs:20170123135812p:plain
XY

XY調XY*11

4


10





structural bias)

使measurement errorXYXY

XYXY


4DAGDAG



DAGJudea Pearl
Causal Inference in Statistics: A Primer

Causal Inference in Statistics: A Primer

 

ちなみに彼は最近、一般向けの因果推論の本を出版しました。因果推論の歴史を平易な言葉で語っているので、読み物として大変面白いです

The Book of Why: The New Science of Cause and Effect

The Book of Why: The New Science of Cause and Effect

 

ハーバード大学教授Miguel Hernanらによる因果推論の教科書(無料オンライン公開中):Causal Inference Book | Miguel Hernan | Harvard T.H. Chan School of Public Health

Hernán, Miguel A., Sonia Hernández-Díaz, and James M. Robins. "A structural approach to selection bias." Epidemiology 15.5 (2004): 615-625.(選択バイアスのDAGによる整理)

http://journals.lww.com/epidem/Abstract/2004/09000/A_Structural_Approach_to_Selection_Bias.20.aspx

HernanによるDAGの無料オンラインコース

Pearce, Neil, and Debbie A. Lawlor. "Causal inference-so much more than statistics." International journal of epidemiology (2017).(DAGに関する解説)

https://academic.oup.com/ije/article-lookup/doi/10.1093/ije/dyw328

Krieger, Nancy, and George Davey Smith. "The tale wagged by the DAG: broadening the scope of causal inference and explanation for epidemiology." International Journal of Epidemiology (2016): dyw114.(DAGや反事実モデルの乱用によって起きうる因果推論上の問題について)

https://academic.oup.com/ije/article-abstract/doi/10.1093/ije/dyw114/2617188/The-tale-wagged-by-the-DAG-broadening-the-scope-of?redirectedFrom=fulltext

林さんのブログ(私とは違う分野ですが、ものすごく丁寧にまとめてあると感じました):因果関係がないのに相関関係があらわれる4つのケースをまとめてみたよ(質問テンプレート付き) - Take a Risk:林岳彦の研究メモ

 


*1:(Correlation)(Association)

*2:

*3:average causal effectAD/

*4:DAG

*5:3DAGXL

*6:

*7:

*8:endogeneity)使

*9:RDD)使

*10:

*11: