G-methods for time-varying treatments (Causal inference: What if, Chapter 21-1)

Keywords: 因果推論, Time-varying, G-formula, IP weighting, Doubly robust estimation

Transcript

1. View Slide

2. ・G-methods for time-fixed treatments
   本日の内容
   ・The g-formula for time-varying treatments
   ・IP weighting for time-varying treatments
   ・A doubly robust estimator for time-varying
   treatments
   

   View Slide

3. ・G-methods for time-fixed treatments
   本日の内容
   ・The g-formula for time-varying treatments
   ・IP weighting for time-varying treatments
   ・A doubly robust estimator for time-varying
   treatments
   

   View Slide

4. Stratification
   effect measure modification (-)
   effect measure modification (+)
   Mantel-Haenszel method
   別々にオッズ比を報告
   

   View Slide

5. Why model?
   effect measure modification (-)
   effect measure modification (+)
   別々にオッズ比を報告（1つの効果を報告できない）
   g-methods
   

   View Slide

6. g-formula
   A=1を代入
   A=0を代入
   

   View Slide

7. IP weighting
   marginal structural model
   

   View Slide

8. Conditional or Marginal?
   outcome regression
   saturated parametric
   stratification
   g-formula
   IP weighting
   g-estimation
   or
   algebraically equivalent
   

   View Slide

9. Time-varying treatment
   g-methods
   

   View Slide

10. ・G-methods for time-fixed treatments
    本日の内容
    ・The g-formula for time-varying treatments
    ・IP weighting for time-varying treatments
    ・A doubly robust estimator for time-varying
    treatments
    

    View Slide

11. 前提
    ・本章ではidentifiability conditions(sequential
    exchangeability, positivity, and consistency)のviolationが
    ないものとする。
    ・static treatment strategies (always treat vs. never treat)
    の効果を推定する。
    

    View Slide

12. g-formula (weighted average)
    ・time-fixed treatment (A1
    の反実アウトカム)
    ・time-varying treatment
    

    View Slide

13. g-formula (weighted average)
    

    View Slide

14. g-formula (weighted average)
    

    View Slide

15. g-formula (simulation)
    のシミュレーション
    と
    

    View Slide

16. g-formulaの注意点
    ・DAGに基づいたcovariates L1
    をモデルに含める
    ・static sequential exchangeabilityが成立すればstatic
    treatment strategyの効果はidentify可能
    

    View Slide

17. g-formulaの一般化
    ・static treatment strategy
    ・dynamic treatment strategy
    linear regression logistic regression
    

    View Slide

18. ・G-methods for time-fixed treatments
    本日の内容
    ・The g-formula for time-varying treatments
    ・IP weighting for time-varying treatments
    ・A doubly robust estimator for time-varying
    treatments
    

    View Slide

19. IP weighting (weights)
    ・nonstabilized IP weights
    ・ stabilized IP weights
    

    View Slide

20. IP weighting (non-stabilized)
    

    View Slide

21. Stabilized weights
    non-stabilized weights: stabilized weights:
    Lと独立であればよい Lと独立であればよい
    

    View Slide

22. IP weighting (stabilized)
    

    View Slide

23. IP weightingの一般化
    ・nonstabilized IP weights
    ・ stabilized IP weights
    logistic regression
    logistic regression
    （misspecifiedでも可）
    

    View Slide

24. Marginal Structural Model
    ・2K > Nのときは推定できない
    ・marginal structural mean model
    stabilized IP weightsを使って推定
    misspecified??
    

    View Slide

25. Effect Measure Modification
    ・baseline variable VによるEMMがある場合、marginal
    structural modelは以下の通り（parametric）
    stabilized IP weightsを使って推定
    Vに入れて良いのはbaseline variableだけ
    

    View Slide

26. ・G-methods for time-fixed treatments
    本日の内容
    ・The g-formula for time-varying treatments
    ・IP weighting for time-varying treatments
    ・A doubly robust estimator for time-varying
    treatments
    

    View Slide

27. Doubly Robust Estimator
    ・g-formula
    ・ IP weighting
    

    View Slide

28. 1.
    Doubly Robust (time-fixed)
    2.
    3. A=1とA=0でそれぞれ
    を推定
    を推定
    ,
    をLについて標準化
    

    View Slide

29. 1.
    Doubly Robust (time-varying)
    2.
    3.
    を推定
    からパラメータ
    を求める。
    を求めておく
    を推定し、Aの値に応じた を求める。
    これを繰り返して を求める。
    always treat
    

    View Slide

30. Discussion
    

    View Slide