医学統計セミナー第4回 傾向スコア分析

1 4 - - ( ) ( vs. )Outcome( ) ( ) ( ) (RCT:RandomizedControlled Trial) ( ) ( ) ( Outcome ) (Propensity score) 2018 12 6 7 6 11 13 25 40 46 66 88 142 169 231 325 369 539 655 897 1,108 1,458 1,925 2,343 2,964 3,949 05001,0001,5002,0002,5003,0003,5004,0004 ,500199519961997199819992000200120022003 2004200520062007200820092010201120122013 20142015201620172018 PubMed propensity score ( 2006)(1) 2 (2) (3) ( ) ( )

2 (propensity score) 55(3), 230-243, 2006. et al., NEJM, 2005, 353, 49-61. Revised Cardiac Risk Index RCRI (1) (2) (3) (4) (5) (6) Neyman-Rubin 1|iy0|iy i |0|1E[E[]]iiyy .. i 1|iy0|iy i 1|iy 0|iy ITE z=0 z=1 (z ) z=1 z=0 |1E[]iy|0E[]iy1||0]E[iyz=0||0]E[iyz=1||] 1E[ziy=0||]1E[ziy= |0|1E[E[]]iiyy z=1 z=0 R2 1||]1E[ziy=|1E[]iy 0||0]E[iyz= |0E[]iy (Leite, 2017) (Treatment Predictor) (Treatment) (Outcome) (Outcome Predictor) (True Confounder) (Mediator) (Brookhart et al.)

3 , 2006) ( ) Leite, W: Practical Propensity Score Methods using R, SAGE, , et al.: variable selection for propensity score models, American Journal of Epidemiology, 163(12), 1149-1156. z=1 z=0 1||]1E[ziy=|1E[]iy 0||0]E[iyz= |0E[]iy PS 1 ( c ( ) z=1 z=0 1||0]E[iyz=0||0]E[iyz=1||]1E[ziy=0||]1E[ ziy= |1E[]iy|0E[]iyATEATT (ATE:AverageTreatment Effect) (ACE:Average Causal Effect) 1|0|]ATE=E[E][iiyy (ATT: Average Treatment effectfor Treated) |0|11 ATT=E[E] |]1|[iiZyyZ== (Leite, 2017)Step 1.

4 2. Random Forests 3. IPW( )4. 5. 6. Rosenbaum(2002) Carnegie et al.(2016) Pr(1 |)i iiez==x iix 1ie ei (Propensity score) ( ) Annals of Surgery, 257(4), ,wefittedalogisticregressionmodelforther eceiptoflaparoscopicgastrectomyasa functionofpatientdemographicandhospitalf actorsincludingage,sex,Charlsoncomorbidi tyindex,bodymassindex,smokingindex,cance rstage(IorII),hospitalvolumecategory,and typeofhospital(teachingornonteaching).

5 TheC-statisticforevaluatingthegoodnessof fitwascalculated. Eachpatientwhoreceivedlaparoscopicgastre ctomywasmatchedwitha patientwhoreceivedopengastrectomywiththe closestestimatedpropensityonthelogitscal ewithina specifiedrange( ofthepooledstandarddeviationofestimatedl ogits)toreducedifferencesbetweentreatmen tgroupsbyat least90%. ( ) vs. hospital volume(1 ) BMI (I/II) (Teaching hospital or Nonteaching hospital) (propensity score) ( ) (propensity score) ( 0 1 ) x z ( ( ) ) ( )

6 ( ) ( ) ( ) (5 ) (IPWE: Inverse Probability Weighting Estimator) (5 ) (nearestneighbor matching) ( ) (optimal matching) 1 1 1 1 (1 k )1 k 1 ( ) ( )

7 ( ) ( ) (Rosenbaum,1989) (2015) 1:1 [one-to-one matching]1 1 (Cohen, 1988) (1:k) [fixed rate matching, one-to-k matching]1 k (Leite, 2017) [ variable rate matching, one-to-many matching]1 ( ) (Leite, 2017) 1:1 (Cepedaet al, 2003; Gu& Rosenbaum, 1993.)

8 Ming & Rosenbaum, 2000) 1:1 ( k=2) ( ) ( ) [Nearest neighbor] [Nearest neighbor within caliper] ( ) ( ) Rosenbaum& Rubin(1985) SD ( 020 ) (Optimal) = (Optimal matching) PS (Genetic matching) (full matching) We performed a one-to-one matching analysis between laparoscopicand open distal gastrectomy groups on the basis of estimatedpropensity scores of each patient.

9 P Lung Cancer, 120, 88-90, 2018 (WBRT) (SRS) ( )Propensity score matching (PSM) wasalso performed. Ten-to-one matching without replacement was completed using the nearest neighbor match on the logit of the propensity score for treatment approach with caliper width set to times the standard deviation of the logit of the propensity score. ( ) VS 2 t 2 McNemar 2 t 2 2 ( ) (Austin, 2008) ( , ICR-Web) (Hill & Stuart, 2008; Ho et al.)

10 , 2007, Schafer and Kang(2998), (2015)) ( 2 ) (WBRT) (SRS) vs. Wilcoxon p (Yang & Dalton, 2012) (Standardizeddifferencescore) ()10cont22102xxsds+ = 95% 102210::::ssxx 2 ()10bin1 10 0(1 2 )(1)ppdp pp p + = 10:: pp 210101 22() +++ J VascSurg Venous LymphatDisord. 5(2),171-176. 2017 P A standardizeddifference (Std diff) of < suggests adequate variable balance after propensity matching (1) (2) (3) (1) (2) ( ) (3) ( ) 110 71 BMI 01(a) [p = ] 01(c) [p = ] 010203040506070 (b) [p = ]051015202530 (d)BMI[p = ]024681012 (e)