High Dimensional Propensity Score Estimation via Covariate Balancing

Transcription

1 High Dimensional Propensity Score Estimation via Covariate Balancing Kosuke Imai Princeton University Talk at Columbia University May 13, 2017 Joint work with Yang Ning and Sida Peng Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 1 / 20

2 Motivation Covariate adjustment in observational studies: 1 Identify and measure a set of potential confounders 2 Weight or match the treated and control units to make them similar Practical questions: 1 Which covariates should we adjust? 2 How should we adjust them? 3 What should we do if we have many potential confounders? Major advances over the last several years in the literature: Tan, Hainmueller, Graham et al., Zubizarreta, Belloni et al., Chan et al., Farrell, Chernozhukov et al., Athey et al., Zhao, etc. Covariate Balancing Propensity Score (CBPS) Imai & Ratkovic JRSSB; Fan et al. Estimate propensity score such that covariates are balanced Extensions: continuous treatment (Fong et al.), dynamic treatment (Imai & Ratkovic JASA), high-dimensional setting (Ning et al.) Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 2 / 20

3 Overview 1 Review of propensity score and CBPS Inverse probability weighting (IPW) Propensity score tautology covariate balancing Optimal covariate balancing for CBPS 2 High-dimensional CBPS (HD-CBPS) Impossible to balance all covariates trade-off between covariates Sparsity assumption + Regularization Balance covariates that are predictive of outcome Remove bias (due to regularization) by covariate balancing Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 3 / 20

4 Propensity Score and CBPS T i {0, 1}: binary treatment X i : d-dimensional pre-treatment covariates Identification assumptions: SUTVA, overlap, unconfoundedness Dual characteristics of propensity score: 1 Predicts treatment assignment: π(x i ) = Pr(T i = 1 X i ) 2 Balances covariates: { } Ti E π(x i ) f (X i) { } 1 Ti = E 1 π(x i ) f (X i) = E{f (X i )} Propensity score tautology CBPS: estimate propensity score such that balance is optimized Covariate balancing as estimating equations Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 4 / 20

5 Optimal Covariate Balancing Equations Which covariates and what functions of them should we balance? The outcome model matters Outcome model: E(Y i (t) X i ) = K t (X i ) for t = 0, 1 Possibly misspecified propensity score π β (X i ) π β o(x i ) Inverse probability weighting (IPW) estimator: 1 n { n i=1 T i Y i π ˆβ (X i) (1 T i)y i 1 π ˆβ (X i) Asymptotic bias for the ATE: [ ( Ti E π β o(x i ) 1 T ) ] i {π β o(x i )K 0 (X i ) + (1 π β o)k 1 (X i )} 1 π β o(x i ) }{{} optimal f (X i ) [ ( = E 1 1 T ) ( ) i Ti ] K 0 (X i ) + 1 π β o(x i ) π β o(x i ) 1 K 1 (X i ) }{{}}{{} balance control group balance treatment group Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 5 / 20 }

6 Covariate Balancing in High-Dimension We focus on the estimation of E(Y i (1)) A similar estimation strategy can be applied to E(Y i (0)) Low dimension strong covariate balancing 1 n n i=1 ( ) Ti 1 X i = 0 ˆπ i asymptotic normality, semiparametric efficiency, double robustness (Fan et al.) High dimension weak covariate balancing 1 n n i=1 ( ) Ti 1 α X i = 0 ˆπ i Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 6 / 20

7 Assumptions for Estimating µ 1 = E(Y i(1)) 1 Linear outcome model: 2 Logistic propensity score model: K 1 (X i ) = α X i Pr(T i = 1 X i ) = π(β X i ) = exp(β X i ) 1 + exp(β X i ) 3 Sparsity of both models: max(s 1, s 2 ) log d log n/n = o(1) where s 1 = #{β j > 0} and s 2 = #{α j > 0} 4 Tuning parameters for Lasso: λ = a log(d)/n and λ = a log(d)/n for some unknown constants a and a Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 7 / 20

8 The Proposed Methodology 1 Fit the penalized logistic regression model for propensity score: 1 ˆβ = argmin β R d n n i=1 { } T i (β X i ) log(1 + exp(β X i )) + λ β 1, 2 Fit the penalized linear regression model for the outcome: 1 n α = argmin T i {Y i α X i } 2 + λ α 1, α R d n i=1 3 Calibrate the estimated propensity score by balancing covariates: 1 n γ = argmin T 2 i n π(γ X i + ˆβ S Sc X i S ) 1 X i S c γ R S where S = {j : α j > 0} i=1 4 Use the estimated propensity score π i = π( γ X i S + ˆβ Sc X i S c ) for the IPW estimator ˆµ 1 = 1 n n i=1 T iy i / π i Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 8 / 20 2

9 Theoretical Properties 1 n-consistency ˆµ 1 µ 1 = 1 n n [ Ti π i=1 i ] (Y i (1) α X i ) + α X i µ 1 + o p (n 1/2 ) 2 asymptotic normality and semiparametric efficiency n(ˆµ1 µ 1 ) [ ]) D 1 N (0, E π E(ɛ2 i1 X) + (α X i µ 1 )2 valid under K -fold cross-validation 3 sample boundedness ˆµ 1 min i:t i =1 Y i n n i=1 T i π i = min i:t i =1 Y i and similarly, ˆµ 1 max i:ti =1 Y i Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 9 / 20

10 Robustness and Generalizations 1 Multi-valued treatment regime use the penalized multinomial logistic regression balance selected covariates for each treatment group 2 Generalized linear models for the outcome exponential family p(y X) = h(y, ψ) exp[{y α X b(α X)}/a(ψ)] balance f (X i ) = (b ( α X i ), b ( α X i )X i S) b (α X i ) b ( α X i ) + b ( α X i )(α α)x i 3 Misspecified propensity score model theoretical properties hold so long as the outcome model is correct sure screening property must hold for S in the outcome model Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 10 / 20

11 Comparison with Other Methods 1 Augmented inverse probability weighting (AIPW) Belloni et al. (2014), Farrell (2015), Chernozhukov et al. (2016) Extends Robins doubly-robust estimator to high-dimension 1 n n { ˆm 1 (X i ) + T } i(y i ˆm 1 (X i )) ˆπ i (X i ) i=1 Double selection method 2 Residual balancing (RB) Athey, Imbens, and Wagner (2016) Assumes sparsity only for the outcome model Critically relies on linearity assumption Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 11 / 20

12 Simulation Studies Inspired by Kang and Schafer (2007) Covariates: X i i.i.d. N (0, Σ) where Σ jk = ρ j k Propensity score model: Pr(T i X i ) = logit 1 ( X i X i2 0.25X i3 0.1X i4 ) Outcome model: Y i (1) = X i X i X i9 + ɛ 1i Y i (0) = X i X i X i X i X i9 + ɛ 0i where ɛ ti i.i.d. N (0, 1) for t = 0, 1 Add irrelevant covariates so that a total number of covariates, d, varies from 10 to 2000 Comparison with Residual Balancing (RB) and regularized augmented inverse probability weighting (AIPW) Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 12 / 20

13 Both Models are Correct Standardized root-mean-squared error E(ˆµ µ ) 2 /µ n = 100 n = 1000 d HD-CBPS RB AIPW HD-CBPS RB AIPW Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 13 / 20

14 Propensity Score Model is Misspecified Misspecification through transformation: X mis = (exp(x 1 /2), X 2 (1 + exp(x 1 )) , (X 1 X 3 / ) 3, (X 2 + X ) 2, X 5,, X d ) n = 100 n = 1000 d HD-CBPS RB AIPW HD-CBPS RB AIPW Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 14 / 20

15 Both Models are Misspecified Outcome model misspecification through another transformation: X mis = (exp(x 1 /2), X 2 (1 + exp(x 1 )) , (X 1 X 3 / ) 3, (X 2 + X ) 2, X 6, exp(x 6 + X 7 ), X 2 9, X , X 9,, X d ) n = 100 n = 1000 d HD-CBPS RB AIPW HD-CBPS RB AIPW Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 15 / 20

16 Coverage of 95% Confidence Intervals Both models are correctly specified Average length of confidence intervals in parentheses d n = 100 n = 500 n = (0.8550) (0.3674) (0.2766) (0.7649) (0.3767) (0.2747) (0.8027) (0.3800) (0.2617) (0.7268) (0.3718) (0.2697) (0.7882) (0.4000) (0.2609) Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 16 / 20

17 Logistic Outcome Model Pr(Y i (1) = 1 X i ) = logit 1 (1 + exp( X i X i X i3 )) Pr(Y i (0) = 1 X i ) = logit 1 (1 + exp( X i X i X i X i X i5 )) n = 100 n = 500 n = 1000 d HD-CBPS AIPW HD-CBPS AIPW HD-CBPS AIPW Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 17 / 20

18 Empirical Illustration Effect of college attendance on political participation (Kam and Palmer 2008 JOP) Data: Youth-Parent Socialization Panel Study (n = 1051) Outcome: a total number of (up to 8) participatory acts Covariates: 81 pre-adult characteristics Propensity score: a total of 204 predictors The authors found little effect of college attendance Henderson and Chatfield (2011) and Mayer (2011) use genetic matching and find a positive and statistically significant effect Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 18 / 20

19 Empirical Results Propensity score: logistic regression with 204 predictors Outcome model: binomial logistic regression with 204 predictors 27 covariates are selected by HD-CBPS Overall (ATE) Overall (ATT) Whites (ATE) HD-CBPS CBPS AIPW (0.1247) (0.2380) (0.1043) (0.1420) (0.3094) (0.1279) (0.1089) Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 19 / 20

20 Concluding Remarks Covariate balancing as an efficient and robust way to estimate propensity score Challenges in high-dimension = cannot balance all covariates High-dimensional covariate balancing propensity score (HD-CBPS) sparsity assumption + weak covariate balancing efficient and robust to misspecification of propensity score model more reliance on the outcome model specification loss of double-robustness HD-CBPS is implemented in the R package CBPS Kosuke Imai (Princeton) Covariate Balancing Propensity Score Columbia (May 13, 2017) 20 / 20