Propensity-score-adjustment method for nonignorable nonresponse Jae Kwang Kim 1 Joint work with Minsun Riddles and Jongho Im 1 Motivating Example Exit Poll: (2012 legislative election in Korea, Gangdong-Gap district ) Gender Male Female Total Truth Kim (ISU) Age 20-29 30-39 40-49 5020-29 30-39 40-49 50- A 93 104 146 560 106 129 170 501 1,809 62,489 Party B 115 233 295 350 159 242 262 218 1,874 57,909 Other 4 8 5 3 8 5 5 7 45 1,624 Refusal 28 82 49 174 62 70 69 211 745 Total 240 427 495 1,087 335 446 506 937 4,473 122,022 Propensity-score-adjustment method for nonignorable nonresponse 2 / 54 Comparison of the methods (%) Method No adjustment Adjustment (Age*Sex) Truth Kim (ISU) A 48.5 49.0 51.2 B 50.3 49.8 47.5 Other 1.2 1.2 1.3 Propensity-score-adjustment method for nonignorable nonresponse 3 / 54 Introduction 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 4 / 54 Introduction Basic Setup (X , Y ): random variable θ: Defined by solving E {U(θ; X , Y )} = 0. yi is subject to missingness 1 δi = 0 if yi responds if yi is missing. Want to find wi such that the solution θ̂w to n X δi wi U(θ; xi , yi ) = 0 i=1 is consistent for θ. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 5 / 54 Introduction Basic Setup Two preliminary results: Result 1: The choice of wi = 1 E (δi | xi , yi ) (1) makes the resulting estimator θ̂w consistent. Result 2: If δi ∼ Bernoulli(πi ), then using wi = 1/πi also makes the resulting estimator consistent, but it is less efficient than θ̂w using wi in (1). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 6 / 54 Introduction Basic Setup May assume that δi ∼ Bernoulli(πi ) where πi = π(xi , yi ; φ). For example, logistic regression model π(xi , yi ; φ) = exp(φ0 + φ1 xi + φ2 yi ) 1 + exp(φ0 + φ1 xi + φ2 yi ) Propensity-score-adjusted (PSA) estimator of θ: solve n X δi i=1 π(xi , yi ; φ̂) U(θ; xi , yi ) = 0 for θ, where φ̂ is a consistent estimator of φ. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 7 / 54 Introduction Basic Setup Questions: 1 2 3 Under what conditions, parameter φ is identifiable (or estimable) ? How to estimate the parameter ? What is the limiting distribution of the PSA estimator θ̂w using wi = {π(xi , yi ; φ̂)}−1 ? Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 8 / 54 Basic Theory 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 9 / 54 Basic Theory Observed likelihood f (y | x; θ): model of y on x g (δ | x, y ; φ): model of δ on (x, y ) Observed likelihood Lobs (θ, φ) = Y f (yi | x i ; θ) g (δi | x i , yi ; φ) δi =1 × YZ f (yi | x i ; θ) g (δi | x i , yi ; φ) dyi δi =0 Under what conditions the parameters are identifiable? Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 10 / 54 Basic Theory Lemma Suppose that we can decompose the covariate vector x into two parts, x 1 and x 2 , such that g (δ|y , x) = g (δ|y , x 1 ) (2) (a) (b) and, for any given x 1 , there exist x 2 6= x 2 (a) such that (b) f (y |x 1 , x 2 = x 2 ) 6= f (y |x 1 , x 2 = x 2 ) (3) almost surely. Under some other minor conditions, all the parameters in f and g are identifiable. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 11 / 54 Basic Theory Remark Condition (2) means δ ⊥ x2 | y , x 1 . That is, given (y , x 1 ), x 2 does not help in explaining δ. Thus, x 2 plays the role of instrumental variable in econometrics: f (y ∗ | x ∗ , z ∗ ) = f (y ∗ | x ∗ ), Cov (z ∗ , x ∗ ) 6= 0. Here, y ∗ = δ, x ∗ = (y , x 1 ), and z ∗ = x 2 . We may call x 2 the nonresponse instrument variable. Rigorous theory developed by Wang et al. (2014). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 12 / 54 Basic Theory Parameter estimation : GMM method Kott and Chang (2010) May assume P(δ = 1 | x, y ) = π(φ0 + φ1 x 1 + φ2 y ) for some (φ0 , φ1 , φ2 ). Construct a set of estimating equations such as n X i=1 δi − 1 (1, x 1i , x 2i ) = 0 π(φ0 + φ1 x 1i + φ2 yi ) that are unbiased to zero. May have overidentified situation: Use the generalized method of moments (GMM). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 13 / 54 Proposed method 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 14 / 54 Proposed method Motivation The GMM approach is based on moment conditions for δ: n X i=1 δi − 1 h(x i ) = 0 π(φ; x 1i , yi ) for some function h(x) of x. The choice of h(x) determines the statistical efficiency and also the computational stability. Idea: Let’s consider the maximum likelihood approach and see whether it works better. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 15 / 54 Proposed method Motivation (Cont’d) Recall, assuming for now that f (y | x) is known, the observed score function is Y Lobs (φ) = f (yi | x i ) g (δi | x 1i , yi ; φ) δi =1 × YZ f (y | x i ) g (δi | x 1i , y ; φ) dy . δi =0 How to obtain the MLE of φ? Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 16 / 54 Proposed method Motivation (Cont’d) Mean score theorem: The MLE can be obtained by solving S̄(φ) ≡ E {Scom (φ) | X , Yobs , δ} = 0 where Scom (φ) is the complete-sample score function of φ and Yobs is the observed part of Y . The above conditional expectation of the complete-sample score function is called mean score function. First discussed by R.A. Fisher in his seminal 1922 paper. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 17 / 54 Proposed method Motivation (Cont’d) Under the IID setup, Scom (φ) = n X Si (φ), i=1 where Si (φ) = ∂ log g (δi | x1i , yi ; φ)/∂φ. The mean score function becomes S̄(φ) = n X [δi Si (φ) + (1 − δi )E {Si (φ) | xi , δi = 0}] , i=1 where R E {Si (φ) | xi , δi = 0} = Kim (ISU) Si (φ)f (y | xi ){1 − π(x1i , y ; φ)}dy R . f (y | xi ){1 − π(x1i , y ; φ)}dy Propensity-score-adjustment method for nonignorable nonresponse 18 / 54 Proposed method Motivation (Cont’d) How to compute the conditional expectation? Classical approach (Greenlees and Zieschang (1982), Baker and Laird (1988), Ibrahim et al. (1999)): Assume a parametric model on f (y | x) = f (y | x; θ) and use the EM to solve the mean score equation of the parameters in the full joint distribution. R E {Si (θ, φ) | xi , δi = 0} = Si (φ)f (y | xi ; θ){1 − π(x1i , y ; φ)}dy R . f (y | xi ; θ){1 − π(x1i , y ; φ)}dy Two problems 1 2 Requires correct specification of f (y | x; θ). Known to be sensitive to the choice of f (y | x; θ). Computationally heavy: Often uses Monte Carlo computation. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 19 / 54 Proposed method Proposed method Remedy (for Problem One) Idea Instead of specifying a parametric model for f (y | x), consider specifying a parametric model for f (y | x, δ = 1), denoted by f1 (y | x). In this case, R Si (φ)f1 (y | xi )O(x1i , y ; φ)dy R E {Si (φ) | xi , δi = 0} = f1 (y | xi )O(x1i , y ; φ)dy where O(x1 , y ; φ) = Kim (ISU) 1 − π(φ; x1 , y ) . π(φ; x1 , y ) Propensity-score-adjustment method for nonignorable nonresponse 20 / 54 Proposed method Remark Based on the following identity f (y | x, δ = 0) = f (y | x, δ = 1) O(x1 , y ; φ) . E {O(x1 , y ; φ) | x, δ = 1} Kim and Yu (2011) considered a Kernel-based nonparametric regression method of estimating f (y | x, δ = 1) to obtain E (Y | x, δ = 0). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 21 / 54 Proposed method Proposed method Problem Two Given that f1 (y | xi ) is known, how to compute R Si (φ)O(x1i, y ; φ)f1 (y | xi )dy E {Si (φ) | xi , δi = 0} = R O(x1i , y ; φ)f1 (y | xi )dy without relying on Monte Carlo computation? Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 22 / 54 Proposed method Proposed method Express E {Si (φ) | xi , δi = 0} = where E1 {Si (φ)O(x1i, Y ; φ) | xi } E1 {O(x1i, Y ; φ) | xi } Z E1 {Q(xi , Y ) | xi } = Q(xi , y )f1 (y | xi )dy . How to estimate E1 {Q(xi , Y ) | xi } without using numerical integration or Monte Carlo methods? Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 23 / 54 Proposed method Proposed method Idea If xi were null, then we would approximate the integration by the empirical distribution among δ = 1. Use Z f1 (y | xi ) f1 (y )dy f1 (y ) X f1 (yj | xi ) ∝ Q(xi , yj ) f1 (yj ) Z Q(xi , y )f1 (y | xi )dy = Q(xi , y ) δj =1 where Z f1 (y ) = f1 (y | x)f (x | δ = 1)dx ∝ X f1 (y | xi ). δi =1 Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 24 / 54 Proposed method Proposed method In practice, f1 (y | x) is unknown and is estimated by fˆ1 (y | x) = f1 (y | x; γ̂). Thus, given γ̂, a fully efficient estimator of φ can be obtained by solving n X S2 (φ, γ̂) ≡ δi S(φ; xi , yi ) + (1 − δi )S̄0 (φ | xi ; γ̂, φ) = 0, (4) i=1 where P S̄0 (φ | xi ; γ̂, φ) = δj =1 S(φ; xi , yj )f1 (yj P δj =1 f1 (yj and fˆ1 (y ) = nR−1 | xi ; γ̂)O(φ; x 1i , yj )/fˆ1 (yj ) | xi ; γ̂)O(φ; x 1i , yj )/fˆ1 (yi ) n X δi f1 (y | xi ; γ̂). i=1 May use EM algorithm to solve (4) for φ. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 25 / 54 Proposed method Proposed method Step 1: Use the responding part of (xi , yi ), obtain γ̂ in the model f1 (y | x; γ). X S1 (γ) ≡ S1 (γ; xi , yi ) = 0. (5) δi =1 Step 2: Given γ̂ from Step 1, obtain φ̂ by solving (4): S2 (φ, γ̂) = 0. Step 3: Using φ̂ computed from Step 2, the PSA estimator of θ can be obtained by solving n X δi U(θ; xi , yi ) = 0, π̂i (6) i=1 where π̂i = πi (φ̂). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 26 / 54 Proposed method Proposed method In many cases, x is categorical and f1 (y | x) can be fully nonparametric. If x has a continuous part, nonparametric Kernel smoothing can be used. The proposed method seems to be robust against the failure of the assumed model on f1 (y | x; γ) . Asymptotic normality of PSA estimator can be obtained & Linearization method can be used for variance estimation. By augmenting the estimating function, we can also impose a calibration constraint such as n n X X δi xi = xi . π̂i i=1 Kim (ISU) i=1 Propensity-score-adjustment method for nonignorable nonresponse 27 / 54 An Example 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 28 / 54 An Example Categorical Data (All dichotomous) Example (SRS, n = 10) ID Weight x1 1 0.1 1 2 0.1 1 3 0.1 0 4 0.1 1 5 0.1 0 6 0.1 1 7 0.1 0 8 0.1 1 9 0.1 0 10 0.1 0 M: Missing Kim (ISU) x2 0 1 1 0 1 0 1 0 1 0 y 1 1 M 0 1 M M 0 1 0 Propensity-score-adjustment method for nonignorable nonresponse 29 / 54 An Example Categorical Data (All dichotomous) Example (Cont’d) Assume P(δ = 1 | x1 , x2 , y ) = π(x1 , y ) ID Weight x1 x2 y 1 0.1 1 0 1 2 0.1 1 1 1 3 0.1 · w3,0 0 1 0 0.1 · w3,1 0 1 1 4 0.1 1 0 0 5 0.1 0 1 1 w3,j = P̂(Y = j | X1 = 0, X2 = 1, δ = 0) ∝ P̂(Y = j | X1 = 0, X2 = 1, δ = 1) Kim (ISU) 1 − π̂(0, j) π̂(0, j) Propensity-score-adjustment method for nonignorable nonresponse 30 / 54 An Example Categorical Data (All dichotomous) Example (Cont’d) ID Weight 6 0.1 · w6,0 0.1 · w6,1 7 0.1 · w7,0 0.1 · w7,1 8 0.1 9 0.1 10 0.1 w6,j w7,j Kim (ISU) x1 1 1 0 0 1 0 0 x2 0 0 1 1 0 1 0 y 0 1 0 1 0 1 0 1 − π̂(1, j) π̂(1, j) 1 − π̂(0, j) ∝ P̂(Y = j | X1 = 0, X2 = 1, δ = 1) π̂(0, j) ∝ P̂(Y = j | X1 = 1, X2 = 0, δ = 1) Propensity-score-adjustment method for nonignorable nonresponse 31 / 54 An Example Categorical Data (All dichotomous) Example (Cont’d) E-step: Compute the conditional probability using the estimated response probability π̂ab . M-step: Update the response probability using the fractional weights. For fully nonparametric model, P δi =1 I (x1i = a, yi = b) π̂ab = P P P1 ∗ δi =1 I (x1i = a, yi = b) + δi =0 j=0 wi,j I (x1i = a, yij = b) Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 32 / 54 An Example Categorical Data (All dichotomous) Example (Cont’d) The proposed method can be viewed as a fractional imputation method of Kim (2011). On the other hand, Kott and Chang method is more close to nonresponse weighting adjustment. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 33 / 54 An Example Categorical Data (All dichotomous) Example Kott and Chang method ID Wgt 1 Wgt2 x1 x2 −1 1 0.1 0.1π̂11 1 0 −1 2 0.1 0.1π̂11 1 1 3 0.1 0.0 0 1 −1 4 0.1 0.1π̂10 1 0 −1 5 0.1 0.1π̂01 0 1 6 0.1 0.0 1 0 7 0.1 0.0 0 1 −1 8 0.1 0.1π̂10 1 0 −1 9 0.1 0.1π̂01 0 1 −1 10 0.1 0.1π̂00 0 0 M: Missing Kim (ISU) y 1 1 M 0 1 M M 0 1 0 Propensity-score-adjustment method for nonignorable nonresponse 34 / 54 An Example Categorical Data (All dichotomous) Kott and Chang (2010) method: Calibration equation X X δi I (x1i = a, x2i = b) = I (x1i = a, x2i = b). π̂i i i 1 2 3 4 X1 X1 X1 X1 = 1, X2 = 1, X2 = 0, X2 = 0, X2 = 1: = 0: = 1: = 0: −1 π̂11 =1 −1 −1 −1 π̂11 + π̂10 + π̂10 =4 −1 −1 π̂01 + π̂01 = 4 −1 π̂00 = 1. The solution of Kott and Chang method is π̂11 = 1, π̂10 = 2/3, π̂01 = 1/2, π̂00 = 1. The solution from the proposed method is π̂11 = 0.2, π̂10 = 0.3, π̂01 = 0.4, π̂00 = 0.1. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 35 / 54 Statistical Properties 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 36 / 54 Statistical Properties Asymptotic properties Writing Up (θ, φ) = n X i=1 δi U(θ; xi , yi ), πi (φ) we can write (θ̂PS , φ̂, γ̂) as the solution to S1 (γ) 0 S2 (φ, γ) = 0 , Up (θ, φ) 0 (7) where S1 (γ) and S2 (φ, γ) are defined in (5) and (4), respectively. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 37 / 54 Statistical Properties Asymptotic properties (Cont’d) Thus, we can use the sandwich formula is, we have γ̂ V φ̂ = T −1 V θ̂PS where to obtain the linearization. That S1 (γ) 0 S2 (φ, γ) T −1 Up (θ, φ) (8) 0 0 E (∂S1 /∂γ 0 ) 0 T = E (∂S2 /∂γ 0 ) E (∂S2 /∂φ0 ) . E (∂Up /∂γ 0 ) E (∂Up /∂φ0 ) E (∂Up /∂θ0 ) Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 38 / 54 Statistical Properties Alternative expression Writing η = (φ, γ) and S(η) = S1 (γ) S2 (φ, γ) , we can write θ̂PS to be solution to Up (θ; η̂) = 0 where η̂ satisfies S(η̂) = 0. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 39 / 54 Statistical Properties Alternative expression (Cont’d) Taylor linearization −1 ∂Up ∂S ∗ ∼ UP (θ, η̂) = UP (θ, η ) − E S(η ∗ ) E ∂η 0 ∂η 0 = UPS (θ, η ∗ ) − Cov Up , S 0 {V (S)}−1 S(η ∗ ), by the property of zero-mean function. (i.e. If E (U) = 0, then E (∂U/∂η 0 ) = −Cov (U, S 0 ).) So, we have V {θ̂PS (η̂)} ∼ = V {θ̂PS (η ∗ ) | S ⊥ } ≤ V {θ̂PS (η ∗ )}, where V {θ̂ | S ⊥ } = V (θ̂) − Cov (θ̂, S 0 ) {V (S)}−1 Cov (S, θ̂). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 40 / 54 Statistical Properties Alternative expression (Cont’d) The solution to Up (θ, η̂) = 0 can be obtained by minimizing QPS (θ) = Up (θ, η̂)0 [Var {Up (θ, η̂)}]−1 Up (θ, η̂) with respect to θ. The solution can also be obtained by minimizing ∗ Q (θ, η) = UP (θ, η) S(η) 0 Var UP (θ, η) S(η) −1 UP (θ, η) S(η) with respect to (θ, η). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 41 / 54 Statistical Properties Alternative expression (Cont’d) Justification 0 −1 UP (θ, η) V11 V12 UP (θ, η) Q (θ, η) = S(η) V21 V22 S(η) = Q1 (θ | η) + Q2 (η) ∗ where 0 −1 −1 −1 ÛP − V12 V22 S V UP | S ⊥ ÛP − V12 V22 S n o−1 Q2 (η) = S(η)0 V̂ (S) S(η) Q1 (θ | η) = For the MLE η̂, we have Q2 (η̂) = 0 and Q1 (θ | η̂) = QPS (θ). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 42 / 54 Statistical Properties Alternative expression (Cont’d) The PS estimator θ̂PS is obtained by the solution to an optimization problem: θ̂PS = arg min QPS (θ) = arg min Q ∗ (θ, η). In computing Q ∗ (θ, η), we can use Pn Pn V̂11 V̂12 δi π̂i−2 (1 − π̂i )Ûi Ûi0 δi π̂i−1 Ûi Ŝi0 i=1 i=1 Pn = Pn −1 0 0 V̂21 V̂22 i=1 δi π̂i Ŝi Ûi i=1 Ŝi Ŝi where Ŝi = Si (η̂), Ûi = U(θ̂PS ; xi , yi ), S(η) = n X i=1 n X δi S1i (γ) Si (η) = S2i (φ, γ) i=1 and, by (4), S2i (φ, γ) = δi S(φ; xi , yi ) + (1 − δi ) S̄0 (φ | xi ; γ, φ) . Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 43 / 54 Statistical Properties Variance estimation Linearized variance estimator n o 0 −1 V̂ (θ̂) = τ̂2−1 V̂11 − V̂12 V̂22 V̂21 τ̂2−1 , (9) where τ̂2 = n X δi π̂i−1 U̇(θ̂; xi , yi ) i=1 and U̇(θ; x, y ) = ∂U(θ; x, y )/∂θ0 . P Instead of using V̂22 = ni=1 Si Si0 , one can use V̂22 = −∂S(η)/∂η evaluated at η = η̂, the observed information using the formula of Oakes (1999). Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 44 / 54 Simulation studies 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 45 / 54 Simulation studies Simulation Study (Cont’d) Simulation setup Case 1: Case 2: √ y = −1 + 2(xi + 1) + ei y = −1 + xi2 + ei ei ∼ N(0, 1), ei ∼ N(0, 1), where xi ∼ N(0, 1). 2,000 MC samples of size n = 500. Response mechanism: logistic regression model, π(x, y ; φ) ≡ P(δ = 1 | x, y ) = {1 + exp(−φ0 − φ1 y )}−1 , where (φ0 , φ1 ) = (1, −0.2). Average response rates for all cases are about 73%. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 46 / 54 Simulation studies Simulation Study (Cont’d) Compare three estimation methods: 1 2 MAR: PSA method assuming MAR. GMM: Kott and Chang (2010)’s method. φ̂ck : a solution to the following calibration condition, n X {δi /π(φck ) − 1} zi = 0, i=1 where zi = (1, xi ) for Case 1 and zi = (1, xi , xi2 ) for Case 2. 3 NEW: Proposed method, using f1 (y | x) ∼ N(β0 + β1 x, σ 2 ) for Case 1. f1 (y | x) ∼ N(β0 + β1 x + β1 x 2 , σ 2 ) for Case 2. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 47 / 54 Simulation studies Simulation Study (Cont’d) Table: Monte Carlo Bias, Variance, and Mean Squared Error (MSE) of the estimators of θ = E (Y ) Case 1 Case 2 Kim (ISU) Method MAR GMM NEW MAR GMM NEW Bias -0.055 -0.001 -0.001 -0.167 -0.000 -0.001 Variance 0.0063 0.0068 0.0068 0.0072 0.0095 0.0085 MSE 0.0092 0.0068 0.0068 0.0352 0.0095 0.0085 Propensity-score-adjustment method for nonignorable nonresponse 48 / 54 Simulation studies Simulation Study (Cont’d) Table: Monte Carlo Bias and Variance of the parameter estimates and Relative bias of variance estimates Parameter Case1 Case2 θ φ0 φ1 θ φ0 φ1 Kim (ISU) Bias -0.00 0.01 -0.00 0.01 0.01 -0.01 GMM Variance 0.0068 0.0111 0.0059 0.0095 0.0147 0.0082 R.Bias 0.05 0.01 -0.03 -0.02 -0.05 -0.08 Bias -0.00 0.01 -0.00 -0.00 0.00 0.00 NEW Variance 0.0068 0.0111 0.0057 0.0085 0.0110 0.0055 Propensity-score-adjustment method for nonignorable nonresponse R.Bias 0.06 0.01 -0.00 -0.03 -0.01 -0.03 49 / 54 Simulation studies Back to Exit Poll Example Table: Prediction Result (Gangdong-Gap) Method No adjustment Adjustment (Age * Sex) New Method Truth Kim (ISU) A 48.5 49.0 51.0 51.2 B 50.3 49.8 47.7 47.5 Other 1.2 1.2 1.2 1.3 Propensity-score-adjustment method for nonignorable nonresponse 50 / 54 Simulation studies Back to Exit Poll Example Table: Predicted Seats in Seoul Method No adjustment Adjustment (Age* Sex) New Method Truth Kim (ISU) A 10 10 15 16 B 36 36 29 30 Other 2 2 4 2 Propensity-score-adjustment method for nonignorable nonresponse 51 / 54 Concluding Remarks 1 Introduction 2 Basic Theory 3 Proposed method 4 An Example 5 Statistical Properties 6 Simulation studies 7 Concluding Remarks Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 52 / 54 Concluding Remarks Concluding Remarks Maximum likelihood (ML) approach to propensity score model parameter estimation under non-ignorable missing. Instrumental variable needed for identifiability of the response model. Unlike the full joint ML approach, the proposed method uses a model for f (y | x, δ = 1) and the result is robust against the failure of the assumption on f (y | x, δ = 1). More efficient than the method based on the method of moments. Directly applicable to complex survey sampling. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 53 / 54 Concluding Remarks Current and Future work AUC estimation through propensity-score-adjustment approach under nonignorable verification bias. Semiparametric estimation of θ based on Kernel regression method for f1 (y | x). Score test and pre-test estimation of φ. Fractional imputation under nonignorable missing data. Extension to longitudinal missing data. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 54 / 54 References REFERENCES Baker, S. G. and N. M. Laird (1988), ‘Regression analysis for categorical variables with outcome subject to nonignorable nonresponse’, Journal of the American Statistical Association 83, 62–69. Greenlees, W. S., Reece J. S. and K. D. Zieschang (1982), ‘Imputation of missing values when the probability of response depends on the variable being imputed’, Journal of the American Statistical Association 77, 251–261. Ibrahim, J. G., S. R. Lipsitz and M. H. Chen (1999), ‘Missing covariates in generalized linear models when the missing data mechanism is non-ignorable’, Journal of the Royal Statistical Society, Series B 61, 173–190. Kim, J. K. (2011), ‘Parametric fractional imputation for missing data analysis’, Biometrika 98, 119–132. Kim, J. K. and C. L. Yu (2011), ‘A semi-parametric estimation of mean functionals with non-ignorable missing data’, Journal of the American Statistical Association 106, 157–165. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 54 / 54 Concluding Remarks Kott, P. S. and T. Chang (2010), ‘Using calibration weighting to adjust for nonignorable unit nonresponse’, Journal of the American Statistical Association 105, 1265–1275. Oakes, D. (1999), ‘Direct calculation of the information matrix via the em algorithm’, Journal of the Royal Statistical Society: Series B 61, 479–482. Wang, S., J. Shao and J. K. Kim (2014), ‘An instrumental variable approach for identifiability and estimation in problems with nonignorable nonresponse’, Statistica Sinica 24, 1097–1116. Kim (ISU) Propensity-score-adjustment method for nonignorable nonresponse 54 / 54