Robustness to Parametric Assumptions in Missing Data Models

Transcription

1 Robustness to Parametric Assumptions in Missing Data Models Bryan Graham NYU Keisuke Hirano University of Arizona April 2011

2 Motivation Motivation We consider the classic missing data problem. In practice covariate dimension is often very high, and conventional asymptotics may be misleading. We consider a finite-sample setting where correct parametric specifications make a big difference. Adopt the Angrist and Hahn (2004) setup, where cells can be small or even empty.

3 Outline Outline Missing Data Model Stratified Sampling Setup Parametric Imputation Empirical Bayes Double Robustness and EB Monte Carlo Some work in progress

4 Missing Data Model MAR Random sample from a population Observe a covariate X, and if D = 1, observe Y. Interested in population mean of Y: θ = E[Y]. Assume Y is missing at random (MAR): Y D X. We are especially interested in cases where cells are small.

5 Missing Data Model Let μ(x) = E[Y X = x] σ 2 (x) = V[Y X = x] e(x) = Pr(D = 1 X = x) bounded away from 0. Semiparametric efficiency bound for estimating θ (Hahn, 1998): VB = E σ (μ(x) θ) 2 2 (X) + E. e(x)

6 Missing Data Model Various efficient estimators proposed: Hahn (1998) Hirano, Imbens, Ridder (2003) Chen, Hong, Tarozzi (2008) Typically involve NP estimation of μ(x), e(x), or both. Typically identifical if X discrete.

7 Stratified sampling Discrete Covariate/Small Cell Model Following Angrist and Hahn (2004): Covariate takes values in {x 1,..., x K }. M k individuals in each cell fixed and known (stratified sampling). Let M k p k = K j=1 M j (distribution of X)

8 Stratified sampling Propensity score: in cell k, observe Y with probability e k > 0. n k observed outcomes per cell, with In cell k, let n k Binomial(e k, M k ). Y k1,..., Y knk iid (μk, σ 2 k ). Estimand: K θ = p k μ k. k=1

9 Poststratification Estimator Poststratification Estimator Let: Estimator: Y k = 1 n k ˆθ PS = n k i=1 Y ki. K p k Y k. k=1 May work poorly if n k small. For empty cells: drop cells (adjust p k ), or combine cells.

10 Parametric Imputation Parametric Imputation Suppose μ k = x k β, where β is a low-dimensional parameter. Then E[Y k n 1,..., n K ] = x k β, V[Y k n 1,..., n K ] = σ 2 k /n k.

11 Parametric Imputation Let ˆβ = WLS of Y k on x k (= OLS of underlying y s on x) Parametric Imputation Estimator: ˆθ PI = K k=1 p k x k ˆβ. Can handle empty cells, and will work well when model is correct. But may be sensitive to parametric specification.

12 Empirical Bayes Empirical Bayes Based on Morris (1983) and Chamberlain (2009) Suppose (temporarily) that μ k N(x k β, τ2 ). Note τ 2 = 0 corresponds to parametric imputation. Further suppose Y k μ k N(μ k, v k ), where v k = σ 2 k /n k.

13 Empirical Bayes Marginal for cell means: Y k N(x k β, v k + τ 2 ). Then posterior mean of μ k is μ k = (1 γ k)y k + γ k (x k β), where γ k = v k v k + τ 2.

14 Empirical Bayes Replace β, v k, τ 2 with estimates, form ˆγ k = ˆv k ˆv k + ˆτ 2 k, and K ˆθ EB1 = p k (1 ˆγk )Y k + ˆγ k (x ˆβ). k k=1 Small cells: if n k = 0 set ˆγ k = 1, modify variance estimate if n k = 1.

15 Empirical Bayes Qualitative features of EB If all ˆγ k near 1 (e.g. ˆτ 2 0), similar to parametric imputation If all ˆγ k near 0 (e.g. if all n k large), similar to poststratification More generally, weighting varies by cell based on v k and τ 2. Similar to adaptive bandwidth kernel.

16 Double Robustness and EB Double Robustness and EB Assumption DR1: μ k = x β for all k. k Assumption DR2: e k = G(x k ) for all k, where G is a known function. Double robustness: estimator is consistent if one or both DR1, DR2 hold. Protection against misspecification of μ k, provided G(x k ) is correct (and vice versa).

17 Double Robustness and EB Bang and Robins (2005): can augment regression with inverse of propensity score. Let α 1, α 2 solve min K α 1,α 2 k=1 p k e k E Yk x k α 1 G 1 2 (x k )α 2. (1) If DR1, DR2, or both hold, then θ = K k=1 p k x k α 1 + G 1 (x k )α 2.

18 Double Robustness and EB Proof Equivalent minimization problem: min α 1,α 2 k=1 K p k e k μk x k α 1 G 1 2 (x k )α 2. (2) If DR1 holds, (2) is solved by setting α 1 = β and α 2 = 0. Then K k=1 p k x k α 1 + G 1 (x k )α 2 = K k=1 p k x k β = K p k μ k = θ. k=1

19 Double Robustness and EB Proof cont d If DR2 holds, first order conditions for (2) implies K k=1 p k e k Then since e k = G(x k ), G(x k ) μk x k α 1 G 1 (x k )α 2 K p k μ k = k=1 K k=1 = 0. p k x k α 1 + G 1 (x k )α 2.

20 Double Robustness and EB Feasible version: ê k = n k M k. Then p k ê k n k, so we could solve min α 1,α 2 k=1 K n k Yk x k α 1 G 1 2 (x k )α 2. This is WLS of Y k on (x k, G 1 (x k )), with weights proportional to n k. ˆθ DR = K k=1 p k x k ˆα 1 + G 1 (x k ) ˆα 2.

21 Double Robustness and EB EB Extension of DR ˆθ DR is parametric imputation estimator with an additional regressor. So we can develop a corresponding empirical Bayes extension ˆθ EB2.

22 Double Robustness and EB EB Extension of DR ˆθ DR is parametric imputation estimator with an additional regressor. So we can develop a corresponding empirical Bayes extension ˆθ EB2. Triply robust?

23 Double Robustness and EB EB Extension of DR ˆθ DR is parametric imputation estimator with an additional regressor. So we can develop a corresponding empirical Bayes extension ˆθ EB2. Triply robust? Caution: DR2 does not imply τ 2 = 0. Could try to engineer alternative estimators of γ k. (But see Monte Carlo evidence below.)

24 Monte Carlo Monte Carlo study Support of X: { J,..., 0,..., J}. Equal size strata: M k = M for all k. μ k = x k β (implies θ = 0). Propensity score: step function, e k =.75 if x k < 0. Overall 1/2 probability of selection. Y ki iid N(μ k, σ 2 ).

25 Monte Carlo Consider various values of J, M with overall sample size Choose σ 2 so that (large sample) efficient estimator should have SE =.1. Specification for μ k : correct model and incorrect model (constant mean) Propensity score correctly specified

26 Monte Carlo Bias, Correct Parametric Mean K PS PI DR EB EBPS K = # of cells For K=375, median of 18 empty cells

27 Monte Carlo RMSE, Correct Parametric Mean K PS PI DR EB EBPS

28 Monte Carlo Bias, Misspecified Mean K PS PI DR EB EBPS

29 Monte Carlo RMSE, Misspecified Parametric Mean K PS PI DR EB EBPS

30 Monte Carlo

31 Monte Carlo

32 Monte Carlo Remarks PS is poor when cells are small. EB is similar to PI when parametric mean correctly specified. With misspecified mean but correct prop score: PI poor, EB better DR good, but EBPS even better shrinking ˆτ all the way to zero would not have helped! In general, EB-type estimators should be (nearly) admissible, but make different risk tradeoffs over parameter space.

33 Extensions and Ongoing Work Extensions and Ongoing Work Continuous covariates: Gaussian process priors. Inference? Some existing literature on EB inference. Finite-sample MSE calculations and finite-sample decision theory. Asymptotics / limit experiments which preserve small/zero cells?

34 Extensions and Ongoing Work Limits of experiments (very very preliminary) Suppose (X i, D i, Y i ) are IID with X i P with density p(x) on X, Pr(D i = 1 X i = x) = e(x), Y i D i, X i = x F(y x, π). Parameters are θ = (e( ), π), and we think of e(x) close to zero.

35 Extensions and Ongoing Work Hellinger Transform Let E = {F θ ; θ Θ} be an experiment. Let α = {α θ ; θ Θ} satisfy α θ 0 θ α θ = 1 only a finite number of the α θ > 0. Define η(α) = (df θ ) α θ. θ

36 Extensions and Ongoing Work Useful properties of Hellinger transform of E: Pointwise convergence of η(α) is equivalent to weak convergence of experiments in the sense of Le Cam. In particular, it implies an asymptotic representation theorem. Hellinger transform of a product experiment is product of Hellinger transforms.

37 Extensions and Ongoing Work For simplicity, drop Y component and focus on (X, D). Local parametrization: Pr(D i = 1 X i = x) = e(x), i = 1,..., n. n Single-observation density: e(x) d f (x, d) = p(x) 1 e(x) n n 1 d wrt dominating measure λ(x, d) = λ d (d)λ x (x).

38 Extensions and Ongoing Work Let α have G non-zero components α 1,..., α G. Hellinger transform (single-obs): η(α) = G f (x, d) α g dλ(x, d) g=1 = e(x) αg + 1 e(x) αg dp(x) g n g n g = E e(x)α g + E 1 e(x) αg. n g n

39 Extensions and Ongoing Work For n observations: η n (α) = {η(α)} n g = E e(x)α g + E 1 e(x) αg n g n E g e(x)α g + 1 E g e(x)α n g n n exp E e(x) α g E e(x)α g. g g n

40 Extensions and Ongoing Work Now, consider a different experiment: Z is a Poisson point process on X, with intensity measure ν(x), indexed by e = e(x), where dν e (x) = e(x)p(x)dλ x (x).

41 Extensions and Ongoing Work By results in Le Cam and Yang (2000): Hellinger transform is G η(α) = exp αg G dνeg α g dν eg = exp g=1 E g g=1 e(x) α g E g α g e(x). Hence the Poisson process experiment characterizes what can be done asymptotically for the small probabilities model. Full model with Y: Poisson process on X Y?