Broadening the Scope of Bootstrap in Complex Problems
Transcripción
Broadening the Scope of Bootstrap in Complex Problems
Broadening the Scope of Bootstrap in Complex Problems Víctor Aguirre Torres ITAM y CIMAT Seminario Dirección General de Investigación Económica, Banco de México 9 de Marzo de 2007 Joint work with Manuel Domínguez, ITAM Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Outline Unconditional Bootstrap Pivots (UBP’s) For large n and using UBP’s, it is possible to draw statistically meaningful inferences (size and power) with a relatively small number of bootstrap replications, say B 50. Examples: procedures that require smoothing or computer intensive methods. Increasing B produces marginal improvement. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Set Up Data Zn = fZ1 , Z2 , ..., Zn g DGP FZ1 Z2 ...Zn . Operator of interest Tn ! d T Limiting distribution unknown. Known but large sample approximation unreliable. Known but unable to compute quantiles. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Bootstrap Approximation Bootstrap the DGP and obtain Zn . Compute Tn . Justi…cation: FT n jZn ! FT in probability, or Tn !d T in probability. With basically the same regularity conditions for Tn to converge to T . Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Main Theorem Let Zn = fZ1 , Z2 , ..., Zn g be a collection of rv´ s, and let Sn (Zn ) be such that FSn ! FS . n o Let Hn = H1 , H2 , ..., Hm (n ) be a collection of rv´ s, m (n) non decreasing integer valued, and let Rn (Zn , Hn ) be such that FR n jZn ! FR in probability, where FR is not stochastic. Then FSn R n ! FS FR Example: Tn = Víctor Aguirre Torres ITAM y CIMAT p θn n b θ and Tn = p θn n b b θn 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Consequence on Bootstrap If Tn !d T and Tn,b !d T in probability for b = 1, 2, ..., B. Then, for …xed B, as n increases to in…nity we have that Tn , Tn,1 , ..., Tn,B !d T B , where T B is a rv with 1 + B iid coordinates. Tn , Tn,b and Tn,b 0 are unconditionally asymptotically independent and equally distributed. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Example: Quantile Regression Data: (Xi , Yi ) i = 1, ..., n. Model: Yi = β1 + β2 Xi + Ui fU unknown. Robust estimation: b β = arg min ∑ qp (Yi qp (u ) = u pIfu 0g (1 β1 p )Ifu <0 g β2 Xi ) 0<p<1 Non di¤erentiable objective function. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 AN in Quantile Regression Take for example p = 1/2, L1 estimator. Under regularity conditions, Koenker and Basset (1978, 1982): p where n b β2 β2 !d N (0, σ2 ) σ2 = fU2 (FU 1 (1/2))V (X )/4 Inference on requires smoothing to estimate fU and then σ. Inference depends on the choice of smoothing constant. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Usual Application of Bootstrap, Quantile Regression Bootstrap resample: (Xi , Yi ) i = 1, ..., n. Under the same regularity conditions for AN, Hahn (1995) p n b β2 b β2 !d N (0, σ2 ) Using percentiles of b β2b directly is equivalent to AN. t-bootstrap requires computation of p zb = n b β2 bb σ b β2 , b = 1, ..., B This demands smoothing for each b = 1, ..., B to estimate b fU and a large B (above 1000). Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Alternative: UBP Consider p β2 n b β2 , p β2,1 n b b β2 , ..., p β2,B n b From the main theorem the limit is i.i.d. N (0, σ2 ). Then: tB = (bβ2 β2 ) 1 b B ∑b =1 ( β2,b B b β2 ) 2 !d t (B ) t b β2 . UBP No smoothing required. No consistency of denominator to estimate σ required. Test of hypothesis. Asymptotic size , global power of one. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 MC exercise, signi…cance level. tB quantile regression, (10,000 MC replications) B=5 300 250 Relative Error 200 t* ; α=0.1 t* ; α=0.05 t* ; α=0.01 N ; α=0.1 N ; α=0.05 N ; α=0.01 150 100 50 0 0 200 400 600 800 1000 1200 -50 Víctor Aguirre Torres ITAM y CIMAT Sample size relative error= αnα α 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Local Power for B Fixed Local alternatives H1n : θ = θ 0 + cσ p . n Power for B and n …nite: π B ,n (c ) = Pc (jtB j > tB ,1 α/2 ) Limit power, B …xed: π B (c ) = limn !∞ π B ,n (c ) π B (c ) = P (jt (B, c )j > tB ,1 α/2 ) Limit power, B ! ∞ π (c ) = P (jN (c, 1)j > z1 Víctor Aguirre Torres ITAM y CIMAT α/2 ) 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Local Power Curves as a Function of B 1 0.9 Power of Unconditional Pivot 0.8 0.7 0.6 5 10 20 50 Limit 0.5 0.4 0.3 0.2 0.1 0 0 1 2 3 4 5 6 Standardized Difference Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Joint Tests Tn !d Np (0, Σ) Then, for …xed B and large n Hn,B = (Tn )T h 1 B ∑Bb=1 Tn,b Tn,b T i 1 Tn !d TB2 ,p Hotelling’s T-square, Bq α B q +1 F (q, B T2 q + 1) UBP Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Another Example Construct an asymptotic pivot by combining Tn , Tn,1 , ..., Tn,B For example, if Tn !d κχ2 (ν) then Hn,B = F Tn (∑Bb=1 T n,b )/B !d F (ν, νB ) UBP Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16 Summary For very complex procedures and large n bootstrap could be applied with drastic reduction in B. For certain procedures that require smoothing, for large n bootstrap could be applied avoiding the use of smoothing constant. Choice of B is nonstochastic. Properties of inferential procedure remain known. Convenient determination of B considering size and power. No need of imposing extra regularity conditions on the DGP or on the asymptotic properties of the operator. Víctor Aguirre Torres ITAM y CIMAT 9 de Marzo de 2007 Joint work with Manuel D Broadening Seminario Dirección the Scope General of Bootstrap de Investigación Económica, Banco de México () / 16