MEI Exam Review. June 7, 2002

Transcription

1 MEI Exam Review June 7, 2002

2 1 Final Exam Revision Notes 1.1 Random Rules and Formulas Linear transformations of random variables. f y (Y ) = f x (X) dx. dg Inverse Proof. (AB)(AB) 1 = I. (B 1 A 1 )(AB)(AB) 1 = (B 1 A 1 ). (AB) 1 = (B 1 A 1 ). Simple linear regression coefficient with only one independent variable and an intercept. [ ] [ ] ˆβ = (X X) 1 X T xt y = yt xt x 2. (1) t xt y t Proof about projections. ˆɛ = M X ɛ = (I X(X X) 1 X )(y Xβ) = y X(X X) 1 X y Xβ + X(X X) 1 X Xβ = y X(X X) 1 X y = M X y = (I X(X X) 1 X )y = y X(X X) 1 X y = y X ˆβ = ˆɛ. Simple linear regression with a constant. R squared. Inverse Rule. y = β 1 + β 2 x t + ɛ t. ˆβ 2 = t(x t x)y t t (x t x). 2 R 2 = 1 RSS T SS = ESS T SS. A 1 = A A. If A is a 2x2, Switch the Main, Negate the Off. 2

3 F test of a restriction. F = (RSS R RSS U )/q RSS U /(N K) With q restrictions and K parameters. F (q, N K). Just algebra. Variance of the OLS estimator. with, s 2 = Multiplying out RSS, RSS N K = RSS = (y X ˆβ) (y X ˆβ) = (y ˆβ X )(y X ˆβ) X X = X ɛ = T x t x t. 1 T x t ɛ t. ˆ V ar( ˆβ) = s 2 (X X) 1, = y y ˆβ X y y X ˆβ + ˆβ X X ˆβ) 1 ˆɛ ˆɛ N K = (y X ˆβ) (y X ˆβ). N K = y y ˆβ X y y X(X X) 1 X y + ((X X) 1 X y) X X(X X) 1 X y = y y ˆβ X y y X(X X) 1 X y + y X (X X) 1 X X(X X) 1 X y }{{} I = y y ˆβ X y y X(X X) 1 X y + y X(X X) 1 X y }{{} 0 = y y ˆβ X y. = y y y X ˆβ. Thus, Mean Lag. (A general lag polynomial.) s 2 = RSS N K = y y ˆβ X y N K. δ(l) = δ (L) = δ j L j. j=0 jδ j L j 1. j=0 3

4 Then, Mean Lag δ (1) δ(1). Lagged dependents PLUS AR(1) errors yields inconsistent estimators! Autoregressive Final Form. 1.2 Characteristic Polynomials Consider an AR(2) process: Rewriting with the lag operator: A(L)y = B(L)X + ɛ. y = A 1 BX + A 1 ɛ. y = A A BX + A A ɛ. A y = A BX + A ɛ. y t = φ 1 y t 1 + φ 2 y t 2 + ɛ t. y t φ 1 Ly t φ 2 L 2 y t = ɛ t. (1 φ 1 L φ 2 L 2 )y t = ɛ t. Factorizing the left hand side: 1 φ 1 L φ 2 L 2 = (1 Z 1 L)(1 Z 2 L). Divide through by L 2 : Multiply out the right hand side: L 2 φ 1 L 1 φ 2 = L 2 (1 Z 1 L)(1 Z 2 L). Let Z = L 1 : L 2 φ 1 L 1 φ 2 = L 2 (1 Z 2 L Z 1 L + Z 1 Z 2 L 2 ). L 2 φ 1 L 1 φ 2 = L 2 Z 2 L 1 Z 1 L 1 + Z 1 Z 2. L 2 φ 1 L 1 φ 2 = L 2 (Z 1 + Z 2 )L 1 + Z 1 Z 2. L 2 φ 1 L 1 φ 2 = (L 1 Z 1 )(L 1 Z 2 ). Z 2 φ 1 Z φ 2 = (Z Z 1 )(Z Z 2 ). Thus the roots of the characteristic polynomial are Z 1 and Z 2. And explicitly: Z 1,2 = φ 1 ± φ φ

5 1.3 Instrumental Variables Consider a regression equation: y = Xβ + ɛ. If E[Xɛ] 0, then use Instrumental Variables (IV). Suppose there exists a set of instruments Z, for the problem X variable. We will use the 2 stage least squares technique to derive the IV estimator. Step 1: Run the following regression: The estimator will be the following form: X = Zγ + u. ˆγ = (Z Z) 1 Z X. Step 2: Compute the fitted values such that: Step 3: Run the following regression: The estimator will be the following: Substitute in for ˆX: Simplifying, Which is the IV estimator. ˆX = Zˆγ = Z(Z Z) 1 Z X. y = ˆXβ + ɛ. ˆβ = ( ˆX ˆX) 1 ˆX y. ˆβ = ((Z(Z Z) 1 Z X) Z(Z Z) 1 Z X) 1 (Z(Z Z) 1 Z X) y. ˆβ = (X Z(Z Z) 1 Z Z(Z Z) 1 Z X) 1 X Z(Z Z) 1 Z y. ˆβ IV = (X Z(Z Z) 1 Z X) 1 X Z(Z Z) 1 Z y. An important note. If X is exactly identified in that Z is the same dimension as X, then we can carry the analysis further. Inverting: ˆβ IV = (Z X) 1 (Z Z) (X Z) 1 X Z(Z Z) 1 Z y. }{{} I ˆβ IV = (Z X) 1 (Z Z)(Z Z) 1 Z y. }{{} I ˆβ IV = (Z X) 1 Z y. 5

6 1.4 Dicky Fuller Tests Use DF tables for tests of stationarity with: H 0 : φ = 1, Non Stationary. H 1 : φ < 1, Stationary. A Guide to Dicky Fuller Tables. First Panel Nothing y = φy t 1 + ɛ t Second Panel Constant y = α + φy t 1 + ɛ t (2) Third Panel Constant and Time Trend y = α + βt + φy t 1 + ɛ t Example. Consider the model: To test: Use: And check against middle panel. y t = α + φy t 1 + ɛ t. H 0 : φ = 1, α = 0, (random walk). H 1 : φ < 1, α <=> 0. τ u = ˆφ 1 SE( ˆφ). If test is anything other than H 0 : φ = 1, use t test! 1.5 Lagrange Multiplier Test We ll use the LM test to test for serially correlated errors. Consider the model: Test: y t = x tβ + u t, u t = φu t 1 + ɛ t, φ < 1, ɛ t iid. H 0 : φ = 0 (Errors NOT Serially Correlated). H 1 : φ 0 (Errors Serially Correlated). Rewrite the model by lagging once, multiplying by φ and subtracting, y t φy t 1 = x tβ φx t 1β + u t φu t 1. y t = φy t 1 + x tβ φx t 1β + ɛ t. 6

7 Log Likelihood: With, Ln(L) = (T 1) ln(2π) 2 (T 1) ln(σ 2 ) 1 2 2σ 2 ɛ t = y t φy t 1 x tβ + φx t 1β. T ɛ 2 t. t=2 First Order Conditions. ln(l) φ = 1 T ɛ σ 2 t ( y t 1 + x t 1β). t=2 ln(l) β = 1 T ɛ σ 2 t ( x t + φx t 1). t=2 To simplify, let: With ψ = (φ β). z t = ɛ t ψ. Thus, z t = ɛ t φ ɛ t β [ yt 1 + x = t 1β x t + φx t 1 ] [ yt 1 x = t 1β x t φx t 1 ]. (3) So the first order conditions simplify to: T z t ɛ t = 0. t=2 So the next step is to calculate the β coefficient under the null hypothesis of non-serially correlated errors. Run OLS of y t on x t and note the estimated coefficient ˆβ 0. Evaluate ɛ t and z t at the restricted values, ˆβ 0 and ˆφ 0 = 0. Thus, ˆɛ t = y t x t ˆβ 0. [ yt 1 x z t = ˆβ t 1 0 x t ]. (4) 7

8 Finally regress ˆɛ t on z t or run the regression: Take R 2 from this regression and test: 1.6 Test for Co-Integration ˆɛ t = y t 1 x ˆβ t 1 0 and x t. }{{} z t LM = T R 2 a χ 2 (1). y t I(1) is cointegrated with x t I(1) iff there exists a vector α, such that, y t α x t = u t I(0). To check, regress y t on x t and save the residuals, û t. û t needs to be I(0) if y t and x t are to be cointegrated. Regress: û t = φ 0 û t + ɛ t + φ j û t j. j }{{} Other Lags The other lags added on make this an Augmented Dicky Fuller Test. Test: H 0 : φ 0 = 0 (NOT Cointegrated). H 1 : φ 0 < 0 (Cointegrated). Why? If φ 0 = 0, then (ignoring the other lags), û t = ɛ t I(0). Thus û t I(0) which means that û t is not I(0). We don t know for sure that it s I(1), but it may be. ˆφ 0 So compute SE( ˆφ 0 ) and reject H 0 if it is less than the critical value in the MacKinnon tables. (NOTE: we can t use the DF tables here). More on Tests for Co-Integration. If u t is a stationary AR(1), then x and y will be conintegrated. Consider the equation for u t : u t = φu t 1 + ɛ t, ɛ t iid φ < 1. 8

9 Thus, u t u t 1 = φu t 1 u t 1 + ɛ t. u t = u t 1 (φ 1) + ɛ t. u t = u t 1 γ + ɛ t, γ = φ 1. Note that u t is stationary if φ < 1 and u t is stationary if γ < 1. Thus test: H 0 : γ = 0 = u t I(0) φ = γ+1 = 1 u t I(1) (x, y) are NOT conintegrated. H 1 : γ < 0 = φ = γ + 1 < 1 u t I(0) (x, y) are conintegrated. 1.7 Properties of Standard Processes AR(1) Model: y t = φy t 1 + ɛ t. Yields: Stationary if φ < 1. E[y t ] = 0. V ar(y t ) = E[yt 2 ] = Stationary Cov(y t, y t 1 ) = φσ2 1 φ 2. Cov(y t, y t s ) = φs σ 2 1 φ 2. σ2 1 φ 2. In another form after backward substitution: s 1 y t = φ s y t s + φ j ɛ t j MA(1) j=0 Model: Yields: y t = θɛ t 1 + ɛ t. E[y t ] = 0. V ar(y t ) = E[yt 2 ] = σ 2 (1 + θ 2 ). Cov(y t, y t 1 ) = θσ 2. Cov(y t, y t s ) = 0 s > 1. 9

10 1.7.3 Random Walk Model: Yields: y t = ɛ t = y t = y t 1 + ɛ t. E[y t ] = 0. V ar(y t ) = tσ Random Walk with Drift Model: Yields: y t = α + ɛ t = y t = α + y t 1 + ɛ t. E[y t ] = αt. V ar(y t ) = tσ Asymptotic Distribution of ˆβ We would first like to check that the ˆβ estimator is consistant. ˆβ = (X X) 1 X y. ˆβ = (X X) 1 X [Xβ + ɛ] = β + (X X) 1 X ɛ. ˆβ β = (X X) 1 X ɛ. Or rewriting: Taking the probability limit: [ 1 ˆβ β = n n ] 1 [ 1 x i x i n n x i ɛ i ]. [ [ plim( ˆβ 1 β) = plim n By Slutsky s Theorem: n ] 1 [ 1 x i x i n n ] ] x i ɛ i. plim( ˆβ β) = [ plim 1 n n ] 1 [ x i x i plim 1 n n x i ɛ i ]. Since x i x i is an iid sequence, and E[x i x i] = Σ xx, then by the Weak Law of Large Numbers (WLLN), 10

11 plim 1 n n x i x i = Σ xx. Also, x i ɛ i is an iid sequence with E[x i ɛ i ] = 0 because of independence. Thus, by the WLLN, plim 1 n n x i ɛ i = 0. Thus, substituting these last two equations in, plim( ˆβ β) = ] 1 [ ] [Σ xx 0 = 0. Thus, ˆβ is a consistant estimator. Note that the two substitutions could have also been done via the ergodic theorem which does not rely on iid but rather on stationarity and limited memory. (iid processes are always Ergodic). 1.9 Lagged Dependents and Serial Correlated Errors Consider the following model: Compute E[y t 1 u t ]. y t = γy t 1 + x tβ + u t, u t = φu t 1 + ɛ t. E[y t 1 u t ] = E[(γy t 2 + x t 1β + u t 1 )(φu t 1 + ɛ t )]. = γφe[y t 2 u t 1 ] + βφe[x t 1u t 1 ] + φe[u 2 t 1] + γe[y t 2 ɛ t ] + βe[x t 1ɛ t ] + E[u t 1 ɛ t ]. = γφe[y t 2 u t 1 ] + βφe[x t 1u t 1 ] + φe[u 2 t 1]. = γφe[y t 1 u t ] + βφe[x tu t ] + φe[u 2 t ]. = βφe[x tu t ] + φe[u 2 t ]. 1 γφ = φe[u2 t ] 1 γφ. = φv ar(u t) 1 γφ. = φ σ2 1 φ 2 1 γφ. = φσ 2 (1 φ 2 )(1 γφ) 0. 11

12 Thus, in this situation, OLS is biased. If φ is unknown, use the C-O transformation to estimate the true parameter, β. If φ is unknown (more realistic), then rewrite model by lagging, multiplying by φ and subtracting which gives us iid errors. Then do MLE on resulting equation. Note that running OLS will work here but you won t be able to distinguish the parameter φ from the β parameter. M LE is equivalent (linearly) to OLS and allows us to discern to two Final Review before Exam Variances. Roots. Roots solve: Thus, Roots solve: Note that Z = 1 satisfies this equation. T SS = RSS + ESS. V ar( ˆβ) = E[ ˆβ β ] 2. y t = φ 1 y t 1 + φ 2 y t 2 + ɛ t. y t φ 1 Ly t φ 2 L 2 y t = ɛ t. Z 2 φ 1 Z φ 2 = 0. y t = φ 1 y t 1 + φ 2 y t 2 + ɛ t. y t φ 1 Ly t φ 2 Ly t = ɛ t. (1 φ 1 L φ 2 L) y t = ɛ t. (1 φ 1 L φ 2 L)(y t y t 1 ) = ɛ t. (1 φ 1 L φ 2 L)(1 L)y t = ɛ t. (Z 2 φ 1 Z φ 2 )(Z 2 Z) = 0. R 2 = ESS T SS = ŷ ŷ y y = T SS RSS T SS = y X ˆβ y y. s 2 = RSS N K = y y y X ˆβ N K. s 2 ML = RSS N. Wald. ( R ˆβ ) ) 1 ( q (Rs 2 (X X) 1 R R ˆβ ) q r F (r, N K). 12

13 Likelihood: LM test. Π T t [ ] 1 ( exp ɛ ) t. 2πσ 2 2σ 2 [ 2 log L ] I(ψ) = E. ψ ψ V ar( ˆψ) = I( ˆψ) [ 1 2 log L ] 1 = E. ψ ψ Under R(ψ) = 0, LM = log L 1 log L I( ˆψ) ψ ψ χ2 (q). Note that I(ψ) is block diagonal with the block corresponding to β as: 1 zt z t. ˆσ 2 0 Thus we can also write the LM stat (seen from regression of ɛ t on z t and computing T R 2 ): LM = ( ) ( ) 1 ( ) zt ɛ t zt z t zt ɛ t. ˆσ 2 0 Note: log L( ˆψ 0 ) = 1ˆσ ( z ψ 0 2 t ɛ t ). 13