Single-Equation GMM: Endogeneity Bias

Transcription

1 Single-Equation GMM: Lecture for Economics 241B Douglas G. Steigerwald UC Santa Barbara January 2012

2 Initial Question Initial Question How valuable is investment in college education? economics - measure value in terms of wage How would you determine the return on investment in college education?

3 Framework Stochastic Model What are the returns to a college education? Random variables of interest W - log of worker wage S - years of schooling A - age M - indicator for male R - indicator for white U - other factors that a ect wages Stochastic Model W = β 0 + β 1 S + β 2 A + β 3 A 2 + β 4 M + β 5 R + U

4 Initial Question Answered? Estimates Stochastic Model W = β 0 + β 1 S + β 2 A + β 3 A 2 + β 4 M + β 5 R + U ˆβ (and signi cantly di erent from zero) each additional year of schooling is worth an additional 8.4% in wages 4 years of college would increase wages by 38% ( ) the median full time worker earns about $550 per week in 2000 wage increase of 38% is $210 per week over 30 year work-life, earnings increase by $170,000 in present value (5% interest), makes public universities a good deal

5 Initial Question Answered? Potential Endogeneity W = β 0 + β 1 S + β 2 A + β 3 A 2 + β 4 M + β 5 R + U β 1 may not capture a causal impact on wages workers who obtained more education may have attributes that would have led to higher earnings even without additional education S is endogenous ) Cov (S, U) 6= 0 ˆβ 1 is biased and inconsistent What is the direction of bias in ˆβ 1? ˆβ 1 is biased upward, does not provide a helpful bound to argue for bene ts of education

6 Background Sources of Covariate-Error Correlation Focus on Cor (S, U) Endogeneity workers who would otherwise have high wage rates are more likely to obtain higher education U = µ + e µ - ability e - random shock S = αµ + w describes how schooling is correlated with ability in this application, likely that Cor (S, U) > 0 Measurement Error S = S + v S - actual schooling S - reported v - measurement error in all applications, Cor (S, U) < 0

7 Background Detail: Measurement Error Correlation Simplify population model W t = βst + U t S t = St + v t estimated model W t = βs t + (U t βv t ) v t is a component of S t ) Cor [S t, (U t βv t )] < 0 in large samples ˆβ tends to Cov (S t, v t ) β 1 Var (S t ) ˆβ = β + n t=1 S t [U t βv t ] n t=1 S 2 t Var (S = t ) β Var (St ) + Var (v t ) where Var (S t ) = Var (S t ) + Var (v t ) Cov (S t, v t ) = Var (v t ) Iron Law of Econometrics - measurement error leads to attenuation bias

8 Background Solutions Instrument z is a (valid) instrument if Cov (S, z) 6= 0 and Cov (U, z) = 0 instruments can address both sources of covariate-error correlation issue - instruments can be di cult to nd Measurement error assumption S = S + v assumptions regard v example: v is symmetric around 0 issue - does not address endogeneity

9 Background Instrument Solutions Standard Instrument Solution implicit model of endogeneity no speci ed model linking endogenous covariates to error yields classic instrumental variable (IV) estimator Model-Based Selection (Endogeneity) Correction explicit model of endogeneity clearly speci ed model linking endogenous covariates to error yields selection-corrected IV estimator

10 Background Standard Instrument Solution : Identi cation X (K 1) covariate vector Z (L1) instrument vector Identi cation Assumption (Rank Condition) The L K matrix E ZX T has rank K. Example X T = (1, S) Z T = (1, z) E ZX T = 1 E (S) E (z) E (Sz) Rank is K if determinant is not zero, Cov (S, z) 6= 0

11 Background Identi cation Identi cation Assumption (Order Condition) There are at least as many instruments as endogenous covariates: L K. Over identi cation rank condition satis ed and L > K Exact identi cation rank condition satis ed and L = K No identi cation L < K (rank condition cannot hold)

12 Initial Question Revisited Selection (Endogeneity) Correction Key - construct E [UjX, Z ] add to regression, remaining error uncorrelated with covariates Wage Regression Application data on twins (indexed by i) who share family characteristics Selection (Endogeneity) Model U i = µ + ε i µ = γs 1 + γs 2 + ω µ - latent family characteristics, correlated with S could relax assumption that γ is constant (use equation for twin 1 to identify γ 2 ) γ - selection e ect : γ > 0 ) families that would otherwise have high wages are more likely to educate their children

13 Initial Question Revisited Selection Correction wage regression (twin 1 C 0 1 = A 1, A 2 1, M 1, R 1 ) W 1 = β 0 + β 1 S 1 + C 0 1δ + (µ + ε 1 ) = β 0 + β 1 S 1 + C 0 1δ + (γs 1 + γs 2 + ω + ε 1 ) identi cation assumption E [U 1 jx, Z ] = γs 1 + γs 2 selection-corrected regression W 1 = β 0 + (β 1 + γ) S 1 + γs 2 + C 0 1δ + (ε 1 + ω) Variable OLS Include S 2 Own education (0.014) (0.015) Sibling s education (0.015) Twins data - endogeneity bias is negative!

14 Review Review Stochastic Model W i = β 0 + β 1 S i + U i What two issues lead to correlation between S i and U i? 1) endogeneity - latent ability that impacts both S i and U i 2) measurement error in S i What is the impact of the correlation on B OLS? biased and inconsistent What is needed to construct a consistent estimator? 1) instrument Z i Cor (Z i, S i ) 6= 0 Cor (Z i, U i ) = 0 2) assumption about measurement error