Day 2A Instrumental Variables, Two-stage Least Squares and Generalized Method of Moments

Transcription

1 Day 2A nstrumental Variables, Two-stage Least Squares and Generalized Method of Moments c A. Colin Cameron Univ. of Calif.- Davis Frontiers in Econometrics Bavarian Graduate Program in Economics. Based on A. Colin Cameron and Pravin K. Trivedi (2009, 2010), Microeconometrics using Stata (MUS), Stata Press. and A. Colin Cameron and Pravin K. Trivedi (2005), Microeconometrics: Methods and Applications (MMA), C.U.P. March 21-25, 2011 c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 1 Colin / 35 C

2 1. ntroduction 1. ntroduction Problem: OLS inconsistent in model y i = xi 0β + u i if Cov[x i, u i ] 6= 0. Solution: Assume there are instruments z i satisfying Cov[z i, u i ] = 0. f #instruments = #regressors instrumental variables (V) estimator f #instruments > #regressors then use two-stage least squares (2SLS) generalized method of moments (GMM). Complications test of assumptions (exogeneity, endogeneity) weak instruments c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 2 Colin / 35 C

3 1. ntroduction Overview 1 ntroduction. 2 V, 2SLS, GMM: De nitions 3 Data Example 4 nstrumental variable methods in practice 5 V Estimator Properties 6 Nonlinear GMM 7 Endogeneity in nonlinear models 8 Stata 9 Appendix: nstrumental Variables ntuition c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 3 Colin / 35 C

4 2. V, 2SLS and GMM Estimators De nitions 2. V, 2SLS and GMM estimators: De nitions Model is y i = x 0 i β + u i OLS is inconsistent as Cov[x i, u i ] 6= 0. Assume there are instruments z i such that Cov[z i, u i ] = 0. Then Cov[z i, u i ] = 0 ) E[z i u i ] = 0 given E[u i jz i ] = 0. We have the population moment condition E[z i (y i x 0 i β)] = 0. Method of moments: solve the corresponding sample moment condition 1 N N i=1 z i (y i x 0 i β) = 0. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 4 Colin / 35 C

5 2. V, 2SLS and GMM Estimators nstrumental Variables Estimator nstrumental variables (V) estimator n just-identi ed case (# instruments = # regressors) solve k equations in k unknowns N 1 i z i (y i xi 0 β) = 0 gives the instrumental variables (V) estimator. bβ V = 1 N i =1 z0 i x i N i =1 z0 i y i = (Z 0 X) 1 Z 0 y estimate using Stata 10 command ivregress 2sls Often just one regressor in x i is endogenous (i.e. correlated with u i ). Then one variable in z i is the instrument for this endogenous regressor. the remaining entries in z i are the exogenous variables i.e. exogenous variables are instruments for themselves. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 5 Colin / 35 C

6 2. V, 2SLS and GMM Estimators GMM Estimator Generalized method of moments estimator n over-identi ed case (# instruments > # regressors) Cannot solve N 1 i z i (y i xi 0 β) = 0. nstead generalized method of moments (GMM) estimator minimizes the quadratic form in N 1 N i =1 z i (y i xi 0 β) Q(β) = 1 0 N i y i xi 0 1 β z i W N N i y i xi 0 β z i = (Z 0 u) 0 W(Z 0 u) The symmetric full-rank weighting matrix W does not depend on β. Then Q(β)/ β = 0 yields the GMM estimator bβ GMM = 1 i x i zi 0 W N N i=1 z i xi 0 i x i zi 0 W N N i=1 z i y i = (X 0 ZW N Z 0 X) 1 X 0 ZW N Z 0 y. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 6 Colin / 35 C

7 2. V, 2SLS and GMM Estimators Optimal GMM and 2SLS Optimal GMM and 2SLS The variance of b β GMM is smallest when the optimal weighting matrix W N is consistent for (Var[Z 0 u]) 1 Though in the just-identi ed (r = K ) GMM = V for any W N. For homoskedastic errors Var[Z 0 u] = σ 2 N i=1 z 0 i z i Two-stage least squares (2SLS) estimator sets W N = ( N i =1 z0 i z i ) 1 Yields β b 2SLS = (X 0 Z(Z 0 Z) 1 Z 0 X) 1 X 0 Z(Z 0 Z) 1 Z 0 y Estimate using Stata 10 command ivregress 2sls but use robust VCE to guard against errors not homoskedastic. For heteroskedastic errors Var[Z 0 u] = σ 2 N i=1 z 0 i z i Optimal GMM estimator if errors are heteroskedastic errors sets W N = ( N i =1 bu2 i z0 i z i ) 1, bu i = y i x 0 i b β 2SLS estimate using Stata 10 command ivregress gmm. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 7 Colin / 35 C

8 2. V, 2SLS and GMM Estimators More on 2SLS More on 2SLS 2SLS gets it s name because it can be computed in two-stages. Suppose y 1 depends in part on scalar y 2 which is endogenous Structural equation for y 1 First-stage equation for y 2 y 1i = β 1 y 2i + x1i 0 β 2 + u i y 2i = x1i 0 π 1 + x2i 0 π 2 + v i where x 2 is one or more instruments for y 2 in earlier notation x i = (y 2i x 0 1i )0 and z i = (x 0 1i x 0 2i )0. OLS of y 1 on y 2 and x 1 is inconsistent. 2SLS can be computed as follows 1. First-stage: by 2 as prediction from OLS of y 2 on x 1 and x Structural: Do OLS of y 2 on by 2 and x 2. But this method does not generalize to nonlinear models. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 8 Colin / 35 C

9 3. Data Example Drug expenditures 3. Data Example: Drug expenditures Example from MUS chapter 6. Drug expenditures for U.S. elderly (ldrugexp) regressed on endogenous private health insurance dummy (hi_empunion) and exogenous regressors de ned by global x2list.. * Read data, define global x2list (exogenous regressors), and summarize d. use mus06data.dta. global x2list totchr age female blhisp linc. keep if linc!=. (302 observations deleted). describe ldrugexp hi_empunion $x2list storage display value variable name type format label variable label ldrugexp float %9.0g log(drugexp) hi_empunion byte %8.0g nsured thro emp/union totchr byte %8.0g Total chronic cond age byte %8.0g Age female byte %8.0g Female blhisp float %9.0g Black or Hispanic linc float %9.0g log(income) c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program in March Economics 21-25,. Based 2011 on A. 9 Colin / 35 C

10 3. Data Example Drug expenditures Summary statistics. summarize ldrugexp hi_empunion $x2list Variable Obs Mean Std. Dev. Min Max ldrugexp hi_empunion totchr age female blhisp linc Sample is % have employer or union-sponsored health insurance. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 10 Colin / 35 C

11 3. Data Example OLS Estimates OLS estimates OLS is inconsistent if hi_empunion endogenous. * OLS. regress ldrugexp hi_empunion $x2list, vce(robust) Linear regression Number of obs = F( 6, 10082) = Prob > F = R-squared = Root MSE = Robust ldrugexp Coef. Std. Err. t P> t [95% Conf. nterval] hi_empunion totchr age female blhisp linc _cons Drug expenditure increases by 7.4% if have private insurance. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 11 Colin / 35 C

12 3. Data Example nstruments nstruments A valid instrument for private health insurance (hi_empunion) must not be directly a cause of ldrugexp (so uncorrelated with u i ) i.e. must not belong in the model for ldrugexp and to be relevant should be correlated with hi_empunion Possible instrument 1 ssiratio = social security income income from all other sources need to assume that the direct role of income is adequately captured by the regressor linc Possible instrument 2 multlc = 1 if rm has multiple locations need to assume that rm size does not e ect ldrugexp c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 12 Colin / 35 C

13 3. Data Example nstruments Two possible instruments ssiratio and multlc. * Two available instruments for hi_empunion. describe ssiratio multlc storage display value variable name type format label variable label ssiratio float %9.0g SS/ncome ratio multlc byte %8.0g Multiple locations. summarize ssiratio multlc Variable Obs Mean Std. Dev. Min Max ssiratio multlc correlate hi_empunion ssiratio multlc (obs=10089) hi_emp~n ssiratio multlc hi_empunion ssiratio multlc Correlation between z and x is low e.g. Cor[z, x] = 0.21 for ssiratio c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 13 Colin / 35 C

14 3. Data Example V estimates V estimates V estimates using the single instrument ssiratio for hi_empunion. * V estimator with ssiratio as single instrument for hi_empunion. ivregress 2sls ldrugexp (hi_empunion = ssiratio) $x2list, vce(robust) nstrumental variables (2SLS) regression Number of obs = Wald chi2(6) = Prob > chi2 = R-squared = Root MSE = Robust ldrugexp Coef. Std. Err. z P> z [95% Conf. nterval] hi_empunion totchr age female blhisp linc _cons nstrumented: hi_empunion nstruments: totchr age female blhisp linc ssiratio Coe cient even changes sign, from (OLS) to (V). Standard error increases from (OLS) to (V). c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 14 Colin / 35 C

15 3. Data Example 2SLS estimates 2SLS Estimates Overidenti ed as two instruments ssiratio and multlc. * 2SLS estimator with ssiratio and multlc as instruments for hi_empunion. ivregress 2sls ldrugexp (hi_empunion = ssiratio multlc) $x2list, vce(robust) nstrumental variables (2SLS) regression Number of obs = Wald chi2(6) = Prob > chi2 = R-squared = Root MSE = Robust ldrugexp Coef. Std. Err. z P> z [95% Conf. nterval] hi_empunion totchr age female blhisp linc _cons nstrumented: hi_empunion nstruments: totchr age female blhisp linc ssiratio multlc Coe cient changes from (V) to (2SLS). Standard error decreases from (V) to (2SLS). c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 15 Colin / 35 C

16 3. Data Example Optimal GMM Estimates Optimal GMM Two instruments ssiratio and multlc optimal GMM if errors are heteroskedastic and start with E[zu] = 0.. * GMM estimator with ssiratio and multlc as instruments for hi_empunion. ivregress gmm ldrugexp (hi_empunion = ssiratio multlc) $x2list, vce(robust) nstrumental variables (GMM) regression Number of obs = Wald chi2(6) = Prob > chi2 = R-squared = GMM weight matrix: Robust Root MSE = Robust ldrugexp Coef. Std. Err. z P> z [95% Conf. nterval] hi_empunion totchr age female blhisp linc _cons nstrumented: hi_empunion nstruments: totchr age female blhisp linc ssiratio multlc Estimate and standard error for hi_empunion are very similar to 2SLS Little e ciency gain compared to 2SLS. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 16 Colin / 35 C

17 3. Data Example Estimator comparison Estimator comparison Compare OLS, V, 2SLS (over-identi ed), GMM (over-identi ed). * Compare estimators. quietly regress ldrugexp hi_empunion $x2list, vce(robust). estimates store OLS. quietly ivregress 2sls ldrugexp (hi_empunion = ssiratio multlc) $x2list, vce(robust). estimates store V. quietly ivregress 2sls ldrugexp (hi_empunion = ssiratio) $x2list, vce(robust). estimates store TWOSLS. quietly ivregress gmm ldrugexp (hi_empunion = ssiratio multlc) $x2list, vce(robust). estimates store GMM. estimates table OLS V TWOSLS GMM, b(%9.4f) se(%9.3f) stats(n r2 F) Variable OLS V TWOSLS GMM hi_empunion totchr age female blhisp linc _cons N r F legend: b/se c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 17 Colin / 35 C

18 4. nstrumental variables in practice 4. nstrumental variables methods in practice Do we need to use instruments? Hausman test of endogeneity. s the instrument valid (uncorrelated with the error)? f model is over-identi ed can do over-identifying restrictions test. What if the instrument is weakly correlated with regressor instrumented Lose e ciency f really weak can have nite-sample bias and wrong test size. How many instruments? Need # instruments # endogenous regressors. n theory more is better but too many can have nite-sample bias. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 18 Colin / 35 C

19 4. nstrumental variables in practice Hausman test of endogeneity Hausman test n general a Hausman test considers two di erent estimators bθ andbθ that have the same plim under H 0 and di erent plim 0 s under H a. H 0 : plim(bθ eθ) = 0 versus H a : plim(bθ eθ) 6= 0. We reject H 0 if the di erence is large, using H = (bθ eθ) 0 (bv[bθ eθ]) 1 (bθ eθ) a χ 2 (q). Tricky bit is estimating V[bθ eθ] = V[bθ]+ V[eθ] 2Cov[bθ, eθ] usual Hausman test implementation assumes one of bθ and eθ is fully e cient under the null. Say eθ: then V[bθ eθ] = V[bθ] V[eθ] such e ciency is not usually the case in practice F e.g. if errors are heteroskedastic then OLS is ine cient instead need to use a robust form of the Hausman test. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 19 Colin / 35 C

20 4. nstrumental variables in practice Hausman test of endogeneity Hausman test of endogeneity: 2SLS (bθ) versus OLS (bθ) H 0 : plim(bθ 2SLS eθ OLS ) = 0 if exogeneity vs. H a : plim(bθ 2SLS eθ OLS ) 6= 0 if endogeneity Use heteroskedasticity-robust version of Hausman test this is command estat endogenous and not hausman. * Robust version of Hausman test using augmented regression. quietly ivregress 2sls ldrugexp (hi_empunion = ssiratio) $x2list, vce(robust). estat endogenous Tests of endogeneity Ho: variables are exogenous Robust score chi2(1) = (p = ) Robust regression F(1,10081) = (p = ) Reject H 0 as p = Conclude that hi_empunion is endogenous. Need to do V. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 20 Colin / 35 C

21 4. nstrumental variables in practice Test of instrument validity Test of instrument validity Cannot test validity in a just identi ed model ntuition: Test based on Cov[z i, bu i ] ' 0 requires bu i based on a consistent estimator of β which requires at least just-identi ed model. Test of overidentifying restrictions (for over-identi ed model) Test H 0 : E[zi 0u i ] = 0 by testing if N 1 i zi 0bu i ' 0. Limited test as assumes instruments in just-identi ed model are valid. n Stata command estat overid after command ivregress gmm. Here one over-identifying restriction (2 instruments for 1 endogenous). * Test of overidentifying restrictions following ivregress gmm. quietly ivregress gmm ldrugexp (hi_empunion = ssiratio multlc) $x2list, wmatrix(robust). estat overid Test of overidentifying restriction: Hansen's J chi2(1) = (p = ) Do not reject H 0 as p = 0.31 < Conclude that, assuming the just-identifying restriction is valid, then the over-identifying restriction is also valid. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 21 Colin / 35 C

22 4. nstrumental variables in practice Weak instruments Weak instruments Weak instrument means that instrument(s) are weakly correlated with endogenous regressor(s), after controlling for exogenous regressors. Then 1. standard errors " greatly as 2SLS much less e cient than OLS. 2. even slight correlation between error and the instrument can lead to 2SLS more inconsistent than OLS. 3. even if instrument(s) are valid so 2SLS is inconsistent, in typical sample sizes usual asymptotic theory can provide a poor approximation e.g. bias. Consequences 1. key coe cient estimate(s) can become statistically insigni cant. 2. even more important to ensure that the instrument is valid. 3. focus of the weak instrument literature. n Stata for 3. use command estat firststage after command ivregress add-on commands condivreg and ivreg2 c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 22 Colin / 35 C

23 5. V Estimator Properties Consistency 5. V estimator properties: consistency Stacking all observations bβ V = Z 0 X 1 Z 0 y. Substitute y = Xβ + u for y yields bβ V = Z 0 X 1 Z 0 [Xβ + u] = β + Z 0 X 1 Z 0 u = β + 1 N Z0 X 1 1 N Z0 u So β b p V! β and βv b is consistent for β if plim N 1 Z0 u = 0 (instruments are valid) and plim N 1 Z0 X 6= 0 (instruments are relevant). c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 23 Colin / 35 C

24 5. V Estimator Properties Asymptotic distribution V estimator: asymptotic distribution nformal derivation: Thus bβ GMM β = (Z 0 X) 1 Z 0 u a (Z 0 X) 1 N [0, V[Z 0 u]] a (Z 0 X) 1 N [0, Z 0 V[ujZ]Z] a (Z 0 X) 1 N [0, Z 0 ΩZ] bβ V a N [β, (Z 0 X) 1 Z 0 ΩZ(X 0 Z) 1 ]; Ω = V[ujZ]. With independent heteroskedastic errors (Stata option vce(robust)) bv[ b β V ] = (Z 0 X) 1 Z 0 bωz(x 0 Z) 1 ; bω = Diag[bu 2 i ]. Note: Cor[Z, X] ) Z 0 X small ) (Z 0 X) 1 large ) b β V is imprecise. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 24 Colin / 35 C

25 5. V Estimator Properties Asymptotic distribution for GMM Asymptotic Distribution of GMM nformal derivation: Thus bβ GMM = (X 0 ZWZ 0 X) 1 X 0 ZWZ 0 (Xβ + u) bβ GMM β = (X 0 ZWZ 0 X) 1 X 0 ZWZ 0 u a (X 0 ZWZ 0 X) 1 X 0 ZW N [0, V[Z 0 u]] a (X 0 ZWZ 0 X) 1 X 0 ZW N [0, Z 0 V[ujZ]Z] a (X 0 ZWZ 0 X) 1 X 0 ZW N [0, Z 0 ΩZ] a bβ GMM N [β, (X 0 ZWZ 0 X) 1 X 0 ZWZ 0 ΩZWZ 0 X(X 0 ZWZ 0 X) 1 ] Ω = V[ujZ]. Optimal W is a consistent estimate of Ω 1. Then bβ OptGMM a N [β, (X 0 ZΩ 1 Z 0 X) 1 ] c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 25 Colin / 35 C

26 6. Nonlinear GMM 6. Nonlinear GMM estimator: De nition Nests LS, MLE, V, GMM,... The way to view estimation. Population unconditional moment condition E[h(w, θ 0 )] = 0; w = (y, x, z) is all observables. bθ solves the corresponding sample moment condition 1 N N i=1 h(w i, bθ) = 0. just-identi ed case (r = q) can solve for β over-identi ed case (r > q) cannot as r equations in k unknowns. The generalized method of moments (GMM) estimator (for r > q) minimizes the quadratic form in N 1 i h(w i, θ) Q(θ) = N N i=1 h(w i, θ) W N N N i=1 h(w i, θ) = g(θ) 0 W N g(θ) where g(θ) = N i =1 h i (θ) and 1 N N i =1 W N is a symmetric full-rank weighting matrix that does not depend on θ. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 26 Colin / 35 C

27 6. Nonlinear GMM Properties Nonlinear GMM estimator: Properties bθ GMM is asymptotically normally distributed with V[bθ GMM ] = N(G 0 WG) 1 G 0 WSWG(G 0 WG) 1. where G = lim E h gn (θ) 0 θ i h = lim E 1 N N i=1 i h i (θ) 0 θ S = Var[ p Ng N (θ)] = Var h 1 p N N i=1 h(w i, θ) i. Optimal GMM: W N = bs 1 where bs p! S. Similar issues as for weighted LS in the linear model. Model choice: specify moment conditions for estimation. Estimator choice: specify a weighting function. Statistical inference: use robust standard errors. Most e cient estimator: a particular choice of weighting function. n Stata 11 use the new command gmm. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 27 Colin / 35 C

28 7. Endogeneity in Nonlinear Models 7. Endogeneity in nonlinear models Example is y i = exp(x 0 i β) + u i and Cov[x i, u i ] 6= 0. Several very di erent methods (and associated models) exist. 1. Nonlinear V (often called nonlinear 2SLS) is nonlinear GMM based on E[z i u i ] = 0 and W = (Z 0 Z) Control function: add estimated rst-stage error bv i as regressor. di ers from 1. in nonlinear models 3. Fully structural approach adds an equation for endogenous regressors and estimates the model Di ers from 1. and 2. in most nonlinear models and is computationally di cult. 4. The following is inconsistent in nonlinear models: get bx i from rst stage regressions and estimate y i = exp(bx i 0 β) + error. The two-stage LS interpretation of 2SLS does not carry over to nonlinear models. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 28 Colin / 35 C

29 8. Stata commands 8. Stata commands V (just-identi ed) 2SLS (over-identi ed) GMM (over-identi ed) Overidentifying restrictions test Hausman test (if i.i.d. error) Hausman test (if heteroskedastic error) Weak instruments (plus user written commands) Static panel V Dynamic panel V Nonlinear GMM (new in Stata 11) ivregress 2sls ivregress 2sls ivregress gmm xtoverid hausman estat endogenous estat firststage condivreg; ivreg2 xtivreg; xthaustaylor xtabond; xtdpdsys; xtdpd gmm c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 29 Colin / 35 C

30 9. Appendix V ntuition 9. Appendix: nstrumental variables ntuition Simplify to scalar regression of y on single regressor x (no intercept). Linear regression model y = βx + u where u is an error term. n general E[yjx] = βx+ E[ujx]. Standard regression: assume E[ujx] = 0 i.e. regressors uncorrelated with error implies the following path analysis diagram x! y % u where there is no association between x and u. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 30 Colin / 35 C

31 9. Appendix V ntuition But there may be an association between regressors and errors. Example: regression of earnings (y) on years of schooling (x). The error u embodies all factors other than schooling that determine earnings, such as ability. Suppose a person has high u, due to high (unobserved) ability. This increases earnings, since y = βx + u. But it may also increase x, since schooling is likely to be higher for those with high ability. So high u (1) directly increases y and (2) indirectly increases y via higher x. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 31 Colin / 35 C

32 9. Appendix V ntuition The path analysis diagram becomes x! y " % u where now there is an association between x and u. Then y = βx + u(x) implies dy dx = β + du dx. OLS is inconsistent for β as it measures dy/dx, not just β. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 32 Colin / 35 C

33 9. Appendix V ntuition Assume there exists an instrument z that has the properties changes in z do not directly lead to changes in y changes in z are associated with changes in x The path analysis diagram becomes z! x! y " % u Note: z does not directly cause y, though z and y are correlated via indirect path of z being correlated with x which in turn determines y. Formally, z is an instrument for regressor x if (1) z is uncorrelated with the error u ; and (2) z is correlated with the regressor x. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 33 Colin / 35 C

34 9. Appendix V ntuition Example: a one unit change in the instrument z is associated with 0.2 more years of schooling (x) and $500 increase in annual earnings (y) (due to z " ) x " ) y ".) Then 0.2 years extra schooling is associated with $500 extra earnings. So a one year increase in schooling is associated with a $500/0.2 = $2, 500 increase in earnings. The causal estimate of β is therefore c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 34 Colin / 35 C

35 9. Appendix V ntuition Mathematically we estimated changes dx/dz and dy/dz and calculated the causal estimator as β V = dy/dz dx/dz. dy/dz estimated by OLS of y on z with slope estimate (z 0 z) 1 z 0 y dx/dz estimated by OLS of x on z with slope estimate (z 0 z) 1 z 0 x. The V estimator is bβ V = (z0 z) 1 z 0 y (z 0 z) 1 z 0 x = (z 0 x) 1 z 0 y 1 = N i=1 z i x i N i=1 z i y i. c A. Colin Cameron Univ. of Calif.- Davis (Frontiers BGPE in Econometrics Course: V, 2SLS, Bavarian GMMGraduate Program inmarch Economics 21-25, Based on A. 35 Colin / 35 C