Instrumental Variables & 2SLS PDF

Instrumental Variables & 2SLS y1 = β0 + β1y2 + β2z1 +... βkzk + u y2 = π0 + π1zk+1 + π2z1 +... πkzk + v Economics 20 - Prof. Schuetze 1 Why Use Instrumental Variables? Instrumental Variables (IV) estimation is used when your model has endogenous x’s i.e. whenever Cov(x,u)...

Instrumental Variables & 2SLS y1 = β0 + β1y2 + β2z1 +... βkzk + u y2 = π0 + π1zk+1 + π2z1 +... πkzk + v Economics 20 - Prof. Schuetze 1 Why Use Instrumental Variables? Instrumental Variables (IV) estimation is used when your model has endogenous x’s i.e. whenever Cov(x,u) ≠ 0 Thus, IV can be used to address the problem of omitted variable bias Also, IV can be used to solve the classic errors-in- variables problem Economics 20 - Prof. Schuetze 2 What Is an Instrumental Variable? In order for a variable, z, to serve as a valid instrument for x, the following must be true 1. The instrument must be exogenous - i.e. Cov(z,u) = 0 -more specifically z should have no “partial” effect on y and should be uncorrelated with u 2. The instrument must be correlated with the endogenous variable x - i.e Cov(z,x) 0 Economics 20 - Prof. Schuetze 3 Difference between IV and Proxy? With IV we will leave the unobserved variable in the error term but use an estimation method that recognizes the presence of the omitted variable With a proxy we were trying to remove the unobserved variable from the error term e.g. IQ IQ would make a poor instrument as it would be correlated with the error in our model (ability in u) Need something correlated with education but uncorrelated with ability (parents education?) Economics 20 - Prof. Schuetze 4 More on Valid Instruments We can’t test if Cov(z,u) = 0 as this is a population assumption Instead, we have to rely on common sense and economic theory to decide if it makes sense However, we can test if Cov(z,x) ≠ 0 using a random sample Simply estimate x = π0 + π1z + v, and test H0: π1 = 0 Refer to this regression as the “first-stage regression” Economics 20 - Prof. Schuetze 5 IV Estimation in the Simple Regression Case We can show that if we have a valid instrument we can get consistent estimates of the parameters For y = β0 + β1x + u Cov(z,y) = β1Cov(z,x) + Cov(z,u), so Thus, β1 = Cov(z,y) / Cov(z,x), since Cov(z,u)=0 and Cov(z,x) ≠ 0 Then the IV estimator for β1 is βˆ1 = ∑ (z i − z )( y i − y ) Same as OLS ∑ (z i − z )( x i − x ) when z = x Economics 20 - Prof. Schuetze 6 Inference with IV Estimation The IV estimator also has an approximate normal distribution in large samples To get estimates of the standard errors we need a slightly different homoskedasticity assumption: E(u2|z) = σ2 = Var(u) (conditioning on z here) If this is true, we can show that the asymptotic variance of β1-hat is: σx2 is the pop variance of x Var (βˆ1 ) = 2 σ σ 2 is the pop variance of u 2 2 nσ x ρ x,z ρ2 xz is the square of the pop correlation between x and z Economics 20 - Prof. Schuetze 7 Inference with IV Estimation Each of the elements in the population variance can be estimated from a random sample The estimated variance is then: Var (βˆ1 ) = 2 σˆ 2 SSTx R x, z Where σ2 = SSR from IV divided by the df, SSTx is the sample variance in x and the R2 is from the first stage regression (x on z) The standard error is just the square root of this Economics 20 - Prof. Schuetze 8 IV versus OLS estimation Standard error in IV case differs from OLS only in the R2 from regressing x on z Since R2 < 1, IV standard errors are larger However, IV is consistent, while OLS is inconsistent, when Cov(x,u) ≠ 0 Notice that the stronger the correlation between z and x, the smaller the IV standard errors Economics 20 - Prof. Schuetze 9 The Effect of Poor Instruments What if our assumption that Cov(z,u) = 0 is false? The IV estimator will be inconsistent also We can compare the asymptotic bias in OLS to that in IV in this case: ˆ Corr ( z, u) σ u IV : plimβ 1 = β 1 + Corr ( z, x) σ x ~ σu OLS : plim β 1 = β 1 + Corr ( x, u ) σ x Even if Corr(z,u) is small the inconsistency can be large if Corr(z,x) is also very small Economics 20 - Prof. Schuetze 10 Effect of Poor Instruments (cont) So, it is not necessarily better to us IV instead of OLS even if z and u are not “highly” correlated Instead, prefer IV only if Corr(z,u)/Corr(z,x) < Corr(x,u) Also notice that the inconsistency gets really large if z and x are only loosely correlated Best to test for correlation in the first stage regression Economics 20 - Prof. Schuetze 11 A Note on R2 in IV R2 after IV estimation can be negative Recall that R2 = 1 - SSR/SST where SSR is the residual sum of IV residuals SSR in this case can be larger than SST making the R2 negative Thus, R2 isn’t very useful here and can’t be used for F-tests Not important as we would prefer consistent estimates of the coefficients Economics 20 - Prof. Schuetze 12 IV Estimation in the Multiple Regression Case IV estimation can be extended to the multiple regression case Estimating: y1 = β0 + β1 y2 + β2 z1 + u1 Where y2 is endogenous and z1 is exogenous Call this the “structural model” If we estimate the structural model the coefficients will be biased and inconsistent Thus, we need an instrument for y2 Can we use z1 if it is correlated with y2 (we know it isn’t correlated with u1)? Economics 20 - Prof. Schuetze 13 Multiple Regression IV (cont) No, because it appears in the structural model Instead, we need an instrument, z2, that: 1. Doesn’t belong in the structural model 2. Is uncorrelated with u1 3. Is correlated with y2 in a particular way - Now because of z1 we need a partial correlation - i.e. for the “reduced form equation” y2 = π0 + π1 z1 + π2 z2 + v2, π2 ≠ 0 If we have such an instrument and u1 is uncorrelated with z1 the model is “identified” Economics 20 - Prof. Schuetze 14 Two Stage Least Squares (2SLS) It is possible to have multiple instruments Consider the structural model, with 1 endogenous, y2, and 1 exogenous, z1, RHS variable Suppose that we have two valid instruments, z2 and z3 Since z1, z2 and z3 are uncorrelated with u1, so is any linear combination of these Thus, any linear combination is also a valid instrument Economics 20 - Prof. Schuetze 15 Best Instrument The best instrument is the one that is most highly correlated with y2 This turns out to be a linear combination of the exogenous variables The reduce form equation is: y2 = π0 + π1 z1 + π2 z2 + π3 z3 + v2 or y2 = y2* + v2 Can think of y2* as the part of y2 that is uncorrelated with u1 and v2 as the part that might be correlated with u1 Thus the best IV for y2 is y2* Economics 20 - Prof. Schuetze 16 More on 2SLS We can estimate y2* by regressing y2 on z1, z2 and z3 – the first stage regression If then substitute 2 for y2 in the structural model, get same coefficient as IV While the coefficients are the same, the standard errors from doing 2SLS by hand are incorrect Also recall that since the R2 can be negative F- tests will be invalid Stata will calculate the correct standard error and F-tests Economics 20 - Prof. Schuetze 17 More on 2SLS (cont) We can extend this method to include multiple endogenous variables However, we need to be sure that we have at least as many excluded exogenous variables (instruments) as there are endogenous variables If not, the model is not identified Economics 20 - Prof. Schuetze 18 Addressing Errors-in-Variables with IV Estimation Recall the classical errors-in-variables problem where we observe x1 instead of x1* Where x1 = x1* + e1, we showed that when x1 and e1 are correlated the OLS estimates are biased We maintain the assumption that u is uncorrelated with x1*, x1 and x2 and that and e1 is uncorrelated with x1* and x2 If we can find an instrument, z, such that Corr(z,u) = 0 and Corr(z,x1) ≠ 0, then we can use IV to remove the attenuation bias Economics 20 - Prof. Schuetze 19 Example of Instrument Suppose that we have a second measure of x1*(z1) Examples: both husband and wife report earnings both employer and employee report earnings z1 will also measure x1* with error However, as long as the measurement error in z1 is uncorrelated with the measurement error in x1, z1 is a valid instrument Economics 20 - Prof. Schuetze 20 Testing for Endogeneity Since OLS is preferred to IV if we do not have an endogeneity problem, then we’d like to be able to test for endogeneity Suppose we have the following structural model: y1 = β0 + β1 y2 + β2 z1 + β3 z2 + u We suspect that y2 is endogenous and we have instruments for y2 (z3, z4) How do we determine if y2 is endogenous? Economics 20 - Prof. Schuetze 21 Testing for Endogeneity (cont) 1. Hausman Test If all variables are exogenous both OLS and 2SLS are consistent If there are statistically significant differences in the coefficients we conclude that y2 is endogenous 2. Regression Test In the first stage equation: y2 = π0 + π1 z1 + π2 z2 + π3 z3 + π3 z3 + v2 Each of the z’s are uncorrelated with u1 Economics 20 - Prof. Schuetze 22 Testing for Endogeneity (cont) Thus y2 will only be correlated with u1 if v2 is correlated with u1 The test is based on this observation So, save the residuals from the first stage Include the residual in the structural equation (which of course has y2 in it) If the coefficient on the residual is statistically different from zero, reject the null of exogeneity If multiple endogenous variables, jointly test the residuals from each first stage Economics 20 - Prof. Schuetze 23 Testing Overidentifying Restrictions How can we determine if we have a good instrument -correlated with y2 uncorrelated with u? Easy to test if z is correlated with y2 If there is just one instrument for our endogenous variable, we can’t test whether the instrument is uncorrelated with the error (u is unobserved) If we have multiple instruments, it is possible to test the overidentifying restrictions i.e. to see if some of the instruments are correlated with the error Economics 20 - Prof. Schuetze 24 The OverID Test Using our previous example, suppose we have two instruments for y2 (z3, z4) We could estimate our structural model using only z3 as an instrument, assuming it is uncorrelated with the error, and get the residuals: uˆ1 = y1 − βˆ 0 − βˆ1 y 2 − βˆ 2 z1 − βˆ 3 z 2 Since z4 hasn’t been used we can check whether it is correlated with u1-hat If they are correlated z4 isn’t a good instrument Economics 20 - Prof. Schuetze 25 The OverID Test We could do the same for z3, as long as we can assume that z4 is uncorrelated with u1 A procedure that allows us to do this is: 1. Estimate the structural model using IV and obtain the residuals 2. Regress the residuals on all the exogenous variables and obtain the R2 to form nR2 3. Under the null that all instruments are uncorrelated with the error, LM ~ χq2 where q is the number of “extra” instruments Economics 20 - Prof. Schuetze 26 Testing for Heteroskedasticity When using 2SLS, we need a slight adjustment to the Breusch-Pagan test Get the residuals from the IV estimation Regress these residuals squared on all of the exogenous variables in the model (including the instruments) Test for the joint significance Note: there are also robust standard errors in the IV setting Economics 20 - Prof. Schuetze 27 Testing for Serial Correlation Also need a slight adjustment to the test for serial correlation when using 2SLS Re-estimate the structural model by 2SLS, including the lagged residuals, and using the same instruments as originally Test if the coefficient on the lagged residual ( ρ) is statistically different than zero Can also correct for serial correlation by doing 2SLS on a quasi-differenced model, using quasi- differenced instruments Economics 20 - Prof. Schuetze 28