300 likes | 482 Views
Lecture 1. Topics in Applied Econometrics B. Instrumental Variables & 2SLS. y 1 = β 0 + β 1 y 2 + β 2 x 1 + . . . β k x k + u y 2 = π 0 + π 1 z + π 2 x 1 + . . . π k x k + v. Why Use Instrumental Variables?. OLS is inconsistent with omitted variable bias Solutions:
E N D
Lecture 1 Topics in Applied Econometrics B
Instrumental Variables & 2SLS y1 = β0 + β1y2 + β2x1 + . . . βkxk + u y2 = π0 + π1z + π2x1 + . . . πkxk + v
Why Use Instrumental Variables? • OLS is inconsistent with omitted variable bias • Solutions: • Ignore the problem and suffer the consequences of biased and inconsistent estimators. Example: you want to show that a coefficient is positive but you know it is biased downward. As long as the estimated coefficient is positive then you know the unbiased coefficient is also positive and larger. • Use a proxy variable for the omitted variable (recall our example of using IQ test results to proxy for ability) • Use panel data techniques and assume that the omitted variable does not change over time • Use Instrumental Variables (IV)
IV Estimation • IV estimation is used when your model has endogenous x’s • Endogenous: A variable is endogenous when it is correlated with the error term (Cov(x,u) ≠ 0) due to an omitted variable, measurement error or simultaneity • IV can be used to address the problem of omitted variable bias if you don’t have a good proxy
What is an Instrumental Variable? • y = b0 + b1x + u • We think x and u are correlated. If we run the regression as is, b1 is biased. • In order to run this regression and obtain an unbiased estimate of b1 , we need some more information. We get this information from a new variable called z. • In order for a variable, z, to serve as a valid instrument for x, the following must be true • The instrument must be exogenous, that is, Cov(z,u) = 0 • The instrument must be correlated with the endogenous variable x, that is, Cov(z,x) ≠ 0
More on Valid Instruments • There is no way to test whether Cov(z,u) = 0 because u is unobserved. So, we have to use common sense and economic theory to decide if it makes sense to assume it. • We can test if Cov(z,x) ≠ 0 • test H0: p1 = 0 in x = p0 + p1z + v • Cov(z,x) ≠ 0 p1 ≠ 0 • Recall our log(wage) on education example with the omitted ability. • Would teudat zehut number make a good IV? • What about IQ? • Sometimes refer to this regression as the first-stage regression
Example • We want to estimate the effect of missing class on final exam score • exam score = b0 + b1skip_class + u • Can we get a good estimate of the causal effect of missing class on exam score from this equation? • What would be considered a good IV?
IV Estimation in the Simple Regression Case • For y = b0 + b1x + u, and given our assumptions • Cov(z,y) = b1Cov(z,x) + Cov(z,u), so • b1 = Cov(z,y) / Cov(z,x) • Then the IV estimator for b1 is • What happens when z = x?
Inference with IV Estimation • As in the OLS case, given the asymptotic variance (the IV estimator has an approximate normal distribution in large sample sizes), we can estimate the standard error • As usual, we start with the assumption of homoskedasticity. In this case, the homoskedasticity assumption is E(u2|z) = s 2 = Var(u) • The assumptotic variance of is: where σ2x is the population variance of x, σ2 is the population variance of u, and ρ2z,x is the square of the population correlation between x and z. This tells us how highly correlated x and z are in the population. • Each of these can be consistently estimated with a random sample
IV versus OLS estimation • Standard error in IV case differs from OLS only in the R2 (ρ) from regressing x on z • Because R2 < 1, IV standard errors are larger, sometimes much larger • The stronger the correlation between z and x, the smaller the IV standard errors • However, IV is consistent, while OLS is inconsistent when Cov(x,u) ≠ 0 • Stata Example 1B-1
The mother of all IV examples • Angrist and Krueger (1991) • Use quarter of birth as an instrument for education. What is the reasoning? • Quarter of birth (Jan-March, April-June, etc) is uncorrelated with ability • Quarter of birth is correlated with education. How? In the US, there are compulsory minimum ages for beginning school and dropping out. • Say the age to begin school is the year the child turns 6. A child with a birthdate of January 1 will be close to 7 years old by the fall and a child born in December of the same year will not even be six years old by the time the school year begins. • If the age to legally drop out of school is 16 years old, you can see that there can be variation in the total amount of schooling people have.
Example cont. • Granted, the correlation between quarter of birth and education is small. So, they compensate by having a very large data set (250,000 men born between 1920 and 1929) • They find that the OLS return to education is 0.0801 (se .0004) and the IV estimate is 0.0715 (se .0219) • Interesting that they don’t differ too much, in fact, the OLS estimate is contained inside the IV estimate
Example cont • Bound, Jaeger, and Baker (1995) • It is not obvious that quarter of birth is unrelated to other factors that affect wage • Turns out that depression, schizophrenia, mental retardation, multiple personalities, etc.. are more likely to occur to people that are born in a particular time in the year. In addition, high income people are less likely to have children in the winter. • Their paper shows that in the case of “weak instruments” when you introduce a small amount of correlation between the IV and the error term—it causes big problems (bias, inconsistency in the IV estimates) • Called the “weak instrument problem”
The Effect of Poor Instruments • What if our assumption that Cov(z,u) = 0 is false? • The IV estimator will be inconsistent too • Can compare asymptotic bias in OLS and IV • Prefer IV if Cor(z,u)/Cor(z,x) < Cor(x,u)
The Effect of Poor Instruments • Suppose x and z are both positively correlated with u and cor(x,z) >0. • Then the asymptotic bias in the IV estimator is less than that for OLS only if cor(z,u)/Cor(z,x) < cor(x,u). • If cor(z,x) is small, then a seemingly small correlation between z and u can be magnified and make IV worse than OLS. • For example: if cor(z,x) = .2, cor(z,u) must be less than .2*cor(x,u) before IV has less asymptotic bias than OLS
IV Estimation in the Multiple Regression Case • IV estimation can be extended to the multiple regression case • Call the model we are interested in estimating “the structural model” • Our problem is that one or more of the variables are endogenous • We need an instrument for each endogenous variable
Multiple Regression IV (cont) • Write the structural model as: • y1 = b0 + b1y2 + b2x1 + u1 • where y2 is endogenous and x1 is exogenous • Let z2 be the instrument, so • Cov(z2,u1) = 0 and • y2 = p0 + p1x1 + p2z2 + v2, where p2≠ 0 • This reduced form equation regresses the endogenous variable on *all* exogenous ones
Two Stage Least Squares (2SLS) • It’s possible to have multiple instruments • Consider our original structural model, and let y2 = p0 + p1x1 + p2z2 + p3z3 + v2 • Here we’re assuming that both z2 and z3 are valid instruments – they do not appear in the structural model and are uncorrelated with the structural error term, u1 (exclusion restrictions)
Best Instrument • Could use either z2 or z3 as an instrument • The best instrument is a linear combination of all of the exogenous variables, y2* = p0 + p1x1 + p2z2 + p3z3
More on 2SLS • While the coefficients are the same, the standard errors from doing 2SLS by hand are incorrect, so let Stata do it for you • Method extends to multiple endogenous variables – need to be sure that we have at least as many excluded exogenous variables (instruments) as there are endogenous variables in the structural equation (necessary condition for identification) • For example if you have two endogenous variables you must have at least two instruments (so 3 is ok but 1 is not). • We call these exclusion restrictions
Multicollinearity and 2SLS • Correlation among regressors can lead to large standard errors for the OLS estimators • This problem is more serious with 2SLS • Recall our formula for the asymptotic variance of 2SLS can be approximated as • Where σ2 = Var(u), SŜT2 is the total variation in ŷ2 and R2 is the R-squared from a regression of ŷ2 on all other exogenous variables appearing in the structural equation
Multicollinearity and 2SLS • Two reasons why variance for the 2SLS estimator is larger than that for OLS • By construction, ŷ2 has less variation than y2 (recall, TSS= ESS + RSS; variation in ŷ2 = ESS and variation in y2 = TSS) • Correlation between ŷ2 and the exogenous variables is often much higher than the correlation between y2 and these variables • Stata Example 1B-1 cont.
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 • If we do not have endogeneity, both OLS and IV are consistent • While it’s a good idea to see if IV and OLS have different implications, it’s easier to use a regression test for endogeneity • If y2 is endogenous, then v2 (from the reduced form equation) and u1 from the structural model will be correlated
Testing for Endogeneity • Suppose we have a single suspected endogenous variable and z1 and z2 are exogenous. Suppose further that we have two additional variables (z3 and z4) to be used as instruments • Since each zj is uncorrelated with u1, y2 is uncorrelated with u1 if and only if v2 is uncorrelated with u1: this is what we wish to test
Testing for Endogeneity • Write u1 = δ1v2 + e1, where e1 is uncorrelated with v2 and has zero mean. • u1 and v2 are uncorrelated if and only if δ1=0. • Of course, v2 is not observed so we can estimate the reduced form for y2 by OLS and obtain the reduced form residuals, . • Then we include these residuals in the structural model and test H0: δ1=0 using a t-statistic
Testing for Endogeneity • One interesting thing to take from this regression is that the estimated coefficients from this OLS regression should be identical to those estimated using 2SLS. • This provides a useful way to check how our coefficients change between OLS and 2SLS when we add the error term from the first stage reduced form.
Testing for Endogeneity: Durbin-Wu-Hausman Test • 1. Run the reduced form regression of the endogenous variable on the instrument and all other exogenous variables • 2. Predict the residuals from this regression • 3. Run the structural model with the endogenous variable and the residuals from the first stage • 4. If the t-stat on the estimated coefficient on the residual variable is significant then we have an endogeneity problem. • If multiple endogenous variables, jointly test the residuals from each first stage • Stata Example 1B-1 cont.
Testing Overidentifying Restrictions • If there is just one instrument for our endogenous variable, we can’t test whether the instrument is uncorrelated with the error • We say the model is just identified • If we have multiple instruments, it is possible to test the overidentifying restrictions – to see if some of the instruments are correlated with the error
Testing Overidentifying Restrictions • The idea: Say we have two instruments. We can then estimate our equation using the first instrument only and then estimate the equation using the second instrument only. • If both instruments are exogenous and both are partially correlated with the endogenous variable then we should obtain similar estimated coefficients on the endogenous variable (that differ only by sampling variation) • If these two estimated coefficients are statistically different from one another then we have no choice but to conclude that either one or both of them fail the exogeneity test • However, even if they pass the test, there is still a more subtle problem. The two IV estimates may be similar though both are inconsistent. • The point is that you should feel too comfortable if your two IV procedures pass the test.
The OverID Test • 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 • Under the null that all instruments are uncorrelated with the error, LM ~ cq2 where q is the number of extra instruments • Stata Example 1B-1 cont.