Lecture 17: Asymptotics (Chapter 12.2–12.5) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. Agenda • Review of Consistency and Probability Limits (Chapter 12.2) • Consistency of OLS (Chapter 12.3) • Replacing Fixed X ’s with Stochastic X ’s (Chapter 12.4) • Asymptotic Efficiency (Chapter 12.5) • Review of Chapter 12.1–12.5 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-2 Consistency (Chapter 12.2) • We would love to be guaranteed that our estimator will be exactly right, or at least very, very close to exactly right. • Certainty is too much to ask for in a stochastic world. • However, if our estimator is consistent, we can be “almost always almost right” in very, very large samples. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-3 Consistency (cont.) • Consistency: as the sample size approaches infinity, the estimator falls arbitrarily close to the true value with probability approaching 1. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-4 Consistency (cont.) • An estimator is consistent if, provided we get the sample size high enough, the probability can be as close to 1 as we like that our estimate is as close to being right as we like. • With finite sample sizes, there is always some probability that a consistent estimator is very wrong. But as the sample size grows, that probability becomes very, very small. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-5 Figure 12.3 The Collapse of a Consistent Estimator’s Distribution as n Grows Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-6 Consistency • What does it take to get consistency? • One often-encountered way is to have an asymptotically unbiased estimator whose variance shrinks to zero as the sample size grows. • In large samples, the estimator on average is right. • In large samples, no single estimate is likely to be very far from the estimator’s average. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-7 Probability Limits (Chapter 12.2) • How can we determine if an estimator is consistent or not? • We need a new mathematical tool, the probability limit (or plim). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-8 Probability Limits (cont.) • A random variable b converges in probability to a constant value c if, as the sample size grows very large, the probability approaches 1 that b takes on a value very close to c. • We call c the probability limit of b. • plim(b ) = c Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-9 Probability Limits (cont.) An estimator of b is consistent if p lim( bˆ ) b Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-10 Algebra of Probability Limits Let k be a constant. , b are random variables. 1. p lim(k ) k 2. p lim( b ) p lim( ) p lim( b ) 3. p lim(b ) p lim( ) p lim( b ) b p lim( b ) 4. p lim if p lim( ) 0 p lim( ) 5. p lim( g ( b )) g ( p lim( b )) for continuous function g ( ). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-11 Checking Understanding 2 1 2 Let n ~ N 3, . Let b n ~ N 2 , 2 n n n i. What is p lim n b n ? n ii. What is p lim ? bn iii. What is p lim b n 2 ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-12 Checking Understanding (cont.) 2 1 2 Let n ~ N 3, . Let b n ~ N 2 , 2 n n n i. What is p lim n b n ? E ( n ) 3. Because n's variance shrinks to 0 as n , p lim( n ) 3. 1 E ( b n ) 2 . As n , E ( b n ) approaches 2. n Because b n's variance shrinks to 0 as n , p lim( b n ) 2. p lim( n b n ) p lim( n ) p lim( b n ) 3 2 1. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-13 Checking Understanding (cont.) 2 1 2 Let n ~ N 3, . Let b n ~ N 2 , 2 n n n n ii. What is p lim ? bn n p lim( n ) 3 p lim 2 b n p lim( b n ) iii. What is p lim b n 2 ? X 2 is a continuous function, so p lim( b n 2 ) ( p lim( b n )) 2 2 2 4 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-14 Consistency of OLS • Is OLS consistent under Gauss–Markov? Yi b 0 b1 X i i E ( i ) 0 Cov( i , j ) 0 if i j Var ( i ) 2 X 's fixed across samples We know OLS is BLUE. xi yi ˆ b1 xi 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-15 Consistency of OLS (cont.) xiYi xi ( b 0 b1 X i i ) ˆ p lim( b1 ) p lim p lim 2 2 x x i i b 0 p lim xi b1 p lim xi2 p lim xi i p lim xi2 1 x n i i b1 p lim 1 2 xi n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-16 Consistency of OLS (cont.) 1 xi i p lim( bˆ1 ) b1 p lim n 1 2 xi n 1 p lim xi i n b1 1 2 p lim xi n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-17 Consistency of OLS (cont.) • We now need to add an assumption to the Gauss–Markov DGP, to guarantee that 1 2 p lim xi Q for some Q 0, Q n • We need to assume that, as n grows large, 1 2 xi Q, a finite, non-zero constant. n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-18 Consistency of OLS (cont.) 1 1 p lim xi i p lim xi i n n ˆ Then p lim( b1 ) b1 b1 Q 1 p lim xi 2 n To show that OLS is consistent, we need to show that 1 p lim xi i 0 n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-19 Consistency of OLS (cont.) We need to show that 1 p lim xi i 0 n We will use the strategy of showing that its expectation is 0 and that its variance declines to 0 as n increases. 1 1 E xi i E(xi i ) n n 1 xi E( i ) 0 n where we have used the assumption that X's are fixed across samples (so we can treat xi as a constant). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-20 Consistency of OLS (cont.) We need to show that 1 Var xi i 0 as n grows large n 1 1 1 Var xi i 2 Var(xi i ) 2 Var(xi i ) Cov 's n n n 2 1 1 2 2 2 xi Var( i ) xi n n n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-21 Consistency of OLS (cont.) We need to show that 1 Var xi i 0 as n grows large n 1 2 1 2 Var xi i xi n n n We have already assumed that 1 2 xi Q as n grows large. n We also require that 2 . 1 2Q As n grows large, Var xi i 0 n n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-22 Consistency of OLS (cont.) 1 1 Because E xi i 0 and Var xi i 0 n n 1 as n grows large, p lim xi i 0. n 0 ˆ Thus p lim( b ) b b Q bˆ is consistent. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-23 Consistency of OLS (cont.) • When determining the assumptions needed for a consistent estimator, notice that we need two new types of assumptions. • First, we typically need to assume that 2 is finite, in order to show that the variance of a term goes to 0 as n increases. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-24 Consistency of OLS (cont.) • Second, we typically need to assume something about what happens to the explanators as the sample size approaches infinity. 1 2 p lim xi Q, a non-zero constant. n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-25 Stochastic X ’s (Chapter 12.4) • It is crucial to pin down the assumptions needed to achieve a consistent estimator when the X ’s are not fixed, because the expectations needed to show unbiasedness will generally be intractable. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-26 Stochastic X ’s (cont.) xiYi xi i ˆ E ( b1 ) E b1 E 2 2 xi xi When the X 's are fixed, we can easily pull the weights to the other side of the expectations operator: xi i xi b1 E b1 2 E ( i ) 2 xi xi When the X 's are stochastic, this step becomes xi i impossible, and we cannot easily eliminate E . 2 xi Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-27 Stochastic X ’s (cont.) When working with consistency, we can take advantage of the properties of probability limits: xi i xiYi ˆ b1 p lim p lim( b1 ) p lim 2 2 xi xi Because the p lim of the ratio is the ratio of the p lim , xi i p lim xi i b1 b1 p lim 2 2 x lim p x i i Then, if we multiply both the numerator and denominator 1 by , we can work with each component separately. n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-28 Stochastic X ’s (cont.) • What key properties do we rely on for OLS to be consistent? 1 p lim xi i 0 n 1 2 p lim xi Q, a non-zero constant n • We have simply added the second property to our list of assumptions. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-29 Stochastic X ’s (cont.) 1 p lim xi i 0 n • For the Gauss–Markov DGP, we established this property using: – The assumption that the X ’s are fixed, to establish the expectation was zero, and – Our usual assumptions about the variance and covariance of and our new assumptions that 2 is finite and xi2 approaches some finite non-zero limit, to show that the variance approaches 0 as n increases Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-30 Stochastic X ’s (cont.) 1 p lim xi i 0 n • When we relax the assumption that X ’s are fixed across samples, the key to achieving consistency will be finding some alternative assumption that makes this plim 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-31 Stochastic X ’s (cont.) • Econometricians generally prefer to base a DGP on statements about expectations. E(xi i ) 0 plus some assumption such that 1 Var xi i n approaches 0 as n increases. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-32 Stochastic X ’s (cont.) • Under the Gauss–Markov conditions, E(xii ) = 0 follows trivially from fixing the X ’s across samples. • When we relax that assumption, we will have a much harder time achieving this property. • The next lecture examines several routine econometric difficulties that violate this key property. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-33 Stochastic X ’s (cont.) • So far, we have shown that OLS is a consistent estimator when we have a single explanator. • What if we have multiple explanators? • The same conditions apply, but for each explanator. • Also, the X ’s must not be multicollinear in the probability limit. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-34 Stochastic X ’s (cont.) • Conditions for OLS to be consistent with multiple explanators: 1 p lim x Ri i 0 for all R n 1 2 p lim x Ri Q, a non-zero constant, for all R n X's are not multicollinear as n grows large Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-35 Asymptotic Efficiency • Earlier in the course, when we focused on cross-sample properties, we looked for the BLUE Estimator: the unbiased linear estimator that was “best” (i.e. with the smallest cross-sample variance). • Now we’re looking for consistent estimators; is there a comparable concept of “best”? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-36 Asymptotic Efficiency (cont.) • Efficiency: for a fixed sample size, the estimator has the smallest crosssample variance. • Asymptotic Efficiency: the consistent estimator has the smallest cross-sample variance for all very large sample sizes (for all n > N, where N is some finite number). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-37 Asymptotic Efficiency (cont.) • Problem: nearly all consistent estimators have variances that collapse to 0 in the limit. • We cannot define asymptotic efficiency as having the smallest variance as n approaches infinity, because by the time n is approaching infinity all the variances have gotten vanishingly close to 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-38 Asymptotic Efficiency (cont.) • Instead of comparing variances as consistent estimators approach infinity, we compare the speed at which the variances collapse towards 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-39 Asymptotic Efficiency (cont.) Suppose we have two estimators, ˆ and bˆ. E (ˆ ) E ( bˆ ) Var (ˆ ) 2 and Var ( bˆ ) 2 2 . n n As n , both estimators collapse around . However, the distribution of bˆ collapses much more quickly. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-40 Asymptotic Efficiency (cont.) Var (ˆ ) 2 and Var ( bˆ ) 2 2 . n n The variance of bˆ shrinks to 0 more quickly than ˆ . At n 10, Var (ˆ ) 2 10 At n 100, Var (ˆ ) while Var ( bˆ ) 2 100 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 2 100 while Var ( bˆ ) . 2 10, 000 . 17-41 Asymptotic Efficiency (cont.) Var (ˆ ) 2 and Var ( bˆ ) 2 2 . n n As n , both collapse around their expectation, . As n gets very large, we cannot easily distinguish between ˆ and bˆ. The differences between them eventually get too small for us to conceptualize. We need a "magnifying glass" to tell them apart. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-42 Asymptotic Efficiency (cont.) Var (ˆ ) 2 and Var ( bˆ ) 2 2 . n n As n , both collapse around their . We need a "magnifying glass" to tell them apart. What if we compared instead n ˆ and n bˆ ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-43 Asymptotic Efficiency (cont.) ˆ Var (ˆ ) and Var ( b ) 2 . n n What if we compared instead n ˆ and n bˆ ? Problem: as n , E (n ˆ ) n E (ˆ ) n . The expectation of n ˆ and n bˆ is growing 2 2 without bound as n approaches infinity! Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-44 Asymptotic Efficiency (cont.) What if we compared instead n ˆ and n bˆ ? Problem: as n , E (n ˆ ) n E (ˆ ) n . Solution: compare the ERRORS of the estimators, which remain centered around 0. E[n (ˆ - )] E[n ( bˆ - )] n - n 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-45 Asymptotic Efficiency (cont.) Compare the ERRORS of the estimators, which remain centered around 0. Var (n (ˆ )) n 2 Var (ˆ ) n Var (ˆ ) n 2 2 2 n 2 n As n , the distribution of n (ˆ ) explodes. The variance increases without bound. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-46 Asymptotic Efficiency (cont.) Var (n ( bˆ )) n 2 Var ( bˆ ) 2 2 ˆ n Var ( b ) n 2 n As n , the variance of n ( bˆ ) 2 2 converges to . 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-47 Asymptotic Efficiency (cont.) As n , the variance of n ( bˆ ) converges to 2 . While the variance of bˆ is shrinking as n increases, we have magnified bˆ by just enough to "keep pace." n ( bˆ ) has a stable asymptotic distribution with mean 0 and variance 2 . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-48 Asymptotic Efficiency (cont.) Var (n (ˆ - )) n 2 and Var (n ( bˆ - )) 2 . As n , the distribution of n (ˆ - ) explodes while the distribution of n ( bˆ - ) converges. A factor of n is "fast enough" to "keep pace" with bˆ - , but it's too fast to keep pace with ˆ - . We need a smaller magnifying glass to look at ˆ - . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-49 Asymptotic Efficiency (cont.) Magnifying ˆ - by n is too much to reach a stable asymptotic distribution. What about n? E[ n (ˆ - )] E[ n ( bˆ - )] n - n 0 Var ( n (ˆ )) n Var (ˆ ) n Var (ˆ ) n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 2 n 2 17-50 Asymptotic Efficiency (cont.) As n , the variance of n (ˆ ) converges to . The asymptotic 2 distribution of n (ˆ ) has expectation 0 and variance . 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-51 Asymptotic Efficiency (cont.) Var ( n ( bˆ )) n Var ( bˆ ) ˆ n Var ( b ) n 2 n n As n , the distribution of n ( bˆ ) collapses around 0. 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 2 17-52 Asymptotic Efficiency (cont.) • In general, if we “blow up” a variable whose distribution is collapsing around 0 by some power of n, we can stabilize the process at some asymptotic distribution. • We have to pick just the right power of n. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-53 Asymptotic Efficiency (cont.) We call the power of n that stabilizes the distribution the "convergence rate." The convergence rate of (ˆ ) is root n. The convergence rate of ( bˆ ) is n. bˆ is clearly more efficient than ˆ. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-54 Asymptotic Efficiency (cont.) Example: let's consider the variance of OLS under the Gauss–Markov assumptions when Yi b 0 b1 X i i Var ( bˆ1 ) 2 xi 2 2 2 p lim(Var ) p lim 2 2 x p lim( x i ) i 1 2 2 We can re-write xi as n xi . n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-55 Asymptotic Efficiency (cont.) 1 2 We can re-write xi as n xi . n (Multiplying and dividing by n is a standard 2 trick when working with probability limits.) 1 2 xi is the variance of X. If, as we keep n increasing the sample size, the variance of X converges to some constant (a reasonable thought, 1 in many applications), then p lim xi Q. n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-56 Asymptotic Efficiency (cont.) ˆ Var ( b1 ) 2 xi 2 2 2 p lim(Var ) p lim 2 2 xi p lim(xi ) 2 1 2 p lim n xi n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 2 p lim(n) Q 0 17-57 Asymptotic Efficiency (cont.) Now, instead of looking at Var ( bˆ1 ), let's look at n Var ( bˆ ). 1 2 2 n ˆ n p lim(Var ( b 1 )) n p lim p lim(n) Q n Q If we blow up the variance by n, then the variance stabilizes. We can move the n inside the variance: p lim(Var ( n b1 ) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 2 Q . 17-58 Asymptotic Efficiency (cont.) ˆ p lim(Var ( n b1 )) Q However, E ( n bˆ ) grows 2 1 without bound. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-59 Asymptotic Efficiency (cont.) Look at the distribution of the ERRORS. p lim( E ( n ( bˆ1 b1 )) 0. p lim(Var ( n ( bˆ1 b1 ))) 2 Q . The convergence rate of bˆ1-b1 is root n. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-60 Asymptotic Efficiency (cont.) • The convergence rate of the OLS errors is root n. • We say that OLS is root n consistent. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-61 Asymptotic Efficiency (cont.) • The variance of an estimator’s magnified errors is called its asymptotic variance. • The asymptotic variance of OLS is 2 Q Q lim n xi 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-62 Asymptotic Efficiency (cont.) • Among estimators with the same convergence rate, the consistent estimator with the smallest asymptotic variance is called asymptotically efficient among estimators with that convergence rate. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-63 Checking Understanding ˆ1 ~ N ( , 2 n x 3 2 i ) 2 ˆ2 ~ N ( , 1 2 2 n 1 (xi ) n where xi2 Q, with 1 Q 3 What is the convergence rate of (ˆ1 )? What is the converge rate of (ˆ )? 2 Which estimator is more efficient? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-64 Checking Understanding (cont.) ˆ Var (1 ) 3 2 n xi 2 ˆ n Var (1 ) 2 xi 2 3 3 2 Var (n (ˆ1 )) 2 x 2 i 2 Q as n 3 2 The convergence rate of (ˆ1 ) is n . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-65 Checking Understanding (cont.) Var (ˆ2 ) 2 1 n 1 (xi2 ) 2 n 3 n3Var (ˆ2 ) 3 2 2 1 2 2 1 ( x i) n Var (n (ˆ2 )) 2 Q 2 as n grows large. 3 2 The convergence rate of (ˆ2 ) is n . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-66 Checking Understanding (cont.) 3 2 Var (n (ˆ1 )) 3 2 Var (n (ˆ2 )) 2 Q 2 Q2 If one estimator converged more quickly than the other, then it would be more efficient. Both estimators converge at the same rate. Because Q 1, ˆ has a smaller asymptotic 2 variance than ˆ1 . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-67 Asymptotic Efficiency • Efficiency is a subtler concept when working at the limit as n approaches infinity. • The distribution of a consistent estimator tends to collapse; the variance shrinks to 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-68 Asymptotic Efficiency (cont.) • We cannot usefully compare the distributions of different estimators or their errors as n approaches infinity. • Every consistent estimator’s distribution has the same asymptotic “destination.” • We CAN compare how quickly they approach that destination. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-69 Asymptotic Efficiency (cont.) • Because the distributions of consistent estimators tend to get very close together as n grows, we need to magnify the differences so we can see them better. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-70 Asymptotic Efficiency (cont.) • We examine the distributions of the errors magnified by their convergence rate, as n approaches infinity. • We can work with this asymptotic distribution. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-71 Asymptotic Efficiency (cont.) • As n approaches infinity, the errors (as blown up by the appropriate magnifier) converge to some distribution. • If the Central Limit Theorem holds, the asymptotic distribution of the errors is Normal. • The most often violated requirement of the CLT is that variances are bounded. • See Appendix 12.4. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-72 Figure 12.4 The Distributions of Errors for bg1, bg2, bg3, and bg4 for Various n and Normal and Skewed, Discrete Disturbances (Panel A) 17-73 Figure 12.4 The Distributions of Errors for bg1, bg2, bg3, and bg4 for Various n and Normal and Skewed, Discrete Disturbances (Panel B) 17-74 Figure 12.4 The Distributions of Errors for bg1, bg2, bg3, and bg4 for Various n and Normal and Skewed, Discrete Disturbances Copyright © 2006 Pearson Addison-Wesley. All rights reserved. Asymptotic Efficiency • In moderate sized samples (say, n > 100), we can use the asymptotic distribution for inference when the finite-sample distribution is unknown. • When OLS is consistent and asymptotically normally distributed, we can use t and F tests. • When OLS is inconsistent, tests are biased. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-76 Review • We have a new strategy for estimation. • When our sample size is reasonably large, we can look not for unbiased estimators but for consistent estimators. • When we relax the assumption that X ’s are fixed across samples, it will be much easier to work with probability limits (for consistency) than with expectations (for unbiasedness). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-77 Review (cont.) • As the sample size grows very large, consistent estimators have a very high probability of giving estimates very close to the correct value. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-78 Review (cont.) • One method for demonstrating that an estimator is consistent: – Demonstrate that the estimator is asymptotically unbiased – Demonstrate that the estimator’s variance converges to 0 as n approaches infinity Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-79 Review (cont.) • To show that an estimator is consistent, we require assumptions that bound key variances. • We also require that error terms and explanators be contemporaneously uncorrelated, i.e. E(xii ) = 0. • This assumption is weaker than requiring the X ’s to be fixed across samples. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-80 Review (cont.) • To evaluate the efficiency of an estimator and draw inferences (i.e. run hypothesis tests), we need to be able to describe an estimator’s distribution as n approaches infinity. • However, the distribution collapses as n grows. • The answer: magnify the errors by just the right amount (the convergence rate) such that the distribution converges. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-81 Review (cont.) • When the Central Limit Theorem applies (as occurs frequently), the asymptotic distribution is Normal. • With asymptotic normality, we can use the same hypothesis tests we developed for unbiased estimators to draw inferences from consistent estimators. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-82 Review (cont.) • We will be working with probability limits as we develop consistent estimators. • You need to be familiar with the algebra of plims. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-83 Algebra of Probability Limits Let k be a constant. , b are random variables. 1. p lim(k ) k 2. p lim( b ) p lim( ) p lim( b ) 3. p lim(b ) p lim( ) p lim( b ) b p lim( b ) 4. p lim if p lim( ) 0 p lim( ) 5. p lim( g ( b )) g ( p lim( b )) for continuous function g ( ). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 17-84