Lecture 27: Cointegration (Chapter 18.6–18.7) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. Agenda • Review • Common Trends (Chapter 18.6) • Cointegration (Chapter 18.6) • The Drunk and Her Dog (Chapter 18.6) • Dynamic OLS (Chapter 18.7) • Example: Deficits (Chapter 18.7) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-2 TABLE 18.1 The Consistency of OLS and the Validity of Conventional Tests in the Face of Stochastic Trends* Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-3 Review • When both e and Z contain stochastic trends, our regression is spurious. • OLS often misleads us into believing Y and Z are related, even when they are not. • We need a test for stochastic trends, and an estimation procedure for variables exhibiting stochastic trends. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-4 Review of Unit Root Tests • The desired test: Zt Zt -1 0 1t vt H 0 : 1 vs. H a : 1 • The problem: – Under the null hypothesis, OLS is not asymptotically normal. • The solution: – Develop a new test statistic whose distribution we know. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-5 Review of Unit Root Tests (cont.) • Dickey and Fuller developed a suitable test. They needed to modify the regression slightly: Zt Zt 1 ( 1)Zt 1 0 1t vt • To test = 1, we test that the coefficient on Zt-1 is 0 against the 1-sided alternative that it is less than 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-6 Review of Unit Root Tests (cont.) • We cannot use our usual t-statistic critical values because the Dickey– Fuller asymptotic distribution is skewed. • Dickey and Fuller determined the appropriate critical values. • The critical values depend on whether the deterministic drift term is included. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-7 Review of Unit Root Tests (cont.) Zt Zt 1 ( 1)Zt 1 0 1t vt • The Dickey–Fuller test assumes that the vt are serially uncorrelated. • If we are concerned that the vt are serially correlated, we need to use the Augmented Dickey–Fuller (ADF) test. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-8 Review of Unit Root Tests (cont.) • The Augmented Dickey–Fuller test adds lagged DZ values to the explanators. DZt Zt Zt 1 ( 1) Zt 1 0 1t 2 DZt 1 3DZt 2 vt Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-9 Review of Unit Root Tests (cont.) • Caution: failing to reject the null of a stochastic trend does NOT establish the presence of a stochastic trend. • The Dickey–Fuller test often has weak power. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-10 Review • The Granger and Newbold strategy: • Both Y and Z contain stochastic trends (for example, let them both follow random walks). • Focus on CHANGES in Y and Z, differencing out the stochastic trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-11 Review (cont.) t Yt Yt 1 vt vs (Y does not depend on Z ) s 0 DYt Yt Yt 1 vt t Z t Z t 1 t vs s 0 DZ t Z t Z t 1 t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-12 Review (cont.) t Yt Yt 1 vt vs s 0 t Z t Z t 1 t vs s 0 If we regress DYt 1DZ t t then t vt Neither the explanator nor the error term contain stochastic trends. The regression satisfies the Gauss–Markov conditions. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-13 Review (cont.) • If Y and Z BOTH receive a large positive shock, then DY and DZ will both be above their means. • However, in the next period, the shock will no longer be evident in DY and DZ. • DY and DZ revert to the mean. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-14 Review (cont.) • Danger: if Y or Z does NOT contain a stochastic trend, we do NOT want to difference. • DY has less variation than Y, hurting efficiency. • Differencing can induce serial correlation. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-15 Review (cont.) • The Dickey–Fuller test warns of the possibility of a stochastic trend. • The Newbold–Granger method is unattractive in the absence of a stochastic trend. • Tests from differenced regressions tend to have lower power. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-16 Review (cont.) • Variables with no stochastic trends are Integrated of Order Zero, written I(0). • The variables we have studied previously in this course are I(0). • The “white noise” innovations in a random walk are I(0). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-17 Review (cont.) • Variables with stochastic trends, but whose first differences contain no stochastic trends, are Integrated of Order One, written I(1). • Variables that follow a random walk are I(1). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-18 Review (cont.) • When both e and Z contain stochastic trends, our regression is spurious. • OLS often misleads us into believing Y and Z are related, even when they are not. • When using a stochastically trending explanator, we can use a Dickey–Fuller test on the residuals. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-19 Common Trends (Chapter 18.6) • Suppose Yt 0 1 X t ut • The Dickey–Fuller test can fail to reject the null hypothesis that Y contains a stochastic trend. • A stochastic trend in Y can come from a stochastic trend in X, a stochastic trend in u, or both. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-20 Common Trends (cont.) • Suppose Yt 0 1 X t ut • We have seen that when both X and u contain stochastic trends, our regression is spurious: we are far too likely to reject the null that 1 = 0, even if it IS zero. • But what if 1 really is non-zero? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-21 Common Trends (cont.) • Suppose Yt 0 1 X t ut • If 1 really is non-zero, and X contains a stochastic trend, then Y inherits that stochastic trend from X. • Moreover, Y and X share that same trend in common. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-22 Figure 18.9 Real per Capita GDP and Real per Capita Consumption, 1948–1998 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-23 Common Trends • Suppose Yt 0 1 Xt ut • If 1 really is non-zero, and X contains a stochastic trend, then Y inherits that stochastic trend from X. • If u ALSO contains a stochastic trend, then Y might contain several stochastic trends. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-24 Cointegration (Chapter 18.6) • Variables that contain stochastic trends exhibit increasing variance over time. • Such variables can get very far away from their ex ante expectations. • However, they may stay relatively close together. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-25 Cointegration (cont.) Yt 0 1 X t ut • If Y and X both contain stochastic trends, and u does not, then Y inherits its only stochastic trending component from X. • We call Y and X cointegrated. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-26 Cointegration (cont.) • Cointegrated variables both wander stochastically, but stay near each other. • They adjust to each other’s locations through a process of error correction. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-27 Cointegration (cont.) • Note: cointegration also extends to the multiple regression case. • The key is that the disturbance term does not contain a stochastic trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-28 Cointegration (cont.) • When there is a linear combination of trending Y ’s and X ’s that itself contains no variable trend, we call the variables “cointegrated.” • i.e., if Yt 0 1 X1t 2 X 2t ut and Y, X1, and X2 contain variable trends, but u does not, we say Y, X1, and X2 are cointegrated. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-29 Cointegration (cont.) • Consider: Yt 0 1 X t ut where both Y and X contain variable trends, but u does not, so Y and X are cointegrated. • Question: can we estimate 1 by analyzing changes in Y and X (the Newbold–Granger method)? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-30 The Drunk and Her Dog • As an example, consider the case of a drunk and her dog. • Both the drunk and the dog, on their own, might tend to wander aimlessly. • However, the dog might make sure to stay close to its drunken master. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-31 The Drunk and Her Dog (cont.) • The instructive tale of the drunk and her dog… – To estimate 1 in: Yt 0 1Zt ut should we study (Yt - Yt -1 ) DYt 1DZt ut - ut -1? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-32 The Drunk and Her Dog (cont.) (Yt - Yt -1 ) DYt 1DZt ut - ut -1 • If two stochastically-trending variables are truly related, a specification solely in changes is mis-specified—it is missing the cointegrating relationship(s) among the variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-33 The Drunk and Her Dog (cont.) • The Lesson: – If we are going to look at changes in Y and X when Y and X are cointegrated, we must also examine the “cointegrating vector” that will be the error correction mechanism holding together Y and X as they trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-34 The Drunk and Her Dog (cont.) • In the example, the drunk and her dog both wander aimlessly. • However, they periodically engage in error correction to move closer together. • Neither one is following a random walk. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-35 The Drunk and Her Dog (cont.) • Consider four possible simulated paths for the drunk and her dog. • The vertical axis measures how far they’ve each wandered from the local pub. • The horizontal axes tracks their steps, from 1 to 1,0000. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-36 Figure 18.11 Four Paths for the Drunk and Her Dog (1 of 2) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-37 Figure 18.11 Four Paths for the Drunk and Her Dog (2 of 2) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-38 Figure 18.11 Four Paths for the Drunk and Her Dog 27-39 The Drunk and Her Dog (cont.) • The order of cointegration depends on the trend in the distance between Y and Z. • If the distance between Y and Z contains no stochastic trend, then Y and Z are cointegrated of order 0. • For Y and Z to be cointegrated of order r, then both Y and Z are integrated I(r+1). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-40 The Drunk and Her Dog (cont.) • We can model the drunk’s and the dog’s cointegrated motion as: Yt Yt 1 e t a1 (Yt 1 Z t 1 ) Z t Z t 1 wt a 2 (Yt 1 Z t 1 ) • a1 and a2 are the speeds of adjustment. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-41 The Drunk and Her Dog (cont.) Yt Yt 1 e t a1 (Yt 1 Z t 1 ) Z t Z t 1 wt a 2 (Yt 1 Z t 1 ) • The second terms on the RHS of each equation are error correction mechanisms. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-42 The Drunk and Her Dog (cont.) • In general, cointegration does NOT require that the Yt – Zt contain no stochastic trend. • We need only some linear combination of the variables to contain no stochastic trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-43 A More General Error Correction Model Long-Run Relationship Yt 1 X t t Short-Run Error Correction DYt a 0 y (Yt 1 1 X t 1 ) t DX t 0 x (Yt 1 1 X t 1 ) vt Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-44 A More General Error Correction Model (cont.) DYt a 0 y (Yt 1 1 X t 1 ) t DX t 0 x (Yt 1 1 X t 1 ) vt Yt 1 X t t • Y and X adjust towards their long-run relationship at speeds y and x. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-45 A More General Error Correction Model (cont.) DYt a 0 y (Yt 1 1 X t 1 ) t DX t 0 x (Yt 1 1 X t 1 ) vt Yt 1 X t t • The (Yt-1 – 1Xt-1) terms are error correction mechanisms to keep Y and X close to their long-run relationship. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-46 A More General Error Correction Model (cont.) In the presence of serial correlation, we need to include lagged differences in the variables. DYt a 0 a1DYt 1 a 2 DX t 1 y (Yt 1 1 X t 1 ) t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-47 Dynamic OLS (Chapter 18.7) • When et contains a stochastic trend, OLS is inconsistent. Use the Granger and Newbold method (differencing Y and X ). • When the et ’s do not contain stochastic trends, but an explanator does, OLS is super consistent, but not asymptotically normal. Our tests do not work (for the trending explanator). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-48 Dynamic OLS (cont.) Yt 0 1Z1t 2 Z2t 3 X t e t • Suppose Y, Z1, and Z2 contain stochastic trends, but X and e do not. • We can estimate all the coefficients consistently. • Only our estimate of 3 is asymptotically normal. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-49 Dynamic OLS (cont.) Yt 0 1Z1t 2 Z2t 3 X t e t • Our hypothesis tests are not asymptotically valid for the variables that contain a stochastic trend (Z1 and Z2). • Stock and Watson’s insight: change the equation so that the 1 and 2 coefficients apply to non-trending variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-50 Dynamic OLS (cont.) • Dynamic OLS: – A method for estimating cointegrated models – Respecify the model so that the coefficients of interest apply to non-trending variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-51 Dynamic OLS (cont.) Case: no serial correlation We wish to estimate Yt 0 1Z1t 2 Z 2t 3 X t e t Yt , Z1t , Z 2t contain stochastic trends. Rewrite the model as Yt 0 1Z1t 2 Z 2t 3 X t 4 DZ1t 5 DZ 2t t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-52 Dynamic OLS (cont.) Case: no serial correlation Rewrite the model as Yt 0 1Z1t 2 Z 2t 3 X t 4 DZ1t 5 DZ 2t t After controlling for DZ1t , there is no stochastic trend in Z1t . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-53 Dynamic OLS (cont.) • Serial correlation complicates Dynamic OLS. • We must add not only DZ but also leads and lags of DZ. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-54 Dynamic OLS (cont.) Case: serial correlation We wish to estimate Yt 0 1Z1t 3 X t e t Yt and Z1t contain stochastic trends. Rewrite the model as Yt 0 1Z1t 2 X t 1DZ1,t 2 2 DZ1,t 1 3 DZ1,t 4 DZ1,t 1 5 DZ1,t 2 t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-55 Dynamic OLS (cont.) • Dynamic OLS allows us to draw inferences about relationships among stochastically trending variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-56 Example: Deficits (Chapter 18.7) • Example: Interest Rates and Deficits • Do large Federal budget deficits drive up long-term interest rates? • We want to regress the 10 year treasury bond rate on the 1 year rate, inflation, the real deficit per capita, and the change in real per capita income. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-57 Example: Deficits (cont.) • Augmented Dickey–Fuller tests suggest that the interest rates, inflation, deficit per capita, and level of income per capita are all integrated I(1). • Our regression may be consistent, depending on whether the disturbances contain a stochastic trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-58 TABLE 18.12 OLS Estimates of a Model of Long-term Interest Rates Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-59 Example: Deficits • The Durbin–Watson statistic of 1.71 suggests that disturbances are not serially correlated. • We apply a Dickey–Fuller test to the residuals to test for a stochastic trend in the disturbances. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-60 TABLE 18.13 Augmented Dickey–Fuller Test for a Stochastic Trend in Disturbances Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-61 Example: Deficits • The Dickey–Fuller test rejects the null hypothesis of a stochastic trend in the disturbances. • We do NOT seem to be in the spurious regression case. • Instead, we are in the cointegration case. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-62 Example: Deficits (cont.) • Because Deficits and 10-year interest rates are cointegrated, our coefficient on USDEF is consistent. • Our t-test on USDEF is incorrect. • We can re-estimate the analysis with Dynamic OLS to obtain proper test statistics. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-63 TABLE 18.14 Dynamic OLS Estimates of a Model of Long-term Interest Rates Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-64 Example: Deficits • The Dynamic OLS t-statistic on USDEF is 5.59 • We can reject the null hypothesis that USDEF = 0 • Deficits appear to raise interest rates Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-65 Example: Deficits (cont.) • An error correction specification, including the speed of adjustment , provides richer information about the short-run behavior of interest rates. • We can use the (consistent) coefficients from the Dynamic OLS model to construct estimated distances, so we can estimate the speed of adjustment. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-66 Example: Deficits (cont.) 1) Use Dynamic OLS to estimate 1 , 2 , 3 2) Use these estimates to construct Distancet Rate10t ˆ1 Rate1t ˆ2 Inf t ˆ3USDEFt 3) Estimate the speed of error correction using DRatet a 0 ( Distancet -1 ) t Note: See the textbook for the case with serial correlation. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-67 Time Trends • What have we learned about time trends? • We know that many macroeconomic variables contain stochastic trends. • We must be on guard against spurious regressions and cointegration. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-68 Time Trends (cont.) 1. Use Dickey–Fuller or ADF tests on time series data to judge which variables are likely to contain stochastic trends. 2. If you use stochastically trending variables in a regression, follow up with a DF or ADF test on the residuals. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-69 Time Trends (cont.) 3. If an explanator and the disturbances both contain stochastic trends, use the Granger and Newbold method for spurious regressions (difference the variables). 4. If the disturbances are not stochastically trending, but Y is, then use dynamic OLS and an error correction regression to examine long-run and short-run behavior. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-70 Time Trends (cont.) 5. Warning: Dickey–Fuller and Augmented Dickey-Fuller tests often have low power. We seldom know with great confidence whether variables are integrated. 6. In addition, regressions with differenced variables as in the Newbold–Granger method are also more likely to produce low-power hypothesis tests. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-71 TABLE 18.1 The Consistency of OLS and the Validity of Conventional Tests in the Face of Stochastic Trends* Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-72 Review • Suppose Yt 0 1 X t ut • The Dickey–Fuller test can fail to reject the null hypothesis that Y contains a stochastic trend. • A stochastic trend in Y can come from a stochastic trend in X, a stochastic trend in u, or both. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-73 Review (cont.) Yt 0 1 X t ut • If Y and X both contain stochastic trends, and u does not, then Y inherits its only stochastic trending component from X. • We call Y and X cointegrated. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-74 Review (cont.) • Cointegrated variables both wander stochastically, but stay near each other. • They adjust to each other’s locations through a process of error correction. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-75 Review (cont.) • The Lesson: – If we are going to look at changes in Y and X when Y and X are cointegrated, we must also examine the “cointegrating vector” that will be the error correction mechanism holding together Y and X as they trend. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-76 Review (cont.) • The order of cointegration depends on the trend in the distance between Y and Z. • If the distance between Y and Z contains no stochastic trend, then Y and Z are cointegrated of order 0. • For Y and Z to be cointegrated of order r, then both Y and Z are integrated I(r + 1). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-77 Review (cont.) • We can model cointegrated motion as Yt Yt 1 e t a1 (Yt 1 Z t 1 ) Z t Z t 1 wt a 2 (Yt 1 Z t 1 ) • a1 and a2 are the speeds of adjustment. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-78 Review (cont.) Long-Run Relationship Yt 1 X t t Short-Run Error Correction DYt a 0 y (Yt 1 1 X t 1 ) t DX t 0 x (Yt 1 1 X t 1 ) vt Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-79 Review (cont.) DYt a 0 y (Yt 1 1 X t 1 ) t DX t 0 x (Yt 1 1 X t 1 ) vt Yt 1 X t t • Y and X adjust towards their long-run relationship at speeds y and x. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-80 Review (cont.) • When et contains a stochastic trend, OLS is inconsistent. Use the Granger and Newbold method (differencing Y and X ). • When the et ’s do not contain stochastic trends, but an explanator does, OLS is super consistent, but not asymptotically normal. Our tests do not work (for the trending explanator). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-81 Review (cont.) Yt 0 1Z1t 2 Z 2t 3 X t e t • Suppose Y, Z1, and Z2 contain stochastic trends, but X and e do not. • We can estimate all the coefficients consistently. • Only our estimate of 3 is asymptotically normal. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-82 Review (cont.) Yt 0 1Z1t 2 Z 2t 3 X t e t • Our hypothesis tests are not asymptotically valid for the variables that contain a stochastic trend (Z1 and Z2). • Stock and Watson’s insight: change the equation so that the 1 and 2 coefficients apply to non-trending variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-83 Review (cont.) • Dynamic OLS: – A method for estimating cointegrated models – Respecify the model so that the coefficients of interest apply to non-trending variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-84 Review (cont.) Case: no serial correlation We wish to estimate Yt 0 1Z1t 2 Z 2t 3 X t e t Yt , Z1t , Z 2t contain stochastic trends. Rewrite the model as Yt 0 1Z1t 2 Z 2t 3 X t 4 DZ1t 5 DZ 2t t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-85 Review (cont.) Case: no serial correlation Rewrite the model as Yt 0 1Z1t 2 Z 2t 3 X t 4 DZ1t 5 DZ 2t t After controlling for DZ1t , there is no stochastic trend in Z1t . Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-86 Review (cont.) • An error correction specification, including the speed of adjustment , provides richer information about the short-run behavior of interest rates. • We can use the (consistent) coefficients from the Dynamic OLS model to construct estimated distances, so we can estimate the speed of adjustment. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-87 Review (cont.) Case: no serial correlation 1) Use Dynamic OLS to estimate 1 , 2 , 3 2) Use these estimates to construct Distance Y ˆ Z ˆ Z ˆ Z t t 1 1t 2 2t 3 3t 3) Estimate the speed of error correction using DYt a 0 ( Distancet -1 ) t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-88 Review (cont.) • What have we learned about time trends? • We know that many macroeconomic variables contain stochastic trends. • We must be on guard against spurious regressions and cointegration. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-89 Review (cont.) 1. Use Dickey–Fuller or ADF tests on time series data to judge which variables are likely to contain stochastic trends. 2. If you use stochastically trending variables in a regression, follow up with a DF or ADF test on the residuals. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-90 Review (cont.) 3. If an explanator and the disturbances both contain stochastic trends, use the Granger and Newbold method for spurious regressions (difference the variables). 4. If the disturbances are not stochastically trending, but Y is, then use dynamic OLS and an error correction regression to examine long-run and short-run behavior. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-91 Review (cont.) 5. Warning: Dickey–Fuller and Augmented Dickey–Fuller tests often have low power. We seldom know with great confidence whether variables are integrated. 6. In addition, regressions with differenced variables as in the Newbold–Granger method are also more likely to produce low-power hypothesis tests. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 27-92