Lecture 12: Joint Hypothesis Tests (Chapter 9.1–9.3, 9.5–9.6) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. Today’s Agenda • Review • Joint Hypotheses (Chapter 9.1) • F-tests (Chapter 9.2–9.3) • Applications of F-tests (Chapter 9.5– 9.6) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-2 Review • Perfect multicollinearity occurs when 2 or more of your explanators are jointly perfectly correlated. • That is, you can write one of your explanators as a linear function of other explanators: X1 aX2 bX 3 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-3 Review (cont.) • OLS breaks down with perfect multicollinearity (and standard errors blow up with near perfect multicollinearity). • Multicollinearity most frequently occurs when you want to include: – Time, age, and birth year effects – A dummy variable for each category, plus a constant Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-4 Review (cont.) • Dummy variables (also called binary variables) take on only the values 0 or 1. • Dummy variables let you estimate separate intercepts and slopes for different groups. • To avoid multicollinearity while including a constant, you need to omit the dummy variable for one group (e.g. males or non-Hispanic whites). You want to pick one of the larger groups to omit. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-5 Review (cont.) Yi 0 1 D _1i 2 D _ 2i 1 X i 2 X i D _1i 3 X i D _ 2i i 0 is the intercept for the omitted category. 0 1 is the intercept for the category coded by D _1. 0 2 is the intercept for the category coded by D _ 2. You can test whether group D _1 has the same intercept as the omitted group by testing H 0 : 1 0. 1 is the slope for the omitted category. 1 2 is the slope for the category coded by D _1. 1 3 is the slope for the category coded by D _ 2. You can test whether group D _1 has the same slope as the omitted group by testing H 0 : 2 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-6 Review (cont.) • You can multiply 2 variables together to create interaction terms. • Interaction terms let the slope of each variable depend on the value of the other variable. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-7 Review (cont.) Yi 0 1 X 1i 2 X 2i 3 X 1i X 2i i Yi 1 3 X 2 i X 1i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-8 Review (cont.) • With many of the specifications covered last time, we encountered hypotheses that required us to test multiple conditions simultaneously. For example, to test: – All categories have the same intercept (with 3 or more categories) – All categories have the same slope (with 3 or more categories) – One explanator has no effect on Y, when that explanator has been used in an interaction term Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-9 Review (cont.) • In economics, many processes are non-linear. Economic theory relies heavily on diminishing marginal returns, decreasing returns to scale, etc. • We want a specification that lets the 50th unit of X have a different marginal effect than the 1st unit of X. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-10 Review (cont.) • If we regress not Yi 0 1 X i i but rather Yi 0 1 X i 2 X i i 2 then the marginal benefit of a unit of X changes to: Y 1 22 X Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-11 Review (cont.) Yi 0 1 X i 2 X i i 2 Y 1 22 X • If 2 > 0, then the marginal impact of X is increasing. If 2 = 0, then X has a constant marginal effect. If 2 < 0, then the marginal impact of X is decreasing. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-12 Review (cont.) log(earnings)i 0 1 Edi 2 Expi 3 Expi 2 i • If 2 > 0 and 3 < 0, then this equation traces an inverse parabola. • Earnings increases quickly in experience at first, but then flattens out. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-13 Joint Hypotheses (Chapter 9.1) log(earnings)i 0 1 Edi 2 Expi 3 Expi 2 i • To test the hypothesis that experience is not an explanator of log(earnings), you need to test H0 : 2 0 AND 3 0 • WARNING: you CANNOT simply look individually at the t-test for 2 = 0 and the t-test for 3 = 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-14 Joint Hypotheses (cont.) • You CANNOT test a JOINT hypothesis by combining multiple t-tests. • Suppose you are testing H 0 : 1 0 AND 2 0 • A t-test rejects 1 = 0 if the data would be very surprising to see, given that 1 = 0. A t-test does NOT reject 1 = 0 if the data would only be pretty surprising. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-15 Joint Hypotheses (cont.) • Each t-test could fail to reject the null if the data would only be “pretty surprising” under each null, taken one at a time. • However, it might be “very surprising” to see two “pretty surprising” events. • We do not know the “size” of a joint test conducted by stacking together many t-tests. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-16 Joint Hypotheses (cont.) • Another problem with t-tests: suppose X1 and X2 are heavily correlated with each other (though not so much as to create perfect multicollinearity). Then each coefficient will have a large standard error. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-17 Joint Hypotheses (cont.) • Another problem with t-tests: suppose X1 and X2 are heavily correlated with each other. • If you remove either variable—leaving in the other—then you lose very little explanatory power. The other variable simply picks up the slack (through the omitted variables bias formula). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-18 Joint Hypotheses (cont.) • However, to test the null hypotheses that neither variable has explanatory power, we want to consider removing both variables at the same time. The two of them together may share a lot of explanatory power, even if either one could do the job nearly as well as both together. • We need a new type of test, that lets us consider multiple hypotheses at once. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-19 Joint Hypotheses (cont.) • Simply including more than one coefficient in the hypothesis does NOT make a joint hypothesis. For example, suppose you believed that X1 and X2 had identical effects. You could test this claim with: H 0 : 1 2 • This test is a single hypotheses, and can be tested using a t-test. The calculation requires you to know the covariance of the two coefficients. See Chapter 7.5. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-20 Joint Hypothesis (cont.) • A joint hypothesis tests more than one condition simultaneously. The easiest way to see how many conditions are being tested is to count the number of equal signs. • E.g. H0 : 1 = 0 AND 2 = 0 has two equal signs, so there are two conditions being tested. This is a joint test. • This hypothesis is often written 1 2 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-21 F-tests (Chapter 9.2–9.3) • How can we test multiple conditions simultaneously? • Intuition: run a regression normally, and then also run a regression where you assume the conditions are true. See if imposing the conditions makes a big difference. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-22 F-tests (cont.) log(earnings)i 0 1 Edi 2 Expi 3 Expi 2 i • To test the hypothesis that experience is not an explanator of log(earnings), you need to test H0 : 2 = 0 AND 3 = 0 • If these conditions are true, then there should be little difference between our “unrestricted” regression and the “restricted” version: log(earnings)i 0 1Edi 0 Expi 0 Expi 2 i 0 1Edi i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-23 F-tests (cont.) • If the conditions we are testing are true, then there should be little difference between our “unrestricted” regression and the “restricted” version. • What do we mean by “little difference”? • Does imposing the restrictions we wish to test greatly affect the model’s ability to fit the data? • We can turn to our measure of fit, R2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-24 F-tests (cont.) • To measure the difference in the quality of fit before and after we impose the restrictions we are testing, we can turn to our measure of fit, R2 ˆ 0 ˆ 1 X ) 2 ( Y SSR 2 R 1 1 2 TSS (Y Y ) • Notice that the Total Sum of Squares is the same for both versions of the regression, so we can focus on the Sum of Squares of the Residuals. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-25 F-tests (cont.) • Does imposing the restrictions from our null hypothesis greatly increase the SSR ? (Remember, we want a low SSR.) • Run both regressions and calculate the SSR. • Call the SSR for the unrestricted version the SSRu • Call the SSR for the restricted version the SSRr Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-26 F-tests (cont.) • Call the SSR for the unconstrained version the SSRu SSRu (log(earnings)i ˆ0 ˆ1Edi ˆ2 Expi ˆ3 Expi 2 )2 • Call the SSR for the constrained version the SSRc SSR c (log(earnings)i ˆ 0 ˆ1 Edi ) 2 • If the null hypothesis (2 = 3 = 0) is true, then imposing the restrictions will not change the SSR much. We will have a “small” SSRc-SSRu Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-27 F-tests (cont.) • If the null hypothesis is true, then imposing the restrictions will not change the SSR much. We will have a “small” SSRc-SSRu • Remember, OLS finds the smallest possible SSR. So SSRc > SSRu • The more restrictions we impose, the larger SSRc will get, even if the restrictions are true. • We need to adjust for the number of restrictions (r) we impose. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-28 F-tests (cont.) • To measure how large an effect our constraints have, look at: SSR SSR r c u • What constitutes a large difference? We want to compare the difference in SSR to the original SSRu. An increase of 100 units is more worrisome if we start from SSRu = 200 than if SSRu = 20,000. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-29 F-tests (cont.) • The more data we have, the more we trust our unconstrained regression. • Also, the more data we have, the more seriously we want to take a deterioration in SSR. • To capture the effect of more data, we weight by n-k-1. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-30 F-tests (cont.) SSR SSR r F u SSR (n k 1) c Copyright © 2006 Pearson Addison-Wesley. All rights reserved. u 12-31 F-tests (cont.) • When the i are distributed normally, the F-statistic will be distributed according to the F-distribution with r, n-k-1 degrees of freedom. • We know how to compute an F-statistic from the data. • We know the distribution of the F-statistic under the null hypothesis. • The F-statistic meets all the needs of a test statistic. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-32 F-tests (cont.) • If our null hypothesis is true, then imposing the hypothesized values as constraints on the regression should not change SSR much. Under the null, we expect a low value of F. • If we see a large value of F, then we can build a compelling case against the null hypothesis. • The F-table tells you the critical values of F for different values of r and n-k-1. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-33 F-tests (cont.) • Let’s return to the earnings test example, with the polynomial specification log(earnings)i 0 1 Edi 2 Expi 3 Expi 2 i • To test the hypothesis that experience is not an explanator of log(earnings), you need to test H 0 : 2 0 AND 3 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-34 F-tests (cont.) H 0 : 2 3 0 r 2; n - k - 1 6540 - 3 - 1 6536 Unconstrained Regression: log(earnings)i 0 1 Edi 2 Expi 3 Expi 2 i SSRu 3844 Constrained Regression: log(earnings)i 0 1 Edi vi SSR c 3959 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-35 F-tests (cont.) H 0 : 2 3 0 SSR c SSRu 3959 3844 r 2 F 97.77 u 3844 SSR 6536 n k 1 The 5% critical value for F2,6536 is 3. We can reject the null hypothesis that experience is not an explanator of log(income). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-36 F-tests (cont.) • Note: we must be able to impose the restrictions as part of an OLS estimation. We can impose only linear restrictions. • For example, we CAN test: 3 14 1 2 3 0 1 42 and 3 - 34 5 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-37 F-tests (cont.) • However, we CANNOT test: 1 ·2 3 1 3 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-38 F-tests (cont.) • Example: Y 0 1 X1 2 X 2 3 X 3 4 X 4 5 X 5 H0 : 1 42 and 3 - 34 5 • There are two equal signs. r = 2. • How do we impose the restrictions? Y 0 (4 2 )X1 2 X2 3 X3 4 X 4 ( 3 - 3 4 )X5 • How do we enter this regression into the computer? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-39 F-tests (cont.) • To enter a regression into the computer, we need to regroup so that all our explanators receive a single coefficient apiece. • We need to transform this expression from one with separated explanators and linear combinations of coefficients to one with separated coefficients and linear combinations of explanators. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-40 F-tests (cont.) Y 0 (42 )X1 2 X2 3 X3 4 X4 ( 3 - 34 )X5 Y 0 2 (4 X1 X2 ) 3 (X3 X5 ) 4 (X4 - 3X5 ) • To find the constrained sum of squares, we need to regress Y on a constant, (4X1+X2), (X3+X5), and (X4 -3X5). The SSR from this regression is our SSRc. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-41 Checking Understanding • You regress Y 0 1 X1 2 X2 • You want to test H 0 : 1 -2 • What are the constrained and unconstrained regressions? What is r ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-42 Checking Understanding (cont.) Y 0 1 X 1 2 X 2 H 0 : 1 -2 Unconstrained regression: Y 0 1 X 1 2 X 2 Constrained regression: Y 0 (- 2 ) X 1 2 X 2 0 2 ( X 2 - X1 ) r 1 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-43 F-tests • Note: when r = 1, you have a choice between using a t-test or an F-test. • When r = 1, F = |t|2. F-tests and t-tests will give the same results. • When r > 1, you cannot use a t-test. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-44 F-tests (cont.) • A frequently encountered test is the null hypothesis that all the coefficients (except the constant) are 0. This test asks whether the entire model is useless. Do our explanators do a better job at predicting Y than simply guessing the mean? • Many econometrics programs automatically calculate this F-statistic when they perform a regression. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-45 An Application of F-tests (Chapter 9.5) • Let’s use F-tests to re-examine the differences in earnings equations between black women and black men in the NLSY. • Regress the following for black workers: log(earnings)i 0 1 Edi 2 Expi 3 D _ Fi i • where Edi = years of education, Expi = years of experience, and D_Fi = 1 if the worker is female Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-46 An Application of F-tests (cont.) log(earnings)i 0 1 Edi 2 Expi 3 D _ Fi i • To test whether black males and black females have the same intercept, we can use a simple t-test with H0 : 3 = 0 • Our estimated coefficient is -0.201 with a standard error of 0.036, yielding a t-statistic of -5.566 • This t-statistic exceeds our critical value of -1.96 • We can reject the null hypothesis at the 5% level Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-47 TABLE 9.1 Earnings Equation for Black Men and Women (NLSY Data) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-48 An Application of F-tests (cont.) • We have rejected the null hypothesis that black men and black women have the same intercept. • Could they also have different slopes for education and experience? • We can use dummy variable interaction terms. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-49 An Application of F-tests (cont.) log(earnings )i 0 1 Edi 2 Expi 3 D _ Fi 4 Edi D _ Fi 5 Expi D _ Fi i Case: worker is male ( D _ Fi 0) : log(earnings )i 0 1 Edi 2 Expi i Case: worker is female ( D _ Fi 1) : log(earnings )i ( 0 3 ) ( 1 4 ) Edi ( 2 5 ) Expi i • To test the null hypothesis that black men and black women have identical earnings equations, we need to test the joint hypothesis: H 0 : 3 4 5 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-50 An Application of F-tests (cont.) H 0 : 3 4 5 0 Unconstrained Regression: log(earnings )i 0 1 Ed i 2 Expi 3 D _ Fi 4 Edi D _ Fi 5 Expi D _ Fi i SSR u 1002.75 Constrained Regression: log(earnings )i 0 1 Ed i 2 Expi vi SSR c 1020.378 r 3, n k 1 1795 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-51 An Application of F-tests (cont.) H 0 : 3 4 5 0 SSR c SSRu 1020.37 1002.75 r 3 F 10.51 u 1002.75 SSR 1795 n k 1 The critical value at the 5% significance level for F3,1795 is 2.60. We can reject the null hypothesis at the 5% level. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-52 An Application of F-tests (cont.) • We can reject the null hypothesis that black men and black women have identical earnings functions. • Do we really need the interaction terms, or do we get the same explanatory power by simply giving black women a different intercept? • Let’s test the null hypothesis that the interaction coefficients are both 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-53 An Application of F-tests (cont.) H 0 : 4 5 0 Unconstrained Regression: log(earnings )i 0 1 Ed i 2 Expi 3 D _ Fi 4 Edi D _ Fi 5 Expi D _ Fi i SSR u 1002.75 Constrained Regression: log(earnings )i 0 1 Ed i 2 Expi 3 D _ Fi vi SSR c 1003.08 r 2, n k 1 1795 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-54 An Application of F-tests (cont.) H 0 : 4 5 0 SSR c SSRu 1003.08 1002.75 r 2 F 0.30 u 1002.75 SSR 1795 n k 1 The critical value at the 5% significance level for F2,1795 is 3.00. We fail to reject the null hypothesis at the 5% level. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-55 F-tests and Regime Shifts (Chapter 9.6) • What is the relationship between Federal budget deficits and long-term interest rates? • We have time-series data from 1960–1994. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-56 F-tests and Regime Shifts (cont.) • Our dependent variable is long-term interest rates (LongTermt) • Our explanators are expected inflation (Inflationt), short-term interest rates (ShortTermt), change in real per-capita income (DeltaInct), and the real per-capita budget deficit (Deficitt). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-57 F-tests and Regime Shifts (cont.) LongTermt 0 1 Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt t • Note that we index observations by t, not i. • 4 is the change in long-term interest rates from a $1 increase in the Federal deficit (measured in 1996 dollars). • Financial market de-regulation began in 1982. • Was the relationship between long-term interest rates and Federal deficits altered by the de-regulation? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-58 F-tests and Regime Shifts (cont.) • We can let the co-efficient on Deficitt vary before and after 1982 by interacting with a dummy variable. • Create the variable D_1982t = 1 if the observation is for year 1983 or later LongTermt 0 1Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt 5 D _1982t 6 Deficitt D _1982t t • To test whether the slope on Deficitt changes after 1982, conduct a t-test of the hypothesis H0 : 6 = 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-59 F-tests and Regime Shifts (cont.) • • • • • • • • • Dependent Variable: LongTermt Independent Variables, with standard errors Inflationt: 0.765 (0.0372) ShortTermt: 0.822 (0.0586) DeltaInct: -6.4·10-6 (1.6·10-4) Deficitt: 0.002 (0.0004) D_1982t: -0.739 (0.6608) Deficitt·D_1982t: 0.0005 (0.0007) Constant: 1.277 (0.2135) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-60 F-tests and Regime Shifts (cont.) • For the period 1960–1982, the slope on Deficitt is 0.0022. A $1 increase in the Federal deficit per capita increases long-term interest rates by 0.0022 points. • For the period 1983–1994, the slope on Deficitt is 0.0022 + 0.0005 = 0.0027. • The t-statistic for Deficitt·D_1982t is 0.63. We fail to reject the null hypothesis that the slopes are different. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-61 F-tests and Regime Shifts (cont.) • For the period 1960–1982, the slope on Deficitt is 0.0022. A $1 increase in the Federal deficit per capita increases long-term interest rates by 0.0022 points. • Is this change important in magnitude? One quick, crude way to assess magnitudes is to ask, “How many standard deviations does Y change when I change X by 1 standard deviation?” Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-62 F-tests and Regime Shifts (cont.) • Standard Deviation of Deficitt = 463 • Standard Deviation of LongTermt = 2.65 • A 1-standard-deviation change in Deficitt is predicted to cause a 463·0.0022 = 1.02 percentage point change in LongTermt, or about a third of a standard deviation • At first glance, the effect of Federal deficits on interest rates is non-negligible, but not massive, either. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-63 F-tests and Regime Shifts (cont.) • Let’s test a more complicated hypothesis. • Does the entire financial regime shift after 1982? • Let’s let every coefficient vary between the two time periods. LongTermt 0 1Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt 5 D _1982t 6 Inflationt D _1982t 7 ShortTermt D _1982t 8 DeltaInct D _1982t 9 Deficitt D _1982t t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-64 F-tests and Regime Shifts (cont.) • Does the entire financial regime shift after 1982? • Test the joint hypothesis that every interaction term is 0: H 0 : 5 6 7 8 9 0 • We need an F-test • There are 5 equal signs, so r = 5 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-65 F-tests and Regime Shifts (cont.) H 0 : 5 6 7 8 9 0 Unconstrained regression: LongTermt 0 1 Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt 5 D _1982t 6 Inflationt D _1982t 7 ShortTermt D _1982t 8 DeltaInct D _1982t 9 Deficitt D _1982t t Constrained regression: LongTermt 0 1 Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt vt Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-66 F-tests and Regime Shifts (cont.) H 0 : 5 6 7 8 9 0 SSR c SSRu 4.93 4.15 r 5 F 0.94. u 4.15 SSR 25 n k 1 The critical value for an F-test with 5,25 degrees of freedom is 2.60. We cannot reject the null hypothesis that there is no regime shift. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-67 F-tests and Regime Shifts (cont.) • Instead of using dummy variables, we could conduct this same test by running the same regression on 3 separate datasets. • For the constrained regression (there is no regime shift), we use all the data, 1960–1994. • For the unconstrained regression (there is a regime shift), we run separate regressions for 1960–1982 and 1983–1994. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-68 F-tests and Regime Shifts (cont.) LongTermt 0 1 Inflationt 2 ShortTermt 3 DeltaInct 4 Deficitt t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-69 F-tests and Regime Shifts (cont.) • For each regression, we record the SSR • SSRc is the SSR from the regression for 1960–1994, SSR1960–1994 • SSRu is the sum SSR1960–1982 + SSR1983–1994 1982 t 1960 et 2 1994 t 1983 et 2 1994 t 1960 et 2 • Using these SSR ’s, we can compute F Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-70 F-tests and Regime Shifts (cont.) SSRu SSR19601982 SSR19831994 2.17 1.98 4.15 SSR c SSR19601994 4.93 r 5 n k 1 25 SSR c SSRu 4.93 4.15 r 5 F 0.94 u 4.15 SSR 25 n k 1 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-71 F-tests and Regime Shifts (cont.) • See Chapter 9.7 for additional tests for regime shifts. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-72 Review • How can we test multiple conditions simultaneously? • Intuition: run a regression normally, and then also run a regression where you constrain the parameters to make the null hypothesis true. See if imposing the conditions makes a big difference. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-73 Review (cont.) • Does imposing the restrictions from our null hypothesis greatly increase the SSR ? (Remember, we want a low SSR.) • Run both regressions and calculate the SSR • Call the SSR for the unrestricted version the SSRu • Call the SSR for the restricted version the SSRr Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-74 Review (cont.) • Run both regressions and calculate the SSR • Call the SSR for the unrestricted version the SSRu • Call the SSR for the restricted version the SSRr • If the null hypothesis is true, then imposing the restrictions will not change the SSR much. We will have a “small” SSRr-SSRu Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-75 Review (cont.) SSR SSR r F u SSR (n k 1) c Copyright © 2006 Pearson Addison-Wesley. All rights reserved. u 12-76 Review (cont.) • When the i are distributed normally, the F-statistic will be distributed according to the F-distribution with r, n-k-1 degrees of freedom • We know how to compute an F-statistic from the data • We know the distribution of the F-statistic under the null hypothesis • The F-statistic meets all the needs of a test statistic Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-77 Review (cont.) • If our null hypothesis is true, then imposing the hypothesized values as constraints on the regression should not change SSR much. Under the null, we expect a low value of F. • If we see a large value of F, then we can build a compelling case against the null hypothesis. • The F table tells you the critical values of F for different values of r and n-k-1. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 12-78