mV eouus tebe JACK JOHNSTON JOHN DiNARDOhttp://www.mhcollege.com P/N 032720-3 PART OF ISBN O-07-913121-2 A McGraw-Hill A Division of The McGraw-Hill Companiesvi ECONOMETRIC METHODS 1.3. To derive cov(a, b) 1.4 Gauss-Markov theorem 1.5. To derive var(eo) Problems Further Aspects of Two-Variable Relationships 2.1 Time as a Regressor 21.1 Constant Growth Curves ~ 2.1.2 Numerical Example 2.2. Transformations of Variables 2.2.1 Log-Log Transformations 2.2.2. Semilog Transformations 2.2.3 Reciprocal Transformations 23. An Empirical Example of a Nonlinear Relation: U.S. Inflation and Unemployment 24° Lagged Dependent Variable as Regressor 24.1 An Introduction to Asymptotics 24.2 Convergence in Probability 24,3 Convergence in Distribution 24.4 The Autoregressive Equation 2.5. Stationary and Nonstationary Series 2.5.1 Unit Root 2.5.2 Numerical IMustration 2.6 Maximum Likelihood Estimation of the Autoregressive Equation 2.6.1 Maximum Likelihood Estimators 2.6.2. Properties of Maximum Likelihood Estimators Appendix 2.1 Change of variables in density functions 2.2 Maximum likelihood estimators for the AR(1) model Problems The k-Variable Linear Equation 3.1 Matrix Formulation of the &-Variable Model 1.1 The Algebra of Least Squares 3.1.2 Decomposition of the Sum of Squares 3.1.3 Equation in Deviation Form 3.2. Partial Correlation Coefficients 3.2.1 Sequential Buildup of the Explained Sum of Squares 3.2.2 Partial Correlation Coefficients and Multiple Regression Coefficients 3.2.3 General Treatment of Partial Correlation and Multiple Regression Coefficients 3.3 The Geometry of Least Squares 3.4 Inference in the k-Variable Equation 3.4.1 Assumptions 3.4.2 Mean and Variance of b 36 36 37 37 41 a2 43 4B 45 47 49 52 53 55 56 37 39 59 61 61 63 65 69 70 70 nD B 6 B 81 82 83 86 86 873.4.3. Estimation of o? 3.4.4 Gauss—Markov Theorem 3.45. Testing Linear Hypotheses about B 3.46 Restricted and Unrestricted Regressions 3.4.7 Fitting the Restricted Regression 3.5. Prediction Appendix 3.1 To prove rig = (riz ~ nist VIF JT 3.2. Solving fora single regression coefficient in a multiple regression 3.3 To show that minimizing a’a subject to X’a = c gives a= XXX) e 3.4 Derivation of the restricted estimator b. Problems Some Tests of the k-Variable Linear Equation for Specification Error 4.1 Specification Error 4.1.1 Possible Problems with w 4.1.2. Possible Problems with X 4.1.3 Possible Problems with B 4.2. Model Evaluation and Diagnostic Tests 4.3. Tests of Parameter Constancy 4.3.1 The Chow Forecast Test 43.2. The Hansen Test 4.3.3. Tests Based on Recursive Estimation 43.4 One-Step Ahead Prediction Errors 43.5 CUSUM and CUSUMSQ Tests, 43.6 A More General Test of Specification Error: ‘The Ramsey RESET Test 44 A Numerical Illustration 45. Tests of Structural Change 4.5.1 Test of One Structural Change 45.2 Tests of Slope Coefficients 45.3. Tests of Intercepts 45.4 Summary 45.5 A Numerical Example 45.6 Extensions 46/ Dummy Variables ’ 4.6.1 Introduction 46.2 Seasonal Dummies 4.6.3 Qualitative Variables 4.6.4 Two ot Mote Sets of Dummy Variables 4.6.5 A Numerical Example Appendix 4.1 To show var(d) = 0? |I,, + Xo(X}Xi)'X3] Problems Contents vii 89 89 90 95 96 99 100 101 103 103 104 109 109 110 10 ul 112 113 113 116 17 118 119 121 121 126 127 128 128 129 130 132 133 133 134 135 137 137 138 139vi BCONOMETRIC METHODS \5 Maximum Likelihood (ML), Generalized Least Squares \ (GLS), and Instrumental Variable (IV) Estimators a 5.1 52 53 54 55 ‘Maximum Likelihood Estimators 5.1.1 Properties of Maximum Likelihood Estimators ML Estimation of the Linear Model Likelihood Ratio, Wald, and Lagrange Multiplier Tests 5.3.1 Likelihood Ratio (LR) Tests 5.3.2 The Wald (W) Test 5.3.3 Lagrange Multiplier (LM) Test ML Estimation of the Linear Model with Nonspherical Disturbances 5.4.1 Generalized Least Squares Instrumental Variable (IV) Estimators 5.5.1 Special Case 5.5.2. Two-Stage Least Squares (2SLS) 5.5.3 Choice of Instruments 5.5.4 Tests of Linear Restrictions Appendix Sd 5.2 53 Change of variables in density functions Centered and uncentered R? To show that €,X(X'X)"!X'e. = ele. — e'e Problems Heteroscedasticity and Autocorrelation 61 62 63 64 65 66 67 Properties of OLS Estimators Tests for Heteroscedasticity 6.2.1 The White Test 6.2.2. The Breusch-Pagan/Godfrey Test 6.2.3. The Goldfeld-Quandt Test 62.4 Extensions of the Goldfeld-Quandt Test Estimation Under Heteroscedasticity 6.3.1 Estimation with Grouped Data 6.3.2 Estimation of the Heteroscedasticity Relation Autocorrelated Disturbances 6.4.1 Forms of Autocorrelation: Autoregressive and Moving Average Schemes 6.4.2 Reasons for Autocorrelated Disturbances OLS and Autocorrelated Disturbances ‘Testing for Autocorrelated Disturbances 6.6.1 Durbin-Watson Test 6.6.2 The Wallis Test for Fourth-Order Autocorrelation 6.6.3 Durbin Tests for a Regression Containing Lagged Values of the Dependent Variable 6.6.4 Breusch-Godfrey Test 66.5 Box-Pierce-Ljung Statistic Estimation of Relationships with Autocorrelated Disturbances 142 142 13 145 147 147 148, 149 151 152 153 156 157 157 158 158 159 160 161 162 163 166 166 167 168 168 170 im im 174 175 176 176 178 179 182 182 185 187 1886.8 Forecasting with Autocorrelated Disturbances 6.9 Autoregressive Conditional Heteroscedasticity (ARCH) Appendix 6.1 LM test for multiplicative heteroscedasticity 6.2. LR test for groupwise homoscedasticity 6.3. Properties of the ARCH(1) process Problems Univariate Time Series Modeling 7.1 A Rationale for Univariate Analysis 71.1 The Lag Operator 7.1.2. ARMA Modeling 7.2 Properties of AR, MA, and ARMA Processes 7.2.1 AR(1) Process 7.2.2 AR(Q) Process 7.23 MA Processes 7.2.4 ARMA Processes 7.3 Testing for Stationarity 73.1 Graphical Inspection 7.3.2. Integrated Series 7.3.3 Trend Stationary (TS) and Difference Stationary (DS) Series 7.3.4 Unit Root Tests 73.5 Numerical Example 7.4 Identification, Estimation, and Testing of ARIMA Models 74.1 Identification 7.4.2 Estimation 7.4.3 Diagnostic Testing 7.5 Forecasting 7.5.1 MA(1) Process 7.5.2. ARMA(I,1) Process 7.5.3. ARIMA(I,1,0) Process 7.6 Seasonality 7.7 A Numerical Example: Monthly Housing Starts Problems Autoregressive Distributed Lag Relationships 8.1 Autoregressive Distributed Lag Relations 8.1.1 A Constant Elasticity Relation 8.1.2 Reparameterization a 8.1.3 Dynamic Equilibrium 8.1.4 Unit Elasticity 8.1.5 Generalizations 8.2 Specification and Testing 8.2.1 General to Simple and Vice Versa 8.2.2 Estimation and Testing 8.2.3 Exogeneity Contents ix 192 195, 198 200 201 202 204 205 206 207 207 207 209 213 214 215 216 220 220 223, 226 228 228 229 230 231 232 233 234 235 236 243 244 244 245 245 246 246 247 248 248, 250 253X__ ECONOMETRIC METHODS 8.2.4 Exogeneity Tests 8.2.5 The Wu-Hausman Test 8.3 Nonstationary Regressors 84 ANumerical Example 8.4.1 Stationarity 8.4.2 Cointegration 8.4.3 A Respecified Relationship 8.4.4 A General ADL Relation 8.4.5 A Reparameterization 8.5 Nonnested Models ‘Appendix 8.1 Nonsingular linear transformations of the variables in an equation 8.2. To establish the equality of the test statistics in Eqs. (8.37) and (8.41) Problems 9 Multiple Equation Models 9.1 Vector Autoregressions (VARs) 9.1.1 A Simple VAR 9.1.2. A Three-Variable VAR 9.1.3 Higher-Order Systems 9.2. Estimation of VARs 9.2.1 Testing the Order of the VAR 9.2.2 Testing for Granger Causality . 9.2.3 Forecasting, Impulse Response Functions, and Variance Decomposition 9.2.4 Impulse Response Functions 9.2.5 Orthogonal Innovations 9.2.6 Variance Decomposition 9.3. Vector Error Correction Models 9.3.1 Testing for Cointegration Rank 9.3.2 Estimation of Cointegrating Vectors 9.3.3 Estimation of a Vector Error Correction Model 9.4 Simultaneous Structural Equation Models 9.5 Identification Conditions 9.6 Estimation of Structural Equations 9.6.1 Nonstationary Variables 9.6.2 System Methods of Estimation Appendix 9.1 Seemingly Unrelated Regressions (SUR) 9.2 Higher-order VARs 9.2.1 A VAR(I) Process 9.2.2. A VAR(2) Process Problems 256 257 259 265 266 266 270 2 215 282 284 285 287 287 287 292 294 295 296 296 297 298 299 301 301 302 303 305 305 309 314 317 317 318 320 320 321 322Contents xi 10 Generalized Method of Moments v nu 12 10.1 10.2 103 10.4 10.5 106 10.7 ‘The Method of Moments OLS as a Moment Problem Instrumental Variables as a Moment Problem GMM and the Orthogonality Condition Distribution of the GMM estimator Applications 10.6.1 Two-Stage Least Squares, and Tests of Overidentifying Restrictions 10.6.2. Wu-Hausman Tests Revisited 10.6.3. Maximum Likelihood 10.6.4 Euler Equations Readings Problems A Smorgasbord of Computationally Intensive Methods Wd 2 3 4 us 116 An Introduction to Monte Carlo Methods 11.1.1 Some Guidelines for Monte Carlo Experiments 11.1.2 An Example 11.1.3 Generating Pseudorandom Numbers 11.1.4 Presenting the Results Monte Carlo Methods and Permutation Tests The Bootstrap 11.3.1 The Standard Error of the Median 11.3.2 An Example 11.3.3. The Parametric Bootstrap 11.3.4 Residual Resampling: Time Series and Forecasting 11.3.5 Data Resampling: Cross-Section Data 11.3.6 Some Remarks on Econometric Applications of the Bootstrap Nonparametric Density Estimation 11.4.1 Some General Remarks on Nonparametric Density Estimation 11.4.2 An Application: The Wage Effects of Unions Nonparametric Regression 11.5.1 Extension: The Partially Linear Regression Model References Problems Panel Data 121 122 123 124 12.5 Sources and Types of Panel Data The Simplest Case—The Pooled Estimator ‘Two Extensions to the Simple Model The Random Effects Model Random Effects as a Combination of Within and Between Estimators 327 328 329 330 333 335 336 336 338 342 343 344 345 348 348 349 350 352 354 359 362 362 363 365 366 369 369 370 375 316 379 383 385 385 388 389 300 390 301 392xii 13 ECONOMETRIC METHODS 12.6 The Fixed Effects Model in the Two-Period Case 12.7 The Fixed Effects Model with More Than Two ‘Time Periods 12.8 The Perils of Fixed Effects Estimation 12.8.1 Example 1: Measurement Error in X 12.8.2 Example 2: Endogenous X 12.9 Fixed Effects or Random Effects? 12.10 A Wu-Hausman Test 12.11 Other Specification Tests and an Introduction to Chamberlain's Approach 12.11. Formalizing the Restrictions 1211.2 Fixed Effects in the General Model 12.113 Testing the Restrictions 12.12, Readings Problems Discrete and Limited Dependent Variable Models 13.1 Types of Discrete Choice Models 13.2. The Linear Probability Model 13.3 Example: A Simple Descriptive Model of Union Participation 13.4 Formulating a Probability Model 13.5 The Probit 13.6 ‘The Logit 13.7 Misspecification in Binary Dependent Models 13.7.1 Heteroscedasticity 13.7.2 Misspecification in the Probit and Logit 13.7.3. Functional Form: What Is the Right Model to Use? 13.8 Extensions to the Basic Model: Grouped Data 13.8.1 Maximum Likelihood Methods 13.8.2 Minimum y* Methods 13.9 Ordered Probit 13.10 Tobit Models 13.10.1_ ‘The Tobit as an Extension of the Probit 13.102. Why Not Ignore “The Problem”? 13.103 Heteroscedasticity and the Tobit 13.11 Two Possible Solutions 13.11.1 Symmetrically Trimmed Least Squares 13.11.2. Censored Least Absolute Deviations (CLAD) Estimator 13.12. Treatment Effects and Two-Step Methods 1312.1 The Simple Heckman Correction 1312.2 Some Cautionary Remarks about Selectivity Bias 13.12.3 The Tobit as a Special Case 13.13 Readings Problems 395 397 399, 399 402 403 403 404 407 407 408 412 412 44 45 418 419 424 426 426 427 430 432 432 433 434 436 436 439 440 441 442 447 449 450 452 452Appendix A Al Vectors A.L1 Multiplication by a Scalar A.1.2 Addition and Subtraction A.13 Linear Combinations A.L4 Some Geometry ALS Vector Multiplication A.L6 Equality of Vectors A2 Matrices A2.1 Matrix Multiplication A.2.2 The Transpose of a Product A23 Some Important Square Matrices A24 Partitioned Matrices A25 — Matrix Differentiation A.2.6 Solution of Equations A27 The Inverse Matrix A28 The Rank of a Matrix A.2.9 Some Properties of Determinants A2.10 Properties of Inverse Matrices A2.11 More on Rank and the Solution of Equations A212 Eigenvalues and Bigenvectors A.2.13 Properties of Eigenvalues and Eigenvectors A214 Quadratic Forms and Positive Definite Matrices Appendix B B.1__ Random Variables and Probability Distributions B.2__ The Univariate Normal Probability Distribution B.3_ Bivariate Distributions B.4__ Relations between the Normal, x, t, and F Distributions B.5 Expectations in Bivariate Distributions B.6 Multivariate Densities B.7 Multivariate Normal pdf B.8__ Distributions of Quadratic Forms B.9 Independence of Quadratic Forms B.10 Independence of a Quadratic Form and a Linear Function Appendix C Appendix D Index Contents xiii 455 456 456 456 456 458 458 459 459 461 463 464 465 467 4m 472 473 416 478 483 485 486 487 489, 490 490 492 493 495 496 497 499 521CHAPTER 1 Relationships between Two Variables The economies literature contains innumerable discussions of relationships be- tween variables in pairs: quantity and price; consumption and income; demand for money and the interest rate; trade balance and the exchange rate; education and income; unemployment and the inflation rate; and many more. This is not to say that economists believe that the world can be analyzed adequately in terms of a collection of bivariate relations. When they leave the two-dimensional diagrams of the text- books behind and take on the analysis of real problems, multivariate relationships abound. Nonetheless, some bivariate relationships are significant in themselves; more importantly for our purposes, the mathematical and statistical tools developed for two-variable relationships are fundamental building blocks for the analysis of more complicated situations, 11 EXAMPLES OF BIVARIATE RELATIONSHIPS Figure 1.1 displays two aspects of the relationship between real personal saving (SAV) and real personal disposable income (INC) in the United States. In Fig. 1.la the value of each series is shown quarterly for the period from 1959.1 to 1992.1. These two series and many of the others in the examples throughout the book come from the DRI Basic Economics Database (formerly Citibase); where relevant, we indicate the correspondence between our labels and the Citibase labels for the vari- ables. Figure I.La is a typical example of a time series plot, in which time is dis- played on the horizontal axis and the values of the series are displayed on the vertical axis. Income shows an upward trend throughout the period, and in the early years, saving does likewise. This pattern, however, is not replicated in the middle and later 1A definition of all series is given in the data disk, which accompanies this volume, Instructions for accessing the disk are given in Appendix C. :2. BCONOMETRIC METHODS “earoouyqeuosiad aiqesodstp (22) ONT 38 8 &€ € 8 8 (smtp 861 40 sto 5 70 (etop 196830 sto1ta “Suyaes youosto 23) AVS Year @ + + at : + we ogee’ t | Tele : owe | we : th | + 3 : : g g g g Avs 2500-3000 3500 INC 1500 2000 1000 Oy FIGURE 1.1 Saving and income.cuarter 1: Relationships between Two Variables 3 years. One might be tempted to conclude from Fig. 1.1a that saving is much more volatile than income, but that does not necessarily follow, since the series have sep- arate scales.” An alternative display of the same information is in terms of a scatter plot, shown in Fig. 1.1b, Here one series is plotted against the other. The time dimension is no longer shown explicitly, but most software programs allow the option of joining successive points on the scatter so that the evolution of the series over time may still be traced. Both parts of Fig. 1.1 indicate a positive association between the variables: increases in one tend to be associated with increases in the other. It is clear that although the association is approximately linear in the early part of the period, it is not so in the second half, Figures 1.2 and 1.3 illustrate various associations between the natural log of real personal expenditure on gasoline (GAS), the natural log of the real price of gasoline (PRICE), and the natural log of real disposable personal income (INCOME). The derivations of the series are described in the data disk. The rationale for the logarith- mic transformations is discussed in Chapter 2. Figure 1.2 gives various time plots of gasoline expenditure, price, and income. The real price series, with 1987 as the base year, shows the two dramatic price hikes of the early and late 1970s, which were subsequently eroded by reductions in the nominal price of oil and by U.S. inflation, so the real price at the end of the period was less than that obtaining at the start, The income and expenditure series are both shown in per capita form, because U.S. population increased by about 44 percent over the period, from 176 million to 254 million, The population series used to deflate the expenditure and income series is the civilian noninstitutional population aged 16 and over, which has increased even faster than the general population. Per capita real expenditure on gasoline increased steadily in the 1960s and early 1970s, as real income grew and real price declined. This steady rise ended.with the price shocks of the 1970s, and per capita gas con- sumption has never regained the peak levels of the early seventies. The scatter plots in Fig. 1.3 further illustrate the upheaval in this market. The plot for the whole period in Fig. 1.3a shows very different associations between ex- penditure and price in the earlier and later periods. The scatter for 1959.1 to 1973.3 in Fig. 1.3b looks like a conventional negative association between price and quantity. This is shattered in the middle period (1973.4 to 1981.4) and reestablished, though with a very different slope, in the last period (1982.1 to 1992.1). This data set will be analyzed econometrically in this and later chapters. These illustrative scatter diagrams have three main characteristics. One is the sign of the association or covariation—that is, do the variables move together in a positive or negative fashion? Another is the strength of the association. A third characteristic is the linearity (or otherwise) of the association—is the general shape of the scatter linear or curvilinear? In Section 1.2 we discuss the extent to which the correlation coefficient measures the first two characteristics for a linear association, and in later chapters we will show how to deal with the linearity question, but first we give an example of a bivariate frequency distribution. 2See Problem 1.14 BCONOMETRIC METHODS. I er Gas 81 — z 4a o 6 70 8 80 85 90 . . 43 & -19 43 @) FIGURE 1.2 Time series plots of natural log of gasoline consumption in 1987 dollars per capita. (@) Gasoline consumption vs. natural log of price in 1987 cents per gallon. (6) Gasoline consumption vs. natural log of income in 1987 dollars per capita.cuarrer 1: Relationships between Two Variables 5 52 41 470 30 465 g g z 4s z 460 455 46 450 1059.1 1992.1 1959.1 0 19733 44 ; 44s ‘31 30 79 «478 497 16 ‘si 80 79 78 7736 cas cas @ © 52 52 30 “ sok Eas Gash S & 46 . ask 1973.40 1981.4 1982.1 0 19921 44 1 n a4 fi \ 480 475-770-765. Seo 77570765 7.80 cas Gas © @ FIGURE 1.3 ‘Scatter plots of price and gasoline consumption. 1.1.1 Bivariate Frequency Distributions The data underlying Figs. 1.1 to 1.3 come in the form of n pairs of observations of the form (Xj, ¥;), i = 1,2, .n. When the sample size n is very large, the data are usually printed as a bivariate frequency distribution; the ranges of X and ¥ are split into subintervals and each cell of the table shows the number of observations6 ECONOMETRIC METHODS TABLE Lt Distribution of heights and chest circumferences of 5732 Scottish militiamen. Chest circumference (inches) 45and Row 3335 36-38 39-41 42-44 over totals 64-65 39 331 326 26 0 2 Height 66-67 40 5911010 170 4 1815 (inches) 68-69 19 3124 488 18 1981 71 3 100 419 290 2B 397 2-13 0 17 120 153 27 317 (Column totals 103 135130791127 2 S732 Source: Edinburgh Medical and Surgical Journal (1817, pp. 260-264). TABLE 1.2 Conditional means for the data in Table 1.1 Mean of heis’ z:venchest inches) 66.31 66.84 67.89 69.16 70.53 ‘Mean of chest given height (inches) 38.41 39.19 40.26 40.76 41.80 in the corresponding pair of subintervals. Table 1.1 provides an example. It is not possible to give a simple, two-dimensional representation of these data. However, inspection of the cell frequencies suggests a positive association between the two measurements. This is confirmed by calculating the conditional means. First of all, each of the five central columns of the table gives a distribution of heights fora given chest measurement. These are conditional frequency distributions, and traditional statistics such as means and variances may be calculated. Similacly, the rows of the table give distributions of chest measurements, conditional on height. The two sets of conditional means are shown in Table 1.2; each mean series increases monotonically with increases in the coaditioning variable, indicating a positive association between the variables. 1.2 THE CORRELATION COEFFICIENT The direction and closeness of the linear association between two variables are mea- sured by the correlation coefficient. Let the observations be denoted by (X;, ¥;) with i = 1,2,...,1. Once the sample means have been calculated, the data may be ex- pressed in deviation form as m=Xi-X y=K-¥ ‘ondensed from Stephen M. Stigler, The History of Statistics, Harvard University Press, 1986, p. 208. “See Stigler, op. cit, fora fascinating and definitive history of the evolution of the correlation coefficient.cHarreR 1: Relationships between Two Variables 7 where X and Y denote the sample means of X and ¥. Figure 1.4 shows an illustrative point on a scatter diagram with the sample means as new axes, giving four quadrants, which are numbered counterclockwise. The product x;y; is positive for all points in quadrants I and III and negative for all points in quadrants I and IV. Since a positive relationship will have points lying for the most part in quadrants I and III, and a negative relationship will have points lying mostly in the other two quadrants, the sign of 377, x;y; will indicate whether the scatter slopes upward or downward. This sum, however, will tend to increase in absolute terms as more data are added to the sample. Thus, it is better to express the sum in average terms, giving the sample covariance, cov(X, ¥) = (Xi, — XK ~ Py/n (a. The value of the covariance depends on the units in which the variables are mea- sured, Changing one variable from dollars to cents will give a new covariance 100 times the old. To obtain a measure of association that is invariant with respect to units of measurement, the deviations are expressed in standard deviation units. The covariance of the standardized deviations is the correlation coefficient, r namely, y Quadrant tt Quadrant Go) negative (9 positive . OY) : t i i ¥ Quadrant Wt Quadrant 1v (a) positive Gay) negative o x x FIGURE 1.4 Coordinates for scatter diagram for paired variables.8 ECONOMETRIC METHODS r = (\R = = Sayilns.y (1.2) a [> 930 Viet Omitting subscripts and the limits of summation (since there is no ambiguity) and performing some algebraic manipulations give three equivalent expressions for the correlation coefficient—two in terms of deviations and one in terms of the raw data: Say nSxSy Soy = 13 vee Sy ad - aD x¥-(SxVEY) VASP = (XP nD = (DP 1.2.1 The Correlation Coefficient for a Bivariate Frequency Distribution In general. a bivariate distribution such as that shown in Table 1.1 may be represented by the paired values X,, ¥ with frequency nj fori = 1,...,mand j = 1,..., p.X; is the midpoint of the ith subinterval on the X axis, and ¥; the midpoint of the jth subinterval on the ¥ axis. If we use a period for a subscript over which summation has taken place. the marginal frequencies for X are given by m = S?_, mj for i= 1,....m.In conjunction with the X; values these marginal frequencies will yield the standard deviation of X. that is. s,. The marginal frequencies for ¥ are nj = SM, nj for j = 1...., p. Thus. the standard deviation of Y, or sy, may be obtained. Finally the covariance is obtained from a ©ov(X, ¥) = 7S) mij(Xi — RY, — Pyyn (1.4) isis where n is the total number of observations. Putting the three elements together, one may express the correlation coefficient for the bivariate frequency distribution in terms of the raw data as a ——————— (1.5) = ™ z z n> n,X? — (> m,X)P fn > ng¥?- (CL ny¥jP y ist Vir j=l m2 = 2 nS S myXi¥i— (EXE 2X) isij=t j=l 1‘carrer 1; Relationships between Two Variables 9 1.2.2 The Limits of r The correlation coefficient must lie in the range from —1 to +1. To see this, let ¢ be any arbitrary constant. Then \(y— ex)? = 0. Now letc = > xy/ 3) x? . Substi- tution in the inequality gives (S xy? = (3 x°)(S y*), thatis,r? < 1. This expres- sion is one form of the Cauchy-Schwarz inequality. The equality will only hold if each and every y deviation is a constant multiple of the corresponding x deviation. In such a case the observations all lie on a single straight line, with a positive slope ( 1) or a negative slope (r = ~1). Figure 1.5 shows two cases in which r is approxi- mately zero. In one case the observations are scattered over all four quadrants; in the other they lie exactly on a quadratic curve, where positive and negative products off- set one another. Thus, the correlation coefficient measures the degree of linear associ- ation. A low value for r does not rule out the possibility of a strong nonlinear associa- tion, and such an association might give positive or negative values for rif the sample observations happen to be located in particular segments of the nonlinear relation. 1.2.3 Nonsense Correlations and Other Matters Correlation coefficients must be interpreted with care. Many coefficients that are both numerically large and also adjudged statistically significant by tests to be de- scribed later may contain no real information. That statistical significance has been achieved does not necessarily imply that a meaningful and useful relationship has been found. The crucial question is, What has caused the observed covariation? If there is a theory about the joint variation of X and Y, the sign and size of the corre- lation coefficient may lend support to that theory, but if no such theory exists or can be devised, the correlation may be classed as a nonsense correlation. y Peo 7 ¥ Pao eI ~ a a @ o FIGURE 1.5 aired variables for which r? = 0. These10 ECONOMETRIC METHODS Our favorite spurious, or nonsense, correlation was given in a beautiful 1926 pa- per by the statistician G. Udny Yule.> Yule took annual data from 1866 to 1911 for the death rate in England and Wales and for the proportion of all marriages solemnized in the Church of England and found the correlation coefficient to be +0.95. How- ever, no British politician proposed closing down the Church of England to confer immortality on the electorate. More recently, using annual data from 1897 to 1958, Plosser and Schwert have found a correlation coefficient of +0.91 between the log of nominal income in the United States and the log of accumulated sunspots. Hendry has noted a very strong. though somewhat nonlinear, positive relationship between the inflation rate and the accumulation of annual rainfall in the United Kingdom.’ It ‘would be nice if the British could reduce their inflation rate and, as a bonus, enjoy the inestimable side effect of improved weather, but such happy conjunctions are not tobe. In these three examples all of the variables are subject to trend-like movements over time.® Presumably some complex set of medical, economic, and social factors contributed to the reduction in the death rate in England and Wales, even as a differ- ent set of factors produced a decline in the proportion of marriages in the Church of England. Cumulative sunspots and cumulative rainfall necessarily trend upward, as do the U.S. nominal income and the British inflation rate. Series responding to essen- tially unrelated generating mechanisms may thus display contemporaneous upward and/or downward movements and thus yield strong correlation coefficients. Trends may be fitted to such series, as will be shown in the next chapter, and the residuals from such trends calculated. Correlations between pairs of residuals for such series will be negligible. An alternative approach to correlating detrended residuals is to correlate the first differences of the series. The first differences are simply the changes in the series between adjacent observations. They are usually denoted by the prefix A. Thus, AX,=Xi-Xe1 0 AY = Me ¥e Many series that show very high correlations between X and Y (the levels) will show very low correlations between AX and AY (the first differences). This result usually indicates a spurious relationship. On the other hand, if there is a causal relationship between the variables, we expect to find correlations between levels and also between first differences. This point has recently been emphasized in an important paper by Stigler and Sherwin.’ The main thesis of the paper is that if 5G. Udny Yule, “Why Do We Sometimes Get Nonsense Correlations between Time Series?”, Jounal of the Royal Statistical Society. Series A, General, 89, 1926, 1-68. “Charles 1, Plosser and G. William Schwert, “Money, Income, and Sunspots: Measuring Economic Re- lationships and the Effects of Differencing,” Journal of Monetary Economics, 4, 1978, 637-660. "David F. Hendry, “Bconometrics—Alchemy or Science?”, Economica, 47, 1980, 387-406, ‘Trends, like most economic phenomena, are often fragile and transitory. The point has been made in lyrical style by Sir Alec Caimeross, one of Britain's most distinguished economists and a former chief economic adviser t the British government, “A trend is a trend, is @ trend, but the question is, will it ‘bend? Will it alter its course, through some unforeseen force and come to a premature end?” George J. Stigler and Robert A. Sherwin, “The Extent of the Market,” Journal of Law and Economics, 28, 1985, 555-585.cuarrex 1. Relationships between Two Variables 11 two goods or services are in the same market their prices should be closely related. However, since most prices, like many economic series, show trend-like movements over time, Stigler and Sherwin wish to guard against being misled by spurious cor- relation. Thus, in addition to correlating price levels they correlate price changes. As one example, the prices of December 1982 silver futures on the New York Com- modity Exchange and the Chicago Board of Trade over a 30-day trading period gave r = 0,997, and the price changes gave r = 0.956. In Minneapolis, Minnesota, and Kansas City, Missouri, two centers of the flour-milling industry, the monthly whole- sale prices of flour over 1971-1981 gave correlations of 0.97 for levels and 0.92 for first differences. In these two cases the first difference correlations strongly reinforce the levels correlations and support the thesis of a single market for these goods. 1.2.4 A Case Study Gasoline is retailed on the West Coast of the United States by the “majors” (Arco, Shell, Texaco, etc.) and by “minors,” or “independents.” Traditionally the majors have offered a greater variety of products, differentiated in terms of grade of gasoline, method of payment, degree of service, and so forth; whereas the minors have sold for cash and offered a smaller range of products. In the spring of 1983 Arco abolished its credit cards and sold for cash only. By the fall of 1983 the other majors had responded by continuing their credit cards but introducing two prices, a credit price and a lower cash price. Subsequently one of the independents sued Arco under the antitrust laws. The essence of the plaintiff's case was that there were really two separate markets for gasoline, one in which the majors competed with each other, and a second in which the minors competed. They further alleged, though not in this precise language, that Arco was like a shark that had jumped out of the big pool into their litle pool with the intention of gobbling them all up. No one questioned that there was competition within the majors and competition within the minors: the crucial question was whether there was competition between majors and minor The problem was a perfect candidate for the Stiglet/Sherwin type of analysis. ‘The Lundberg Survey reports detailed information twice a month on the prices of all types and grades of gasoline at a very large sample of stations. These data are also averaged for majors and minors. Twelve differentiated products were defined for the majors and four for the minors. This step allowed the calculation of 66 correlation coefficients for all pairs of products within the majors and 6 correlation coefficients within the minors. Each set of coefficients would be expected to consist of very high numbers, reflecting the intensity of competition inside each group. However, it was also possible to calculate 48 correlation coefficients for all cross-pairs of a major price and a minor price. If the plaintiff's argument were correct, these 48 coefficients would be of negligible size. On the other hand, if there were just a single large mar- ket for gasoline, the cross correlations should not be markedly less than correlations within each group. A nice feature of the problem was that the within-group corre- lations provided a standard of reference for the assessment of the cross correlations. In the cases discussed in the Stigler/Sherwin paper only subjective judgments could be made about the size of correlation coefficient required to establish that two goods were in the same market.12. EcoNomETRIC METHODS The preceding approach yielded a matrix of 120 correlation coefficients. In or der to guard against possible spurious correlation, such a matrix was computed for levels, for first differences, for logs of levels, and for first differences of logs (which measure percent changes in price). In addition, regression analysis was used to adjust for possible common influences from the price of crude oil or from general inflation, and matrices were produced for correlations between the residuals from these regres- sions. In all cases the matrices showed “forests” of tall trees (that is, high correlation coefficients), and the trees were just as tall in the rectangle of cross correlations as in the triangles of within correlations. The simple correlation coefficients thus provided conclusive evidence for the existence of a single market for retail gasoline. 13 PROBABILITY MODELS FOR TWO VARIABLES Classical statistical inference is based on the presumption that there exists some population distribution of all possible observations on the variables of interest. That distribution is characterized by certain crucial parameter values. From a sample of n observations sample statistics are computed and these serve as a basis for inference about the population parameters. Ever since the work of Haavelmo in the 1940s the probability approach has been extensively used in econometrics.! Indeed the development of econometrics in the past half century has been driven mainly by the effort to adapt and extend classical inference procedures to deal with the special problems raised by the nature of the data generation process in economics and the general unavailability of controlled economic experiments. 1.3.1 Discrete Bivariate Probability Distribution To introduce some of the main ideas, consider a discrete bivariate probability dis- tribution as shown in Table 1.3. The cell entries indicate the probability of the joint occurrence of the associated X, ¥ values. Thus, pi; = probability that X = X; and ¥ = ¥;. The column and row totals, where a period indicates the subscript over which summation has taken place, give the marginal probabilities for X and Y, re- spectively. There are six important population parameters for the bivariate distribu- tion. The means are defined by Be = EQ) = > piXi and wy = EY) = >opjh (1.6) i 7 ‘The variances are defined as o} = var(X) = EU(X ~ wx) = > p(X - wa? 7 2 (1.7) oF = var(Y) = ET ~ wy => ely 7 'Trygve Haavelmo, The Probability Approach in Econometrics, supplement to Econometrica, 12, July, 1944.cuarren 1; Relationships between Two Variables 13 TABLE 1.3 ‘ A bivariate probability distribution Marginal X x Xm probability ¥ Pu Pa Pt Pa % a Pr Py y, Pin Pip Pap Ps Marginal p, Peo Pe 1 The covariance is 8 0 cov(X, ¥) = ELK - wo - w,)] DD pi - why - wy) (8) 75 Finally, the population correlation coefficient is defined as a Ory com(X,Y) = p (1.9) oxy In these formulae >”; and >” ; indicate summation over the relevant subscripts. Conditional probabilities 7 Consider the X; column in Table 1.3. Each cell probability may be divided by the column total, p;,, to give a conditional probability for ¥ given X;. Thus, a = probability that ¥ = ¥; given that X = X; fa = prob(Y; | X;) ‘The mean of this distribution is the conditional expectation of Y, given X;, that is, My, = EY |X) = > (Ze) (iy +P Similarly, the variance of this distribution is a conditional variance, or oy, = var(Y | X;) = > (2) Byix? (1.12) Vp. The conditional means and variances are both functions of X, so there is a set of m conditional means and variances. In a similar fashion one may use the row probabil- ities to study the conditional distributions of X given ¥.14 BCONoMETRIC METHODS TABLE 1.4 Bivariate distribution of income (X) and vacation expenditure (Y) X ($7000) 20 30 40 1 ae 2 08 1508 y 3 06 06.06 00) 4 0 5 As . 5 | 0 0° © 6 0° 0 3 Marginal probability 7"40™) 30 30 Mean (Y |X) 25 39 Vary X) 44° 851.09 Oat Tt Roto < 14 . TABLE LS. Conditional probabilities from Table 1.4 ¥ aT aX %4 0 ee oe 002 ea eee X % OI—05 02 02 0 0 @ 0 01 02 05 O11 On A numerical example Table 1.4 presents hypothetical data on income and vacation expenditure for an imaginary population. There are just three levels of income and six possible levels of vacation expenditure. Everyone. no matter how humble, gets to spend at least $1,000 on vacation. The marginal probabilities show that 40 percent of this population have incomes of $20,000, 30 percent have incomes of $30,000, and 30 percent have in- comes of 40.000. The conditional probabilities derived from these data are shown in Table 1.5. These conditional probabilities are used to calculate the conditional means and variances shown in the last two rows of Table 1.4. Mean vacation expenditure rises with income but the increase is not linear, being greater for the increase from $30,000 to $40,000 than for the increase from $20,000 to $30,000. The conditional variance also increases with income. One could carry out the parallel analysis for X given Y. This might be of interest to a travel agent concerned with the distribution of income for people with a given vacation expenditure. 1.3.2: The Bivariate Normal Distribution The previous examples have been in terms of discrete variables. For continuous vari- ables the most famous distribution is the bivariate normal. When X and ¥ follow a bivariate normal distribution, the probability density function (pdf) is given bycuarren 1: Relationships between Two Variables 15, f(x.) = ——__ x = 2mo,oy J1- PP 1 XT Mex yo my), (y= mY -5 > |(*—# te (Po By) (2 Be) I of 20 — p>) ( oy if oy ey ae In this equation we have used x and y to indicate the values taken by the variables X and Y. The lower-case letters here do tot measure deviations from sample means, as they do in the discussion of the correlation coefficient in Section 1.2. The range of variation for both variables is from minus to plus infinity. Integrating over y in Eq. (1.13) gives the marginal distribution for X, which is (1.14) Thus, the marginal distribution of X is seen to be normal with mean jx and stan- dard deviation o-,. Likewise, the marginal distribution of Y is normal with mean sy and standard deviation oy. The remaining parameter in Eq. (1.13) is p, which can be shown to be the correlation coefficient between X and Y. Finally, from the joint distribution (Eq. (1.13)] and the marginal distribution [Eq. (1.14)], the conditional distribution of Y given X may be obtained" as FO1D = FW F@ 1 1 (¥— My jl = ex ae (1.15) Qmay\x of 2 ( yx ‘The conditional distribution is also seen to be normal. The conditional mean is Hyx = a + Bx (1.16) where a = pby—By, and (L172) The conditional mean is thus a linear function of the X variable. The conditional variance is invariant with X and is given by Oj, = OL — p?) (1.18) This condition of constant variance is referred to as homoscedasticity. Finally, the conditional mean and variance for X given Y may be obtained by interchanging x and y in the last three formulae. 14 THE TWO-VARIABLE LINEAR REGRESSION MODEL In many bivariate situations the variables are treated in a symmetrical fashion. For the Scottish soldiers of Table 1.1 the conditional distribution of height, given chest “See Problem 1.4.16 ECONOMETRIC METHODS size, is just as meaningful and interesting as the conditional distribution of chest size, given height. These are two aspects of the joint variation, However, in the va- cation expenditure/income example we have already tended to show more interest in the conditional distribution of expenditure, given income, than in the distribution of income, given expenditure. This example is typical of many economic situations, Economists often have explicit notions, derived from theoretical models, of causal- ity running from X, say, to ¥. Thus. the theory of consumer behavior leads one to expect that household income will be a major determinant of household vacation expenditure, but labor economics does not give equal strength to the proposition that household vacation expenditure is a major determinant of household income. Although it is formally true that a joint distribution can always be factored in two different ways into the product of a marginal and a conditional distribution, one fac- torization will often be of more interest to an economist than the other, Thus, in the expenditure/income case the factorization f(X, ¥) = f(X) + f(¥ | X) will be of greater interest than the factorization f(X, ¥) = f(¥)- f(X | ¥). Moreover, in the first factorization the conditional distribution of expenditure, given income, will usually receive much more attention and analysis than the marginal distribution for income. 1.4.1 A Conditional Model To formulate a model for vacation expenditure that is conditional on income, let us consider how data on such variables might be obtained. One possibility is that a sample of households from the N’ households in the population was taken and the values of ¥ and X recorded for the year in question. This is an example of cros section data. There will be some—presumably complex end certainly unknown— bivariate distribution for all N households. This bivariate distribution itself will be some marginalization of a multivariate distribution covering income and all cate~ gories of expenditure, Concentrating on the conditional distribution, economic the- ory would suggest EY |X) = g(X) where g(X) is expected to be an increasing function of X. If the conditional expec- tation is linear in X, as in the case of a bivariate normal distribution, then EW |X) =a + BX (1.19) For the ith household this expectation gives E(Y | X)) = @ + BX, The actual vacation expenditure of the ith household is denoted by ¥;, so we define a discrepancy or disturbance u; as uy = Yj — E(Y | Xi) = Yi a ~ BX; (1.20) "We now return to the earlier convention of using X and ¥ to indicate both the label fora variable and the values that it may assume,ccuarrex |; Relationships between Two Variables 17 ‘The disturbance u; must therefore represent the net influence of everything other than the income of the ith household. These other factors might include such things as the number and ages of household members, accumulated savings, and so forth. Such factors might be measured and included in Eq. (1.19), but with any finite number of explanatory factors we still cannot expect perfect agreement between individual ob- servations and expected values. Thus, the need to specify a disturbance term remains. Taking conditional expectations of both sides of Eq. (1.20) gives E(u; | X;) = 0. The variance of 1; is also seen to be the variance of the conditional distribution, 72, If we look at the jth household, the disturbance u; will have zero expectation and variance Fle; These conditional variances may well vary with income. In the hypo- thetical data of Table 1.4 they are positively associated with income. For the present, however, we will make the homoscedasticity assumption that the disturbance vari- ances are constant and independent of income. Finally, we make the assumption that the disturbances are distributed independently of one another. This rules out such things as “vacation mania,” where everyone rushes off to Europe and large positive disturbances become apparent. This assumption implies that the disturbances are pairwise uncorrelated.'’ Collecting these assumptions together gives Elu;) = 0 for all i var(uj) = E(u?) = a? for all i (1.21) cov(uj, uj) = Eluiuj) = 0 fori # j These assumptions are embodied in the simple statement The uj are iid(0, 0?) (1.22) which reads “the u; are independently and identically distributed with zero mean and variance 0.” Now suppose the available data come in time series form and that X, = aggregate real disposable personal income in year ¢ ¥, = aggregate real vacation expenditure in year t 1,2,...,m. The series {X,} is no longer a set of sample values from the mn of all N incomes in any year: it is the actual sum of all incomes in each "Two variables are said to be independently distributed, or stochastically independent, ifthe conditional distributions are equal (o the corresponding marginal distributions. This statement is equivalent to the joint probabilities being the product of the marginal probabilities. For the discrete case. the covariance between X and ¥ is then covlX ¥) = 0S) p(X — wd - wy) = > ii ~ nd S p AY ws) using Bq. (1.6) =o The converse is not necessarily true since the covariance measures linear association; but substituting 1p = Oin Eg, (1.13) shows that it is true for the bivariate normal distribution, since the bivariate density then collapses into the product of the two marginal densities.18. ECONOMETRIC METHODS year. It might be regarded as a sample of n observations from the “population” of all possible aggregate income numbers, but this interpretation seems to be putting some strain on the meaning of both sample and population. Moreover, the usual time series “sample” consists of data for n adjacent years. We would be rather suspicious of cross-section samples that always consisted only of n adjacent households. They could be from Millionaires’ Row or from Skid Row. Thus, it is difficult to give an unambiguous and useful interpretation of f(X), the marginal distribution of X over time, However, the conditional distribution f(Y | X) is still important and must be given a probabilistic formulation. To see this reasoning, retum to the cross section formulation and introduce the time subscript. Thus, Yin = @ + BXin + Uir (1.23) real vacation expenditure by the ith household in year t real disposable income of the ith household in year r where Yir Xir Making the (implausible) assumption that the a and 8 parameters are the same for all households and aggregating Eq. (1.23) over all N households in the economy, we find promerolEr) which may be rewritten as > Ur ¥, = Na + BX, + U; (1.24) where ¥ and X denote aggregate expenditure and aggregate income and U is an ag- gregate disturbance. The assumptions made about the household u's imply that U; is a stochastic variable with zero mean and variance No”. In the context of time series, one needs to make a further assumption about the independence, or lack thereof, of the U's. If the independence assumption is chosen, then the statement is that the U; are iid(0, No”), 1.4.2. Estimates and Estimators Whether the sample data are of cross section or time series form, the simplest version of the two-variable model is ¥; = a + BX; + uj, with the u; being iid(0, 02). There are thus three parameters to be estimated in the model, namely, a, 8, and o?, The parameters a and f are taken as a pair, since numerical values of both are required to fit a specific line. Once such a line has been fitted, the residuals from that line may be used to form an estimate of 0. An estimator is a formula, method, or recipe for estimating an unknown popu- lation parameter; and an estimate is the numerical value obtained when sample data are substituted in the formula, The first step in fitting a straight line to sample data is to plot the scatter diagram and make sure from visual inspection that the scatter is approximately linear. The treatment of nonlinear scatters is discussed in the next chapter. Let the straight line fitted to the data be denoted by ?; = a + bX;, where ¥; indicates the height of the line at X;. The actual ¥; value will in general deviate from Y;, Many estimators of the pair a,b may be devised.cuarrex |: Relationships between Two Variables 19 1. Fit line by eye and read off the implied values for the intercept a and slope b. Different “artists” may, of course, draw different lines, so it is preferable to have an estimator that will yield the same result for a given data set, irrespective of the investigator. 2. Pass a line through the leftmost point and the rightmost point of the scatter. If Xe denotes the smallest value of X in the sample and X.e the largest and Y., Yee the associated ¥ values, this estimator is b = (Yor — Ye M(Xee — Xe) a = Y,~bXy = Yuu — Xue This estimator can hardly be expected to perform very well since it uses only two of the sample points and ignores the rest. 3. The last criticism may be met by averaging the X and ¥ coordinates of the m left ‘most and the m rightmost points, where m is some integer between I and n/2, and passing a line through the resultant average points. Such an estimator with m set at n/3 or n/2 has been proposed in the literature on errors in variables, as will be discussed later. This type of estimator does not easily lend itself to mathematical manipulation, and some of its properties in repeated applications are difficult to determine, 1.43 Least-Squares Estimators ‘The dominant and powerful estimating principle, which emerged in the early years of the nineteenth century for this and other problems, is that of least squares.'* Let the residuals from any fitted straight line be denoted by = bX; G2 1,2.00 (1.25) From the definition of P; and from Fig. 1.6 these residuals are seen to be measured in the vertical (Y) direction. Each pair of a, b values defines a different line and hence a different set of residuals. The residual sum of squares is thus a function of a and +b. The least squares principle is Select a, b to minimize the residual sum of squares. RSS = Se? = f(a,b) The necessary conditions for a stationary value of RSS are!® See again the unfolding story Press, 1986. 'Sin obtaining the derivatives we leave the summation sign in place and differentiate the typical term. with respect to @ and b in tur, and simply observe the rule that any constant can be moved in front of the summation sign but anything that Varies from one sample point to another must be kept to the right of the summation sign. Finally, we have dropped the subscripts and range of summation since there is ‘no ambiguity. Strictly speaking, one should also distinguish between the a and b values that appear in the expression to be minimized and the specific values that actually do minimize the residual sum of squares, but again there is litte risk of ambiguity and we have kept the expressions uncluttered. Stephen M. Stigler, The History of Statistics, Harvard University20 ECONOMETRIC METHODS y. ~ ° > & x FIGURE 1.6 Residuals from a fitted straight line. 2 92) 2 27-0 bx) = 2D e = 0 (1.26) ade - 3 2° Xe = Simplifying gives the normal equations for the linear regression of ¥ on X. That is, DY =na+b> x (128) Dx =aSxt+o> xX? ‘The reason for the adjective normal will become clear when we discuss the geometry of least squares later. The first normal equation may be rewritten as a =Y-bx (1.29) and = -20X(Y - a- bx) = (1.27) ‘Substituting for a in the second normal equation gives < ee S32 (1.30)cuarrer 1: Relationships between Two Variables 21 Thus, the least-squares slope may first of all be estimated by Eq. (1.30) from the sample deviations, and the intercept then obtained from substituting for b in Eq. (1.29). Notice that these two expressions have exactly the same form as those given in Eq. (1.17) for the intercept and slope of the conditional mean in the bivariate normal distribution. The only difference is that Eqs. (1.29) and (1.30) are in terms of sample statistics, whereas Eq. (1.17) is in terms of population statistics. ‘To summarize, the least-squares line has three important properties. It minimizes the sum of the squared residuals. It passes through the mean point (X, Y), as shown by Eq, (1.29). Finally, the least-squares residuals have zero correlation in the sample with the values of X.'6 The disturbance variance a? cannot be estimated from a sample of u values, since these depend on the unknown a and B values and are thus unobservable. Anes- timate can be based on the calculated residuals (the ¢;). Two possibilities are 5. e?/n or >, e°/(n — 2). For reasons to be explained in Chapter 3 the usual choice is - 22 7 ad (1.31), 1.4.4 Decomposition of the Sum of Squares Using Eqs. (1.25) and (1.29), one may express the residuals in terms of the x, y deviations, namely ei = Yi — bx; ee (1.32) Squaring both sides, followed by summing over the sample observations, gives De = Vy = 2 Say t PS? The residual sum of squares is thus seen to be a quadratic function of b. Since x? = 0, and the equality would only hold in the pathological case of zero variation in the X variable, the single stationary point is necessarily a minimum. Substitution from Eq. (1.30) gives DFP LP +e =b> y+ De . (1.33) Vyr+D This famous decomposition of the sum of squares is usually written as TSS = ESS + RSS Dike = Drs Re = Set De =D using Eq. (1.26) Hence, cov(X,e) = 0 using Eq. (1.27)22. ECONOMETRIC METHODS where!’ TSS RSS ‘otal sum of squared deviations in the Y variable esidual, or unexplained, sum of squares from the regression of ¥ on X ESS = explained sum of squares from the regression of ¥ on X The last line of Eq. (1.33) may be rearranged to give 2 RSS _ ESS Sloss. Tss ~ TSS ‘Thus, 7? may be interpreted as the proportion of the Y variation attributable to the linear regression on X. Equation (1.34) provides an alternative demonstration that the limits of r are +1 and that in the limiting case the sample points all lie on a single straight line. (1.34) 1.45 A Numerical Example Table 1.6 gives some simple data to illustrate the application of these formulae. Sub- stitution in Eq. (1.28) then gives the normal equations 40 = 5a+20b 230 = 20a + 1206 with solution P=1+175X The same data in deviation form are shown in Table 1.7. The regression coefficients may be obtained from Say _ 70 b= S45 = 7175 and a=Y~bxX =8-1.75(4) =1 The explained sum of squares may be calculated as ESS = bS xy = 1.75(70) = 122.5 and the residual sum of squares is given by subtraction as RSS = TSS ~ ESS = 124 ~ 122.5 = 1.5 Finally, the proportion of the ¥ variation explained by the linear regression is, ESS _ 122.5 P= FSS > pap 7 0.9879 "TUnfortunately there is no uniform notation for sums of squares. Some authors use SSR to indicate the sum of squares due to the regression (our ESS), and SSE to indicate the sum of squares due to error (our RSS),carrer 1; Relationships between Two Variables 23 TABLE 1.6 xX oy x wf 204 8 4 450 -050 0 -1 307 2 9 625) 07S 2.25 eee ee 273 75) 0.25 Seo 525) 975) 075) 375) 9 17 153 8116.75 0252.25 Sums 20 40230-12040 0 0 TABLE 1.7 ay 5 xe 8 16-350 -050 1.00 1 1 el) 0.75, 0.75) 15 2-525 0.25 -0.75 14 ot 1 1475 -0.75 -0.75 5.9 45 Gl ea ee Sums 0 0 70 40 124 0 0 0 15 INFERENCE IN THE TWO-VARIABLE, LEAST-SQUARES MODEL The least-squares (LS) estimators of a and B have been defined in Eqs. (1.28) to (1.30). There are now two important questions: 1. What are the properties of these estimators? 2. How may these estimators be used to make inferences about a and 8? 1.5.1 Properties of LS Estimators The answers to both questions depend on the sampling distribution of the LS es- timators. A sampling distribution describes the behavior of the estimator(s) in re- peated applications of the estimating formulae. A given sample yields a specific ‘numerical estimate. Another sample from the same population will yield another numerical estimate. A sampling distribution describes the results that will be ob- tained for the estimator(s) over the potentially infinite set of samples that may be drawn from the population. The parameters of interest are a, 8, and o? of the conditional distribution, ‘f(Y | X). In that conditional distribution the only source of variation from one hy- pothetical sample to another is variation in the stochastic disturbance (u), which in conjunction with the given X values will determine the ¥ values and hence the sam- ple values of a, b, and s®, Analyzing ¥ conditional on X thus treats the X1, Xp...» X, values as fixed in repeated sampling. This treatment rests on the implicit assump-24 ECONOMETRIC METHODS tion that the marginal distribution for X, thatis, f(X), does not involve the parameters of interest or, in other words, that f(X) contains no information on @, 8, and 0. This is called the fixed regressor case, or the case of nonstochastic X. From Eq. (1.30) the LS slope may be written b= Swi where the weights w; are given by xi a 3. =x (1.35) These weights are fixed in repeated sampling and have the following properties: we = 0 we and wix; = > wiX; = 1 (1.36) Day It then follows that b= Swi (137) so that the LS slope is a linear combination of the ¥ values. ‘The sampling distribution of b is derived from Eq, (1.37) by substituting ¥; = @ + BX; + u; and using the stochastic properties of u to determine the stochastic properties of b. Thus, : a(S) mi) +B (Swix) + wi b = B+ Swans (1.38) and so E(b) = B (1.39) that is, the LS slope is an unbiased estimator of 8. From Eq. (1.38) the variance of Bis seen to be var(b) = El(b ~ BY?) = £[(S mau) From the properties of the w’s it may be shown! that 2 o vantb) = So (1.40) By similar methods it may beshown!? that E(a)=a (1.41) 2 and var(a) = o? [4+ * (1.42) nt SH These four formulae give the means and variances of the marginal distributions of and b. The two estimators, however, are in general not stochastically independent, "See Appendix 1.1 "See Appendix 1.2.chapter |: Relationships between Two Variables 25 for the covariance is*® cov(a, b) = (1.43) This covariance only vanishes if X = 0. One can always rearrange an LS regression to have a zero mean for the right-hand-side variable. By using Eq. (1.29), ¥ = a + bX + e can be rewritten as Y = ¥ +bx+e, which gives cov(¥, b) = cov(i,b) = 0. 1.5.2 Gauss-Markoy Theorem The LS estimators are seen to be linear combinations of the Y variable and hence linear combinations of the stochastic u variable. Because they are also unbiased, they belong to the class of linear unbiased estimators. Their great importance in the theory and practice of statistics is that their sampling variances are the smallest that can be achieved by any linear unbiased estimator. Looking at estimators of B, for example, let b= Sal; denote any arbitrary linear unbiased estimator of B. The unbiasedness criterion im- poses two linear constraints on the weights, (c;), leaving (n ~ 2) weights “free.” It can be shown?! that var(b*) = var(b) + 0? (ci — wi? Since S(c; — w;)? = 0, var(b*) = var(b). Equality only holds when ¢; = w; for all i, that is, when b* = b. The least-squares estimator thus has minimum variance in the class of linear unbiased estimators and is said to be a best linear unbiased estimator, or BLUE. 1.5.3 Inference Procedures The results established so far have required the assumption that the u; are iid(0, 0). The derivation of inference procedures requires a further assumption about the form of the probability distribution of the u's. The standard assumption is that of normal- ity, which may be justified by appeal to the Central Limit Theorem, since the u's represent the net effect of many separate but unmeasured influences. Linear combi- nations of normal variables are themselves normally distributed. Thus. the sampling distribution of a,b is bivariate normal, as in the formula in Eq, (1.13), The marginal distributions are therefore also normal and are determined by the means and vari- ances already obtained. Thus, b~NB.071>. 2) (1.44) See Appendix 1.3, See Appendix 1.4.26 ECONOMETRIC METHODS to be read, “b is normally distributed with mean and variance 0/5: x2.” The square root of this variance, that is, the standard deviation of the sampling distribu- tion, is often referred to as the standard error of b and denoted by s.e.(b). The sampling distribution of the intercept term is a~nfaor(h+ 25) (1.45) If o? were known, these results could be put to practical use. For example, a 95 percent confidence interval for would be provided by b+ 1.9601) 2 Italso follows from Eq, (1.44) that b-B = ——= ~ NO) 1.46) 7 sa ey where (0, 1) denotes the standard normal distribution (a normally distributed vari- able with zero mean and unit variance). Thus, a test of the hypothesis Ho: B = Bo is carried out by computing and contrasting this statistic with a preselected critical value from the standard nor- mal distribution. If, for example, the absolute value of this statistic exceeded 1.96, Ho would be rejected at the 5 percent level of significance. When o? is unknown these procedures are not feasible. Te derive an operational procedure we need two further results. They will be stated here and proved for the general case of multiple regression in Chapter 3. The relevant results are = ~ x(n - 2) (1.47) to be read “S. e*/o? is distributed as x? with (n — 2) degrees of freedom,” and De? is distributed independently of f(a, b) (1.48) As shown in Appendix B the f distribution is defined as a combination of a standard normal variable and an independent y” variable. Thus, Eqs. (1.46) through (1.48) give Die: sl JS where s? = > e7/(n — 2), the estimator of o? defined in Eq. (1.31). Notice that Eq. (1.49) has the same structure as Eq. (1.46), the only difference being that the un- known ois replaced by the estimate s. This causes a shift from the normal distribu- tion to the distribution. For degrees of freedom in excess of about 30, the differences between the critical values of the f distribution and the standard normal distribution are negligible. A 95 percent confidence interval for B is ~X(n-2) (1.49)y cuarreR t: Relationships between Two Variables 27 b* tors! > (1.50) and Ho: B = Bo would be rejected if b- Bo FO) > tooas(n - 2) 151 vise foors(n ~ 2) (51) where fno2s(1t — 2) indicates the 2.5 percent point of the ¢ distribution with (n — 2) degrees of freedom. The conditions in Eqs. (1.50) and (1.51) are opposite sides of the same coin. If Eq. (1.51) leads to a rejection of Ho, then Bp lies outside the confidence interval given in Eq. (1.50). Likewise, if By lies inside the 95 percent confidence interval, Eq. (1.51) will not lead to the rejection of Ho at the 5 percent level. The most commonly used test of significance is that of Ho: B = 0, The test statistic is then b b "TES ew (1.52) and Hy would be rejected at the 5 percent level of significance if the absolute value of b exceeded fo 2s times its standard error. The abbreviation s.c.(b) is used to de- note both the true and the estimated standard error of f.. The test statistic in Eq. (1.52) is a routine output of most regression packages and usually has some label such as T-STAT. Many programs also report a P-value, which is the probability of obtaining a coefficient as far or farther from zero as the sample value if, in fact, the true value of the coefficient is zero. This number may be labeled P- VALUE or 2-TAIL SIG. By a similar development, tests on the intercept are based on the f distribution: a-a s/n + X77 xt Thus, a 100(1 ~ €) percent confidence interval for a is given by a+ tens /Un + 27>" 2 (1.54) and the hypothesis Hy:a = ag would be rejected at the 100€ percent level of sig- nificance if Pees (1.53) a-ay | | > kn s Jin + XY 2 Tests on o? may be derived from the result stated in Eq. (1.47). Using that result one may, for example, write =2)s% prob abs < one < xm} = 0.95 (1.55) ‘nich states that 95 percent of the values of a x? variable will lie between the values (teat cut off 2.5 percent in each tail of the distribution, The critical values are read off from the x? distribution with (n — 2) degrees of freedom, or accessed through any appropriate software package. The only unknown in Eq. (1.55) is 02, and the ‘28 ECONOMETRIC METHODS contents of the probability statement may be rearranged to give a 95 percent confi- dence interval for o? as ( De n= Xbors Xb.oxs 1.5.4 Numerical Example (Continued from Section 1.4.5) From the data in Tables 1.6 and 1.7 we have already calculated n=5 a=1 b=175 TSS = 124 ESS=1225 RSS=15 7 = 0.9879 We now obtain 8 = RSSi(n — 2) = 1.5/3 = 0.5 var(b) = 3° 0.5/40 = 0.0125 vant = 05 (3 + a) 0.3 40 The estimated standard errors of the regression coefficients are thus se(a) = J0.3 = 0.5477 s.e.(b) = 0.0125 = 0.1118 A preselected critical value from the # distribution with 3 degrees of freedom is 4.025 = 3.182. Thus, a 95 percent confidence interval for a is 1 + 3.1820.5477) that is, -074 to 2.74 and a 95 percent confidence interval for B is 7 1,75 + 3.182(0.1118) that is, 139 t © 211 The intercept is not significantly different from zero since a 1 sela) 03477 whereas the slope is strongly significant since _b _ 175 se(b) O1118 As indicated earlier, once confidence intervals have been computed, actually com- puting the significance tests is unnecessary, since a confidence interval that includes zero is equivalent to accepting the hypothesis that the true value of the parameter is, = 1.826 < 3.182 = 15.653 > 3.182a cnarrer 1: Relationships between Two Variables 29 zero, and an interval that does not embrace zero is equivalent to rejecting the null hypothesis. From the x distribution with 3 degrees of freedom x} 995 = 0.216 and X}.o75 = 9.35. We also have > e = 1.5, Thus, a 95 percent confidence interval for o? is AS 15 935 0.216 that is, 0.16 to 634 1.6 ANALYSIS OF VARIANCE IN THE TWO-VARIABLE REGRESSION MODEL The test for the significance of X, (Ho: B = 0), derived in the previous section may also be set out in an analysis of variance framework, and this alternative approach will be especially helpful later when we treat problems of multiple regression. We have seen in Eq, (1.46) that the ratio of (b — 8) to the true standard error of b is a standard normal variable. From the definition of the x? variable in Appendix B it follows that (b= BY asa xO It has also been stated in Eq, (1.47) that ye Gr 7-2) and that this statistic is distributed independently of b. Recalling from Appendix B thatthe ratio of two independent y* variables, each divided by the associated degrees of freedom, follows the F distribution, we then have pe Gm BY Se Den — 2) As with the shift from the normal to the ¢ distribution in the preceding section. this development has the felicitous result that the unknown a” disappears from the ex- pression for F. To test the hypothesis Ho: B = 0, we make this substitution in Eq. (1.56), giving ~ F(,n- 2) (1.56) _ PS? ~ Teln-D By referring to the decomposition of the sum of squares in Eq. (1.33), the F statistic in Eq. (1.57) is seen to be F ~ F(l,n—2) : (57) _ __ESS/i ~ RSSKin = 2) (1.58)30 ECONOMETRIC METHODS TABLE 1.8 ANOVA for two-variable regression Source of variation Sums of squares Degrees of freedom Mean squares @ ® @ x 1 ESsiI Residual (1-2) RSSi(n ~ 2) Total (m=) Following this approach, we can set out the data in an analysis of variance (ANOVA) table (Table 1.8). The entries in columns 2 and 3 of the table are additive. The mean squares in the final column are obtained by dividing the sum of squares in each row by the corresponding number of degrees of freedom. An intuitive explanation of the degrees of freedom concept is that it is equal to the number of values that may be set arbitrarily. Thus. we may set n ~ I values of y at will, but the nth is then determined by the condition that 5° ewise, we may set n ~ 2 values of e at will, but the least-squares fit imposes two conditions on e, namely, Se = > Xe = 0, and finally there is only 1 degree of freedom attached to the explained sum of squares since that depends only on a single parameter 6. The F statistic in Eq. (1.58) is seen to be the ratio of the mean square due to X to the residual mean square. The latter may be regarded as a measure of the “noise” in the system, and thus an X effect is only detected if it is greater than the inherent noise level. The significance of X is thus tested by examining whether the sample F exceeds the appropriate critical value of F taken from the upper tail of the F dis- tribution. The test procedure is then as follows: Reject Ho: 8 = 0 at the 5 percent level of significance if, ESS/I P= RSsin—D > Fogs(l, n — 2) where Fo9s indicates the value of F such that just 5 percent of the distribution lies to the right of the ordinate at Fos. Other levels of significance may be handled ina similar fashion. For the simple numerical example from Tables 1.6 and 1.7 Sample F = 122.5/0.5 = 245.0 and Foos(1, 3) = 10.1. Thus, we reject Ho: B = 0, as before. We now have two ways of testing the significance of X, one based on a statistic that follows a 1 distri- bution, and another based on a statistic with an F distribution. The tests, however, are identical since, as shown in Appendix B, P(m) = F(1,m) (1.59) A little arithmetic shows that this relation holds for both sample statistics and critical values in the foregoing example. There is also a third version of this test, which is sometimes used. By using the decomposition of the sum of squares in Eq. (1.33), the F statistic in Eq. (1.58) may be writtencuarter 1: Relationships between Two Variables 31 P(n-2) a=?) Taking the square root will give a t statistic, ria=d va=P) Either statistic may be used and referred to the appropriate F or ¢ distribution to test the significance of X. Thus, we may base a test on the correlation coefficient, on the regression slope, or on the decomposition of the sum of squares; but all three approaches give the same answer to the single question: Does X play a statistically significant role in the explanation of Y? F (1.60) (61) 17 PREDICTION IN THE TWO-VARIABLE REGRESSION MODEL After having estimated a regression line from the sample of n observations, our in- terest often centers on some specific value Xo of the regressor variable and we are required to predict the value Yo likely to be associated with Xo. For instance, if Y denotes the consumption of gasoline and X the price, we might be interested in pre- dicting the demand for gasoline at some future higher price. The value of Xp may lie within the range of sample X values or, more frequently, we may be concerned with predicting ¥ for a value of X outside the sample observations, In either case the prediction involves the assumption that the relationship presumed to have generated the sample data still holds for the new observation, whether it relates to a future time period or to a unit that was not included in a sample cross section. Alternatively, ‘we may have a new observation (Xo, Yo), where the actual ¥ value is known, and the question arises whether this observation may be presumed to have come from the same population as the sample data. For example, does the introduction of a speed limit reduce the demand for gasoline, conditional on the price being charged? Prediction theory enables us to address both questions. We may make two kinds of predictions, a point prediction or an interval prediction, in just the same way as we can give a point estimate or an interval estimate of a parameter 8. But in practice, a point estimate is of little use without some indication of its precision, so one should always provide an estimate of the prediction error. The point prediction is given by the regression value corresponding to Xo, that is, Po = a + bXy = P + bxo (1.62) where xp = Xo — X. The true value of ¥ for the prediction period or observation is Yo = a + BXo + uo The average value of Y taken over the n sample observations is Ysa+pX+a x32. BCONOMETRIC METHODS Subtracting gives P+ Brot u-& (1.63) The prediction error is defined as 0 = Yo~ Yo = ~(b~ B)x0 + uo — (1.64) Clearly the expected prediction error is zero, so Po is a linear unbiased predictor of Yo. The variance of ey may be shown”? to be 1, var(eo) = 07 ( a se) (1.65) It will be shown in Chapter 3 that the error variance in Eq. (1.65) is a minimum in the class of linear unbiased predictors. Thus, the optimum property of LS estimates of coefficients carries over to prediction based on the LS regression. From Eq. (1.64) we see that eo is a linear combination of normally distributed variables (b, uo, and @). Thus, it is also normally distributed, and so | oft + Unt S02 Replacing the unknown o” by its estimate s? then gives Yo — Po s JTF In + (Xo — XPS Everything in Eg. (1.66) is known except Yo, and so, in the usual way, we derive a 95 percent confidence interval for Yo as N(,1) se (1.66) (1.67) (a + Xo) = tooss + : Even if we know a and 6 with certainty, there is an inherent element of uncer- tainty in predicting Yo, owing to the random drawing uo that occurs in the prediction period. Thus, interest often centers more realistically on the prediction of the mean value of Yo, that is, E(Yo) = a + BXo This eliminates the term up from the prediction error. Following through the same analysis gives a 95 percent confidence interval for E(¥o) as Co - 29 =e ‘The width of the confidence interval in both Eqs. (1.67) and (1.68) is seen to increase symmetrically the further Xo is from the sample mean X. 1 (@ + Ko) * toorss 7 (1.68) See Appendix 15.ccuapter |: Relationships between Two Variables 33 If we continue the previous numerical example, a 95 percent confidence interval for ¥ conditional on X = 10is [10-4 1 d0-4F 1+ 1.75(10) = 3.182 V0.5) + 3 which is 15.24 t0 21.76 The 95 percent interval for E(Y | X = 10) is _ ras 3.82/05! + besa which is 16.14 to 2086 To test whether a new observation (Xo, Yo) comes from the structure generating the sample data, one contrasts the observation with the contidence interval for Yo. For example, the observation (10, 25) gives a ¥ value that lies outside the interval 15.24 to 21,76; and one would reject at the 5 percent level the hypothesis that it came from the same structure as the sample data, 18 GASOLINE CONSUMPTION: A PRELIMINARY ANALYSIS This preliminary look at gasoline consumption cannot be expected to be economi- cally realistic, since we are currently restricted to using just one explanatory vari- able. One does not need to be a Nobel laureate in economics to suspect that price and income both influence consumption. The main purpose is to illustrate the vari- ous descriptive and test statistics in a typical computer printout Consider first the price and gasoline consumption (GAS) scatter for the period 1959.1 to 1973.3, shown in Fig. 1.35. Fitting a linear regression to these series gives the results in Table 1.9.3 The dependent variable and the sample time span are indicated at the top of the table. Then follows the estimated intercept (coefficient on the C variable) and the estimated coefficient on the regressor PRICE. or (X2). with their standard errors and the ¢ statistics for testing the hypothesis that the true value of each coefficient is zero, R-squared is the statistic defined in Eq. (1.34). Adjusted R-squared will be explained in Chapter 3. S.E. of regression is the square root of the statistic defined in Eq. (1.31), and Sum squared resid is the residual sum of squares (RSS). The log likelihood will be discussed in Chapter 2. and the Durbin-Watson statistic in Chapter 6. The Akaike and Schwarz information criteria will be discussed in Chapter 3. Finally the F-statistic, defined in three equivalent forms in Eqs. (1.56), (1.57), and "This regression output, like most of the empirical results in Chapters 1 to 9, comes from EViews, a ‘Windows software program from Quantitative Micro Software, Irvine, California34 ECONOMETRIC METHODS TABLE 1.9 Regression of gasoline consumption on price, 1959.1 to 1973.3, SL LS 1! Dependent Variable is GAS Sample: 1951:1 1973:3 Included observations: 59 Variable Coefficient Std. Error T-Statistic. Prob. ic 2.121645 0.548643 3.867078 0.0003 x2 -2150563 0.118430 -18.15899 0,000 R-squared 0.852618 | Mean dependent var —7.840375 Adjusted R-squared 0.850032 SD. dependent var 0.136145 S.E. of regression 0.052723 Akaike info criterion —5.852095 Sum squared resid 0.158444 Schwarz criterion 5.781670 Log likelihood 9091943 F-statistic 329.7489 Durbin-Watson stat 0.290306 —_-Prob(F-statistic) 0.000000 (1.58), tests the significance of the overall regression, which in the two-variable case is equivalent to testing the significance of the single regressor. The numerical value of the F statistic is the square of the r statistic on the regressor, as may be checked from the regression output. The Prob(F-statistic) is the P-value that tests the null hypothesis of no relationship between the two variables, and this hypothesis is seen to be decisively rejected. A similar regression for the period 1982.1 to 1992.1, for which the scatter is shown in Fig. 1.3d. gives the results shown in Table 1.10. The fit is much less good than in the earlier period, and the price coefficient is numerically much smaller. However, these two-variable regressions have no economic significance. A proper analysis requires a multivariate approach with attention to the dynamics of the de- ‘mand function, and this will be attempted in Chapter 8. TABLE 1.10 Regression of gasoline consumption on price, 1982.1 to 1992.1 LLL LS // Dependent Variable is GAS Sample: 1982:1 1992:1 Included observations: 41 Variable Coefficient Std. Error T-Statistic Prob. ic 7.055535 0.091918 -76.75925 0.0000 X2 0.137704 0.019344. =7.118512 0,000 R-squared 0.565088 Mean dependent var —7.709410 Adjusted R-squared 0.553936 —_S.D. dependent var 0.032420 S.E. of regression 0.021653 Akaike info criterion ~7.617707 ‘Sum squared resid 0.018285. Schwarz criterion =1,534118 Log likelihood 9.98651 F-statistic 50.67324 Durbin-Watson stat 0.669097 __Prob(F-statistic) 0.000000ccuaprex 1; Relationships between Two Variables 35 APPENDIX. APPENDIX 1.1 To prove var(6) = o/Sx? As shown in the text var(b) = E (= win where the w; and their properties are defined in Egs. (1.35) and (1.36). Expanding the right-hand side (Se) (ov tyr wrtta 2+ + Wally)? = wha + wud +b whe + Qi juiuy + where all the remaining terms are cross products of the form wrwjuiu; (i < j). Taking expectations, we find elm] tw oe “Se by virtue of Eqs. (1.21) and (1.36). APPENDIX 1.2 To derive the mean and variance of the sampling distribution of a From Eq. (1.29) a= P- oR =a+ BX +a- bX =a-(b-B)X+a Since E(b) = B and E(i) = 0, it follows that E(a)=a@ Then var(a) = El(a - @)*] = SPE[(b — BY] + Ela?) ~ 2XE(b - Aya) Now Elb- BY) = 07/236 ECONoMETRIC METHODS and E[@] = 07 ln since @ is the mean of a random sample of n drawings from the u distribution, which has zero mean and variance o?. Finally, El(b - B)al] = Elem) Du) = (t > wind + cross-product terms in wu] Thus, APPENDIX 1.3 To derive cov(a, b) cov(a,b) = El(a ~ a\(b ~ B)) = Ella ~ (b- B)XXb - B)) = El(b — B)ii) — XE((b - BY] APPENDIX 1.4 Gauss—Markov theorem A linear estimator of B is b = 3 ci¥i, where the cj are to be determined. Un- biasedness requires E(b*) = B. Now B= Sela + BX, + us) (Sa)+e(Sax)+ Nom ‘Thus, b* will be a linear unbiased estimator if and only if Da=0 and Saki = Sax =1 When these conditions are satisfied b= B+ > cui 2 and var(b") = li cu) | =0°S To compare this variance with that of the least squares b, write cy = wi + (ci — wi)CHAPTER 1: Relationships between Two Variables 37 Thus, Did = Dov} + Si — wi? + 2 wile; - wd) 7 The properties of the w; and the conditions on the c; ensure that DS wici - wi) = 0 and so var(b") = var(b) + 6? “(ci — wi? as given in the text, which proves the theorem. APPENDIX 1.5 To derive var(eo) From Eq. (1.64) e9 = —(b— B)xy + ug - 7 ‘Square both sides and take expectations. The expectations of all three cross-product terms vanish. The independence of the u's means that uo is uncorrelated with @ and also with b, which is a function of 1; to up, and we have seen in 1.3 above that E{(b — B)i) = 0. Thus var(e) = E(ui) + El] + xZEL(b — BY] af.t, 3 cole ts st] PROBLEMS 1.1., How might the volatility of the savings and income series in Fig. 1.1a be measured? A possible measure is the coefficient of variation, which is the standard devia- tion of a series expressed as a percentage of the mean, From your software package compute the coefficient of variation for each series. Also compute the savings ratio (savings/income) and comment on its movement over time. 1.2. Verify the conditional means in Table 1.2, 1.3. Verify the equivalence of the three expressions for the correlation coefficient in Eq (3), 1.4, Verify the results for the conditional normal distribution in Eqs. (1.15), (1.16), and (1.17) by taking the ratio of Eq. (1.13) to Eq. (1.14). 1.5. Compute the correlation coefficients for the various scatter plots in Fig. 1.3 and com- ment on your results, 1.6, Carry out regression analyses for the data on gasoline consumption and price (illus- trated in Fig. 1.3) using any subperiods you consider appropriate and comment on your results.38 ECONOMETRIC METHODS 1.7. A container holds three balls numbered 1, 2, and 3. A ball is drawn at random and the number (X) noted. A second ball is drawn at random and the number (¥) noted. Construct tables showing the bivariate distribution f(X, Y) when the drawing is (@) with replacement (6) without replacement In each case compute the conditional means and correlation coefficient. 1.8. Three discrete variables, X, Y, and Z have possible values: get) tee 248 ‘The trivariate distribution is shown in the accompanying table. Construct the bivariate distribution f(X, ¥) obtained by “integrating out” Z from the trivariate distribution. In this case the bivariate distribution is itself a marginal distribution, the result of a ‘marginalization with respect to Z. x a eo 3 Y 202 °5 22 5 Zz 4 8 4 4.8 8 LK%2Z) 20 1 10 q 1.9, This problem requires the derivation of a sampling distribution from first principles. The postulated relation is Y= BXi+u; with =1,2 and X= 1%) =2 The sample size is 2, and the bivariate probability distribution for u is given in the accompanying table. uw -2 2525 25 25 Verify that the u's have zero mean and zero covariance. Derive the sampling dis- tribution of the estimator b= SxS X? Show that the expected value of the estimator is 8, and compute the variance of the sampling distribution. (This problem comes from “A Simple Approach to Teaching Generalized Least Squares Theory,” by E. H. Oksanen, The American Statistician, 48, August 1991, 229-233.) 1.10. The fixed values of X in a problem are as follows:Lu. 1.12. 1.13. 114, cuaprer |; Relationships between Two Variables 39 ‘An econometrician with no calculator and an aversion to arithmetic proposes to estimate the slope of the linear relation between Y and X by e+ ¥s-%-%) Show that this estimator is unbiased. Deduce its sampling variance and compare this with the sampling variance of the least-squares estimator of the slope. ‘We know that the least-squares estimator has minimum variance in the class of linear unbiased estimators. The ratio of the two variances, with that of the LS estima- tor in the numerator, defines the efficiency of the alternative estimator. Compute the eficiency of the lazy econometrician’s estimator. Let us consider the “other” or “reverse” regression. The process of defining residuals in the X direction and minimizing the new RSS gives the regression of X on ¥. There are thus at least two possible regressions. but only one r? can be computed for a given data set. What is the interpretation of rin the regression of X on ¥? Prove that eBay where the b’s are the LS slopes in the respective regressions. Hence, show that by = Vow (provided both slopes are positive) and that, viewed in the X,Y plane with ¥ on the vertical axis and X on the horizontal. the regression of ¥ on X will have a smaller slope than the regression of X on ¥. What can you say if > ~~ slopes are negative? Notice that in any specific example, both slopes must have the same sign. since the sign of the slope is given by the sign of the sample covariance From a sample of 200 observations the following quantities were calculated: Dx=u34 Sy = 207 Dk s6 Sy = 3496 Noxy = 2213 Estimate both regression equations, compute 7°, and confirm the foregoing statements. ‘Show that if r is the correlation coefficient between n pairs of variables (X,, ¥,), then the squared correlation between the n pairs (aX; + b, c¥, +d), where a, b, c, and dare constants, is also r?, Data on aggregate income ¥ and aggregate consumption C yield the follo sions, expressed in deviation form: 1g regres 1.2c 0.6y If ¥ is identically equal to C + Z, where Z is aggregate saving. compute the correla- tion between Y and Z, the correlation between C and Z, and the ratio of the standard deviations of Z and ¥. d é The accompanying table gives the means and standard deviations of two variables X and ¥ and the correlation between them for each of two samples. Calculate the corre- lation between X and ¥ for the composite sample consisting of the two samples taken together. Why is this correlation smaller than either of the correlations in the subsam- ples?40. ECONOMETRIC METHODS Number in 8 Sample sample X Ys 1 C00: geet . 40027 103 1.15. An investigator is interested in the accompanying two series for 1935-1946. _ Year 353637 «38 «30 40 41 4243444546 X,deaths ofchildren 60 62 61 55 53 60 63 53 52 48 49 43 ‘under I year (000) Y, consumption of 23-«-23-« «25-25 26 H_—«2D:—s3HH—s3 3233. beer (bulk barrels) (a) Calculate the coefficient of correlation between X and Y. (6) A linear time trend may be fitted to X (or Y) by calculating an LS regression of X (or Y) on time 1. This process requires choosing an origin and a unit of measurement for the time variable. For example, if the origin is set at mid-1935 and the unit of measurement is | year, then the year 1942 corresponds to t = 7, and so forth for the other years. If the origin is set at end-1940 (beginning of 1941) and the unit ‘of measurement is 6 months, then 1937 corresponds to t Show that any computed trend value X, = a + bt is unaffected by the choice of origin and unit of measurement. (c) Let ex, and e,,, denote the residuals of X and Y from their trend values. Calcu- late the correlation coefficient between e,., and ¢,,,. Compare this value with that obtained in part (a), and comment on the difference. 1.16., A sample of 20 observations corresponding to the model Y=a+pxeu where the u’s are normally and independently distributed with zero mean and constant variance, gave the following data: Dy=29 Sw-Pfr= 869 Sx- RHy- Ff) = 1064 DX = 186.2 Sx - Xf = 2154 Estimate a and B and calculate their standard errors. Estimate the conditional mean value of ¥ corresponding to X = 10 and find a 95 percent confidence interval for this. mean,CHAPTER 2 Further Aspects of Two- Variable Relationships Coapter 1 presented a set of inference procedures associated with least-squares (LS) estimators in the context of bivariate relationships. The derivation of these proce- dures was based on two crucial assumptions, one about the form of the conditional expectation E(¥ | X) and the other about the stochastic properties of the disturbance term u. The specific assumptions were EY |X) = a+ BX ap and . E(u) = 0 for all 7 EQ) =o? for alli (2.2) Etujuj)=0 fori # j Tt also follows from Eq, (2.2) and the fixed regressor assumption that E(Xiu)) =XiE(uj) =O — forall, j (2.3) Adding the assumption of normality to Eq. (2.2) gives ‘The uw; are iid N(0,07) (4) which reads, “The wu; are independently and identically distributed normal variables with zero mean and variance o?.” The validity of the inference procedures obviously depends on the correctness of the underpinning assumptions. Most of this chapter deals with various possible respecifications of the condi- tional expectation assumption in Eq. (2.1). We will first look at some of the issues raised when the regressor (explanatory) variable is time. This leads naturally to the consideration of constant growth curves, where the logarithm: of the depen- dent variable is expressed as a linear function of time, We then consider cases where transformations of the dependent and/or explanatory variable may be useful. ‘Many relationships that are nonlinear in the original variables may be linearized by 4L42. ECONOMETRIC METHODS suitable transformations. In such cases the simple techniques developed in Chapter 1 for the linear model may be applied to the transformed variables. Next we consider the bivariate model where the explanatory variable is sim- ply the lagged value of the dependent variable. This is the first-order, autoregres- sive AR(1) scheme. The change seems innocuous enough but it moves us into fun- damentally new territory. The least-squares estimators are no longer unbiased; and the exact, finite sample results of Chapter I are no longer strictly valid. The least- squares procedures of Chapter | can still be applied to the autoregressive equation but they now have only a large-sample, or asymptotic validity. An appreciation of this result requires an introduction to some basic and very important ideas relating to large-sample theory, namely, asymptotic distributions, asymptotic efficiency, and consistency. The simple autoregressive equation also raises the issue of the sta- tionarity of a time series. These issues will be developed extensively in later chap- ters, especially Chapters 5, 7. and 8. We hope their introduction in the context of a two-variable relation will Keep the initial exposition as simple as possible and serve as a bridge to the later treatment. 21 TIME AS A REGRESSOR Ina time series plot. as shown in Chapter 1, a variable ¥; on the vertical axis is plotted against time on the horizontal axis. This may also be regarded as a scatter plot, the only difference from the conventional scatter plot being that the X (time) variable in- creases monotonically by one unit with each observation, Many economic variables increase or decrease with time. A linear trend relationship would be modeled as Y=a+prtu (2.5) where T indicates time. The 7 variable may be specified in many fashions, but each specification requires one to define the origin from which time is measured and the unit of measurement that is used. For example, if we had annual observations on some variable for the (a = 13) years from 1980 to 1992, possible specifications of the T variable would be T = 1980, 1981, 1982,..., 1992 T = 1,2,3,...,13 T = ~6,-5,-4,....6 Inalll three cases the unit of measurement is a year. The origins are, respectively, the start of the Gregorian calendar, 1979, and 1986. The third scheme is advantageous for small-scale calculations since in this case T has zero mean, so the normal equations for fitting Eq. (2.5) simplify to a=F and b= ST/ST? Many software packages will generate a TREND variable for use in regression anal- ysis. This is the second specification for T above.‘cuApter2: Further Aspects of Two-Variable Relationships 43 2.1.1 Constant Growth Curves ‘Taking first differences of Eq. (2.5) gives AY, = B + (u)~ 4-1) If we ignore the disturbances, the implication of Eq. (2.5) is that the series increases (decreases) by a constant amount each period. For an increasing series (8 > 0), this implies a decreasing growth rate, and for a decreasing series (8 < 0), the specifica- tion gives an increasing decline rate. For series with an underlying constant growth rate, whether positive or negative, Eq. (2.5) is then an inappropriate specification. ‘The appropriate specification expresses the logarithm of the series as a linear func- tion of time. This result may be seen as follows. Without disturbances a constant growth series is given by the equation Y, = Yo + gy (2.6) where g = (¥;~ Y;-1)/¥;-1 is the constant proportionate rate of growth per period. Taking logs of both sides of Eq. (2.6) gives! InY, =a+ Bt (2.7) where a= In¥o and B = In(l +g) (2.8) Tf one suspects that a series has a constant growth rate, plotting the log of the series against time provides a quick check. If the scatter is approximately linear, Eq. (2.7) can be fitted by least squares, regressing the log of Y against time. The resultant slope coefficient then provides an estimate @ of the growth rate, namely, b=Inl+Q) giving = g=e-1 The B coefficient of Eq. (2.7) represents the continuous rate of change ain ¥,/3t, whereas g represents the discrete rate. Formulating a constant growth series in con- tinuous time gives Y= Ye or In, =a + Bt Finally, note that taking first differences of Eq. (2.7) gives Any, = B =In(l+g)~g 2.9) Thus, taking first differences of logs gives the continuous growth rate, which in turn is an approximation to the discrete growth rate. This approximation is only reason- ably accurate for small values of g. . Ae ne 2.1.2 Numerical Example Table 2.1 gives data on bituminous coal output in the United States by decades from 1841 to 1910. Plotting the log of output against time, we find a linear relationship. “We use In to indicate logs to the natural base e.44. BCONOMETRIC METHODS TABLE 2.1 Bituminous coal output in the United States, 1841-1910 ‘Average annual output (1,000 net tons), Decade ¥ 1 ny 1841-1850 1.837 -3 22.5457 1851-1860 4.868 -2 =16.9809 1861-1870 12411 -1 -9.4263 1871-1880 32,617 0 0 1881-1890 82,770 1 113238 1891-1900 148,457 2 23.8161 1901-1910 322,958 338.0558 Sum 0 24.2408 So we will fita constant growth curve and estimate the annual growth rate. Setting the origin for time at the center of the 1870s and taking a unit of time to be 10 years, we obtain the ¢ series shown in the table. From the data in the table = [MY _ 71.7424 _ 1.469 n 7 _ Sain _ 24.2408 b= Sp = ag = 08657 The r? for this regression is 0.9945, confirming the linearity of the scatter. The esti- ‘mated growth rate per decade is obtained from = 1 = 1.3768 ‘Thus the constant growth rate is almost 140 percent per decade. The annual growth rate (agr) is then found from : (1 + age)! = 2.3768 which gives agr = 0.0904, or just over 9 percent per annum. The equivalent contin- uous rate is 0.0866. The time variable may be treated as a fixed regressor, and so the inference pro- cedures of Chapter 1 are applicable to equations like (2.5) and (2.7).? 22 ‘TRANSFORMATIONS OF VARIABLES ‘The log transformation of the dependent variable in growth studies leads naturally to the consideration of other transformations. These transformations may be of the 2For a very useful discussion of the use of time as a regressor, see Russell Davidson and James G. MacKinnon, Estimation and Inference in Econometrics, Oxford University Press, 1993, pp. 115-118,cuapter 2: Further Aspects of Two-Variable Relationships 45 dependent variable, the regressor variable, or both. Their main purpose is to achieve a linearizing transformation so that the simple techniques of Chapter | may be applied to suitably transformed variables and thus obviate the need to fit more com- plicated relations. 2.2.1 Log-Log Transformations The growth equation has employed a transformation of the dependent variable, Many important econometric applications involve the logs of both variables. The relevant functional specification is Y=Ax® or = In¥ =a@+BInX (2.10) where a = In A, The elasticity of ¥ with respect to X is defined as. weiny - AY X Elasticity = Oy It measures the percent change in ¥ for a I percent change in X. Applying the elas- ticity formula to the first expression in Eq. (2.10) shows that the elasticity of this function is simply B, and the second expression in Eg. (2.10) shows that the slope of the log-log specification is the elasticity. Thus Eq. (2.10) specifies a constant elasticity function. Such specifications frequently appear in applied work, possibly because of their simplicity and ease of interpretation, since slopes in log-log regres sions are direct estimates of (constant) elasticities. Figure 2.1 shows some typical shapes in the ¥,X plane for various Bs. y y A A ° 0 1 x FIGURE 2.1 ¥ = Ax8,46 economeTRic MeTHoDs 2.2.2, Semilog Transformations One example has already been given in the constant growth equation, The general formulation is? In¥=a+PX+u Qi ‘This specification is widely used in human capital models, where ¥ denotes earnings and X years of schooling or work experience.* It follows from Eq. (2.11) that lay _ Ydx ‘Thus the slope in the semilog regression estimates the proportionate change in ¥ per unit change in X. An illustration for positive 8 is shown in Fig. 2.2a. Reversing the lnYea+Bx B>0 Yeaspinx p>0 @ (0) FIGURE 2.2 Semilog model. The astute reader will have noticed that in discussing various transformations we are playing fast and loose with the disturbance term, inserting it in some equations and not in others in order to simplify the transformations. The only justifications for such a (common) practice are ignorance and convenience. ‘The late Sir Julian Huxley (distinguished biologist and brother of the novelist Aldous Huxley) once described God as a “personified symbol for man’s residual ignorance.” The disturbance term plays 2 similar role in econometrics, being a stochastic symbol for the econometrician’s residual ignorance. ‘And, just as one ofien does with God, one ascribes to the inscrutable and unknowable the properties ‘most convenient for the purpose at hand. “This specification is derived from theoretical considerations in J. Mincer, School, Experience, and Eam- ‘ngs, Columbia University Press, New York, 1974,cuarren 2: Further Aspects of Two-Variable Relationships 47 axes gives: ¥=a+BinX (2.12) An illustration for positive 8 appears in Fig. 2.2b. In a cross-section study of house- hold budgets such a curve might represent the relation between a class of expenditure ¥ and income X, A certain threshold level of income (e~“#) is needed before any- thing is spent on this commodity. Expenditure then increases monotonically with . but at a diminishing rate. The marginal propensity (B/X) to consume this 's with increasing income, and the elasticity (B/Y) also declines as in- come increases. 2.2.3 Reciprocal Transformations Reciprocal transformations are useful in modeling situations where there are asymp- totes for one or both variables. Consider (F = ai (X — a2) = a3 (2.13) This describes a rectangular hyperbola with asymptotes at ¥ = a) and X = a9. Figure 2.3 shows some typical shapes for positive and negative a3. Equation (2.13) may be rewritten as Y=a,+ (2.14) Xap a> L : ' ' ' q @ o FIGURE 2.3 Rectangular hyperbola. . -48 ECONOMETRIC METHODS The result of adding an error term to Eq. (2.14) and attempting to minimize the residual sum of squares gives equations that are nonlinear in the a’s. In this case there is no possible linearizing transformation that will take us back to the simple routines of Chapter 1. However, there are two special cases of Eq. (2.14) where linearizing transformations are available. Setting a to zero gives y=a+a(y) (2.15) where a = a and 8 = a, Alternatively, setting a to zero gives () =a+ BX 2.16) where @ = —axfax and 8 = |/ee. Illustrations are shown in Figs. 2.4 and 2.5. Figure 2.4a has been fitted frequently in the study of Phillips curves, with ¥ rep- resenting the rate of wage or price change and X the unemployment rate. This spec- ification carries the unrealistic implication that the asymptote for the unemployment rate is zero. The alternative simplification in Eg. (2.16) permits a positive minimum, unemployment rate, but at the cost of imposing a zero minimum for wage change. veara(}) aro veara(}) aco @ ® FIGURE 24 Yy=a+p (3) As will be seen later, the equation may be fited directly by nonlinear least squares.cuarrer 2; Further Aspects of Two-Variable Relationships 49 y @ » FIGURE 28 yp a+ BX The more general specification illustrated in Fig. 2.3a removes both restrictions and allows the possibility of a positive minimum rate of unemployment and a negative wage change. Figure 2.4b might represent a cross-section expenditure function. A certain threshold level of income is required before there is any expenditure on, say, restaurant meals, but such expenditure tends toward some upper limit, where the billionaire spends only infinitesimally more than the millionaire. 23 AN EMPIRICAL EXAMPLE OF A NONLINEAR RELATION: U.S. INFLATION AND UNEMPLOYMENT The publication of the “Phillips curve” article in 1958 launched a new growth in- dustry, whose practitioners searched for (and found) Phillips curves in various coun tries.° In the original article Phillips plotted the annual percentage wage change in the United Kingdom against the unemployment rate for the period 1861 to 1913. The seatter revealed a negative nonlinear relation, which Phillips summarized in the form of a curved line. Most remarkably, data for two subsequent periods. 1913-1948, and 1948-1957, lay close to the curve derived from the 1861-1913 data. This simple Phillips curve has not survived the passage of time and has been subject to both sta- tistical and theoretical attack and reformulation. Thus a simple two-variable analysis, SA. W. Phillips, “The Relation between Unemployment and the Rate of Change of Money Wages in the United Kingdom, 1861~1957," Economica, New Series 25, 1958, 283-299.50 ECONOMETRIC METHODS of wage (or price) inflation and unemployment can no longer be regarded as a serious piece of econometrics. However, in this chapter we are still restricted to two-variable relations, and the following example should only be taken as an illustration of the statistical steps in fitting nonlinear relations in two variables. The data used are annual data for the United States from 1957 to 1970. The in- flation variable (INF) is the annual percentage change in the Consumer Price Index (CPI). The unemployment variable (UNR) is the unemployment rate for civilian workers 16 years and over, Inflation ranges from a low of 0.69 percent in 1959 to a high of 5.72 percent in 1970, with a mean of 2.58 percent. The unemployment rate was 4.3 percent in 1957, rising to a peak of 6.7 percent in 1961 and falling steadily 6 al + z +7 2 2 oF oe + + ° . . 37 @ SO UNR @ 6 ab 2 2 ° 0.10 O15 020 0.25 0.30 Reciprocal UNR, lagged © FIGURE 26 U.S. inflation and unemployment, 1957-1970.cuiaprer2: Further Aspects of Two-Variable Relationships 51 TABLE 2.2 ‘Various inflation/unemployment regressions, 1957-1970* Explanatory variable Constant Slope SER. a9! UNR 6.92 0.8764 033 140 (3.82) (-2.45) UNR(-1) 9.13 13386 081074 (9.80) {(-7.19) VUNR(-1) 4.48 32,9772 0.90 054 (-651) (10.50) "The statistics are in parentheses, $.E.R- isthe standard error ofthe regression, through the rest of the 1960s. Figure 2.6a shows the scatter of inflation against the current unemployment rate. The slope is negative but the scatter is dispersed. In Fig. 2.6b inflation is plotted against the previous year's unemploymentrate. A lagged response is not unreasonable since time is required for unemployment to affect wages and further time for wage changes to filter through to the prices of final goods. The scatter is now much tighter and there is an indication of nonlinearity. The same figure shows the fit of a linear regression of inflation on lagged unemployment. The linear specification is clearly an inadequate representation. Of the 14 residuals from the regression, 5 are positive and 9 are negative. The 5 positive residuals occur at the lowest and highest values of the explanatory variable. Inspection of the residuals can thus indicate possible misspecification. Runs of positive or negative residuals suggest misspecification. Figure 2.6c shows the result of fitting the reciprocal relation INF = o+rlmecyl + ary The residuals are somewhat smaller than in Fig. 2.6b and the scatter is more nearly inear, but not totally so. Table 2.2 summarizes the main results from the regressions associated with Fig, 2.6. We notice the substantial jump in 7? on changing the ex- planatory variable from current to lagged unemployment, and a still further increase from 0.81 to 0.90 on using the reciprocal transformation. Finally, we note the result of fitting the nonlinear relation. 3 INF = @1 + ORR oa t (2.18) This is fitted by nonlinear least squares, which is an iterative estimation process, commencing with some arbitrary values for the unknown parameters. calculating the residual sum of squares, then searching for changes in the parameters to re- duce the RSS, and continuing in this way until successive changes in the estimated parameters and in the associated RSS are negligibly small. Standard errors and other test statistics can be produced at the final stage, just as in linear least squares: but, as will be explained shortly, they now have an asymptotic justification rather than exact, finite sample properties. As noted earlier the linearizing transformations obtained by setting a or 2 to zero impose theoretically inappropriate constraints on52 ECONOMETRIC METHODS, the shape of the relationship. Setting a to zero, as in the third regression in Table 2.2, gives a lower asymptote of zero for the unemployment rate, which is implausibly small, On the other hand, setting a to zero gives a lower asymptote of zero for the inflation rate, which implies that the price level could not fall no matter how great the level of unemployment. which again is an implausible restriction. Using nonlinear least squares to estimate the relation without these restrictions gives 4.8882, UNR(— 1) — 2.6917 with 7? = 0.95 and S.E.R ~ 0.40. This expression provides the best fit and low- est standard error of all the regressions, The intercept term, which is the estimated asymptote for the inflation rate, is slightly negative but not significantly different from zero. The estimated asymptote for the unemployment rate is 2.69 percent. The contrast between the unemployment asymptotes in fitting Eqs. (2.17) and (2.18) is a striking reminder of the fact that each specification imposes a particular shape on the estimated relation. Equation (2.18) implies dramatically increasing inflation rates for unemployment rates just below 3 percent, whereas Eq. (2.17) requires un- employment rates below 1 percent to give similar inflation numbers, The fits to the sample data are not very different, but extrapolations outside the range of the sample data give dramatically different pictures. INF = -0.32+ 24 LAGGED DEPENDENT VARIABLE AS REGRESSOR, When variables display trends as in Section 2.1, successive values tend to be fairly close together. Another way of modeling such behavior is by means of an autore- gression. The simplest autoregressive scheme is Y= a+ BYitu (2.19) This is called a first-order, autoregressive scheme and is frequently denoted by the notation AR(1). The order indicates the (maximum) iag in the equation. If, for ex- ample, ¥ were measured quarterly. Y, = a + BY,—4 +, embodies the assumption that the current ¥ is related to the value in the same quarter of the previous year, and this is a special case of an AR(4) scheme. The LS equations for fitting Eq. (2.19) are Diy = na +b SY DK =a Sh tbo ¥2y The range of summation in Eq. (2.20) has not been indicated explicitly. If ¢ ranges from 1 to n, the implementation of Eq, (2.20) requires a value for Yo. If this is not available, then ¥; is the starting value and the effective sample size is n — 1. LS estimates of the parameters of Eq. (2.19) can thus be computed; but the properties of LS estimators and the associated inference procedures derived in Chapter 1 are not strictly applicable here, even though we continue to make the same assumptions (2.20)54 EconomeTRic METHODS. is how a random variable such as %, and its pdf behave as n > ». There are two main aspects of this behavior, the first relating to convergence in probability and the second to convergence in distribution, 2.4.2. Convergence in Probability ‘The x’s are iid( 2, 07) by assumption. It follows directly that e E(x) = wand var ¥q) = TF Thus %, is an unbiased estimator for any sample size, and the variance tends to zero as n increases indefinitely. It is then intuitively clear that the distribution of %,, whatever its precise form, becomes more and more concentrated in the neighborhood of 4 as n increases. Formally, if one defines a neighborhood around px as ju + €, the expression Pru —€ Prilin — ul 1-8 (2.23) The random variable #, is then said to converge in probability to the constant yz. An equivalent statement is lim Priliq — | 139 >} | | 120 no) | 100) |-————--—_———_ oo) J] > i: = 6 sof + 50 40 0 30 ? ; 5” 20 L B® f wh 10 ° a __aed 0 0 20 400 6009001000 5 4 5 6 7 8D City size (housands) Milos @ o FIGURE 2.7 Influence of original distribution and sample size on approach to normality. (a) Frequency distribution of sizes of 196 US. cities in 1920; (6) frequency distribution of totals of 200 simple random samples with n = 49, (Reprinted by permission of Joha Wiley & Sons, Inc.)‘cnarteR2: Further Aspects of Two-Variable Relationships 57 with zero means and finite variances and covariance.’ Thus the least-squares esti- mates are consistent for @ and 8. Moreover the limiting variances and covariance may be consistently estimated by the LS formulae of Chapter 1. Consequently we may apply the LS techniques of Chapter 1 to the autoregressive model; and they now have an asymptotic, or large-sample, justification rather than exact, finite-sample va- lidity. ‘The Mann—Wald result depends on two crucial assumptions. The first assump- tion is that the disturbances in the relation are independently and identically dis- tributed with zero mean and finite variance, as in Eq. (2.2). Note, however, that there is no assumption of normality for the disturbances. The second assumption is that the {¥;} series is stationary. This raises new considerations, which are explored in the next section, 25 STATIONARY AND NONSTATIONARY SERIES We return to the relationship specified in Eq. (2.19), Y¥,=a@+ BY) +m and make the assumptions about the u variable stated in Eq. (2.2). These assumptions define a white noise series. The crucial question is, how does the Y series behave over time? Equation (2.21) shows Y, as a function of a, 8, Yo, and the current and previous disturbances. Assuming that the process started a very long time ago, we rewrite Eq, (2.21) as Y= a(l+ B+ BP +--) +(e, + Burt Burt) (2.29) The stochastic properties of the ¥ series are determined by the stochastic properties of the u series. Taking expectations of both sides of Eq. (2.29) gives E(¥,) = a(l + B + B2 +--+) This expectation only exists if the infinite geometric series on the right-hand side has a limit. The necessary and sufficient condition is ial <1 (2.30) The expectation is then BW) = a= 725 231) and so the Y series has a constant unconditional mean yz at all points. To determine the variance we can now write °H, B. Mann and A. Wald, “On the Statistical Treatment of Linear Stochastic Difference Equations.” Econometrica, 11, 1943, pp. 173-220. The article is long and very technical, but the results are of great practical importance.58 ECONOMETRIC METHODS (fp =p) = ty + Bly + Bua to (2.32) ‘Squaring both sides and taking expectations, we find var(¥,) = Els ~ 47] = Elud + Bud, + Btudy +--+ 2B. + 2BPurty-2 +2 The assumption in Eq. (2.2) then yields varlY) = 03 = (2.33) B Thus the ¥ series has a constant unconditional variance, independent of time. ‘A new concept is that of autocovariance, which is the covariance of Y with a lagged value of itself. The first-lag autocovariance is defined as, m= EL — WO - 2) = Bot using Eq. (2.32) Ina similar fashion the second-lag autocovariance is, , Ya = EY, ~ wha - WI = p02 and, in general, y= Bo, 5 = 0,12... (2.34) ‘The autocovariances thus depend only on the lag length and are independent of 1, Clearly yo(= 02) is another symbol for the variance. Dividing the covariances by the variance gives the set of autocerrelation coefficients, also known as serial cor- relation coefficients, which we will designate by ps=ylyo $= 012... (2.35) 107 = 08k a + 2 Boot | § 2 Bosak * i + 3 + o2b +, tHe t+ oo + FIGURE 28 ° 5 10-15-20 Correlogram of an AR(I) series Lag (parameter = 0.75).Harter 2: Further Aspects of Two-Variable Relationships 59 Plotting the autocorrelation coefficients against the lag lengths gives the correlo- gram of the series. For the first-order AR(1) scheme the autocorrelations decline exponentially from one toward zero, as illustrated in Fig. 2.8. To summarize, when |B| <1 the mean, variance, and covariances of the ¥ se- Ties are constants, independent of time. The Y series is then said to be weakly or covariance stationary. In particular, it satisfies the stationarity condition required for the asymptotic, results of Mann and Wald to hold true. 2.5.1 Unit Root When 8 = I the AR(1) process is said to have a unit root. The equation becomes Ys at¥nr tm (2.36) which is called a random walk with drift.'° From Eq. (2.21) the conditional ex- peciation is E(Y,| Yo) = at + Yo which increases or decreases without limit as t increases. The conditional variance is var(Y; | Yo) = El(¥; — EM | ¥o))"1 = Elle + mya to + a)) = which increases without limit. Referring to Eq. (2.29) we clearly see that, in the unit root case, the unconditional mean and variance of ¥ do not exist, The ¥ series is then said to be nonstationary, and the asymptotic results previously described no longer hold. The treatment of nonstationary series will be taken up in Chapters 7 and 8. ‘When |8| > | the Y series will exhibit explosive behavior, as will be illustrated in the following example. 2.5.2. Numerical Illustration Three series have been generated as follows: A A 0.05 +0.95A,-; +4, Autoregressive (stationary) series RR, = 005+ Rei tay Random walk with drift, BE: £, = 0.05+1.05E,1+ 4, Explosive series All series are started at zero and 500 terms generated by random drawings from a standard normal distribution u,. Figure 2.9 contrasts the autoregressive and random "When a = 0, Bg. (2.36) reduces to a simple random walk, Fy = Your + de60 ECONOMETRIC METHODS 2 z < % 160 240 320 +400 480 Observation number FIGURE 29 A stationary AR(1) series and a random walk. walk series, The theoretical mean of the A series is 1, and its theoretical standard deviation is 3.2. The series displays a fairly steady mean level, and all observations are contained in the range of + 10. The random walk series does not look much dif- ferent from the stable A series: through most of the early observations, but does drift markedly away in the later observations. Figure 2.10 shows the A and R series again, ._ but this time the explosive E series is plotted on the same graph, Notice the dramatic difference in the vertical scale compared with Fig. 2.9. On the scale required to accommodate the E series, the other two appear as a single straight line. The B pa- rameter for E only exceeds unity by 0.05, the same amount as the B parameter for A falls short of unity. Thus, the unit root case is a watershed. The A and R series have more the typical look of an economic time series than does the E series. As Fig. 2.9 suggests it may in practice be very difficult to distinguish between stationary series, such as A, and nonstationary series like R. Tt might appear simple to test for a unit root by fitting Eg. (2.19), computing the test statistic (b — 1)/s.e.(b), and referring this to the conventional critical values from the ¢ distribution, Unfortunately this procedure is not valid since, under the full hypothesis, the ¥ series does not satisfy, even asymptotically, the conditions assumed in deriving the test. The distribution of the test statistic is nonstandard, and critical values can only be obtained by Monte Carlo simulations. Unit root tests will be dealt with in Chapter 7. A more informal test for stationarity is based on inspection of the autocorrelation coefficients. As shown earlier, the correlogram of a stationary AR series should decline exponentially. This will not be true for a nonsta- tionary series. Table 2.3 shows selected autocorrelation coefficients for the A and Rccharrer2: Further Aspects of Two-Variable Relationships 61 AR and E Observation number FIGURE 2.10 An explosive series. TABLE 2.3 Autocorrelations for the A and R series Lag A series R series 1 0.936 0.992 5 0.678 0.958 9 0395, 0.929 13 0.202 o9it 18 0.108 0.882 series. There is a clear difference in the patterns of the two sets of autocorrelations, confirming the stationarity of the first series and the nonstationarity of the second. However, these coefficients have been calculated from observations 101 to $00. The differences would not be so clear in smaller samples. We will return to the issue of stationarity and the validity of standard inference procedures in more realistic ‘multivariate situations in later chapters. 2.6 MAXIMUM LIKELIHOOD ESTIMATION ™ OF THE AUTOREGRESSIVE EQUATION 2.6.1 Maximum Likelihood Estimators The derivation of the Mann—Wald results in Section 2.4 did not require any assump- tion about the specific form of the paf for the disturbance term. Ifsuch an assumption62 BCONOMETRIC METHODS. can be made, it is then possible to derive maximum likelihood estimators of the pa- rameters of the autoregressive model. Maximum likelihood estimators, or MLEs, are consistent and asymptotically normal, as are the Mann—Wald estimators; but the MLEs have the additional property of asymptotic efficiency, as will be explained next. A more complete treatment of ML estimation will be given in Chapter 5, to which the present treatment of the simple autoregressive case will serve as an intro- duction. The comton approach is to assume, as in Eq, (2.4), that the disturbances are identically and independently distributed normal variables with zero mean and con- stant variance. The probability density function for w is then Lane Flai) = ee GaLQ..m olin We will further postulate some arbitrary initial value Yo, the precise value being irrelevant in an asymptotic analysis, Any observed set of sample values ¥1, ¥2. ..-. Yn is then generated by some set of uw values. Given Eq. (2.4), the probability of a set of u values is Pru, ua,» Surf (ua) + fe) = [fe rl 1 Epatwne?y Croty? From Eq, (2.19) the joint density of the ¥ values, conditional on yp, is then!” 1 aero gS, = a-p¥..1)*) 237) a Pr(V 1 Yas Pa) = This density may be interpreted in two ways. For given «, 8, and o? it indicates the probability of a set of sample outcomes. Alternatively it may be interpreted as a function of a. B, and o*, conditional on a set of sample outcomes. In the latter interpretation it is referred to as a likelihood function and written Likelihood function = L(a, B, 0°; ¥) (2.38) with the order of the symbols in the parentheses reflecting the emphasis on the pa- rameters being conditional on the observations. Maximizing the likelihood with re~ spect to the three parameters gives specific values @, 8 and 6?, which maximize the probability of obtaining the sample values that have actually been observed. These are the maximum likelihood estimators (MLEs) of the parameters of Eq. (2.19). They are obtained by solving ob ol _ Jl yg aa Op dt "See Appendix 2.1 on the transformation of variables in pdf's,‘charter 2: Further Aspects of Two-Variable Relationships 63 In most applications it is simpler to maximize the logarithm of the likelihood function. We will denote the log-li Since | is a monotonic transformation of L the MLEs may equally well be obtained by solving ald a Ga 3B” Go? For Eq. (2.19) the log-likelihood (conditional on Yo) is T= 5 ln@a) — Fine? ina Bry (2.39) The formal derivations are given in Appendix 2.2, but it is intuitively clear that the 4, B values that maximize / are those that minimize 0"_(¥, — @ — BY,-1)*. Thus, in this case the LS and ML estimates of « and B are identical. The normal equations for the LS estimates appear in Eq. (2.20). The ML estimate of o is @ =! Su,-a- srw 240) a 2.6.2 Properties of Maximum Likelihood Estimators The major properties of ML estimators are large-sample, or asymptotic, ones. They hold under fairly general conditions. 1. Consistency. MLEs are consistent. Thus, Eqs. (2.20) and (2.40) yield consistent estimates of , B, and a. . 2. Asymptotic normality. The estimators @, A, and 6? have asymptotically normal distributions centered at the true parameter values. The asymptotic variances are derived from the information matrix, which will be explained in Chapter 5. In the present application it may be shown that the asymptotic variance of is estimated by é # Est. avar(B) = ——3—7-—, 5 YP, - Sh ha? where 6? is defined in Eq. (2.40) and the abbreviation avar denotes asymptotic variance. If we let ¥_, = (/n)(¥o + ¥i +--+ Yaea)and ya = Ya — Fo, the estimated asymptotic variance becomes Est. avar(6) = <——~ Sree which is seen, by comparison with Eq. (1.40), to be the variance that would be calculated by the usual LS procedure, except for the trivial difference that 67 has a divisor of m rather than (n — 2). .64 ECONOMETRICMETHODS 3. Asymptotic efficiency. No other consistent and asymptotically normal estimator can have a smaller asymptotic variance. This property mirtors the finite sample minimum variance property of LS estimators. To summarize, the lagged dependent variable model of Eq. (2.19) may be esti- mated by LS and standard errors calculated in the usual way. However, significance levels and confidence coefficients are no longer known precisely. The calculated values are only approximately correct and there is no way of assessing the degree of error in any particular application. Even this somewhat cautious conclusion still rests on two important assumptions, The first is the zero covariance assumption for all pairs of disturbances. which was used in the derivation of Eq. (2.37). The second is that S. y?_,/7 has a finite limit as » tends to infinity, which is part of the require- ments for the stationarity of the ¥ series. APPENDIX APPENDIX 2.1 Change of variables in density functions The basic idea may be simply illustrated for the univariate case. Suppose uw is a random variable with density function p(w). Let a new variable y be defined by the relation y = ftw). The y variable must then also have a density funetion, which will obviously depend on both p(u) and f(u). Suppose that the relation between y and w is monotonically increasing, as in Fig, A-2.1. Whenever u lies in the interval Au, y will be in the corresponding interval Ay. Thus Pr(y lies in Ay) = Pr(w lies in Au) or poy')Ay = plu’yAu where uw" and y" denote appropriate values of w and y in the intervals Au and Ay, and p(y) indicates the postulated density function for y: Taking limits as Au goes to zero gives du POY = Pw 7 Ify were a decreasing function of u, the derivative in this last expression would be negative, thus implying an impossible negative value for the density function. There- fore, the absolute value of the derivative must be taken and the result reformulated toread Py) = plu) du day Ify = f(s) were nota monotonic function this last result would require amendment, but we are only concerned with monotonic transformations. The relevant relationshipCHAPTER 2: Further Aspects of Two- Variable Relationships 65 yas) ° a FIGURE A-2.1 in the text is Y= a+ Bhat (2.19) which gives du, a7) forall which gives the joint density shown in Eg. (2.37). APPENDIX 2.2 Maximum likelihood estimators for the AR(1) model The log-likelihood is given in Eq. (2.39) as ~finQn)~ Fino*— 5 aoe -a- BYP The partial derivatives with respect to the three parameters are Z. eter e BY,-1)ald D> Yie >a — BY) wt = al 1< a + gt a Blea? ao? ~ 20? Equating the first two derivatives to zero gives the ML (LS) estimators in Eq. (2.20), and equating the third to zero gives the estimator of the variance in Eq. (2.40). PROBLEMS 2.1, Fit a constant growth curve to the accompanying data, using two different specifications of the time variable, Marijuana crop ‘Year (10,000 tons) 1985 38.1 1986 30.0 . 1987 W704 1988 354.5 1989 744.4 Estimate the annual growth rate, and forecast marijuana production in 1995 from each specification. 2.2. Show that log ¥ = a + BlogX + w gives the same estimate of i whether logs are taken to base 10 orto base ¢. Is this true ofthe estimate of a? Do your conclusions need modification if log X is replaced by 1? 2.3. Discuss briefly the advantages and disadvantages of the relation vy = at Blog ve as a representation of an Engel curve, where v; is expenditure per person on commodity i, and vp is income per week. Fit such a curve to the following data, and from your results estimate the income elasticity at an income of $500 per week. Does it make any difference to your estitnate of the income elasticity if the logarithms of vp are taken to base 10 or to base e? $°00 per week y 08 12 15 18 22 23° 26 31 w 17 27 36 465.7 67 81120 2.4. Prove that 4YX _ diin¥) _ dog ¥) aX ¥ ~ dilnX) ~ aogX) Note that this relationship holds generally and not just for constant elasticity functions.cuapTeR 2: Further Aspects of Two-Variable Relationships 67 2.5. A response rate Y to a stimulus X is modeled by the funetion 1 8 io-¥ ~°*¥ where ¥ is measured in percentage terms. Outline the properties of this function aad sketch its graph. Fit the function to the accompanying data. ee xX 3 7 2 17 2 3 45 55 7 120 y % 7% 1% 6 6 62 52 SI SI 48 26. Discuss the properties of the following functions and sketch their graphs: x ye ax-B outs Tre » Find the transformations that will linearize each function, 2.1. The firmness of cheese depends on the time allowed for a certain process in its manu- facture. In an experiment 18 cheeses were taken, and at each of several times, firmness was determined on samples from three of the cheeses. The results were as follows. Tume (hours) Firmness os 102 1051S Il no 120115 Ls 126° 128119 2 B20 143139 3 160 49147 4 164 166172 Estimate the parameters of a linear regression of firmness on time, Compute 7? and the standard errors of the estimates. Calculate the conditional means at each value of time ‘and regress these conditional means on time. Compare the coefficients, standard errors, and ? with the previous regression. Can you think of conditions where the results you have obtained in this example would not hold? 2.8. A theorist postulates that the following functional form will provide a good fit to a set of data on ¥ and X: a+o/ ‘Sketch the graph of this function when @ and 6 are both positive. Three sample obser- vations give these values:68 ECONOMETRIC METHODS Fit the foregoing function to these data. If Y denotes the per capita consumption of peanuts and X denotes income, give a point estimate of the peanut consumption of a millionaire. 2.9, An economist hypothesizes that the average production cost of an article declines with increasing batch size, tending toward an asymptotic minimum value. Some sample data from the process are as follows, Batch size 1 5 to 20 Average cost(S) 311412 v-ase(z) to these data. where Y denotes average cost and X indicates batch size. What is the estimated minimum cost level? Estimate the batch size that would be required to get the average cost to within 10 percent of this minimum level. Fit a curve of the form 2.10. A variable x, has the following pdf: x 1 n RO) - Un in Determine £(x,) and var(x,), and investigate their limits as n becomes infinitely large. Does plim x, exist? If so, what is its value? 2.1L, Let xy.43.....%, be a random sample from each of the following pdfs: (a) Bemoulli distribution fro) = ei OS A<1 x =01 (b) Poisson distribution son = FE 012. secs (©) Exponential distribution F008) = Be" O (oxy, nb +b: X= SY or bX +h > xX = xy as in Eq, (1.28). EXAMPLE3,2, THREE-VARIABLE EQUATION. Ina similar fashion it may be shown that the normal equations for fitting a three-variable equation by least squares are nby + by >) X2 + by Xa = DY72 BCONOMETRIC METHODS by D> Xa + be X93 + bs XOXs = SX by 31% + ba > XaXs + ba KF = GY Ify in Eq. (3.4) is replaced by Xb + e the result is (XX) = X'(Xb +e) = (XX) + Xe Thus, Xe=0 @.5) which is another fundamental least-squares result. The first element in Eq. (3.5) gives Se; = 0, that is, 2 = Pb — boXp — ++ — bX, = 0 The residuals have zero mean, and the regression plane passes through the point of ‘means in & dimensional space. The remaining elements in Eq. (3.5) are of the form De Ones! ek As seen in footnote 16 of Chapter 1 this condition means that each regressor has zero sample correlation with the residuals, This in turn implies that (= Xb), the vector of regression values for Y, is uncorrelated with e, for Sie = (Xbye = b'X'e = 0 3.1.2 Decomposition of the Sum of Squares The zero covariances between regressors and the residual underlie the decomposi- tion of the sum of squares. Decomposing the y vector into the part explained by the regression and the unexplained part, y=jte=Xbte it follows that JY = Gt eG +e) =HP tele = HXX +e'e However, y'y = of, ¥? is the sum of squares of the actual ¥ values. Interest nor- mally centers on analyzing the variation in ¥. measured by the sum of the squared deviations from the sample mean, namely, Di - HF = Sv? - nv? Thus, subtracting n¥? from each side of the previous decomposition gives a revised decomposition, } Oy ~ n¥?) = ('X'Xb — n¥*) + e'e f TSS = ESS +RSS where TSS indicates the total sum of squares in Y, and ESS and RSS the explained and residual (unexplained) sum of squares. j (3.6)cuarter 3: The k-Variable Linear Equation 73 3.1.3 Equation in Deviation Form An alternative approach is to begin by expressing all the data in the form of devia- tions from the sample means. The least-squares equation is Yp = by + BaXy + sXe to + OXy te t= Len Averaging over the sample observations gives ¥ = by + byXp + b3Xy too + By which contains no term in e, since @ is zero. Subtracting the second equation from the first gives Ye = baxa + baxze too + exe + Cr where, as in Chapter 1, lowercase letters denote deviations from sample means. The intercept b; disappears from the deviation form of the equation, but it may be recov- ered from by =P bnRy— bake The least-squares slope coefficients by gression equation, as are the residuals. Collecting all n observations, the deviation form of the equation may be written compactly using a transformation matrix, snl where i is a column vector of n ones. As shown in Appendix A. this is a symmetric, idempotent matrix, which, on premultiplication of a vector of n observations, trans- forms that vector into deviation form. Thus it follows that Ae = e and Ai = 0. Write the least-squares equations as. by are identical in both forms of the re~ (3.7) y=Xbt+e=li xilp']+« where X2 is the m X (k — 1) matrix of observations on the regressors and by is the k— 1 element vector containing the coefficients, bz, bs, ..., by. Premultiplying by A gives Ay = 10 Axa [ft] +40 = (AX2)b2 +e 2 or ye = Xb +e (3.8) where ye = Ay and X. = AX? give the data in deviation form. Since X’e = 0, it follows that Xe = 0. Thus premultiplying Eq. (3.8) by X! gives Xiy. = (X.X.)br which are the familiar normal equations, as in Eq. (3.4), except that now the data have all been expressed in deviation form and the 52 vector contains the k ~ 1 slope coefficients and excludes the intercept term. By using Eq. (3.8), the decomposition74 ECONOMETRIC METHODS of the sum of squares may be expressed as Vn = DAXLKb2 + ele TSS= ESS +RSS (3.9) The coefficient of multiple correlation R is defined as the positive square root of > _ ESS _|_ RSS hl Thus R? measures the proportion of the total variation in Y explained by the linear combination of the regressors. Most computer programs also routinely produce an adjusted R?, denoted by R*. This statistic takes explicit account of the number of regressors used in the equation. It is useful for comparing the fit of specifications that differ in the addition or deletion of explanatory variables. The unadjusted R? will never decrease with the addition of any variable to the set of regressors. If the added variable is totally irrelevant the ESS simply remains constant. The adjusted coefficient, however, may decrease with the addition of variables of low explanatory power. It is defined as 3.10) a | TsSiin— 1) As will be seen later, the numerator and denominator on the right-hand side of Eq. (3.11) are unbiased estimates of the disturbance variance and the variance of Y. The relation between the adjusted and unadjusted coefficients is (3.11) (3.12) ‘Two other frequently used criteria for comparing the fit of various specifications involving different numbers of regressors are the Schwarz criterion? sc=n@! + inn non and the Akaike information criterion? AIC = In— + ze One looks for specifications that will reduce the residual sum of squares, but each criterion adds on a penalty, which increases with the number of regressors. Schwarz, G., “Estimating the Dimension of a Model,” Annals of Statistics, 6, 1978, 461-464. 3 Akaike, H., “Information Theory and an Extension of the Maximum Likelihood Principle,” in B. Petrov and F. Csake, eds., 2nd International Symposium on Information Theory, Budapest, Akademiai Kiado, 1973.cuapter 3: The k-Variable Linear Equation 75 EXAMPLE 3.3. To illustrate these formulae, here is a brief numerical example. The numbers have been kept very simple so as not fo obscure the nature of the operations with cumbersome arithmetic. More realistic computer applications will follow later. The sample data are 13 i and = X=/1 5 12 14 araau 5! where we have already inserted a column of ones in the first column of X. From these data we readily compute 515 25 20 NSe95) 1 ABN | ands || 70. 25 81 129 109. The normal equations, Eq. (3.4), are then 5 15 25)[bi 207 15 55 81]|d> 16 25 81 129JLbs] 109. To solve by Gaussian elimination we first subtract three times the first row from the second row and five times the first row from the third, These steps give the revised system, 5 15 25)7b,) (20) 0 10 6))/| =|16 0 6 libs 9] Next, subtract six-tenths of the second row from the third to get 5 15 25 jf] [20 ( 10 6 |el- 16 0 0 oallon) [-o6 The third equation gives 0.44; = ~0.6, that is by = -15 Substituting for bs in the second equation, we find that 10b2 + 6b; = 16 which gives by = 25 Finally, the first equation 5b, + 15b; + 25bs = 20 ives bad ‘The regression equation is thus ¥ = 4+25X2— L5Xa “The sample data have yielded a unique solution forthe B vector. The condition required fora unique solution will be examined later in this chapter.16 EcoNoMETRIC METHODS Alternatively, transforming the data into deviation form gives <1 0 -3 -1 =| 4] and 1 -1 -1 1 1 i The relevant normal equations are then (" 4 2 _ ['s 6 allo) 1L9 ‘These are the second and third equations obtained in the first step of the Gaussian elim- ination above.’ Thus the solutions for by and bs coincide with those already obtained, Likewise 6, will be the same as before, since the final equation in the foregoing back substitution is readily seen to be by = ¥ — boX2 — biXs From the y. vector, TSS is seen to be 28. ESS may be computed from [10 all 25 HKD. = 125 1.51) 6 ai| 26.5 ‘or, more simply, from biXly, = [25 -rsi['5| 26.5 ‘Then RSS is 1.5, R? = 0.95 and R? = 0.89. 3.2 PARTIAL CORRELATION COEFFICIENTS With two or more regressors, partial correlations become relevant. Consider again the Plosser/Schwert example in Chapter I of a high positive correlation between the logs of nominal income in the United States and accumulated sunspots. It was suggested there that each variable displayed independent time trends and that the influence of this common variable (time) was basically responsible for the observed correlation between income and sunspots. This supposition can be checked by fitting time trends to each variable separately, computing the unexplained residuals from these time trends, and examining the correlation between the residuals. To simplify the notation let ¥ = log of nominal income Xp = log of accumulated sunspots X = time (measured in years) See Problem 3.1.cuarrer3. The k-Variable Linear Equation 77 ‘We will also use the index 1 to refer to Y, 2 for Xp, and 3 for X3. Then 712 = correlation between ¥ and X2 123 = correlation between X» and X3, etc. biz = slope of the regression of ¥ on Xp b32 = slope of the regression of X3 on X2, etc. 1, = residual from the regression of ¥ on X> €32 = residual from the regression of X3 on X3, etc. Working with the data in deviation form, the unexplained residual from the regres- sion of log income on time is e. y — bi3x3 where To keep the equations uncluttered, we omit the observation subscripts. Likewise the unexplained residual from the regression of the log of accumulated sunspots on time is Dass za The partial correlation coefficient between income and sunspots. with the influence of time removed or held constant, is defined as the correlation coefficient between these two sets of residuals. It is denoted by rj23. Thus, €23 = X2—bysxx where bas Eaves a VZ4/E4s Since the e’s are least-squares residuals, they have zero means. so there is no need for a correction for means in Eq. (3.13). One could implement Eq. (3.13) directly by computing the two sets of residuals and then calculating the correlation coefficient between them. In practice, however, itis simpler to express rz 3 in terms of the three simple correlation coefficients, ry2, r13, and rz3.° That is, nia eee teats G14) I-rb 1-4 The simple correlation coefficients are often called zero-order coefficients and coefficients like 7) 3 first-order coefficients, since the common influence of one other variable has been taken into account. In a typical spurious correlation case the zero order correlations tend to be large, whereas 12.3 will be negligible. In general, how- ever, first-order correlations may be larger or smaller than the corresponding zero- See Appendix 3.178 BCONOMETRIC METHODS order coefficients, and they may not even have the same sign. In the three-variable case there are two other first-order coefficients, namely, ri3.2 and ro.1. The first mea- sures the net association between ¥ and X3 when any common influence from X> has been allowed for, and the second measures the net association between X2 and X3 when any common effect from ¥ has been removed. In a single-equation specifica- tion, rj23 and ry39 are usually the first-order coefficients of interest. The formula for ri32 may be derived from first principles or from Eq. (3.14) by interchanging subscripts 2 and 3. Thus, n3— rons na = @.15) Viol 3.2.1 Sequential Buildup of the Explained Sum of Squares The main objective of the single-equation analysis is to explain the variation Sy? in the dependent variable. Running a regression of Y on X2 gives an explained sum of squares of r7, Sy, leaving a residual sum of squares of (1 ~r},) > y? = Sef. The proportion of this residual variation explained by using the adjusted X variable, that is. e3.2 = x3 —b32.x2, is r},.9. Thus the increment in the ESS at the second stage is 1}, 9(1 — 173) Sy. Adding the ESS at each stage gives a total ESS of [r?, +r; .(1 r}>)] = y*. Alternatively, the multiple regression of Y on Xj and X3 gives a total ESS. which may be written as R}., > y*, where Rj 23 is the coefficient of multiple correlation. using the index notation to show explicitly which variables are involved. In this approach the increment in ESS due to the addition of X3 is (R?.25 — 7?) >. y*. ‘The two expressions for the incremental ESS are identical, for it may be shown’ that Rin = rh + riaa(l ~ rh) 3.16) The sequence may alternatively be started with X3, and the increment due to Xp computed. Both sequences are shown in Table 3.1. TapLE 3 Buildup of the explained sum of squares ‘Variable Sumofaquares Variable Sum of guares % aze % REP ——hLrrT———C—CCM Xpand Xs Rly Xpand Xs Rin Sy Residual (I~ R}y,) Ny? Residual (UR) DP 7See Problem 3.3.cHaprer 3: The k-Variable Linear Equation 79 EXAMPLE 3.4. The data of Example 3.3 may be used to illustrate Table 3.1. We have _ 6 - f= ary 7 O93 re = 0.9562 oy _ = 7h = Gaya 7 07232 na = 0.8504 or _ _ Tl ray = 0.9487 Remember that the signs of the correlation coefficients must be determined from the covariances, If one calculates the squared correlations, one must not simply take the positive square roots in calculating the coefficients. The partial correlations are then 0.9562 — (0.8504)(0.9487) rag = 29362 = OSS0O9887) 9 g082 ma i= 07232 Jt — 0.9000 A 0.8504 — (0.9562)(0.9487) = -0.6130 v1 — 0.9143 J/1 = 0.9000 Thus r3, = 0.8067 andr}, = 0.3758. The various sums of squares for Table 3.1 may then be computed as my = 256 ad -) Dy A> y= 2025 Rd -ySyY Table 3.2 collects the results. ‘The total ESS was shown in Example 3.3 to be 26.5, and the same result is reached here by use of the simple and partial correlation coefficients =09 6.25 With two (or more) explanatory variables there is no unambiguous way of de- termining the relative importance of each variable in explaining the movement in Y, except in the extreme and unlikely case of zero correlation between the explanatory variables. When r3 is zero, the variables are said to be orthogonal, and it may be shown’ that R},, = 77, + rj; Thus in this special case ESS may be split into two components, each of which is unambiguously attributed to an explanatory vai able. When the X's are correlated, no such split is possible. Kruskal considers various TABLE 3.2 Sum of squares from Example 3.3 Vari Sum of squares % 20.25 ae Increment Increment due to Xs te due 10 X: o Xs and Xs 265 Xpand X 265 Residual 15 Residual 15 *See Problem 3.3.80 ECONOMETRIC METHODS methods of assessing the relative importance of different explanatory variables.? He pays most attention to averaging the squared simple and partial correlation coeffi- cients over the various possible orders in which the X’s might be introduced. The rationale is that at each stage the relevant squared coefficient indicates the propor- tion of the remaining variance that is explained by a specific X variable. For this example his method gives ha + MagW2 (0.9143 + 0.8067)/2 Average proportion for X3 = (r}3 + r732)/2 = (0.7232 + 0.3758)/2 = 0.55 Kruskal has some reservations about applying this technique in a regression frame- work, but his article does contain an interesting application of the method to data from Friedman and Meiselman on the perennial topic of the relative importance of money and autonomous expenditure in the determination of income. A different way of illustrating the relative contributions of the explanatory variables is a Tinbergen diagram. These diagrams were used extensively in Tin- bergen’s pioneering study of business cycles.'° The appropriate diagram for this numerical example is shown in Fig. 3.1. Working with deviations, Fig. 3.14 shows Average proportion for X2 0.86 6 rk byste ‘ .? ae & 2 2 4 Observation woh aol @ (o) af a 2k 2 Zo a, +o Observation Sale Observation ar aif wih © @ FIGURE 3.1 Tinbergen Diagram for Example 3.3 *William Kruskal, “Relative Importance by Averaging over Orderings,” The American Statistician, 1987, 41, 6-10 "J, Tinbergen, Business Cycles in the United States of America, 1919-1932, League of Nations, 1939.cHarreRs: The &-Variable Linear Equation 81 the actual and calculated y values; parts (b) and (c) show by x> and b33x3, respectively: and part (d) shows the residuals from the regression. All parts should have the same vertical scale to facilitate visual comparisons. This diagram and the Kruskal average coefficients both suggest a larger role for X> than for X; in the determination of ¥ in this example, 3.2.2 Partial Correlation Coefficients and Multiple Regression Coefficients There are two regression slope coefficients in the three-variable equation y= box thx te Alternatively, one could obtain a slope coefficient from the regression of e1.3 on e23, and another slope coefficient from the regression of e),2 on e3.. Let us denote these two coefficients by b12.3 and by3,2, respectively, since they come from the series used to calculate the corresponding partial correlation coefficients, What is the relation of these latter regression coefficients to b> and b3 from the multiple regression? The answer is, they are identical: bj23 = bz and b132 = 63. The multiple regression coefficients come from the normal equations (a, SP] Ee] Solving for by shows TA dyn -Tawd yx YBTA- (Lay? we find Deze. _— DO = braxs\(ar - basxs) Dey Dl = bax b= From the first pair of residuals bi23 = Algebraic simplification of this last expression gives by23 = bp. The equality of the other two coefficients may be shown in a similar fashion. In the early days of econometrics there was some confusion over the use of time in regression analysis. The preceding result shows that it does not matter whether time is included among the explanatory variables or whether the variables are “de- trended” before being put into the regression. Suppose, for example. that a demand function is specified as Q = APS eB:T where Q measures the quantity demanded, P indicates price, and T denotes time. The price elasticity is 82, and the rate of shift in the quantity demanded per unit of time is 3. Taking logs gives InQ = B; + ByInP + B3T ‘The price elasticity may be estimated direcily by fitting this multiple regression, or by removing a linear time trend from both In@ and In P and estimating the slope82. ECONOMETRIC METHODS of the regression of the first residual on the second. Notice, however, that neither of the coefficients of time in the separate trend analyses is an estimate of the 83 shift parameter.!! 3.2.3 General Treatment of Partial Correlation and Multiple Regression Coefficients Under the conditions shown in the next section, the normal equations solve for b = (X'X)|X'y. The residuals from the LS regression may then be expressed as e =y—Xb = y- X(X'X) ''X'y = My (3.17) where M =1-X(X'X)'X' Itis readily seen that M is a symmetric, idempotent matrix. It also has the properties that MX = 0 and Me = e. Now write the general regression in partitioned form as yak Xl) [p|+e In this partitioning x2 is the » 1 vector of observations on X, with coefficient bp, and X, is the n x(k ~ 1) matrix of all the other variables (including the column of ones) with coefficient vector b,2).'? The normal equations for this setup are ixhxr x5X-][ bo] _ [xy Xe. XX.| (be) (XY. We wish to solve for b> and interpret the result in terms of a simple regression slope. The solution is!® by = (Man) My) (3.18) =1- XXX) X, M. is a symmetric, idempotent matrix with the properties M.X. = 0 and Me = e. Now by analogy with Eq. (3.17) it follows that where M, My is the vector of residuals when y is regressed on X. and ‘M.x; is the vector of residuals when x) is regressed on X. Regressing the first vector on the second gives a slope coefficient, which, using the symmetry and idempotency of M., gives the b coefficient already defined in Eq. (3.18). This general result has already been illustrated for the three-variable case. A simpler and elegant way of proving the same result is as follows. Write the partitioned regression as Y = Xb. + Xba, +e "See Problem 3.4 Note that this is a different use of the star subscript than in an earlier section where it was used to indicate data in deviation form. "See Appendix 3.2cuaprers: The k-Variable Linear Equation 83 Premultiply by M. to obtain My = (Mx)by +e Finally, premultiply by x}, which gives xyMay = (xMox2)by which replicates Eq. (3.18). The partial correlation coefficient between ¥ and Xp, conditional on X3, is defined as 1123,.k = correlation coefficient between (My) and (M.x2) x\Moy Ve\Mox2 Jy Moy Comparison of Eq. (3.19) with Eq, (3.18) gives (3.19) i by = 1234.8 VeiMx, $1.34. = Nase 3.20) $2.34..k where 51.34. is the standard deviation of the residuals from the regression of ¥ on a constant and X5, ..., X,iand s2 34, «is the standard deviation of the residuals from the regression of Xo on the same variables. Equation (3.20) is the multivariate version of Eq. (1.30) for the two-variable model. The other partial correlation coefficients and multiple regression coefficients may be derived by replacing x2 by x; (i in Eqs. (3.19) and (3.20) and making the corresponding changes in M.. a THE GEOMETRY OF LEAST SQUARES The simplest case is shown in Fig. 3.2, with ay vector and an x vector for a single explanatory variable. The two vectors lie in E”, that is, n-dimensional Euclidean space.'* The y vector may be expressed asy = +e, where = bx is some multiple y $ FIGURE 3.2 e * To distinguish between the use of the same letter for Euclidean space and for the expectation operator ‘we will use bold E” for the former and italic, nonbold £ for the latter.84 ECONOMETRIC METHODS. of the x vector. Three possibilities are shown in the figure. The least-squares principle is to choose b to make § as close as possible to y. This is achieved by making the length of e a minimum, that is, by dropping a perpendicular from y to x. As shown in Appendix A, the condition for x and ¢ to be orthogonal is x'e = 0. This gives xy — bx) = 0, rb = x'yix'x. Then 5 =x0 = (22) = (\y = @er'sy x’ ) Gw'xy Ie'y =Py where P= x(x'x) lx’ Notice that xx’ is an n X n matrix, whereas x'x is a scalar. The matrix P is seen to be symmetric. idempotent. It is called a projection matrix, since postmultiplication by y gives the projection of the y vector onto the x vector. Figure 3.3 shows the case where there are two explanatory variables, with vec- tors.x, and x2. All linear combinations of these two vectors define a two-dimensional subspace of E”. This is the column space of X. The residual vector must be perpen- dicular to this column space, which requires e to be orthogonal to both x; and x, which may be stated compactly as X'e=0 This condition was derived algebraically in Eq. (3.5), which in turn gives the normal equations (X'X)b = X'y. From the parallelogram law for the addition of vectors it is clear from Fig. 3.3 that | may be expressed as a unique linear combination of x1 and x2. The equivalent algebraic condition is that the normal equations solve for a unique b vector. The x vectors in Fig. 3.3 are linearly independent; thus the column. space of X has dimension two, which is the rank of X. As shown in Appendix A, Rank of X = rank of (X'X) = rank of (XX') Thus (X'X) is nonsingular, and the normal equations solve for = (X'X)"!X'y. Fig- ure 3.4 shows a case where the two x vectors are linearly dependent, The § point is still uniquely determined by the perpendicular from y to the line through the x vec- tors, but there is no unique representation of § in terms of x1 and xp. y FIGURE 3.3cuaprer s: The k-Variable Linear Equation 85 FIGURE 3.4 a g a In the general case the X matrix is a set of k column vectors Kaley xp om Any linear combination of these vectors lies in the column space of X. Ify, the vector of observations on the dependent variable, were also in the column space, then there would be at least one k X | vector b satisfying y = Xb. The linear combination of the explanatory variables would account for all of the variation in Y, leaving zero unexplained variation. Such an outcome is totally unlikely in practice; the typical situation is shown in Fig. 3.5, where the y vector lies outside the column space of X. By denoting any arbitrary linear combination of the columns of X by 9 = Xb, the y vector can be expressed as y = § + e. The least-squares principle is to choose § to minimize the length of the e vector. This is achieved when the and e vectors are orthogonal. Since § is a linear combination of the columns of X. this requires that e be orthogonal to each x; vector, giving xJe=0 f=1,2, or, written more compactly, Xe=0 which is the same equation as that derived earlier for the case of just two explana- tory variables. If the columns of X are linearly independent (ie.. X has full column rank) then can be expressed as a unique linear combination of the x; vectors (the ’ FIGURE 3.586 ECONOMETRIC METHODS normal equations solve for a unique 6). If, however, the column vectors are linearly dependent, § is still given uniquely by the perpendicular from y to the column space, but one can find one or more ¢ vectors satisfying Xe = 0. Then 5 = Xb = Xb+Xe =Xb+0) which says that § cannot be expressed as a unique linear combination of the x;’s, and so the normal equations do not solve for a unique 6. To sum up, least squares requires that X has rank k, so that (X'X) is nonsingular and the normal equations solve for a unique b. 34 INFERENCE IN THE k-VARIABLE EQUATION Next we need to establish the statistical properties of the least-squares estimator and to derive appropriate inference procedures. These depend on what assumptions are made in the specification of the relationship. 3.4.1 Assumptions 1. X is nonstochastic and has full column rank k. Inferences will be conditional on the sample values of the X variables, so the elements of the X matrix are treated as fixed in repeated sampling. As shown in Section 3.3, linear independence of the columns of X is required for a unique determination of the & vector. 2. The disturbances have the properties E(u) = 0 (G.21) and var(u) = E(uu') = 07 (3.22) When the expectation operator E is applied to a vector or matrix, it is applied to every element in that vector or matrix. Thus, Eq. (3.21) gives my} fEGy)] 10 mw} |E@2)| _ |0 Eu) = E ; =0 wnt Lewes Lo, and Eq. (3.22) gives 4 Eup) Elna) ++ Elita 2) owe Buu’) = £4}? |tu up ug) = Fam) Be) E(uzun) tn Elui) Elune) var(ur) Covi, Hr) = COV(t, Mn) o 0 _feoviwa, ti) varia) + cov(uarand\ [0 o? OV(Un, U1) COV(Un, 2) var(n) 0 0cuapter3: The k-Variable Linear Equation 87 This matrix is the variance-covariance matrix of the disturbance term. The variances are displayed on the main diagonal and the covariances in the off-diagonal positions, We will denote it by the shorthand notation var(u). Itis sometimes indicated by V(t) or by cov(u).. This variance matrix embodies two strong assumptions. The first is that the disturbance variance is constant at each sample point. This condition is termed ho- moscedasticity, and the obverse condition, where the disturbance variances are not the same at all points, is termed heteroscedasticity. The second assumption is that the disturbances are painvise uncorrelated. In a cross-section study of, say, the expenditure patterns of households, this implies zero covariances between the dis- turbances of different households. With time series data, the implication is zero covariances between disturbances at different time periods. When this condition fails, the disturbances are said to be autocorrelated or serially correlated. 3.4.2. Mean and Variance of b For theoretical derivations it is simpler to rewrite the normal equations as b= (XXX Substituting for y gives b= (XX) 'X(KB +0) = B+ (XXX from which b-B=(X'X)'Xu (3.23) In taking expectations the expectation operator may be moved to the right past non- stochastic terms such as X, but must be applied to any stochastic variable. Thus, E(b— B) = (X'X)'X'EW) = 0 giving E@) = B (3.24) Thus, under the assumptions of this model, the LS coefficients are unbiased esti- mates of the parameters. The variance-covariance matrix of the LS estimates is established as follows. From first principles, as in the development of Eq. (3.22), var(b) = E[(b ~ B)(b— B)'] If we substitute from Eq. (3.23), El — Bb — B)') = EUX'X) Xu’ X(X'X)!) = (XX) 1X'Eluu'|X(X'X)! o(X'Xy! Thus, var(b) = 0°(X'X)! (3.25) This expression is a k k matrix with the sampling variances of the b; displayed on the main diagonal and the covariances in the off-diagonal positions. EXAMPLE 3. STANDARD ERRORS IN THE TWO-VARIABLE EQUATION. As al- ready shown in Example 3.1, X°X in this case is88 ECONOMETRIC METHODS wre [ty Fe] Thus, wx ae al where D is the determinant of (X'X), 2 DeanDe-(Sxf ene ‘Then, denoting the LS intercept and slope by a and b, respectively, we have ze 1, ots S at SH which confirms Eq. (1.42). The square roots of these variances are frequently referred to as standard errors. They are the standard deviations of the marginal distributions of a * and b.'’ Finally it may be seen that x cov(a, b) = —o* =e which confirms Equation (1.43). EXAMPLE3.6. A THREE-VARIABLE EQUATION. In most economic applications in- terest centers on the regression slope coefficients rather than on the intercept term. One ccan then work with the data in deviation form. Expressions like Eq. (3.25) still hold, and the problem becomes two-dimensional rather than three-dimensional. Thus, vaniby = | Y2t(b2) —cov(ba,bs)] 24 Lexy]! Jcov(b:, bs) —_var(bs) Yann Day Some algebra then shows that var(b:) = and var(bs) = a THC Sxa- where rz; is the correlation between X> and X3. If the explanatory variables are uncor- related, these sampling variances reduce to those for the simple regressions of ¥ on X2 and ¥ on X3. However, increasing correlation between the explanatory variables inflates the standard errors beyond those that would pertain in the orthogonal case. The more the 7h) 7) "These formulae are nonoperational since? is unknown. When it is replaced by the estimators derived in Section 3.4.3, we have estimated standard errors. The term standard error is thus used in- terchangeably to refer to true or estimated standard errors, the choice usually being obvious from the context.cuavters: The k-Variable Linear Equation 89 X’s look alike, the more imprecise is the attempt to estimate their relative effects, This situation is referred to as multicollinearity or collinearity. With perfect or exact collinear~ ity the standard errors go to infinity. Exact collinearity means that the columns of X are linearly dependent, and so the LS vector cannot be estimated, 3.4.3 Estimation of 0? The variance-covariance matrix in Eq. (3.25) involves the disturbance variance 02, which is unknown. Itis reasonable to base an estimate on the residual sum of squares from the fitted regression. Following Eq. (3.17), we writee = My = M(XB + u) Mu, since MX = Thus, E(e'e) = E(u'M'Mu) = E(u'Mu) Utilizing the fact that the trace of a scalar is the scalar, we write E(u'Mu) = Elt(u'Mu)] E{tr(uu'M)) = o*tr(M) = ot — a tr X(X'X) |X") = ot — 07 trf(X'X)'(X'X)] = o(n-b) ee SS 6.26) Thus, defines an unbiased estimator of a7. The square root s is the standard deviation of the ¥ values about the regression plane. It is often referred to as the standard error of estimate or the standard error of the regression (SER). 3.4.4 Gauss-Markov Theorem This is the fundamental least-squares theorem. It states that, conditional on the as- sumptions made, no other linear, unbiased estimator of the 8 coefficients can have smaller sampling variances than those of the least-squares estimator, given in Eq. (3.25). We prove a more general result relating to any linear combination of the B coefficients. Let ¢ denote an arbitrary k-element vector of known constants, and de- fine a scalar quantity p. as wa=cB If we choose c' = (0 1 0 +++ OJ, then y = By. Thus, we can pick out any single element in B.. If we choose =U Xone Xanti 1° Xensil then w= E(¥n+i)90 ECONOMETRIC METHODS which is the expected value of the dependent variable Y in period n + 1, conditional on the X values in that period. In general, 2 represents any linear combination of the elements of B. We wish to consider the class of linear unbiased estimators of 1. Thus, we define a scalar m that will serve as a linear estimator of jz, that is, m= a'y =a'XB +a'u where a is some n-element column vector. The definition ensures linearity. To ensure unbiasedness we have E(m) = @'XB +a'Eu) =a'xp =c'B only if ax c 3.27) The problem is to find an n-vector a that will minimize the variance of m, subject to the k side conditions given by Eq. (3.27). The variance of m is seen to be var(m) ea'a where the derivation uses the fact that, since a’w is a scalar, its square may be written as the product of its transpose and itself. The problem is then to find @ to minimize a'a subject to X'a = . The solution is'® E(a'uu'a) a=XXX)'e which yields m=a'y = (XX) IX'y =c'b This result specifically means the following: 1. Each LS coefficient 6; is a best linear unbiased estimator of the corresponding population parameter Bj. 2. The best linear unbiased estimate (BLUE) of any linear combination of ’s is that same linear combination of the b's. 3. The BLUE of E(¥,) is Py = by + brXrq + bsX3p + 00+ + DX ts which is the value found by inserting a relevant vector of X values into the re- gression equation. 3.4.5 Testing Linear Hypotheses about B We have established the properties of the LS estimator of B. It remains to show how to use this estimator to test various hypotheses about B. Consider the following examples of typical hypotheses. "See Appendix 3.3.caper 3: The k-Variable Linear Equation 91 (0 Ho: Bi = 0. This sets up the hypothesis that the regressor X; has no influence on ¥. This type of test is very common and is often referred to simply as a significance test. (i) Ho: B; = Bjo. Here Bio is some specified value. If, for instance, B; denotes a price elasticity one might wish to test that the elasticity is —1 (iii) Ho: By + B3 = 1. If the B's indicate labor and capital elasticities in a pro- duction function, this formulation hypothesizes constant returns to scale. (iv) Ho: Bs ~ Bs, of Bs — By = 0. This hypothesizes that X3 and X, have the same coefficient. () Ho: By 0 Bs) _|o Bul 0. This sets up the hypothesis that the complete set of regressors has no effect on Y. It tests the significance of the overall relation. The intercept term does not, enter into the hypothesis, since interest centers on the variation of Y around its mean and the level of the series is usually of no specific relevance. (vi) Ho: Bz = 0. Here the B vector is partitioned into two subvectors, By and Bz, containing, respectively, ky and ky(= k ~ ky) elements. This sets up the hypothesis that a specified subset of regressors plays no role in the determi- nation of ¥. All six examples fit into the general linear framework RB =r (3.28) where R is a q X k matrix of known constants, with q < k, and r is a q-vector of known constants. Each null hypothesis determines the relevant elements in R andr. For the foregoing examples we have @ R= - 010 with 1 in the ith position. Gi) R=(0 - 0 1 0 -+ Of with 1 in the ith position, Gi) R=[0 11 0 + O} (iv) R=[0 0 1 -1 0 + O} @) R= Ta) where 0 is a vector of k — 1 zeros. (i) R= 0rxK Te] r=0 qzkh The efficient way to proceed is to derive a testing procedure for the general linear hypothesis 0 The general test may then be specialized to deal with any specific application. Given the LS estimator, an obvious step is to compute the vector (Rb ~ r). This vector measures the discrepancy between expectation and observation. If this vector is, in Hy: RB-r92 ECONOMETRIC METHODS some sense, “large,” it casts doubt on the null hypothesis, and conversely, if it is “small” it tends not to contradict the null. As in all conventional testing procedures, the distinction between large and small is determined from the relevant sampling distribution under the null, in this case, the distribution of Rb when RB = r. From Eq, (3.24) it follows directly that E(Rb) = RB (3.29) Therefore, from Eq. (3.25) var(Rb) = E[R(b — B)(b - B)'R'] = Rvar(b)R' = PRX'X)'R' (3.30) ‘We thus know the mean and variance of the Rb vector. One further assumption is re- quired to determine the form of the sampling distribution. Since b is a function of the u vector, the sampling distribution of Rb will be determined by the distribution of u. The assumptions made so far about w are given in Eqs. (3.21) and (3.22). By making the additional assumption that the ; are normally distributed, all three assumptions may be combined in the single statement u~ N01) 31) Since linear combinations of normal variables are also normally distributed, it fol- lows directly that b ~ N(B,0(X'X)"] (3.32) Then Rb ~ N[RB, 0? R(X'X)"'R') (3.33) and so Rb ~ B) ~ N(0,0?R(X'X)'R'] (3.34) If the null hypothesis RB = r is true, then (Rb ~ r) ~ N[O, 0? ROX'X)"'R’] (3.35) This equation gives us the sampling distribution of RB; and, as seen in Appendix B, we may then derive a x? variable, namely, (Rb ~ 7) '(o?7ROCX)'R'T (RB ~ 1) ~ 7Q) (3.36) The only problem hindering practical application of Eq. (3.36) is the presence of the unknown o?, However, it is shown in Appendix B that S ~¥a-b 37) and that this statistic is distributed independently of b. Thus, Eqs. (3.36) and (3.37) may be combined to form a computable statistic, which has an F distribution under the null hypothesis, namely, (Rb = [ROX RT RG - lg eat Fqn-b (3.38)carrer s: The k-Variable Linear Equation 93 The test procedure is then to reject the hypothesis RB = r if the computed F value exceeds a preselected critical value. Now we must see what this test procedure amounts to in the specific applications indicated previously. For some cases it is helpful to rewrite Eq. (3.38) as (Rb — r)'[s*R(X'X)'R') (RB — r)/q ~ F(qen - &) 3.39) where s? was defined in Eq. (3.26). Thus, s?(X'X)~! is the estimated variance- covariance matrix of b. If we let cj; denote the i, jth element in (X'X)~! then Sei = var(b;) — and sey = cov(bi bj) i, f = 1,2... In each application the specific forms of R and r are substituted in Eq. (3.38) or 3.39). ( Ho: Bi = 0. Rb picks out b; and R(X'X)-R' picks out ci, the ith diagonal element in (X'X)~'. Thus Eq. (3.39) becomes bb bf = Bi > Tatby 7 FeO of, taking the square root, bi 1 Sa 3eAbo) ~t(n-k ‘Thus the null hypothesis that X; has no association with Y is tested by dividing the ith estimated coefficient by its estimated standard error and referring the ratio to the ¢ distribution i) Ho: Bi = Bio. Ina lar fashion this hypothesis is tested by bi- Bo | s.elbi) Instead of testing specific hypotheses about B; one may compute, say, a 95 percent confidence interval for B;. It is given by t(n— k) bj © ton58.0.(bi) (iii) Hy: Bo + Bs = 1.Rb gives the sum of the two estimated coefficients, b2 +3 Premultiplying (X'X)~! by R gives a row vector whose elements are the sum. of the corresponding elements in the second and third rows of (X"X)~'. Form- ing the inner product with R' gives the sum of the second and third elements of the row vector, that is, ¢2 + 2c23 + ¢33, noting that C23 = c3. Thus, RS(X'X)'R! = (cx + 2en + 033) = var(by) + 2cov(bo, bs) + var(bs) = vat(b> + bs) The test statistic is then p= Hatbhs—l) un—k) /var(by + b3)94 ECONOMETRIC METHIODS Altematively one may compute, say, a 95 percent confidence interval for the sum (Bz + B3) as (bz + b3) © toons vvar(by + b3) (iv) Ho: Bs = Bs. Ina similar fashion to the derivation in (iii), the test statistic here is b3 — by ~ tn - k) (v) Ho: Bo = Bs . Bx = 0. The first four examples have each involved just a single hypothesis, which is why we had a choice of equivalent F and 1 tests. This example involves a composite hypothesis about all k— | regressor coefficients. Now R(X'X)~'R' picks out the square submatrix of order k—- 1 in the bottom right-hand corner of (X’X)~!. To evaluate this submatrix, partition the X matrix as [i X>] where X> is the matrix of observations on all k — 1 regressors. Then wel@]y yr ef #& xx= [fp] a1 [ie x38] From the formula in Appendix A for the inverse of a partitioned matrix, we can then express the required submatrix as (XX — X3in7'7Q)"! = QAM)! = VOX.) where A is the matrix, already defined in Eq. (3.7), which transforms observa- tions into deviation form, and X, = AX). Rb = bp, which is the vector of LS estimates of the coefficients of the k— I regressors. Apart from the divisor q, the numerator in Eq. (3.38) is then &5X”X,b2, which has already been shown in Eq. (3.9) to be the ESS from the regression. Thus the F statistic for testing the joint significance of the complete set of regressors is _ ESSMk- 1) F = ssn 7 FEW ba ® (3.40) By using Eq. (3.10), this statistic may also be expressed as — _RUK-V on P= Gym 7 FRO be 8) G41) The test essentially asks whether the mean square due to the regression is significantly larger than the residual mean square. (vi) Ho: Bz = 0. This hypothesis postulates that a subset of regressor coeffi- cients is a zero vector, in contrast with the previous example, where all regres- sor coefficients were hypothesized to be zero. Partition the regression equation as follows: Xb, + Xob2 +e 1x xaifp!] +cuarreR3: The &-Variable Linear Equation 95 where X; has ky columns, including a column of ones, X2 has k (= k ~ ki) columns, and 6; and 63 are the corresponding subvectors of regression coef- ficients. The partitioning of the X matrix gives XX) xe 4X = Nyy, XG R(X’X) 'R' picks out the square matrix of order kp in the bottom right-hand corner of (X'X)~!. Ina similar fashion to Example (¥), this may be shown to be (XM) X2)"! where My = I~ X,(X}X1)~!X}. This is a symmetric, idempo- tent matrix of the type already defined in Eg, (3.17). It also has the properties that M\X; = Oand Mye = e. Further, My gives the vector of residuals when 9 is regressed on X,. The numerator in Eq. (3.38) is then BS(X3M 1X2 )bo/ ky To appreciate what is measured by this numerator, premultiply the partitioned regression equation by Mj to obtain Miy = M\X2b) +e Squaring both sides gives Y'Myy = b5(X)MX2)b2 + ee The term on the left of this equation is the RSS when y is regressed just on X). The last term, e'e, is the RSS when y is regressed on [XX]. Thus the middle term measures the increment in ESS (or equivalently, the reduction in RSS) when Xp is added to the set of regressors. The hypothesis may thus be tested by running two separate regressions. First regress y on X, and denote the RSS by e(e.. Then run the regression on all the Xs, obtaining the RSS, denoted as usual by e'e. From Eq, (3.38) the test statistic is = ler —e'ei/ks Dolman 7 Fkun 8) 42) 3.4.6 Restricted and Unrestricted Regressions Clearly, (v) is a special case of (vi), and so (v) may also be interpreted as the outcome of two separate regressions. Recall from Eq. (3.9) that ESS may be expressed as ESS = yly. ~e’e, where y. = Ay with A defined in Eq. (3.7), It may be shown that yY« is the RSS when y, is regressed on x; (= i).7 With this substitution for ESS in Eq. (3.40), that test statistic has the same form as Eq. (3.42). In both cases (v) and (vi) the first regression may be regarded as a restricted re- gression and the second as an unrestricted regression. Likewise ee, is the restricted RSS and e'e is the unrestricted RSS. In the restricted regression the restrictions in Ho "See Problem 3.5.96 ECONOMETRIC METHODS are actually imposed on the estimated equation. Thus, the restricted regression in (v) omits Xp, X3, ..., Xx from the regression, or equivalently, b, bs, ..., be are set to zero. In (vi) the restricted regression uses only the variables in X1. The unrestricted regression in each case uses all the variables in the X matrix. By the same argument, Example (i) is also a special case of (vi). In the restricted regression all variables except X; are used. Thus the significance test for 8 asks whether there is a significant reduction in RSS (increase in ESS) upon adding X; to the set of regressors. Students sometimes have difficulty in determining the correct value for g in these tests. It may be calculated in several equivalent ways: 1, The number of rows in the R matrix 2. The number of elements in the null hypothesis 3. The difference between the number of B coefficients estimated in the unrestricted and restricted equations 4, The difference in the degrees of freedom attaching to ee. and e’e In all six exampies, test statistics have been derived involving the 6; coefficients from the unrestricted regression. However, we have seen that in Examples (i), (v). and (vi) the test statistics may also be expressed in terms of the difference between the restricted and unrestricted RSS. In all three cases the restricted regression was easily obtained by excluding the relevant variables from the regression. The question naturally arises as to whether the tests in Examples (ii), (ii), and (iv) have a similar interpretation in terms of the difference between two residual sums of squares. This requires an examination of how to fit the restricted regression in these cases. 3.4.7 Fitting the Restricted Regression This may be done in two ways. One is to work out each specific case from first principles; the other is to derive a general formula into which specific cases can then be fitted. As an example of the first approach consider Example (iii) with the regression in deviation form, y = box + bay te We wish to impose the restriction that by + bs = 1 . Substituting the restriction in the regression gives a reformulated equation as YY = byx2 + (1 — by) x3 + ee or (y ~ 33) = ba (x2 — 343) + ee Thus two new variables, (y—.x3) and (x) ~ x3), are formed; and the simple regression of the first on the second (without an intercept term) gives the restricted estimate of bp. The RSS from this regression is the restricted RSS, e/e.. The general approach requires a b. vector that minimizes the RSS subject to the restrictions Rb. = r. To do this we set up the function $ = (y —Xb,)'(y — Xb,) — 20'(RB, — 1) where A is a q-vector of Lagrange multipliers. The first-order conditions arecuapter 3: The k-Variable Linear Equation 97 OD © oxy 42071. - RA = 0 ob. op 5a 72(Rbe -) The solution for b, is!* be = b+ (XX) IR'IR(X'X) Rr — RB) (3.43) where d is the usual, unrestricted LS estimator (X'X)~!X'y. The residuals from the restricted regression are e. =y— Xb. =y—Xb-X(b. -b) e— X(b, ~ 6) ‘Transposing and multiplying, we obtain ele. = e'e + (b. ~ bY X'X(b. ~ b) The process of substituting for (b. — 6) from Eq. (3.43) and simplifying gives ele. —e'e = (r — Rb)[R(X'X) !R'}'r — Rb) where, apart from g, the expression on the right-hand side is the same as the numer- ator in the F statistic in Eq. (3.38). Thus an alternative expression of the test statistic for Ho: Rb = ris (ele, ~ e'evlq e'e(n — k) Consequently all six examples fit into the same framework. In summary, tests of RB =r may be implemented by fitting the unrestricted regression and substituting the resultant b vector in Eq. (3.38). Alternatively, a re- stricted regression may be fitted as well, and tests based on the difference (ee. —e'e) between the restricted and unrestricted RSS, as in Eq. (3.44). The following numer- ical examples illustrate these procedures. ~Fqn-b (3.44) EXAMPLE 3.7. We will continue with the data of Example 3.3. (Ho: By ~ 0. The appropriate test statistic is t = bs/s /e, where c33 is the bottom right-hand element in (X’X)"!, From the results already obtained, bs = — 1.5, and s = JRSSim— 8 = V1.5 = J0.75. The term cx may be obtained by caleu- lating the determinant of the 3 x 3 matrix (X’X) directly and dividing the relevant cofactor by this determinant, Evaluating the determinant directly is tedious. Since adding multiples of rows (columns) to a row (column) of a matrix does not alter the determinant, it is simpler to find the determinant of the echelon matrix already obtained in the Gaussian elimination in Example 3.3, namely,98 ECONOMETRIC METHODS 5 15 25 x= |o 10 6 |=20 lo 0 oa where the determinant is simply the product of the elements on the principal diagonal. ‘The relevant cofactor is 515 [IS 55| 50 Thus ¢33 = 50/20 = 2.5 and -15 70 0.75/25 which falls well short of any conventional critical value. When a hypothesis does not involve the intercept term, it is often simpler to work with the data in deviation form, since doing so reduces the dimensionality of, the problem. In this case cs; will be the bottom right-hand element in the inverse of, 10 6 6 4. Alternatively, one may examine the problem in terms of the reduction in RSS when Xs is added to the set of regressors. From Table 3.2, e1e, — e'e = 0.9. Thus, ~ V12 = -1.095 (XLX,). Referring to Example 3.3 gives ae 1 “| pee [ Eg 25 which is the same result as before _ tence 09 * vein isn ~*? andr = JF = \/1.2 as before. (ii) Ho: By = —1. The relevant test statistic is now -15-(-) -0.5 ~0.365 s fess Jo.7s V2.5 which is insignificant, Alternatively one might compute, say, a 95 percent confi- dence interval for 3. From the t distribution, ta025(2) = 4.303. The interval is then bs = toorss.e.(bs) which is ~1.5 + 4,303 0.75 V2.5 that is -139 to 439 ‘The wide interval confirms the nonrejection of hypotheses (i) and (ii). The example is, of course, extremely simplistic, and the small sample size does not permit any sharp inferences. (iii) Ho: Bo + Bs = 0. From (X,X.)~" in case (i) var(b2 + bs) = 0.75[1 + 2.5 ~ 2(1.5)] = 0.375 giving t=- 0375 = 1.63 which is insignificant. (iv) Ho: Bx = Bs = 0. Notice carefully the distinction between this and the previous hypothesis. Substitution in Eq. (3.40) givescarrer s: The k-Variable Linear Equation 99 26.512 isi However, Foos(2, 2) = 19.00, so the hypothesis would not be rejected at the 5 per- cent level, despite the impressive R°. 17.67 35 PREDICTION Suppose that we have fitted a regression equation, and we now consider some specific vector of regressor values, CHU Ky oe Kyl ‘The Xs may be hypothetical if an investigator is exploring possible effects of dif- ferent scenarios, or they may be newly observed values. In either case we wish to predict the value of ¥ conditional on c. Any such prediction is based on the assump- tion that the fitted mode! still holds in the prediction period. When a new value ¥ is also observed it is possible to test this stability assumption. An appealing point prediction is obtained by inserting the given X values into the regression equation, giving Pp = by + byXrp to + OyXep = eb G45) In the discussion of the Gauss-Markov theorem it was shown that e’b is a best lin- ear unbiased estimator of c’B. In the present context c'B = E(Y,). Thus P; is an optimal predictor of EY). Moreover, it was shown in Eq. (3.30) that var(Rb) = Rvar(b)R'. Replacing R by c' gives var(e'b) = e' var(b)e If we assume normality for the disturbance term, it follows that cece Ywarle'b) When the unknown o? in var(b) is replaced by s?, the usual shift to the ¢ distribution occurs, giving ~NO,1) ~t(n-k) 6.46) from which a 95 percent confidence interval for E(Y f) is Py + tooass Ve(X'X) 'e G47) EXAMPLE 3.8. Let us continue with the data of Example 3.3. We wish to predict E(Y) if Xx = 10 and X3 = 10. The point prediction is Py = 4+2.5(10) — 1.5(10) = 14 [* 45 “S| 45 10 -15 “80 -15 25, Inverting (X'X) gives (X'X)"100 ECONOMETRIC METHODS. 26.7 45 -8.0 Thus, (XX) 'e=[1 10 10]] 45° 10 15 -80 -15 25 10| = 6.7 10, and s? = 0.75 as before. Thus the 95 percent confidence interval for E(Y,) is 14 + 4,303 0.75 V6.7 or 435 to 23.65 When separated from a PC one prefers not to have to invert 3 x 3 matrices. Example 3.8 may be reworked in terms of deviations, which lowers the dimension ality by one.'? Sometimes one wishes to obtain a confidence interval for Yy rather than E(Y ;). The two differ only by the disturbance w, that happens to appear in the prediction period. This is unpredictable, so the point prediction remains as in Eq. (3.45). Itis still a best linear unbiased predictor, but the uncertainty of the prediction is increased. We have ¥; = ¢'b as before, and now Yr = c'B + uy. The prediction error is thus ey = ¥,-P, = uy eb - B) The process of squaring both sides and taking expectations gives the variance of the prediction error as var(es) = a? + ¢'var(b)e = o(1+0'(X'X) 'c) from which we derive a r statistic eet ieal sJ1+e(X'X) le Comparison with Eq. (3.46) shows that the only difference is an increase of one in the square root term, Thus the revised confidence interval for the data in Example 3.8 is ~t(n- kb (3.48) 14 = 4.303 Y0.75. V7.7 or 66) eo | At APPENDIX ——— APPENDIX 3.1 To prove ris = (2 — nists JT - 13, JT Recall from Chapter 1 that for a two-variable regression, y = bx + e, in devia- tion form, b = rsy/sy and > e? = S y?(1—?) = ns3(1 — 7°), where s denotes the See Problem 3.6.charter 3: The k-Variable Linear Equation 101 sample standard deviation of the subscripted variable. Thus, Dieis = asi ri) and Se} = asi - 7) 2 Also and €2.3, = X2—~ M2303 3 After some simplification, we obtain DS erse23 = nsis(n2 ~ nisr23) risras and so SSS =oe ee Vinal 4s APPENDIX 3.2 Solving for a single regression coefficient in a multiple regression. The normal equations are yep xX. ][ be] _ fay Xi X1X.)|be)] ~ [Xy where x2 is the n-vector of observations on Xz and Xz is the n x (k ~ 1) matrix of observations on all the other right-hand side variables, including a column of ones, The scalar by is the LS coefficient of Xp, and ba) denotes the coefficients of the remaining k — 1 variables. Inverting the matrix in the normal equations gives b as by = cry) + en(Xy) where c1y is the first element in the top row of the inverse matrix and ¢12 contains the remaining k — 1 elements of the first row. From the formulae for the inverse of a partitioned matrix cn = (ena — XOX) Xin)! = Man)! where Me =1-XAXX) XL Also 12 = —(ehMox2) xh X(XLX.) | Thus, By = (yeaa) 'xhy — ()Maxr) Uh X.(XXe) XY = (x)M.x2)"'x3M.y Notice that this coefficient has two possible interpretations. The vector M.x denotes the residuals in x» when the linear influence of all the variables in X. has been re- moved, Similarly M.y denotes the vector of residuals in y when X. has been allowed for. The formula then shows that by is the slope in the regression of the latter residu- als on the former, However, the idempotency of M. ensures that the same coefficient would also be obtained by regressing y on M.x2102 ECONOMETRIC METHODS This is an example of a general result, the Frisch-Waugh-Lovell Theorem.2° Suppose the regressors are partitioned into two submatrices, X = [XX]. The regression may then be written yam xallpi] te = bs + Habs +e (A3.1) Premultiply this equation by My = I — X\(X}X.)"!X}, where My is the type of symmetric, idempotent matrix developed in Eq. (3.17). This multiplication gives My = MiXobr +e (43.2) since MX, = 0 and Mye = e. Premultiplying this equation by X} yields XGMiy = (X4MX)b> or, equivalently, (M1X2)'\(Miy) = (M)X2)'(M1X2)b2 (A3.3) This last equation shows that the subvector b) from the regression in Eq. (A3.1) may also be obtained from the regression of (My) on (MX). From the properties of My it follows that Muy = vector of r MiX2 ‘A comparison of Eqs. (A3.1) and (A3.2) shows that the residual vector is the same in each regression. Thus, the RSS may be obtained from either regression In a similar fashion one may define Mz = I — X2(X}X2)"1X) and obtain the regression equation May = MzXyb, + e. In this case y and X; have been adjusted for the linear influence of the variables in X2, and the coefficient vector from the regression of the residuals coincides with the by subvector in Eq. (A3.1). Earlier results, which have been separately derived from first principles, fol- low simply from the general theorem. If X = i, a column of ones, and X2 is the nX (k — 1) matrix of regressors, then duals when y is regressed on X; matrix of residuals when each variable in X is regressed on X M, =1-i@iy |i =1- sai) =A where A is the deviation-producing matrix defined in Eq. (3.7). Thus the slope coeffi- cients may be obtained from the regression using the variables in deviation form, and this regression also yields the same RSS as the regression with an intercept and the regressors in raw form, If X; represents time, the coefficients on the other regressors may be obtained by first removing a linear trend from all variables and regressing the trend-corrected variables, or by using the regressors in raw form and including 2°R. Frisch and F. V, Waugh, “Partial Time Regressions as Compared with Individual Trends.” Econo- ‘metrica, 1933, 1, 387-401; and M. C. Lovell, “Seasonal Adjustment of Economic Time Series,” Jour- nal of the American Statistical Association, 1963, 58, 993-1010. The theorem is discussed and applied extensively in Russell Davidson and James MacKinnon, Estimation and Inference in Econometrics, ‘Oxford University Press, 1993.‘cuarrer3: The k-Variable Linear Equation 103 time explicitly as one of the regressors. Finally, the general theorem may be used to illustrate the sequential buildup of ESS or, equivalently, the sequential reduction in RSS as more variables are added to the regression. Squaring Eq. (A3.2) gives Y'Miy = bi(XjM1Xp)b2 + e'e Now y'Miy = (Miy)'(Miy) = RSS from the regression of y on X1, and e’e is the RSS from the regression of y on [X, __X2]. Thus b5(X5M)X2)b2 measures the reduc- tion in RSS due to the addition of Xo to the set of regressors. APPENDIX 3.3, To show that minimizing a’a subject to X'a = c gives a = X(X'X)"'c. In problem a is an unknown n-vector, ¢ is a known k-vector, and X is a known nx k matrix. Let $= a'a—20(X'a~ 0) The partial derivatives are Ib _ oq : _ _yyig— rr, LC cs Equating these derivatives to zero and premultiplying the first equation by X’ gives A= (XX) 'X'a = (X'X) le = XA= XXX) , then nt = ¢'b, which is a best linear unbiased Thus If a scalar m is defined as m estimator of c’B. 8 APPENDIX 3.4 Derivation of the restricted estimator b.. Define = (y — Xb,)'(y — Xb.) — 2A'(RB. ~ 7) where A is a g-vector of Lagrange multipliers. Equating the partial derivatives of & to zero gives the equations (XX). = Xy +R'A Rb =r ‘The first equation yields be = b+ (XX RA where b is the unrestricted LS estimator. Premultiplying by R gives104 [BCONOMETRIC METHODS. Rb. = Rb +|R(X'X) RA whence A= [ROX RT | - Rb) and so b+ (X'X) IR’ [ROX) IR] | (r — Rb) PROBLEMS 3.1. 35. 3.6, Show algebraically that in general the two equations in terms of deviations in Example 3.3 must be identical with the second and third equations obtained in the first step of Gaussian elimination applied to the normal equations in terms of raw data. Derive Eq. (3.15) for ri32 from first principles. Prove Eq. (3.16) that Rj 23 —r]y = rfy2(1~ rfp). Also prove that Rj.33— is = iaa(l— 72,). Corollary: Show that if ry = 0, then Rt», = 73, + Ry Consider the hypothetical demand function InQ = B; + Bln P + BsT. Denote three of the simple regression coefficients by 5 z za = =p = =e ose Se mo Sa where q, p, and t denote In Q, In P, and T in deviation form. Prove that the LS estimate of the shift parameter fs in the multiple regression may be written bua which, in general, is not equal to either of the coefficients of time, bys and bs, in the two-variable trend analyses. What happens when bas = 0? Prove that y.y., where y, = Ay with A defined in Eq. (3.7), is the RSS when y is regressed on x; = i. Show also that the estimated coefficient of the regression is ¥. Consider an alternative setup for the regression equation where the right-hand side regressors are expressed in deviation form, that is, X = [i X.], where X. is the n X (k — 1) matrix of deviations. Show that when y is regressed on X the LS vector is of] where by is the (k — 1) element vector, by = (KX) 'Xiy = OK) Xe A point prediction may then be obtained from Py = P+ berry to + bany Show that 1 var(¥,) = 2? gt RK) tx where x, = [12 xuy],and hence rework the prediction problem in Example 3.8,cuapter3: The k-Variable Linear Equation 105 3.7. Your research assistant reports the following results in several different regression prob- Jems. In which cases could you be certain that an error had been committed? Explain, (a) R}2; = 0.89 and R?>,, = 0.86 (b) 2, = 0.227, 73, = 0.126, and R3>, = © (LPL Y)-& ayy = -1732.86 (University of Michigan, 1980) 3.8 Sometimes variables are standardized before the computation of regression coeffi- cients. Standardization is achieved by dividing each observation on a variable by its standard deviation, so that the standard deviation of the transformed variable is unity. Ifthe original relation is, say, Y = Bi + BoX2 + BsXa tu and the corresponding relation between the transformed variables is Y= BU BIKE + BiXS+ where Y* = Yisy X= Xs i= 23 ‘what is the relationship between B3, 8; and B», By? Show that the partial correlation coefficients are unaffected by the transformation. The 8” coefficients are often referred to in the statistical literature as beta coefficients. They measure the effect of a one- standard-deviation change in a regressor on the dependent variable (also measured in standard deviation units). 3.9. Test each of the hypotheses By = 1, 8: = 1, Bs = —2,in the regression model Y, = Bo + BiX1 + B2Xx + BsXx + uy given the following sums of squares and products of deviations from means for 24 observations: Dyr=0 SAx=00 SHx=30 Y= 20 Syn-7) Syn= Syn = -26 San=10 Saw =5 Sen = 15 ‘Test the hypothesis that By + B + Bs = 0. How does this differ from the hypothesis that [B; BBs] = [1 1 —2)? Test the latter hypothesis. 3.10. The following sums were obtained from 10 sets of observations on ¥, X;, and Xo: DSv=2 Smu=30 Se =40 DY = 8820 xp = 92 SXF = 163 Dyn =59 Sx. = 88 MX = 119 Estimate the regression of ¥ on Xi and X2, including an intercept term, and test the hypothesis that the coefficient of Xo is zero,106 3.1L. 3.12. 3.13. 3.14, ECONOMETRIC METHODS ‘The following regression equation is estimated as a production function for Q: InQ = 1.37 +0.632inK + 0.452InL (0.257) 0.219) R= 0.98 cov(by, by) = 0.055 where the standard errors are given in parentheses. Test the following hypotheses: (a) The capital and labor elasticities of output are identical. () There are constant returns to scale, (University of Washington, 1980) Note: The problem does not give the number of sample observations. Does this omis- sion affect your conclusions? Consider a multiple regression model for which all classical assumptions hold, but in which there is no constant term. Suppose you wish to test the null hypothesis that there is no relationship between y and X, that is, Hy: Br =" = By = 0 against the alternative that at least one of the B's is nonzero. Present the appropriate test statistic and state its distributions (including the number{s] of degrees of freedom). (University of Michigan, 1978) One aspect of the rational expectations hypothesis involves the claim that expectations are unbiased, that is, that the average prediction is equal to the observed realization of the variable under investigation, This claim can be tested by reference to announced predictions and (o actual values of the rate of interest on three-month U.S. Treasury Bills published in The Goldsmith-Nagan Bond and Money Market Letter. The results of least-squares estimation (based on 30 quarterly observations) of the regression of the actual on the predicted interest rates were as follows: ry = 0.24 + 0.941! + (0.86) (0.14) RSS = 28.56 where 7; is the observed interest rate, and r+ is the average expectation of 7, held at the end of the preceding quarter. Figures in parentheses are estimated standard errors. The sample data on r* give Dra0=10 Se Carry out the test, assuming that all basic assumptions of the classical regression model are satisfied. “= 52 (University of Michigan, 1981) Consider the following regression model in deviation form: Ye = Bite + Box + ur with sample data 493 n=10 Sy = SH= D8 DSay=30 Swy=20 Sum =03.15. 3.16. 3.17, 3.18, 3.19. ‘cuarrer3: The k-Variable Linear Equation 107 (@ Compute the LS estimates of B; and Bx, and also calculate R?. (b) Test the hypothesis Ho: B2 = 7 against Hy: Bo #7. (©) Test the hypothesis Ho: Bi = Bz = O against 41: By # 0 or Br ¥ 0. (@ Test the hypothesis Ho: B = 7B against Hy: Ba # 7B. (University of London, 1981) Given the following least-squares estimates, GQ constant + 0,92Y, + ex constant + 0.84C,- + ex constant + 0.78¥; + ey = constant + 0.55C)-1 + ea 0 calculate the least-squares estimates of B2 and Bs in Cr = Bi + Ba¥e + BC + uy (University of Michigan, 1989) Prove that R? is the square of the simple correlation between y and §, where § = X(M'N) XY. Prove that if a regression is fitted without a constant term, the residuals will not neces- sarily sum to zero, and R?, if calculated as 1 ~ e'e/(y'y ~ n¥*), may be negative Prove that A? increases with the addition of an extra explanatory variable only if the F (= #) statistic for testing the significance ofthat variable exceeds unity. If r denotes the partial correlation coefficient associated with X,, prove that 2 et tee Fedf P+df where F and # are the values of the test statis of freedom in the regression. ics for X; and df is the number of degrees An economist is studying the variation in fatalities from road traffic in different states. ‘She hypothesizes that the fatality rate depends on the average speed and the standard deviation of speed in each state, No data are available on the standard deviation, but in a normal distribution the standard deviation can be approximated by the difference between the 85th percentile and the 50th percentile, Thus, the specified model is Y = Bi + BoX2 + Bx(Xa ~ X2) + fatality rate average speed Xs = 85th percentile speed where Instead of regressing ¥ on X2 and (X3 — X2) as in the specified model, the research. assistant fits ¥ = ay tank, + 05K; tu with the following results: Y = constant — 0.24X> + 0.20X3 +e108 3.20. [ECONOMETRIC METHODS with R? = 0,62. The f statistics for the two slope coefficients are 1.8 and 2.3, respec- tively, and the covariance of the regression slopes is 0.003. Use these regression results to calculate a point estimate of By and test the hypothesis that average speed has no ef- fect on fatalities. Data on a three-variable problem yield the following results: 33000 «0 132 XX=|0 40 20 Xy=| 24 Dc - PP = 150 0 20 60. 92. (a) What is the sample size? () Compute the regression equation. (c) Estimate the standard error of by and test the hypothesis that 2 is zero. (@) Test the same hypothesis by running the appropriate restricted regression and ex- amining the difference in the residual sums of squares. (©) Compute a conditional prediction for ¥;, given x2, = —4 and xs¢ = 2. Obtain also a 95 percent interval for this prediction. If the actual value of ¥ turned out to be 12, would you think it came from the relationship underlying the sample data?CHAPTER 4 eT Some Tests of the k- Variable Linear Equation for Specification Error The least-squares technique of Chapter 3 is the workhorse of econometrics and ap- plied statistics, routinely used in the analysis of myriad data sets. Its often referred to as Ordinary Least Squares (OLS) because it is derived from the si implest set of as- sumptions about the equation. Given the assumptions of Chapter 3, the least-squares estimators have the desirable properties enumerated and can also be employed in an attractive array of exact inference procedures. However, there is a crucial question. How do we know if the assumptions underlying OLS are valid for a given data set? How do we know the properties of the unobservable disturbance term? How do we know which variables should be included in the X matrix and in what functional form? If any of the underpinning assumptions are invalid, what happens to the OLS estimators? Are they still useful in any sense, or are they seriously flawed and mis- Teading? Are there alternative estimators and inference procedures that may be more appropriate under alternative assumptions? These questions will be pursued in this and subsequent chapters. If any of the underlying assumptions are wrong, there is a specification er- Tor. Although some specification errors may have minor implications, others may be much more serious; and it is extremely important to be alert to the possibility of specification errors and to test for their presence. It will be seen in later chap- ters that more complex specifications than those underlying the OLS technique are often required and lead to the corresponding development of appropriate inference procedures. 41 SPECIFICATION ERROR The specification of the linear model centers on the disturbance vector w and the X matrix. The assumptions made in Chapter 3 were as follows: 109110 ECoNOMETRIC METHODS 41) 4; are iid (0,07) Saat (4.2a) or 4; are iid N(O, 07) san (4.26) E(Xis) = Oforall i = 1,...,kandts = 1. (4.3) X is nonstochastic with full column rank & (4.4) Assumption (4.2a) postulates that the disturbances are white noise and (4.2) that they are Gaussian white noise. Given the fixed regressor assumption, Eq. (4.3) fol- lows trivially from the assumed zero mean for the disturbance term. What might go wrong? We will indicate some of the major possibilities for de- parture from these assumptions. This outline, however, is only preliminary, and sev- eral important topics will be dealt with in later chapters. 4.1.1 Possible Problems with w 1. Assumption (4.2) holds but (4.25) does not. As already indicated in Chapter 2, this does not destroy the BLUE property of OLS, but the inference procedures are now only asymptotically valid. 2. E(uu') = diaglo} --- 03]. The variance-covariance matrix for w is diagonal with different variances on the main diagonal and zeros everywhere else, so the assumption of homoscedasticity is violated. This is the simplest form of het- eroscedasticity, frequently found in cross-section applications, although this and more complicated forms may also be found in time series applications. The de- tection of heteroscedasticity and the development of appropriate inference proce- dures will be taken up in Chapter 6. 3. E(ujm)-s) # 0.(s ¥ 0). Here the disturbances are assumed to be pairwise cor- related. In time series applications there may be strong correlations between adjacent disturbances and. perhaps, smaller correlations between disturbances further apart. Similarly in cross-section data certain units may share common disturbances. Tests for autocorrelated disturbances and relevant inference procedures will also be discussed in Chapter 6. 4.1.2 Possible Problems with X 1, Exclusion of relevant variables. Economic theory teaches that income and prices jointly affect demand, so we would not expect to obtain a good estimate of a price elasticity if we left income out of a demand equation. However, in more complicated situations, which variables should be incorporated into a relation is often not clear, and this becomes an important specification problem. 2. Inclusion of irrelevant variables. This is the converse of Problem 1. Now the maintained hypothesis contains some variables that have no business there. There are some consequences for inference procedures, but they are generally less se- rious than those pertaining to the exclusion of relevant variables.‘carrer: Some Tests of the k-Variable Linear Equation for Specification Error 111 3. Incorrect functional form. We may have an appropriate list of variables but have embedded them in an incorrect functional form. Sometimes this can still be dealt, with in the context of the linear model. For instance, a relation ¥ = f(X2,X3) might be specified as Y = Bi + BoX2 + BaXy + u or perhaps Y = Bi + BoX2 + BsX3 + yoX} + ysX} + 6(K2N3) + w The second equation allows for both a quadratic response to the regressors and an interaction effect. The interaction effect is based on a new variable, which is the product of the two regressors. Thus, the expected effect of a unit change in X is By + 2y2X2 + 5X3, so it depends on By and the current levels of both X> and X;. The expected effect of a unit change in X, likewise, will depend on the level of Xz as well as X3, If the specification error consists of using the first equation rather than the second, it can easily be corrected by adding terms in X3, X3, and (X2X3). Other times an intrinsically nonlinear specification may be required, as in some of the examples in Chapter 2. 4. The X matrix has less than full column rank. This precludes estimation of a unique 4 vector. Ofien the regressors may be close to linear dependence, in which case the OLS vector can be computed. but the elements are likely to have large standard errors. This is known as the collinearity problem. 5. Nonzero correlations between the regressors and the disturbance. This is a break- down of Assumption (4.3). It may occur ina variety of ways. As shown in Chapter 2itcan happen when a lagged value of ¥ appears as a regressor, Such a value will be uncorrelated with the current and future disturbances. but will be correlated with past disturbances. OLS estimates will now be biased in finite samples, but will be consistent and asymptotically normally distributed, since the Mann-Wald theorem, described in Chapter 2, is applicable. A more serious breakdown occurs, if a regressor is correlated with the current disturbance. The OLS estimates are then biased and inconsistent. Such a condition occurs when there are measure- ment errors in the regressor(s) or when the equation under consideration is one of a system of simultaneous equations. These situations will be discussed in later chapters. 6. Nonstationary variables. As mentioned briefly in Chapter 2 most traditional in- ference procedures implicitly assume stationary variables. When this is not the case inference procedures are nonstandard, and we enter the realm of integrated variables, cointegration, error correction models, and the rest, all to be discussed at a later stage, 4.1.3 Possible Problems with 6 The assumption implicit in Eq. (4.1) is that the B vector is constant over all actual or possible sample observations. There may, however, be sudden structural breaks in coefficients, or slow evolution in response to changing social and environmental112. EcoNomeTRic METHODS factors. It would not be reasonable to expect the elasticity of the demand for apples to be the same in the Garden of Eden as in Los Angeles toward the close of the twentieth century. That circumstance, however, would not preclude the development of a demand function that would have useful practical applications in the current situation. 4.2 MODEL EVALUATION AND DIAGNOSTIC TESTS Standard econometric practice for a long time was to (i) formulate a model on the basis of theory or previous econometric findings, (i) estimate the parameters of the model using what relevant sample data one could obtain, and (iii) inspect the resul- tant estimates and associated statistics to judge the adequacy of the specified model. That inspection typically focused on the overall fit, the agreement of the signs of the coefficients with a priori expectation, the statistical significance of the coefficients, and a test of autocorrelation in the disturbances. If the model were deemed “satis- factory” on these criteria. a new equation would be added to the literature and might well be used to make predictions for data points outside the time scale or empirical range of the sample. If the estimated model were deemed “unsatisfactory,” the inves- tigator would engage in a specification search, trying out different reformulations in an attempt to reach a “satisfactory” equation, That search process went largely unre- ported, for it smacked of data mining, which was held to be reprehensible, and also because it was practically impossible to determine correct P-values and confidence coefficients for the final statistics.' In recent years there has been a substantial shift in opinion about good economet- ric practice. largely initiated by Denis Sargan of the London School of Economics, who wrote in 1975. + Despite the problems associated with “data mining” I consider that a suggested specifi- cation should be tested in all possible ways, and only those specifications which survive and correspond to a reasonable economic model should be used? This approach has been actively developed. especially by David Hendry and asso- iates.? The result is a battery of available diagnostic tests. Their use is not routine or automatic but requires judgment and economic intuition or good sense. Some tests may point to a particular specification error or errors. Others may indicate that a specification does not perform very well without locating a precise problem, and a specification may survive some tests and not others. 'See, for example, Michael C. Lovell, “Data Mining.” The Review of Economics and Statistics, LXV, 1983, 1-12. 2 J.D. Sargan, Discussion on Misspecification in Modelling the Economy, ed. G. A. Renton, Heinemann, 1975, quoted in Adrian R. Pagan, “Model Evaluation by Variable Addition,” Chapter 5, Econometrics ‘and Quantitative Economics, eds. David F. Hendry and Kenneth Wallis, Blackwell, 1984, p. 131 Por details see any recent PC-GIVE manual (David F. Hendry etal, Institute of Economics and Statis- tics, University of Oxford, UK).HAPTER «: Some Tests of the k-Variable Linear Equation for Specification Error 113 43 vi dead TESTS OF PARAMETER CONSTANCY f One of the most important criteria for an estimated equation is that it should have relevance for data outside the sample data used in the estimation. This criterion is embodied in the notion of parameter constancy, that is, that the vector should ap- ply both outside and within the sample data. Parameter constancy may be examined in various ways. One of the most useful is a test of predictive accuracy. 4.3.1 The Chow Forecast Test* If the parameter vector is constant, out-of-sample predictions will have specified probabilities of lying within bounds calculated from the sample data. “Large” pre- diction errors therefore cast doubt on the constancy hypothesis, and the converse for “small” prediction errors. Instead of using all the sample observations for estima- tion, the suggested procedure is to divide the data set of n sample observations into ‘ny observations to be used for estimation and m2 = n — ny, observations to be used for testing. With time series data one usually takes the first m, observations for es- timation and the last 2 for testing. In cross-section applications the data could be partitioned by the values of a size variable, such as household income or firm rev- enue, profits, employment, etc, There are no hard and fast rules for determining the relative sizes of nm and nz. It is not uncommon to reserve 5, 10, or 15 percent of the observations for testing. The test of predictive accuracy, widely referred to as the Chow test in honor of Chow's influential 1960 article, is as follows: 1, Estimate the OLS vector from the n, observations, obtaining by = Ok) Xin (45) where X;,y: (i = 1,2) indicate the partitioning of the data into m;, nz observa- tions, 2. Use by to obtain a prediction of the y2 vector, namely, Sa = Xabi (4.6) 3. Obtain the vector of prediction errors and analyze its sampling distribution under the null hypothesis of parameter constancy. The vector of prediction errors is d = yo ~ $2 = yn — Xa (4.7) Ifthe equation y = XB +u , with E(uu’) = 07 . holds for both data sets, the vector of prediction errors may be written d = yx ~ Xpby = uz — Xr, - B) “G.C. Chow, “Tests of Equality between Sets of Coefficients in Two Linear Regressions,” Econometrica, 52, 1960, 211-22.114 ECONOMETRIC METHODS Thus E(d) = 0, and it may be shown’ that the variance-covariance matrix for d is var(d) = E(dd') O7Tq, + Xp + var(by) + X5 (48) = 07 In, + X2(XjX1)'X3] If we assume Gaussian disturbances, d ~ N(0,var(a)] and so d'ivatayy ld ~ 20m) Further, ejer/o? ~ x°(m — kK) where e'¢1 is the residual sum of squares from the estimated regression, The two statistics are distributed independently. Thus, under the hypothesis of parameter constancy, a Fm + KK” ejeil(m; — k) Large values of this F statistic would reject the hypothesis that the same B vector applies within and outside the estimation data. The derivation of var(d) has assumed that o? is the same in each subset of data. The F test in Eq. (4.9) is therefore condi- tional on that assumption. If the disturbances are heteroscedastic, the F statistic may overstate the true significance level. ‘There is an illuminating alternative way of deriving this test of predictive ac- curacy. due to Pagan and Nicholls, following on an earlier article by Salkever.6 It also provides a simpler calculation. Suppose we allow for the possibility of a differ- ent coefficient vector in the forecast period. The complete model could then be writ- ten as : Y= XB + uy Yo = Xre +a = XB + Xe By tu =P +y +m, where y = Xx(a—B).Ify = 0.then@ = B, and the coefficient vector is constant over the estimation and forecast periods. The model is written compactly as ya X OV/B), [an Bl=be AE] +E amp If X denotes the augmented matrix in Eq. (4.11), then ry = [XIX + XGX2 XY xXx f pa : J ding F ~ F(nz,m — k) (4.9) (4.10) ‘See Appendix 4.1 “Adrian Pagan and D. Nicholls, “Estimating Pres Constructed Variables," Journal of Econometrics Dummy Variables to Compute Predictions, Prediction Errors, and Confidence Intervals, Econometrics, 4, 1976, 393-397, ion Errors and Their Standard Deviations Using 1, 1984, 293-310; and D. Salkever, “The Use of Journal ofcurren s. Some Tests of the k-Variable Linear Equation for Specification Error 115 where the np subscript on J has been omitted for simplicity. The inverse is (xy! ~ OU XG SKK 1+ QIK XY Estimating Eq. (4.11) by OLS gives | = [ (xy! (XXX ip + Xn] ie] = |- (XxX! = [ MAXX)! 1+ KAK LX XG, y2 =| %xXD xy | = Al b. ~¥0KiXi) Xi] * [a em Thus the first k OLS coefficients replicate the b; estimate of , obtained from the ‘nj data points; and the n remaining coefficients. which estimate the y vector, are simply the prediction errors defined in Eq. (4.7). The OLS estimation of Eg. (4.11) may be written fyi] _]X1 O]fbi] , fer Bl ila} fe cae However, the second equation in Eq, (4.13) is y2 = Xabi +d+er Substituting for d gives e: = 0, Thus the RSS from fitting Eq. (4.11) is simply eje1. Itis clear from Eq, (4.10) that the hypothesis of constant B is equivalent to Hy:y = 0. This hypothesis is tested by examining the joint significance of the last ‘ng variables in the augmented regression (4.13). This is a direct application of the general procedure for testing a set of linear restrictions developed in Eq. (3.38). Mak- ing the appropriate substitutions in Eq. (3.38) gives the test statistic as = d'[var(d)|"'ding ciel, — m—®) a The degrees of freedom in the denominator of this expression are obtained from (m + mz) observations less (k + nz) estimated parameters. The variance matrix vat (d) is given by 0 times the submatrix in the bottom right-hand comer of (X'X)~! and is the same as the expression already obtained in Eq. (4.8). Substitution in Eq. (4.14) replicates the Chow statistic in Eq. (4.9). Finally, an even simpler calculation is available in terms of a restricted and an unrestricted regression, As seen in Chapter 3. tests of linear restrictions can always be reformulated in this way. In the present case the unrestricted regression is, Eq. (4.13), with RSS = ¢/¢;. The restricted regression is obtained by setting ¥ to zero, that is, by estimating lyn] _ [X Bl fib +« with RSS = e/e.. The resultant test statistic for y = O is116 ECoNoMETRIC METHODS (eles ~ eer ing eeiAadtit F(m,m — &) (4.15) ‘The Chow test may thus be implemented as follows: . Using the designated m, observations, regress y on X; and obtain the RSS, ey). . Fit the same regression to all (mj + 2) observations and obtain the restricted RSS, ele. 3. Substitute in Eq, (4.15) and reject the hypothesis of parameter constancy if F exceeds a preselected critical value. pe 4.3.2 The Hansen Test” A difficulty with the Chow test is the arbitrary nature of the partitioning of the data set. One such partitioning might reject the nuil hypothesis and another fail to reject. This difficulty does not apply to the Hansen test, which fits the linear equation to all 1 observations. It allows a somewhat more general specification of the model than that used in Chapter 3. although it does rule out nonstationary regressors. The technical derivation of the test is beyond the level of this book, but it is possible to give an outline of the procedure. Write the equation as Bixis + Boxy +07 + Bax + t an (4.16) The lowercase letters now denote the actual levels of the variables, not deviations from sample means; and the first variable, 2x,,, is typically equal to one, Denote the OLS residuals, as usual, by y r= M1 Byte — bray 0 — beet The OLS fit gives the conditions t DS xuer = i 4.17) ii and Dé - 6) =0 (4.18) mi This latter condition defines an estimate of the disturbance variance as 0? = 7.1 e7/n, which differs slightly from the unbiased estimator s? = S17, e7/(n— b. Define Wit fa = | He i te Lee i Combining Eqs. (4.17) and (4.18) gives 7Bruce E. Hansen, “Testing for Parameter Instability in Linear Models,” Journal of Policy Modeling, 14, 1992, 517-533. .‘cuarreR: Some Tests ofthe &-Variable Linear Equation for Specification Exror 117 Dfe=0 Fake 4.19) = ‘The Hansen test statistics are based on cumulative sums of the fiz, namely, Su => fis (4.20) mi He develops tests for the stability of each parameter individually and for the joint stability of all parameters. The individual test statistics are Iso Ly mwa (4.21) where = ># (4.22) a For the joint stability test let Se= Ui > fasta! o ee The joint stability test statistic is then 1 oy Le = > 8iV'sy (4.23) where , ve fifi (4.24) imi Under the null hypothesis the cumulative sums will tend to be distributed around zero. Thus “large” values of the test statistics suggest rejection of the null. The dis- tribution theory is nonstandard, and only asymptotic critical values are available. These are tabulated in Table 7 in Appendix D. There is a line for each number of parameters from | to 20. The first line of the table thus gives critical values for the in- dividual coefficient test. The 5 percent critical value is 0.470, leading to a rough rule of thumb that a test value in excess of one-half suggests an unstable parameter. For five parameters, including the variance, the 5 percent critical value is approximately 1.5. The Hansen testis already incorporated in some econometric packages. 4.3.3 Tests Based on Recursive Estimation. The model in Eq. (4.1) may be written yexbtuy 1-1 (4.25) where y, is the sth observation on the dependent variable, and x; = [1 x2, +++ xu]118 ECONOMETRIC METHODS is the row vector of regressors at the th sample point, again using lowercase letters to denote the levels of variables. The complete sample matrix of regressors is x iL X= = s The idea behind recursive estimation is very simple. Fit the model to the first k observations. The fit will be perfect since there are only k regression coefficients to be estimated. Next use the first k + 1 data points and compute the coefficient vector again. Proceed in this way, adding one sample point at a time, until the final coefficient vector is obtained, based on all n sample points. This process generates a sequence of vectors. by. Bi1, ...,b,, Where the subscript indicates the number of sample points used in the estimation. In general, b, = (X/X) Xr (4.26) where X, is the 1 x k matrix of regressors for the first ¢ sample points, and yy is the t-vector of the first 1 observations on the dependent variable. The standard errors of the various coefficients may be computed at each stage of the recursion, except at the first step. since the RSS is zero when ¢ = k. Some software packages ini- tialize the recursive calculations at some m > k, generating the sequence Bra, Bm 1s -+b,. Graphs may be prepared showing the evolution of each coefficient, plus and minus two standard errors. Visual inspection of the graphs may suggest parameter constancy. or its reverse. As data are added, graphs sometimes display substantial vertical movement. to a level outside previously estimated confidence limits. This phenomenon is usually the result of the model trying to digest a structural change and leads one to suspect parameter inconstancy. Recursive estimation is an appealing procedure with time series data. since time gives a unique ordering of the data. How- ever, the procedure is readily applicable to cross-section data, which can be ordered by a suitable “size” variable. if required. 4.3.4 One-Step Ahead Prediction Errors By using all data up to and including period r ~ 1, the one-step ahead prediction of yr is x1b,-). The one-step ahead prediction error is thus Y= xe (4.27) From Eq. (4.8) the variance of the one-step ahead prediction error is var(vy) = 0? [1 +)(X)_ X11] (4.28) The unknown ? in Eq. (4.28) can be replaced by the residual variance estimated from the first (1— 1) observations, provided r— 1 > &. Taking the square root gives the estimated standard error of regression (S.E.R.). Plus or minus twice these recursively estimated standard errors can be plotted around the zero line and the actual prediction errors (also referred to as recursive residuals) shown on the same graph. ResidualsCharrer4: Some Tests of the k-Variable Linear Equation for Specification Error 119 lying outside the standard error bands are suggestive of parameter inconstancy. At each point the probability of the observed error under the null hypothesis can be calculated from the appropriate t distribution, as in Eq. (3.48). 4.3.5 CUSUM and CUSUMSQ Tests Scaled recursive residuals are defined as ys [1 + (XX, (4.29) tx] Under the assumptions in Eqs. (4.1) and (4.26) ww ~ N(O,07) Tt can also be shown that the scaled recursive residuals are pairwise uncorrelated.® Thus, w = N(O,07f,-1) (4.30) Brown et al. suggest a pair of tests of parameter constancy, based on these scaled recursive residuals. The first test statistic is the CUSUM quantity, We= So wile t= kta 431) ie where 6? = RSS,/(n— &) , with RSS, being the residual sum of squares calculated from the full-sample regression. W, is a cumulative sum, and it is plotted against t. With constant parameters, E(W,) = 0, but with nonconstant parameters W, will tend to diverge from the zero mean value line. The significance of the departure from the Zero line may be assessed by reference to a pair of straight lines that pass through the points (k, taJn— k) and (n, t3ayn—k) where a is a parameter depending on the significance level a chosen for the test. The correspondence for some conventional significance levels is @=001 a= 1.143 a = 0.05 a = 0.948 a@=010 a=0,850 ‘The lines are shown in Fig. 4.1. The second test statistic is based on cumulative sums of squared residuals, namely, 'R. L. Brown, J. Durbin, and J. M. Evans, “Techniques for Testing the Constancy of Regression Rela- tionships over Time," Journal of the Royal Statistical Sociery, Series B, 38, 1975, 149-192.120 EcoNoMErRIC METHODS FIGURE 41 CUSUM plot Ss tek+h.n (4.32) Under the null hypothesis the squared w’s are independent y?(1) variables. The nu- merator thus has an expected value of 1 — k, and the denominator an expected value of n ~ k. The mean value line, giving the approximate expected value of the test statistic under the null hypothesis, is ES) = nk which goes from zero att = ktounity at = 7. The significance of departures from the expected value line is assessed by reference to a pair of lines drawn parallel to the £(S,) line at a distance co above and below. Values of co from the Brown, Durbin, and Evans article for various sample sizes and significance levels are tabulated in Appendix D. Hansen (op. cit.) suggests that the CUSUM test is akin to his Ly test (of the stability of the intercept), and that the CUSUMSQ test is akin to his Ly test (of the stability of the variance).cuarrer ¢: Some Tests of the k-Variable Linear Equation for Specification Error 121 4.3.6 A More General Test of Specification Error: ‘The Ramsey RESET Test Ramsey has argued that various specification errors listed in Section 4.1 (omitted variables, incorrect functional form, correlation between X and u) give rise to a nonzero w vector.) Thus the null and alternative hypotheses are Hy:u ~ NO, 071) Hiu~N(woD w40 The test of Ho is based on an augmented regression Y=XB+Za+u The test for specification error is then a = 0. Ramsey's suggestion is that Z should contain powers of the predicted values of the dependent variable. Using the second, third, and fourth powers gives z=" P94 where § = Xb, and $? = [97,93 ... $]' ete. The first power, 9, is not included since it is an exact linear combination of the columns of X. Its inclusion would make the regressor matrix [X Z] have less than full rank. 44 A NUMERICAL ILLUSTRATION This numerical example is not econometrically realistic. It is meant merely to il- lustrate the tests outlined in the previous section. The variables are those already introduced in Chapter 1, that is, Y x2 X3 = log of per capita real disposable personal income Jog of per capita real expenditure on gasoline and oil log of the real price index for gasoline and oil The first oil shock hit in 1973.4, so for this exercise we chose a sample period from 1959.1 to 1973.3, a period for which it might seem reasonable to postulate parame- ter constancy. As shown in Fig. 4.2, consumption and income trended fairly steadily upward during this period, and price trended downward with a greater rate of de- crease in the second half of the sample period. The pairwise correlations are obvi- ously fairly high, so it may be difficult to disentangle the relative contributions of the two explanatory variables. The first 51 observations were used for estimation and the remaining 8 reserved for the Chow forecast test, The simple specification the Royal Statistical Society, Series B, 31, 1969, 350-371. See also I. B. Ramsey and P. Schmit, “Some Further Results on the Use of OLS and BLUS Residuals in Specification Error Tests,” Journal of the ‘American Statistical Association, 71, 1976, 389-390.122. ECONOMETRIC METHODS Log (variable) FIGURE 4.2 Gasoline consumption (Y), price (X2), and income (X3). Y By + Br(X2) + B3(X3) + u was employed. The results are shown in Table 4.1. They are a mixture of good news and bad news. Looking at the good news first, we see the specification appears economically sensible. The price elasticity is negative (~-0.66), and the income elas- ticity is positive (0.85). Both coefficients are well determined by conventional stan- dards, and the overall F statistic overwhelmingly rejects the null bypothesis that price and income have nothing to do with consumption. Furthermore, the specifi- cation passes the Chow test with flying colors, the F statistic being only 0.18. The one-step forecasts in Table 4.1 should not be confused with the one-step ahead pre- diction errors discussed in Section 4.2. The latter come from recursive estimation. The forecasts in Table 4.1 take the form f, = x,'b, where b is the coefficient vector estimated from the 51 observations and x; is the vector of explanatory variables in the forecast period, The forecast SE is given by s J/I + x;(X’X)-!x;, where is the estimated standard error of the regression and X is the matrix of regressors for the first 51 sample points. This section of the table tests each forecast point individually, and we sce that the two standard error range about the forecast includes each actual value of the dependent variable. Now for the bad news. Two items of bad news are already contained in Table 4.1, The column headed Instab contains the Hansen statistics for testing the sta- bility of individual coefficients. The hypothesis of stability is rejected for all three coefficients, and, not surprisingly, the joint stability test decisively rejects the nullcuarter «: Some Tests of the k-Variable Linear Equation for Specification Error 123 TABLE 4.1 OLS Regression of ¥ on X2 and X3 ————— es Present sample: 1959.1 to 1973.3 less 8 forecasts Forecast period: 1971.4 to 1973.3 Variable Coe Std Error value Instab Constant 1.1041 0.47043 -2347 og x2 =0.66234 0.14546 4.553 0.92" x3 0.84791 0.060387 14.081 ogi (#: significant at 1% level) P = 0.962856 F(2, 48) = 622.13 10.0000] = 000232118 DW = 0316 Variance instability test: 0.12107; joint instability test: 4,0695** Analysis of one-step forecasts Date Actual Forecast y-? ForecastSE —_¢-Value 19714 -7.65916 00161520 0.0249789 0.646624 19721 =7.65555 0.009076820.0264099 0.343690 1972.2 -7.65785 0,00917203 0.0276412 -0,331825 19723 -7.65144 =0.00555639 0.0263309- -0.211021 19724 =7.63462 —0.00373734 0.0251797*-0,148427 1973.1 -7,60615 00180611 0.0251543, 0.718011 19732 ~7.62518 63150 0,00631639-0.0247320 0.255304 1973.3 -7.62061 7.62581 0.00519761 0.024790 -—0.209890 ‘Tests of parameter constancy over: 1971.4 to 1973.3 Chow F(8, 48) = 0.18421 [0.9920] of parameter constancy.'? A second disturbing piece of information is the very low value of the Durbin-Watson (DW) statistic. As will be explained in Chapter 6, this indicates substantial autocorrelation in the disturbance term, so a lot of action is not being explained by the included regressors. This also vitiates the estimated and F “values. The presence of major specification error in this too simple equation is more firmly demonstrated by the recursive tests. Figure 4.3 shows the recursive residuals along with two standard error bands. The calculations underlying this figure are de- fined in Eqs. (4.27) and (4.28). A point on the graph lying outside the standard error bands is equivalent to a r statistic [v,/s.e(v,)] being numerically greater than two and thus suggestive of parameter inconstancy. There is one such point in 1966, and a number of similar points from 1968 through 1970. Figure 4.4, generated by PC-GIVE, is an alternative way of showing the same information as that given by Fig. 4.3, which is generated by EViews. The test implicit in Fig. 4.3 is ar test, whereas that implicit in Fig. 4.4 is an F test, The F statistic is the ‘square of the corresponding r statistic. However, by following Eq, (4.15), the one-step Chow test for parameter constancy through the first j observations is based on ‘The Hansen test may be inappropriate here, since Fig. 4.2 seems to suggest nonstationary vasiables. However, a more extended data sample will show price and consumption both reversing course, though income tends to move slowly upward.0.08 — Recursive residuals 22S, 0.06 F Residuals FIGURE 43 Recursive residuals and standard error bands for the gasoline equation of Section 4.4, 33 27 24 2 Sealed F-test value 12 09 06 03 1965 1970 1975 Year FIGURE 4.4 One-step forecast errors (scaled by 5 percent critical values),cuarter 4. Some Tests of the k-Variable Linear Equation for Specification Error 125 jametlun (4.33) where m is the number of observations used in the initial recursion. Under the null this statistic follows F(1, j — k ~ 1) . Dividing the F statistic in Eq. (4.33) by the 5 percent critical value from F(1, j — k— 1) gives the series plotted in Fig. 4.4. Any point lying above the horizontal line at 1 implies rejection of parameter constancy, whereas points below do not lead to rejection. As in Fig. 4.3, there is one rejection in 1966 and a group of rejections in 1968 through 1970, The three panels in Fig. 4.5 show the recursively estimated coefficients, with two standard error bands. As might be anticipated from Figs. 4.3 and 4.4, there are dramatic changes in the late 1960s, especially in the constant, C(1), and the price elasticity, C(2). In the first half of the sample, the price elasticity is not significantly different from zero, and the point estimate of price elasticity is positive. Only when data for the 1970s are included does the price elasticity turn negative, and signifi- cantly so. The income elasticity, C(3), is positive and reasonably stable. 08 Mh descLetel estaba slau — Recursive C1) eatmates ian 22 SE. — Recursive C12) estimates 22SE. ty Hie lasLgaasl saad 62 63 64 65 66 67 68 69 7071 72 73 62 63 64 65 66.67 68 69 70 7172 73 Year Year @ © Reainive Ci) eines 3 62 63 64 65 66 67 68 69 70 7172 73 Year © FIGURE 45 Recursively estimated coefficients: (a) Constant, C(1); (b) price elasticity, C(2); (c) income elasticity, C(3),126 EcONOMETRIC METHODS 40 — asim 16 Gis a squares sof == supe Seslicane ab a 3 5 —— z g a a Z a g g é 0 baat PAIR ELELELAHA] I AnrerN aU RTOTATER Ge ees eSec ches IN TD Osi aaasecraown nn Year Yeu @ () FIGURE 4.6 CUSUM tests of the gasoline equation in Section 4.4. The CUSUM tests reported in Fig. 4.6 confirm the message of the previous fig- ures. Finally. the Ramsey RESET test, using just §?, gives F = 47.2, which is a very strong indicator of specification error. The process is somewhat akin to the medical examination of a patient by a doctor who is (one hopes) skillful. Some vital signs may be in the acceptable range, even as others give disturbing readings. The doctor must assess the patient's overall state of health and what. if anything. she is fit for. More importantly, can the doctor suggest appropriate remedial measures for the serious problems? In this case it seems clear that we have a very sick patient indeed. It remains to be seen in future chapters what remedies may be available. 45 “TESTS OF STRUCTURAL CHANGE ‘The Chow forecast test leads naturally to more general tests of structural change. A structural change or structural break occurs if the parameters underlying a relation- ship differ from one subset of the data to another. There may, of course, be several relevant subsets of the data, with the possibility of several structural breaks. For the moment we will consider just two subsets of 2 and > observations making up the total sample of n = nj + np observations. Suppose, for example, that one wishes to investigate whether the aggregate consumption in a country differs between peace- time and wartime and that we have observations on the relevant variables for n, peacetime years and mz wartime years. A Chow test could be performed by using the estimated peacetime function to forecast consumption in the wartime years. How- ever, provided n> > k, one might alternatively use the estimated wartime function to forecast consumption in the peacetime years. It is not clear which choice should be made, and the two procedures might well yield different answers. If the subsets are Jarge enough it is better to estimate both functions and test for common parameters.cuarrer «: Some Tests of the k-Variable Linear Equation for Specification Error 127 4.5.1 Test of One Structural Change The test of structural change may be carried out in three different, but equivalent, ways. Let yi, Xi (i = 1, 2) indicate the appropriate partitioning of the data. The un- restricted model may be written B= Fo llBiyjee w~veoorn (434) where B; and B> are the k-vectors of peacetime and wartime coefficients respec tively. The null hypothesis of no structural break is Ho: B, = Bo (4.35) The first approach is a straightforward application of the test for linear restrictions, defined in Eqs. (3.28) and (3.38). The null hypothesis defines R = [Ix - fu] and 1 = 0. Substituting in Eq. (3.38) gives Rb —r = by — bp, where b and bp are the OLS estimates of the coefficient vectors in Eq. (4.34). Fitting Eq. (4.34) also provides the unrestricted RSS, e’e. The OLS coefficients may be written i - es 0 I xin| Ee [aie a] ba 0 XX), Xyy2 (X20) Xp ‘Thus the unrestricted model may be estimated by setting up the data as in Eq. (4.34) and running OLS estimation once, or by fitting the equation separately to the peace- time data and to the wartime data. In the latter case the two RSSs must be summed to give the unrestricted RSS, that is, e'e = e/e; + ee. Substitution in Eq. (3.38) tests the linear restrictions. We have seen in Chapter 3 that a test of linear restrictions may also be formulated in terms of an unrestricted RSS and a restricted RSS. In this case the null hypothesis gives the resiricted model as Pl eles 9 Denoting the RSS from fitting Eq. (4.36) as e1e., the test of the null is given by _ Wiles - e'eytk © “een 2k) ~ F(k,n = 2k) For a third possibility, consider an alternative setup of the unrestricted model, B= be lle. a an Now the test of Ho is simply atest of the joint significance of the last k regressors. The choice of the most convenient procedure largely depends on the software package one is using.128 ECONOMETRIC METHODS 4.5.2 Tests of Slope Coefficients Frequently in economic research interest centers on the slope coefficients and one does not wish to impose restrictions on the intercept term. To explain such tests, partition the X matrices as X=[h Mi] Male Hi] where i, i2 are ny and nz vectors of ones, and the X} are matrices of the k — 1 regressor variables. The conformable partitioning of the B vectors is Bi =a: Bi] Bi=(a2 Be) a ‘The null hypothesis is now Bj = B3 The unrestricted model is ay Mi i 0 X O}\a> ble 2 allele oe B; ‘The restricted model is nef Oo MD 4.39) 4 li i Xy eo ‘ ‘The test of the null can be based on the RSS from these two regressions. Some re- gression packages automatically supply an intercept term (by inserting a column of nes in the regressors). This step must be suppressed in fitting Eqs. (4.38) and (4.39) to avoid generating linearly dependent regressors. An alternative formulation of the unrestricted model is ay p]=fk @ olor Bil-[e a x alle] a 3 Bi The test of the joint significance of the last k— 1 regressors in this setup is a test of the null hypothesis. The same caveat about the intercept term applies to the estimation of Eq. (4.40). 4.5.3 Tests of Intercepts It might appear that a test of Ho: a1 = a2 is given by testing the significance of the second regressor in Eq. (4.40). However, such a test would normally make little sense. Since the estimation of Eq. (4.40) places no restrictions on the slope coefficients, the hypothesis of equal intercepts is then asking whether two different regression surfaces intersect the y axis at the same point. A test of differential in- tercepis makes more sense if it is made conditional on the assumption of commonCHAPTER 4: Some Tests of the k-Variable Linear Equation for Specification Error 129 regression slopes. It now amounts to asking whether one regression plane is higher or lower than the other in the direction of the dependent variable. The unrestricted model for this test is that already specified in Eq. (4.39), where it appeared as the restricted model in the test of slope coefficients. The restricted model is ia Z tt lel (441) Contrasting RSS between Eqs. (4.39) and (4.41) then provides a test of equality of intercepts, given equal regression slopes. The alternative setup of the unrestricted model is : ya yi} fi o xf ™ al (i x a +H (4.42) Now a test of the significance of the second regressor tests the conditional hypothesis that the intercepts are equal. 45.4 Summary There is a hierarchy of three models, namely, Bl Sele« Common parameters np 9% a aa Differential intercepts, 2" 0 i Xs]| fe common slope vectors a mf] =f ® Xf Ofer], Differensat imerceprs, ‘bal lo & 0 x3]/ BF differential slope vectors B: Application of OLS to each mode! will yield a residual sum of squares, RSS, with associated degrees of freedom, respectively, of n — k,n — k — I, and n — 2k. The test statistics for various hypotheses are then as follows: Ho: a) = a Test of differential intercepts RSS) ~ RSS. ee Fm Resa kop 7 Pb kD) Ho: Bi = Bj Test of differential slope vectors p= RSS. RSSVE—N peg 4 24) RSS3(n— 2k) Ho: B, = Bz Test of differential parameters (intercepts and slopes) = (RSS — RSS)Vk os P= “RSSofn— 2k) ~ P24)130 ECONOMETRIC METHODS The degrees of freedom in the numerator of these expressions are the number of restrictions imposed in going from the relevant unrestricted model to the restricted model. This number is also equal to the difference in the degrees of freedom of the residual sums of squares in the numerator, 45.5 A Numerical Example Suppose we have the following data: x m=5 m=10 ne a= v= = Bsaan Renee COV ARV Ss 8 From these data we form the following n-vectors: tn _fo a=) «-(2] _)] 2-/*) gef) yen a-fe] e-[e] «-B) o-B » The hierarchy of three models is then as follows: I: Regression of y oni and x Il: Regression of y on dy. d>, and x Hil: Regression of y on d).d>. 21. and 2) Table 4.2 shows these three regressions. From the third part of the table the separate regressions are 5) = -0.0625 + 0.4375x, So = 0.4000 + 0.5091x2 In each regression the RSS is given by the entry labeled Sum of squared resid. The test of equality of the B vectors is thus 3.160015 = 4) From the tables of the F distribution, Foss ll) = 3.98 and Fas9(2)11) = 7.21cuarter é: Some Tests of the k-Variable Linear Equation for Specification Error TABLE 4.2 LS /! Dependent Variable is ¥ VARIABLE COEFFICIENT _STD.ERROR _TSTAT. —_2-TAIL SIG. c =0.0697842 03678736 —0.1896961. 0.8525 x 05244604 0.032917 15,896733 0.0000 Resquared 0.951074 Mean of dependent var 5.000000 Adjusted R-squared 0.947310 SD. of dependent var 3.093773 SEE. of regression 0.710152 Sum of squared resid 6.556115 Log likelihood =15.07669 statistic 252.7061 ‘Durbin-Watson stat 1.343686 __Prob(F statistic) 0.000000 LS// Dependent Variable is Y VARIABLE COEFFICIENT STD.ERROR STAT. —_>-TAILSIG. DI =0.4658537 03048463 —1.5281589 0.1524 Dz 0,5536585 0.390025 1.633192 0.1284 x 0.495120 0.026345 18589463, 0.0000 R-squared, 0.973953 Mean of dependent var 5.000000 Adjusted R-squared 0.969612 S.D. of dependent var 3.093773 S.E. of regression 0.539309 Sum of squared resid 3.490244 Log likelihood 1034849 F-statistie 228.3564 ‘Durbin-Watson stat 2.462522 Prob(F-statistic) 0.000000 LS /{ Dependent Variable is Y VARIABLE COEFFICIENT STD.ERROR —TSTAT. _2-TAIL SIG, Di -0.0625000 04831417 =0.1293616 0.8994 D2 0.400000 0.3661560 10924304 0.2980 ZI 0.437500 0.0599263, 73006283 0.0000 . 2 0,5090909 0.029507 __17.253988 0.0000 Resquared 0.976416 Mcan of dependent var 5.000000 Adjusted R-squared 0.969984 S.D. of dependent var 3.093773 S.E. of regression 0.535998 Sum of squared resid 3.160227 Log likelihood 9.603531 Festatistic 151.8074 Durbin-Watson stat 2.820099 Prob(F-statistic) 0.000000 31 ‘Thus the hypothesis of no structural change would be rejected at the 5 percent level of significance, but not at the 1 percent level. The test of change in the regression slope is based on 3.4902 — 3.1602 _ 1602/11 with Fygs(1,11) = 4.84, Thus the null of a common regression slope is not rejected, Given the assumption of a common regression slope, it is possible to test for common intercepts. The appropriate test statistic is _ 6.5561 — 3.4902 P= Sanna ~ 1054132. ECONoMETRIC METHODS TABLE LS // Dependent Variable is Y VARIABLE COEFFICIENT STD.ERROR —_TSTAT, _2-TAILSIG. c 00625000 04831417 —0.1293616 0.8994 D2 04625000 0.6062146—0.7628312 0.4616 x 0.437500 0.059263 -—=—7.3006283——0.0000 2 0.0715909__0.0667964 LOTI7786 0.3068 R-squared 0.976416 Mean of dependent var 5.000000 Adjusted R-squared «0.969984 S.D. of dependent var. 3.093773 SLE. of regression 0.535998 Sum of squared resid 3.160227 Log likelihood 9.603531 F-statistic 151.8074 Durbin-Watson stat 2.820099 _Prob(F-statistic) 0.000000 LS // Dependent Variable is Y VARIABLE COEFFICIENT STD.ERROR STAT. _2-TAIL SIG. c =0.4658537 03048463 =1,5281589 0.1524 D2 1.0195122 03140167 3.246815 (0.0070 x 0.495120 00266345 __18.589463 0.0000 R-squared 0.973953 Mean of dependent var 5.000000 Adjusted R-squared 0.969612 __S.D. of dependent var 3.093773 SEE. of regression 0.539309 Sum of squared resid 3.490244 Log likelihood ~10.34849 statistic 224.3564 urbin-Watson stat 2.462522 Prob(F-statistic) ‘0.000000 with Foq9(1,12) = 9.33, so the difference in intercepts is significant at the 1 percent level. Table 4.3 illustrates the alternative approach to the same three tests. The first part, of the table gives the fit of the model in Eq. (4,40). Testing the joint significance of the second and fourth variables tests whether the B vectors (intercept and slope) are the same in the two subsets. MicroTSP (a pre-Windows program from Quantitative Micro Software) returns an F statistic for this test of 5.91, the same as that obtained in the preceding paragraph. The same table also provides a test of the equality of the regression slopes. This merely involves testing the significance of the fourth variable. ‘The t statistic is 1.0718, which is clearly insignificant. Squaring the # statistic gives an F value of 1.15 as before. The second part of Table 4.3 reports the results of fitting Eq. (4.42). The hy- pothesis of equal intercepts is tested by examining the significance of the second variable. The f statistic is 3.2467, and its square is 10.54, as before, so the null is rejected. 4.5.6 Extensions There are two major ways in which these tests of structural change may be extended. The first is to split the total sample data into more than two subsets. One mightCHapreR «: Some Tests of the &-Variable Linear Equation for Specification Error 133, examine the stability of a relation across several subperiods, (World War II, Cold War, Post-Cold War), or across countries, industries, social groups, or whatever. The classification of the subsets need not necessarily be time related, nor need the data within each subset be time series data. The same hierarchy of models applies, but now there are p > 2 subvectors in each column. The second extension is concerned with testing for common subsets of parameters. Testing for common intercepts and for common slope vectors are special cases, and we already know how to deal with them. More generally, we may wish to test any subset of the k parameters. The test procedure follows the general principles already established. Fit the restricted model with the subset of coefficiemts whose stability is under test taking the same value in each data subset. Leave all other coefficients free to vary across data subsets, Then fit the completely unrestricted model. The test is based on the contrast between the two residual sums of squares. 46 DUMMY VARIABLES 4.6.1 Introduction We have already encountered dummy variables but have not so labeled them. The Jast nz variables in the augmented matrix in Eq. (4.11) take the form (?] The 0 matrix is of order ny X nz , and I, is the identity matrix of order mp. Each n-vector column is a dummy variable, where a single element is one and the other n— 1 elements are all zero. As shown in Eq. (4.12) the effect of the dummies is to exclude the last n observations from the estimation of the B vector. The coefficients 6f the dummy variables are the forecast errors for the last nz observations, and the regression residuals are zero at these points. Sometimes a single dummy variable of this type is defined for an observation that is thought to be unusual. For example, 1973.4 was the quarter in which the Organization of Petroleum Exporting Countries (OPEC) oil embargo hit. In estimat- ing an energy demand function, one might define a dummy variable with the value of one for this quarter and zero for all other quarters. The effect is shown in the first panel of Fig. 4.7, where a two-variable relation has been assumed for simplicity. The regression line is determined from the observations other than 1973.4. and for that point the regression line shifts to give the actual value of the dependent variable (Fig. 4.7a). Fig. 4.7b shows the case of three such dummy variables. The basic regression line is estimated from the n ~ 3 observations, and it makes three one-period shifts to pass through the three chosen values. A second type of dummy variable takes the form «-[134 BCONOMETRIC METHODS y @ ® ‘Scatter of n> observations. Scatter of m, observations oO FIGURE 4.7 Regressions with dummy variables. as in the models of structural change. The effect of fitting a model such as Eq. (4.42) « is shown in Fig. 4.7c. There are two parallel regression lines, with all n observations being used to estimate the common slope. 4.6.2 Seasonal Dummies In working with, say, quarterly data one may wish to allow for seasonal shifts in a relation, Vacation expenditure will depend on income but may have a positive or negative shift in certain quarters. This requires the specification of quarterly dummy variables, such as, Qu = 1 if observation is in quarter i =0 otherwise for i = 1,..., 4. For the four quarters of each year these dummies are QQ Os Os 1 0 0 0 o 1 0 0 o o 1 0 oo 0 1‘cuapter4: Some Tests of the k-Variable Linear Equation for Specification Exror 135 The relationship might then be written Yr, = Que +0 + @4Qae +x; B + uy (4.43) where x} contains observations on relevant regressors but must not contain an el- ement of one, since a column of ones would be perfectly collinear with the four seasonal dummies, yielding a singular data matrix, The function has four intercepts, denoted by the a’s, An alternative specification is Y= @y + ¥2Qx + ¥30s + Yau +2;8 + 4, (4.44) Comparing coefficients of the dummy variables in the two equations gives y2 = @2— ar ya = a3 — ay V4 = @q— ay The 7's thus measure differential intercepts, by reference to a1. The hypothesis of interest is usually Ho: ay This could be tested by estimating Eq. (4.43) and testing the appropriate linear re~ strictions. Alternatively, the null hypothesis may be expressed as @ = 03 = a4 Ho wey = 7-0 This is easily tested by fitting Eq. (4.44) and testing the joint significance of the three quarterly dummies. The test statistic is invariant to the choice of which quarterly dummy to omit in moving from Eq. (4.43) to Eg. (4.44). 4.6.3 Qualitative Variables Suppose a labor economist postulates an earnings function as Income = (sex, race, educational level, age) The first two explanatory variables are qualitative rather than quantitative, That is, they are not subject to cardinal measurement. They may, however. be represented by dummy variables. Sex is represented by two variables, namely, 5, = Lif male = Ootherwise and S, = Lif female = otherwise The two categories are mutually exclusive and exhaustive. For each individual the sum of S) and S is one. Suppose that race is defined by, say, three mutually exclu- sive and exhaustive categories (for example, Caucasian, Black, and other). Typical entries for the S and R dummy variables would look like136 ECoNoMETRIC METHODS Si Sp Ri Ra Ry ee ea 10 0 1 0 The first entry is for a female Caucasian, and the second is a male Black. The R columns sum to one. as do the § columns. Educational level is a different type of variable. It could be represented in a car- dinal fashion by years of education. Alternatively, it could be expressed in terms of dummy variables. One way is to classify educational level by the highest level of diploma awarded. say. high school diploma, bachelor diploma, or graduate diploma. Individuals with diplomas would have an entry of one for the relevant dummy vari- able and zero for the other two. Although the three categories are mutually exclusive by definition, they are not exhaustive, since there is no category for people without a diploma, Let E be the dummy variable for dropouts, and E, E3, Ey be the dummy variables for the highest diploma awarded. If one modeled income just as a function of educational level, we would have Y = aE) + aE + 383 + agEy tu 4.45) The expected level of income, conditional on educational level, is EV|E)=a i If we suppress £}. an alternative specification is Y = ay + y2k2 + yaks + Yaka + u (4.46) The y's measure the marginal increment in expected income for a diploma over the no-diploma level. The marginal increment for a bachelor diploma over a high school diploma is y3 — y2 and the marginal increment for a graduate diploma over a bachelor diploma is ys — y3. The significance of each marginal increment may be tested by testing the relevant linear restriction. Alternatively, a reformulation of the dummy variables can provide direct estimates of the stepwise marginal increments. _ Let the dummy variable have a value of one if person has the relevant diploma, irrespective of whether he or she has one or more higher diplomas. Also define E, as a pre-high school diploma dummy so that every individual has an entry of one for this dummy. The educational dummies for a person with only a high school diploma would then be [1 100]. and a person with a graduate degree would show as [1 1 1 1]. ‘The equation to be fitted is ¥ = ay + 82E> + 533 + 64s + u (4.47) The expected values are E(¥ | pre-HS diploma) = ay " E(Y | HS diploma) = a + 52 E(Y | bachelor diploma) = a; + 52 + 8 E(¥ | graduate diploma) = a + 6) + 63 + 84 Now the 8’s provide direct estimates of the marginal increment from one level to the next higher level, and the corresponding standard errors provide a direct test of significance.carrer ¢: Some Tests of the k-Variable Linear Equation for Specification Error 137 4.6.4 Two or More Sets of Dummy Variables ‘As we have seen, a mutually exclusive and exhaustive set of dummy variables sums to give the unit vector, i,. Estimating a constant in a relationship is done by inserting a unit vector into the set of regressors. To avoid a singular data matrix the constant must be suppressed if a complete set of dummy variables is used; or, if the con- stant is retained, one of the dummy variables must be dropped. If there are two sets of dummy variables in a relationship and the constant is suppressed the estimation procedure still breaks down, because the included dummy variables are linearly de- pendent (the sum of the first set minus the sum of the second set gives the zero vector). If a constant is retained, one dummy variable must be dropped from each set; and this rule obviously extends to three or more sets. 4.6.5 A Numerical Example Table 4.4 shows some hypothetical data on income (¥), conditional on sex (S), and education (E). For example, three persons in the first category for both sex and edu- cation have incomes of 8, 10, and 12. Inserting a constant and suppressing the first dummy variable in each set specifies the equation to be estimated as Y = p+ ark) + a3 + BrS2 tu (4.48) TABLE 4.4 Ey Ey Si: 810,12 12,14 20,22 6 10,12 20,24 The relevant variables appear in column form as Y & S 8 0 10 12 12 14 20 22 5 6 10 12 20 24 wr coco-Kococom ocr eocconHoce] HHH HHH ccocco] The estimated relationship is P =9+44E, + 13.5E — Sy138 ECONOMETRIC METHODS TABLE 4.5 TABLE 4.6 EE EB Bs S913 S321 S77 ou S58 MB The resultant estimated mean incomes for various combinations of sex and education are shown in Table 4.5. Every specification forces the data to conform to the strait- jacket implicit in the specification. In this case a possibly undesirable feature of the specification is that the marginal increments for education are the same for each sex, and conversely, the sex difference is the same at all educational levels. It is prefer- able to test for this possibility rather than to impose it. The appropriate test is a test for interaction effects, It is carried out by adding two new dummy variables to the specification. These are the products (£252) and (E352). The revised specification is Y = p+ ak) + a3s + BoSp + y2(E2S2) + y(EsS2) tu (4.49) Expected values are now ‘ BQ’ |S, Ei) = a » E(Y | $1, E2) = wt an EW | $1, E3) = w+ a3 E(Y | Sx, E\) = 2+ Bo E(Y | Sz, E2) = w +02 + Prt y2 E(Y | Sz, E3) = wt+ast Brot y3 Fitting Eq. (4.49) gives P = 10 + 3E> + IE; — 4.582 + 2.5(EgSp) + 5.5(E3S2) Table 4.6 shows the resultant estimated mean incomes. The entries in the cells are seen to be the arithmetic means of the raw data in Table 4.4. This special result is due to two factors. First. the only regressors are dummy variables; and allowing for interaction effects gives as many parameters to be estimated as there are cells in the original table. In practice. of course, most specifications contain cardinal regressors in addition to any dummy variables, so this effect would not be observed in general. APPENDIX APPENDIX 4.1 To show var(d) = 07 [In + Xo(XiX1)"1X3] As shown in the text d = uy ~ X2(b, — B)cuarter «: Some Tests of the k-Variable Linear Equation for Specification Error 139 From Eq. (3.23) by - B= (XX) Xm Thus d= uz, ~ XXX) Xiu, Then Ed) =0 and var(d) = E(dd') = E luz ~ X(XjX 1) Xj] [ue — XXX! Xan |! = E(upu) + Xo(X}Xi) 1X) - Elwyus) «Xi (KiKi) XS The cross-product terms vanish since E(uju3) = 0 by assumption. Substituting E(uju}) = oF, then gives the desired result. The variance of a single prediction error in Chapter 3, shown just before Eq. (3.48), isa special case of this result with m= 1 PROBLEMS 44. 42. 43. A bivariate regression is fitted to 20 sample observations on Y and X, where the data are expressed as 10) usesia |) pataaeraadt aa fio 30 xy-[f] wae ‘A new observation is obtained, showing X = 2 and ¥ = 4, Calculate a Chow test for parameter constancy with these data, A four-variable regression using quarterly data from 1958 to 1976 inclusive gave an estimated equation = 2.20 + 0.104X> — 3.48Xy + 0.34X4 The explained sum of squares was 109.6, and the residual sum of squares, 18.48, When. the equation was re-estimated with three seasonal dummies added to the specification, the explained sum of squares rose to 114.8. Test for the presence of seasonality. ‘Two further regressions based on the original specification were run for the sub- periods 1958.1 to 1968.4 and 1969.1 to 1976.4, yielding residual sums of squares of 9.32 and 7.46, respectively. Test for the constancy of the relationship over the two sub- periods. Gasoline sales in a regional market were modeled by the following regression equation, estimated with quarterly dat Q = 10 - 0.01P + 0.2¥ ~ 1.55) + 3.65; + 4.78; where Q is sales, P is price, ¥ is disposable income, and the S; are quarterly dummy variables. The expected paths of P and Y for the next year are as follows: Quirter 1234 Po 062214 ¥ 100102104103 Calculate the sales of gasoline to be expected in each quarter of the year.140 EcoNomerRic METHODS Suppose another researcher proposes to use the same data to estimate an equation of the same form, except that she wishes to employ dummy variables So, Ss, and Ss. Write down the equation that will come from her calculations. Yet another investigator proposes to suppress the intercept and use all four seasonal dummies. Write down the results of his estimation, 44, The model ¥ = ay + yok + y3ks tu is estimated by OLS, where E2 and £3 are dummy variables indicating membership of the second and third educational classes. Show that the OLS estimates are - a yi | 7 li a al Ws where 7, denotes the mean value of ¥ in the ith educational class. 4.5. Using the data in Table 4.4 estimate the specification Y = aE + an; + a3Es + Bada + u and estimate the resultant mean values for Table 4.5. Compare your results with the values given in the text and comment, Repeat the exercise for the specification. Y = pt aE) + ak; + BS +4 46. Using the data of Table 4.4 estimate a specification of your own choosing without a constant term but with appropriate dummy variables to allow for interaction effects. Calculate the resultant version of Table 4.6 and compare with the results in the text, 47. Survey records for a sample of 12 families show the following weekly consumption expenditures (Y) and weekly incomes (X): —_— Y 1 7 9 100 105 113 122 120 146 135 147 155 X 80 95 105 1S 135 «4515S 165, S850 Families with an asterisk (*) reported that their income is higher than in the previous year, Using a linear consumption function, test whether the consumption behavior of families experiencing an increase in income is different from that of families who did not experience an increase. 48. Annual data for 1929-1967 (39 observations) are obtained for Q = anindex of U.S. GDP in constant dollars L = an index of labor input K = an index of capital input ‘A production function is estimated for the whole period ascuarter 4. Some Tests of the k-Variable Linear Equation for Specification Error 141 log Q = —3,8766 + 1.4106 log L + 0.4162 log K with R? = 0.9937 and s = 0.03755. Regressions for two subperiods yield 1929-1948 log Q = ~4.0576 + 1.6167 log L. + 0.2197 log K with R? = 0.9759 and s = 0.04573, 1949-1967 : log @ = - 1.9564 + 0.8336 10g L + 0.6631 log K with R? = 0,9904 and s = 0.02185 Test for the stability of the production function over the two subperiods. 4.9. The usual two-variable linear model is postulated, and a sample of 20 observations is drawn from an urban area and another sample of 10 observations from a rural area. The sample information in raw form is summarized xs follows: Urban 20 20] ry _ [10 20. 25. 7 -| xx-[ 50] yy = 30 Rural 1 10 10 las lio A xy By eee [ool yy ‘Test the hypothesis that the same relationship holds in both urban and rural areas, 4.10. A study of vacation expenditures in relation to income was based on data for 256 house- holds, which were grouped into three separate income classes. Log linear regressions (with an intercept term) were computed for each income group and for all households with the following results: Regression Residual Number of slope variance households Low income 0.02 0.26 102 Middle income 0.09 0.42 102 High income 0.14 0.30 52 Al households 0.07 0.38 256 Test whether the expenditure function is the same for all income groups. ‘What additional information would you need to test whether the expenditure elas- ticity is the same across income groups? Given that the variance of the log of income in the complete sample is 24, test the hypothesis that the expenditure elasticity for all households is 0.10.CHAPTER 5 Maximum Likelihood (ML), Generalized Least Squares (GLS), and Instrumental Variable (IV) Estimators The maximum likelihood principle was introduced in Chapter 2, Now is the time to give a more comprehensive treatment. 5.1 MAXIMUM LIKELIHOOD ESTIMATORS In recent years there has been a rapid development of new econometric tests, vari- ously based on the Wald and Lagrange multiplier approaches. This has also led to a resurgence of interest in the maximum likelihood approach. Let y’ = [y: yx -*+ yn] be an n-vector of sample values, dependent on some k-vector of unknown parameters, 8” = 8, 02 ++ 04). Let the joint den- sity be written f(y:0). which indicates the dependence on 8. This density may be interpreted in two different ways. For a given 8 it indicates the probability of a set of sample outcomes. Alternatively, it may be interpreted as a function of 8, conditional ona set of sample outcomes. In the latter interpretation itis referred to as a likelihood function. The formal definition is Likelihood function = L(0;y) = f(y:8) (5.1) It is customary to reverse the order of the symbols in writing the likelihood function to emphasize the new focus of interest. Maximizing the likelihood function with respect to 6 amounts to finding a specific value. say 8, that maximizes the probability of obtaining the sample values that have actually been observed. Then 6 is said to be the maximum likelihood estimator (MLE) of the unknown parameter vector 8. In most applications it is simpler to maximize the log of the likelihood function. We will denote the log-likelihood by 142charters: ML, GLS, and IV Estimators 143 a lal 3 La and the 6 that maximizes / will also maximize L. The derivative of / with respect to @ is known as the score, s(@:y). The MLE, 6, is obtained by setting the score to zero, that is, by finding the value of 6 that solves a 0 Then s(0;y) = 6.2) The widespread use of maximum likelihood estimators is largely due to a range of desirable properties, which are summarized in the next section. 5.1.1 Properties of Maximum Likelihood Estimators The major properties of MLEs are large-sample, or asymptotic, ones. They hold under fairly general conditions. 1. Consistency plim(O) = @ 2. Asymptotic normality 62£NO1'(0)) ‘This states that the asymptotic distribution of @ is normal with mean @ and vari- ance given by the inverse of 1(@). 1(8) is the information matrix and is defined in two equivalent ways by . al\/at\ Fl 10) = e((3\(%)| ~~ 5535: 63) In practice it is usually much easier to evaluate the second expression, When @ is a K-vector, 41/48 denotes a column vector of k partial derivatives, that is, 31/30 al _ | 01/06 a7 61/305 Each element in this score (or gradient) vector is itself a function of @, and so may be differentiated partially with respect to each element in @. For example, aaa) fl LL 38, 30? 50,883 0,50, where the second-order derivatives have been written as a row vector. Proceeding in this way yields a square, symmetric matrix of second-order derivatives, known.144. ECONOMETRIC METHODS as the Hessian matrix, a fl a 30% 000 0000, #t et el 7! | 56208, 32808; 5030" : ; ; #t Fw #l 80,50; 50,082 303 Notice that this is nor the same matrix as (41/40)(91/0)'. The latter is also a square, symmetric k X k matrix, but its i,jth element is the product (01/86,) (61/88). 3. Asymptotic efficiency. If 6 is the maximum likelihood estimator of a single parameter @, the previous property means that Jn - 0) NO,0?) for some finite constant o”. If 6 denotes any other consistent, asymptotically nor- mal estimator of @, then ,/n6 has anormal limiting distribution whose variance is, greater than or equal to o*. The MLE has minimum variance in the class of consis- tent, asymptotically normal estimators. The term asymptotic variance refers to the variance of a limiting distribution. Thus the asymptotic variance of //n0 is o?. However. the term is also used to describe the variance of the asymptotic approx- imation to the unknown finite sample distribution. Thus an equivalent statement is that the asymptotic variance of 6 is o/n. When @ is a vector of parameters and 6 is the MLE, nd - 0)-4 NO.) for some positive definite matrix V. If V denotes the variance matrix of any other lent, asymptotically normal estimator, then V— V is a positive semidefinite matrix. 4. Invariance. If 6 is the MLE of @ and g(@) is a continuous function of @, then g() is the MLE of g(9), 5. The score has zero mean and variance /(). ‘To demonstrate the zero mean ‘we note that integrating the joint density over all possible values of y gives a value of one, that is, “dyn = [| Lay = 1 Differentiating both sides with respect to @ yields aL [-[a 0‘cuarrers: ML, GLS, and [V Estimators 145 But Es) -{ : {fee : aL ~[--[%e =0 Ceca It then follows that the variance of s is var(s) = Biss’) = #[(0)(%)| = 16) 5.2 ML ESTIMATION OF THE LINEAR MODEL This section covers the maximum likelihood estimation of the linear model, which comprises many of the econometric applications. The equation is y=XB+u with ~ u~ No) The multivaciate normal density for w is 1 (Ro? \u'u) 1) ~ Gagan The multivariate density for y conditional on X is then FX) = foa| where |(du/dy)| is the absolute value of the determinant formed from the n X n matrix of partial derivatives of the elements of w with respect to the elements of y.! Here this matrix is simply the identity matrix. Thus the log-likelihood function is 1=Infy|X) = nfw) = ~5ln2n— 5 Ino? = shu ~fin2e 5 Ino? = Lg ~ XB)(y — XB) (5.4) The vector of unknown parameters, @, has k + | elements, namely, = [B'.0"] Taking partial derivatives gives 'See Appendix 5.1 on the change of variables in density functions.146. ECONOMETRIC METHODS a op al nd ; do? ~ ~2g2 * 280 ~ XB ~ XB) Setting these partial derivatives to zero gives the MLEs as B= (xx) X'y (5.5) and & = y— XY — Xpyn 6.6) ‘The MLE. B. is seen to be the OLS estimator b, and 6? is e’e/n where e = y — Xbis the vector of OLs residuals.? We know from least- squares theory that E(e'el(n — k)) = a. Thus E(6?) = o°(n — k/n, so that 6 is biased for 0, though B is unbiased for 8. The second-order derivatives are L yy 4 ye ~ gal Xy + XXB) Al XX yi acre el xX 7 apop'}~ “a a with == eeu) =0 Fl n uw i fl \_on Tey Ie oe With ~£ (525) * oF since E(u'u) = na. The information matrix is 16) = (8) = 1 ys pk” 0 0 n 2K! 0 and its inverse is r (8) = [’ 2] " Oia ‘The zero off-diagonal terms indicate that B and G? are distributed independently of ‘one another. Substituting the MLE values from Eqs. (5.5) and (5.6) in the log-likelihood func- tion, Eq. (5.4), and exponentiating gives the maximum of the likelihood function as UB, 6) = (2mey- (62? 2 =) (ey 67) = constant - (e’e)"" ; where the constant does not depend on any of the parameters of the model. 2Do not confuse ¢ or its elements with the mathematical constant e = 2.71828,‘cuarrers: ML, GLS, and IV Estimators 147 53 7 LIKELIHOOD RATIO, WALD, AND LAGRANGE MULTIPLIER TESTS We will illustrate these tests in the context of linear hypotheses about B. A general linear hypothesis takes the form Ho: RB =r (68) where R is aq X k (q < k) matrix of known constants and r aq X 1 known vector, ‘The main tests are LR, W, and LM tests. 5.3.1 Likelihood Ratio (LR) Tests The MLEs in Eqs. (5.5) and (5.6) maximize the likelihood function without impos- ing any restrictions. The resultant value of L(B, 6?) in Eq. (5.7) is the unrestricted maximum likelihood and is expressible as a function of the unrestricted residual swm of squares, e’e. The model may also be estimated in restricted form by maximizing the likelihood subject to the restrictions, RB = r. Let the resultant estimators be denoted by B and G*. The relevant maximum of the likelihood is then obtained by substituting these values in the likelihood function to get L(B, 6). The restricted maximum cannot exceed the unrestricted maximum, but if the restrictions are valid one would expect the restricted maximum to be “close” to the unrestricted maximum. The likelihood ratio is defined as LB, 6?) LB, 6) and intuitively we expect to reject the null hypothesis if A is “small.” In some cases exact finite-sample tests of the “smallness” of A can be derived for some special transformations of A. However, a large-sample test of general applicability is avail- able in that R = ~2Ind = 2nL(B, 6?) - in LB, e147) (5.9) The restricted MLEs are derived by maximizing I =1-p (RB -r) (5.10) where yt is ag X 1 vector of Lagrange multipliers. and / is the log-likelihood spec- ified in Eq, (5.4). It can be shown that is simply the restricted b. vector already derived in the standard least squares analysis (see Eq. (3.43)]. This vector satisfies the constraints Rb. = r. If we denote the corresponding residuals by 7 e. = y— Xb. eles/n, and so the restricted MLE of o? is 6? = 1B, 67) = constant - (e,e.)-"™ (5.11)148 ECONOMETRIC METHODS Substitution of Eqs. (5.7) and (5.11) into Eq. (5.9) gives the LR test statistic as LR = n(Inele. — Ine’e). For future reference we note some alternative forms of the LR statistic, namely, LR n(Ineje, ~ Ine’e) mn(t ewer ee =) ee 6.12) ae (; — ee. wane) ‘The calculation of the LR statistic thus requires the fitting of both the restricted and the unrestricted model. 5.3.2 The Wald (W) Test In the Wald procedure only the unrestricted is calculated. The vector (RB — r) then indicates the extent to which the unrestricted ML estimates fit the null hy- pothesis. This vector being “close” to zero would tend to support the null hy- pothesis: “large” values would tend to contradict it. Since B is asymptotically normally distributed with mean vector B and variance-covariance matrix I~'(B) it follows that, under Ho, (RB ~ r) is asymptotically distributed as multivariate normal with zero mean vector and variance-covariance matrix RI~'(B)R’, where I-'(B) = 0°(X'X)"'. As shown earlier, the information matrix for the linear re- gression model is block diagonal, so we can concentrate on the submatrix relating to B. It then follows that® (RB ~ ry (RE"(B)R'T "RB - 1) 2 79) (5.13) where q is the number of restrictions in R. The asymptotic distribution still holds when the unknown o” in J-!(B) is replaced by a consistent estimator 6? = e’e/n. ‘The result is the Wald statistic, Rate ig . w= BB PRUX RT RB 1) 2 214) 6.14) et It was shown earlier [see Eq. (3.44)] that the numerator in Eq. (5.14) can also be expressed as (e,e, — e'e), so an alternative form of the Wald statistic for testing (5.8) is w= Meee) ¢ ig) 6.15) ee 2As seen in Chapter 3, this statistic also has an exact, finite sample x(q) distribution when the distur- bances are normally distributed.charter s: ML, GLS, and IV Estimators 149 5.3.3 Lagrange Multiplier (LM) Test The LM test, also known as the score test, is based on the score (or gradient) vector, ainL — al s() 0 0 The unrestricted estimator, 6, is found by solving s(@) = 0, where the notation s(0) indicates the score vector evaluared at 8. When the score vector is evaluated at 8, the restricted estimator, it general not be zero. However, if the restrictions are valid, the restricted maximum, 1(6), should be close to the unrestricted maximum, 1(8), and so the gradient at the former should be close to zero, As shown earlier, the score vector has zero mean and variance-covariance matrix given by the information matrix, [(@). The quadratic form, s'(@)f'(8)s(@), will then have a y? distribution. Evaluating this quadratic form at 8 = 6 provides a test of the null hypothesis. The basic result is that, under the null hypothesis, LM = s'(6)I '(8)s(6) ~ x(q) (5.16) Notice that each element in Eq. (5.16) is evaluated at 8. In contrast to the Wald test, we now need calculate only the restricted estimator. The popularity of LM tests is due to the fact that in many cases it is much easier to calculate the restricted estimator than the unrestricted estimator. From the development leading up to Eqs. (5.5) and (5.6) the score vector is al 1 = SXu = -|Bl-| @ sO)=|F l=) wwf me a 207 * 264 To evaluate the score vector at the restricted estimator , we replace u by és = a XB and 0? by 5? = e/e./n. The B vector satisfies RB = r. Thus, ral ‘The inverse of the information matrix was given before. Evaluating this at 0 gives 5(6) = a er 0 T(@) = 26* eeoare Substitution in Eq, (5.16) then gives, waxy! 0 I Ly, 1X fy ; Pall aa Ml" o n um =| & " eXW'X) Xe. “oo ne,X(X'X)"1X' eles i (5.17)150 ECONoMETRIC METHODS By recalling the expressions for the explained and total sum of squares from a mul- tiple regression in Chapter 3, the LM statistic in Eq. (5.17) is LM = nR* where R? is the squared multiple correlation coefficient from the regression of e. on X. Ife, does not have zero mean, R” is the uncentered R?. However, if the restrictions involve only the slope coefficients, 2, 83. .... Bx. as is usually the case, then ex will have zero mean and the R? in this expression is the centered statistic from the conventional regression packages.t The LM test can therefore be implemented in two steps. First compute the restricted estimator B and obtain the resultant residual vector es. Then regress e. on X and refer nR? from this regression to 7(q). This two-step procedure occurs frequently in cases where maximization of the likelihood is equivalent to minimization of a sum of squares. It may be shown’ that Eq. (5.17) may be rewritten as LM = Mees =e’) (5.18) eles ‘We can now illustrate the famous inequality for these three test statistics in the linear model. namely, W = LR = LM. Applying the first two terms of the logarithmic expansion In(1 + z) = z— 32° +--+ to the second expression for LR in Eq. (5.12) gives te. —e! Facer LR = Meee mee) _ 3 (ts: “) ee 2\ ee "which yields LR = W. Similarly, using the third expression in Eq. (5.12) gives -nin(i - £828) ee. Fi / et en —e'e) i = Mest — ee) , (ees — ee 2\ ele. ele. 2 so that LR = LM and finally W = LR = LM. The tests are asymptotically equiva- lent but in general will give different numerical results in finite samples EXAMPLE 5.1. We return to the data of Example 3.3 and test Ho: By = 0 by these asymptotic tests. From Table 3.2 we see that the unrestricted regression of ¥ on Xz and Xa givese’e = 1.5; and the restricted regression, when X; is excluded, givesele. = 24, Substitution in Eqs. (5.15), (5.12), and (5.18) in that order gives W = 5(24- 1.5/1.5 = 3.00 LR = 51n(2.4/1.5) = 235 : i LM = 5(24~ 1.59/24 = 1.875 “See Appendix 5.2 for an explanation of centered and uncentered F°. $See Appendix 5.3.cuapters) ML, GLS, and IV Estimators 151 The 5 percent point of x*(1) is 3.841, so the null hypothesis is not rejected by any of the tests, although not much can be expected from the use of asymptotic tests on such a small sample. When the residual series from the restricted regression of ¥ on X3 1s used as the dependent variable in a regression on X2 and X3, the resultant R° is 0.375. giving nR? = 5(0.375) = 1.875, in agreement with the value of the preceding LM statistic. 5.4 ML ESTIMATION OF THE LINEAR MODEL WITH NONSPHERICAL DISTURBANCES The postulated model is now y=XB+u with u~N(@00°A) (5.19) where O is a positive definite matrix of order n. This model is referred to as the case of nonspherical disturbances, compared with var(#) = 071, which is the case of spherical disturbances. For the present the elements of © will be assumed to be known, For example, if the disturbance variance at each sample point is proportional to the square of one of the regressors, say, X2, we have var(u)) = 0? = 0X3, §=1,2...,0 where o? is a scaling factor. The variance-covariance matrix of the disturbance is then 0 0 Oo x}, 7 var(u) = 07 2 = ofdiagiX3, Xb XB} 0 0 x __ From Eq, (5.19), the multivariate normal density for w is f(a) = 2) 0? A? exp[— tuo?) a} Noting that |o?Q| = o?"|Q|, we may rewrite the density as Fa), = ayo?) |O/-? expl(- 1/207 u'O"u] ‘The log-likelihood is then n ae : eae - gy —xpya-'y- = inn) gine 7 iniQ| qi XpyQ"'y-XB) (5.20) Differentiating with respect to B and a? gives OPE eee 5B 7 GiX'O ty XO xB) abe tine eee ee eee eet and Jot = ~3g2 + Fgi ~ XBYQ'y - XB) Setting the partial derivatives to zero gives the ML estimators