Lecture 18: Major Challenges for Consistency Measurement Error, Omitted Variables Bias, Simultaneous Causality, and Lagged Dependent Variables (Chapter 8.2, 13.1, 13.3) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. Agenda • Stochastic X ’s • Omitted Variables Bias (Chapter 8.2, 13.3) • Measurement Error (Chapter 13.1) • Simultaneous Causality (Chapter 13.3) • Using Lagged Values of the Dependent Variable as Explanators (Chapter 13.3) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-2 Review: Fixed X ’s • The Gauss–Markov DGP Yi 0 1 X i i E ( i ) 0 Var ( i ) 2 Cov( i , j ) 0 if i j X ’s fixed across samples OLS is BLUE. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-3 Review: Fixed X ’s (cont.) xiYi xiYi ˆ E ( 1 ) E 2 E 2 xi xi xi E 2 [ 0 1 X i i ] xi xi xi i xi 2 0 2 1 X i E 2 xi xi xi xi xi X i xi E ( i ) 0 1 0 1 0 2 2 2 xi xi xi Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-4 Stochastic X ’s • The assumption that X ’s are fixed across samples is very strong. • We are assuming that the process that generates the i ’s is completely separate from the process that generates the Xi ’s. • What happens to the expectation of OLS if we do not make this assumption? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-5 Stochastic X ’s (cont.) xiYi xiYi ˆ E ( 1 ) E 2 E 2 xi xi xi E 2 [ 0 1 X i i ] xi xi xi i xi 2 0 2 1 X i E 2 xi xi xi xi i 0 1 E 2 xi Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-6 Stochastic X ’s (cont.) • Consistency is an easier goal. • We need to assume 2 1 2 p lim xi Q, where Q is a non-zero constant. n E(xi i ) 0 which, by the Law of Large Numbers, 1 implies p lim xi i 0. n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-7 Stochastic X ’s (cont.) xiYi xi [ 0 1 X 1 i ] ˆ p lim( 1 ) p lim 2 p lim 2 x x i i 1 n xi i xi i 1 p lim 1 p lim 2 1 xi xi2 n 1 p lim xi i 0 n 1 1 1 Q 1 2 p lim xi n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-8 Stochastic X ’s (cont.) • Asking that key variances be non-zero and bounded is often reasonable (except in certain time series cases). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-9 Stochastic X ’s (cont.) • How reasonable an assumption is E(Xii ) 0 (i.e. Xi and i are uncorrelated)? • In practice, this assumption is often terrible. • When an explanator is correlated with the error term, we call the explanator a “troublesome variable.” Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-10 Stochastic X ’s (cont.) • This lecture outlines 4 major causes of correlation between Xi and i • One or more of these problems is at least a plausible concern in most applications. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-11 Stochastic X ’s (cont.) 1. Omitted Variables Bias 2. Measurement Error 3. Simultaneous Causality 4. Using Lagged Values of the Dependent Variable as Explanators, in the presence of serial correlation • Any of these conditions will make OLS inconsistent (and biased). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-12 Omitted Variables Bias • We have already learned that omitting an explanator of Y from the regression can create an Omitted Variables Bias in OLS. • The omitted explanator must be correlated with an included explanator. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-13 Omitted Variables Bias (cont.) • The DGP includes X2i. Yi 0 1 X 1i 2 X 2i i • Y depends on both X1 and X2 , but we regress Y on only X1. Yi 0 1 X 1i i Is E (ˆ 1) 1 ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-14 Omitted Variables Bias (cont.) E (ˆ1 ) E ( wiYi ) E ( wiYi ) wi E (Yi ) wi E ( 0 1 X 1i 2 X 2i 1 ) 0 wi 1 wi X 1i 2 wi X 1i E ( i ) 1 2 wi X 2i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-15 Omitted Variables Bias (cont.) • We can reconceptualize OVB as case of concurrent correlation between X and the error term: xii 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-16 Omitted Variables Bias (cont.) When Y 0 1 X1 2 X 2 but we instead regress Y 0 1 X1 then = 2 X 2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-17 Omitted Variables Bias (cont.) Y 0 1 X1 2 X 2 Cov( X1 , ) Cov( X1 , 2 X 2 ) Cov( X1 , 2 X 2 ) Cov( X1 , ) 2Cov( X1 , X 2 ) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-18 Omitted Variables Bias (cont.) • X is correlated with • Contemporaneous correlation implies inconsistency. • Using probability limits, we can re-derive the formula for omitted variables bias and OLS. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-19 Omitted Variables Bias (cont.) 1 p lim xii n p lim(ˆ1 ) 1 1 2 p lim xi n E ( xii ) Cov ( X 1 , ) 1 1 2 E ( xi ) Var ( X 1 ) Cov( X 1 , X 2 ) 1 2 Var ( X 1 ) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-20 Omitted Variables Bias (cont.) Cov( X 1 , X 2 ) p lim(ˆ1 ) 1 2 Var ( X 1 ) Cov( X 1 , X 2 ) Note: is the OLS formula Var ( X 1 ) for 1 in the regression X 2 0 1 X 1 v Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-21 Omitted Variables Bias (cont.) • Example: heart attacks and treatments • Do cardiac catheterization, revascularization, and other intensive medical procedures reduce mortality among the elderly? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-22 Omitted Variables Bias (cont.) • Patients who receive more intensive cardiac treatment are on average younger, urban, white males. • These traits may account for the lower mortality rates among the intensively treated. • We can easily include these observable variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-23 Omitted Variables Bias (cont.) Mortalityi 0 1 D IntensivelyTreated i 2 D Female i ... kUrbani i • If these observables were the only differences between patients who received intensive treatment and those who did not, then DIntensivelyTreated would not tell us anything about . OLS would be consistent. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-24 Omitted Variables Bias (cont.) • However, we reasonably believe that a doctor’s choice to perform intensive cardiac procedures are correlated with many other variables. • In general, whenever an explanator is actively chosen by an economic agent, we worry a great deal about selection biases. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-25 Omitted Variables Bias (cont.) • Doctors might select patients to receive treatment based on their underlying health status. • If health status is an unobservable determinant of mortality, then it is a component of . • OLS will give an inconsistent estimate of the benefits of intensive treatment. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-26 Omitted Variables Bias (cont.) • OLS will give an inconsistent estimate of the benefits of intensive treatment. • The extent of the bias depends on the extent to which health status affects mortality and the extent to which intensive treatment predicts health status. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-27 Measurement Error (Chapter 13.1) • Another source of correlation between X and is the mis-measurement of X (also called “errors in variables”). • To the extent that the X we observe differs from the X we are modeling, our coefficients will be wrong. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-28 Measurement Error (cont.) • Another source of correlation between X and is the mis-measurement of X. • Measurement error might arise because: – The data are flawed (for example, survey respondents mis-remember how long they’ve worked at their current jobs) – The economist is using an imperfect proxy (for example, using total taxable income as a measure of total income) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-29 Measurement Error (cont.) • Another source of correlation between X and is the mis-measurement of X. • Measurement error might arise because: – The economist does not properly understand the nature of X (for example, focusing on changes in annual income instead of changes in total lifetime income, the relevant variable in Friedman’s “permanent income hypothesis”) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-30 Measurement Error (cont.) We model Yi 0 1 X i i but instead of observing X i we observe M i X i vi where vi is some error E(vi ) 0, Cov( X i ,vi ) 0 We regress Yi 0 1 M i i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-31 Measurement Error (cont.) We model Yi 0 1 X i i We regress Yi 0 1 M i i Substituting, our regression equation is equivalent to 0 1 ( X i vi ) i 0 1 X i 1vi i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-32 Measurement Error (cont.) We model Yi 0 1 X i i We regress Yi 0 1M i i which is, in essence, 0 1 X i 1vi i ˆ1 has to perform two jobs. It estimates the coefficient on X i (1 ) but it ALSO estimates the coefficient on vi . Since vi does not appear in our model, its true coefficient is 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-33 Measurement Error (cont.) In essence, we regress Yi 0 1 X i 1vi i ˆ1 has to perform two jobs. It estimates the coefficient on X i ( 1 ) but it ALSO estimates the coefficient on vi . Since vi does not appear in our model, its true coefficient is 0. ˆ1 will be a weighted average of 1 and 0. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-34 Measurement Error (cont.) • How can measurement error be conceptualized as a correlation between X and ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-35 Measurement Error (cont.) Our model is Yi 0 1 X i i Instead of X i , we observe M i X i i Yi 0 1 ( M i vi ) i 0 1 M i 1vi i 0 1 M i i where i 1vi i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-36 Measurement Error (cont.) Our model is Yi 0 1 X i i Instead of X i , we observe M i X i vi We regress Yi 0 1M i i where i 1vi i Cov( M i ,i ) Cov( X i vi , 1vi i ) Cov(vi , 1vi ) 1 Cov(vi , vi ) 1 Var (vi ) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-37 Measurement Error (cont.) • Measurement error implies a correlation between our observed explanator M i Xi vi and the error term i -1vi i • What bias does this correlation create in OLS? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-38 DGP with Measurement Error Yi 0 1 X i i E( i ) 0 Var( i ) 2 Cov( i , j ) 0 for i j E( X i i ) 0, M i X i vi 1 (xi 2 ) 2X n E(vi ) 0 Var(vi ) v2 Cov(vi ,v j ) 0 for i j Cov(vi , X i ) 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-39 Measurement Error Regress Yi 0 1M i i Is ˆ1 a consistent estimator of 1 ? Yi mi p lim(ˆ1 ) p lim where mi M i M 2 m j ( 0 1M i 1vi i )mi p lim m j 2 1 p lim ( 0 1M i 1vi i )mi n 1 p lim ( x j v j ) 2 n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-40 Measurement Error (cont.) 1 p lim ( 0 1 ( X i vi ) 1vi i )mi n p lim(ˆ1 ) 1 p lim ( x j v j ) 2 n 1 1 1 p lim 0 mi 1 X i mi mi i n n n 1 2 1 2 1 p lim x j v j xi vi n n n Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-41 Measurement Error (cont.) 1 1 2 1 1 p lim 0 mi 1 xi xi vi mi i n n n n p lim(ˆ1 ) 1 2 1 2 1 p lim x j v j xi vi n n n Applying the Law of Large Numbers.... 1Var ( X i ) 1Cov( X i , vi ) Cov( xi vi , i ) Var ( X i ) Var (vi ) Cov( X i , vi ) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-42 Measurement Error (cont.) p lim(ˆ1 ) 1Var ( X i ) 1Cov( X i , vi ) Cov( xi vi , i ) Var ( X i ) Var (vi ) Cov( X i , vi ) Var ( X i ) 1 Var ( X i ) Var (vi ) 1 X2 X2 v2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-43 Measurement Error (cont.) p lim(ˆ1 1 ) 1 X2 2 X 2 v 1 v2 2 1 2 X v v2 Notice that 2 1 2 X v The bias will always pull ˆ1 PARTWAY from 1 to 0. We call this bias ATTENUATION BIAS. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-44 Measurement Error (cont.) • Mismeasuring X leads to ATTENUATION BIAS. The estimated coefficient is biased towards 0. • The magnitude of the bias depends on the relative variances of X and v. • A small amount of random measurement noise will not bias the estimate very much. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-45 Measurement Error (cont.) • Mismeasuring X leads to ATTENUATION BIAS. The estimated coefficient is biased towards 0. • Note: mismeasuring Y does NOT lead to measurement error bias (though it does increase the variance of the error term, thus increasing standard errors). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-46 Checking Understanding (cont.) • Suppose you are advising policy makers on the effect of a one-time tax rebate on consumption. Using a cross section of 50,000 households, you regress reported current consumption against reported current income. • Your results suggest that the tax rebate in question will NOT have a large enough impact on consumption to justify the policy. (continued on next slide) Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-47 Checking Understanding (cont.) • Suppose your results found that the marginal propensity to consume was too small to justify the tax rebate. • A proponent of the measure argues that you should be regressing consumption against Friedman’s “permanent income,” not current income. The resulting measurement error renders your results irrelevant. • Assess this argument. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-48 Checking Understanding (cont.) • Answer: If your regression suffers from measurement error, then the true effect of the tax rebate on consumption is larger than what you found. • The coefficient with attenuation bias is too small to justify the policy. The larger, true coefficient might or might not be large enough to support the tax rebate. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-49 Simultaneous Causality (Chapter 13.3) • Another common source of correlation between Xi and i is SIMULTANEOUS CAUSALITY • This complication is also called JOINTLY DETERMINED VARIABLES or ENDOGENEITY Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-50 Simultaneous Causality (cont.) • Both X and Y are jointly determined • The process that generates Y also generates X at the same time • Because X and Y are determined simultaneously, X can adjust in response to shocks to Y () • Thus X will be correlated with Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-51 Simultaneous Causality (cont.) • The classic example of simultaneous causality in economics is supply and demand. • Both prices and quantities adjust until supply and demand are in equilibrium. • A shock to demand or supply causes BOTH prices and quantities to move. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-52 Simultaneous Causality (cont.) • Thus, any attempt to estimate the relationship between prices and quantities (say, to estimate a demand elasticity) suffers from SIMULTANEITY BIAS. • Econometricians have a frequent interest in estimating elasticities resulting from such an equilibrium process. Simultaneity bias is a MAJOR problem. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-53 Simultaneous Causality (cont.) • For example, consider the market for wheat. • The quantity demanded for wheat is a function of the price consumers pay and the income of the population: QiD 0 1Pi D 2 I i iD • i indexes separate markets Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-54 Simultaneous Causality (cont.) • The quantity of wheat supplied is a function of the price suppliers receive and the weather (which affects crop yields). Q 0 1Pi 2Wi S i S Copyright © 2006 Pearson Addison-Wesley. All rights reserved. S i 18-55 Simultaneous Causality (cont.) S D Q Q • In equilibrium, and Pi Pi S i D i • Let’s focus on the demand equation. Is PiD correlated with iD ? Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-56 Simultaneous Causality (cont.) Q 0 1Pi 2Wi S i S S i QiD 0 1Pi D 2 I i iD • Suppose iD > 0 (there is a positive shock to demand). This shock makes QiD greater than usual. In equilibrium, QiD = QiS • To balance the supply equation, PiS must increase. Suppliers must be paid a higher price to supply the greater demanded quantity. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-57 Simultaneous Causality (cont.) QiS 0 1Pi S 2Wi iS QiD 0 1Pi D 2 I i iD • In equilibrium, PiS = PiD. The consumers must pay a higher price to enjoy the higher quantity of wheat they demand • Thus, a positive shock to iD induces a higher PiD Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-58 Simultaneous Causality (cont.) E(Pi • ) 0 D D i • A positive demand shock increases the quantity demanded. In order to increase supply, the price must go up. The demand shock and the price are correlated. • OLS will be inconsistent. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-59 Simultaneous Causality (cont.) • When we have a system of equations (as with supply and demand), all the variables that are jointly determined are called endogenous variables. Price and quantity are endogenous variables. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-60 Simultaneous Causality (cont.) • Variables that are determined outside the system of equations are called exogenous variables. The weather is an exogenous variable. In partial equilibrium (such as the supply and demand for wheat), the population’s income is also exogenous. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-61 Simultaneous Causality (cont.) • It is arbitrary which endogenous variables we write on the left-hand side. • We could write both equations with either Price or Quantity on the lefthand side. • For convenience, let us use Price on the LHS of the Supply equation and Quantity on the LHS of the Demand equation. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-62 Simultaneous Causality (cont.) (1) Q 0 1 Pi 2 I i D i D D i (2) Pi S 0 1QiS 2Wi iS D determines QiD . i QiD Qi S , so Cov( iD ,QiS ) 0 Q determines Pi S i S Pi Pi , so Cov(Pi ,Q ) 0 S D D S i Therefore, Cov(Pi D , iD ) 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-63 Simultaneous Causality (cont.) (1) Q 0 1 Pi 2 I i D i D D i (2) Pi 0 1Q 2Wi S S i Q Q Pi Pi D i D i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. S i S S i D 18-64 Lagged Dependent Variables (Chapter 13.3) • Using lagged dependent variables as explanators is another potential source of correlation between an explanator and the error term. • For example, you try to predict next period’s inflation as a function of this period’s inflation. Inft Inft 1 t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-65 Lagged Dependent Variables (cont.) • Lagged dependent variables present a problem in the presence of serial correlation. • Example: suppose there is first order serial correlation: Inft Inft 1 t t t 1 vt 0 | | 1 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-66 Lagged Dependent Variables (cont.) Inft Inft 1 t Substituting in t t 1 vt : Inft Inft 1 t 1 vt However, Inft 1 Inft 2 t 1 t 1 is a determinant of BOTH t and Yt 1 Cov(Inft 1 , t ) 0 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-67 Lagged Dependent Variables (cont.) • Including lagged dependent variables as an explanator does NOT lead to inconsistency in the absence of firstorder serial correlation. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-68 Review • Instead of assuming fixed X ’s, we can instead assume E(Xii ) 0 (i.e. Xi and i are uncorrelated). • In practice, this assumption is often terrible. • When an explanator is correlated with the error term, we call the explanator a “troublesome variable.” Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-69 Review (cont.) • This lecture outlined 4 major causes of correlation between Xi and i • One or more of these problems is at least a plausible concern in most applications. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-70 Review (cont.) 1. Omitted Variables Bias 2. Measurement Error 3. Simultaneous Causality 4. Using Lagged Values of the Dependent Variable as Explanators (in the presence of serial correlation) • Any of these conditions will make OLS inconsistent (and biased). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-71 Review (cont.) • Omitting a variable creates a bias only if: 1. X2 is an explanator of Y (so, when omitted, it becomes a component of the error term) 2. X2 is correlated with X1 (so that X2 creates a correlation between X1 and the error term). Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-72 Review (cont.) Yi 0 1 X 1i 2 X 2i i Regress: Yi 0 1 X 1i i Cov( X 1 , X 2 ) p lim(ˆ1 ) 1 2 Var ( X 1 ) Cov( X 1 , X 2 ) Note: is the OLS formula Var ( X 1 ) for 1 in the regression X 2 0 1 X 1 v Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-73 Review (cont.) • Measurement error also induces a correlation between our included explanator and the error term. • Instead of observing Xi , we observe M i Xi vi Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-74 Review (cont.) • Measurement error implies a correlation between our observed explanator M i Xi vi and the error term, i -1vi i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-75 Review (cont.) Instead of observing X i , we observe M i X i vi Var ( X i ) X2 p lim(ˆ1) 1 1 2 Var ( X i ) Var (vi ) X v2 Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-76 Review (cont.) • Mismeasuring X leads to ATTENUATION BIAS. The estimated coefficient is biased towards 0. • The magnitude of the bias depends on the relative variances of X and v. Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-77 Review (cont.) • Under simultaneous causality, Q and P are jointly determined • Because Q and P are determined simultaneously, P can adjust in response to shocks to Q ( D) • Thus P will be correlated with D Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-78 Simultaneous Causality (cont.) (1) QiD 0 1 Pi D 2 I i iD (2) Pi S 0 1QiS 2Wi iS Q Q Pi Pi D i D i Copyright © 2006 Pearson Addison-Wesley. All rights reserved. S i S D 18-79 Review (cont.) • Using lagged dependent variables as explanators is another potential source of correlation between an explanator and the error term. • If there is first-order serial correlation in the error terms, then t-1 is correlated with t • Therefore Yt-1 is correlated with t Copyright © 2006 Pearson Addison-Wesley. All rights reserved. 18-80